2020-08-16 18:49:28 -07:00
|
|
|
//===- IRModules.cpp - IR Submodules of pybind module ---------------------===//
|
|
|
|
|
//
|
|
|
|
|
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
|
|
|
|
|
// See https://llvm.org/LICENSE.txt for license information.
|
|
|
|
|
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
|
|
|
|
|
//
|
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
#include <optional>
|
|
|
|
|
#include <utility>
|
2024-12-18 19:31:32 +00:00
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
#include "Globals.h"
|
|
|
|
|
#include "IRModule.h"
|
|
|
|
|
#include "NanobindUtils.h"
|
2020-12-03 17:22:57 -08:00
|
|
|
#include "mlir-c/BuiltinAttributes.h"
|
2021-04-19 13:37:01 +02:00
|
|
|
#include "mlir-c/Debug.h"
|
[mlir][python] Capture error diagnostics in exceptions
This updates most (all?) error-diagnostic-emitting python APIs to
capture error diagnostics and include them in the raised exception's
message:
```
>>> Operation.parse('"arith.addi"() : () -> ()'))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
mlir._mlir_libs.MLIRError: Unable to parse operation assembly:
error: "-":1:1: 'arith.addi' op requires one result
note: "-":1:1: see current operation: "arith.addi"() : () -> ()
```
The diagnostic information is available on the exception for users who
may want to customize the error message:
```
>>> try:
... Operation.parse('"arith.addi"() : () -> ()')
... except MLIRError as e:
... print(e.message)
... print(e.error_diagnostics)
... print(e.error_diagnostics[0].message)
...
Unable to parse operation assembly
[<mlir._mlir_libs._mlir.ir.DiagnosticInfo object at 0x7fed32bd6b70>]
'arith.addi' op requires one result
```
Error diagnostics captured in exceptions aren't propagated to diagnostic
handlers, to avoid double-reporting of errors. The context-level
`emit_error_diagnostics` option can be used to revert to the old
behaviour, causing error diagnostics to be reported to handlers instead
of as part of exceptions.
API changes:
- `Operation.verify` now raises an exception on verification failure,
instead of returning `false`
- The exception raised by the following methods has been changed to
`MLIRError`:
- `PassManager.run`
- `{Module,Operation,Type,Attribute}.parse`
- `{RankedTensorType,UnrankedTensorType}.get`
- `{MemRefType,UnrankedMemRefType}.get`
- `VectorType.get`
- `FloatAttr.get`
closes #60595
depends on D144804, D143830
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D143869
2023-02-07 16:07:50 -05:00
|
|
|
#include "mlir-c/Diagnostics.h"
|
2021-04-22 00:07:30 -06:00
|
|
|
#include "mlir-c/IR.h"
|
2023-04-30 22:11:02 -07:00
|
|
|
#include "mlir-c/Support.h"
|
2024-12-20 23:32:32 -05:00
|
|
|
#include "mlir/Bindings/Python/Nanobind.h"
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
#include "mlir/Bindings/Python/NanobindAdaptors.h"
|
2024-12-20 23:32:32 -05:00
|
|
|
#include "mlir-c/Bindings/Python/Interop.h" // This is expected after nanobind.
|
2021-10-13 10:02:12 +02:00
|
|
|
#include "llvm/ADT/ArrayRef.h"
|
2020-09-02 04:40:28 +00:00
|
|
|
#include "llvm/ADT/SmallVector.h"
|
2020-08-18 17:23:46 -07:00
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
namespace nb = nanobind;
|
|
|
|
|
using namespace nb::literals;
|
2020-08-18 17:23:46 -07:00
|
|
|
using namespace mlir;
|
2020-08-16 20:53:45 -07:00
|
|
|
using namespace mlir::python;
|
|
|
|
|
|
2020-09-02 04:40:28 +00:00
|
|
|
using llvm::SmallVector;
|
2020-11-29 13:52:11 -08:00
|
|
|
using llvm::StringRef;
|
|
|
|
|
using llvm::Twine;
|
2020-09-02 04:40:28 +00:00
|
|
|
|
2020-08-16 20:53:45 -07:00
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
// Docstrings (trivial, non-duplicated docstrings are included inline).
|
|
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
|
2020-10-31 23:40:25 -07:00
|
|
|
static const char kContextParseTypeDocstring[] =
|
|
|
|
|
R"(Parses the assembly form of a type.
|
2020-09-20 21:25:46 -07:00
|
|
|
|
[mlir][python] Capture error diagnostics in exceptions
This updates most (all?) error-diagnostic-emitting python APIs to
capture error diagnostics and include them in the raised exception's
message:
```
>>> Operation.parse('"arith.addi"() : () -> ()'))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
mlir._mlir_libs.MLIRError: Unable to parse operation assembly:
error: "-":1:1: 'arith.addi' op requires one result
note: "-":1:1: see current operation: "arith.addi"() : () -> ()
```
The diagnostic information is available on the exception for users who
may want to customize the error message:
```
>>> try:
... Operation.parse('"arith.addi"() : () -> ()')
... except MLIRError as e:
... print(e.message)
... print(e.error_diagnostics)
... print(e.error_diagnostics[0].message)
...
Unable to parse operation assembly
[<mlir._mlir_libs._mlir.ir.DiagnosticInfo object at 0x7fed32bd6b70>]
'arith.addi' op requires one result
```
Error diagnostics captured in exceptions aren't propagated to diagnostic
handlers, to avoid double-reporting of errors. The context-level
`emit_error_diagnostics` option can be used to revert to the old
behaviour, causing error diagnostics to be reported to handlers instead
of as part of exceptions.
API changes:
- `Operation.verify` now raises an exception on verification failure,
instead of returning `false`
- The exception raised by the following methods has been changed to
`MLIRError`:
- `PassManager.run`
- `{Module,Operation,Type,Attribute}.parse`
- `{RankedTensorType,UnrankedTensorType}.get`
- `{MemRefType,UnrankedMemRefType}.get`
- `VectorType.get`
- `FloatAttr.get`
closes #60595
depends on D144804, D143830
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D143869
2023-02-07 16:07:50 -05:00
|
|
|
Returns a Type object or raises an MLIRError if the type cannot be parsed.
|
2020-09-20 21:25:46 -07:00
|
|
|
|
2020-10-31 23:40:25 -07:00
|
|
|
See also: https://mlir.llvm.org/docs/LangRef/#type-system
|
2020-09-20 21:25:46 -07:00
|
|
|
)";
|
|
|
|
|
|
2021-10-13 10:02:12 +02:00
|
|
|
static const char kContextGetCallSiteLocationDocstring[] =
|
|
|
|
|
R"(Gets a Location representing a caller and callsite)";
|
|
|
|
|
|
2020-10-31 23:40:25 -07:00
|
|
|
static const char kContextGetFileLocationDocstring[] =
|
|
|
|
|
R"(Gets a Location representing a file, line and column)";
|
|
|
|
|
|
2021-12-11 10:12:29 -08:00
|
|
|
static const char kContextGetFusedLocationDocstring[] =
|
|
|
|
|
R"(Gets a Location representing a fused location with optional metadata)";
|
|
|
|
|
|
2021-10-12 12:45:57 -07:00
|
|
|
static const char kContextGetNameLocationDocString[] =
|
|
|
|
|
R"(Gets a Location representing a named location with optional child location)";
|
|
|
|
|
|
2020-10-31 23:40:25 -07:00
|
|
|
static const char kModuleParseDocstring[] =
|
2020-08-16 20:53:45 -07:00
|
|
|
R"(Parses a module's assembly format from a string.
|
|
|
|
|
|
[mlir][python] Capture error diagnostics in exceptions
This updates most (all?) error-diagnostic-emitting python APIs to
capture error diagnostics and include them in the raised exception's
message:
```
>>> Operation.parse('"arith.addi"() : () -> ()'))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
mlir._mlir_libs.MLIRError: Unable to parse operation assembly:
error: "-":1:1: 'arith.addi' op requires one result
note: "-":1:1: see current operation: "arith.addi"() : () -> ()
```
The diagnostic information is available on the exception for users who
may want to customize the error message:
```
>>> try:
... Operation.parse('"arith.addi"() : () -> ()')
... except MLIRError as e:
... print(e.message)
... print(e.error_diagnostics)
... print(e.error_diagnostics[0].message)
...
Unable to parse operation assembly
[<mlir._mlir_libs._mlir.ir.DiagnosticInfo object at 0x7fed32bd6b70>]
'arith.addi' op requires one result
```
Error diagnostics captured in exceptions aren't propagated to diagnostic
handlers, to avoid double-reporting of errors. The context-level
`emit_error_diagnostics` option can be used to revert to the old
behaviour, causing error diagnostics to be reported to handlers instead
of as part of exceptions.
API changes:
- `Operation.verify` now raises an exception on verification failure,
instead of returning `false`
- The exception raised by the following methods has been changed to
`MLIRError`:
- `PassManager.run`
- `{Module,Operation,Type,Attribute}.parse`
- `{RankedTensorType,UnrankedTensorType}.get`
- `{MemRefType,UnrankedMemRefType}.get`
- `VectorType.get`
- `FloatAttr.get`
closes #60595
depends on D144804, D143830
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D143869
2023-02-07 16:07:50 -05:00
|
|
|
Returns a new MlirModule or raises an MLIRError if the parsing fails.
|
2020-08-18 17:23:46 -07:00
|
|
|
|
|
|
|
|
See also: https://mlir.llvm.org/docs/LangRef/
|
|
|
|
|
)";
|
|
|
|
|
|
2020-10-31 23:40:25 -07:00
|
|
|
static const char kOperationCreateDocstring[] =
|
|
|
|
|
R"(Creates a new operation.
|
2020-08-18 17:23:46 -07:00
|
|
|
|
2020-10-31 23:40:25 -07:00
|
|
|
Args:
|
|
|
|
|
name: Operation name (e.g. "dialect.operation").
|
|
|
|
|
results: Sequence of Type representing op result types.
|
|
|
|
|
attributes: Dict of str:Attribute.
|
|
|
|
|
successors: List of Block for the operation's successors.
|
|
|
|
|
regions: Number of regions to create.
|
|
|
|
|
location: A Location object (defaults to resolve from context manager).
|
|
|
|
|
ip: An InsertionPoint (defaults to resolve from context manager or set to
|
|
|
|
|
False to disable insertion, even with an insertion point set in the
|
|
|
|
|
context manager).
|
2023-07-23 21:26:52 -07:00
|
|
|
infer_type: Whether to infer result types.
|
2020-10-31 23:40:25 -07:00
|
|
|
Returns:
|
|
|
|
|
A new "detached" Operation object. Detached operations can be added
|
|
|
|
|
to blocks, which causes them to become "attached."
|
2020-08-16 20:53:45 -07:00
|
|
|
)";
|
|
|
|
|
|
2020-10-20 23:20:04 -07:00
|
|
|
static const char kOperationPrintDocstring[] =
|
|
|
|
|
R"(Prints the assembly form of the operation to a file like object.
|
|
|
|
|
|
|
|
|
|
Args:
|
|
|
|
|
file: The file like object to write to. Defaults to sys.stdout.
|
|
|
|
|
binary: Whether to write bytes (True) or str (False). Defaults to False.
|
|
|
|
|
large_elements_limit: Whether to elide elements attributes above this
|
|
|
|
|
number of elements. Defaults to None (no limit).
|
|
|
|
|
enable_debug_info: Whether to print debug/location information. Defaults
|
|
|
|
|
to False.
|
|
|
|
|
pretty_debug_info: Whether to format debug information for easier reading
|
|
|
|
|
by a human (warning: the result is unparseable).
|
|
|
|
|
print_generic_op_form: Whether to print the generic assembly forms of all
|
|
|
|
|
ops. Defaults to False.
|
|
|
|
|
use_local_Scope: Whether to print in a way that is more optimized for
|
|
|
|
|
multi-threaded access but may not be consistent with how the overall
|
|
|
|
|
module prints.
|
2021-11-28 15:33:03 -08:00
|
|
|
assume_verified: By default, if not printing generic form, the verifier
|
|
|
|
|
will be run and if it fails, generic form will be printed with a comment
|
|
|
|
|
about failed verification. While a reasonable default for interactive use,
|
|
|
|
|
for systematic use, it is often better for the caller to verify explicitly
|
|
|
|
|
and report failures in a more robust fashion. Set this to True if doing this
|
|
|
|
|
in order to avoid running a redundant verification. If the IR is actually
|
|
|
|
|
invalid, behavior is undefined.
|
2024-06-20 17:15:08 +02:00
|
|
|
skip_regions: Whether to skip printing regions. Defaults to False.
|
2020-10-20 23:20:04 -07:00
|
|
|
)";
|
|
|
|
|
|
2023-11-13 10:21:21 -08:00
|
|
|
static const char kOperationPrintStateDocstring[] =
|
|
|
|
|
R"(Prints the assembly form of the operation to a file like object.
|
|
|
|
|
|
|
|
|
|
Args:
|
|
|
|
|
file: The file like object to write to. Defaults to sys.stdout.
|
|
|
|
|
binary: Whether to write bytes (True) or str (False). Defaults to False.
|
|
|
|
|
state: AsmState capturing the operation numbering and flags.
|
|
|
|
|
)";
|
|
|
|
|
|
2020-10-20 23:20:04 -07:00
|
|
|
static const char kOperationGetAsmDocstring[] =
|
|
|
|
|
R"(Gets the assembly form of the operation with all options available.
|
|
|
|
|
|
|
|
|
|
Args:
|
|
|
|
|
binary: Whether to return a bytes (True) or str (False) object. Defaults to
|
|
|
|
|
False.
|
|
|
|
|
... others ...: See the print() method for common keyword arguments for
|
|
|
|
|
configuring the printout.
|
|
|
|
|
Returns:
|
|
|
|
|
Either a bytes or str object, depending on the setting of the 'binary'
|
|
|
|
|
argument.
|
|
|
|
|
)";
|
|
|
|
|
|
2022-09-05 11:54:19 +00:00
|
|
|
static const char kOperationPrintBytecodeDocstring[] =
|
|
|
|
|
R"(Write the bytecode form of the operation to a file like object.
|
|
|
|
|
|
|
|
|
|
Args:
|
|
|
|
|
file: The file like object to write to.
|
2023-04-29 05:35:53 -07:00
|
|
|
desired_version: The version of bytecode to emit.
|
|
|
|
|
Returns:
|
|
|
|
|
The bytecode writer status.
|
2022-09-05 11:54:19 +00:00
|
|
|
)";
|
|
|
|
|
|
2020-08-16 20:53:45 -07:00
|
|
|
static const char kOperationStrDunderDocstring[] =
|
2020-10-20 23:20:04 -07:00
|
|
|
R"(Gets the assembly form of the operation with default options.
|
2020-08-16 20:53:45 -07:00
|
|
|
|
|
|
|
|
If more advanced control over the assembly formatting or I/O options is needed,
|
2020-10-20 23:20:04 -07:00
|
|
|
use the dedicated print or get_asm method, which supports keyword arguments to
|
|
|
|
|
customize behavior.
|
2020-08-16 20:53:45 -07:00
|
|
|
)";
|
|
|
|
|
|
|
|
|
|
static const char kDumpDocstring[] =
|
|
|
|
|
R"(Dumps a debug representation of the object to stderr.)";
|
|
|
|
|
|
2020-09-20 21:25:46 -07:00
|
|
|
static const char kAppendBlockDocstring[] =
|
|
|
|
|
R"(Appends a new block, with argument types as positional args.
|
|
|
|
|
|
|
|
|
|
Returns:
|
|
|
|
|
The created block.
|
|
|
|
|
)";
|
|
|
|
|
|
2020-10-20 11:21:05 +02:00
|
|
|
static const char kValueDunderStrDocstring[] =
|
|
|
|
|
R"(Returns the string form of the value.
|
|
|
|
|
|
|
|
|
|
If the value is a block argument, this is the assembly form of its type and the
|
|
|
|
|
position in the argument list. If the value is an operation result, this is
|
|
|
|
|
equivalent to printing the operation that produced it.
|
|
|
|
|
)";
|
|
|
|
|
|
2023-05-07 18:19:46 -05:00
|
|
|
static const char kGetNameAsOperand[] =
|
|
|
|
|
R"(Returns the string form of value as an operand (i.e., the ValueID).
|
|
|
|
|
)";
|
|
|
|
|
|
2023-04-26 09:55:27 -05:00
|
|
|
static const char kValueReplaceAllUsesWithDocstring[] =
|
|
|
|
|
R"(Replace all uses of value with the new value, updating anything in
|
|
|
|
|
the IR that uses 'self' to use the other value instead.
|
|
|
|
|
)";
|
|
|
|
|
|
[mlir,python] Expose replaceAllUsesExcept to Python bindings (#115850)
Problem originally described in [the forums
here](https://discourse.llvm.org/t/mlir-python-expose-replaceallusesexcept/83068/1).
Using the MLIR Python bindings, the method
[`replaceAllUsesWith`](https://mlir.llvm.org/doxygen/classmlir_1_1Value.html#ac56b0fdb6246bcf7fa1805ba0eb71aa2)
for `Value` is exposed, e.g.,
```python
orig_value.replace_all_uses_with(
new_value
)
```
However, in my use-case I am separating a block into multiple blocks, so
thus want to exclude certain Operations from having their Values
replaced (since I want them to diverge).
Within Value, we have
[`replaceAllUsesExcept`](https://mlir.llvm.org/doxygen/classmlir_1_1Value.html#a9ec8d5c61f8a6aada4062f609372cce4),
where we can pass the Operations which should be skipped.
This is not currently exposed in the Python bindings: this PR fixes
this. Adds `replace_all_uses_except`, which works with individual
Operations, and lists of Operations.
2024-11-20 00:00:57 +00:00
|
|
|
static const char kValueReplaceAllUsesExceptDocstring[] =
|
|
|
|
|
R"("Replace all uses of this value with the 'with' value, except for those
|
|
|
|
|
in 'exceptions'. 'exceptions' can be either a single operation or a list of
|
|
|
|
|
operations.
|
|
|
|
|
)";
|
|
|
|
|
|
2020-09-02 04:40:28 +00:00
|
|
|
//------------------------------------------------------------------------------
|
2020-10-21 23:34:01 -07:00
|
|
|
// Utilities.
|
2020-09-02 04:40:28 +00:00
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
|
2021-04-19 13:37:01 +02:00
|
|
|
/// Helper for creating an @classmethod.
|
2021-01-18 11:27:19 -08:00
|
|
|
template <class Func, typename... Args>
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object classmethod(Func f, Args... args) {
|
|
|
|
|
nb::object cf = nb::cpp_function(f, args...);
|
|
|
|
|
return nb::borrow<nb::object>((PyClassMethod_New(cf.ptr())));
|
2021-01-18 11:27:19 -08:00
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
static nb::object
|
2020-10-21 23:34:01 -07:00
|
|
|
createCustomDialectWrapper(const std::string &dialectNamespace,
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object dialectDescriptor) {
|
2020-10-21 23:34:01 -07:00
|
|
|
auto dialectClass = PyGlobals::get().lookupDialectClass(dialectNamespace);
|
|
|
|
|
if (!dialectClass) {
|
|
|
|
|
// Use the base class.
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
return nb::cast(PyDialect(std::move(dialectDescriptor)));
|
2020-10-21 23:34:01 -07:00
|
|
|
}
|
2020-09-02 04:40:28 +00:00
|
|
|
|
2020-10-21 23:34:01 -07:00
|
|
|
// Create the custom implementation.
|
|
|
|
|
return (*dialectClass)(std::move(dialectDescriptor));
|
|
|
|
|
}
|
2020-11-24 18:35:22 +00:00
|
|
|
|
|
|
|
|
static MlirStringRef toMlirStringRef(const std::string &s) {
|
|
|
|
|
return mlirStringRefCreate(s.data(), s.size());
|
|
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
static MlirStringRef toMlirStringRef(const nb::bytes &s) {
|
|
|
|
|
return mlirStringRefCreate(static_cast<const char *>(s.data()), s.size());
|
|
|
|
|
}
|
|
|
|
|
|
2023-05-24 21:51:36 -04:00
|
|
|
/// Create a block, using the current location context if no locations are
|
|
|
|
|
/// specified.
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
static MlirBlock createBlock(const nb::sequence &pyArgTypes,
|
|
|
|
|
const std::optional<nb::sequence> &pyArgLocs) {
|
2023-05-24 21:51:36 -04:00
|
|
|
SmallVector<MlirType> argTypes;
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
argTypes.reserve(nb::len(pyArgTypes));
|
2023-05-24 21:51:36 -04:00
|
|
|
for (const auto &pyType : pyArgTypes)
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
argTypes.push_back(nb::cast<PyType &>(pyType));
|
2023-05-24 21:51:36 -04:00
|
|
|
|
|
|
|
|
SmallVector<MlirLocation> argLocs;
|
|
|
|
|
if (pyArgLocs) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
argLocs.reserve(nb::len(*pyArgLocs));
|
2023-05-24 21:51:36 -04:00
|
|
|
for (const auto &pyLoc : *pyArgLocs)
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
argLocs.push_back(nb::cast<PyLocation &>(pyLoc));
|
2023-05-24 21:51:36 -04:00
|
|
|
} else if (!argTypes.empty()) {
|
|
|
|
|
argLocs.assign(argTypes.size(), DefaultingPyLocation::resolve());
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (argTypes.size() != argLocs.size())
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error(("Expected " + Twine(argTypes.size()) +
|
2023-05-24 21:51:36 -04:00
|
|
|
" locations, got: " + Twine(argLocs.size()))
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.str()
|
|
|
|
|
.c_str());
|
2023-05-24 21:51:36 -04:00
|
|
|
return mlirBlockCreate(argTypes.size(), argTypes.data(), argLocs.data());
|
|
|
|
|
}
|
|
|
|
|
|
2021-04-19 13:37:01 +02:00
|
|
|
/// Wrapper for the global LLVM debugging flag.
|
|
|
|
|
struct PyGlobalDebugFlag {
|
2025-01-12 18:30:42 +00:00
|
|
|
static void set(nb::object &o, bool enable) { mlirEnableGlobalDebug(enable); }
|
2021-04-19 13:37:01 +02:00
|
|
|
|
2025-01-12 18:30:42 +00:00
|
|
|
static bool get(const nb::object &) { return mlirIsGlobalDebugEnabled(); }
|
2021-04-19 13:37:01 +02:00
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
static void bind(nb::module_ &m) {
|
2021-04-19 13:37:01 +02:00
|
|
|
// Debug flags.
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::class_<PyGlobalDebugFlag>(m, "_GlobalDebug")
|
|
|
|
|
.def_prop_rw_static("flag", &PyGlobalDebugFlag::get,
|
|
|
|
|
&PyGlobalDebugFlag::set, "LLVM-wide debug flag")
|
2024-05-24 23:15:18 +02:00
|
|
|
.def_static(
|
|
|
|
|
"set_types",
|
|
|
|
|
[](const std::string &type) {
|
|
|
|
|
mlirSetGlobalDebugType(type.c_str());
|
|
|
|
|
},
|
|
|
|
|
"types"_a, "Sets specific debug types to be produced by LLVM")
|
|
|
|
|
.def_static("set_types", [](const std::vector<std::string> &types) {
|
|
|
|
|
std::vector<const char *> pointers;
|
|
|
|
|
pointers.reserve(types.size());
|
|
|
|
|
for (const std::string &str : types)
|
|
|
|
|
pointers.push_back(str.c_str());
|
|
|
|
|
mlirSetGlobalDebugTypes(pointers.data(), pointers.size());
|
|
|
|
|
});
|
2021-04-19 13:37:01 +02:00
|
|
|
}
|
|
|
|
|
};
|
|
|
|
|
|
2022-12-21 16:22:39 -08:00
|
|
|
struct PyAttrBuilderMap {
|
|
|
|
|
static bool dunderContains(const std::string &attributeKind) {
|
|
|
|
|
return PyGlobals::get().lookupAttributeBuilder(attributeKind).has_value();
|
|
|
|
|
}
|
2025-01-07 16:33:01 +01:00
|
|
|
static nb::callable dunderGetItemNamed(const std::string &attributeKind) {
|
2022-12-21 16:22:39 -08:00
|
|
|
auto builder = PyGlobals::get().lookupAttributeBuilder(attributeKind);
|
|
|
|
|
if (!builder)
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::key_error(attributeKind.c_str());
|
2022-12-21 16:22:39 -08:00
|
|
|
return *builder;
|
|
|
|
|
}
|
2025-01-07 16:33:01 +01:00
|
|
|
static void dunderSetItemNamed(const std::string &attributeKind,
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::callable func, bool replace) {
|
[mlir][python bindings] generate all the enums
This PR implements python enum bindings for *all* the enums - this includes `I*Attrs` (including positional/bit) and `Dialect/EnumAttr`.
There are a few parts to this:
1. CMake: a small addition to `declare_mlir_dialect_python_bindings` and `declare_mlir_dialect_extension_python_bindings` to generate the enum, a boolean arg `GEN_ENUM_BINDINGS` to make it opt-in (even though it works for basically all of the dialects), and an optional `GEN_ENUM_BINDINGS_TD_FILE` for handling corner cases.
2. EnumPythonBindingGen.cpp: there are two weedy aspects here that took investigation:
1. If an enum attribute is not a `Dialect/EnumAttr` then the `EnumAttrInfo` record is canonical, as far as both the cases of the enum **and the `AttrDefName`**. On the otherhand, if an enum is a `Dialect/EnumAttr` then the `EnumAttr` record has the correct `AttrDefName` ("load bearing", i.e., populates `ods.ir.AttributeBuilder('<NAME>')`) but its `enum` field contains the cases, which is an instance of `EnumAttrInfo`. The solution is to generate an one enum class for both `Dialect/EnumAttr` and "independent" `EnumAttrInfo` but to make that class interopable with two builder registrations that both do the right thing (see next sub-bullet).
2. Because we don't have a good connection to cpp `EnumAttr`, i.e., only the `enum class` getters are exposed (like `DimensionAttr::get(Dimension value)`), we have to resort to parsing e.g., `Attribute.parse(f'#gpu<dim {x}>')`. This means that the set of supported `assemblyFormat`s (for the enum) is fixed at compile of MLIR (currently 2, the only 2 I saw). There might be some things that could be done here but they would require quite a bit more C API work to support generically (e.g., casting ints to enum cases and binding all the getters or going generically through the `symbolize*` methods, like `symbolizeDimension(uint32_t)` or `symbolizeDimension(StringRef)`).
A few small changes:
1. In addition, since this patch registers default builders for attributes where people might've had their own builders already written, I added a `replace` param to `AttributeBuilder.insert` (`False` by default).
2. `makePythonEnumCaseName` can't handle all the different ways in which people write their enum cases, e.g., `llvm.CConv.Intel_OCL_BI`, which gets turned into `INTEL_O_C_L_B_I` (because `llvm::convertToSnakeFromCamelCase` doesn't look for runs of caps). So I dropped it. On the otherhand regularization does need to done because some enums have `None` as a case (and others might have other python keywords).
3. I turned on `llvm` dialect generation here in order to test `nvvm.WGMMAScaleIn`, which is an enum with [[ https://github.com/llvm/llvm-project/blob/d7e26b56207cbd8995296c5bb7c11ce676b649da/mlir/include/mlir/IR/EnumAttr.td#L22-L25 | no explicit discriminator ]] for the `neg` case.
Note, dialects that didn't get a `GEN_ENUM_BINDINGS` don't have any enums to generate.
Let me know if I should add more tests (the three trivial ones I added exercise both the supported `assemblyFormat`s and `replace=True`).
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D157934
2023-08-23 13:27:08 -05:00
|
|
|
PyGlobals::get().registerAttributeBuilder(attributeKind, std::move(func),
|
|
|
|
|
replace);
|
2022-12-21 16:22:39 -08:00
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
static void bind(nb::module_ &m) {
|
|
|
|
|
nb::class_<PyAttrBuilderMap>(m, "AttrBuilder")
|
2022-12-21 16:22:39 -08:00
|
|
|
.def_static("contains", &PyAttrBuilderMap::dunderContains)
|
2025-01-07 16:33:01 +01:00
|
|
|
.def_static("get", &PyAttrBuilderMap::dunderGetItemNamed)
|
|
|
|
|
.def_static("insert", &PyAttrBuilderMap::dunderSetItemNamed,
|
[mlir][python bindings] generate all the enums
This PR implements python enum bindings for *all* the enums - this includes `I*Attrs` (including positional/bit) and `Dialect/EnumAttr`.
There are a few parts to this:
1. CMake: a small addition to `declare_mlir_dialect_python_bindings` and `declare_mlir_dialect_extension_python_bindings` to generate the enum, a boolean arg `GEN_ENUM_BINDINGS` to make it opt-in (even though it works for basically all of the dialects), and an optional `GEN_ENUM_BINDINGS_TD_FILE` for handling corner cases.
2. EnumPythonBindingGen.cpp: there are two weedy aspects here that took investigation:
1. If an enum attribute is not a `Dialect/EnumAttr` then the `EnumAttrInfo` record is canonical, as far as both the cases of the enum **and the `AttrDefName`**. On the otherhand, if an enum is a `Dialect/EnumAttr` then the `EnumAttr` record has the correct `AttrDefName` ("load bearing", i.e., populates `ods.ir.AttributeBuilder('<NAME>')`) but its `enum` field contains the cases, which is an instance of `EnumAttrInfo`. The solution is to generate an one enum class for both `Dialect/EnumAttr` and "independent" `EnumAttrInfo` but to make that class interopable with two builder registrations that both do the right thing (see next sub-bullet).
2. Because we don't have a good connection to cpp `EnumAttr`, i.e., only the `enum class` getters are exposed (like `DimensionAttr::get(Dimension value)`), we have to resort to parsing e.g., `Attribute.parse(f'#gpu<dim {x}>')`. This means that the set of supported `assemblyFormat`s (for the enum) is fixed at compile of MLIR (currently 2, the only 2 I saw). There might be some things that could be done here but they would require quite a bit more C API work to support generically (e.g., casting ints to enum cases and binding all the getters or going generically through the `symbolize*` methods, like `symbolizeDimension(uint32_t)` or `symbolizeDimension(StringRef)`).
A few small changes:
1. In addition, since this patch registers default builders for attributes where people might've had their own builders already written, I added a `replace` param to `AttributeBuilder.insert` (`False` by default).
2. `makePythonEnumCaseName` can't handle all the different ways in which people write their enum cases, e.g., `llvm.CConv.Intel_OCL_BI`, which gets turned into `INTEL_O_C_L_B_I` (because `llvm::convertToSnakeFromCamelCase` doesn't look for runs of caps). So I dropped it. On the otherhand regularization does need to done because some enums have `None` as a case (and others might have other python keywords).
3. I turned on `llvm` dialect generation here in order to test `nvvm.WGMMAScaleIn`, which is an enum with [[ https://github.com/llvm/llvm-project/blob/d7e26b56207cbd8995296c5bb7c11ce676b649da/mlir/include/mlir/IR/EnumAttr.td#L22-L25 | no explicit discriminator ]] for the `neg` case.
Note, dialects that didn't get a `GEN_ENUM_BINDINGS` don't have any enums to generate.
Let me know if I should add more tests (the three trivial ones I added exercise both the supported `assemblyFormat`s and `replace=True`).
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D157934
2023-08-23 13:27:08 -05:00
|
|
|
"attribute_kind"_a, "attr_builder"_a, "replace"_a = false,
|
|
|
|
|
"Register an attribute builder for building MLIR "
|
|
|
|
|
"attributes from python values.");
|
2022-12-21 16:22:39 -08:00
|
|
|
}
|
|
|
|
|
};
|
|
|
|
|
|
2023-07-12 22:11:14 -07:00
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
// PyBlock
|
|
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object PyBlock::getCapsule() {
|
|
|
|
|
return nb::steal<nb::object>(mlirPythonBlockToCapsule(get()));
|
2023-07-12 22:11:14 -07:00
|
|
|
}
|
|
|
|
|
|
2020-09-19 22:02:32 -07:00
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
// Collections.
|
|
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
|
|
|
|
|
namespace {
|
|
|
|
|
|
|
|
|
|
class PyRegionIterator {
|
|
|
|
|
public:
|
|
|
|
|
PyRegionIterator(PyOperationRef operation)
|
|
|
|
|
: operation(std::move(operation)) {}
|
|
|
|
|
|
|
|
|
|
PyRegionIterator &dunderIter() { return *this; }
|
|
|
|
|
|
|
|
|
|
PyRegion dunderNext() {
|
|
|
|
|
operation->checkValid();
|
|
|
|
|
if (nextIndex >= mlirOperationGetNumRegions(operation->get())) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::stop_iteration();
|
2020-09-19 22:02:32 -07:00
|
|
|
}
|
|
|
|
|
MlirRegion region = mlirOperationGetRegion(operation->get(), nextIndex++);
|
|
|
|
|
return PyRegion(operation, region);
|
|
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
static void bind(nb::module_ &m) {
|
|
|
|
|
nb::class_<PyRegionIterator>(m, "RegionIterator")
|
2020-09-19 22:02:32 -07:00
|
|
|
.def("__iter__", &PyRegionIterator::dunderIter)
|
|
|
|
|
.def("__next__", &PyRegionIterator::dunderNext);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
private:
|
|
|
|
|
PyOperationRef operation;
|
|
|
|
|
int nextIndex = 0;
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
/// Regions of an op are fixed length and indexed numerically so are represented
|
|
|
|
|
/// with a sequence-like container.
|
|
|
|
|
class PyRegionList {
|
|
|
|
|
public:
|
|
|
|
|
PyRegionList(PyOperationRef operation) : operation(std::move(operation)) {}
|
|
|
|
|
|
2023-05-24 22:05:06 -04:00
|
|
|
PyRegionIterator dunderIter() {
|
|
|
|
|
operation->checkValid();
|
|
|
|
|
return PyRegionIterator(operation);
|
|
|
|
|
}
|
|
|
|
|
|
2020-09-19 22:02:32 -07:00
|
|
|
intptr_t dunderLen() {
|
|
|
|
|
operation->checkValid();
|
|
|
|
|
return mlirOperationGetNumRegions(operation->get());
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
PyRegion dunderGetItem(intptr_t index) {
|
|
|
|
|
// dunderLen checks validity.
|
|
|
|
|
if (index < 0 || index >= dunderLen()) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::index_error("attempt to access out of bounds region");
|
2020-09-19 22:02:32 -07:00
|
|
|
}
|
|
|
|
|
MlirRegion region = mlirOperationGetRegion(operation->get(), index);
|
|
|
|
|
return PyRegion(operation, region);
|
|
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
static void bind(nb::module_ &m) {
|
|
|
|
|
nb::class_<PyRegionList>(m, "RegionSequence")
|
2020-09-19 22:02:32 -07:00
|
|
|
.def("__len__", &PyRegionList::dunderLen)
|
2023-05-24 22:05:06 -04:00
|
|
|
.def("__iter__", &PyRegionList::dunderIter)
|
2020-09-19 22:02:32 -07:00
|
|
|
.def("__getitem__", &PyRegionList::dunderGetItem);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
private:
|
|
|
|
|
PyOperationRef operation;
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
class PyBlockIterator {
|
|
|
|
|
public:
|
|
|
|
|
PyBlockIterator(PyOperationRef operation, MlirBlock next)
|
|
|
|
|
: operation(std::move(operation)), next(next) {}
|
|
|
|
|
|
|
|
|
|
PyBlockIterator &dunderIter() { return *this; }
|
|
|
|
|
|
|
|
|
|
PyBlock dunderNext() {
|
|
|
|
|
operation->checkValid();
|
|
|
|
|
if (mlirBlockIsNull(next)) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::stop_iteration();
|
2020-09-19 22:02:32 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
PyBlock returnBlock(operation, next);
|
|
|
|
|
next = mlirBlockGetNextInRegion(next);
|
|
|
|
|
return returnBlock;
|
|
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
static void bind(nb::module_ &m) {
|
|
|
|
|
nb::class_<PyBlockIterator>(m, "BlockIterator")
|
2020-09-19 22:02:32 -07:00
|
|
|
.def("__iter__", &PyBlockIterator::dunderIter)
|
|
|
|
|
.def("__next__", &PyBlockIterator::dunderNext);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
private:
|
|
|
|
|
PyOperationRef operation;
|
|
|
|
|
MlirBlock next;
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
/// Blocks are exposed by the C-API as a forward-only linked list. In Python,
|
2020-10-29 04:03:15 +09:00
|
|
|
/// we present them as a more full-featured list-like container but optimize
|
2020-09-19 22:02:32 -07:00
|
|
|
/// it for forward iteration. Blocks are always owned by a region.
|
|
|
|
|
class PyBlockList {
|
|
|
|
|
public:
|
|
|
|
|
PyBlockList(PyOperationRef operation, MlirRegion region)
|
|
|
|
|
: operation(std::move(operation)), region(region) {}
|
|
|
|
|
|
|
|
|
|
PyBlockIterator dunderIter() {
|
|
|
|
|
operation->checkValid();
|
|
|
|
|
return PyBlockIterator(operation, mlirRegionGetFirstBlock(region));
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
intptr_t dunderLen() {
|
|
|
|
|
operation->checkValid();
|
|
|
|
|
intptr_t count = 0;
|
|
|
|
|
MlirBlock block = mlirRegionGetFirstBlock(region);
|
|
|
|
|
while (!mlirBlockIsNull(block)) {
|
|
|
|
|
count += 1;
|
|
|
|
|
block = mlirBlockGetNextInRegion(block);
|
|
|
|
|
}
|
|
|
|
|
return count;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
PyBlock dunderGetItem(intptr_t index) {
|
|
|
|
|
operation->checkValid();
|
|
|
|
|
if (index < 0) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::index_error("attempt to access out of bounds block");
|
2020-09-19 22:02:32 -07:00
|
|
|
}
|
|
|
|
|
MlirBlock block = mlirRegionGetFirstBlock(region);
|
|
|
|
|
while (!mlirBlockIsNull(block)) {
|
|
|
|
|
if (index == 0) {
|
|
|
|
|
return PyBlock(operation, block);
|
|
|
|
|
}
|
|
|
|
|
block = mlirBlockGetNextInRegion(block);
|
|
|
|
|
index -= 1;
|
|
|
|
|
}
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::index_error("attempt to access out of bounds block");
|
2020-09-19 22:02:32 -07:00
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
PyBlock appendBlock(const nb::args &pyArgTypes,
|
|
|
|
|
const std::optional<nb::sequence> &pyArgLocs) {
|
2020-09-20 21:25:46 -07:00
|
|
|
operation->checkValid();
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
MlirBlock block =
|
|
|
|
|
createBlock(nb::cast<nb::sequence>(pyArgTypes), pyArgLocs);
|
2020-09-20 21:25:46 -07:00
|
|
|
mlirRegionAppendOwnedBlock(region, block);
|
|
|
|
|
return PyBlock(operation, block);
|
|
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
static void bind(nb::module_ &m) {
|
|
|
|
|
nb::class_<PyBlockList>(m, "BlockList")
|
2020-09-19 22:02:32 -07:00
|
|
|
.def("__getitem__", &PyBlockList::dunderGetItem)
|
|
|
|
|
.def("__iter__", &PyBlockList::dunderIter)
|
2020-09-20 21:25:46 -07:00
|
|
|
.def("__len__", &PyBlockList::dunderLen)
|
2023-05-24 21:51:36 -04:00
|
|
|
.def("append", &PyBlockList::appendBlock, kAppendBlockDocstring,
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("args"), nb::kw_only(),
|
|
|
|
|
nb::arg("arg_locs") = std::nullopt);
|
2020-09-19 22:02:32 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
private:
|
|
|
|
|
PyOperationRef operation;
|
|
|
|
|
MlirRegion region;
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
class PyOperationIterator {
|
|
|
|
|
public:
|
|
|
|
|
PyOperationIterator(PyOperationRef parentOperation, MlirOperation next)
|
|
|
|
|
: parentOperation(std::move(parentOperation)), next(next) {}
|
|
|
|
|
|
|
|
|
|
PyOperationIterator &dunderIter() { return *this; }
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object dunderNext() {
|
2020-09-19 22:02:32 -07:00
|
|
|
parentOperation->checkValid();
|
|
|
|
|
if (mlirOperationIsNull(next)) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::stop_iteration();
|
2020-09-19 22:02:32 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
PyOperationRef returnOperation =
|
|
|
|
|
PyOperation::forOperation(parentOperation->getContext(), next);
|
|
|
|
|
next = mlirOperationGetNextInBlock(next);
|
2020-11-01 23:05:36 -08:00
|
|
|
return returnOperation->createOpView();
|
2020-09-19 22:02:32 -07:00
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
static void bind(nb::module_ &m) {
|
|
|
|
|
nb::class_<PyOperationIterator>(m, "OperationIterator")
|
2020-09-19 22:02:32 -07:00
|
|
|
.def("__iter__", &PyOperationIterator::dunderIter)
|
|
|
|
|
.def("__next__", &PyOperationIterator::dunderNext);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
private:
|
|
|
|
|
PyOperationRef parentOperation;
|
|
|
|
|
MlirOperation next;
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
/// Operations are exposed by the C-API as a forward-only linked list. In
|
|
|
|
|
/// Python, we present them as a more full-featured list-like container but
|
2020-10-29 04:03:15 +09:00
|
|
|
/// optimize it for forward iteration. Iterable operations are always owned
|
2020-09-19 22:02:32 -07:00
|
|
|
/// by a block.
|
|
|
|
|
class PyOperationList {
|
|
|
|
|
public:
|
|
|
|
|
PyOperationList(PyOperationRef parentOperation, MlirBlock block)
|
|
|
|
|
: parentOperation(std::move(parentOperation)), block(block) {}
|
|
|
|
|
|
|
|
|
|
PyOperationIterator dunderIter() {
|
|
|
|
|
parentOperation->checkValid();
|
|
|
|
|
return PyOperationIterator(parentOperation,
|
|
|
|
|
mlirBlockGetFirstOperation(block));
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
intptr_t dunderLen() {
|
|
|
|
|
parentOperation->checkValid();
|
|
|
|
|
intptr_t count = 0;
|
|
|
|
|
MlirOperation childOp = mlirBlockGetFirstOperation(block);
|
|
|
|
|
while (!mlirOperationIsNull(childOp)) {
|
|
|
|
|
count += 1;
|
|
|
|
|
childOp = mlirOperationGetNextInBlock(childOp);
|
|
|
|
|
}
|
|
|
|
|
return count;
|
|
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object dunderGetItem(intptr_t index) {
|
2020-09-19 22:02:32 -07:00
|
|
|
parentOperation->checkValid();
|
|
|
|
|
if (index < 0) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::index_error("attempt to access out of bounds operation");
|
2020-09-19 22:02:32 -07:00
|
|
|
}
|
|
|
|
|
MlirOperation childOp = mlirBlockGetFirstOperation(block);
|
|
|
|
|
while (!mlirOperationIsNull(childOp)) {
|
|
|
|
|
if (index == 0) {
|
|
|
|
|
return PyOperation::forOperation(parentOperation->getContext(), childOp)
|
2020-11-01 23:05:36 -08:00
|
|
|
->createOpView();
|
2020-09-19 22:02:32 -07:00
|
|
|
}
|
|
|
|
|
childOp = mlirOperationGetNextInBlock(childOp);
|
|
|
|
|
index -= 1;
|
|
|
|
|
}
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::index_error("attempt to access out of bounds operation");
|
2020-09-19 22:02:32 -07:00
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
static void bind(nb::module_ &m) {
|
|
|
|
|
nb::class_<PyOperationList>(m, "OperationList")
|
2020-09-19 22:02:32 -07:00
|
|
|
.def("__getitem__", &PyOperationList::dunderGetItem)
|
|
|
|
|
.def("__iter__", &PyOperationList::dunderIter)
|
2020-10-28 23:16:36 -07:00
|
|
|
.def("__len__", &PyOperationList::dunderLen);
|
2020-09-19 22:02:32 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
private:
|
|
|
|
|
PyOperationRef parentOperation;
|
|
|
|
|
MlirBlock block;
|
|
|
|
|
};
|
|
|
|
|
|
2022-12-06 19:30:56 -07:00
|
|
|
class PyOpOperand {
|
|
|
|
|
public:
|
|
|
|
|
PyOpOperand(MlirOpOperand opOperand) : opOperand(opOperand) {}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object getOwner() {
|
2022-12-06 19:30:56 -07:00
|
|
|
MlirOperation owner = mlirOpOperandGetOwner(opOperand);
|
|
|
|
|
PyMlirContextRef context =
|
|
|
|
|
PyMlirContext::forContext(mlirOperationGetContext(owner));
|
|
|
|
|
return PyOperation::forOperation(context, owner)->createOpView();
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
size_t getOperandNumber() { return mlirOpOperandGetOperandNumber(opOperand); }
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
static void bind(nb::module_ &m) {
|
|
|
|
|
nb::class_<PyOpOperand>(m, "OpOperand")
|
|
|
|
|
.def_prop_ro("owner", &PyOpOperand::getOwner)
|
|
|
|
|
.def_prop_ro("operand_number", &PyOpOperand::getOperandNumber);
|
2022-12-06 19:30:56 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
private:
|
|
|
|
|
MlirOpOperand opOperand;
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
class PyOpOperandIterator {
|
|
|
|
|
public:
|
|
|
|
|
PyOpOperandIterator(MlirOpOperand opOperand) : opOperand(opOperand) {}
|
|
|
|
|
|
|
|
|
|
PyOpOperandIterator &dunderIter() { return *this; }
|
|
|
|
|
|
|
|
|
|
PyOpOperand dunderNext() {
|
|
|
|
|
if (mlirOpOperandIsNull(opOperand))
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::stop_iteration();
|
2022-12-06 19:30:56 -07:00
|
|
|
|
|
|
|
|
PyOpOperand returnOpOperand(opOperand);
|
|
|
|
|
opOperand = mlirOpOperandGetNextUse(opOperand);
|
|
|
|
|
return returnOpOperand;
|
|
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
static void bind(nb::module_ &m) {
|
|
|
|
|
nb::class_<PyOpOperandIterator>(m, "OpOperandIterator")
|
2022-12-06 19:30:56 -07:00
|
|
|
.def("__iter__", &PyOpOperandIterator::dunderIter)
|
|
|
|
|
.def("__next__", &PyOpOperandIterator::dunderNext);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
private:
|
|
|
|
|
MlirOpOperand opOperand;
|
|
|
|
|
};
|
|
|
|
|
|
2020-09-19 22:02:32 -07:00
|
|
|
} // namespace
|
|
|
|
|
|
2020-09-18 00:21:09 -07:00
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
// PyMlirContext
|
|
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
|
2020-09-18 18:38:21 -07:00
|
|
|
PyMlirContext::PyMlirContext(MlirContext context) : context(context) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::gil_scoped_acquire acquire;
|
2020-09-18 18:38:21 -07:00
|
|
|
auto &liveContexts = getLiveContexts();
|
|
|
|
|
liveContexts[context.ptr] = this;
|
2020-09-18 00:21:09 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
PyMlirContext::~PyMlirContext() {
|
|
|
|
|
// Note that the only public way to construct an instance is via the
|
|
|
|
|
// forContext method, which always puts the associated handle into
|
|
|
|
|
// liveContexts.
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::gil_scoped_acquire acquire;
|
2025-01-12 18:30:42 +00:00
|
|
|
getLiveContexts().erase(context.ptr);
|
2020-09-18 00:21:09 -07:00
|
|
|
mlirContextDestroy(context);
|
|
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object PyMlirContext::getCapsule() {
|
|
|
|
|
return nb::steal<nb::object>(mlirPythonContextToCapsule(get()));
|
2020-09-28 09:08:09 -07:00
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object PyMlirContext::createFromCapsule(nb::object capsule) {
|
2020-09-28 09:08:09 -07:00
|
|
|
MlirContext rawContext = mlirPythonCapsuleToContext(capsule.ptr());
|
|
|
|
|
if (mlirContextIsNull(rawContext))
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::python_error();
|
Revert "[mlir][python] Make the Context/Operation capsule creation methods work as documented. (#76010)"
This reverts commit bbc29768683b394b34600347f46be2b8245ddb30.
This change seems to be at odds with the non-owning part semantics of
MlirOperation in C API. Since downstream clients can only take and
return MlirOperation, it does not sound correct to force all returns of
MlirOperation transfer ownership. Specifically, this makes it impossible
for downstreams to implement IR-traversing functions that, e.g., look at
neighbors of an operation.
The following patch triggers the exception, and there does not seem to
be an alternative way for a downstream binding writer to express this:
```
diff --git a/mlir/lib/Bindings/Python/IRCore.cpp b/mlir/lib/Bindings/Python/IRCore.cpp
index 39757dfad5be..2ce640674245 100644
--- a/mlir/lib/Bindings/Python/IRCore.cpp
+++ b/mlir/lib/Bindings/Python/IRCore.cpp
@@ -3071,6 +3071,11 @@ void mlir::python::populateIRCore(py::module &m) {
py::arg("successors") = py::none(), py::arg("regions") = 0,
py::arg("loc") = py::none(), py::arg("ip") = py::none(),
py::arg("infer_type") = false, kOperationCreateDocstring)
+ .def("_get_first_in_block", [](PyOperation &self) -> MlirOperation {
+ MlirBlock block = mlirOperationGetBlock(self.get());
+ MlirOperation first = mlirBlockGetFirstOperation(block);
+ return first;
+ })
.def_static(
"parse",
[](const std::string &sourceStr, const std::string &sourceName,
diff --git a/mlir/test/python/ir/operation.py b/mlir/test/python/ir/operation.py
index f59b1a26ba48..6b12b8da5c24 100644
--- a/mlir/test/python/ir/operation.py
+++ b/mlir/test/python/ir/operation.py
@@ -24,6 +24,25 @@ def expect_index_error(callback):
except IndexError:
pass
+@run
+def testCustomBind():
+ ctx = Context()
+ ctx.allow_unregistered_dialects = True
+ module = Module.parse(
+ r"""
+ func.func @f1(%arg0: i32) -> i32 {
+ %1 = "custom.addi"(%arg0, %arg0) : (i32, i32) -> i32
+ return %1 : i32
+ }
+ """,
+ ctx,
+ )
+ add = module.body.operations[0].regions[0].blocks[0].operations[0]
+ op = add.operation
+ # This will get a reference to itself.
+ f1 = op._get_first_in_block()
+
+
# Verify iterator based traversal of the op/region/block hierarchy.
# CHECK-LABEL: TEST: testTraverseOpRegionBlockIterators
```
2023-12-21 10:01:44 +00:00
|
|
|
return forContext(rawContext).releaseObject();
|
2020-09-28 09:08:09 -07:00
|
|
|
}
|
|
|
|
|
|
2020-09-18 00:21:09 -07:00
|
|
|
PyMlirContextRef PyMlirContext::forContext(MlirContext context) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::gil_scoped_acquire acquire;
|
2020-09-18 00:21:09 -07:00
|
|
|
auto &liveContexts = getLiveContexts();
|
|
|
|
|
auto it = liveContexts.find(context.ptr);
|
|
|
|
|
if (it == liveContexts.end()) {
|
Revert "[mlir][python] Make the Context/Operation capsule creation methods work as documented. (#76010)"
This reverts commit bbc29768683b394b34600347f46be2b8245ddb30.
This change seems to be at odds with the non-owning part semantics of
MlirOperation in C API. Since downstream clients can only take and
return MlirOperation, it does not sound correct to force all returns of
MlirOperation transfer ownership. Specifically, this makes it impossible
for downstreams to implement IR-traversing functions that, e.g., look at
neighbors of an operation.
The following patch triggers the exception, and there does not seem to
be an alternative way for a downstream binding writer to express this:
```
diff --git a/mlir/lib/Bindings/Python/IRCore.cpp b/mlir/lib/Bindings/Python/IRCore.cpp
index 39757dfad5be..2ce640674245 100644
--- a/mlir/lib/Bindings/Python/IRCore.cpp
+++ b/mlir/lib/Bindings/Python/IRCore.cpp
@@ -3071,6 +3071,11 @@ void mlir::python::populateIRCore(py::module &m) {
py::arg("successors") = py::none(), py::arg("regions") = 0,
py::arg("loc") = py::none(), py::arg("ip") = py::none(),
py::arg("infer_type") = false, kOperationCreateDocstring)
+ .def("_get_first_in_block", [](PyOperation &self) -> MlirOperation {
+ MlirBlock block = mlirOperationGetBlock(self.get());
+ MlirOperation first = mlirBlockGetFirstOperation(block);
+ return first;
+ })
.def_static(
"parse",
[](const std::string &sourceStr, const std::string &sourceName,
diff --git a/mlir/test/python/ir/operation.py b/mlir/test/python/ir/operation.py
index f59b1a26ba48..6b12b8da5c24 100644
--- a/mlir/test/python/ir/operation.py
+++ b/mlir/test/python/ir/operation.py
@@ -24,6 +24,25 @@ def expect_index_error(callback):
except IndexError:
pass
+@run
+def testCustomBind():
+ ctx = Context()
+ ctx.allow_unregistered_dialects = True
+ module = Module.parse(
+ r"""
+ func.func @f1(%arg0: i32) -> i32 {
+ %1 = "custom.addi"(%arg0, %arg0) : (i32, i32) -> i32
+ return %1 : i32
+ }
+ """,
+ ctx,
+ )
+ add = module.body.operations[0].regions[0].blocks[0].operations[0]
+ op = add.operation
+ # This will get a reference to itself.
+ f1 = op._get_first_in_block()
+
+
# Verify iterator based traversal of the op/region/block hierarchy.
# CHECK-LABEL: TEST: testTraverseOpRegionBlockIterators
```
2023-12-21 10:01:44 +00:00
|
|
|
// Create.
|
|
|
|
|
PyMlirContext *unownedContextWrapper = new PyMlirContext(context);
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object pyRef = nb::cast(unownedContextWrapper);
|
|
|
|
|
assert(pyRef && "cast to nb::object failed");
|
Revert "[mlir][python] Make the Context/Operation capsule creation methods work as documented. (#76010)"
This reverts commit bbc29768683b394b34600347f46be2b8245ddb30.
This change seems to be at odds with the non-owning part semantics of
MlirOperation in C API. Since downstream clients can only take and
return MlirOperation, it does not sound correct to force all returns of
MlirOperation transfer ownership. Specifically, this makes it impossible
for downstreams to implement IR-traversing functions that, e.g., look at
neighbors of an operation.
The following patch triggers the exception, and there does not seem to
be an alternative way for a downstream binding writer to express this:
```
diff --git a/mlir/lib/Bindings/Python/IRCore.cpp b/mlir/lib/Bindings/Python/IRCore.cpp
index 39757dfad5be..2ce640674245 100644
--- a/mlir/lib/Bindings/Python/IRCore.cpp
+++ b/mlir/lib/Bindings/Python/IRCore.cpp
@@ -3071,6 +3071,11 @@ void mlir::python::populateIRCore(py::module &m) {
py::arg("successors") = py::none(), py::arg("regions") = 0,
py::arg("loc") = py::none(), py::arg("ip") = py::none(),
py::arg("infer_type") = false, kOperationCreateDocstring)
+ .def("_get_first_in_block", [](PyOperation &self) -> MlirOperation {
+ MlirBlock block = mlirOperationGetBlock(self.get());
+ MlirOperation first = mlirBlockGetFirstOperation(block);
+ return first;
+ })
.def_static(
"parse",
[](const std::string &sourceStr, const std::string &sourceName,
diff --git a/mlir/test/python/ir/operation.py b/mlir/test/python/ir/operation.py
index f59b1a26ba48..6b12b8da5c24 100644
--- a/mlir/test/python/ir/operation.py
+++ b/mlir/test/python/ir/operation.py
@@ -24,6 +24,25 @@ def expect_index_error(callback):
except IndexError:
pass
+@run
+def testCustomBind():
+ ctx = Context()
+ ctx.allow_unregistered_dialects = True
+ module = Module.parse(
+ r"""
+ func.func @f1(%arg0: i32) -> i32 {
+ %1 = "custom.addi"(%arg0, %arg0) : (i32, i32) -> i32
+ return %1 : i32
+ }
+ """,
+ ctx,
+ )
+ add = module.body.operations[0].regions[0].blocks[0].operations[0]
+ op = add.operation
+ # This will get a reference to itself.
+ f1 = op._get_first_in_block()
+
+
# Verify iterator based traversal of the op/region/block hierarchy.
# CHECK-LABEL: TEST: testTraverseOpRegionBlockIterators
```
2023-12-21 10:01:44 +00:00
|
|
|
liveContexts[context.ptr] = unownedContextWrapper;
|
|
|
|
|
return PyMlirContextRef(unownedContextWrapper, std::move(pyRef));
|
2020-09-18 00:21:09 -07:00
|
|
|
}
|
2020-09-18 18:38:21 -07:00
|
|
|
// Use existing.
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object pyRef = nb::cast(it->second);
|
2020-09-18 18:38:21 -07:00
|
|
|
return PyMlirContextRef(it->second, std::move(pyRef));
|
2020-09-18 00:21:09 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
PyMlirContext::LiveContextMap &PyMlirContext::getLiveContexts() {
|
|
|
|
|
static LiveContextMap liveContexts;
|
|
|
|
|
return liveContexts;
|
|
|
|
|
}
|
|
|
|
|
|
2025-01-12 18:30:42 +00:00
|
|
|
size_t PyMlirContext::getLiveCount() { return getLiveContexts().size(); }
|
2020-09-18 00:21:09 -07:00
|
|
|
|
2020-09-18 18:38:21 -07:00
|
|
|
size_t PyMlirContext::getLiveOperationCount() { return liveOperations.size(); }
|
|
|
|
|
|
2024-02-08 11:39:06 -08:00
|
|
|
std::vector<PyOperation *> PyMlirContext::getLiveOperationObjects() {
|
|
|
|
|
std::vector<PyOperation *> liveObjects;
|
|
|
|
|
for (auto &entry : liveOperations)
|
|
|
|
|
liveObjects.push_back(entry.second.second);
|
|
|
|
|
return liveObjects;
|
|
|
|
|
}
|
|
|
|
|
|
2022-04-19 15:03:15 -07:00
|
|
|
size_t PyMlirContext::clearLiveOperations() {
|
|
|
|
|
for (auto &op : liveOperations)
|
|
|
|
|
op.second.second->setInvalid();
|
|
|
|
|
size_t numInvalidated = liveOperations.size();
|
|
|
|
|
liveOperations.clear();
|
|
|
|
|
return numInvalidated;
|
|
|
|
|
}
|
|
|
|
|
|
2023-10-25 07:17:56 +02:00
|
|
|
void PyMlirContext::clearOperation(MlirOperation op) {
|
|
|
|
|
auto it = liveOperations.find(op.ptr);
|
|
|
|
|
if (it != liveOperations.end()) {
|
|
|
|
|
it->second.second->setInvalid();
|
|
|
|
|
liveOperations.erase(it);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void PyMlirContext::clearOperationsInside(PyOperationBase &op) {
|
|
|
|
|
typedef struct {
|
|
|
|
|
PyOperation &rootOp;
|
|
|
|
|
bool rootSeen;
|
|
|
|
|
} callBackData;
|
|
|
|
|
callBackData data{op.getOperation(), false};
|
|
|
|
|
// Mark all ops below the op that the passmanager will be rooted
|
|
|
|
|
// at (but not op itself - note the preorder) as invalid.
|
|
|
|
|
MlirOperationWalkCallback invalidatingCallback = [](MlirOperation op,
|
|
|
|
|
void *userData) {
|
|
|
|
|
callBackData *data = static_cast<callBackData *>(userData);
|
|
|
|
|
if (LLVM_LIKELY(data->rootSeen))
|
|
|
|
|
data->rootOp.getOperation().getContext()->clearOperation(op);
|
|
|
|
|
else
|
|
|
|
|
data->rootSeen = true;
|
2024-04-17 15:09:47 +09:00
|
|
|
return MlirWalkResult::MlirWalkResultAdvance;
|
2023-10-25 07:17:56 +02:00
|
|
|
};
|
|
|
|
|
mlirOperationWalk(op.getOperation(), invalidatingCallback,
|
|
|
|
|
static_cast<void *>(&data), MlirWalkPreOrder);
|
2023-10-20 20:28:32 -05:00
|
|
|
}
|
2024-02-21 11:01:00 +01:00
|
|
|
void PyMlirContext::clearOperationsInside(MlirOperation op) {
|
|
|
|
|
PyOperationRef opRef = PyOperation::forOperation(getRef(), op);
|
|
|
|
|
clearOperationsInside(opRef->getOperation());
|
|
|
|
|
}
|
2023-10-20 20:28:32 -05:00
|
|
|
|
2024-05-30 10:06:02 +02:00
|
|
|
void PyMlirContext::clearOperationAndInside(PyOperationBase &op) {
|
|
|
|
|
MlirOperationWalkCallback invalidatingCallback = [](MlirOperation op,
|
|
|
|
|
void *userData) {
|
|
|
|
|
PyMlirContextRef &contextRef = *static_cast<PyMlirContextRef *>(userData);
|
|
|
|
|
contextRef->clearOperation(op);
|
|
|
|
|
return MlirWalkResult::MlirWalkResultAdvance;
|
|
|
|
|
};
|
|
|
|
|
mlirOperationWalk(op.getOperation(), invalidatingCallback,
|
|
|
|
|
&op.getOperation().getContext(), MlirWalkPreOrder);
|
|
|
|
|
}
|
|
|
|
|
|
2020-10-12 21:19:13 -07:00
|
|
|
size_t PyMlirContext::getLiveModuleCount() { return liveModules.size(); }
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object PyMlirContext::contextEnter(nb::object context) {
|
|
|
|
|
return PyThreadContextEntry::pushContext(context);
|
2020-10-31 23:40:25 -07:00
|
|
|
}
|
2020-09-20 21:25:46 -07:00
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
void PyMlirContext::contextExit(const nb::object &excType,
|
|
|
|
|
const nb::object &excVal,
|
|
|
|
|
const nb::object &excTb) {
|
2020-10-31 23:40:25 -07:00
|
|
|
PyThreadContextEntry::popContext(*this);
|
|
|
|
|
}
|
2020-09-20 21:25:46 -07:00
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object PyMlirContext::attachDiagnosticHandler(nb::object callback) {
|
2022-01-03 16:39:58 -08:00
|
|
|
// Note that ownership is transferred to the delete callback below by way of
|
|
|
|
|
// an explicit inc_ref (borrow).
|
|
|
|
|
PyDiagnosticHandler *pyHandler =
|
|
|
|
|
new PyDiagnosticHandler(get(), std::move(callback));
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object pyHandlerObject =
|
|
|
|
|
nb::cast(pyHandler, nb::rv_policy::take_ownership);
|
2022-01-03 16:39:58 -08:00
|
|
|
pyHandlerObject.inc_ref();
|
|
|
|
|
|
|
|
|
|
// In these C callbacks, the userData is a PyDiagnosticHandler* that is
|
|
|
|
|
// guaranteed to be known to pybind.
|
|
|
|
|
auto handlerCallback =
|
|
|
|
|
+[](MlirDiagnostic diagnostic, void *userData) -> MlirLogicalResult {
|
|
|
|
|
PyDiagnostic *pyDiagnostic = new PyDiagnostic(diagnostic);
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object pyDiagnosticObject =
|
|
|
|
|
nb::cast(pyDiagnostic, nb::rv_policy::take_ownership);
|
2022-01-03 16:39:58 -08:00
|
|
|
|
|
|
|
|
auto *pyHandler = static_cast<PyDiagnosticHandler *>(userData);
|
|
|
|
|
bool result = false;
|
|
|
|
|
{
|
|
|
|
|
// Since this can be called from arbitrary C++ contexts, always get the
|
|
|
|
|
// gil.
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::gil_scoped_acquire gil;
|
2022-01-03 16:39:58 -08:00
|
|
|
try {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
result = nb::cast<bool>(pyHandler->callback(pyDiagnostic));
|
2022-01-03 16:39:58 -08:00
|
|
|
} catch (std::exception &e) {
|
|
|
|
|
fprintf(stderr, "MLIR Python Diagnostic handler raised exception: %s\n",
|
|
|
|
|
e.what());
|
|
|
|
|
pyHandler->hadError = true;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
pyDiagnostic->invalidate();
|
|
|
|
|
return result ? mlirLogicalResultSuccess() : mlirLogicalResultFailure();
|
|
|
|
|
};
|
|
|
|
|
auto deleteCallback = +[](void *userData) {
|
|
|
|
|
auto *pyHandler = static_cast<PyDiagnosticHandler *>(userData);
|
|
|
|
|
assert(pyHandler->registeredID && "handler is not registered");
|
|
|
|
|
pyHandler->registeredID.reset();
|
|
|
|
|
|
|
|
|
|
// Decrement reference, balancing the inc_ref() above.
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object pyHandlerObject = nb::cast(pyHandler, nb::rv_policy::reference);
|
2022-01-03 16:39:58 -08:00
|
|
|
pyHandlerObject.dec_ref();
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
pyHandler->registeredID = mlirContextAttachDiagnosticHandler(
|
|
|
|
|
get(), handlerCallback, static_cast<void *>(pyHandler), deleteCallback);
|
|
|
|
|
return pyHandlerObject;
|
|
|
|
|
}
|
|
|
|
|
|
[mlir][python] Capture error diagnostics in exceptions
This updates most (all?) error-diagnostic-emitting python APIs to
capture error diagnostics and include them in the raised exception's
message:
```
>>> Operation.parse('"arith.addi"() : () -> ()'))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
mlir._mlir_libs.MLIRError: Unable to parse operation assembly:
error: "-":1:1: 'arith.addi' op requires one result
note: "-":1:1: see current operation: "arith.addi"() : () -> ()
```
The diagnostic information is available on the exception for users who
may want to customize the error message:
```
>>> try:
... Operation.parse('"arith.addi"() : () -> ()')
... except MLIRError as e:
... print(e.message)
... print(e.error_diagnostics)
... print(e.error_diagnostics[0].message)
...
Unable to parse operation assembly
[<mlir._mlir_libs._mlir.ir.DiagnosticInfo object at 0x7fed32bd6b70>]
'arith.addi' op requires one result
```
Error diagnostics captured in exceptions aren't propagated to diagnostic
handlers, to avoid double-reporting of errors. The context-level
`emit_error_diagnostics` option can be used to revert to the old
behaviour, causing error diagnostics to be reported to handlers instead
of as part of exceptions.
API changes:
- `Operation.verify` now raises an exception on verification failure,
instead of returning `false`
- The exception raised by the following methods has been changed to
`MLIRError`:
- `PassManager.run`
- `{Module,Operation,Type,Attribute}.parse`
- `{RankedTensorType,UnrankedTensorType}.get`
- `{MemRefType,UnrankedMemRefType}.get`
- `VectorType.get`
- `FloatAttr.get`
closes #60595
depends on D144804, D143830
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D143869
2023-02-07 16:07:50 -05:00
|
|
|
MlirLogicalResult PyMlirContext::ErrorCapture::handler(MlirDiagnostic diag,
|
|
|
|
|
void *userData) {
|
|
|
|
|
auto *self = static_cast<ErrorCapture *>(userData);
|
|
|
|
|
// Check if the context requested we emit errors instead of capturing them.
|
|
|
|
|
if (self->ctx->emitErrorDiagnostics)
|
|
|
|
|
return mlirLogicalResultFailure();
|
|
|
|
|
|
|
|
|
|
if (mlirDiagnosticGetSeverity(diag) != MlirDiagnosticError)
|
|
|
|
|
return mlirLogicalResultFailure();
|
|
|
|
|
|
|
|
|
|
self->errors.emplace_back(PyDiagnostic(diag).getInfo());
|
|
|
|
|
return mlirLogicalResultSuccess();
|
|
|
|
|
}
|
|
|
|
|
|
2020-10-31 23:40:25 -07:00
|
|
|
PyMlirContext &DefaultingPyMlirContext::resolve() {
|
|
|
|
|
PyMlirContext *context = PyThreadContextEntry::getDefaultContext();
|
|
|
|
|
if (!context) {
|
2023-05-22 17:30:12 -05:00
|
|
|
throw std::runtime_error(
|
2020-10-31 23:40:25 -07:00
|
|
|
"An MLIR function requires a Context but none was provided in the call "
|
|
|
|
|
"or from the surrounding environment. Either pass to the function with "
|
|
|
|
|
"a 'context=' argument or establish a default using 'with Context():'");
|
2020-09-20 21:25:46 -07:00
|
|
|
}
|
2020-10-31 23:40:25 -07:00
|
|
|
return *context;
|
2020-10-28 23:16:36 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
// PyThreadContextEntry management
|
|
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
|
|
|
|
|
std::vector<PyThreadContextEntry> &PyThreadContextEntry::getStack() {
|
|
|
|
|
static thread_local std::vector<PyThreadContextEntry> stack;
|
|
|
|
|
return stack;
|
|
|
|
|
}
|
|
|
|
|
|
2020-10-31 23:40:25 -07:00
|
|
|
PyThreadContextEntry *PyThreadContextEntry::getTopOfStack() {
|
2020-10-28 23:16:36 -07:00
|
|
|
auto &stack = getStack();
|
|
|
|
|
if (stack.empty())
|
|
|
|
|
return nullptr;
|
|
|
|
|
return &stack.back();
|
|
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
void PyThreadContextEntry::push(FrameKind frameKind, nb::object context,
|
|
|
|
|
nb::object insertionPoint,
|
|
|
|
|
nb::object location) {
|
2020-10-28 23:16:36 -07:00
|
|
|
auto &stack = getStack();
|
2020-10-31 23:40:25 -07:00
|
|
|
stack.emplace_back(frameKind, std::move(context), std::move(insertionPoint),
|
|
|
|
|
std::move(location));
|
|
|
|
|
// If the new stack has more than one entry and the context of the new top
|
|
|
|
|
// entry matches the previous, copy the insertionPoint and location from the
|
|
|
|
|
// previous entry if missing from the new top entry.
|
|
|
|
|
if (stack.size() > 1) {
|
|
|
|
|
auto &prev = *(stack.rbegin() + 1);
|
|
|
|
|
auto ¤t = stack.back();
|
|
|
|
|
if (current.context.is(prev.context)) {
|
|
|
|
|
// Default non-context objects from the previous entry.
|
|
|
|
|
if (!current.insertionPoint)
|
|
|
|
|
current.insertionPoint = prev.insertionPoint;
|
|
|
|
|
if (!current.location)
|
|
|
|
|
current.location = prev.location;
|
|
|
|
|
}
|
|
|
|
|
}
|
2020-10-28 23:16:36 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
PyMlirContext *PyThreadContextEntry::getContext() {
|
|
|
|
|
if (!context)
|
|
|
|
|
return nullptr;
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
return nb::cast<PyMlirContext *>(context);
|
2020-10-28 23:16:36 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
PyInsertionPoint *PyThreadContextEntry::getInsertionPoint() {
|
|
|
|
|
if (!insertionPoint)
|
|
|
|
|
return nullptr;
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
return nb::cast<PyInsertionPoint *>(insertionPoint);
|
2020-10-28 23:16:36 -07:00
|
|
|
}
|
|
|
|
|
|
2020-10-31 23:40:25 -07:00
|
|
|
PyLocation *PyThreadContextEntry::getLocation() {
|
|
|
|
|
if (!location)
|
|
|
|
|
return nullptr;
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
return nb::cast<PyLocation *>(location);
|
2020-10-31 23:40:25 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
PyMlirContext *PyThreadContextEntry::getDefaultContext() {
|
|
|
|
|
auto *tos = getTopOfStack();
|
|
|
|
|
return tos ? tos->getContext() : nullptr;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
PyInsertionPoint *PyThreadContextEntry::getDefaultInsertionPoint() {
|
|
|
|
|
auto *tos = getTopOfStack();
|
|
|
|
|
return tos ? tos->getInsertionPoint() : nullptr;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
PyLocation *PyThreadContextEntry::getDefaultLocation() {
|
|
|
|
|
auto *tos = getTopOfStack();
|
|
|
|
|
return tos ? tos->getLocation() : nullptr;
|
|
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object PyThreadContextEntry::pushContext(nb::object context) {
|
|
|
|
|
push(FrameKind::Context, /*context=*/context,
|
|
|
|
|
/*insertionPoint=*/nb::object(),
|
|
|
|
|
/*location=*/nb::object());
|
|
|
|
|
return context;
|
2020-10-31 23:40:25 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void PyThreadContextEntry::popContext(PyMlirContext &context) {
|
|
|
|
|
auto &stack = getStack();
|
|
|
|
|
if (stack.empty())
|
2023-05-22 17:30:12 -05:00
|
|
|
throw std::runtime_error("Unbalanced Context enter/exit");
|
2020-10-31 23:40:25 -07:00
|
|
|
auto &tos = stack.back();
|
|
|
|
|
if (tos.frameKind != FrameKind::Context && tos.getContext() != &context)
|
2023-05-22 17:30:12 -05:00
|
|
|
throw std::runtime_error("Unbalanced Context enter/exit");
|
2020-10-31 23:40:25 -07:00
|
|
|
stack.pop_back();
|
2020-10-28 23:16:36 -07:00
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object
|
|
|
|
|
PyThreadContextEntry::pushInsertionPoint(nb::object insertionPointObj) {
|
|
|
|
|
PyInsertionPoint &insertionPoint =
|
|
|
|
|
nb::cast<PyInsertionPoint &>(insertionPointObj);
|
|
|
|
|
nb::object contextObj =
|
2020-10-31 23:40:25 -07:00
|
|
|
insertionPoint.getBlock().getParentOperation()->getContext().getObject();
|
|
|
|
|
push(FrameKind::InsertionPoint,
|
|
|
|
|
/*context=*/contextObj,
|
|
|
|
|
/*insertionPoint=*/insertionPointObj,
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
/*location=*/nb::object());
|
2020-10-31 23:40:25 -07:00
|
|
|
return insertionPointObj;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void PyThreadContextEntry::popInsertionPoint(PyInsertionPoint &insertionPoint) {
|
|
|
|
|
auto &stack = getStack();
|
|
|
|
|
if (stack.empty())
|
2023-05-22 17:30:12 -05:00
|
|
|
throw std::runtime_error("Unbalanced InsertionPoint enter/exit");
|
2020-10-31 23:40:25 -07:00
|
|
|
auto &tos = stack.back();
|
|
|
|
|
if (tos.frameKind != FrameKind::InsertionPoint &&
|
|
|
|
|
tos.getInsertionPoint() != &insertionPoint)
|
2023-05-22 17:30:12 -05:00
|
|
|
throw std::runtime_error("Unbalanced InsertionPoint enter/exit");
|
2020-10-31 23:40:25 -07:00
|
|
|
stack.pop_back();
|
|
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object PyThreadContextEntry::pushLocation(nb::object locationObj) {
|
|
|
|
|
PyLocation &location = nb::cast<PyLocation &>(locationObj);
|
|
|
|
|
nb::object contextObj = location.getContext().getObject();
|
2020-10-31 23:40:25 -07:00
|
|
|
push(FrameKind::Location, /*context=*/contextObj,
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
/*insertionPoint=*/nb::object(),
|
2020-10-31 23:40:25 -07:00
|
|
|
/*location=*/locationObj);
|
|
|
|
|
return locationObj;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void PyThreadContextEntry::popLocation(PyLocation &location) {
|
|
|
|
|
auto &stack = getStack();
|
|
|
|
|
if (stack.empty())
|
2023-05-22 17:30:12 -05:00
|
|
|
throw std::runtime_error("Unbalanced Location enter/exit");
|
2020-10-31 23:40:25 -07:00
|
|
|
auto &tos = stack.back();
|
|
|
|
|
if (tos.frameKind != FrameKind::Location && tos.getLocation() != &location)
|
2023-05-22 17:30:12 -05:00
|
|
|
throw std::runtime_error("Unbalanced Location enter/exit");
|
2020-10-31 23:40:25 -07:00
|
|
|
stack.pop_back();
|
2020-09-20 21:25:46 -07:00
|
|
|
}
|
|
|
|
|
|
2022-01-03 16:39:58 -08:00
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
// PyDiagnostic*
|
|
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
|
|
|
|
|
void PyDiagnostic::invalidate() {
|
|
|
|
|
valid = false;
|
|
|
|
|
if (materializedNotes) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
for (nb::handle noteObject : *materializedNotes) {
|
|
|
|
|
PyDiagnostic *note = nb::cast<PyDiagnostic *>(noteObject);
|
2022-01-03 16:39:58 -08:00
|
|
|
note->invalidate();
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
PyDiagnosticHandler::PyDiagnosticHandler(MlirContext context,
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object callback)
|
2022-01-03 16:39:58 -08:00
|
|
|
: context(context), callback(std::move(callback)) {}
|
|
|
|
|
|
2022-01-08 20:37:39 +00:00
|
|
|
PyDiagnosticHandler::~PyDiagnosticHandler() = default;
|
2022-01-03 16:39:58 -08:00
|
|
|
|
|
|
|
|
void PyDiagnosticHandler::detach() {
|
|
|
|
|
if (!registeredID)
|
|
|
|
|
return;
|
|
|
|
|
MlirDiagnosticHandlerID localID = *registeredID;
|
|
|
|
|
mlirContextDetachDiagnosticHandler(context, localID);
|
|
|
|
|
assert(!registeredID && "should have unregistered");
|
|
|
|
|
// Not strictly necessary but keeps stale pointers from being around to cause
|
|
|
|
|
// issues.
|
|
|
|
|
context = {nullptr};
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void PyDiagnostic::checkValid() {
|
|
|
|
|
if (!valid) {
|
|
|
|
|
throw std::invalid_argument(
|
|
|
|
|
"Diagnostic is invalid (used outside of callback)");
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
MlirDiagnosticSeverity PyDiagnostic::getSeverity() {
|
|
|
|
|
checkValid();
|
|
|
|
|
return mlirDiagnosticGetSeverity(diagnostic);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
PyLocation PyDiagnostic::getLocation() {
|
|
|
|
|
checkValid();
|
|
|
|
|
MlirLocation loc = mlirDiagnosticGetLocation(diagnostic);
|
|
|
|
|
MlirContext context = mlirLocationGetContext(loc);
|
|
|
|
|
return PyLocation(PyMlirContext::forContext(context), loc);
|
|
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::str PyDiagnostic::getMessage() {
|
2022-01-03 16:39:58 -08:00
|
|
|
checkValid();
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object fileObject = nb::module_::import_("io").attr("StringIO")();
|
2022-01-03 16:39:58 -08:00
|
|
|
PyFileAccumulator accum(fileObject, /*binary=*/false);
|
|
|
|
|
mlirDiagnosticPrint(diagnostic, accum.getCallback(), accum.getUserData());
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
return nb::cast<nb::str>(fileObject.attr("getvalue")());
|
2022-01-03 16:39:58 -08:00
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::tuple PyDiagnostic::getNotes() {
|
2022-01-03 16:39:58 -08:00
|
|
|
checkValid();
|
|
|
|
|
if (materializedNotes)
|
|
|
|
|
return *materializedNotes;
|
|
|
|
|
intptr_t numNotes = mlirDiagnosticGetNumNotes(diagnostic);
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::tuple notes = nb::steal<nb::tuple>(PyTuple_New(numNotes));
|
2022-01-03 16:39:58 -08:00
|
|
|
for (intptr_t i = 0; i < numNotes; ++i) {
|
|
|
|
|
MlirDiagnostic noteDiag = mlirDiagnosticGetNote(diagnostic, i);
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object diagnostic = nb::cast(PyDiagnostic(noteDiag));
|
|
|
|
|
PyTuple_SET_ITEM(notes.ptr(), i, diagnostic.release().ptr());
|
2022-01-03 16:39:58 -08:00
|
|
|
}
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
materializedNotes = std::move(notes);
|
|
|
|
|
|
2022-01-03 16:39:58 -08:00
|
|
|
return *materializedNotes;
|
|
|
|
|
}
|
|
|
|
|
|
[mlir][python] Capture error diagnostics in exceptions
This updates most (all?) error-diagnostic-emitting python APIs to
capture error diagnostics and include them in the raised exception's
message:
```
>>> Operation.parse('"arith.addi"() : () -> ()'))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
mlir._mlir_libs.MLIRError: Unable to parse operation assembly:
error: "-":1:1: 'arith.addi' op requires one result
note: "-":1:1: see current operation: "arith.addi"() : () -> ()
```
The diagnostic information is available on the exception for users who
may want to customize the error message:
```
>>> try:
... Operation.parse('"arith.addi"() : () -> ()')
... except MLIRError as e:
... print(e.message)
... print(e.error_diagnostics)
... print(e.error_diagnostics[0].message)
...
Unable to parse operation assembly
[<mlir._mlir_libs._mlir.ir.DiagnosticInfo object at 0x7fed32bd6b70>]
'arith.addi' op requires one result
```
Error diagnostics captured in exceptions aren't propagated to diagnostic
handlers, to avoid double-reporting of errors. The context-level
`emit_error_diagnostics` option can be used to revert to the old
behaviour, causing error diagnostics to be reported to handlers instead
of as part of exceptions.
API changes:
- `Operation.verify` now raises an exception on verification failure,
instead of returning `false`
- The exception raised by the following methods has been changed to
`MLIRError`:
- `PassManager.run`
- `{Module,Operation,Type,Attribute}.parse`
- `{RankedTensorType,UnrankedTensorType}.get`
- `{MemRefType,UnrankedMemRefType}.get`
- `VectorType.get`
- `FloatAttr.get`
closes #60595
depends on D144804, D143830
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D143869
2023-02-07 16:07:50 -05:00
|
|
|
PyDiagnostic::DiagnosticInfo PyDiagnostic::getInfo() {
|
|
|
|
|
std::vector<DiagnosticInfo> notes;
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
for (nb::handle n : getNotes())
|
|
|
|
|
notes.emplace_back(nb::cast<PyDiagnostic>(n).getInfo());
|
|
|
|
|
return {getSeverity(), getLocation(), nb::cast<std::string>(getMessage()),
|
|
|
|
|
std::move(notes)};
|
[mlir][python] Capture error diagnostics in exceptions
This updates most (all?) error-diagnostic-emitting python APIs to
capture error diagnostics and include them in the raised exception's
message:
```
>>> Operation.parse('"arith.addi"() : () -> ()'))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
mlir._mlir_libs.MLIRError: Unable to parse operation assembly:
error: "-":1:1: 'arith.addi' op requires one result
note: "-":1:1: see current operation: "arith.addi"() : () -> ()
```
The diagnostic information is available on the exception for users who
may want to customize the error message:
```
>>> try:
... Operation.parse('"arith.addi"() : () -> ()')
... except MLIRError as e:
... print(e.message)
... print(e.error_diagnostics)
... print(e.error_diagnostics[0].message)
...
Unable to parse operation assembly
[<mlir._mlir_libs._mlir.ir.DiagnosticInfo object at 0x7fed32bd6b70>]
'arith.addi' op requires one result
```
Error diagnostics captured in exceptions aren't propagated to diagnostic
handlers, to avoid double-reporting of errors. The context-level
`emit_error_diagnostics` option can be used to revert to the old
behaviour, causing error diagnostics to be reported to handlers instead
of as part of exceptions.
API changes:
- `Operation.verify` now raises an exception on verification failure,
instead of returning `false`
- The exception raised by the following methods has been changed to
`MLIRError`:
- `PassManager.run`
- `{Module,Operation,Type,Attribute}.parse`
- `{RankedTensorType,UnrankedTensorType}.get`
- `{MemRefType,UnrankedMemRefType}.get`
- `VectorType.get`
- `FloatAttr.get`
closes #60595
depends on D144804, D143830
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D143869
2023-02-07 16:07:50 -05:00
|
|
|
}
|
|
|
|
|
|
2020-10-21 23:34:01 -07:00
|
|
|
//------------------------------------------------------------------------------
|
[mlir] Overhaul C/Python registration APIs to properly scope registration/loading activities.
Since the very first commits, the Python and C MLIR APIs have had mis-placed registration/load functionality for dialects, extensions, etc. This was done pragmatically in order to get bootstrapped and then just grew in. Downstreams largely bypass and do their own thing by providing various APIs to register things they need. Meanwhile, the C++ APIs have stabilized around this and it would make sense to follow suit.
The thing we have observed in canonical usage by downstreams is that each downstream tends to have native entry points that configure its installation to its preferences with one-stop APIs. This patch leans in to this approach with `RegisterEverything.h` and `mlir._mlir_libs._mlirRegisterEverything` being the one-stop entry points for the "upstream packages". The `_mlir_libs.__init__.py` now allows customization of the environment and Context by adding "initialization modules" to the `_mlir_libs` package. If present, `_mlirRegisterEverything` is treated as such a module. Others can be added by downstreams by adding a `_site_initialize_{i}.py` module, where '{i}' is a number starting with zero. The number will be incremented and corresponding module loaded until one is not found. Initialization modules can:
* Perform load time customization to the global environment (i.e. registering passes, hooks, etc).
* Define a `register_dialects(registry: DialectRegistry)` function that can extend the `DialectRegistry` that will be used to bootstrap the `Context`.
* Define a `context_init_hook(context: Context)` function that will be added to a list of callbacks which will be invoked after dialect registration during `Context` initialization.
Note that the `MLIRPythonExtension.RegisterEverything` is not included by default when building a downstream (its corresponding behavior was prior). For downstreams which need the default MLIR initialization to take place, they must add this back in to their Python CMake build just like they add their own components (i.e. to `add_mlir_python_common_capi_library` and `add_mlir_python_modules`). It is perfectly valid to not do this, in which case, only the things explicitly depended on and initialized by downstreams will be built/packaged. If the downstream has not been set up for this, it is recommended to simply add this back for the time being and pay the build time/package size cost.
CMake changes:
* `MLIRCAPIRegistration` -> `MLIRCAPIRegisterEverything` (renamed to signify what it does and force an evaluation: a number of places were incidentally linking this very expensive target)
* `MLIRPythonSoure.Passes` removed (without replacement: just drop)
* `MLIRPythonExtension.AllPassesRegistration` removed (without replacement: just drop)
* `MLIRPythonExtension.Conversions` removed (without replacement: just drop)
* `MLIRPythonExtension.Transforms` removed (without replacement: just drop)
Header changes:
* `mlir-c/Registration.h` is deleted. Dialect registration functionality is now in `IR.h`. Registration of upstream features are in `mlir-c/RegisterEverything.h`. When updating MLIR and a couple of downstreams, I found that proper usage was commingled so required making a choice vs just blind S&R.
Python APIs removed:
* mlir.transforms and mlir.conversions (previously only had an __init__.py which indirectly triggered `mlirRegisterTransformsPasses()` and `mlirRegisterConversionPasses()` respectively). Downstream impact: Remove these imports if present (they now happen as part of default initialization).
* mlir._mlir_libs._all_passes_registration, mlir._mlir_libs._mlirTransforms, mlir._mlir_libs._mlirConversions. Downstream impact: None expected (these were internally used).
C-APIs changed:
* mlirRegisterAllDialects(MlirContext) now takes an MlirDialectRegistry instead. It also used to trigger loading of all dialects, which was already marked with a TODO to remove -- it no longer does, and for direct use, dialects must be explicitly loaded. Downstream impact: Direct C-API users must ensure that needed dialects are loaded or call `mlirContextLoadAllAvailableDialects(MlirContext)` to emulate the prior behavior. Also see the `ir.c` test case (e.g. ` mlirContextGetOrLoadDialect(ctx, mlirStringRefCreateFromCString("func"));`).
* mlirDialectHandle* APIs were moved from Registration.h (which now is restricted to just global/upstream registration) to IR.h, arguably where it should have been. Downstream impact: include correct header (likely already doing so).
C-APIs added:
* mlirContextLoadAllAvailableDialects(MlirContext): Corresponds to C++ API with the same purpose.
Python APIs added:
* mlir.ir.DialectRegistry: Mapping for an MlirDialectRegistry.
* mlir.ir.Context.append_dialect_registry(MlirDialectRegistry)
* mlir.ir.Context.load_all_available_dialects()
* mlir._mlir_libs._mlirAllRegistration: New native extension that exposes a `register_dialects(MlirDialectRegistry)` entry point and performs all upstream pass/conversion/transforms registration on init. In this first step, we eagerly load this as part of the __init__.py and use it to monkey patch the Context to emulate prior behavior.
* Type caster and capsule support for MlirDialectRegistry
This should make it possible to build downstream Python dialects that only depend on a subset of MLIR. See: https://github.com/llvm/llvm-project/issues/56037
Here is an example PR, minimally adapting IREE to these changes: https://github.com/iree-org/iree/pull/9638/files In this situation, IREE is opting to not link everything, since it is already configuring the Context to its liking. For projects that would just like to not think about it and pull in everything, add `MLIRPythonExtension.RegisterEverything` to the list of Python sources getting built, and the old behavior will continue.
Reviewed By: mehdi_amini, ftynse
Differential Revision: https://reviews.llvm.org/D128593
2022-07-16 16:09:03 -07:00
|
|
|
// PyDialect, PyDialectDescriptor, PyDialects, PyDialectRegistry
|
2020-10-21 23:34:01 -07:00
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
|
|
|
|
|
MlirDialect PyDialects::getDialectForKey(const std::string &key,
|
|
|
|
|
bool attrError) {
|
2021-07-28 20:32:47 +00:00
|
|
|
MlirDialect dialect = mlirContextGetOrLoadDialect(getContext()->get(),
|
|
|
|
|
{key.data(), key.size()});
|
2020-10-21 23:34:01 -07:00
|
|
|
if (mlirDialectIsNull(dialect)) {
|
2023-05-22 17:30:12 -05:00
|
|
|
std::string msg = (Twine("Dialect '") + key + "' not found").str();
|
|
|
|
|
if (attrError)
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::attribute_error(msg.c_str());
|
|
|
|
|
throw nb::index_error(msg.c_str());
|
2020-10-21 23:34:01 -07:00
|
|
|
}
|
|
|
|
|
return dialect;
|
|
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object PyDialectRegistry::getCapsule() {
|
|
|
|
|
return nb::steal<nb::object>(mlirPythonDialectRegistryToCapsule(*this));
|
[mlir] Overhaul C/Python registration APIs to properly scope registration/loading activities.
Since the very first commits, the Python and C MLIR APIs have had mis-placed registration/load functionality for dialects, extensions, etc. This was done pragmatically in order to get bootstrapped and then just grew in. Downstreams largely bypass and do their own thing by providing various APIs to register things they need. Meanwhile, the C++ APIs have stabilized around this and it would make sense to follow suit.
The thing we have observed in canonical usage by downstreams is that each downstream tends to have native entry points that configure its installation to its preferences with one-stop APIs. This patch leans in to this approach with `RegisterEverything.h` and `mlir._mlir_libs._mlirRegisterEverything` being the one-stop entry points for the "upstream packages". The `_mlir_libs.__init__.py` now allows customization of the environment and Context by adding "initialization modules" to the `_mlir_libs` package. If present, `_mlirRegisterEverything` is treated as such a module. Others can be added by downstreams by adding a `_site_initialize_{i}.py` module, where '{i}' is a number starting with zero. The number will be incremented and corresponding module loaded until one is not found. Initialization modules can:
* Perform load time customization to the global environment (i.e. registering passes, hooks, etc).
* Define a `register_dialects(registry: DialectRegistry)` function that can extend the `DialectRegistry` that will be used to bootstrap the `Context`.
* Define a `context_init_hook(context: Context)` function that will be added to a list of callbacks which will be invoked after dialect registration during `Context` initialization.
Note that the `MLIRPythonExtension.RegisterEverything` is not included by default when building a downstream (its corresponding behavior was prior). For downstreams which need the default MLIR initialization to take place, they must add this back in to their Python CMake build just like they add their own components (i.e. to `add_mlir_python_common_capi_library` and `add_mlir_python_modules`). It is perfectly valid to not do this, in which case, only the things explicitly depended on and initialized by downstreams will be built/packaged. If the downstream has not been set up for this, it is recommended to simply add this back for the time being and pay the build time/package size cost.
CMake changes:
* `MLIRCAPIRegistration` -> `MLIRCAPIRegisterEverything` (renamed to signify what it does and force an evaluation: a number of places were incidentally linking this very expensive target)
* `MLIRPythonSoure.Passes` removed (without replacement: just drop)
* `MLIRPythonExtension.AllPassesRegistration` removed (without replacement: just drop)
* `MLIRPythonExtension.Conversions` removed (without replacement: just drop)
* `MLIRPythonExtension.Transforms` removed (without replacement: just drop)
Header changes:
* `mlir-c/Registration.h` is deleted. Dialect registration functionality is now in `IR.h`. Registration of upstream features are in `mlir-c/RegisterEverything.h`. When updating MLIR and a couple of downstreams, I found that proper usage was commingled so required making a choice vs just blind S&R.
Python APIs removed:
* mlir.transforms and mlir.conversions (previously only had an __init__.py which indirectly triggered `mlirRegisterTransformsPasses()` and `mlirRegisterConversionPasses()` respectively). Downstream impact: Remove these imports if present (they now happen as part of default initialization).
* mlir._mlir_libs._all_passes_registration, mlir._mlir_libs._mlirTransforms, mlir._mlir_libs._mlirConversions. Downstream impact: None expected (these were internally used).
C-APIs changed:
* mlirRegisterAllDialects(MlirContext) now takes an MlirDialectRegistry instead. It also used to trigger loading of all dialects, which was already marked with a TODO to remove -- it no longer does, and for direct use, dialects must be explicitly loaded. Downstream impact: Direct C-API users must ensure that needed dialects are loaded or call `mlirContextLoadAllAvailableDialects(MlirContext)` to emulate the prior behavior. Also see the `ir.c` test case (e.g. ` mlirContextGetOrLoadDialect(ctx, mlirStringRefCreateFromCString("func"));`).
* mlirDialectHandle* APIs were moved from Registration.h (which now is restricted to just global/upstream registration) to IR.h, arguably where it should have been. Downstream impact: include correct header (likely already doing so).
C-APIs added:
* mlirContextLoadAllAvailableDialects(MlirContext): Corresponds to C++ API with the same purpose.
Python APIs added:
* mlir.ir.DialectRegistry: Mapping for an MlirDialectRegistry.
* mlir.ir.Context.append_dialect_registry(MlirDialectRegistry)
* mlir.ir.Context.load_all_available_dialects()
* mlir._mlir_libs._mlirAllRegistration: New native extension that exposes a `register_dialects(MlirDialectRegistry)` entry point and performs all upstream pass/conversion/transforms registration on init. In this first step, we eagerly load this as part of the __init__.py and use it to monkey patch the Context to emulate prior behavior.
* Type caster and capsule support for MlirDialectRegistry
This should make it possible to build downstream Python dialects that only depend on a subset of MLIR. See: https://github.com/llvm/llvm-project/issues/56037
Here is an example PR, minimally adapting IREE to these changes: https://github.com/iree-org/iree/pull/9638/files In this situation, IREE is opting to not link everything, since it is already configuring the Context to its liking. For projects that would just like to not think about it and pull in everything, add `MLIRPythonExtension.RegisterEverything` to the list of Python sources getting built, and the old behavior will continue.
Reviewed By: mehdi_amini, ftynse
Differential Revision: https://reviews.llvm.org/D128593
2022-07-16 16:09:03 -07:00
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
PyDialectRegistry PyDialectRegistry::createFromCapsule(nb::object capsule) {
|
[mlir] Overhaul C/Python registration APIs to properly scope registration/loading activities.
Since the very first commits, the Python and C MLIR APIs have had mis-placed registration/load functionality for dialects, extensions, etc. This was done pragmatically in order to get bootstrapped and then just grew in. Downstreams largely bypass and do their own thing by providing various APIs to register things they need. Meanwhile, the C++ APIs have stabilized around this and it would make sense to follow suit.
The thing we have observed in canonical usage by downstreams is that each downstream tends to have native entry points that configure its installation to its preferences with one-stop APIs. This patch leans in to this approach with `RegisterEverything.h` and `mlir._mlir_libs._mlirRegisterEverything` being the one-stop entry points for the "upstream packages". The `_mlir_libs.__init__.py` now allows customization of the environment and Context by adding "initialization modules" to the `_mlir_libs` package. If present, `_mlirRegisterEverything` is treated as such a module. Others can be added by downstreams by adding a `_site_initialize_{i}.py` module, where '{i}' is a number starting with zero. The number will be incremented and corresponding module loaded until one is not found. Initialization modules can:
* Perform load time customization to the global environment (i.e. registering passes, hooks, etc).
* Define a `register_dialects(registry: DialectRegistry)` function that can extend the `DialectRegistry` that will be used to bootstrap the `Context`.
* Define a `context_init_hook(context: Context)` function that will be added to a list of callbacks which will be invoked after dialect registration during `Context` initialization.
Note that the `MLIRPythonExtension.RegisterEverything` is not included by default when building a downstream (its corresponding behavior was prior). For downstreams which need the default MLIR initialization to take place, they must add this back in to their Python CMake build just like they add their own components (i.e. to `add_mlir_python_common_capi_library` and `add_mlir_python_modules`). It is perfectly valid to not do this, in which case, only the things explicitly depended on and initialized by downstreams will be built/packaged. If the downstream has not been set up for this, it is recommended to simply add this back for the time being and pay the build time/package size cost.
CMake changes:
* `MLIRCAPIRegistration` -> `MLIRCAPIRegisterEverything` (renamed to signify what it does and force an evaluation: a number of places were incidentally linking this very expensive target)
* `MLIRPythonSoure.Passes` removed (without replacement: just drop)
* `MLIRPythonExtension.AllPassesRegistration` removed (without replacement: just drop)
* `MLIRPythonExtension.Conversions` removed (without replacement: just drop)
* `MLIRPythonExtension.Transforms` removed (without replacement: just drop)
Header changes:
* `mlir-c/Registration.h` is deleted. Dialect registration functionality is now in `IR.h`. Registration of upstream features are in `mlir-c/RegisterEverything.h`. When updating MLIR and a couple of downstreams, I found that proper usage was commingled so required making a choice vs just blind S&R.
Python APIs removed:
* mlir.transforms and mlir.conversions (previously only had an __init__.py which indirectly triggered `mlirRegisterTransformsPasses()` and `mlirRegisterConversionPasses()` respectively). Downstream impact: Remove these imports if present (they now happen as part of default initialization).
* mlir._mlir_libs._all_passes_registration, mlir._mlir_libs._mlirTransforms, mlir._mlir_libs._mlirConversions. Downstream impact: None expected (these were internally used).
C-APIs changed:
* mlirRegisterAllDialects(MlirContext) now takes an MlirDialectRegistry instead. It also used to trigger loading of all dialects, which was already marked with a TODO to remove -- it no longer does, and for direct use, dialects must be explicitly loaded. Downstream impact: Direct C-API users must ensure that needed dialects are loaded or call `mlirContextLoadAllAvailableDialects(MlirContext)` to emulate the prior behavior. Also see the `ir.c` test case (e.g. ` mlirContextGetOrLoadDialect(ctx, mlirStringRefCreateFromCString("func"));`).
* mlirDialectHandle* APIs were moved from Registration.h (which now is restricted to just global/upstream registration) to IR.h, arguably where it should have been. Downstream impact: include correct header (likely already doing so).
C-APIs added:
* mlirContextLoadAllAvailableDialects(MlirContext): Corresponds to C++ API with the same purpose.
Python APIs added:
* mlir.ir.DialectRegistry: Mapping for an MlirDialectRegistry.
* mlir.ir.Context.append_dialect_registry(MlirDialectRegistry)
* mlir.ir.Context.load_all_available_dialects()
* mlir._mlir_libs._mlirAllRegistration: New native extension that exposes a `register_dialects(MlirDialectRegistry)` entry point and performs all upstream pass/conversion/transforms registration on init. In this first step, we eagerly load this as part of the __init__.py and use it to monkey patch the Context to emulate prior behavior.
* Type caster and capsule support for MlirDialectRegistry
This should make it possible to build downstream Python dialects that only depend on a subset of MLIR. See: https://github.com/llvm/llvm-project/issues/56037
Here is an example PR, minimally adapting IREE to these changes: https://github.com/iree-org/iree/pull/9638/files In this situation, IREE is opting to not link everything, since it is already configuring the Context to its liking. For projects that would just like to not think about it and pull in everything, add `MLIRPythonExtension.RegisterEverything` to the list of Python sources getting built, and the old behavior will continue.
Reviewed By: mehdi_amini, ftynse
Differential Revision: https://reviews.llvm.org/D128593
2022-07-16 16:09:03 -07:00
|
|
|
MlirDialectRegistry rawRegistry =
|
|
|
|
|
mlirPythonCapsuleToDialectRegistry(capsule.ptr());
|
|
|
|
|
if (mlirDialectRegistryIsNull(rawRegistry))
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::python_error();
|
[mlir] Overhaul C/Python registration APIs to properly scope registration/loading activities.
Since the very first commits, the Python and C MLIR APIs have had mis-placed registration/load functionality for dialects, extensions, etc. This was done pragmatically in order to get bootstrapped and then just grew in. Downstreams largely bypass and do their own thing by providing various APIs to register things they need. Meanwhile, the C++ APIs have stabilized around this and it would make sense to follow suit.
The thing we have observed in canonical usage by downstreams is that each downstream tends to have native entry points that configure its installation to its preferences with one-stop APIs. This patch leans in to this approach with `RegisterEverything.h` and `mlir._mlir_libs._mlirRegisterEverything` being the one-stop entry points for the "upstream packages". The `_mlir_libs.__init__.py` now allows customization of the environment and Context by adding "initialization modules" to the `_mlir_libs` package. If present, `_mlirRegisterEverything` is treated as such a module. Others can be added by downstreams by adding a `_site_initialize_{i}.py` module, where '{i}' is a number starting with zero. The number will be incremented and corresponding module loaded until one is not found. Initialization modules can:
* Perform load time customization to the global environment (i.e. registering passes, hooks, etc).
* Define a `register_dialects(registry: DialectRegistry)` function that can extend the `DialectRegistry` that will be used to bootstrap the `Context`.
* Define a `context_init_hook(context: Context)` function that will be added to a list of callbacks which will be invoked after dialect registration during `Context` initialization.
Note that the `MLIRPythonExtension.RegisterEverything` is not included by default when building a downstream (its corresponding behavior was prior). For downstreams which need the default MLIR initialization to take place, they must add this back in to their Python CMake build just like they add their own components (i.e. to `add_mlir_python_common_capi_library` and `add_mlir_python_modules`). It is perfectly valid to not do this, in which case, only the things explicitly depended on and initialized by downstreams will be built/packaged. If the downstream has not been set up for this, it is recommended to simply add this back for the time being and pay the build time/package size cost.
CMake changes:
* `MLIRCAPIRegistration` -> `MLIRCAPIRegisterEverything` (renamed to signify what it does and force an evaluation: a number of places were incidentally linking this very expensive target)
* `MLIRPythonSoure.Passes` removed (without replacement: just drop)
* `MLIRPythonExtension.AllPassesRegistration` removed (without replacement: just drop)
* `MLIRPythonExtension.Conversions` removed (without replacement: just drop)
* `MLIRPythonExtension.Transforms` removed (without replacement: just drop)
Header changes:
* `mlir-c/Registration.h` is deleted. Dialect registration functionality is now in `IR.h`. Registration of upstream features are in `mlir-c/RegisterEverything.h`. When updating MLIR and a couple of downstreams, I found that proper usage was commingled so required making a choice vs just blind S&R.
Python APIs removed:
* mlir.transforms and mlir.conversions (previously only had an __init__.py which indirectly triggered `mlirRegisterTransformsPasses()` and `mlirRegisterConversionPasses()` respectively). Downstream impact: Remove these imports if present (they now happen as part of default initialization).
* mlir._mlir_libs._all_passes_registration, mlir._mlir_libs._mlirTransforms, mlir._mlir_libs._mlirConversions. Downstream impact: None expected (these were internally used).
C-APIs changed:
* mlirRegisterAllDialects(MlirContext) now takes an MlirDialectRegistry instead. It also used to trigger loading of all dialects, which was already marked with a TODO to remove -- it no longer does, and for direct use, dialects must be explicitly loaded. Downstream impact: Direct C-API users must ensure that needed dialects are loaded or call `mlirContextLoadAllAvailableDialects(MlirContext)` to emulate the prior behavior. Also see the `ir.c` test case (e.g. ` mlirContextGetOrLoadDialect(ctx, mlirStringRefCreateFromCString("func"));`).
* mlirDialectHandle* APIs were moved from Registration.h (which now is restricted to just global/upstream registration) to IR.h, arguably where it should have been. Downstream impact: include correct header (likely already doing so).
C-APIs added:
* mlirContextLoadAllAvailableDialects(MlirContext): Corresponds to C++ API with the same purpose.
Python APIs added:
* mlir.ir.DialectRegistry: Mapping for an MlirDialectRegistry.
* mlir.ir.Context.append_dialect_registry(MlirDialectRegistry)
* mlir.ir.Context.load_all_available_dialects()
* mlir._mlir_libs._mlirAllRegistration: New native extension that exposes a `register_dialects(MlirDialectRegistry)` entry point and performs all upstream pass/conversion/transforms registration on init. In this first step, we eagerly load this as part of the __init__.py and use it to monkey patch the Context to emulate prior behavior.
* Type caster and capsule support for MlirDialectRegistry
This should make it possible to build downstream Python dialects that only depend on a subset of MLIR. See: https://github.com/llvm/llvm-project/issues/56037
Here is an example PR, minimally adapting IREE to these changes: https://github.com/iree-org/iree/pull/9638/files In this situation, IREE is opting to not link everything, since it is already configuring the Context to its liking. For projects that would just like to not think about it and pull in everything, add `MLIRPythonExtension.RegisterEverything` to the list of Python sources getting built, and the old behavior will continue.
Reviewed By: mehdi_amini, ftynse
Differential Revision: https://reviews.llvm.org/D128593
2022-07-16 16:09:03 -07:00
|
|
|
return PyDialectRegistry(rawRegistry);
|
|
|
|
|
}
|
|
|
|
|
|
2020-10-31 23:40:25 -07:00
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
// PyLocation
|
|
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object PyLocation::getCapsule() {
|
|
|
|
|
return nb::steal<nb::object>(mlirPythonLocationToCapsule(*this));
|
2020-11-29 13:30:23 -08:00
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
PyLocation PyLocation::createFromCapsule(nb::object capsule) {
|
2020-11-29 13:30:23 -08:00
|
|
|
MlirLocation rawLoc = mlirPythonCapsuleToLocation(capsule.ptr());
|
|
|
|
|
if (mlirLocationIsNull(rawLoc))
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::python_error();
|
2020-11-29 13:30:23 -08:00
|
|
|
return PyLocation(PyMlirContext::forContext(mlirLocationGetContext(rawLoc)),
|
|
|
|
|
rawLoc);
|
|
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object PyLocation::contextEnter(nb::object locationObj) {
|
|
|
|
|
return PyThreadContextEntry::pushLocation(locationObj);
|
2020-10-31 23:40:25 -07:00
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
void PyLocation::contextExit(const nb::object &excType,
|
|
|
|
|
const nb::object &excVal,
|
|
|
|
|
const nb::object &excTb) {
|
2020-10-31 23:40:25 -07:00
|
|
|
PyThreadContextEntry::popLocation(*this);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
PyLocation &DefaultingPyLocation::resolve() {
|
|
|
|
|
auto *location = PyThreadContextEntry::getDefaultLocation();
|
|
|
|
|
if (!location) {
|
2023-05-22 17:30:12 -05:00
|
|
|
throw std::runtime_error(
|
2020-10-31 23:40:25 -07:00
|
|
|
"An MLIR function requires a Location but none was provided in the "
|
|
|
|
|
"call or from the surrounding environment. Either pass to the function "
|
|
|
|
|
"with a 'loc=' argument or establish a default using 'with loc:'");
|
|
|
|
|
}
|
|
|
|
|
return *location;
|
|
|
|
|
}
|
|
|
|
|
|
2020-09-18 18:38:21 -07:00
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
// PyModule
|
|
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
|
2020-10-12 21:19:13 -07:00
|
|
|
PyModule::PyModule(PyMlirContextRef contextRef, MlirModule module)
|
|
|
|
|
: BaseContextObject(std::move(contextRef)), module(module) {}
|
|
|
|
|
|
|
|
|
|
PyModule::~PyModule() {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::gil_scoped_acquire acquire;
|
2020-10-12 21:19:13 -07:00
|
|
|
auto &liveModules = getContext()->liveModules;
|
|
|
|
|
assert(liveModules.count(module.ptr) == 1 &&
|
|
|
|
|
"destroying module not in live map");
|
|
|
|
|
liveModules.erase(module.ptr);
|
|
|
|
|
mlirModuleDestroy(module);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
PyModuleRef PyModule::forModule(MlirModule module) {
|
|
|
|
|
MlirContext context = mlirModuleGetContext(module);
|
|
|
|
|
PyMlirContextRef contextRef = PyMlirContext::forContext(context);
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::gil_scoped_acquire acquire;
|
2020-10-12 21:19:13 -07:00
|
|
|
auto &liveModules = contextRef->liveModules;
|
|
|
|
|
auto it = liveModules.find(module.ptr);
|
|
|
|
|
if (it == liveModules.end()) {
|
|
|
|
|
// Create.
|
|
|
|
|
PyModule *unownedModule = new PyModule(std::move(contextRef), module);
|
|
|
|
|
// Note that the default return value policy on cast is automatic_reference,
|
|
|
|
|
// which does not take ownership (delete will not be called).
|
|
|
|
|
// Just be explicit.
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object pyRef = nb::cast(unownedModule, nb::rv_policy::take_ownership);
|
2020-10-12 21:19:13 -07:00
|
|
|
unownedModule->handle = pyRef;
|
|
|
|
|
liveModules[module.ptr] =
|
|
|
|
|
std::make_pair(unownedModule->handle, unownedModule);
|
|
|
|
|
return PyModuleRef(unownedModule, std::move(pyRef));
|
|
|
|
|
}
|
|
|
|
|
// Use existing.
|
|
|
|
|
PyModule *existing = it->second.second;
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object pyRef = nb::borrow<nb::object>(it->second.first);
|
2020-10-12 21:19:13 -07:00
|
|
|
return PyModuleRef(existing, std::move(pyRef));
|
|
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object PyModule::createFromCapsule(nb::object capsule) {
|
2020-10-12 21:19:13 -07:00
|
|
|
MlirModule rawModule = mlirPythonCapsuleToModule(capsule.ptr());
|
|
|
|
|
if (mlirModuleIsNull(rawModule))
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::python_error();
|
2020-10-12 21:19:13 -07:00
|
|
|
return forModule(rawModule).releaseObject();
|
2020-09-18 18:38:21 -07:00
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object PyModule::getCapsule() {
|
|
|
|
|
return nb::steal<nb::object>(mlirPythonModuleToCapsule(get()));
|
2020-09-28 09:08:09 -07:00
|
|
|
}
|
|
|
|
|
|
2020-08-26 23:48:42 -07:00
|
|
|
//------------------------------------------------------------------------------
|
2020-09-18 18:38:21 -07:00
|
|
|
// PyOperation
|
|
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
|
|
|
|
|
PyOperation::PyOperation(PyMlirContextRef contextRef, MlirOperation operation)
|
|
|
|
|
: BaseContextObject(std::move(contextRef)), operation(operation) {}
|
|
|
|
|
|
|
|
|
|
PyOperation::~PyOperation() {
|
2021-04-23 20:54:04 -06:00
|
|
|
// If the operation has already been invalidated there is nothing to do.
|
|
|
|
|
if (!valid)
|
|
|
|
|
return;
|
2024-05-30 10:06:02 +02:00
|
|
|
|
|
|
|
|
// Otherwise, invalidate the operation and remove it from live map when it is
|
|
|
|
|
// attached.
|
|
|
|
|
if (isAttached()) {
|
|
|
|
|
getContext()->clearOperation(*this);
|
|
|
|
|
} else {
|
|
|
|
|
// And destroy it when it is detached, i.e. owned by Python, in which case
|
|
|
|
|
// all nested operations must be invalidated at removed from the live map as
|
|
|
|
|
// well.
|
|
|
|
|
erase();
|
2020-09-18 18:38:21 -07:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
PyOperationRef PyOperation::createInstance(PyMlirContextRef contextRef,
|
|
|
|
|
MlirOperation operation,
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object parentKeepAlive) {
|
2020-09-18 18:38:21 -07:00
|
|
|
auto &liveOperations = contextRef->liveOperations;
|
|
|
|
|
// Create.
|
|
|
|
|
PyOperation *unownedOperation =
|
|
|
|
|
new PyOperation(std::move(contextRef), operation);
|
|
|
|
|
// Note that the default return value policy on cast is automatic_reference,
|
|
|
|
|
// which does not take ownership (delete will not be called).
|
|
|
|
|
// Just be explicit.
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object pyRef = nb::cast(unownedOperation, nb::rv_policy::take_ownership);
|
2020-09-18 18:38:21 -07:00
|
|
|
unownedOperation->handle = pyRef;
|
|
|
|
|
if (parentKeepAlive) {
|
|
|
|
|
unownedOperation->parentKeepAlive = std::move(parentKeepAlive);
|
|
|
|
|
}
|
|
|
|
|
liveOperations[operation.ptr] = std::make_pair(pyRef, unownedOperation);
|
|
|
|
|
return PyOperationRef(unownedOperation, std::move(pyRef));
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
PyOperationRef PyOperation::forOperation(PyMlirContextRef contextRef,
|
|
|
|
|
MlirOperation operation,
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object parentKeepAlive) {
|
2020-09-18 18:38:21 -07:00
|
|
|
auto &liveOperations = contextRef->liveOperations;
|
|
|
|
|
auto it = liveOperations.find(operation.ptr);
|
|
|
|
|
if (it == liveOperations.end()) {
|
|
|
|
|
// Create.
|
|
|
|
|
return createInstance(std::move(contextRef), operation,
|
|
|
|
|
std::move(parentKeepAlive));
|
|
|
|
|
}
|
|
|
|
|
// Use existing.
|
|
|
|
|
PyOperation *existing = it->second.second;
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object pyRef = nb::borrow<nb::object>(it->second.first);
|
2020-09-18 18:38:21 -07:00
|
|
|
return PyOperationRef(existing, std::move(pyRef));
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
PyOperationRef PyOperation::createDetached(PyMlirContextRef contextRef,
|
|
|
|
|
MlirOperation operation,
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object parentKeepAlive) {
|
2020-09-18 18:38:21 -07:00
|
|
|
auto &liveOperations = contextRef->liveOperations;
|
|
|
|
|
assert(liveOperations.count(operation.ptr) == 0 &&
|
|
|
|
|
"cannot create detached operation that already exists");
|
|
|
|
|
(void)liveOperations;
|
|
|
|
|
|
|
|
|
|
PyOperationRef created = createInstance(std::move(contextRef), operation,
|
|
|
|
|
std::move(parentKeepAlive));
|
|
|
|
|
created->attached = false;
|
|
|
|
|
return created;
|
|
|
|
|
}
|
|
|
|
|
|
2022-11-08 16:55:06 -05:00
|
|
|
PyOperationRef PyOperation::parse(PyMlirContextRef contextRef,
|
|
|
|
|
const std::string &sourceStr,
|
|
|
|
|
const std::string &sourceName) {
|
[mlir][python] Capture error diagnostics in exceptions
This updates most (all?) error-diagnostic-emitting python APIs to
capture error diagnostics and include them in the raised exception's
message:
```
>>> Operation.parse('"arith.addi"() : () -> ()'))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
mlir._mlir_libs.MLIRError: Unable to parse operation assembly:
error: "-":1:1: 'arith.addi' op requires one result
note: "-":1:1: see current operation: "arith.addi"() : () -> ()
```
The diagnostic information is available on the exception for users who
may want to customize the error message:
```
>>> try:
... Operation.parse('"arith.addi"() : () -> ()')
... except MLIRError as e:
... print(e.message)
... print(e.error_diagnostics)
... print(e.error_diagnostics[0].message)
...
Unable to parse operation assembly
[<mlir._mlir_libs._mlir.ir.DiagnosticInfo object at 0x7fed32bd6b70>]
'arith.addi' op requires one result
```
Error diagnostics captured in exceptions aren't propagated to diagnostic
handlers, to avoid double-reporting of errors. The context-level
`emit_error_diagnostics` option can be used to revert to the old
behaviour, causing error diagnostics to be reported to handlers instead
of as part of exceptions.
API changes:
- `Operation.verify` now raises an exception on verification failure,
instead of returning `false`
- The exception raised by the following methods has been changed to
`MLIRError`:
- `PassManager.run`
- `{Module,Operation,Type,Attribute}.parse`
- `{RankedTensorType,UnrankedTensorType}.get`
- `{MemRefType,UnrankedMemRefType}.get`
- `VectorType.get`
- `FloatAttr.get`
closes #60595
depends on D144804, D143830
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D143869
2023-02-07 16:07:50 -05:00
|
|
|
PyMlirContext::ErrorCapture errors(contextRef);
|
2022-11-08 16:55:06 -05:00
|
|
|
MlirOperation op =
|
|
|
|
|
mlirOperationCreateParse(contextRef->get(), toMlirStringRef(sourceStr),
|
|
|
|
|
toMlirStringRef(sourceName));
|
|
|
|
|
if (mlirOperationIsNull(op))
|
[mlir][python] Capture error diagnostics in exceptions
This updates most (all?) error-diagnostic-emitting python APIs to
capture error diagnostics and include them in the raised exception's
message:
```
>>> Operation.parse('"arith.addi"() : () -> ()'))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
mlir._mlir_libs.MLIRError: Unable to parse operation assembly:
error: "-":1:1: 'arith.addi' op requires one result
note: "-":1:1: see current operation: "arith.addi"() : () -> ()
```
The diagnostic information is available on the exception for users who
may want to customize the error message:
```
>>> try:
... Operation.parse('"arith.addi"() : () -> ()')
... except MLIRError as e:
... print(e.message)
... print(e.error_diagnostics)
... print(e.error_diagnostics[0].message)
...
Unable to parse operation assembly
[<mlir._mlir_libs._mlir.ir.DiagnosticInfo object at 0x7fed32bd6b70>]
'arith.addi' op requires one result
```
Error diagnostics captured in exceptions aren't propagated to diagnostic
handlers, to avoid double-reporting of errors. The context-level
`emit_error_diagnostics` option can be used to revert to the old
behaviour, causing error diagnostics to be reported to handlers instead
of as part of exceptions.
API changes:
- `Operation.verify` now raises an exception on verification failure,
instead of returning `false`
- The exception raised by the following methods has been changed to
`MLIRError`:
- `PassManager.run`
- `{Module,Operation,Type,Attribute}.parse`
- `{RankedTensorType,UnrankedTensorType}.get`
- `{MemRefType,UnrankedMemRefType}.get`
- `VectorType.get`
- `FloatAttr.get`
closes #60595
depends on D144804, D143830
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D143869
2023-02-07 16:07:50 -05:00
|
|
|
throw MLIRError("Unable to parse operation assembly", errors.take());
|
2022-11-08 16:55:06 -05:00
|
|
|
return PyOperation::createDetached(std::move(contextRef), op);
|
|
|
|
|
}
|
|
|
|
|
|
2020-11-29 13:52:11 -08:00
|
|
|
void PyOperation::checkValid() const {
|
2020-09-18 18:38:21 -07:00
|
|
|
if (!valid) {
|
2023-05-22 17:30:12 -05:00
|
|
|
throw std::runtime_error("the operation has been invalidated");
|
2020-09-18 18:38:21 -07:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2023-11-13 10:21:21 -08:00
|
|
|
void PyOperationBase::print(std::optional<int64_t> largeElementsLimit,
|
2020-11-01 23:05:36 -08:00
|
|
|
bool enableDebugInfo, bool prettyDebugInfo,
|
2021-11-28 15:33:03 -08:00
|
|
|
bool printGenericOpForm, bool useLocalScope,
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
bool assumeVerified, nb::object fileObject,
|
2024-06-20 17:15:08 +02:00
|
|
|
bool binary, bool skipRegions) {
|
2020-11-01 23:05:36 -08:00
|
|
|
PyOperation &operation = getOperation();
|
|
|
|
|
operation.checkValid();
|
2020-10-20 23:20:04 -07:00
|
|
|
if (fileObject.is_none())
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
fileObject = nb::module_::import_("sys").attr("stdout");
|
2020-12-03 17:45:28 +00:00
|
|
|
|
2020-10-20 23:20:04 -07:00
|
|
|
MlirOpPrintingFlags flags = mlirOpPrintingFlagsCreate();
|
|
|
|
|
if (largeElementsLimit)
|
|
|
|
|
mlirOpPrintingFlagsElideLargeElementsAttrs(flags, *largeElementsLimit);
|
|
|
|
|
if (enableDebugInfo)
|
2022-11-17 20:44:27 -08:00
|
|
|
mlirOpPrintingFlagsEnableDebugInfo(flags, /*enable=*/true,
|
|
|
|
|
/*prettyForm=*/prettyDebugInfo);
|
2020-10-20 23:20:04 -07:00
|
|
|
if (printGenericOpForm)
|
|
|
|
|
mlirOpPrintingFlagsPrintGenericOpForm(flags);
|
2022-06-15 21:59:57 -07:00
|
|
|
if (useLocalScope)
|
|
|
|
|
mlirOpPrintingFlagsUseLocalScope(flags);
|
2023-02-25 03:51:31 -05:00
|
|
|
if (assumeVerified)
|
|
|
|
|
mlirOpPrintingFlagsAssumeVerified(flags);
|
2024-06-20 17:15:08 +02:00
|
|
|
if (skipRegions)
|
|
|
|
|
mlirOpPrintingFlagsSkipRegions(flags);
|
2020-10-20 23:20:04 -07:00
|
|
|
|
|
|
|
|
PyFileAccumulator accum(fileObject, binary);
|
2020-11-29 13:52:11 -08:00
|
|
|
mlirOperationPrintWithFlags(operation, flags, accum.getCallback(),
|
2020-10-20 23:20:04 -07:00
|
|
|
accum.getUserData());
|
|
|
|
|
mlirOpPrintingFlagsDestroy(flags);
|
|
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
void PyOperationBase::print(PyAsmState &state, nb::object fileObject,
|
2023-11-13 10:21:21 -08:00
|
|
|
bool binary) {
|
|
|
|
|
PyOperation &operation = getOperation();
|
|
|
|
|
operation.checkValid();
|
|
|
|
|
if (fileObject.is_none())
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
fileObject = nb::module_::import_("sys").attr("stdout");
|
2023-11-13 10:21:21 -08:00
|
|
|
PyFileAccumulator accum(fileObject, binary);
|
|
|
|
|
mlirOperationPrintWithState(operation, state.get(), accum.getCallback(),
|
|
|
|
|
accum.getUserData());
|
|
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
void PyOperationBase::writeBytecode(const nb::object &fileObject,
|
2023-04-30 22:11:02 -07:00
|
|
|
std::optional<int64_t> bytecodeVersion) {
|
2022-09-05 11:54:19 +00:00
|
|
|
PyOperation &operation = getOperation();
|
|
|
|
|
operation.checkValid();
|
|
|
|
|
PyFileAccumulator accum(fileObject, /*binary=*/true);
|
2023-04-29 05:35:53 -07:00
|
|
|
|
|
|
|
|
if (!bytecodeVersion.has_value())
|
|
|
|
|
return mlirOperationWriteBytecode(operation, accum.getCallback(),
|
|
|
|
|
accum.getUserData());
|
|
|
|
|
|
|
|
|
|
MlirBytecodeWriterConfig config = mlirBytecodeWriterConfigCreate();
|
|
|
|
|
mlirBytecodeWriterConfigDesiredEmitVersion(config, *bytecodeVersion);
|
2023-04-30 22:11:02 -07:00
|
|
|
MlirLogicalResult res = mlirOperationWriteBytecodeWithConfig(
|
2023-04-29 05:35:53 -07:00
|
|
|
operation, config, accum.getCallback(), accum.getUserData());
|
2023-06-21 09:40:31 -07:00
|
|
|
mlirBytecodeWriterConfigDestroy(config);
|
2023-04-30 22:11:02 -07:00
|
|
|
if (mlirLogicalResultIsFailure(res))
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error((Twine("Unable to honor desired bytecode version ") +
|
2023-04-30 22:11:02 -07:00
|
|
|
Twine(*bytecodeVersion))
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.str()
|
|
|
|
|
.c_str());
|
2022-09-05 11:54:19 +00:00
|
|
|
}
|
|
|
|
|
|
2024-04-17 15:09:47 +09:00
|
|
|
void PyOperationBase::walk(
|
|
|
|
|
std::function<MlirWalkResult(MlirOperation)> callback,
|
|
|
|
|
MlirWalkOrder walkOrder) {
|
|
|
|
|
PyOperation &operation = getOperation();
|
|
|
|
|
operation.checkValid();
|
2024-04-18 15:09:31 +01:00
|
|
|
struct UserData {
|
|
|
|
|
std::function<MlirWalkResult(MlirOperation)> callback;
|
|
|
|
|
bool gotException;
|
|
|
|
|
std::string exceptionWhat;
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object exceptionType;
|
2024-04-18 15:09:31 +01:00
|
|
|
};
|
|
|
|
|
UserData userData{callback, false, {}, {}};
|
2024-04-17 15:09:47 +09:00
|
|
|
MlirOperationWalkCallback walkCallback = [](MlirOperation op,
|
|
|
|
|
void *userData) {
|
2024-04-18 15:09:31 +01:00
|
|
|
UserData *calleeUserData = static_cast<UserData *>(userData);
|
|
|
|
|
try {
|
|
|
|
|
return (calleeUserData->callback)(op);
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
} catch (nb::python_error &e) {
|
2024-04-18 15:09:31 +01:00
|
|
|
calleeUserData->gotException = true;
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
calleeUserData->exceptionWhat = std::string(e.what());
|
|
|
|
|
calleeUserData->exceptionType = nb::borrow(e.type());
|
2024-04-18 15:09:31 +01:00
|
|
|
return MlirWalkResult::MlirWalkResultInterrupt;
|
|
|
|
|
}
|
2024-04-17 15:09:47 +09:00
|
|
|
};
|
2024-04-18 15:09:31 +01:00
|
|
|
mlirOperationWalk(operation, walkCallback, &userData, walkOrder);
|
|
|
|
|
if (userData.gotException) {
|
|
|
|
|
std::string message("Exception raised in callback: ");
|
|
|
|
|
message.append(userData.exceptionWhat);
|
|
|
|
|
throw std::runtime_error(message);
|
|
|
|
|
}
|
2024-04-17 15:09:47 +09:00
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object PyOperationBase::getAsm(bool binary,
|
2023-01-14 01:25:58 -08:00
|
|
|
std::optional<int64_t> largeElementsLimit,
|
2020-11-01 23:05:36 -08:00
|
|
|
bool enableDebugInfo, bool prettyDebugInfo,
|
2021-11-28 15:33:03 -08:00
|
|
|
bool printGenericOpForm, bool useLocalScope,
|
2024-06-20 17:15:08 +02:00
|
|
|
bool assumeVerified, bool skipRegions) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object fileObject;
|
2020-10-20 23:20:04 -07:00
|
|
|
if (binary) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
fileObject = nb::module_::import_("io").attr("BytesIO")();
|
2020-10-20 23:20:04 -07:00
|
|
|
} else {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
fileObject = nb::module_::import_("io").attr("StringIO")();
|
2020-10-20 23:20:04 -07:00
|
|
|
}
|
2023-11-13 10:21:21 -08:00
|
|
|
print(/*largeElementsLimit=*/largeElementsLimit,
|
2020-10-20 23:20:04 -07:00
|
|
|
/*enableDebugInfo=*/enableDebugInfo,
|
|
|
|
|
/*prettyDebugInfo=*/prettyDebugInfo,
|
|
|
|
|
/*printGenericOpForm=*/printGenericOpForm,
|
2021-11-28 15:33:03 -08:00
|
|
|
/*useLocalScope=*/useLocalScope,
|
2023-11-13 10:21:21 -08:00
|
|
|
/*assumeVerified=*/assumeVerified,
|
|
|
|
|
/*fileObject=*/fileObject,
|
2024-06-20 17:15:08 +02:00
|
|
|
/*binary=*/binary,
|
|
|
|
|
/*skipRegions=*/skipRegions);
|
2020-10-20 23:20:04 -07:00
|
|
|
|
|
|
|
|
return fileObject.attr("getvalue")();
|
|
|
|
|
}
|
|
|
|
|
|
2021-10-31 09:37:20 +01:00
|
|
|
void PyOperationBase::moveAfter(PyOperationBase &other) {
|
|
|
|
|
PyOperation &operation = getOperation();
|
|
|
|
|
PyOperation &otherOp = other.getOperation();
|
|
|
|
|
operation.checkValid();
|
|
|
|
|
otherOp.checkValid();
|
|
|
|
|
mlirOperationMoveAfter(operation, otherOp);
|
|
|
|
|
operation.parentKeepAlive = otherOp.parentKeepAlive;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void PyOperationBase::moveBefore(PyOperationBase &other) {
|
|
|
|
|
PyOperation &operation = getOperation();
|
|
|
|
|
PyOperation &otherOp = other.getOperation();
|
|
|
|
|
operation.checkValid();
|
|
|
|
|
otherOp.checkValid();
|
|
|
|
|
mlirOperationMoveBefore(operation, otherOp);
|
|
|
|
|
operation.parentKeepAlive = otherOp.parentKeepAlive;
|
|
|
|
|
}
|
|
|
|
|
|
[mlir][python] Capture error diagnostics in exceptions
This updates most (all?) error-diagnostic-emitting python APIs to
capture error diagnostics and include them in the raised exception's
message:
```
>>> Operation.parse('"arith.addi"() : () -> ()'))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
mlir._mlir_libs.MLIRError: Unable to parse operation assembly:
error: "-":1:1: 'arith.addi' op requires one result
note: "-":1:1: see current operation: "arith.addi"() : () -> ()
```
The diagnostic information is available on the exception for users who
may want to customize the error message:
```
>>> try:
... Operation.parse('"arith.addi"() : () -> ()')
... except MLIRError as e:
... print(e.message)
... print(e.error_diagnostics)
... print(e.error_diagnostics[0].message)
...
Unable to parse operation assembly
[<mlir._mlir_libs._mlir.ir.DiagnosticInfo object at 0x7fed32bd6b70>]
'arith.addi' op requires one result
```
Error diagnostics captured in exceptions aren't propagated to diagnostic
handlers, to avoid double-reporting of errors. The context-level
`emit_error_diagnostics` option can be used to revert to the old
behaviour, causing error diagnostics to be reported to handlers instead
of as part of exceptions.
API changes:
- `Operation.verify` now raises an exception on verification failure,
instead of returning `false`
- The exception raised by the following methods has been changed to
`MLIRError`:
- `PassManager.run`
- `{Module,Operation,Type,Attribute}.parse`
- `{RankedTensorType,UnrankedTensorType}.get`
- `{MemRefType,UnrankedMemRefType}.get`
- `VectorType.get`
- `FloatAttr.get`
closes #60595
depends on D144804, D143830
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D143869
2023-02-07 16:07:50 -05:00
|
|
|
bool PyOperationBase::verify() {
|
|
|
|
|
PyOperation &op = getOperation();
|
|
|
|
|
PyMlirContext::ErrorCapture errors(op.getContext());
|
|
|
|
|
if (!mlirOperationVerify(op.get()))
|
|
|
|
|
throw MLIRError("Verification failed", errors.take());
|
|
|
|
|
return true;
|
|
|
|
|
}
|
|
|
|
|
|
2023-01-14 01:25:58 -08:00
|
|
|
std::optional<PyOperationRef> PyOperation::getParentOperation() {
|
2021-04-23 20:54:04 -06:00
|
|
|
checkValid();
|
2020-10-28 23:16:36 -07:00
|
|
|
if (!isAttached())
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error("Detached operations have no parent");
|
2020-10-28 23:16:36 -07:00
|
|
|
MlirOperation operation = mlirOperationGetParentOperation(get());
|
|
|
|
|
if (mlirOperationIsNull(operation))
|
2021-08-16 22:37:14 -07:00
|
|
|
return {};
|
2020-10-28 23:16:36 -07:00
|
|
|
return PyOperation::forOperation(getContext(), operation);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
PyBlock PyOperation::getBlock() {
|
2021-04-23 20:54:04 -06:00
|
|
|
checkValid();
|
2023-01-14 01:25:58 -08:00
|
|
|
std::optional<PyOperationRef> parentOperation = getParentOperation();
|
2020-10-28 23:16:36 -07:00
|
|
|
MlirBlock block = mlirOperationGetBlock(get());
|
|
|
|
|
assert(!mlirBlockIsNull(block) && "Attached operation has null parent");
|
2021-08-16 22:37:14 -07:00
|
|
|
assert(parentOperation && "Operation has no parent");
|
|
|
|
|
return PyBlock{std::move(*parentOperation), block};
|
2020-10-28 23:16:36 -07:00
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object PyOperation::getCapsule() {
|
2021-04-23 20:54:04 -06:00
|
|
|
checkValid();
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
return nb::steal<nb::object>(mlirPythonOperationToCapsule(get()));
|
2021-04-06 14:15:22 -07:00
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object PyOperation::createFromCapsule(nb::object capsule) {
|
2021-04-06 14:15:22 -07:00
|
|
|
MlirOperation rawOperation = mlirPythonCapsuleToOperation(capsule.ptr());
|
|
|
|
|
if (mlirOperationIsNull(rawOperation))
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::python_error();
|
2021-04-06 14:15:22 -07:00
|
|
|
MlirContext rawCtxt = mlirOperationGetContext(rawOperation);
|
Revert "[mlir][python] Make the Context/Operation capsule creation methods work as documented. (#76010)"
This reverts commit bbc29768683b394b34600347f46be2b8245ddb30.
This change seems to be at odds with the non-owning part semantics of
MlirOperation in C API. Since downstream clients can only take and
return MlirOperation, it does not sound correct to force all returns of
MlirOperation transfer ownership. Specifically, this makes it impossible
for downstreams to implement IR-traversing functions that, e.g., look at
neighbors of an operation.
The following patch triggers the exception, and there does not seem to
be an alternative way for a downstream binding writer to express this:
```
diff --git a/mlir/lib/Bindings/Python/IRCore.cpp b/mlir/lib/Bindings/Python/IRCore.cpp
index 39757dfad5be..2ce640674245 100644
--- a/mlir/lib/Bindings/Python/IRCore.cpp
+++ b/mlir/lib/Bindings/Python/IRCore.cpp
@@ -3071,6 +3071,11 @@ void mlir::python::populateIRCore(py::module &m) {
py::arg("successors") = py::none(), py::arg("regions") = 0,
py::arg("loc") = py::none(), py::arg("ip") = py::none(),
py::arg("infer_type") = false, kOperationCreateDocstring)
+ .def("_get_first_in_block", [](PyOperation &self) -> MlirOperation {
+ MlirBlock block = mlirOperationGetBlock(self.get());
+ MlirOperation first = mlirBlockGetFirstOperation(block);
+ return first;
+ })
.def_static(
"parse",
[](const std::string &sourceStr, const std::string &sourceName,
diff --git a/mlir/test/python/ir/operation.py b/mlir/test/python/ir/operation.py
index f59b1a26ba48..6b12b8da5c24 100644
--- a/mlir/test/python/ir/operation.py
+++ b/mlir/test/python/ir/operation.py
@@ -24,6 +24,25 @@ def expect_index_error(callback):
except IndexError:
pass
+@run
+def testCustomBind():
+ ctx = Context()
+ ctx.allow_unregistered_dialects = True
+ module = Module.parse(
+ r"""
+ func.func @f1(%arg0: i32) -> i32 {
+ %1 = "custom.addi"(%arg0, %arg0) : (i32, i32) -> i32
+ return %1 : i32
+ }
+ """,
+ ctx,
+ )
+ add = module.body.operations[0].regions[0].blocks[0].operations[0]
+ op = add.operation
+ # This will get a reference to itself.
+ f1 = op._get_first_in_block()
+
+
# Verify iterator based traversal of the op/region/block hierarchy.
# CHECK-LABEL: TEST: testTraverseOpRegionBlockIterators
```
2023-12-21 10:01:44 +00:00
|
|
|
return forOperation(PyMlirContext::forContext(rawCtxt), rawOperation)
|
2021-04-06 14:15:22 -07:00
|
|
|
.releaseObject();
|
|
|
|
|
}
|
|
|
|
|
|
2022-03-28 15:45:40 +02:00
|
|
|
static void maybeInsertOperation(PyOperationRef &op,
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
const nb::object &maybeIp) {
|
2022-03-28 15:45:40 +02:00
|
|
|
// InsertPoint active?
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
if (!maybeIp.is(nb::cast(false))) {
|
2022-03-28 15:45:40 +02:00
|
|
|
PyInsertionPoint *ip;
|
|
|
|
|
if (maybeIp.is_none()) {
|
|
|
|
|
ip = PyThreadContextEntry::getDefaultInsertionPoint();
|
|
|
|
|
} else {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
ip = nb::cast<PyInsertionPoint *>(maybeIp);
|
2022-03-28 15:45:40 +02:00
|
|
|
}
|
|
|
|
|
if (ip)
|
|
|
|
|
ip->insert(*op.get());
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object PyOperation::create(const std::string &name,
|
2023-01-14 01:25:58 -08:00
|
|
|
std::optional<std::vector<PyType *>> results,
|
|
|
|
|
std::optional<std::vector<PyValue *>> operands,
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
std::optional<nb::dict> attributes,
|
2023-01-14 01:25:58 -08:00
|
|
|
std::optional<std::vector<PyBlock *>> successors,
|
|
|
|
|
int regions, DefaultingPyLocation location,
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
const nb::object &maybeIp, bool inferType) {
|
2020-10-31 23:40:25 -07:00
|
|
|
llvm::SmallVector<MlirValue, 4> mlirOperands;
|
|
|
|
|
llvm::SmallVector<MlirType, 4> mlirResults;
|
|
|
|
|
llvm::SmallVector<MlirBlock, 4> mlirSuccessors;
|
|
|
|
|
llvm::SmallVector<std::pair<std::string, MlirAttribute>, 4> mlirAttributes;
|
|
|
|
|
|
|
|
|
|
// General parameter validation.
|
|
|
|
|
if (regions < 0)
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error("number of regions must be >= 0");
|
2020-10-31 23:40:25 -07:00
|
|
|
|
|
|
|
|
// Unpack/validate operands.
|
|
|
|
|
if (operands) {
|
|
|
|
|
mlirOperands.reserve(operands->size());
|
|
|
|
|
for (PyValue *operand : *operands) {
|
|
|
|
|
if (!operand)
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error("operand value cannot be None");
|
2020-10-31 23:40:25 -07:00
|
|
|
mlirOperands.push_back(operand->get());
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Unpack/validate results.
|
|
|
|
|
if (results) {
|
|
|
|
|
mlirResults.reserve(results->size());
|
|
|
|
|
for (PyType *result : *results) {
|
|
|
|
|
// TODO: Verify result type originate from the same context.
|
|
|
|
|
if (!result)
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error("result type cannot be None");
|
2020-11-29 13:30:23 -08:00
|
|
|
mlirResults.push_back(*result);
|
2020-10-31 23:40:25 -07:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
// Unpack/validate attributes.
|
|
|
|
|
if (attributes) {
|
|
|
|
|
mlirAttributes.reserve(attributes->size());
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
for (std::pair<nb::handle, nb::handle> it : *attributes) {
|
2020-12-05 01:28:41 +00:00
|
|
|
std::string key;
|
|
|
|
|
try {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
key = nb::cast<std::string>(it.first);
|
|
|
|
|
} catch (nb::cast_error &err) {
|
2020-12-05 01:28:41 +00:00
|
|
|
std::string msg = "Invalid attribute key (not a string) when "
|
|
|
|
|
"attempting to create the operation \"" +
|
|
|
|
|
name + "\" (" + err.what() + ")";
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::type_error(msg.c_str());
|
2020-12-05 01:28:41 +00:00
|
|
|
}
|
|
|
|
|
try {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
auto &attribute = nb::cast<PyAttribute &>(it.second);
|
2020-12-05 01:28:41 +00:00
|
|
|
// TODO: Verify attribute originates from the same context.
|
|
|
|
|
mlirAttributes.emplace_back(std::move(key), attribute);
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
} catch (nb::cast_error &err) {
|
|
|
|
|
std::string msg = "Invalid attribute value for the key \"" + key +
|
|
|
|
|
"\" when attempting to create the operation \"" +
|
|
|
|
|
name + "\" (" + err.what() + ")";
|
|
|
|
|
throw nb::type_error(msg.c_str());
|
|
|
|
|
} catch (std::runtime_error &) {
|
2020-12-05 01:28:41 +00:00
|
|
|
// This exception seems thrown when the value is "None".
|
|
|
|
|
std::string msg =
|
|
|
|
|
"Found an invalid (`None`?) attribute value for the key \"" + key +
|
|
|
|
|
"\" when attempting to create the operation \"" + name + "\"";
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw std::runtime_error(msg);
|
2020-12-05 01:28:41 +00:00
|
|
|
}
|
2020-10-31 23:40:25 -07:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
// Unpack/validate successors.
|
|
|
|
|
if (successors) {
|
|
|
|
|
mlirSuccessors.reserve(successors->size());
|
|
|
|
|
for (auto *successor : *successors) {
|
|
|
|
|
// TODO: Verify successor originate from the same context.
|
|
|
|
|
if (!successor)
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error("successor block cannot be None");
|
2020-10-31 23:40:25 -07:00
|
|
|
mlirSuccessors.push_back(successor->get());
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Apply unpacked/validated to the operation state. Beyond this
|
|
|
|
|
// point, exceptions cannot be thrown or else the state will leak.
|
2020-11-24 18:35:22 +00:00
|
|
|
MlirOperationState state =
|
2020-11-29 13:30:23 -08:00
|
|
|
mlirOperationStateGet(toMlirStringRef(name), location);
|
2020-10-31 23:40:25 -07:00
|
|
|
if (!mlirOperands.empty())
|
|
|
|
|
mlirOperationStateAddOperands(&state, mlirOperands.size(),
|
|
|
|
|
mlirOperands.data());
|
2023-07-23 21:26:52 -07:00
|
|
|
state.enableResultTypeInference = inferType;
|
2020-10-31 23:40:25 -07:00
|
|
|
if (!mlirResults.empty())
|
|
|
|
|
mlirOperationStateAddResults(&state, mlirResults.size(),
|
|
|
|
|
mlirResults.data());
|
|
|
|
|
if (!mlirAttributes.empty()) {
|
|
|
|
|
// Note that the attribute names directly reference bytes in
|
|
|
|
|
// mlirAttributes, so that vector must not be changed from here
|
|
|
|
|
// on.
|
|
|
|
|
llvm::SmallVector<MlirNamedAttribute, 4> mlirNamedAttributes;
|
|
|
|
|
mlirNamedAttributes.reserve(mlirAttributes.size());
|
|
|
|
|
for (auto &it : mlirAttributes)
|
2020-12-11 18:50:04 +00:00
|
|
|
mlirNamedAttributes.push_back(mlirNamedAttributeGet(
|
|
|
|
|
mlirIdentifierGet(mlirAttributeGetContext(it.second),
|
|
|
|
|
toMlirStringRef(it.first)),
|
|
|
|
|
it.second));
|
2020-10-31 23:40:25 -07:00
|
|
|
mlirOperationStateAddAttributes(&state, mlirNamedAttributes.size(),
|
|
|
|
|
mlirNamedAttributes.data());
|
|
|
|
|
}
|
|
|
|
|
if (!mlirSuccessors.empty())
|
|
|
|
|
mlirOperationStateAddSuccessors(&state, mlirSuccessors.size(),
|
|
|
|
|
mlirSuccessors.data());
|
|
|
|
|
if (regions) {
|
|
|
|
|
llvm::SmallVector<MlirRegion, 4> mlirRegions;
|
|
|
|
|
mlirRegions.resize(regions);
|
|
|
|
|
for (int i = 0; i < regions; ++i)
|
|
|
|
|
mlirRegions[i] = mlirRegionCreate();
|
|
|
|
|
mlirOperationStateAddOwnedRegions(&state, mlirRegions.size(),
|
|
|
|
|
mlirRegions.data());
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Construct the operation.
|
|
|
|
|
MlirOperation operation = mlirOperationCreate(&state);
|
2023-07-23 21:26:52 -07:00
|
|
|
if (!operation.ptr)
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error("Operation creation failed");
|
2020-10-31 23:40:25 -07:00
|
|
|
PyOperationRef created =
|
|
|
|
|
PyOperation::createDetached(location->getContext(), operation);
|
2022-03-28 15:45:40 +02:00
|
|
|
maybeInsertOperation(created, maybeIp);
|
2020-10-31 23:40:25 -07:00
|
|
|
|
[mlir:python] Change PyOperation::create to actually return a PyOperation. (#114542)
In the tablegen-generated Python bindings, we typically see a pattern
like:
```
class ConstantOp(_ods_ir.OpView):
...
def __init__(self, value, *, loc=None, ip=None):
...
super().__init__(self.build_generic(attributes=attributes, operands=operands, successors=_ods_successors, regions=regions, loc=loc, ip=ip))
```
i.e., the generated code calls `OpView.__init__()` with the output of
`build_generic`. The purpose of `OpView` is to wrap another operation
object, and `OpView.__init__` can accept any `PyOperationBase` subclass,
and presumably the intention is that `build_generic` returns a
`PyOperation`, so the user ends up with a `PyOpView` wrapping a
`PyOperation`.
However, `PyOpView::buildGeneric` calls `PyOperation::create`, which
does not just build a PyOperation, but it also calls `createOpView` to
wrap that operation in a subclass of `PyOpView` and returns that view.
But that's rather pointless: we called this code from the constructor of
an `OpView` subclass, so we already have a view object ready to go; we
don't need to build another one!
If we change `PyOperation::create` to return the underlying
`PyOperation`, rather than a view wrapper, we can save allocating a
useless `PyOpView` object for each ODS-generated Python object.
This saves approximately 1.5s of Python time in a JAX LLM benchmark that
generates a mixture of upstream dialects and StableHLO.
2024-11-04 20:19:18 -05:00
|
|
|
return created.getObject();
|
2020-10-31 23:40:25 -07:00
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object PyOperation::clone(const nb::object &maybeIp) {
|
2022-03-28 15:45:40 +02:00
|
|
|
MlirOperation clonedOperation = mlirOperationClone(operation);
|
|
|
|
|
PyOperationRef cloned =
|
|
|
|
|
PyOperation::createDetached(getContext(), clonedOperation);
|
|
|
|
|
maybeInsertOperation(cloned, maybeIp);
|
|
|
|
|
|
|
|
|
|
return cloned->createOpView();
|
|
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object PyOperation::createOpView() {
|
2021-04-23 20:54:04 -06:00
|
|
|
checkValid();
|
2020-11-01 23:05:36 -08:00
|
|
|
MlirIdentifier ident = mlirOperationGetName(get());
|
|
|
|
|
MlirStringRef identStr = mlirIdentifierStr(ident);
|
2023-01-22 23:31:18 -05:00
|
|
|
auto operationCls = PyGlobals::get().lookupOperationClass(
|
2020-11-29 13:52:11 -08:00
|
|
|
StringRef(identStr.data, identStr.length));
|
2023-01-22 23:31:18 -05:00
|
|
|
if (operationCls)
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
return PyOpView::constructDerived(*operationCls, getRef().getObject());
|
|
|
|
|
return nb::cast(PyOpView(getRef().getObject()));
|
2020-11-01 23:05:36 -08:00
|
|
|
}
|
|
|
|
|
|
2021-04-23 20:54:04 -06:00
|
|
|
void PyOperation::erase() {
|
|
|
|
|
checkValid();
|
2024-05-30 10:06:02 +02:00
|
|
|
getContext()->clearOperationAndInside(*this);
|
2021-04-23 20:54:04 -06:00
|
|
|
mlirOperationDestroy(operation);
|
|
|
|
|
}
|
|
|
|
|
|
2021-01-18 11:27:19 -08:00
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
// PyOpView
|
|
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
static void populateResultTypes(StringRef name, nb::list resultTypeList,
|
|
|
|
|
const nb::object &resultSegmentSpecObj,
|
2023-07-23 21:26:52 -07:00
|
|
|
std::vector<int32_t> &resultSegmentLengths,
|
|
|
|
|
std::vector<PyType *> &resultTypes) {
|
2021-01-18 11:27:19 -08:00
|
|
|
resultTypes.reserve(resultTypeList.size());
|
|
|
|
|
if (resultSegmentSpecObj.is_none()) {
|
|
|
|
|
// Non-variadic result unpacking.
|
2022-01-02 22:02:14 +00:00
|
|
|
for (const auto &it : llvm::enumerate(resultTypeList)) {
|
2021-01-18 11:27:19 -08:00
|
|
|
try {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
resultTypes.push_back(nb::cast<PyType *>(it.value()));
|
2021-01-18 11:27:19 -08:00
|
|
|
if (!resultTypes.back())
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::cast_error();
|
|
|
|
|
} catch (nb::cast_error &err) {
|
|
|
|
|
throw nb::value_error((llvm::Twine("Result ") +
|
2021-01-18 11:27:19 -08:00
|
|
|
llvm::Twine(it.index()) + " of operation \"" +
|
|
|
|
|
name + "\" must be a Type (" + err.what() + ")")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.str()
|
|
|
|
|
.c_str());
|
2021-01-18 11:27:19 -08:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
} else {
|
|
|
|
|
// Sized result unpacking.
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
auto resultSegmentSpec = nb::cast<std::vector<int>>(resultSegmentSpecObj);
|
2021-01-18 11:27:19 -08:00
|
|
|
if (resultSegmentSpec.size() != resultTypeList.size()) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error((llvm::Twine("Operation \"") + name +
|
2021-01-18 11:27:19 -08:00
|
|
|
"\" requires " +
|
|
|
|
|
llvm::Twine(resultSegmentSpec.size()) +
|
2021-11-05 12:05:02 +01:00
|
|
|
" result segments but was provided " +
|
2021-01-18 11:27:19 -08:00
|
|
|
llvm::Twine(resultTypeList.size()))
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.str()
|
|
|
|
|
.c_str());
|
2021-01-18 11:27:19 -08:00
|
|
|
}
|
|
|
|
|
resultSegmentLengths.reserve(resultTypeList.size());
|
2022-01-02 22:02:14 +00:00
|
|
|
for (const auto &it :
|
2021-01-18 11:27:19 -08:00
|
|
|
llvm::enumerate(llvm::zip(resultTypeList, resultSegmentSpec))) {
|
|
|
|
|
int segmentSpec = std::get<1>(it.value());
|
|
|
|
|
if (segmentSpec == 1 || segmentSpec == 0) {
|
|
|
|
|
// Unpack unary element.
|
|
|
|
|
try {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
auto *resultType = nb::cast<PyType *>(std::get<0>(it.value()));
|
2021-01-18 11:27:19 -08:00
|
|
|
if (resultType) {
|
|
|
|
|
resultTypes.push_back(resultType);
|
|
|
|
|
resultSegmentLengths.push_back(1);
|
|
|
|
|
} else if (segmentSpec == 0) {
|
|
|
|
|
// Allowed to be optional.
|
|
|
|
|
resultSegmentLengths.push_back(0);
|
|
|
|
|
} else {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error(
|
|
|
|
|
(llvm::Twine("Result ") + llvm::Twine(it.index()) +
|
|
|
|
|
" of operation \"" + name +
|
|
|
|
|
"\" must be a Type (was None and result is not optional)")
|
|
|
|
|
.str()
|
|
|
|
|
.c_str());
|
2021-01-18 11:27:19 -08:00
|
|
|
}
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
} catch (nb::cast_error &err) {
|
|
|
|
|
throw nb::value_error((llvm::Twine("Result ") +
|
2021-01-18 11:27:19 -08:00
|
|
|
llvm::Twine(it.index()) + " of operation \"" +
|
|
|
|
|
name + "\" must be a Type (" + err.what() +
|
|
|
|
|
")")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.str()
|
|
|
|
|
.c_str());
|
2021-01-18 11:27:19 -08:00
|
|
|
}
|
|
|
|
|
} else if (segmentSpec == -1) {
|
|
|
|
|
// Unpack sequence by appending.
|
|
|
|
|
try {
|
|
|
|
|
if (std::get<0>(it.value()).is_none()) {
|
|
|
|
|
// Treat it as an empty list.
|
|
|
|
|
resultSegmentLengths.push_back(0);
|
|
|
|
|
} else {
|
|
|
|
|
// Unpack the list.
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
auto segment = nb::cast<nb::sequence>(std::get<0>(it.value()));
|
|
|
|
|
for (nb::handle segmentItem : segment) {
|
|
|
|
|
resultTypes.push_back(nb::cast<PyType *>(segmentItem));
|
2021-01-18 11:27:19 -08:00
|
|
|
if (!resultTypes.back()) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::type_error("contained a None item");
|
2021-01-18 11:27:19 -08:00
|
|
|
}
|
|
|
|
|
}
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
resultSegmentLengths.push_back(nb::len(segment));
|
2021-01-18 11:27:19 -08:00
|
|
|
}
|
|
|
|
|
} catch (std::exception &err) {
|
|
|
|
|
// NOTE: Sloppy to be using a catch-all here, but there are at least
|
|
|
|
|
// three different unrelated exceptions that can be thrown in the
|
|
|
|
|
// above "casts". Just keep the scope above small and catch them all.
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error((llvm::Twine("Result ") +
|
2021-01-18 11:27:19 -08:00
|
|
|
llvm::Twine(it.index()) + " of operation \"" +
|
|
|
|
|
name + "\" must be a Sequence of Types (" +
|
|
|
|
|
err.what() + ")")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.str()
|
|
|
|
|
.c_str());
|
2021-01-18 11:27:19 -08:00
|
|
|
}
|
|
|
|
|
} else {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error("Unexpected segment spec");
|
2021-01-18 11:27:19 -08:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
2023-07-23 21:26:52 -07:00
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object PyOpView::buildGeneric(
|
|
|
|
|
const nb::object &cls, std::optional<nb::list> resultTypeList,
|
|
|
|
|
nb::list operandList, std::optional<nb::dict> attributes,
|
2023-07-23 21:26:52 -07:00
|
|
|
std::optional<std::vector<PyBlock *>> successors,
|
|
|
|
|
std::optional<int> regions, DefaultingPyLocation location,
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
const nb::object &maybeIp) {
|
2023-07-23 21:26:52 -07:00
|
|
|
PyMlirContextRef context = location->getContext();
|
|
|
|
|
// Class level operation construction metadata.
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
std::string name = nb::cast<std::string>(cls.attr("OPERATION_NAME"));
|
2023-07-23 21:26:52 -07:00
|
|
|
// Operand and result segment specs are either none, which does no
|
|
|
|
|
// variadic unpacking, or a list of ints with segment sizes, where each
|
|
|
|
|
// element is either a positive number (typically 1 for a scalar) or -1 to
|
|
|
|
|
// indicate that it is derived from the length of the same-indexed operand
|
|
|
|
|
// or result (implying that it is a list at that position).
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object operandSegmentSpecObj = cls.attr("_ODS_OPERAND_SEGMENTS");
|
|
|
|
|
nb::object resultSegmentSpecObj = cls.attr("_ODS_RESULT_SEGMENTS");
|
2023-07-23 21:26:52 -07:00
|
|
|
|
|
|
|
|
std::vector<int32_t> operandSegmentLengths;
|
|
|
|
|
std::vector<int32_t> resultSegmentLengths;
|
|
|
|
|
|
|
|
|
|
// Validate/determine region count.
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
auto opRegionSpec = nb::cast<std::tuple<int, bool>>(cls.attr("_ODS_REGIONS"));
|
2023-07-23 21:26:52 -07:00
|
|
|
int opMinRegionCount = std::get<0>(opRegionSpec);
|
|
|
|
|
bool opHasNoVariadicRegions = std::get<1>(opRegionSpec);
|
|
|
|
|
if (!regions) {
|
|
|
|
|
regions = opMinRegionCount;
|
|
|
|
|
}
|
|
|
|
|
if (*regions < opMinRegionCount) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error(
|
2023-07-23 21:26:52 -07:00
|
|
|
(llvm::Twine("Operation \"") + name + "\" requires a minimum of " +
|
|
|
|
|
llvm::Twine(opMinRegionCount) +
|
|
|
|
|
" regions but was built with regions=" + llvm::Twine(*regions))
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.str()
|
|
|
|
|
.c_str());
|
2023-07-23 21:26:52 -07:00
|
|
|
}
|
|
|
|
|
if (opHasNoVariadicRegions && *regions > opMinRegionCount) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error(
|
2023-07-23 21:26:52 -07:00
|
|
|
(llvm::Twine("Operation \"") + name + "\" requires a maximum of " +
|
|
|
|
|
llvm::Twine(opMinRegionCount) +
|
|
|
|
|
" regions but was built with regions=" + llvm::Twine(*regions))
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.str()
|
|
|
|
|
.c_str());
|
2023-07-23 21:26:52 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Unpack results.
|
|
|
|
|
std::vector<PyType *> resultTypes;
|
|
|
|
|
if (resultTypeList.has_value()) {
|
|
|
|
|
populateResultTypes(name, *resultTypeList, resultSegmentSpecObj,
|
|
|
|
|
resultSegmentLengths, resultTypes);
|
|
|
|
|
}
|
2021-01-18 11:27:19 -08:00
|
|
|
|
|
|
|
|
// Unpack operands.
|
|
|
|
|
std::vector<PyValue *> operands;
|
|
|
|
|
operands.reserve(operands.size());
|
|
|
|
|
if (operandSegmentSpecObj.is_none()) {
|
|
|
|
|
// Non-sized operand unpacking.
|
2022-01-02 22:02:14 +00:00
|
|
|
for (const auto &it : llvm::enumerate(operandList)) {
|
2021-01-18 11:27:19 -08:00
|
|
|
try {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
operands.push_back(nb::cast<PyValue *>(it.value()));
|
2021-01-18 11:27:19 -08:00
|
|
|
if (!operands.back())
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::cast_error();
|
|
|
|
|
} catch (nb::cast_error &err) {
|
|
|
|
|
throw nb::value_error((llvm::Twine("Operand ") +
|
2021-01-18 11:27:19 -08:00
|
|
|
llvm::Twine(it.index()) + " of operation \"" +
|
|
|
|
|
name + "\" must be a Value (" + err.what() + ")")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.str()
|
|
|
|
|
.c_str());
|
2021-01-18 11:27:19 -08:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
} else {
|
|
|
|
|
// Sized operand unpacking.
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
auto operandSegmentSpec = nb::cast<std::vector<int>>(operandSegmentSpecObj);
|
2021-01-18 11:27:19 -08:00
|
|
|
if (operandSegmentSpec.size() != operandList.size()) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error((llvm::Twine("Operation \"") + name +
|
2021-01-18 11:27:19 -08:00
|
|
|
"\" requires " +
|
|
|
|
|
llvm::Twine(operandSegmentSpec.size()) +
|
|
|
|
|
"operand segments but was provided " +
|
|
|
|
|
llvm::Twine(operandList.size()))
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.str()
|
|
|
|
|
.c_str());
|
2021-01-18 11:27:19 -08:00
|
|
|
}
|
|
|
|
|
operandSegmentLengths.reserve(operandList.size());
|
2022-01-02 22:02:14 +00:00
|
|
|
for (const auto &it :
|
2021-01-18 11:27:19 -08:00
|
|
|
llvm::enumerate(llvm::zip(operandList, operandSegmentSpec))) {
|
|
|
|
|
int segmentSpec = std::get<1>(it.value());
|
|
|
|
|
if (segmentSpec == 1 || segmentSpec == 0) {
|
|
|
|
|
// Unpack unary element.
|
|
|
|
|
try {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
auto *operandValue = nb::cast<PyValue *>(std::get<0>(it.value()));
|
2021-01-18 11:27:19 -08:00
|
|
|
if (operandValue) {
|
|
|
|
|
operands.push_back(operandValue);
|
|
|
|
|
operandSegmentLengths.push_back(1);
|
|
|
|
|
} else if (segmentSpec == 0) {
|
|
|
|
|
// Allowed to be optional.
|
|
|
|
|
operandSegmentLengths.push_back(0);
|
|
|
|
|
} else {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error(
|
|
|
|
|
(llvm::Twine("Operand ") + llvm::Twine(it.index()) +
|
|
|
|
|
" of operation \"" + name +
|
|
|
|
|
"\" must be a Value (was None and operand is not optional)")
|
|
|
|
|
.str()
|
|
|
|
|
.c_str());
|
2021-01-18 11:27:19 -08:00
|
|
|
}
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
} catch (nb::cast_error &err) {
|
|
|
|
|
throw nb::value_error((llvm::Twine("Operand ") +
|
2021-01-18 11:27:19 -08:00
|
|
|
llvm::Twine(it.index()) + " of operation \"" +
|
|
|
|
|
name + "\" must be a Value (" + err.what() +
|
|
|
|
|
")")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.str()
|
|
|
|
|
.c_str());
|
2021-01-18 11:27:19 -08:00
|
|
|
}
|
|
|
|
|
} else if (segmentSpec == -1) {
|
|
|
|
|
// Unpack sequence by appending.
|
|
|
|
|
try {
|
|
|
|
|
if (std::get<0>(it.value()).is_none()) {
|
|
|
|
|
// Treat it as an empty list.
|
|
|
|
|
operandSegmentLengths.push_back(0);
|
|
|
|
|
} else {
|
|
|
|
|
// Unpack the list.
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
auto segment = nb::cast<nb::sequence>(std::get<0>(it.value()));
|
|
|
|
|
for (nb::handle segmentItem : segment) {
|
|
|
|
|
operands.push_back(nb::cast<PyValue *>(segmentItem));
|
2021-01-18 11:27:19 -08:00
|
|
|
if (!operands.back()) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::type_error("contained a None item");
|
2021-01-18 11:27:19 -08:00
|
|
|
}
|
|
|
|
|
}
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
operandSegmentLengths.push_back(nb::len(segment));
|
2021-01-18 11:27:19 -08:00
|
|
|
}
|
|
|
|
|
} catch (std::exception &err) {
|
|
|
|
|
// NOTE: Sloppy to be using a catch-all here, but there are at least
|
|
|
|
|
// three different unrelated exceptions that can be thrown in the
|
|
|
|
|
// above "casts". Just keep the scope above small and catch them all.
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error((llvm::Twine("Operand ") +
|
2021-01-18 11:27:19 -08:00
|
|
|
llvm::Twine(it.index()) + " of operation \"" +
|
|
|
|
|
name + "\" must be a Sequence of Values (" +
|
|
|
|
|
err.what() + ")")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.str()
|
|
|
|
|
.c_str());
|
2021-01-18 11:27:19 -08:00
|
|
|
}
|
|
|
|
|
} else {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error("Unexpected segment spec");
|
2021-01-18 11:27:19 -08:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Merge operand/result segment lengths into attributes if needed.
|
|
|
|
|
if (!operandSegmentLengths.empty() || !resultSegmentLengths.empty()) {
|
|
|
|
|
// Dup.
|
|
|
|
|
if (attributes) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
attributes = nb::dict(*attributes);
|
2021-01-18 11:27:19 -08:00
|
|
|
} else {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
attributes = nb::dict();
|
2021-01-18 11:27:19 -08:00
|
|
|
}
|
2023-08-09 12:47:13 -07:00
|
|
|
if (attributes->contains("resultSegmentSizes") ||
|
|
|
|
|
attributes->contains("operandSegmentSizes")) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error("Manually setting a 'resultSegmentSizes' or "
|
2023-08-09 12:47:13 -07:00
|
|
|
"'operandSegmentSizes' attribute is unsupported. "
|
2021-01-18 11:27:19 -08:00
|
|
|
"Use Operation.create for such low-level access.");
|
|
|
|
|
}
|
|
|
|
|
|
2023-08-09 12:47:13 -07:00
|
|
|
// Add resultSegmentSizes attribute.
|
2021-01-18 11:27:19 -08:00
|
|
|
if (!resultSegmentLengths.empty()) {
|
2022-08-12 15:43:03 -04:00
|
|
|
MlirAttribute segmentLengthAttr =
|
|
|
|
|
mlirDenseI32ArrayGet(context->get(), resultSegmentLengths.size(),
|
|
|
|
|
resultSegmentLengths.data());
|
2023-08-09 12:47:13 -07:00
|
|
|
(*attributes)["resultSegmentSizes"] =
|
2021-01-18 11:27:19 -08:00
|
|
|
PyAttribute(context, segmentLengthAttr);
|
|
|
|
|
}
|
|
|
|
|
|
2023-08-09 12:47:13 -07:00
|
|
|
// Add operandSegmentSizes attribute.
|
2021-01-18 11:27:19 -08:00
|
|
|
if (!operandSegmentLengths.empty()) {
|
2022-08-12 15:43:03 -04:00
|
|
|
MlirAttribute segmentLengthAttr =
|
|
|
|
|
mlirDenseI32ArrayGet(context->get(), operandSegmentLengths.size(),
|
|
|
|
|
operandSegmentLengths.data());
|
2023-08-09 12:47:13 -07:00
|
|
|
(*attributes)["operandSegmentSizes"] =
|
2021-01-18 11:27:19 -08:00
|
|
|
PyAttribute(context, segmentLengthAttr);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Delegate to create.
|
2022-01-02 22:02:18 +00:00
|
|
|
return PyOperation::create(name,
|
2021-01-18 11:27:19 -08:00
|
|
|
/*results=*/std::move(resultTypes),
|
2021-01-24 14:46:56 -08:00
|
|
|
/*operands=*/std::move(operands),
|
2021-01-18 11:27:19 -08:00
|
|
|
/*attributes=*/std::move(attributes),
|
|
|
|
|
/*successors=*/std::move(successors),
|
2023-07-23 21:26:52 -07:00
|
|
|
/*regions=*/*regions, location, maybeIp,
|
|
|
|
|
!resultTypeList);
|
2021-01-18 11:27:19 -08:00
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object PyOpView::constructDerived(const nb::object &cls,
|
|
|
|
|
const nb::object &operation) {
|
|
|
|
|
nb::handle opViewType = nb::type<PyOpView>();
|
|
|
|
|
nb::object instance = cls.attr("__new__")(cls);
|
2023-01-22 23:31:18 -05:00
|
|
|
opViewType.attr("__init__")(instance, operation);
|
|
|
|
|
return instance;
|
|
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
PyOpView::PyOpView(const nb::object &operationObject)
|
2020-11-01 23:05:36 -08:00
|
|
|
// Casting through the PyOperationBase base-class and then back to the
|
|
|
|
|
// Operation lets us accept any PyOperationBase subclass.
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
: operation(nb::cast<PyOperationBase &>(operationObject).getOperation()),
|
2020-11-01 23:05:36 -08:00
|
|
|
operationObject(operation.getRef().getObject()) {}
|
2020-10-21 23:34:01 -07:00
|
|
|
|
2020-10-28 23:16:36 -07:00
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
// PyInsertionPoint.
|
|
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
|
|
|
|
|
PyInsertionPoint::PyInsertionPoint(PyBlock &block) : block(block) {}
|
|
|
|
|
|
2020-11-01 23:05:36 -08:00
|
|
|
PyInsertionPoint::PyInsertionPoint(PyOperationBase &beforeOperationBase)
|
|
|
|
|
: refOperation(beforeOperationBase.getOperation().getRef()),
|
|
|
|
|
block((*refOperation)->getBlock()) {}
|
2020-10-28 23:16:36 -07:00
|
|
|
|
2020-11-01 23:05:36 -08:00
|
|
|
void PyInsertionPoint::insert(PyOperationBase &operationBase) {
|
|
|
|
|
PyOperation &operation = operationBase.getOperation();
|
2020-10-28 23:16:36 -07:00
|
|
|
if (operation.isAttached())
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error(
|
2023-05-22 17:30:12 -05:00
|
|
|
"Attempt to insert operation that is already attached");
|
2020-10-28 23:16:36 -07:00
|
|
|
block.getParentOperation()->checkValid();
|
|
|
|
|
MlirOperation beforeOp = {nullptr};
|
|
|
|
|
if (refOperation) {
|
|
|
|
|
// Insert before operation.
|
|
|
|
|
(*refOperation)->checkValid();
|
|
|
|
|
beforeOp = (*refOperation)->get();
|
2021-01-24 14:46:56 -08:00
|
|
|
} else {
|
|
|
|
|
// Insert at end (before null) is only valid if the block does not
|
|
|
|
|
// already end in a known terminator (violating this will cause assertion
|
|
|
|
|
// failures later).
|
|
|
|
|
if (!mlirOperationIsNull(mlirBlockGetTerminator(block.get()))) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::index_error("Cannot insert operation at the end of a block "
|
2021-01-24 14:46:56 -08:00
|
|
|
"that already has a terminator. Did you mean to "
|
|
|
|
|
"use 'InsertionPoint.at_block_terminator(block)' "
|
|
|
|
|
"versus 'InsertionPoint(block)'?");
|
|
|
|
|
}
|
2020-10-28 23:16:36 -07:00
|
|
|
}
|
2020-11-29 13:52:11 -08:00
|
|
|
mlirBlockInsertOwnedOperationBefore(block.get(), beforeOp, operation);
|
2020-10-28 23:16:36 -07:00
|
|
|
operation.setAttached();
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
PyInsertionPoint PyInsertionPoint::atBlockBegin(PyBlock &block) {
|
|
|
|
|
MlirOperation firstOp = mlirBlockGetFirstOperation(block.get());
|
|
|
|
|
if (mlirOperationIsNull(firstOp)) {
|
|
|
|
|
// Just insert at end.
|
|
|
|
|
return PyInsertionPoint(block);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Insert before first op.
|
|
|
|
|
PyOperationRef firstOpRef = PyOperation::forOperation(
|
|
|
|
|
block.getParentOperation()->getContext(), firstOp);
|
|
|
|
|
return PyInsertionPoint{block, std::move(firstOpRef)};
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
PyInsertionPoint PyInsertionPoint::atBlockTerminator(PyBlock &block) {
|
|
|
|
|
MlirOperation terminator = mlirBlockGetTerminator(block.get());
|
|
|
|
|
if (mlirOperationIsNull(terminator))
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error("Block has no terminator");
|
2020-10-28 23:16:36 -07:00
|
|
|
PyOperationRef terminatorOpRef = PyOperation::forOperation(
|
|
|
|
|
block.getParentOperation()->getContext(), terminator);
|
|
|
|
|
return PyInsertionPoint{block, std::move(terminatorOpRef)};
|
|
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object PyInsertionPoint::contextEnter(nb::object insertPoint) {
|
|
|
|
|
return PyThreadContextEntry::pushInsertionPoint(insertPoint);
|
2020-10-28 23:16:36 -07:00
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
void PyInsertionPoint::contextExit(const nb::object &excType,
|
|
|
|
|
const nb::object &excVal,
|
|
|
|
|
const nb::object &excTb) {
|
2020-10-31 23:40:25 -07:00
|
|
|
PyThreadContextEntry::popInsertionPoint(*this);
|
2020-10-28 23:16:36 -07:00
|
|
|
}
|
|
|
|
|
|
2020-08-19 15:33:02 -07:00
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
// PyAttribute.
|
|
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
|
2023-03-28 11:05:00 -04:00
|
|
|
bool PyAttribute::operator==(const PyAttribute &other) const {
|
2020-08-19 15:33:02 -07:00
|
|
|
return mlirAttributeEqual(attr, other.attr);
|
|
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object PyAttribute::getCapsule() {
|
|
|
|
|
return nb::steal<nb::object>(mlirPythonAttributeToCapsule(*this));
|
2020-11-29 13:30:23 -08:00
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
PyAttribute PyAttribute::createFromCapsule(nb::object capsule) {
|
2020-11-29 13:30:23 -08:00
|
|
|
MlirAttribute rawAttr = mlirPythonCapsuleToAttribute(capsule.ptr());
|
|
|
|
|
if (mlirAttributeIsNull(rawAttr))
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::python_error();
|
2020-11-29 13:30:23 -08:00
|
|
|
return PyAttribute(
|
|
|
|
|
PyMlirContext::forContext(mlirAttributeGetContext(rawAttr)), rawAttr);
|
|
|
|
|
}
|
|
|
|
|
|
2020-08-19 15:33:02 -07:00
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
// PyNamedAttribute.
|
|
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
|
|
|
|
|
PyNamedAttribute::PyNamedAttribute(MlirAttribute attr, std::string ownedName)
|
|
|
|
|
: ownedName(new std::string(std::move(ownedName))) {
|
2020-12-11 18:50:04 +00:00
|
|
|
namedAttr = mlirNamedAttributeGet(
|
|
|
|
|
mlirIdentifierGet(mlirAttributeGetContext(attr),
|
|
|
|
|
toMlirStringRef(*this->ownedName)),
|
|
|
|
|
attr);
|
2020-08-19 15:33:02 -07:00
|
|
|
}
|
|
|
|
|
|
2020-08-16 18:49:28 -07:00
|
|
|
//------------------------------------------------------------------------------
|
2020-08-18 17:23:46 -07:00
|
|
|
// PyType.
|
2020-08-16 18:49:28 -07:00
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
|
2023-03-28 11:05:00 -04:00
|
|
|
bool PyType::operator==(const PyType &other) const {
|
2020-08-18 17:23:46 -07:00
|
|
|
return mlirTypeEqual(type, other.type);
|
2020-08-16 18:49:28 -07:00
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object PyType::getCapsule() {
|
|
|
|
|
return nb::steal<nb::object>(mlirPythonTypeToCapsule(*this));
|
2020-11-29 13:30:23 -08:00
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
PyType PyType::createFromCapsule(nb::object capsule) {
|
2020-11-29 13:30:23 -08:00
|
|
|
MlirType rawType = mlirPythonCapsuleToType(capsule.ptr());
|
|
|
|
|
if (mlirTypeIsNull(rawType))
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::python_error();
|
2020-11-29 13:30:23 -08:00
|
|
|
return PyType(PyMlirContext::forContext(mlirTypeGetContext(rawType)),
|
|
|
|
|
rawType);
|
|
|
|
|
}
|
|
|
|
|
|
2023-05-22 11:12:53 -05:00
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
// PyTypeID.
|
|
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object PyTypeID::getCapsule() {
|
|
|
|
|
return nb::steal<nb::object>(mlirPythonTypeIDToCapsule(*this));
|
2023-05-22 11:12:53 -05:00
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
PyTypeID PyTypeID::createFromCapsule(nb::object capsule) {
|
2023-05-22 11:12:53 -05:00
|
|
|
MlirTypeID mlirTypeID = mlirPythonCapsuleToTypeID(capsule.ptr());
|
|
|
|
|
if (mlirTypeIDIsNull(mlirTypeID))
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::python_error();
|
2023-05-22 11:12:53 -05:00
|
|
|
return PyTypeID(mlirTypeID);
|
|
|
|
|
}
|
|
|
|
|
bool PyTypeID::operator==(const PyTypeID &other) const {
|
|
|
|
|
return mlirTypeIDEqual(typeID, other.typeID);
|
|
|
|
|
}
|
|
|
|
|
|
2020-10-20 11:21:05 +02:00
|
|
|
//------------------------------------------------------------------------------
|
[mlir][python] value casting (#69644)
This PR adds "value casting", i.e., a mechanism to wrap `ir.Value` in a
proxy class that overloads dunders such as `__add__`, `__sub__`, and
`__mul__` for fun and great profit.
This is thematically similar to
https://github.com/llvm/llvm-project/commit/bfb1ba752655bf09b35c486f6cc9817dbedfb1bb
and
https://github.com/llvm/llvm-project/commit/9566ee280607d91fa2e5eca730a6765ac84dfd0f.
The example in the test demonstrates the value of the feature (no pun
intended):
```python
@register_value_caster(F16Type.static_typeid)
@register_value_caster(F32Type.static_typeid)
@register_value_caster(F64Type.static_typeid)
@register_value_caster(IntegerType.static_typeid)
class ArithValue(Value):
__add__ = partialmethod(_binary_op, op="add")
__sub__ = partialmethod(_binary_op, op="sub")
__mul__ = partialmethod(_binary_op, op="mul")
a = arith.constant(value=FloatAttr.get(f16_t, 42.42))
b = a + a
# CHECK: ArithValue(%0 = arith.addf %cst, %cst : f16)
print(b)
a = arith.constant(value=FloatAttr.get(f32_t, 42.42))
b = a - a
# CHECK: ArithValue(%1 = arith.subf %cst_0, %cst_0 : f32)
print(b)
a = arith.constant(value=FloatAttr.get(f64_t, 42.42))
b = a * a
# CHECK: ArithValue(%2 = arith.mulf %cst_1, %cst_1 : f64)
print(b)
```
**EDIT**: this now goes through the bindings and thus supports automatic
casting of `OpResult` (including as an element of `OpResultList`),
`BlockArgument` (including as an element of `BlockArgumentList`), as
well as `Value`.
2023-11-07 10:49:41 -06:00
|
|
|
// PyValue and subclasses.
|
2020-10-20 11:21:05 +02:00
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object PyValue::getCapsule() {
|
|
|
|
|
return nb::steal<nb::object>(mlirPythonValueToCapsule(get()));
|
2021-04-22 00:07:30 -06:00
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object PyValue::maybeDownCast() {
|
[mlir][python] value casting (#69644)
This PR adds "value casting", i.e., a mechanism to wrap `ir.Value` in a
proxy class that overloads dunders such as `__add__`, `__sub__`, and
`__mul__` for fun and great profit.
This is thematically similar to
https://github.com/llvm/llvm-project/commit/bfb1ba752655bf09b35c486f6cc9817dbedfb1bb
and
https://github.com/llvm/llvm-project/commit/9566ee280607d91fa2e5eca730a6765ac84dfd0f.
The example in the test demonstrates the value of the feature (no pun
intended):
```python
@register_value_caster(F16Type.static_typeid)
@register_value_caster(F32Type.static_typeid)
@register_value_caster(F64Type.static_typeid)
@register_value_caster(IntegerType.static_typeid)
class ArithValue(Value):
__add__ = partialmethod(_binary_op, op="add")
__sub__ = partialmethod(_binary_op, op="sub")
__mul__ = partialmethod(_binary_op, op="mul")
a = arith.constant(value=FloatAttr.get(f16_t, 42.42))
b = a + a
# CHECK: ArithValue(%0 = arith.addf %cst, %cst : f16)
print(b)
a = arith.constant(value=FloatAttr.get(f32_t, 42.42))
b = a - a
# CHECK: ArithValue(%1 = arith.subf %cst_0, %cst_0 : f32)
print(b)
a = arith.constant(value=FloatAttr.get(f64_t, 42.42))
b = a * a
# CHECK: ArithValue(%2 = arith.mulf %cst_1, %cst_1 : f64)
print(b)
```
**EDIT**: this now goes through the bindings and thus supports automatic
casting of `OpResult` (including as an element of `OpResultList`),
`BlockArgument` (including as an element of `BlockArgumentList`), as
well as `Value`.
2023-11-07 10:49:41 -06:00
|
|
|
MlirType type = mlirValueGetType(get());
|
|
|
|
|
MlirTypeID mlirTypeID = mlirTypeGetTypeID(type);
|
|
|
|
|
assert(!mlirTypeIDIsNull(mlirTypeID) &&
|
|
|
|
|
"mlirTypeID was expected to be non-null.");
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
std::optional<nb::callable> valueCaster =
|
[mlir][python] value casting (#69644)
This PR adds "value casting", i.e., a mechanism to wrap `ir.Value` in a
proxy class that overloads dunders such as `__add__`, `__sub__`, and
`__mul__` for fun and great profit.
This is thematically similar to
https://github.com/llvm/llvm-project/commit/bfb1ba752655bf09b35c486f6cc9817dbedfb1bb
and
https://github.com/llvm/llvm-project/commit/9566ee280607d91fa2e5eca730a6765ac84dfd0f.
The example in the test demonstrates the value of the feature (no pun
intended):
```python
@register_value_caster(F16Type.static_typeid)
@register_value_caster(F32Type.static_typeid)
@register_value_caster(F64Type.static_typeid)
@register_value_caster(IntegerType.static_typeid)
class ArithValue(Value):
__add__ = partialmethod(_binary_op, op="add")
__sub__ = partialmethod(_binary_op, op="sub")
__mul__ = partialmethod(_binary_op, op="mul")
a = arith.constant(value=FloatAttr.get(f16_t, 42.42))
b = a + a
# CHECK: ArithValue(%0 = arith.addf %cst, %cst : f16)
print(b)
a = arith.constant(value=FloatAttr.get(f32_t, 42.42))
b = a - a
# CHECK: ArithValue(%1 = arith.subf %cst_0, %cst_0 : f32)
print(b)
a = arith.constant(value=FloatAttr.get(f64_t, 42.42))
b = a * a
# CHECK: ArithValue(%2 = arith.mulf %cst_1, %cst_1 : f64)
print(b)
```
**EDIT**: this now goes through the bindings and thus supports automatic
casting of `OpResult` (including as an element of `OpResultList`),
`BlockArgument` (including as an element of `BlockArgumentList`), as
well as `Value`.
2023-11-07 10:49:41 -06:00
|
|
|
PyGlobals::get().lookupValueCaster(mlirTypeID, mlirTypeGetDialect(type));
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
// nb::rv_policy::move means use std::move to move the return value
|
[mlir][python] value casting (#69644)
This PR adds "value casting", i.e., a mechanism to wrap `ir.Value` in a
proxy class that overloads dunders such as `__add__`, `__sub__`, and
`__mul__` for fun and great profit.
This is thematically similar to
https://github.com/llvm/llvm-project/commit/bfb1ba752655bf09b35c486f6cc9817dbedfb1bb
and
https://github.com/llvm/llvm-project/commit/9566ee280607d91fa2e5eca730a6765ac84dfd0f.
The example in the test demonstrates the value of the feature (no pun
intended):
```python
@register_value_caster(F16Type.static_typeid)
@register_value_caster(F32Type.static_typeid)
@register_value_caster(F64Type.static_typeid)
@register_value_caster(IntegerType.static_typeid)
class ArithValue(Value):
__add__ = partialmethod(_binary_op, op="add")
__sub__ = partialmethod(_binary_op, op="sub")
__mul__ = partialmethod(_binary_op, op="mul")
a = arith.constant(value=FloatAttr.get(f16_t, 42.42))
b = a + a
# CHECK: ArithValue(%0 = arith.addf %cst, %cst : f16)
print(b)
a = arith.constant(value=FloatAttr.get(f32_t, 42.42))
b = a - a
# CHECK: ArithValue(%1 = arith.subf %cst_0, %cst_0 : f32)
print(b)
a = arith.constant(value=FloatAttr.get(f64_t, 42.42))
b = a * a
# CHECK: ArithValue(%2 = arith.mulf %cst_1, %cst_1 : f64)
print(b)
```
**EDIT**: this now goes through the bindings and thus supports automatic
casting of `OpResult` (including as an element of `OpResultList`),
`BlockArgument` (including as an element of `BlockArgumentList`), as
well as `Value`.
2023-11-07 10:49:41 -06:00
|
|
|
// contents into a new instance that will be owned by Python.
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object thisObj = nb::cast(this, nb::rv_policy::move);
|
[mlir][python] value casting (#69644)
This PR adds "value casting", i.e., a mechanism to wrap `ir.Value` in a
proxy class that overloads dunders such as `__add__`, `__sub__`, and
`__mul__` for fun and great profit.
This is thematically similar to
https://github.com/llvm/llvm-project/commit/bfb1ba752655bf09b35c486f6cc9817dbedfb1bb
and
https://github.com/llvm/llvm-project/commit/9566ee280607d91fa2e5eca730a6765ac84dfd0f.
The example in the test demonstrates the value of the feature (no pun
intended):
```python
@register_value_caster(F16Type.static_typeid)
@register_value_caster(F32Type.static_typeid)
@register_value_caster(F64Type.static_typeid)
@register_value_caster(IntegerType.static_typeid)
class ArithValue(Value):
__add__ = partialmethod(_binary_op, op="add")
__sub__ = partialmethod(_binary_op, op="sub")
__mul__ = partialmethod(_binary_op, op="mul")
a = arith.constant(value=FloatAttr.get(f16_t, 42.42))
b = a + a
# CHECK: ArithValue(%0 = arith.addf %cst, %cst : f16)
print(b)
a = arith.constant(value=FloatAttr.get(f32_t, 42.42))
b = a - a
# CHECK: ArithValue(%1 = arith.subf %cst_0, %cst_0 : f32)
print(b)
a = arith.constant(value=FloatAttr.get(f64_t, 42.42))
b = a * a
# CHECK: ArithValue(%2 = arith.mulf %cst_1, %cst_1 : f64)
print(b)
```
**EDIT**: this now goes through the bindings and thus supports automatic
casting of `OpResult` (including as an element of `OpResultList`),
`BlockArgument` (including as an element of `BlockArgumentList`), as
well as `Value`.
2023-11-07 10:49:41 -06:00
|
|
|
if (!valueCaster)
|
|
|
|
|
return thisObj;
|
|
|
|
|
return valueCaster.value()(thisObj);
|
|
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
PyValue PyValue::createFromCapsule(nb::object capsule) {
|
2021-04-22 00:07:30 -06:00
|
|
|
MlirValue value = mlirPythonCapsuleToValue(capsule.ptr());
|
|
|
|
|
if (mlirValueIsNull(value))
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::python_error();
|
2021-04-22 00:07:30 -06:00
|
|
|
MlirOperation owner;
|
|
|
|
|
if (mlirValueIsAOpResult(value))
|
|
|
|
|
owner = mlirOpResultGetOwner(value);
|
|
|
|
|
if (mlirValueIsABlockArgument(value))
|
|
|
|
|
owner = mlirBlockGetParentOperation(mlirBlockArgumentGetOwner(value));
|
|
|
|
|
if (mlirOperationIsNull(owner))
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::python_error();
|
2021-04-22 00:07:30 -06:00
|
|
|
MlirContext ctx = mlirOperationGetContext(owner);
|
|
|
|
|
PyOperationRef ownerRef =
|
|
|
|
|
PyOperation::forOperation(PyMlirContext::forContext(ctx), owner);
|
|
|
|
|
return PyValue(ownerRef, value);
|
|
|
|
|
}
|
|
|
|
|
|
2021-11-02 12:39:36 +01:00
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
// PySymbolTable.
|
|
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
|
|
|
|
|
PySymbolTable::PySymbolTable(PyOperationBase &operation)
|
|
|
|
|
: operation(operation.getOperation().getRef()) {
|
|
|
|
|
symbolTable = mlirSymbolTableCreate(operation.getOperation().get());
|
|
|
|
|
if (mlirSymbolTableIsNull(symbolTable)) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::type_error("Operation is not a Symbol Table.");
|
2021-11-02 12:39:36 +01:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object PySymbolTable::dunderGetItem(const std::string &name) {
|
2021-11-02 12:39:36 +01:00
|
|
|
operation->checkValid();
|
|
|
|
|
MlirOperation symbol = mlirSymbolTableLookup(
|
|
|
|
|
symbolTable, mlirStringRefCreate(name.data(), name.length()));
|
|
|
|
|
if (mlirOperationIsNull(symbol))
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::key_error(
|
|
|
|
|
("Symbol '" + name + "' not in the symbol table.").c_str());
|
2021-11-02 12:39:36 +01:00
|
|
|
|
|
|
|
|
return PyOperation::forOperation(operation->getContext(), symbol,
|
|
|
|
|
operation.getObject())
|
|
|
|
|
->createOpView();
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void PySymbolTable::erase(PyOperationBase &symbol) {
|
|
|
|
|
operation->checkValid();
|
|
|
|
|
symbol.getOperation().checkValid();
|
|
|
|
|
mlirSymbolTableErase(symbolTable, symbol.getOperation().get());
|
|
|
|
|
// The operation is also erased, so we must invalidate it. There may be Python
|
|
|
|
|
// references to this operation so we don't want to delete it from the list of
|
|
|
|
|
// live operations here.
|
|
|
|
|
symbol.getOperation().valid = false;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void PySymbolTable::dunderDel(const std::string &name) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object operation = dunderGetItem(name);
|
|
|
|
|
erase(nb::cast<PyOperationBase &>(operation));
|
2021-11-02 12:39:36 +01:00
|
|
|
}
|
|
|
|
|
|
2023-07-04 16:06:17 -04:00
|
|
|
MlirAttribute PySymbolTable::insert(PyOperationBase &symbol) {
|
2021-11-02 12:39:36 +01:00
|
|
|
operation->checkValid();
|
|
|
|
|
symbol.getOperation().checkValid();
|
|
|
|
|
MlirAttribute symbolAttr = mlirOperationGetAttributeByName(
|
|
|
|
|
symbol.getOperation().get(), mlirSymbolTableGetSymbolAttributeName());
|
|
|
|
|
if (mlirAttributeIsNull(symbolAttr))
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error("Expected operation to have a symbol name.");
|
2023-07-04 16:06:17 -04:00
|
|
|
return mlirSymbolTableInsert(symbolTable, symbol.getOperation().get());
|
2021-11-02 12:39:36 +01:00
|
|
|
}
|
|
|
|
|
|
2023-07-04 16:06:17 -04:00
|
|
|
MlirAttribute PySymbolTable::getSymbolName(PyOperationBase &symbol) {
|
2021-11-28 20:30:18 -08:00
|
|
|
// Op must already be a symbol.
|
|
|
|
|
PyOperation &operation = symbol.getOperation();
|
|
|
|
|
operation.checkValid();
|
|
|
|
|
MlirStringRef attrName = mlirSymbolTableGetSymbolAttributeName();
|
|
|
|
|
MlirAttribute existingNameAttr =
|
|
|
|
|
mlirOperationGetAttributeByName(operation.get(), attrName);
|
|
|
|
|
if (mlirAttributeIsNull(existingNameAttr))
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error("Expected operation to have a symbol name.");
|
2023-07-04 16:06:17 -04:00
|
|
|
return existingNameAttr;
|
2021-11-28 20:30:18 -08:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void PySymbolTable::setSymbolName(PyOperationBase &symbol,
|
|
|
|
|
const std::string &name) {
|
|
|
|
|
// Op must already be a symbol.
|
|
|
|
|
PyOperation &operation = symbol.getOperation();
|
|
|
|
|
operation.checkValid();
|
|
|
|
|
MlirStringRef attrName = mlirSymbolTableGetSymbolAttributeName();
|
|
|
|
|
MlirAttribute existingNameAttr =
|
|
|
|
|
mlirOperationGetAttributeByName(operation.get(), attrName);
|
|
|
|
|
if (mlirAttributeIsNull(existingNameAttr))
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error("Expected operation to have a symbol name.");
|
2021-11-28 20:30:18 -08:00
|
|
|
MlirAttribute newNameAttr =
|
|
|
|
|
mlirStringAttrGet(operation.getContext()->get(), toMlirStringRef(name));
|
|
|
|
|
mlirOperationSetAttributeByName(operation.get(), attrName, newNameAttr);
|
|
|
|
|
}
|
|
|
|
|
|
2023-07-04 16:06:17 -04:00
|
|
|
MlirAttribute PySymbolTable::getVisibility(PyOperationBase &symbol) {
|
2021-11-28 20:30:18 -08:00
|
|
|
PyOperation &operation = symbol.getOperation();
|
|
|
|
|
operation.checkValid();
|
|
|
|
|
MlirStringRef attrName = mlirSymbolTableGetVisibilityAttributeName();
|
|
|
|
|
MlirAttribute existingVisAttr =
|
|
|
|
|
mlirOperationGetAttributeByName(operation.get(), attrName);
|
|
|
|
|
if (mlirAttributeIsNull(existingVisAttr))
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error("Expected operation to have a symbol visibility.");
|
2023-07-04 16:06:17 -04:00
|
|
|
return existingVisAttr;
|
2021-11-28 20:30:18 -08:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void PySymbolTable::setVisibility(PyOperationBase &symbol,
|
|
|
|
|
const std::string &visibility) {
|
|
|
|
|
if (visibility != "public" && visibility != "private" &&
|
|
|
|
|
visibility != "nested")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error(
|
2021-11-28 20:30:18 -08:00
|
|
|
"Expected visibility to be 'public', 'private' or 'nested'");
|
|
|
|
|
PyOperation &operation = symbol.getOperation();
|
|
|
|
|
operation.checkValid();
|
|
|
|
|
MlirStringRef attrName = mlirSymbolTableGetVisibilityAttributeName();
|
|
|
|
|
MlirAttribute existingVisAttr =
|
|
|
|
|
mlirOperationGetAttributeByName(operation.get(), attrName);
|
|
|
|
|
if (mlirAttributeIsNull(existingVisAttr))
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error("Expected operation to have a symbol visibility.");
|
2021-11-28 20:30:18 -08:00
|
|
|
MlirAttribute newVisAttr = mlirStringAttrGet(operation.getContext()->get(),
|
|
|
|
|
toMlirStringRef(visibility));
|
|
|
|
|
mlirOperationSetAttributeByName(operation.get(), attrName, newVisAttr);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void PySymbolTable::replaceAllSymbolUses(const std::string &oldSymbol,
|
|
|
|
|
const std::string &newSymbol,
|
|
|
|
|
PyOperationBase &from) {
|
|
|
|
|
PyOperation &fromOperation = from.getOperation();
|
|
|
|
|
fromOperation.checkValid();
|
|
|
|
|
if (mlirLogicalResultIsFailure(mlirSymbolTableReplaceAllSymbolUses(
|
|
|
|
|
toMlirStringRef(oldSymbol), toMlirStringRef(newSymbol),
|
|
|
|
|
from.getOperation())))
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error("Symbol rename failed");
|
2021-11-28 20:30:18 -08:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void PySymbolTable::walkSymbolTables(PyOperationBase &from,
|
|
|
|
|
bool allSymUsesVisible,
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object callback) {
|
2021-11-28 20:30:18 -08:00
|
|
|
PyOperation &fromOperation = from.getOperation();
|
|
|
|
|
fromOperation.checkValid();
|
|
|
|
|
struct UserData {
|
|
|
|
|
PyMlirContextRef context;
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object callback;
|
2021-11-28 20:30:18 -08:00
|
|
|
bool gotException;
|
|
|
|
|
std::string exceptionWhat;
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object exceptionType;
|
2021-11-28 20:30:18 -08:00
|
|
|
};
|
|
|
|
|
UserData userData{
|
|
|
|
|
fromOperation.getContext(), std::move(callback), false, {}, {}};
|
|
|
|
|
mlirSymbolTableWalkSymbolTables(
|
|
|
|
|
fromOperation.get(), allSymUsesVisible,
|
|
|
|
|
[](MlirOperation foundOp, bool isVisible, void *calleeUserDataVoid) {
|
|
|
|
|
UserData *calleeUserData = static_cast<UserData *>(calleeUserDataVoid);
|
|
|
|
|
auto pyFoundOp =
|
|
|
|
|
PyOperation::forOperation(calleeUserData->context, foundOp);
|
|
|
|
|
if (calleeUserData->gotException)
|
|
|
|
|
return;
|
|
|
|
|
try {
|
|
|
|
|
calleeUserData->callback(pyFoundOp.getObject(), isVisible);
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
} catch (nb::python_error &e) {
|
2021-11-28 20:30:18 -08:00
|
|
|
calleeUserData->gotException = true;
|
|
|
|
|
calleeUserData->exceptionWhat = e.what();
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
calleeUserData->exceptionType = nb::borrow(e.type());
|
2021-11-28 20:30:18 -08:00
|
|
|
}
|
|
|
|
|
},
|
|
|
|
|
static_cast<void *>(&userData));
|
|
|
|
|
if (userData.gotException) {
|
|
|
|
|
std::string message("Exception raised in callback: ");
|
|
|
|
|
message.append(userData.exceptionWhat);
|
2022-01-02 22:02:18 +00:00
|
|
|
throw std::runtime_error(message);
|
2021-11-28 20:30:18 -08:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2020-10-20 11:21:05 +02:00
|
|
|
namespace {
|
|
|
|
|
/// CRTP base class for Python MLIR values that subclass Value and should be
|
|
|
|
|
/// castable from it. The value hierarchy is one level deep and is not supposed
|
|
|
|
|
/// to accommodate other levels unless core MLIR changes.
|
2021-03-19 11:57:01 -07:00
|
|
|
template <typename DerivedTy>
|
|
|
|
|
class PyConcreteValue : public PyValue {
|
2020-10-20 11:21:05 +02:00
|
|
|
public:
|
|
|
|
|
// Derived classes must define statics for:
|
|
|
|
|
// IsAFunctionTy isaFunction
|
|
|
|
|
// const char *pyClassName
|
|
|
|
|
// and redefine bindDerived.
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
using ClassTy = nb::class_<DerivedTy, PyValue>;
|
2020-11-29 18:31:11 -08:00
|
|
|
using IsAFunctionTy = bool (*)(MlirValue);
|
2020-10-20 11:21:05 +02:00
|
|
|
|
|
|
|
|
PyConcreteValue() = default;
|
|
|
|
|
PyConcreteValue(PyOperationRef operationRef, MlirValue value)
|
|
|
|
|
: PyValue(operationRef, value) {}
|
|
|
|
|
PyConcreteValue(PyValue &orig)
|
|
|
|
|
: PyConcreteValue(orig.getParentOperation(), castFrom(orig)) {}
|
|
|
|
|
|
|
|
|
|
/// Attempts to cast the original value to the derived type and throws on
|
|
|
|
|
/// type mismatches.
|
|
|
|
|
static MlirValue castFrom(PyValue &orig) {
|
|
|
|
|
if (!DerivedTy::isaFunction(orig.get())) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
auto origRepr = nb::cast<std::string>(nb::repr(nb::cast(orig)));
|
|
|
|
|
throw nb::value_error((Twine("Cannot cast value to ") +
|
2023-05-22 17:30:12 -05:00
|
|
|
DerivedTy::pyClassName + " (from " + origRepr +
|
|
|
|
|
")")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.str()
|
|
|
|
|
.c_str());
|
2020-10-20 11:21:05 +02:00
|
|
|
}
|
|
|
|
|
return orig.get();
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/// Binds the Python module objects to functions of this class.
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
static void bind(nb::module_ &m) {
|
|
|
|
|
auto cls = ClassTy(m, DerivedTy::pyClassName);
|
|
|
|
|
cls.def(nb::init<PyValue &>(), nb::keep_alive<0, 1>(), nb::arg("value"));
|
2021-11-28 14:08:06 -08:00
|
|
|
cls.def_static(
|
|
|
|
|
"isinstance",
|
|
|
|
|
[](PyValue &otherValue) -> bool {
|
|
|
|
|
return DerivedTy::isaFunction(otherValue);
|
|
|
|
|
},
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("other_value"));
|
[mlir][python] value casting (#69644)
This PR adds "value casting", i.e., a mechanism to wrap `ir.Value` in a
proxy class that overloads dunders such as `__add__`, `__sub__`, and
`__mul__` for fun and great profit.
This is thematically similar to
https://github.com/llvm/llvm-project/commit/bfb1ba752655bf09b35c486f6cc9817dbedfb1bb
and
https://github.com/llvm/llvm-project/commit/9566ee280607d91fa2e5eca730a6765ac84dfd0f.
The example in the test demonstrates the value of the feature (no pun
intended):
```python
@register_value_caster(F16Type.static_typeid)
@register_value_caster(F32Type.static_typeid)
@register_value_caster(F64Type.static_typeid)
@register_value_caster(IntegerType.static_typeid)
class ArithValue(Value):
__add__ = partialmethod(_binary_op, op="add")
__sub__ = partialmethod(_binary_op, op="sub")
__mul__ = partialmethod(_binary_op, op="mul")
a = arith.constant(value=FloatAttr.get(f16_t, 42.42))
b = a + a
# CHECK: ArithValue(%0 = arith.addf %cst, %cst : f16)
print(b)
a = arith.constant(value=FloatAttr.get(f32_t, 42.42))
b = a - a
# CHECK: ArithValue(%1 = arith.subf %cst_0, %cst_0 : f32)
print(b)
a = arith.constant(value=FloatAttr.get(f64_t, 42.42))
b = a * a
# CHECK: ArithValue(%2 = arith.mulf %cst_1, %cst_1 : f64)
print(b)
```
**EDIT**: this now goes through the bindings and thus supports automatic
casting of `OpResult` (including as an element of `OpResultList`),
`BlockArgument` (including as an element of `BlockArgumentList`), as
well as `Value`.
2023-11-07 10:49:41 -06:00
|
|
|
cls.def(MLIR_PYTHON_MAYBE_DOWNCAST_ATTR,
|
|
|
|
|
[](DerivedTy &self) { return self.maybeDownCast(); });
|
2020-10-20 11:21:05 +02:00
|
|
|
DerivedTy::bindDerived(cls);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/// Implemented by derived classes to add methods to the Python subclass.
|
|
|
|
|
static void bindDerived(ClassTy &m) {}
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
/// Python wrapper for MlirBlockArgument.
|
|
|
|
|
class PyBlockArgument : public PyConcreteValue<PyBlockArgument> {
|
|
|
|
|
public:
|
|
|
|
|
static constexpr IsAFunctionTy isaFunction = mlirValueIsABlockArgument;
|
|
|
|
|
static constexpr const char *pyClassName = "BlockArgument";
|
|
|
|
|
using PyConcreteValue::PyConcreteValue;
|
|
|
|
|
|
|
|
|
|
static void bindDerived(ClassTy &c) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
c.def_prop_ro("owner", [](PyBlockArgument &self) {
|
2020-10-20 11:21:05 +02:00
|
|
|
return PyBlock(self.getParentOperation(),
|
|
|
|
|
mlirBlockArgumentGetOwner(self.get()));
|
|
|
|
|
});
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
c.def_prop_ro("arg_number", [](PyBlockArgument &self) {
|
2020-10-20 11:21:05 +02:00
|
|
|
return mlirBlockArgumentGetArgNumber(self.get());
|
|
|
|
|
});
|
2021-11-28 14:08:06 -08:00
|
|
|
c.def(
|
|
|
|
|
"set_type",
|
|
|
|
|
[](PyBlockArgument &self, PyType type) {
|
|
|
|
|
return mlirBlockArgumentSetType(self.get(), type);
|
|
|
|
|
},
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("type"));
|
2020-10-20 11:21:05 +02:00
|
|
|
}
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
/// Python wrapper for MlirOpResult.
|
|
|
|
|
class PyOpResult : public PyConcreteValue<PyOpResult> {
|
|
|
|
|
public:
|
|
|
|
|
static constexpr IsAFunctionTy isaFunction = mlirValueIsAOpResult;
|
|
|
|
|
static constexpr const char *pyClassName = "OpResult";
|
|
|
|
|
using PyConcreteValue::PyConcreteValue;
|
|
|
|
|
|
|
|
|
|
static void bindDerived(ClassTy &c) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
c.def_prop_ro("owner", [](PyOpResult &self) {
|
2020-10-20 11:21:05 +02:00
|
|
|
assert(
|
|
|
|
|
mlirOperationEqual(self.getParentOperation()->get(),
|
|
|
|
|
mlirOpResultGetOwner(self.get())) &&
|
|
|
|
|
"expected the owner of the value in Python to match that in the IR");
|
2021-04-23 20:32:54 -06:00
|
|
|
return self.getParentOperation().getObject();
|
2020-10-20 11:21:05 +02:00
|
|
|
});
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
c.def_prop_ro("result_number", [](PyOpResult &self) {
|
2020-10-20 11:21:05 +02:00
|
|
|
return mlirOpResultGetResultNumber(self.get());
|
|
|
|
|
});
|
|
|
|
|
}
|
|
|
|
|
};
|
|
|
|
|
|
2021-10-04 11:38:20 +02:00
|
|
|
/// Returns the list of types of the values held by container.
|
|
|
|
|
template <typename Container>
|
2023-05-26 10:23:17 -05:00
|
|
|
static std::vector<MlirType> getValueTypes(Container &container,
|
|
|
|
|
PyMlirContextRef &context) {
|
|
|
|
|
std::vector<MlirType> result;
|
2022-07-21 14:00:37 +00:00
|
|
|
result.reserve(container.size());
|
|
|
|
|
for (int i = 0, e = container.size(); i < e; ++i) {
|
2023-05-26 10:23:17 -05:00
|
|
|
result.push_back(mlirValueGetType(container.getElement(i).get()));
|
2021-10-04 11:38:20 +02:00
|
|
|
}
|
|
|
|
|
return result;
|
|
|
|
|
}
|
|
|
|
|
|
2020-10-20 11:21:05 +02:00
|
|
|
/// A list of block arguments. Internally, these are stored as consecutive
|
|
|
|
|
/// elements, random access is cheap. The argument list is associated with the
|
|
|
|
|
/// operation that contains the block (detached blocks are not allowed in
|
|
|
|
|
/// Python bindings) and extends its lifetime.
|
2021-09-29 21:42:10 +02:00
|
|
|
class PyBlockArgumentList
|
|
|
|
|
: public Sliceable<PyBlockArgumentList, PyBlockArgument> {
|
2020-10-20 11:21:05 +02:00
|
|
|
public:
|
2021-09-29 21:42:10 +02:00
|
|
|
static constexpr const char *pyClassName = "BlockArgumentList";
|
[mlir][python] value casting (#69644)
This PR adds "value casting", i.e., a mechanism to wrap `ir.Value` in a
proxy class that overloads dunders such as `__add__`, `__sub__`, and
`__mul__` for fun and great profit.
This is thematically similar to
https://github.com/llvm/llvm-project/commit/bfb1ba752655bf09b35c486f6cc9817dbedfb1bb
and
https://github.com/llvm/llvm-project/commit/9566ee280607d91fa2e5eca730a6765ac84dfd0f.
The example in the test demonstrates the value of the feature (no pun
intended):
```python
@register_value_caster(F16Type.static_typeid)
@register_value_caster(F32Type.static_typeid)
@register_value_caster(F64Type.static_typeid)
@register_value_caster(IntegerType.static_typeid)
class ArithValue(Value):
__add__ = partialmethod(_binary_op, op="add")
__sub__ = partialmethod(_binary_op, op="sub")
__mul__ = partialmethod(_binary_op, op="mul")
a = arith.constant(value=FloatAttr.get(f16_t, 42.42))
b = a + a
# CHECK: ArithValue(%0 = arith.addf %cst, %cst : f16)
print(b)
a = arith.constant(value=FloatAttr.get(f32_t, 42.42))
b = a - a
# CHECK: ArithValue(%1 = arith.subf %cst_0, %cst_0 : f32)
print(b)
a = arith.constant(value=FloatAttr.get(f64_t, 42.42))
b = a * a
# CHECK: ArithValue(%2 = arith.mulf %cst_1, %cst_1 : f64)
print(b)
```
**EDIT**: this now goes through the bindings and thus supports automatic
casting of `OpResult` (including as an element of `OpResultList`),
`BlockArgument` (including as an element of `BlockArgumentList`), as
well as `Value`.
2023-11-07 10:49:41 -06:00
|
|
|
using SliceableT = Sliceable<PyBlockArgumentList, PyBlockArgument>;
|
2020-10-20 11:21:05 +02:00
|
|
|
|
2021-09-29 21:42:10 +02:00
|
|
|
PyBlockArgumentList(PyOperationRef operation, MlirBlock block,
|
|
|
|
|
intptr_t startIndex = 0, intptr_t length = -1,
|
|
|
|
|
intptr_t step = 1)
|
|
|
|
|
: Sliceable(startIndex,
|
|
|
|
|
length == -1 ? mlirBlockGetNumArguments(block) : length,
|
|
|
|
|
step),
|
|
|
|
|
operation(std::move(operation)), block(block) {}
|
|
|
|
|
|
2022-07-21 14:00:37 +00:00
|
|
|
static void bindDerived(ClassTy &c) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
c.def_prop_ro("types", [](PyBlockArgumentList &self) {
|
2022-07-21 14:00:37 +00:00
|
|
|
return getValueTypes(self, self.operation->getContext());
|
|
|
|
|
});
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
private:
|
|
|
|
|
/// Give the parent CRTP class access to hook implementations below.
|
|
|
|
|
friend class Sliceable<PyBlockArgumentList, PyBlockArgument>;
|
|
|
|
|
|
2021-09-29 21:42:10 +02:00
|
|
|
/// Returns the number of arguments in the list.
|
2022-07-21 14:00:37 +00:00
|
|
|
intptr_t getRawNumElements() {
|
2020-10-20 11:21:05 +02:00
|
|
|
operation->checkValid();
|
|
|
|
|
return mlirBlockGetNumArguments(block);
|
|
|
|
|
}
|
|
|
|
|
|
2022-07-21 14:00:37 +00:00
|
|
|
/// Returns `pos`-the element in the list.
|
|
|
|
|
PyBlockArgument getRawElement(intptr_t pos) {
|
2021-09-29 21:42:10 +02:00
|
|
|
MlirValue argument = mlirBlockGetArgument(block, pos);
|
|
|
|
|
return PyBlockArgument(operation, argument);
|
2020-10-20 11:21:05 +02:00
|
|
|
}
|
|
|
|
|
|
2021-09-29 21:42:10 +02:00
|
|
|
/// Returns a sublist of this list.
|
|
|
|
|
PyBlockArgumentList slice(intptr_t startIndex, intptr_t length,
|
|
|
|
|
intptr_t step) {
|
|
|
|
|
return PyBlockArgumentList(operation, block, startIndex, length, step);
|
2020-10-20 11:21:05 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
PyOperationRef operation;
|
|
|
|
|
MlirBlock block;
|
|
|
|
|
};
|
|
|
|
|
|
2020-11-06 15:09:54 +01:00
|
|
|
/// A list of operation operands. Internally, these are stored as consecutive
|
2023-10-02 15:37:25 -05:00
|
|
|
/// elements, random access is cheap. The (returned) operand list is associated
|
|
|
|
|
/// with the operation whose operands these are, and thus extends the lifetime
|
|
|
|
|
/// of this operation.
|
2020-11-06 15:09:54 +01:00
|
|
|
class PyOpOperandList : public Sliceable<PyOpOperandList, PyValue> {
|
2020-10-21 23:34:01 -07:00
|
|
|
public:
|
2020-11-06 15:09:54 +01:00
|
|
|
static constexpr const char *pyClassName = "OpOperandList";
|
[mlir][python] value casting (#69644)
This PR adds "value casting", i.e., a mechanism to wrap `ir.Value` in a
proxy class that overloads dunders such as `__add__`, `__sub__`, and
`__mul__` for fun and great profit.
This is thematically similar to
https://github.com/llvm/llvm-project/commit/bfb1ba752655bf09b35c486f6cc9817dbedfb1bb
and
https://github.com/llvm/llvm-project/commit/9566ee280607d91fa2e5eca730a6765ac84dfd0f.
The example in the test demonstrates the value of the feature (no pun
intended):
```python
@register_value_caster(F16Type.static_typeid)
@register_value_caster(F32Type.static_typeid)
@register_value_caster(F64Type.static_typeid)
@register_value_caster(IntegerType.static_typeid)
class ArithValue(Value):
__add__ = partialmethod(_binary_op, op="add")
__sub__ = partialmethod(_binary_op, op="sub")
__mul__ = partialmethod(_binary_op, op="mul")
a = arith.constant(value=FloatAttr.get(f16_t, 42.42))
b = a + a
# CHECK: ArithValue(%0 = arith.addf %cst, %cst : f16)
print(b)
a = arith.constant(value=FloatAttr.get(f32_t, 42.42))
b = a - a
# CHECK: ArithValue(%1 = arith.subf %cst_0, %cst_0 : f32)
print(b)
a = arith.constant(value=FloatAttr.get(f64_t, 42.42))
b = a * a
# CHECK: ArithValue(%2 = arith.mulf %cst_1, %cst_1 : f64)
print(b)
```
**EDIT**: this now goes through the bindings and thus supports automatic
casting of `OpResult` (including as an element of `OpResultList`),
`BlockArgument` (including as an element of `BlockArgumentList`), as
well as `Value`.
2023-11-07 10:49:41 -06:00
|
|
|
using SliceableT = Sliceable<PyOpOperandList, PyValue>;
|
2020-10-21 23:34:01 -07:00
|
|
|
|
2020-11-06 15:09:54 +01:00
|
|
|
PyOpOperandList(PyOperationRef operation, intptr_t startIndex = 0,
|
|
|
|
|
intptr_t length = -1, intptr_t step = 1)
|
|
|
|
|
: Sliceable(startIndex,
|
|
|
|
|
length == -1 ? mlirOperationGetNumOperands(operation->get())
|
|
|
|
|
: length,
|
|
|
|
|
step),
|
|
|
|
|
operation(operation) {}
|
|
|
|
|
|
2022-07-21 14:00:37 +00:00
|
|
|
void dunderSetItem(intptr_t index, PyValue value) {
|
|
|
|
|
index = wrapIndex(index);
|
|
|
|
|
mlirOperationSetOperand(operation->get(), index, value.get());
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void bindDerived(ClassTy &c) {
|
|
|
|
|
c.def("__setitem__", &PyOpOperandList::dunderSetItem);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
private:
|
|
|
|
|
/// Give the parent CRTP class access to hook implementations below.
|
|
|
|
|
friend class Sliceable<PyOpOperandList, PyValue>;
|
|
|
|
|
|
|
|
|
|
intptr_t getRawNumElements() {
|
2020-10-21 23:34:01 -07:00
|
|
|
operation->checkValid();
|
|
|
|
|
return mlirOperationGetNumOperands(operation->get());
|
|
|
|
|
}
|
|
|
|
|
|
2022-07-21 14:00:37 +00:00
|
|
|
PyValue getRawElement(intptr_t pos) {
|
2021-07-14 20:19:27 -07:00
|
|
|
MlirValue operand = mlirOperationGetOperand(operation->get(), pos);
|
|
|
|
|
MlirOperation owner;
|
|
|
|
|
if (mlirValueIsAOpResult(operand))
|
|
|
|
|
owner = mlirOpResultGetOwner(operand);
|
|
|
|
|
else if (mlirValueIsABlockArgument(operand))
|
|
|
|
|
owner = mlirBlockGetParentOperation(mlirBlockArgumentGetOwner(operand));
|
|
|
|
|
else
|
|
|
|
|
assert(false && "Value must be an block arg or op result.");
|
|
|
|
|
PyOperationRef pyOwner =
|
|
|
|
|
PyOperation::forOperation(operation->getContext(), owner);
|
|
|
|
|
return PyValue(pyOwner, operand);
|
2020-10-21 23:34:01 -07:00
|
|
|
}
|
|
|
|
|
|
2020-11-06 15:09:54 +01:00
|
|
|
PyOpOperandList slice(intptr_t startIndex, intptr_t length, intptr_t step) {
|
|
|
|
|
return PyOpOperandList(operation, startIndex, length, step);
|
2020-10-21 23:34:01 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
PyOperationRef operation;
|
|
|
|
|
};
|
|
|
|
|
|
2020-10-20 11:21:05 +02:00
|
|
|
/// A list of operation results. Internally, these are stored as consecutive
|
2023-10-02 15:37:25 -05:00
|
|
|
/// elements, random access is cheap. The (returned) result list is associated
|
|
|
|
|
/// with the operation whose results these are, and thus extends the lifetime of
|
|
|
|
|
/// this operation.
|
2020-11-06 15:09:54 +01:00
|
|
|
class PyOpResultList : public Sliceable<PyOpResultList, PyOpResult> {
|
2020-10-20 11:21:05 +02:00
|
|
|
public:
|
2020-11-06 15:09:54 +01:00
|
|
|
static constexpr const char *pyClassName = "OpResultList";
|
[mlir][python] value casting (#69644)
This PR adds "value casting", i.e., a mechanism to wrap `ir.Value` in a
proxy class that overloads dunders such as `__add__`, `__sub__`, and
`__mul__` for fun and great profit.
This is thematically similar to
https://github.com/llvm/llvm-project/commit/bfb1ba752655bf09b35c486f6cc9817dbedfb1bb
and
https://github.com/llvm/llvm-project/commit/9566ee280607d91fa2e5eca730a6765ac84dfd0f.
The example in the test demonstrates the value of the feature (no pun
intended):
```python
@register_value_caster(F16Type.static_typeid)
@register_value_caster(F32Type.static_typeid)
@register_value_caster(F64Type.static_typeid)
@register_value_caster(IntegerType.static_typeid)
class ArithValue(Value):
__add__ = partialmethod(_binary_op, op="add")
__sub__ = partialmethod(_binary_op, op="sub")
__mul__ = partialmethod(_binary_op, op="mul")
a = arith.constant(value=FloatAttr.get(f16_t, 42.42))
b = a + a
# CHECK: ArithValue(%0 = arith.addf %cst, %cst : f16)
print(b)
a = arith.constant(value=FloatAttr.get(f32_t, 42.42))
b = a - a
# CHECK: ArithValue(%1 = arith.subf %cst_0, %cst_0 : f32)
print(b)
a = arith.constant(value=FloatAttr.get(f64_t, 42.42))
b = a * a
# CHECK: ArithValue(%2 = arith.mulf %cst_1, %cst_1 : f64)
print(b)
```
**EDIT**: this now goes through the bindings and thus supports automatic
casting of `OpResult` (including as an element of `OpResultList`),
`BlockArgument` (including as an element of `BlockArgumentList`), as
well as `Value`.
2023-11-07 10:49:41 -06:00
|
|
|
using SliceableT = Sliceable<PyOpResultList, PyOpResult>;
|
2020-10-20 11:21:05 +02:00
|
|
|
|
2020-11-06 15:09:54 +01:00
|
|
|
PyOpResultList(PyOperationRef operation, intptr_t startIndex = 0,
|
|
|
|
|
intptr_t length = -1, intptr_t step = 1)
|
|
|
|
|
: Sliceable(startIndex,
|
|
|
|
|
length == -1 ? mlirOperationGetNumResults(operation->get())
|
|
|
|
|
: length,
|
|
|
|
|
step),
|
[mlir][python] value casting (#69644)
This PR adds "value casting", i.e., a mechanism to wrap `ir.Value` in a
proxy class that overloads dunders such as `__add__`, `__sub__`, and
`__mul__` for fun and great profit.
This is thematically similar to
https://github.com/llvm/llvm-project/commit/bfb1ba752655bf09b35c486f6cc9817dbedfb1bb
and
https://github.com/llvm/llvm-project/commit/9566ee280607d91fa2e5eca730a6765ac84dfd0f.
The example in the test demonstrates the value of the feature (no pun
intended):
```python
@register_value_caster(F16Type.static_typeid)
@register_value_caster(F32Type.static_typeid)
@register_value_caster(F64Type.static_typeid)
@register_value_caster(IntegerType.static_typeid)
class ArithValue(Value):
__add__ = partialmethod(_binary_op, op="add")
__sub__ = partialmethod(_binary_op, op="sub")
__mul__ = partialmethod(_binary_op, op="mul")
a = arith.constant(value=FloatAttr.get(f16_t, 42.42))
b = a + a
# CHECK: ArithValue(%0 = arith.addf %cst, %cst : f16)
print(b)
a = arith.constant(value=FloatAttr.get(f32_t, 42.42))
b = a - a
# CHECK: ArithValue(%1 = arith.subf %cst_0, %cst_0 : f32)
print(b)
a = arith.constant(value=FloatAttr.get(f64_t, 42.42))
b = a * a
# CHECK: ArithValue(%2 = arith.mulf %cst_1, %cst_1 : f64)
print(b)
```
**EDIT**: this now goes through the bindings and thus supports automatic
casting of `OpResult` (including as an element of `OpResultList`),
`BlockArgument` (including as an element of `BlockArgumentList`), as
well as `Value`.
2023-11-07 10:49:41 -06:00
|
|
|
operation(std::move(operation)) {}
|
2020-11-06 15:09:54 +01:00
|
|
|
|
2022-07-21 14:00:37 +00:00
|
|
|
static void bindDerived(ClassTy &c) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
c.def_prop_ro("types", [](PyOpResultList &self) {
|
2022-07-21 14:00:37 +00:00
|
|
|
return getValueTypes(self, self.operation->getContext());
|
|
|
|
|
});
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
c.def_prop_ro("owner", [](PyOpResultList &self) {
|
2023-08-11 20:39:56 -05:00
|
|
|
return self.operation->createOpView();
|
|
|
|
|
});
|
2022-07-21 14:00:37 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
private:
|
|
|
|
|
/// Give the parent CRTP class access to hook implementations below.
|
|
|
|
|
friend class Sliceable<PyOpResultList, PyOpResult>;
|
|
|
|
|
|
|
|
|
|
intptr_t getRawNumElements() {
|
2020-10-20 11:21:05 +02:00
|
|
|
operation->checkValid();
|
|
|
|
|
return mlirOperationGetNumResults(operation->get());
|
|
|
|
|
}
|
|
|
|
|
|
2022-07-21 14:00:37 +00:00
|
|
|
PyOpResult getRawElement(intptr_t index) {
|
2020-10-20 11:21:05 +02:00
|
|
|
PyValue value(operation, mlirOperationGetResult(operation->get(), index));
|
|
|
|
|
return PyOpResult(value);
|
|
|
|
|
}
|
|
|
|
|
|
2020-11-06 15:09:54 +01:00
|
|
|
PyOpResultList slice(intptr_t startIndex, intptr_t length, intptr_t step) {
|
|
|
|
|
return PyOpResultList(operation, startIndex, length, step);
|
2020-10-20 11:21:05 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
PyOperationRef operation;
|
|
|
|
|
};
|
|
|
|
|
|
2023-10-02 15:37:25 -05:00
|
|
|
/// A list of operation successors. Internally, these are stored as consecutive
|
|
|
|
|
/// elements, random access is cheap. The (returned) successor list is
|
|
|
|
|
/// associated with the operation whose successors these are, and thus extends
|
|
|
|
|
/// the lifetime of this operation.
|
|
|
|
|
class PyOpSuccessors : public Sliceable<PyOpSuccessors, PyBlock> {
|
|
|
|
|
public:
|
|
|
|
|
static constexpr const char *pyClassName = "OpSuccessors";
|
|
|
|
|
|
|
|
|
|
PyOpSuccessors(PyOperationRef operation, intptr_t startIndex = 0,
|
|
|
|
|
intptr_t length = -1, intptr_t step = 1)
|
|
|
|
|
: Sliceable(startIndex,
|
|
|
|
|
length == -1 ? mlirOperationGetNumSuccessors(operation->get())
|
|
|
|
|
: length,
|
|
|
|
|
step),
|
|
|
|
|
operation(operation) {}
|
|
|
|
|
|
|
|
|
|
void dunderSetItem(intptr_t index, PyBlock block) {
|
|
|
|
|
index = wrapIndex(index);
|
|
|
|
|
mlirOperationSetSuccessor(operation->get(), index, block.get());
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void bindDerived(ClassTy &c) {
|
|
|
|
|
c.def("__setitem__", &PyOpSuccessors::dunderSetItem);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
private:
|
|
|
|
|
/// Give the parent CRTP class access to hook implementations below.
|
|
|
|
|
friend class Sliceable<PyOpSuccessors, PyBlock>;
|
|
|
|
|
|
|
|
|
|
intptr_t getRawNumElements() {
|
|
|
|
|
operation->checkValid();
|
|
|
|
|
return mlirOperationGetNumSuccessors(operation->get());
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
PyBlock getRawElement(intptr_t pos) {
|
|
|
|
|
MlirBlock block = mlirOperationGetSuccessor(operation->get(), pos);
|
|
|
|
|
return PyBlock(operation, block);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
PyOpSuccessors slice(intptr_t startIndex, intptr_t length, intptr_t step) {
|
|
|
|
|
return PyOpSuccessors(operation, startIndex, length, step);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
PyOperationRef operation;
|
|
|
|
|
};
|
|
|
|
|
|
2020-11-06 11:25:41 +01:00
|
|
|
/// A list of operation attributes. Can be indexed by name, producing
|
|
|
|
|
/// attributes, or by index, producing named attributes.
|
|
|
|
|
class PyOpAttributeMap {
|
|
|
|
|
public:
|
2022-01-02 01:26:44 +00:00
|
|
|
PyOpAttributeMap(PyOperationRef operation)
|
|
|
|
|
: operation(std::move(operation)) {}
|
2020-11-06 11:25:41 +01:00
|
|
|
|
2023-07-04 16:06:17 -04:00
|
|
|
MlirAttribute dunderGetItemNamed(const std::string &name) {
|
2020-11-24 18:35:22 +00:00
|
|
|
MlirAttribute attr = mlirOperationGetAttributeByName(operation->get(),
|
|
|
|
|
toMlirStringRef(name));
|
2020-11-06 11:25:41 +01:00
|
|
|
if (mlirAttributeIsNull(attr)) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::key_error("attempt to access a non-existent attribute");
|
2020-11-06 11:25:41 +01:00
|
|
|
}
|
2023-07-04 16:06:17 -04:00
|
|
|
return attr;
|
2020-11-06 11:25:41 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
PyNamedAttribute dunderGetItemIndexed(intptr_t index) {
|
|
|
|
|
if (index < 0 || index >= dunderLen()) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::index_error("attempt to access out of bounds attribute");
|
2020-11-06 11:25:41 +01:00
|
|
|
}
|
|
|
|
|
MlirNamedAttribute namedAttr =
|
|
|
|
|
mlirOperationGetAttribute(operation->get(), index);
|
2020-12-11 18:50:04 +00:00
|
|
|
return PyNamedAttribute(
|
|
|
|
|
namedAttr.attribute,
|
2021-11-11 01:44:58 +00:00
|
|
|
std::string(mlirIdentifierStr(namedAttr.name).data,
|
|
|
|
|
mlirIdentifierStr(namedAttr.name).length));
|
2020-11-06 11:25:41 +01:00
|
|
|
}
|
|
|
|
|
|
2022-01-02 01:26:44 +00:00
|
|
|
void dunderSetItem(const std::string &name, const PyAttribute &attr) {
|
2020-11-24 18:35:22 +00:00
|
|
|
mlirOperationSetAttributeByName(operation->get(), toMlirStringRef(name),
|
2020-11-29 13:30:23 -08:00
|
|
|
attr);
|
2020-11-17 18:28:16 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void dunderDelItem(const std::string &name) {
|
2020-11-24 18:35:22 +00:00
|
|
|
int removed = mlirOperationRemoveAttributeByName(operation->get(),
|
|
|
|
|
toMlirStringRef(name));
|
2020-11-17 18:28:16 +01:00
|
|
|
if (!removed)
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::key_error("attempt to delete a non-existent attribute");
|
2020-11-17 18:28:16 +01:00
|
|
|
}
|
|
|
|
|
|
2020-11-06 11:25:41 +01:00
|
|
|
intptr_t dunderLen() {
|
|
|
|
|
return mlirOperationGetNumAttributes(operation->get());
|
|
|
|
|
}
|
|
|
|
|
|
[mlir] Add basic support for attributes in ODS-generated Python bindings
In ODS, attributes of an operation can be provided as a part of the "arguments"
field, together with operands. Such attributes are accepted by the op builder
and have accessors generated.
Implement similar functionality for ODS-generated op-specific Python bindings:
the `__init__` method now accepts arguments together with operands, in the same
order as in the ODS `arguments` field; the instance properties are introduced
to OpView classes to access the attributes.
This initial implementation accepts and returns instances of the corresponding
attribute class, and not the underlying values since the mapping scheme of the
value types between C++, C and Python is not yet clear. Default-valued
attributes are not supported as that would require Python to be able to parse
C++ literals.
Since attributes in ODS are tightely related to the actual C++ type system,
provide a separate Tablegen file with the mapping between ODS storage type for
attributes (typically, the underlying C++ attribute class), and the
corresponding class name. So far, this might look unnecessary since all names
match exactly, but this is not necessarily the cases for non-standard,
out-of-tree attributes, which may also be placed in non-default namespaces or
Python modules. This also allows out-of-tree users to generate Python bindings
without having to modify the bindings generator itself. Storage type was
preferred over the Tablegen "def" of the attribute class because ODS
essentially encodes attribute _constraints_ rather than classes, e.g. there may
be many Tablegen "def"s in the ODS that correspond to the same attribute type
with additional constraints
The presence of the explicit mapping requires the change in the .td file
structure: instead of just calling the bindings generator directly on the main
ODS file of the dialect, it becomes necessary to create a new file that
includes the main ODS file of the dialect and provides the mapping for
attribute types. Arguably, this approach offers better separability of the
Python bindings in the build system as the main dialect no longer needs to know
that it is being processed by the bindings generator.
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D91542
2020-11-16 16:17:03 +01:00
|
|
|
bool dunderContains(const std::string &name) {
|
2020-11-24 18:35:22 +00:00
|
|
|
return !mlirAttributeIsNull(mlirOperationGetAttributeByName(
|
|
|
|
|
operation->get(), toMlirStringRef(name)));
|
[mlir] Add basic support for attributes in ODS-generated Python bindings
In ODS, attributes of an operation can be provided as a part of the "arguments"
field, together with operands. Such attributes are accepted by the op builder
and have accessors generated.
Implement similar functionality for ODS-generated op-specific Python bindings:
the `__init__` method now accepts arguments together with operands, in the same
order as in the ODS `arguments` field; the instance properties are introduced
to OpView classes to access the attributes.
This initial implementation accepts and returns instances of the corresponding
attribute class, and not the underlying values since the mapping scheme of the
value types between C++, C and Python is not yet clear. Default-valued
attributes are not supported as that would require Python to be able to parse
C++ literals.
Since attributes in ODS are tightely related to the actual C++ type system,
provide a separate Tablegen file with the mapping between ODS storage type for
attributes (typically, the underlying C++ attribute class), and the
corresponding class name. So far, this might look unnecessary since all names
match exactly, but this is not necessarily the cases for non-standard,
out-of-tree attributes, which may also be placed in non-default namespaces or
Python modules. This also allows out-of-tree users to generate Python bindings
without having to modify the bindings generator itself. Storage type was
preferred over the Tablegen "def" of the attribute class because ODS
essentially encodes attribute _constraints_ rather than classes, e.g. there may
be many Tablegen "def"s in the ODS that correspond to the same attribute type
with additional constraints
The presence of the explicit mapping requires the change in the .td file
structure: instead of just calling the bindings generator directly on the main
ODS file of the dialect, it becomes necessary to create a new file that
includes the main ODS file of the dialect and provides the mapping for
attribute types. Arguably, this approach offers better separability of the
Python bindings in the build system as the main dialect no longer needs to know
that it is being processed by the bindings generator.
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D91542
2020-11-16 16:17:03 +01:00
|
|
|
}
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
static void bind(nb::module_ &m) {
|
|
|
|
|
nb::class_<PyOpAttributeMap>(m, "OpAttributeMap")
|
[mlir] Add basic support for attributes in ODS-generated Python bindings
In ODS, attributes of an operation can be provided as a part of the "arguments"
field, together with operands. Such attributes are accepted by the op builder
and have accessors generated.
Implement similar functionality for ODS-generated op-specific Python bindings:
the `__init__` method now accepts arguments together with operands, in the same
order as in the ODS `arguments` field; the instance properties are introduced
to OpView classes to access the attributes.
This initial implementation accepts and returns instances of the corresponding
attribute class, and not the underlying values since the mapping scheme of the
value types between C++, C and Python is not yet clear. Default-valued
attributes are not supported as that would require Python to be able to parse
C++ literals.
Since attributes in ODS are tightely related to the actual C++ type system,
provide a separate Tablegen file with the mapping between ODS storage type for
attributes (typically, the underlying C++ attribute class), and the
corresponding class name. So far, this might look unnecessary since all names
match exactly, but this is not necessarily the cases for non-standard,
out-of-tree attributes, which may also be placed in non-default namespaces or
Python modules. This also allows out-of-tree users to generate Python bindings
without having to modify the bindings generator itself. Storage type was
preferred over the Tablegen "def" of the attribute class because ODS
essentially encodes attribute _constraints_ rather than classes, e.g. there may
be many Tablegen "def"s in the ODS that correspond to the same attribute type
with additional constraints
The presence of the explicit mapping requires the change in the .td file
structure: instead of just calling the bindings generator directly on the main
ODS file of the dialect, it becomes necessary to create a new file that
includes the main ODS file of the dialect and provides the mapping for
attribute types. Arguably, this approach offers better separability of the
Python bindings in the build system as the main dialect no longer needs to know
that it is being processed by the bindings generator.
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D91542
2020-11-16 16:17:03 +01:00
|
|
|
.def("__contains__", &PyOpAttributeMap::dunderContains)
|
2020-11-06 11:25:41 +01:00
|
|
|
.def("__len__", &PyOpAttributeMap::dunderLen)
|
|
|
|
|
.def("__getitem__", &PyOpAttributeMap::dunderGetItemNamed)
|
2020-11-17 18:28:16 +01:00
|
|
|
.def("__getitem__", &PyOpAttributeMap::dunderGetItemIndexed)
|
|
|
|
|
.def("__setitem__", &PyOpAttributeMap::dunderSetItem)
|
|
|
|
|
.def("__delitem__", &PyOpAttributeMap::dunderDelItem);
|
2020-11-06 11:25:41 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
private:
|
|
|
|
|
PyOperationRef operation;
|
|
|
|
|
};
|
|
|
|
|
|
2021-12-07 18:27:58 +00:00
|
|
|
} // namespace
|
2020-10-20 11:21:05 +02:00
|
|
|
|
2020-08-19 15:33:02 -07:00
|
|
|
//------------------------------------------------------------------------------
|
2021-03-19 11:57:01 -07:00
|
|
|
// Populates the core exports of the 'ir' submodule.
|
2020-08-19 15:33:02 -07:00
|
|
|
//------------------------------------------------------------------------------
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
void mlir::python::populateIRCore(nb::module_ &m) {
|
2024-12-29 09:13:10 -08:00
|
|
|
// disable leak warnings which tend to be false positives.
|
|
|
|
|
nb::set_leak_warnings(false);
|
2022-01-03 16:39:58 -08:00
|
|
|
//----------------------------------------------------------------------------
|
|
|
|
|
// Enums.
|
|
|
|
|
//----------------------------------------------------------------------------
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::enum_<MlirDiagnosticSeverity>(m, "DiagnosticSeverity")
|
2022-01-03 16:39:58 -08:00
|
|
|
.value("ERROR", MlirDiagnosticError)
|
|
|
|
|
.value("WARNING", MlirDiagnosticWarning)
|
|
|
|
|
.value("NOTE", MlirDiagnosticNote)
|
|
|
|
|
.value("REMARK", MlirDiagnosticRemark);
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::enum_<MlirWalkOrder>(m, "WalkOrder")
|
2024-04-17 15:09:47 +09:00
|
|
|
.value("PRE_ORDER", MlirWalkPreOrder)
|
|
|
|
|
.value("POST_ORDER", MlirWalkPostOrder);
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::enum_<MlirWalkResult>(m, "WalkResult")
|
2024-04-17 15:09:47 +09:00
|
|
|
.value("ADVANCE", MlirWalkResultAdvance)
|
|
|
|
|
.value("INTERRUPT", MlirWalkResultInterrupt)
|
|
|
|
|
.value("SKIP", MlirWalkResultSkip);
|
|
|
|
|
|
2022-01-03 16:39:58 -08:00
|
|
|
//----------------------------------------------------------------------------
|
|
|
|
|
// Mapping of Diagnostics.
|
|
|
|
|
//----------------------------------------------------------------------------
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::class_<PyDiagnostic>(m, "Diagnostic")
|
|
|
|
|
.def_prop_ro("severity", &PyDiagnostic::getSeverity)
|
|
|
|
|
.def_prop_ro("location", &PyDiagnostic::getLocation)
|
|
|
|
|
.def_prop_ro("message", &PyDiagnostic::getMessage)
|
|
|
|
|
.def_prop_ro("notes", &PyDiagnostic::getNotes)
|
|
|
|
|
.def("__str__", [](PyDiagnostic &self) -> nb::str {
|
2022-01-03 16:39:58 -08:00
|
|
|
if (!self.isValid())
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
return nb::str("<Invalid Diagnostic>");
|
2022-01-03 16:39:58 -08:00
|
|
|
return self.getMessage();
|
|
|
|
|
});
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::class_<PyDiagnostic::DiagnosticInfo>(m, "DiagnosticInfo")
|
|
|
|
|
.def("__init__",
|
|
|
|
|
[](PyDiagnostic::DiagnosticInfo &self, PyDiagnostic diag) {
|
|
|
|
|
new (&self) PyDiagnostic::DiagnosticInfo(diag.getInfo());
|
|
|
|
|
})
|
|
|
|
|
.def_ro("severity", &PyDiagnostic::DiagnosticInfo::severity)
|
|
|
|
|
.def_ro("location", &PyDiagnostic::DiagnosticInfo::location)
|
|
|
|
|
.def_ro("message", &PyDiagnostic::DiagnosticInfo::message)
|
|
|
|
|
.def_ro("notes", &PyDiagnostic::DiagnosticInfo::notes)
|
[mlir][python] Capture error diagnostics in exceptions
This updates most (all?) error-diagnostic-emitting python APIs to
capture error diagnostics and include them in the raised exception's
message:
```
>>> Operation.parse('"arith.addi"() : () -> ()'))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
mlir._mlir_libs.MLIRError: Unable to parse operation assembly:
error: "-":1:1: 'arith.addi' op requires one result
note: "-":1:1: see current operation: "arith.addi"() : () -> ()
```
The diagnostic information is available on the exception for users who
may want to customize the error message:
```
>>> try:
... Operation.parse('"arith.addi"() : () -> ()')
... except MLIRError as e:
... print(e.message)
... print(e.error_diagnostics)
... print(e.error_diagnostics[0].message)
...
Unable to parse operation assembly
[<mlir._mlir_libs._mlir.ir.DiagnosticInfo object at 0x7fed32bd6b70>]
'arith.addi' op requires one result
```
Error diagnostics captured in exceptions aren't propagated to diagnostic
handlers, to avoid double-reporting of errors. The context-level
`emit_error_diagnostics` option can be used to revert to the old
behaviour, causing error diagnostics to be reported to handlers instead
of as part of exceptions.
API changes:
- `Operation.verify` now raises an exception on verification failure,
instead of returning `false`
- The exception raised by the following methods has been changed to
`MLIRError`:
- `PassManager.run`
- `{Module,Operation,Type,Attribute}.parse`
- `{RankedTensorType,UnrankedTensorType}.get`
- `{MemRefType,UnrankedMemRefType}.get`
- `VectorType.get`
- `FloatAttr.get`
closes #60595
depends on D144804, D143830
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D143869
2023-02-07 16:07:50 -05:00
|
|
|
.def("__str__",
|
|
|
|
|
[](PyDiagnostic::DiagnosticInfo &self) { return self.message; });
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::class_<PyDiagnosticHandler>(m, "DiagnosticHandler")
|
2022-01-03 16:39:58 -08:00
|
|
|
.def("detach", &PyDiagnosticHandler::detach)
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro("attached", &PyDiagnosticHandler::isAttached)
|
|
|
|
|
.def_prop_ro("had_error", &PyDiagnosticHandler::getHadError)
|
2022-01-03 16:39:58 -08:00
|
|
|
.def("__enter__", &PyDiagnosticHandler::contextEnter)
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def("__exit__", &PyDiagnosticHandler::contextExit,
|
|
|
|
|
nb::arg("exc_type").none(), nb::arg("exc_value").none(),
|
|
|
|
|
nb::arg("traceback").none());
|
2022-01-03 16:39:58 -08:00
|
|
|
|
2021-04-16 12:54:43 +00:00
|
|
|
//----------------------------------------------------------------------------
|
2021-04-19 13:37:01 +02:00
|
|
|
// Mapping of MlirContext.
|
[mlir] Overhaul C/Python registration APIs to properly scope registration/loading activities.
Since the very first commits, the Python and C MLIR APIs have had mis-placed registration/load functionality for dialects, extensions, etc. This was done pragmatically in order to get bootstrapped and then just grew in. Downstreams largely bypass and do their own thing by providing various APIs to register things they need. Meanwhile, the C++ APIs have stabilized around this and it would make sense to follow suit.
The thing we have observed in canonical usage by downstreams is that each downstream tends to have native entry points that configure its installation to its preferences with one-stop APIs. This patch leans in to this approach with `RegisterEverything.h` and `mlir._mlir_libs._mlirRegisterEverything` being the one-stop entry points for the "upstream packages". The `_mlir_libs.__init__.py` now allows customization of the environment and Context by adding "initialization modules" to the `_mlir_libs` package. If present, `_mlirRegisterEverything` is treated as such a module. Others can be added by downstreams by adding a `_site_initialize_{i}.py` module, where '{i}' is a number starting with zero. The number will be incremented and corresponding module loaded until one is not found. Initialization modules can:
* Perform load time customization to the global environment (i.e. registering passes, hooks, etc).
* Define a `register_dialects(registry: DialectRegistry)` function that can extend the `DialectRegistry` that will be used to bootstrap the `Context`.
* Define a `context_init_hook(context: Context)` function that will be added to a list of callbacks which will be invoked after dialect registration during `Context` initialization.
Note that the `MLIRPythonExtension.RegisterEverything` is not included by default when building a downstream (its corresponding behavior was prior). For downstreams which need the default MLIR initialization to take place, they must add this back in to their Python CMake build just like they add their own components (i.e. to `add_mlir_python_common_capi_library` and `add_mlir_python_modules`). It is perfectly valid to not do this, in which case, only the things explicitly depended on and initialized by downstreams will be built/packaged. If the downstream has not been set up for this, it is recommended to simply add this back for the time being and pay the build time/package size cost.
CMake changes:
* `MLIRCAPIRegistration` -> `MLIRCAPIRegisterEverything` (renamed to signify what it does and force an evaluation: a number of places were incidentally linking this very expensive target)
* `MLIRPythonSoure.Passes` removed (without replacement: just drop)
* `MLIRPythonExtension.AllPassesRegistration` removed (without replacement: just drop)
* `MLIRPythonExtension.Conversions` removed (without replacement: just drop)
* `MLIRPythonExtension.Transforms` removed (without replacement: just drop)
Header changes:
* `mlir-c/Registration.h` is deleted. Dialect registration functionality is now in `IR.h`. Registration of upstream features are in `mlir-c/RegisterEverything.h`. When updating MLIR and a couple of downstreams, I found that proper usage was commingled so required making a choice vs just blind S&R.
Python APIs removed:
* mlir.transforms and mlir.conversions (previously only had an __init__.py which indirectly triggered `mlirRegisterTransformsPasses()` and `mlirRegisterConversionPasses()` respectively). Downstream impact: Remove these imports if present (they now happen as part of default initialization).
* mlir._mlir_libs._all_passes_registration, mlir._mlir_libs._mlirTransforms, mlir._mlir_libs._mlirConversions. Downstream impact: None expected (these were internally used).
C-APIs changed:
* mlirRegisterAllDialects(MlirContext) now takes an MlirDialectRegistry instead. It also used to trigger loading of all dialects, which was already marked with a TODO to remove -- it no longer does, and for direct use, dialects must be explicitly loaded. Downstream impact: Direct C-API users must ensure that needed dialects are loaded or call `mlirContextLoadAllAvailableDialects(MlirContext)` to emulate the prior behavior. Also see the `ir.c` test case (e.g. ` mlirContextGetOrLoadDialect(ctx, mlirStringRefCreateFromCString("func"));`).
* mlirDialectHandle* APIs were moved from Registration.h (which now is restricted to just global/upstream registration) to IR.h, arguably where it should have been. Downstream impact: include correct header (likely already doing so).
C-APIs added:
* mlirContextLoadAllAvailableDialects(MlirContext): Corresponds to C++ API with the same purpose.
Python APIs added:
* mlir.ir.DialectRegistry: Mapping for an MlirDialectRegistry.
* mlir.ir.Context.append_dialect_registry(MlirDialectRegistry)
* mlir.ir.Context.load_all_available_dialects()
* mlir._mlir_libs._mlirAllRegistration: New native extension that exposes a `register_dialects(MlirDialectRegistry)` entry point and performs all upstream pass/conversion/transforms registration on init. In this first step, we eagerly load this as part of the __init__.py and use it to monkey patch the Context to emulate prior behavior.
* Type caster and capsule support for MlirDialectRegistry
This should make it possible to build downstream Python dialects that only depend on a subset of MLIR. See: https://github.com/llvm/llvm-project/issues/56037
Here is an example PR, minimally adapting IREE to these changes: https://github.com/iree-org/iree/pull/9638/files In this situation, IREE is opting to not link everything, since it is already configuring the Context to its liking. For projects that would just like to not think about it and pull in everything, add `MLIRPythonExtension.RegisterEverything` to the list of Python sources getting built, and the old behavior will continue.
Reviewed By: mehdi_amini, ftynse
Differential Revision: https://reviews.llvm.org/D128593
2022-07-16 16:09:03 -07:00
|
|
|
// Note that this is exported as _BaseContext. The containing, Python level
|
|
|
|
|
// __init__.py will subclass it with site-specific functionality and set a
|
|
|
|
|
// "Context" attribute on this module.
|
2021-03-19 11:57:01 -07:00
|
|
|
//----------------------------------------------------------------------------
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::class_<PyMlirContext>(m, "_BaseContext")
|
|
|
|
|
.def("__init__",
|
|
|
|
|
[](PyMlirContext &self) {
|
|
|
|
|
MlirContext context = mlirContextCreateWithThreading(false);
|
|
|
|
|
new (&self) PyMlirContext(context);
|
|
|
|
|
})
|
2021-03-19 11:57:01 -07:00
|
|
|
.def_static("_get_live_count", &PyMlirContext::getLiveCount)
|
|
|
|
|
.def("_get_context_again",
|
|
|
|
|
[](PyMlirContext &self) {
|
|
|
|
|
PyMlirContextRef ref = PyMlirContext::forContext(self.get());
|
|
|
|
|
return ref.releaseObject();
|
|
|
|
|
})
|
|
|
|
|
.def("_get_live_operation_count", &PyMlirContext::getLiveOperationCount)
|
2024-02-08 11:39:06 -08:00
|
|
|
.def("_get_live_operation_objects",
|
|
|
|
|
&PyMlirContext::getLiveOperationObjects)
|
2022-04-19 15:03:15 -07:00
|
|
|
.def("_clear_live_operations", &PyMlirContext::clearLiveOperations)
|
2024-02-21 11:01:00 +01:00
|
|
|
.def("_clear_live_operations_inside",
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::overload_cast<MlirOperation>(
|
2024-02-21 11:01:00 +01:00
|
|
|
&PyMlirContext::clearOperationsInside))
|
2021-03-19 11:57:01 -07:00
|
|
|
.def("_get_live_module_count", &PyMlirContext::getLiveModuleCount)
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro(MLIR_PYTHON_CAPI_PTR_ATTR, &PyMlirContext::getCapsule)
|
2021-03-19 11:57:01 -07:00
|
|
|
.def(MLIR_PYTHON_CAPI_FACTORY_ATTR, &PyMlirContext::createFromCapsule)
|
|
|
|
|
.def("__enter__", &PyMlirContext::contextEnter)
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def("__exit__", &PyMlirContext::contextExit, nb::arg("exc_type").none(),
|
|
|
|
|
nb::arg("exc_value").none(), nb::arg("traceback").none())
|
|
|
|
|
.def_prop_ro_static(
|
2021-03-19 11:57:01 -07:00
|
|
|
"current",
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
[](nb::object & /*class*/) {
|
2021-03-19 11:57:01 -07:00
|
|
|
auto *context = PyThreadContextEntry::getDefaultContext();
|
|
|
|
|
if (!context)
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
return nb::none();
|
|
|
|
|
return nb::cast(context);
|
2021-03-19 11:57:01 -07:00
|
|
|
},
|
|
|
|
|
"Gets the Context bound to the current thread or raises ValueError")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro(
|
2021-03-19 11:57:01 -07:00
|
|
|
"dialects",
|
|
|
|
|
[](PyMlirContext &self) { return PyDialects(self.getRef()); },
|
|
|
|
|
"Gets a container for accessing dialects by name")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro(
|
2021-03-19 11:57:01 -07:00
|
|
|
"d", [](PyMlirContext &self) { return PyDialects(self.getRef()); },
|
|
|
|
|
"Alias for 'dialect'")
|
|
|
|
|
.def(
|
|
|
|
|
"get_dialect_descriptor",
|
|
|
|
|
[=](PyMlirContext &self, std::string &name) {
|
|
|
|
|
MlirDialect dialect = mlirContextGetOrLoadDialect(
|
|
|
|
|
self.get(), {name.data(), name.size()});
|
|
|
|
|
if (mlirDialectIsNull(dialect)) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error(
|
|
|
|
|
(Twine("Dialect '") + name + "' not found").str().c_str());
|
2020-12-09 02:15:02 +00:00
|
|
|
}
|
2021-03-19 11:57:01 -07:00
|
|
|
return PyDialectDescriptor(self.getRef(), dialect);
|
|
|
|
|
},
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("dialect_name"),
|
2021-03-19 11:57:01 -07:00
|
|
|
"Gets or loads a dialect by name, returning its descriptor object")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_rw(
|
2021-03-19 11:57:01 -07:00
|
|
|
"allow_unregistered_dialects",
|
|
|
|
|
[](PyMlirContext &self) -> bool {
|
|
|
|
|
return mlirContextGetAllowUnregisteredDialects(self.get());
|
|
|
|
|
},
|
|
|
|
|
[](PyMlirContext &self, bool value) {
|
|
|
|
|
mlirContextSetAllowUnregisteredDialects(self.get(), value);
|
2021-03-30 22:19:10 -07:00
|
|
|
})
|
2022-01-03 16:39:58 -08:00
|
|
|
.def("attach_diagnostic_handler", &PyMlirContext::attachDiagnosticHandler,
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("callback"),
|
2022-01-03 16:39:58 -08:00
|
|
|
"Attaches a diagnostic handler that will receive callbacks")
|
2021-11-28 14:08:06 -08:00
|
|
|
.def(
|
|
|
|
|
"enable_multithreading",
|
|
|
|
|
[](PyMlirContext &self, bool enable) {
|
|
|
|
|
mlirContextEnableMultithreading(self.get(), enable);
|
|
|
|
|
},
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("enable"))
|
2021-11-28 14:08:06 -08:00
|
|
|
.def(
|
|
|
|
|
"is_registered_operation",
|
|
|
|
|
[](PyMlirContext &self, std::string &name) {
|
|
|
|
|
return mlirContextIsRegisteredOperation(
|
|
|
|
|
self.get(), MlirStringRef{name.data(), name.size()});
|
|
|
|
|
},
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("operation_name"))
|
[mlir] Overhaul C/Python registration APIs to properly scope registration/loading activities.
Since the very first commits, the Python and C MLIR APIs have had mis-placed registration/load functionality for dialects, extensions, etc. This was done pragmatically in order to get bootstrapped and then just grew in. Downstreams largely bypass and do their own thing by providing various APIs to register things they need. Meanwhile, the C++ APIs have stabilized around this and it would make sense to follow suit.
The thing we have observed in canonical usage by downstreams is that each downstream tends to have native entry points that configure its installation to its preferences with one-stop APIs. This patch leans in to this approach with `RegisterEverything.h` and `mlir._mlir_libs._mlirRegisterEverything` being the one-stop entry points for the "upstream packages". The `_mlir_libs.__init__.py` now allows customization of the environment and Context by adding "initialization modules" to the `_mlir_libs` package. If present, `_mlirRegisterEverything` is treated as such a module. Others can be added by downstreams by adding a `_site_initialize_{i}.py` module, where '{i}' is a number starting with zero. The number will be incremented and corresponding module loaded until one is not found. Initialization modules can:
* Perform load time customization to the global environment (i.e. registering passes, hooks, etc).
* Define a `register_dialects(registry: DialectRegistry)` function that can extend the `DialectRegistry` that will be used to bootstrap the `Context`.
* Define a `context_init_hook(context: Context)` function that will be added to a list of callbacks which will be invoked after dialect registration during `Context` initialization.
Note that the `MLIRPythonExtension.RegisterEverything` is not included by default when building a downstream (its corresponding behavior was prior). For downstreams which need the default MLIR initialization to take place, they must add this back in to their Python CMake build just like they add their own components (i.e. to `add_mlir_python_common_capi_library` and `add_mlir_python_modules`). It is perfectly valid to not do this, in which case, only the things explicitly depended on and initialized by downstreams will be built/packaged. If the downstream has not been set up for this, it is recommended to simply add this back for the time being and pay the build time/package size cost.
CMake changes:
* `MLIRCAPIRegistration` -> `MLIRCAPIRegisterEverything` (renamed to signify what it does and force an evaluation: a number of places were incidentally linking this very expensive target)
* `MLIRPythonSoure.Passes` removed (without replacement: just drop)
* `MLIRPythonExtension.AllPassesRegistration` removed (without replacement: just drop)
* `MLIRPythonExtension.Conversions` removed (without replacement: just drop)
* `MLIRPythonExtension.Transforms` removed (without replacement: just drop)
Header changes:
* `mlir-c/Registration.h` is deleted. Dialect registration functionality is now in `IR.h`. Registration of upstream features are in `mlir-c/RegisterEverything.h`. When updating MLIR and a couple of downstreams, I found that proper usage was commingled so required making a choice vs just blind S&R.
Python APIs removed:
* mlir.transforms and mlir.conversions (previously only had an __init__.py which indirectly triggered `mlirRegisterTransformsPasses()` and `mlirRegisterConversionPasses()` respectively). Downstream impact: Remove these imports if present (they now happen as part of default initialization).
* mlir._mlir_libs._all_passes_registration, mlir._mlir_libs._mlirTransforms, mlir._mlir_libs._mlirConversions. Downstream impact: None expected (these were internally used).
C-APIs changed:
* mlirRegisterAllDialects(MlirContext) now takes an MlirDialectRegistry instead. It also used to trigger loading of all dialects, which was already marked with a TODO to remove -- it no longer does, and for direct use, dialects must be explicitly loaded. Downstream impact: Direct C-API users must ensure that needed dialects are loaded or call `mlirContextLoadAllAvailableDialects(MlirContext)` to emulate the prior behavior. Also see the `ir.c` test case (e.g. ` mlirContextGetOrLoadDialect(ctx, mlirStringRefCreateFromCString("func"));`).
* mlirDialectHandle* APIs were moved from Registration.h (which now is restricted to just global/upstream registration) to IR.h, arguably where it should have been. Downstream impact: include correct header (likely already doing so).
C-APIs added:
* mlirContextLoadAllAvailableDialects(MlirContext): Corresponds to C++ API with the same purpose.
Python APIs added:
* mlir.ir.DialectRegistry: Mapping for an MlirDialectRegistry.
* mlir.ir.Context.append_dialect_registry(MlirDialectRegistry)
* mlir.ir.Context.load_all_available_dialects()
* mlir._mlir_libs._mlirAllRegistration: New native extension that exposes a `register_dialects(MlirDialectRegistry)` entry point and performs all upstream pass/conversion/transforms registration on init. In this first step, we eagerly load this as part of the __init__.py and use it to monkey patch the Context to emulate prior behavior.
* Type caster and capsule support for MlirDialectRegistry
This should make it possible to build downstream Python dialects that only depend on a subset of MLIR. See: https://github.com/llvm/llvm-project/issues/56037
Here is an example PR, minimally adapting IREE to these changes: https://github.com/iree-org/iree/pull/9638/files In this situation, IREE is opting to not link everything, since it is already configuring the Context to its liking. For projects that would just like to not think about it and pull in everything, add `MLIRPythonExtension.RegisterEverything` to the list of Python sources getting built, and the old behavior will continue.
Reviewed By: mehdi_amini, ftynse
Differential Revision: https://reviews.llvm.org/D128593
2022-07-16 16:09:03 -07:00
|
|
|
.def(
|
|
|
|
|
"append_dialect_registry",
|
|
|
|
|
[](PyMlirContext &self, PyDialectRegistry ®istry) {
|
|
|
|
|
mlirContextAppendDialectRegistry(self.get(), registry);
|
|
|
|
|
},
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("registry"))
|
|
|
|
|
.def_prop_rw("emit_error_diagnostics", nullptr,
|
|
|
|
|
&PyMlirContext::setEmitErrorDiagnostics,
|
|
|
|
|
"Emit error diagnostics to diagnostic handlers. By default "
|
|
|
|
|
"error diagnostics are captured and reported through "
|
|
|
|
|
"MLIRError exceptions.")
|
[mlir] Overhaul C/Python registration APIs to properly scope registration/loading activities.
Since the very first commits, the Python and C MLIR APIs have had mis-placed registration/load functionality for dialects, extensions, etc. This was done pragmatically in order to get bootstrapped and then just grew in. Downstreams largely bypass and do their own thing by providing various APIs to register things they need. Meanwhile, the C++ APIs have stabilized around this and it would make sense to follow suit.
The thing we have observed in canonical usage by downstreams is that each downstream tends to have native entry points that configure its installation to its preferences with one-stop APIs. This patch leans in to this approach with `RegisterEverything.h` and `mlir._mlir_libs._mlirRegisterEverything` being the one-stop entry points for the "upstream packages". The `_mlir_libs.__init__.py` now allows customization of the environment and Context by adding "initialization modules" to the `_mlir_libs` package. If present, `_mlirRegisterEverything` is treated as such a module. Others can be added by downstreams by adding a `_site_initialize_{i}.py` module, where '{i}' is a number starting with zero. The number will be incremented and corresponding module loaded until one is not found. Initialization modules can:
* Perform load time customization to the global environment (i.e. registering passes, hooks, etc).
* Define a `register_dialects(registry: DialectRegistry)` function that can extend the `DialectRegistry` that will be used to bootstrap the `Context`.
* Define a `context_init_hook(context: Context)` function that will be added to a list of callbacks which will be invoked after dialect registration during `Context` initialization.
Note that the `MLIRPythonExtension.RegisterEverything` is not included by default when building a downstream (its corresponding behavior was prior). For downstreams which need the default MLIR initialization to take place, they must add this back in to their Python CMake build just like they add their own components (i.e. to `add_mlir_python_common_capi_library` and `add_mlir_python_modules`). It is perfectly valid to not do this, in which case, only the things explicitly depended on and initialized by downstreams will be built/packaged. If the downstream has not been set up for this, it is recommended to simply add this back for the time being and pay the build time/package size cost.
CMake changes:
* `MLIRCAPIRegistration` -> `MLIRCAPIRegisterEverything` (renamed to signify what it does and force an evaluation: a number of places were incidentally linking this very expensive target)
* `MLIRPythonSoure.Passes` removed (without replacement: just drop)
* `MLIRPythonExtension.AllPassesRegistration` removed (without replacement: just drop)
* `MLIRPythonExtension.Conversions` removed (without replacement: just drop)
* `MLIRPythonExtension.Transforms` removed (without replacement: just drop)
Header changes:
* `mlir-c/Registration.h` is deleted. Dialect registration functionality is now in `IR.h`. Registration of upstream features are in `mlir-c/RegisterEverything.h`. When updating MLIR and a couple of downstreams, I found that proper usage was commingled so required making a choice vs just blind S&R.
Python APIs removed:
* mlir.transforms and mlir.conversions (previously only had an __init__.py which indirectly triggered `mlirRegisterTransformsPasses()` and `mlirRegisterConversionPasses()` respectively). Downstream impact: Remove these imports if present (they now happen as part of default initialization).
* mlir._mlir_libs._all_passes_registration, mlir._mlir_libs._mlirTransforms, mlir._mlir_libs._mlirConversions. Downstream impact: None expected (these were internally used).
C-APIs changed:
* mlirRegisterAllDialects(MlirContext) now takes an MlirDialectRegistry instead. It also used to trigger loading of all dialects, which was already marked with a TODO to remove -- it no longer does, and for direct use, dialects must be explicitly loaded. Downstream impact: Direct C-API users must ensure that needed dialects are loaded or call `mlirContextLoadAllAvailableDialects(MlirContext)` to emulate the prior behavior. Also see the `ir.c` test case (e.g. ` mlirContextGetOrLoadDialect(ctx, mlirStringRefCreateFromCString("func"));`).
* mlirDialectHandle* APIs were moved from Registration.h (which now is restricted to just global/upstream registration) to IR.h, arguably where it should have been. Downstream impact: include correct header (likely already doing so).
C-APIs added:
* mlirContextLoadAllAvailableDialects(MlirContext): Corresponds to C++ API with the same purpose.
Python APIs added:
* mlir.ir.DialectRegistry: Mapping for an MlirDialectRegistry.
* mlir.ir.Context.append_dialect_registry(MlirDialectRegistry)
* mlir.ir.Context.load_all_available_dialects()
* mlir._mlir_libs._mlirAllRegistration: New native extension that exposes a `register_dialects(MlirDialectRegistry)` entry point and performs all upstream pass/conversion/transforms registration on init. In this first step, we eagerly load this as part of the __init__.py and use it to monkey patch the Context to emulate prior behavior.
* Type caster and capsule support for MlirDialectRegistry
This should make it possible to build downstream Python dialects that only depend on a subset of MLIR. See: https://github.com/llvm/llvm-project/issues/56037
Here is an example PR, minimally adapting IREE to these changes: https://github.com/iree-org/iree/pull/9638/files In this situation, IREE is opting to not link everything, since it is already configuring the Context to its liking. For projects that would just like to not think about it and pull in everything, add `MLIRPythonExtension.RegisterEverything` to the list of Python sources getting built, and the old behavior will continue.
Reviewed By: mehdi_amini, ftynse
Differential Revision: https://reviews.llvm.org/D128593
2022-07-16 16:09:03 -07:00
|
|
|
.def("load_all_available_dialects", [](PyMlirContext &self) {
|
|
|
|
|
mlirContextLoadAllAvailableDialects(self.get());
|
|
|
|
|
});
|
2020-10-19 22:29:35 -07:00
|
|
|
|
2021-03-19 11:57:01 -07:00
|
|
|
//----------------------------------------------------------------------------
|
|
|
|
|
// Mapping of PyDialectDescriptor
|
|
|
|
|
//----------------------------------------------------------------------------
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::class_<PyDialectDescriptor>(m, "DialectDescriptor")
|
|
|
|
|
.def_prop_ro("namespace",
|
|
|
|
|
[](PyDialectDescriptor &self) {
|
|
|
|
|
MlirStringRef ns = mlirDialectGetNamespace(self.get());
|
|
|
|
|
return nb::str(ns.data, ns.length);
|
|
|
|
|
})
|
2021-03-19 11:57:01 -07:00
|
|
|
.def("__repr__", [](PyDialectDescriptor &self) {
|
|
|
|
|
MlirStringRef ns = mlirDialectGetNamespace(self.get());
|
|
|
|
|
std::string repr("<DialectDescriptor ");
|
|
|
|
|
repr.append(ns.data, ns.length);
|
|
|
|
|
repr.append(">");
|
|
|
|
|
return repr;
|
|
|
|
|
});
|
2020-11-08 09:28:35 +08:00
|
|
|
|
2021-03-19 11:57:01 -07:00
|
|
|
//----------------------------------------------------------------------------
|
|
|
|
|
// Mapping of PyDialects
|
|
|
|
|
//----------------------------------------------------------------------------
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::class_<PyDialects>(m, "Dialects")
|
2021-03-19 11:57:01 -07:00
|
|
|
.def("__getitem__",
|
|
|
|
|
[=](PyDialects &self, std::string keyName) {
|
|
|
|
|
MlirDialect dialect =
|
|
|
|
|
self.getDialectForKey(keyName, /*attrError=*/false);
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object descriptor =
|
|
|
|
|
nb::cast(PyDialectDescriptor{self.getContext(), dialect});
|
2021-03-19 11:57:01 -07:00
|
|
|
return createCustomDialectWrapper(keyName, std::move(descriptor));
|
|
|
|
|
})
|
|
|
|
|
.def("__getattr__", [=](PyDialects &self, std::string attrName) {
|
|
|
|
|
MlirDialect dialect =
|
|
|
|
|
self.getDialectForKey(attrName, /*attrError=*/true);
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object descriptor =
|
|
|
|
|
nb::cast(PyDialectDescriptor{self.getContext(), dialect});
|
2021-03-19 11:57:01 -07:00
|
|
|
return createCustomDialectWrapper(attrName, std::move(descriptor));
|
|
|
|
|
});
|
2020-11-08 09:28:35 +08:00
|
|
|
|
2021-03-19 11:57:01 -07:00
|
|
|
//----------------------------------------------------------------------------
|
|
|
|
|
// Mapping of PyDialect
|
|
|
|
|
//----------------------------------------------------------------------------
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::class_<PyDialect>(m, "Dialect")
|
|
|
|
|
.def(nb::init<nb::object>(), nb::arg("descriptor"))
|
|
|
|
|
.def_prop_ro("descriptor",
|
|
|
|
|
[](PyDialect &self) { return self.getDescriptor(); })
|
|
|
|
|
.def("__repr__", [](nb::object self) {
|
2021-03-19 11:57:01 -07:00
|
|
|
auto clazz = self.attr("__class__");
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
return nb::str("<Dialect ") +
|
|
|
|
|
self.attr("descriptor").attr("namespace") + nb::str(" (class ") +
|
|
|
|
|
clazz.attr("__module__") + nb::str(".") +
|
|
|
|
|
clazz.attr("__name__") + nb::str(")>");
|
2021-03-19 11:57:01 -07:00
|
|
|
});
|
2020-10-19 22:29:35 -07:00
|
|
|
|
[mlir] Overhaul C/Python registration APIs to properly scope registration/loading activities.
Since the very first commits, the Python and C MLIR APIs have had mis-placed registration/load functionality for dialects, extensions, etc. This was done pragmatically in order to get bootstrapped and then just grew in. Downstreams largely bypass and do their own thing by providing various APIs to register things they need. Meanwhile, the C++ APIs have stabilized around this and it would make sense to follow suit.
The thing we have observed in canonical usage by downstreams is that each downstream tends to have native entry points that configure its installation to its preferences with one-stop APIs. This patch leans in to this approach with `RegisterEverything.h` and `mlir._mlir_libs._mlirRegisterEverything` being the one-stop entry points for the "upstream packages". The `_mlir_libs.__init__.py` now allows customization of the environment and Context by adding "initialization modules" to the `_mlir_libs` package. If present, `_mlirRegisterEverything` is treated as such a module. Others can be added by downstreams by adding a `_site_initialize_{i}.py` module, where '{i}' is a number starting with zero. The number will be incremented and corresponding module loaded until one is not found. Initialization modules can:
* Perform load time customization to the global environment (i.e. registering passes, hooks, etc).
* Define a `register_dialects(registry: DialectRegistry)` function that can extend the `DialectRegistry` that will be used to bootstrap the `Context`.
* Define a `context_init_hook(context: Context)` function that will be added to a list of callbacks which will be invoked after dialect registration during `Context` initialization.
Note that the `MLIRPythonExtension.RegisterEverything` is not included by default when building a downstream (its corresponding behavior was prior). For downstreams which need the default MLIR initialization to take place, they must add this back in to their Python CMake build just like they add their own components (i.e. to `add_mlir_python_common_capi_library` and `add_mlir_python_modules`). It is perfectly valid to not do this, in which case, only the things explicitly depended on and initialized by downstreams will be built/packaged. If the downstream has not been set up for this, it is recommended to simply add this back for the time being and pay the build time/package size cost.
CMake changes:
* `MLIRCAPIRegistration` -> `MLIRCAPIRegisterEverything` (renamed to signify what it does and force an evaluation: a number of places were incidentally linking this very expensive target)
* `MLIRPythonSoure.Passes` removed (without replacement: just drop)
* `MLIRPythonExtension.AllPassesRegistration` removed (without replacement: just drop)
* `MLIRPythonExtension.Conversions` removed (without replacement: just drop)
* `MLIRPythonExtension.Transforms` removed (without replacement: just drop)
Header changes:
* `mlir-c/Registration.h` is deleted. Dialect registration functionality is now in `IR.h`. Registration of upstream features are in `mlir-c/RegisterEverything.h`. When updating MLIR and a couple of downstreams, I found that proper usage was commingled so required making a choice vs just blind S&R.
Python APIs removed:
* mlir.transforms and mlir.conversions (previously only had an __init__.py which indirectly triggered `mlirRegisterTransformsPasses()` and `mlirRegisterConversionPasses()` respectively). Downstream impact: Remove these imports if present (they now happen as part of default initialization).
* mlir._mlir_libs._all_passes_registration, mlir._mlir_libs._mlirTransforms, mlir._mlir_libs._mlirConversions. Downstream impact: None expected (these were internally used).
C-APIs changed:
* mlirRegisterAllDialects(MlirContext) now takes an MlirDialectRegistry instead. It also used to trigger loading of all dialects, which was already marked with a TODO to remove -- it no longer does, and for direct use, dialects must be explicitly loaded. Downstream impact: Direct C-API users must ensure that needed dialects are loaded or call `mlirContextLoadAllAvailableDialects(MlirContext)` to emulate the prior behavior. Also see the `ir.c` test case (e.g. ` mlirContextGetOrLoadDialect(ctx, mlirStringRefCreateFromCString("func"));`).
* mlirDialectHandle* APIs were moved from Registration.h (which now is restricted to just global/upstream registration) to IR.h, arguably where it should have been. Downstream impact: include correct header (likely already doing so).
C-APIs added:
* mlirContextLoadAllAvailableDialects(MlirContext): Corresponds to C++ API with the same purpose.
Python APIs added:
* mlir.ir.DialectRegistry: Mapping for an MlirDialectRegistry.
* mlir.ir.Context.append_dialect_registry(MlirDialectRegistry)
* mlir.ir.Context.load_all_available_dialects()
* mlir._mlir_libs._mlirAllRegistration: New native extension that exposes a `register_dialects(MlirDialectRegistry)` entry point and performs all upstream pass/conversion/transforms registration on init. In this first step, we eagerly load this as part of the __init__.py and use it to monkey patch the Context to emulate prior behavior.
* Type caster and capsule support for MlirDialectRegistry
This should make it possible to build downstream Python dialects that only depend on a subset of MLIR. See: https://github.com/llvm/llvm-project/issues/56037
Here is an example PR, minimally adapting IREE to these changes: https://github.com/iree-org/iree/pull/9638/files In this situation, IREE is opting to not link everything, since it is already configuring the Context to its liking. For projects that would just like to not think about it and pull in everything, add `MLIRPythonExtension.RegisterEverything` to the list of Python sources getting built, and the old behavior will continue.
Reviewed By: mehdi_amini, ftynse
Differential Revision: https://reviews.llvm.org/D128593
2022-07-16 16:09:03 -07:00
|
|
|
//----------------------------------------------------------------------------
|
|
|
|
|
// Mapping of PyDialectRegistry
|
|
|
|
|
//----------------------------------------------------------------------------
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::class_<PyDialectRegistry>(m, "DialectRegistry")
|
|
|
|
|
.def_prop_ro(MLIR_PYTHON_CAPI_PTR_ATTR, &PyDialectRegistry::getCapsule)
|
[mlir] Overhaul C/Python registration APIs to properly scope registration/loading activities.
Since the very first commits, the Python and C MLIR APIs have had mis-placed registration/load functionality for dialects, extensions, etc. This was done pragmatically in order to get bootstrapped and then just grew in. Downstreams largely bypass and do their own thing by providing various APIs to register things they need. Meanwhile, the C++ APIs have stabilized around this and it would make sense to follow suit.
The thing we have observed in canonical usage by downstreams is that each downstream tends to have native entry points that configure its installation to its preferences with one-stop APIs. This patch leans in to this approach with `RegisterEverything.h` and `mlir._mlir_libs._mlirRegisterEverything` being the one-stop entry points for the "upstream packages". The `_mlir_libs.__init__.py` now allows customization of the environment and Context by adding "initialization modules" to the `_mlir_libs` package. If present, `_mlirRegisterEverything` is treated as such a module. Others can be added by downstreams by adding a `_site_initialize_{i}.py` module, where '{i}' is a number starting with zero. The number will be incremented and corresponding module loaded until one is not found. Initialization modules can:
* Perform load time customization to the global environment (i.e. registering passes, hooks, etc).
* Define a `register_dialects(registry: DialectRegistry)` function that can extend the `DialectRegistry` that will be used to bootstrap the `Context`.
* Define a `context_init_hook(context: Context)` function that will be added to a list of callbacks which will be invoked after dialect registration during `Context` initialization.
Note that the `MLIRPythonExtension.RegisterEverything` is not included by default when building a downstream (its corresponding behavior was prior). For downstreams which need the default MLIR initialization to take place, they must add this back in to their Python CMake build just like they add their own components (i.e. to `add_mlir_python_common_capi_library` and `add_mlir_python_modules`). It is perfectly valid to not do this, in which case, only the things explicitly depended on and initialized by downstreams will be built/packaged. If the downstream has not been set up for this, it is recommended to simply add this back for the time being and pay the build time/package size cost.
CMake changes:
* `MLIRCAPIRegistration` -> `MLIRCAPIRegisterEverything` (renamed to signify what it does and force an evaluation: a number of places were incidentally linking this very expensive target)
* `MLIRPythonSoure.Passes` removed (without replacement: just drop)
* `MLIRPythonExtension.AllPassesRegistration` removed (without replacement: just drop)
* `MLIRPythonExtension.Conversions` removed (without replacement: just drop)
* `MLIRPythonExtension.Transforms` removed (without replacement: just drop)
Header changes:
* `mlir-c/Registration.h` is deleted. Dialect registration functionality is now in `IR.h`. Registration of upstream features are in `mlir-c/RegisterEverything.h`. When updating MLIR and a couple of downstreams, I found that proper usage was commingled so required making a choice vs just blind S&R.
Python APIs removed:
* mlir.transforms and mlir.conversions (previously only had an __init__.py which indirectly triggered `mlirRegisterTransformsPasses()` and `mlirRegisterConversionPasses()` respectively). Downstream impact: Remove these imports if present (they now happen as part of default initialization).
* mlir._mlir_libs._all_passes_registration, mlir._mlir_libs._mlirTransforms, mlir._mlir_libs._mlirConversions. Downstream impact: None expected (these were internally used).
C-APIs changed:
* mlirRegisterAllDialects(MlirContext) now takes an MlirDialectRegistry instead. It also used to trigger loading of all dialects, which was already marked with a TODO to remove -- it no longer does, and for direct use, dialects must be explicitly loaded. Downstream impact: Direct C-API users must ensure that needed dialects are loaded or call `mlirContextLoadAllAvailableDialects(MlirContext)` to emulate the prior behavior. Also see the `ir.c` test case (e.g. ` mlirContextGetOrLoadDialect(ctx, mlirStringRefCreateFromCString("func"));`).
* mlirDialectHandle* APIs were moved from Registration.h (which now is restricted to just global/upstream registration) to IR.h, arguably where it should have been. Downstream impact: include correct header (likely already doing so).
C-APIs added:
* mlirContextLoadAllAvailableDialects(MlirContext): Corresponds to C++ API with the same purpose.
Python APIs added:
* mlir.ir.DialectRegistry: Mapping for an MlirDialectRegistry.
* mlir.ir.Context.append_dialect_registry(MlirDialectRegistry)
* mlir.ir.Context.load_all_available_dialects()
* mlir._mlir_libs._mlirAllRegistration: New native extension that exposes a `register_dialects(MlirDialectRegistry)` entry point and performs all upstream pass/conversion/transforms registration on init. In this first step, we eagerly load this as part of the __init__.py and use it to monkey patch the Context to emulate prior behavior.
* Type caster and capsule support for MlirDialectRegistry
This should make it possible to build downstream Python dialects that only depend on a subset of MLIR. See: https://github.com/llvm/llvm-project/issues/56037
Here is an example PR, minimally adapting IREE to these changes: https://github.com/iree-org/iree/pull/9638/files In this situation, IREE is opting to not link everything, since it is already configuring the Context to its liking. For projects that would just like to not think about it and pull in everything, add `MLIRPythonExtension.RegisterEverything` to the list of Python sources getting built, and the old behavior will continue.
Reviewed By: mehdi_amini, ftynse
Differential Revision: https://reviews.llvm.org/D128593
2022-07-16 16:09:03 -07:00
|
|
|
.def(MLIR_PYTHON_CAPI_FACTORY_ATTR, &PyDialectRegistry::createFromCapsule)
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def(nb::init<>());
|
[mlir] Overhaul C/Python registration APIs to properly scope registration/loading activities.
Since the very first commits, the Python and C MLIR APIs have had mis-placed registration/load functionality for dialects, extensions, etc. This was done pragmatically in order to get bootstrapped and then just grew in. Downstreams largely bypass and do their own thing by providing various APIs to register things they need. Meanwhile, the C++ APIs have stabilized around this and it would make sense to follow suit.
The thing we have observed in canonical usage by downstreams is that each downstream tends to have native entry points that configure its installation to its preferences with one-stop APIs. This patch leans in to this approach with `RegisterEverything.h` and `mlir._mlir_libs._mlirRegisterEverything` being the one-stop entry points for the "upstream packages". The `_mlir_libs.__init__.py` now allows customization of the environment and Context by adding "initialization modules" to the `_mlir_libs` package. If present, `_mlirRegisterEverything` is treated as such a module. Others can be added by downstreams by adding a `_site_initialize_{i}.py` module, where '{i}' is a number starting with zero. The number will be incremented and corresponding module loaded until one is not found. Initialization modules can:
* Perform load time customization to the global environment (i.e. registering passes, hooks, etc).
* Define a `register_dialects(registry: DialectRegistry)` function that can extend the `DialectRegistry` that will be used to bootstrap the `Context`.
* Define a `context_init_hook(context: Context)` function that will be added to a list of callbacks which will be invoked after dialect registration during `Context` initialization.
Note that the `MLIRPythonExtension.RegisterEverything` is not included by default when building a downstream (its corresponding behavior was prior). For downstreams which need the default MLIR initialization to take place, they must add this back in to their Python CMake build just like they add their own components (i.e. to `add_mlir_python_common_capi_library` and `add_mlir_python_modules`). It is perfectly valid to not do this, in which case, only the things explicitly depended on and initialized by downstreams will be built/packaged. If the downstream has not been set up for this, it is recommended to simply add this back for the time being and pay the build time/package size cost.
CMake changes:
* `MLIRCAPIRegistration` -> `MLIRCAPIRegisterEverything` (renamed to signify what it does and force an evaluation: a number of places were incidentally linking this very expensive target)
* `MLIRPythonSoure.Passes` removed (without replacement: just drop)
* `MLIRPythonExtension.AllPassesRegistration` removed (without replacement: just drop)
* `MLIRPythonExtension.Conversions` removed (without replacement: just drop)
* `MLIRPythonExtension.Transforms` removed (without replacement: just drop)
Header changes:
* `mlir-c/Registration.h` is deleted. Dialect registration functionality is now in `IR.h`. Registration of upstream features are in `mlir-c/RegisterEverything.h`. When updating MLIR and a couple of downstreams, I found that proper usage was commingled so required making a choice vs just blind S&R.
Python APIs removed:
* mlir.transforms and mlir.conversions (previously only had an __init__.py which indirectly triggered `mlirRegisterTransformsPasses()` and `mlirRegisterConversionPasses()` respectively). Downstream impact: Remove these imports if present (they now happen as part of default initialization).
* mlir._mlir_libs._all_passes_registration, mlir._mlir_libs._mlirTransforms, mlir._mlir_libs._mlirConversions. Downstream impact: None expected (these were internally used).
C-APIs changed:
* mlirRegisterAllDialects(MlirContext) now takes an MlirDialectRegistry instead. It also used to trigger loading of all dialects, which was already marked with a TODO to remove -- it no longer does, and for direct use, dialects must be explicitly loaded. Downstream impact: Direct C-API users must ensure that needed dialects are loaded or call `mlirContextLoadAllAvailableDialects(MlirContext)` to emulate the prior behavior. Also see the `ir.c` test case (e.g. ` mlirContextGetOrLoadDialect(ctx, mlirStringRefCreateFromCString("func"));`).
* mlirDialectHandle* APIs were moved from Registration.h (which now is restricted to just global/upstream registration) to IR.h, arguably where it should have been. Downstream impact: include correct header (likely already doing so).
C-APIs added:
* mlirContextLoadAllAvailableDialects(MlirContext): Corresponds to C++ API with the same purpose.
Python APIs added:
* mlir.ir.DialectRegistry: Mapping for an MlirDialectRegistry.
* mlir.ir.Context.append_dialect_registry(MlirDialectRegistry)
* mlir.ir.Context.load_all_available_dialects()
* mlir._mlir_libs._mlirAllRegistration: New native extension that exposes a `register_dialects(MlirDialectRegistry)` entry point and performs all upstream pass/conversion/transforms registration on init. In this first step, we eagerly load this as part of the __init__.py and use it to monkey patch the Context to emulate prior behavior.
* Type caster and capsule support for MlirDialectRegistry
This should make it possible to build downstream Python dialects that only depend on a subset of MLIR. See: https://github.com/llvm/llvm-project/issues/56037
Here is an example PR, minimally adapting IREE to these changes: https://github.com/iree-org/iree/pull/9638/files In this situation, IREE is opting to not link everything, since it is already configuring the Context to its liking. For projects that would just like to not think about it and pull in everything, add `MLIRPythonExtension.RegisterEverything` to the list of Python sources getting built, and the old behavior will continue.
Reviewed By: mehdi_amini, ftynse
Differential Revision: https://reviews.llvm.org/D128593
2022-07-16 16:09:03 -07:00
|
|
|
|
2021-03-19 11:57:01 -07:00
|
|
|
//----------------------------------------------------------------------------
|
|
|
|
|
// Mapping of Location
|
|
|
|
|
//----------------------------------------------------------------------------
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::class_<PyLocation>(m, "Location")
|
|
|
|
|
.def_prop_ro(MLIR_PYTHON_CAPI_PTR_ATTR, &PyLocation::getCapsule)
|
2021-03-19 11:57:01 -07:00
|
|
|
.def(MLIR_PYTHON_CAPI_FACTORY_ATTR, &PyLocation::createFromCapsule)
|
|
|
|
|
.def("__enter__", &PyLocation::contextEnter)
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def("__exit__", &PyLocation::contextExit, nb::arg("exc_type").none(),
|
|
|
|
|
nb::arg("exc_value").none(), nb::arg("traceback").none())
|
2021-03-19 11:57:01 -07:00
|
|
|
.def("__eq__",
|
|
|
|
|
[](PyLocation &self, PyLocation &other) -> bool {
|
|
|
|
|
return mlirLocationEqual(self, other);
|
|
|
|
|
})
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def("__eq__", [](PyLocation &self, nb::object other) { return false; })
|
|
|
|
|
.def_prop_ro_static(
|
2021-03-19 11:57:01 -07:00
|
|
|
"current",
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
[](nb::object & /*class*/) {
|
2021-03-19 11:57:01 -07:00
|
|
|
auto *loc = PyThreadContextEntry::getDefaultLocation();
|
|
|
|
|
if (!loc)
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error("No current Location");
|
2021-03-19 11:57:01 -07:00
|
|
|
return loc;
|
|
|
|
|
},
|
|
|
|
|
"Gets the Location bound to the current thread or raises ValueError")
|
|
|
|
|
.def_static(
|
|
|
|
|
"unknown",
|
|
|
|
|
[](DefaultingPyMlirContext context) {
|
|
|
|
|
return PyLocation(context->getRef(),
|
|
|
|
|
mlirLocationUnknownGet(context->get()));
|
|
|
|
|
},
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("context").none() = nb::none(),
|
2021-03-19 11:57:01 -07:00
|
|
|
"Gets a Location representing an unknown location")
|
2021-10-13 10:02:12 +02:00
|
|
|
.def_static(
|
|
|
|
|
"callsite",
|
|
|
|
|
[](PyLocation callee, const std::vector<PyLocation> &frames,
|
|
|
|
|
DefaultingPyMlirContext context) {
|
|
|
|
|
if (frames.empty())
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error("No caller frames provided");
|
2021-10-13 10:02:12 +02:00
|
|
|
MlirLocation caller = frames.back().get();
|
2021-10-19 17:13:54 +00:00
|
|
|
for (const PyLocation &frame :
|
2023-01-09 18:11:07 +01:00
|
|
|
llvm::reverse(llvm::ArrayRef(frames).drop_back()))
|
2021-10-13 10:02:12 +02:00
|
|
|
caller = mlirLocationCallSiteGet(frame.get(), caller);
|
|
|
|
|
return PyLocation(context->getRef(),
|
|
|
|
|
mlirLocationCallSiteGet(callee.get(), caller));
|
|
|
|
|
},
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("callee"), nb::arg("frames"),
|
|
|
|
|
nb::arg("context").none() = nb::none(),
|
2021-10-13 10:02:12 +02:00
|
|
|
kContextGetCallSiteLocationDocstring)
|
2021-03-19 11:57:01 -07:00
|
|
|
.def_static(
|
|
|
|
|
"file",
|
|
|
|
|
[](std::string filename, int line, int col,
|
|
|
|
|
DefaultingPyMlirContext context) {
|
|
|
|
|
return PyLocation(
|
|
|
|
|
context->getRef(),
|
|
|
|
|
mlirLocationFileLineColGet(
|
|
|
|
|
context->get(), toMlirStringRef(filename), line, col));
|
|
|
|
|
},
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("filename"), nb::arg("line"), nb::arg("col"),
|
|
|
|
|
nb::arg("context").none() = nb::none(),
|
|
|
|
|
kContextGetFileLocationDocstring)
|
2021-12-11 10:12:29 -08:00
|
|
|
.def_static(
|
|
|
|
|
"fused",
|
2022-01-03 16:39:58 -08:00
|
|
|
[](const std::vector<PyLocation> &pyLocations,
|
2023-01-14 01:25:58 -08:00
|
|
|
std::optional<PyAttribute> metadata,
|
2021-12-11 10:12:29 -08:00
|
|
|
DefaultingPyMlirContext context) {
|
|
|
|
|
llvm::SmallVector<MlirLocation, 4> locations;
|
|
|
|
|
locations.reserve(pyLocations.size());
|
|
|
|
|
for (auto &pyLocation : pyLocations)
|
|
|
|
|
locations.push_back(pyLocation.get());
|
|
|
|
|
MlirLocation location = mlirLocationFusedGet(
|
|
|
|
|
context->get(), locations.size(), locations.data(),
|
|
|
|
|
metadata ? metadata->get() : MlirAttribute{0});
|
|
|
|
|
return PyLocation(context->getRef(), location);
|
|
|
|
|
},
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("locations"), nb::arg("metadata").none() = nb::none(),
|
|
|
|
|
nb::arg("context").none() = nb::none(),
|
|
|
|
|
kContextGetFusedLocationDocstring)
|
2021-10-12 12:45:57 -07:00
|
|
|
.def_static(
|
|
|
|
|
"name",
|
2023-01-14 01:25:58 -08:00
|
|
|
[](std::string name, std::optional<PyLocation> childLoc,
|
2021-10-12 12:45:57 -07:00
|
|
|
DefaultingPyMlirContext context) {
|
|
|
|
|
return PyLocation(
|
|
|
|
|
context->getRef(),
|
|
|
|
|
mlirLocationNameGet(
|
|
|
|
|
context->get(), toMlirStringRef(name),
|
|
|
|
|
childLoc ? childLoc->get()
|
|
|
|
|
: mlirLocationUnknownGet(context->get())));
|
|
|
|
|
},
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("name"), nb::arg("childLoc").none() = nb::none(),
|
|
|
|
|
nb::arg("context").none() = nb::none(),
|
|
|
|
|
kContextGetNameLocationDocString)
|
2023-01-19 15:01:46 -08:00
|
|
|
.def_static(
|
|
|
|
|
"from_attr",
|
|
|
|
|
[](PyAttribute &attribute, DefaultingPyMlirContext context) {
|
|
|
|
|
return PyLocation(context->getRef(),
|
|
|
|
|
mlirLocationFromAttribute(attribute));
|
|
|
|
|
},
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("attribute"), nb::arg("context").none() = nb::none(),
|
2023-01-19 15:01:46 -08:00
|
|
|
"Gets a Location from a LocationAttr")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro(
|
2021-03-19 11:57:01 -07:00
|
|
|
"context",
|
|
|
|
|
[](PyLocation &self) { return self.getContext().getObject(); },
|
|
|
|
|
"Context that owns the Location")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro(
|
2023-01-19 15:01:46 -08:00
|
|
|
"attr",
|
2023-05-31 15:52:46 -05:00
|
|
|
[](PyLocation &self) { return mlirLocationGetAttribute(self); },
|
2023-01-19 15:01:46 -08:00
|
|
|
"Get the underlying LocationAttr")
|
2022-01-03 16:39:58 -08:00
|
|
|
.def(
|
|
|
|
|
"emit_error",
|
|
|
|
|
[](PyLocation &self, std::string message) {
|
|
|
|
|
mlirEmitError(self, message.c_str());
|
|
|
|
|
},
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("message"), "Emits an error at this location")
|
2021-03-19 11:57:01 -07:00
|
|
|
.def("__repr__", [](PyLocation &self) {
|
|
|
|
|
PyPrintAccumulator printAccum;
|
|
|
|
|
mlirLocationPrint(self, printAccum.getCallback(),
|
|
|
|
|
printAccum.getUserData());
|
|
|
|
|
return printAccum.join();
|
|
|
|
|
});
|
2020-08-16 18:49:28 -07:00
|
|
|
|
2020-10-21 23:34:01 -07:00
|
|
|
//----------------------------------------------------------------------------
|
2020-08-18 17:23:46 -07:00
|
|
|
// Mapping of Module
|
2020-10-21 23:34:01 -07:00
|
|
|
//----------------------------------------------------------------------------
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::class_<PyModule>(m, "Module", nb::is_weak_referenceable())
|
|
|
|
|
.def_prop_ro(MLIR_PYTHON_CAPI_PTR_ATTR, &PyModule::getCapsule)
|
2020-10-12 21:19:13 -07:00
|
|
|
.def(MLIR_PYTHON_CAPI_FACTORY_ATTR, &PyModule::createFromCapsule)
|
2020-10-31 23:40:25 -07:00
|
|
|
.def_static(
|
|
|
|
|
"parse",
|
[mlir][python] Capture error diagnostics in exceptions
This updates most (all?) error-diagnostic-emitting python APIs to
capture error diagnostics and include them in the raised exception's
message:
```
>>> Operation.parse('"arith.addi"() : () -> ()'))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
mlir._mlir_libs.MLIRError: Unable to parse operation assembly:
error: "-":1:1: 'arith.addi' op requires one result
note: "-":1:1: see current operation: "arith.addi"() : () -> ()
```
The diagnostic information is available on the exception for users who
may want to customize the error message:
```
>>> try:
... Operation.parse('"arith.addi"() : () -> ()')
... except MLIRError as e:
... print(e.message)
... print(e.error_diagnostics)
... print(e.error_diagnostics[0].message)
...
Unable to parse operation assembly
[<mlir._mlir_libs._mlir.ir.DiagnosticInfo object at 0x7fed32bd6b70>]
'arith.addi' op requires one result
```
Error diagnostics captured in exceptions aren't propagated to diagnostic
handlers, to avoid double-reporting of errors. The context-level
`emit_error_diagnostics` option can be used to revert to the old
behaviour, causing error diagnostics to be reported to handlers instead
of as part of exceptions.
API changes:
- `Operation.verify` now raises an exception on verification failure,
instead of returning `false`
- The exception raised by the following methods has been changed to
`MLIRError`:
- `PassManager.run`
- `{Module,Operation,Type,Attribute}.parse`
- `{RankedTensorType,UnrankedTensorType}.get`
- `{MemRefType,UnrankedMemRefType}.get`
- `VectorType.get`
- `FloatAttr.get`
closes #60595
depends on D144804, D143830
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D143869
2023-02-07 16:07:50 -05:00
|
|
|
[](const std::string &moduleAsm, DefaultingPyMlirContext context) {
|
|
|
|
|
PyMlirContext::ErrorCapture errors(context->getRef());
|
2020-11-24 18:35:22 +00:00
|
|
|
MlirModule module = mlirModuleCreateParse(
|
|
|
|
|
context->get(), toMlirStringRef(moduleAsm));
|
[mlir][python] Capture error diagnostics in exceptions
This updates most (all?) error-diagnostic-emitting python APIs to
capture error diagnostics and include them in the raised exception's
message:
```
>>> Operation.parse('"arith.addi"() : () -> ()'))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
mlir._mlir_libs.MLIRError: Unable to parse operation assembly:
error: "-":1:1: 'arith.addi' op requires one result
note: "-":1:1: see current operation: "arith.addi"() : () -> ()
```
The diagnostic information is available on the exception for users who
may want to customize the error message:
```
>>> try:
... Operation.parse('"arith.addi"() : () -> ()')
... except MLIRError as e:
... print(e.message)
... print(e.error_diagnostics)
... print(e.error_diagnostics[0].message)
...
Unable to parse operation assembly
[<mlir._mlir_libs._mlir.ir.DiagnosticInfo object at 0x7fed32bd6b70>]
'arith.addi' op requires one result
```
Error diagnostics captured in exceptions aren't propagated to diagnostic
handlers, to avoid double-reporting of errors. The context-level
`emit_error_diagnostics` option can be used to revert to the old
behaviour, causing error diagnostics to be reported to handlers instead
of as part of exceptions.
API changes:
- `Operation.verify` now raises an exception on verification failure,
instead of returning `false`
- The exception raised by the following methods has been changed to
`MLIRError`:
- `PassManager.run`
- `{Module,Operation,Type,Attribute}.parse`
- `{RankedTensorType,UnrankedTensorType}.get`
- `{MemRefType,UnrankedMemRefType}.get`
- `VectorType.get`
- `FloatAttr.get`
closes #60595
depends on D144804, D143830
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D143869
2023-02-07 16:07:50 -05:00
|
|
|
if (mlirModuleIsNull(module))
|
|
|
|
|
throw MLIRError("Unable to parse module assembly", errors.take());
|
2020-10-31 23:40:25 -07:00
|
|
|
return PyModule::forModule(module).releaseObject();
|
|
|
|
|
},
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("asm"), nb::arg("context").none() = nb::none(),
|
|
|
|
|
kModuleParseDocstring)
|
|
|
|
|
.def_static(
|
|
|
|
|
"parse",
|
|
|
|
|
[](nb::bytes moduleAsm, DefaultingPyMlirContext context) {
|
|
|
|
|
PyMlirContext::ErrorCapture errors(context->getRef());
|
|
|
|
|
MlirModule module = mlirModuleCreateParse(
|
|
|
|
|
context->get(), toMlirStringRef(moduleAsm));
|
|
|
|
|
if (mlirModuleIsNull(module))
|
|
|
|
|
throw MLIRError("Unable to parse module assembly", errors.take());
|
|
|
|
|
return PyModule::forModule(module).releaseObject();
|
|
|
|
|
},
|
|
|
|
|
nb::arg("asm"), nb::arg("context").none() = nb::none(),
|
2020-10-31 23:40:25 -07:00
|
|
|
kModuleParseDocstring)
|
|
|
|
|
.def_static(
|
|
|
|
|
"create",
|
|
|
|
|
[](DefaultingPyLocation loc) {
|
2020-11-29 13:30:23 -08:00
|
|
|
MlirModule module = mlirModuleCreateEmpty(loc);
|
2020-10-31 23:40:25 -07:00
|
|
|
return PyModule::forModule(module).releaseObject();
|
|
|
|
|
},
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("loc").none() = nb::none(), "Creates an empty module")
|
|
|
|
|
.def_prop_ro(
|
2020-10-12 21:19:13 -07:00
|
|
|
"context",
|
|
|
|
|
[](PyModule &self) { return self.getContext().getObject(); },
|
|
|
|
|
"Context that created the Module")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro(
|
2020-09-18 18:38:21 -07:00
|
|
|
"operation",
|
|
|
|
|
[](PyModule &self) {
|
|
|
|
|
return PyOperation::forOperation(self.getContext(),
|
|
|
|
|
mlirModuleGetOperation(self.get()),
|
|
|
|
|
self.getRef().releaseObject())
|
|
|
|
|
.releaseObject();
|
|
|
|
|
},
|
|
|
|
|
"Accesses the module as an operation")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro(
|
2020-10-28 05:57:17 +00:00
|
|
|
"body",
|
|
|
|
|
[](PyModule &self) {
|
2021-12-20 19:45:05 +00:00
|
|
|
PyOperationRef moduleOp = PyOperation::forOperation(
|
2020-10-28 05:57:17 +00:00
|
|
|
self.getContext(), mlirModuleGetOperation(self.get()),
|
|
|
|
|
self.getRef().releaseObject());
|
2021-12-20 19:45:05 +00:00
|
|
|
PyBlock returnBlock(moduleOp, mlirModuleGetBody(self.get()));
|
2020-10-28 05:57:17 +00:00
|
|
|
return returnBlock;
|
|
|
|
|
},
|
|
|
|
|
"Return the block for this module")
|
2020-08-18 17:23:46 -07:00
|
|
|
.def(
|
|
|
|
|
"dump",
|
|
|
|
|
[](PyModule &self) {
|
2020-09-18 18:38:21 -07:00
|
|
|
mlirOperationDump(mlirModuleGetOperation(self.get()));
|
2020-08-18 17:23:46 -07:00
|
|
|
},
|
|
|
|
|
kDumpDocstring)
|
2020-08-16 20:53:45 -07:00
|
|
|
.def(
|
|
|
|
|
"__str__",
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
[](nb::object self) {
|
2021-11-28 15:33:03 -08:00
|
|
|
// Defer to the operation's __str__.
|
|
|
|
|
return self.attr("operation").attr("__str__")();
|
2020-08-16 20:53:45 -07:00
|
|
|
},
|
|
|
|
|
kOperationStrDunderDocstring);
|
2020-08-18 17:23:46 -07:00
|
|
|
|
2020-10-21 23:34:01 -07:00
|
|
|
//----------------------------------------------------------------------------
|
2020-09-18 18:38:21 -07:00
|
|
|
// Mapping of Operation.
|
2020-10-21 23:34:01 -07:00
|
|
|
//----------------------------------------------------------------------------
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::class_<PyOperationBase>(m, "_OperationBase")
|
|
|
|
|
.def_prop_ro(MLIR_PYTHON_CAPI_PTR_ATTR,
|
|
|
|
|
[](PyOperationBase &self) {
|
|
|
|
|
return self.getOperation().getCapsule();
|
|
|
|
|
})
|
2020-11-01 23:05:36 -08:00
|
|
|
.def("__eq__",
|
|
|
|
|
[](PyOperationBase &self, PyOperationBase &other) {
|
|
|
|
|
return &self.getOperation() == &other.getOperation();
|
|
|
|
|
})
|
|
|
|
|
.def("__eq__",
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
[](PyOperationBase &self, nb::object other) { return false; })
|
2021-11-02 17:04:42 +01:00
|
|
|
.def("__hash__",
|
|
|
|
|
[](PyOperationBase &self) {
|
|
|
|
|
return static_cast<size_t>(llvm::hash_value(&self.getOperation()));
|
|
|
|
|
})
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro("attributes",
|
|
|
|
|
[](PyOperationBase &self) {
|
|
|
|
|
return PyOpAttributeMap(self.getOperation().getRef());
|
|
|
|
|
})
|
|
|
|
|
.def_prop_ro(
|
2023-09-18 21:30:41 -07:00
|
|
|
"context",
|
|
|
|
|
[](PyOperationBase &self) {
|
|
|
|
|
PyOperation &concreteOperation = self.getOperation();
|
|
|
|
|
concreteOperation.checkValid();
|
|
|
|
|
return concreteOperation.getContext().getObject();
|
|
|
|
|
},
|
|
|
|
|
"Context that owns the Operation")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro("name",
|
|
|
|
|
[](PyOperationBase &self) {
|
|
|
|
|
auto &concreteOperation = self.getOperation();
|
|
|
|
|
concreteOperation.checkValid();
|
|
|
|
|
MlirOperation operation = concreteOperation.get();
|
|
|
|
|
MlirStringRef name =
|
|
|
|
|
mlirIdentifierStr(mlirOperationGetName(operation));
|
|
|
|
|
return nb::str(name.data, name.length);
|
|
|
|
|
})
|
|
|
|
|
.def_prop_ro("operands",
|
|
|
|
|
[](PyOperationBase &self) {
|
|
|
|
|
return PyOpOperandList(self.getOperation().getRef());
|
|
|
|
|
})
|
|
|
|
|
.def_prop_ro("regions",
|
|
|
|
|
[](PyOperationBase &self) {
|
|
|
|
|
return PyRegionList(self.getOperation().getRef());
|
|
|
|
|
})
|
|
|
|
|
.def_prop_ro(
|
2020-10-20 11:21:05 +02:00
|
|
|
"results",
|
2020-11-01 23:05:36 -08:00
|
|
|
[](PyOperationBase &self) {
|
|
|
|
|
return PyOpResultList(self.getOperation().getRef());
|
|
|
|
|
},
|
2020-10-20 11:21:05 +02:00
|
|
|
"Returns the list of Operation results.")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro(
|
2020-11-29 13:52:11 -08:00
|
|
|
"result",
|
|
|
|
|
[](PyOperationBase &self) {
|
|
|
|
|
auto &operation = self.getOperation();
|
|
|
|
|
auto numResults = mlirOperationGetNumResults(operation);
|
|
|
|
|
if (numResults != 1) {
|
|
|
|
|
auto name = mlirIdentifierStr(mlirOperationGetName(operation));
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error(
|
2023-05-22 17:30:12 -05:00
|
|
|
(Twine("Cannot call .result on operation ") +
|
|
|
|
|
StringRef(name.data, name.length) + " which has " +
|
|
|
|
|
Twine(numResults) +
|
|
|
|
|
" results (it is only valid for operations with a "
|
|
|
|
|
"single result)")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.str()
|
|
|
|
|
.c_str());
|
2020-11-29 13:52:11 -08:00
|
|
|
}
|
|
|
|
|
return PyOpResult(operation.getRef(),
|
[mlir][python] value casting (#69644)
This PR adds "value casting", i.e., a mechanism to wrap `ir.Value` in a
proxy class that overloads dunders such as `__add__`, `__sub__`, and
`__mul__` for fun and great profit.
This is thematically similar to
https://github.com/llvm/llvm-project/commit/bfb1ba752655bf09b35c486f6cc9817dbedfb1bb
and
https://github.com/llvm/llvm-project/commit/9566ee280607d91fa2e5eca730a6765ac84dfd0f.
The example in the test demonstrates the value of the feature (no pun
intended):
```python
@register_value_caster(F16Type.static_typeid)
@register_value_caster(F32Type.static_typeid)
@register_value_caster(F64Type.static_typeid)
@register_value_caster(IntegerType.static_typeid)
class ArithValue(Value):
__add__ = partialmethod(_binary_op, op="add")
__sub__ = partialmethod(_binary_op, op="sub")
__mul__ = partialmethod(_binary_op, op="mul")
a = arith.constant(value=FloatAttr.get(f16_t, 42.42))
b = a + a
# CHECK: ArithValue(%0 = arith.addf %cst, %cst : f16)
print(b)
a = arith.constant(value=FloatAttr.get(f32_t, 42.42))
b = a - a
# CHECK: ArithValue(%1 = arith.subf %cst_0, %cst_0 : f32)
print(b)
a = arith.constant(value=FloatAttr.get(f64_t, 42.42))
b = a * a
# CHECK: ArithValue(%2 = arith.mulf %cst_1, %cst_1 : f64)
print(b)
```
**EDIT**: this now goes through the bindings and thus supports automatic
casting of `OpResult` (including as an element of `OpResultList`),
`BlockArgument` (including as an element of `BlockArgumentList`), as
well as `Value`.
2023-11-07 10:49:41 -06:00
|
|
|
mlirOperationGetResult(operation, 0))
|
|
|
|
|
.maybeDownCast();
|
2020-11-29 13:52:11 -08:00
|
|
|
},
|
|
|
|
|
"Shortcut to get an op result if it has only one (throws an error "
|
|
|
|
|
"otherwise).")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro(
|
2021-10-18 16:00:39 +02:00
|
|
|
"location",
|
|
|
|
|
[](PyOperationBase &self) {
|
|
|
|
|
PyOperation &operation = self.getOperation();
|
|
|
|
|
return PyLocation(operation.getContext(),
|
|
|
|
|
mlirOperationGetLocation(operation.get()));
|
|
|
|
|
},
|
|
|
|
|
"Returns the source location the operation was defined or derived "
|
|
|
|
|
"from.")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro("parent",
|
|
|
|
|
[](PyOperationBase &self) -> nb::object {
|
|
|
|
|
auto parent = self.getOperation().getParentOperation();
|
|
|
|
|
if (parent)
|
|
|
|
|
return parent->getObject();
|
|
|
|
|
return nb::none();
|
|
|
|
|
})
|
2020-09-18 18:38:21 -07:00
|
|
|
.def(
|
|
|
|
|
"__str__",
|
2020-11-01 23:05:36 -08:00
|
|
|
[](PyOperationBase &self) {
|
2020-10-20 23:20:04 -07:00
|
|
|
return self.getAsm(/*binary=*/false,
|
2022-12-06 00:03:44 -08:00
|
|
|
/*largeElementsLimit=*/std::nullopt,
|
2020-10-20 23:20:04 -07:00
|
|
|
/*enableDebugInfo=*/false,
|
|
|
|
|
/*prettyDebugInfo=*/false,
|
|
|
|
|
/*printGenericOpForm=*/false,
|
2021-11-28 15:33:03 -08:00
|
|
|
/*useLocalScope=*/false,
|
2024-06-20 17:15:08 +02:00
|
|
|
/*assumeVerified=*/false,
|
|
|
|
|
/*skipRegions=*/false);
|
2020-09-18 18:38:21 -07:00
|
|
|
},
|
2020-10-20 23:20:04 -07:00
|
|
|
"Returns the assembly form of the operation.")
|
2023-11-13 10:21:21 -08:00
|
|
|
.def("print",
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::overload_cast<PyAsmState &, nb::object, bool>(
|
2023-11-13 10:21:21 -08:00
|
|
|
&PyOperationBase::print),
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("state"), nb::arg("file").none() = nb::none(),
|
|
|
|
|
nb::arg("binary") = false, kOperationPrintStateDocstring)
|
2023-11-13 10:21:21 -08:00
|
|
|
.def("print",
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::overload_cast<std::optional<int64_t>, bool, bool, bool, bool,
|
|
|
|
|
bool, nb::object, bool, bool>(
|
2024-06-20 17:15:08 +02:00
|
|
|
&PyOperationBase::print),
|
2020-10-20 23:20:04 -07:00
|
|
|
// Careful: Lots of arguments must match up with print method.
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("large_elements_limit").none() = nb::none(),
|
|
|
|
|
nb::arg("enable_debug_info") = false,
|
|
|
|
|
nb::arg("pretty_debug_info") = false,
|
|
|
|
|
nb::arg("print_generic_op_form") = false,
|
|
|
|
|
nb::arg("use_local_scope") = false,
|
|
|
|
|
nb::arg("assume_verified") = false,
|
|
|
|
|
nb::arg("file").none() = nb::none(), nb::arg("binary") = false,
|
|
|
|
|
nb::arg("skip_regions") = false, kOperationPrintDocstring)
|
|
|
|
|
.def("write_bytecode", &PyOperationBase::writeBytecode, nb::arg("file"),
|
|
|
|
|
nb::arg("desired_version").none() = nb::none(),
|
2022-09-05 11:54:19 +00:00
|
|
|
kOperationPrintBytecodeDocstring)
|
2020-11-01 23:05:36 -08:00
|
|
|
.def("get_asm", &PyOperationBase::getAsm,
|
2020-10-20 23:20:04 -07:00
|
|
|
// Careful: Lots of arguments must match up with get_asm method.
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("binary") = false,
|
|
|
|
|
nb::arg("large_elements_limit").none() = nb::none(),
|
|
|
|
|
nb::arg("enable_debug_info") = false,
|
|
|
|
|
nb::arg("pretty_debug_info") = false,
|
|
|
|
|
nb::arg("print_generic_op_form") = false,
|
|
|
|
|
nb::arg("use_local_scope") = false,
|
|
|
|
|
nb::arg("assume_verified") = false, nb::arg("skip_regions") = false,
|
2024-06-20 17:15:08 +02:00
|
|
|
kOperationGetAsmDocstring)
|
[mlir][python] Capture error diagnostics in exceptions
This updates most (all?) error-diagnostic-emitting python APIs to
capture error diagnostics and include them in the raised exception's
message:
```
>>> Operation.parse('"arith.addi"() : () -> ()'))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
mlir._mlir_libs.MLIRError: Unable to parse operation assembly:
error: "-":1:1: 'arith.addi' op requires one result
note: "-":1:1: see current operation: "arith.addi"() : () -> ()
```
The diagnostic information is available on the exception for users who
may want to customize the error message:
```
>>> try:
... Operation.parse('"arith.addi"() : () -> ()')
... except MLIRError as e:
... print(e.message)
... print(e.error_diagnostics)
... print(e.error_diagnostics[0].message)
...
Unable to parse operation assembly
[<mlir._mlir_libs._mlir.ir.DiagnosticInfo object at 0x7fed32bd6b70>]
'arith.addi' op requires one result
```
Error diagnostics captured in exceptions aren't propagated to diagnostic
handlers, to avoid double-reporting of errors. The context-level
`emit_error_diagnostics` option can be used to revert to the old
behaviour, causing error diagnostics to be reported to handlers instead
of as part of exceptions.
API changes:
- `Operation.verify` now raises an exception on verification failure,
instead of returning `false`
- The exception raised by the following methods has been changed to
`MLIRError`:
- `PassManager.run`
- `{Module,Operation,Type,Attribute}.parse`
- `{RankedTensorType,UnrankedTensorType}.get`
- `{MemRefType,UnrankedMemRefType}.get`
- `VectorType.get`
- `FloatAttr.get`
closes #60595
depends on D144804, D143830
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D143869
2023-02-07 16:07:50 -05:00
|
|
|
.def("verify", &PyOperationBase::verify,
|
|
|
|
|
"Verify the operation. Raises MLIRError if verification fails, and "
|
|
|
|
|
"returns true otherwise.")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def("move_after", &PyOperationBase::moveAfter, nb::arg("other"),
|
2021-10-31 09:37:20 +01:00
|
|
|
"Puts self immediately after the other operation in its parent "
|
|
|
|
|
"block.")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def("move_before", &PyOperationBase::moveBefore, nb::arg("other"),
|
2021-10-31 09:37:20 +01:00
|
|
|
"Puts self immediately before the other operation in its parent "
|
|
|
|
|
"block.")
|
2023-09-18 21:30:41 -07:00
|
|
|
.def(
|
|
|
|
|
"clone",
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
[](PyOperationBase &self, nb::object ip) {
|
2023-09-18 21:30:41 -07:00
|
|
|
return self.getOperation().clone(ip);
|
|
|
|
|
},
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("ip").none() = nb::none())
|
2021-10-31 09:37:20 +01:00
|
|
|
.def(
|
|
|
|
|
"detach_from_parent",
|
|
|
|
|
[](PyOperationBase &self) {
|
|
|
|
|
PyOperation &operation = self.getOperation();
|
|
|
|
|
operation.checkValid();
|
|
|
|
|
if (!operation.isAttached())
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error("Detached operation has no parent.");
|
2021-10-31 09:37:20 +01:00
|
|
|
|
|
|
|
|
operation.detachFromParent();
|
|
|
|
|
return operation.createOpView();
|
|
|
|
|
},
|
2023-09-18 21:30:41 -07:00
|
|
|
"Detaches the operation from its parent block.")
|
2024-04-17 15:09:47 +09:00
|
|
|
.def("erase", [](PyOperationBase &self) { self.getOperation().erase(); })
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def("walk", &PyOperationBase::walk, nb::arg("callback"),
|
|
|
|
|
nb::arg("walk_order") = MlirWalkPostOrder);
|
|
|
|
|
|
|
|
|
|
nb::class_<PyOperation, PyOperationBase>(m, "Operation")
|
|
|
|
|
.def_static("create", &PyOperation::create, nb::arg("name"),
|
|
|
|
|
nb::arg("results").none() = nb::none(),
|
|
|
|
|
nb::arg("operands").none() = nb::none(),
|
|
|
|
|
nb::arg("attributes").none() = nb::none(),
|
|
|
|
|
nb::arg("successors").none() = nb::none(),
|
|
|
|
|
nb::arg("regions") = 0, nb::arg("loc").none() = nb::none(),
|
|
|
|
|
nb::arg("ip").none() = nb::none(),
|
|
|
|
|
nb::arg("infer_type") = false, kOperationCreateDocstring)
|
2022-11-08 16:55:06 -05:00
|
|
|
.def_static(
|
|
|
|
|
"parse",
|
|
|
|
|
[](const std::string &sourceStr, const std::string &sourceName,
|
|
|
|
|
DefaultingPyMlirContext context) {
|
|
|
|
|
return PyOperation::parse(context->getRef(), sourceStr, sourceName)
|
|
|
|
|
->createOpView();
|
|
|
|
|
},
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("source"), nb::kw_only(), nb::arg("source_name") = "",
|
|
|
|
|
nb::arg("context").none() = nb::none(),
|
2022-11-08 16:55:06 -05:00
|
|
|
"Parses an operation. Supports both text assembly format and binary "
|
|
|
|
|
"bytecode format.")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro(MLIR_PYTHON_CAPI_PTR_ATTR, &PyOperation::getCapsule)
|
2021-04-06 14:15:22 -07:00
|
|
|
.def(MLIR_PYTHON_CAPI_FACTORY_ATTR, &PyOperation::createFromCapsule)
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro("operation", [](nb::object self) { return self; })
|
|
|
|
|
.def_prop_ro("opview", &PyOperation::createOpView)
|
|
|
|
|
.def_prop_ro(
|
2023-10-02 15:37:25 -05:00
|
|
|
"successors",
|
|
|
|
|
[](PyOperationBase &self) {
|
|
|
|
|
return PyOpSuccessors(self.getOperation().getRef());
|
|
|
|
|
},
|
|
|
|
|
"Returns the list of Operation successors.");
|
2020-11-01 23:05:36 -08:00
|
|
|
|
2021-01-18 11:27:19 -08:00
|
|
|
auto opViewClass =
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::class_<PyOpView, PyOperationBase>(m, "OpView")
|
|
|
|
|
.def(nb::init<nb::object>(), nb::arg("operation"))
|
|
|
|
|
.def_prop_ro("operation", &PyOpView::getOperationObject)
|
|
|
|
|
.def_prop_ro("opview", [](nb::object self) { return self; })
|
2023-10-02 15:37:25 -05:00
|
|
|
.def(
|
|
|
|
|
"__str__",
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
[](PyOpView &self) { return nb::str(self.getOperationObject()); })
|
|
|
|
|
.def_prop_ro(
|
2023-10-02 15:37:25 -05:00
|
|
|
"successors",
|
|
|
|
|
[](PyOperationBase &self) {
|
|
|
|
|
return PyOpSuccessors(self.getOperation().getRef());
|
|
|
|
|
},
|
|
|
|
|
"Returns the list of Operation successors.");
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
opViewClass.attr("_ODS_REGIONS") = nb::make_tuple(0, true);
|
|
|
|
|
opViewClass.attr("_ODS_OPERAND_SEGMENTS") = nb::none();
|
|
|
|
|
opViewClass.attr("_ODS_RESULT_SEGMENTS") = nb::none();
|
2021-01-24 14:46:56 -08:00
|
|
|
opViewClass.attr("build_generic") = classmethod(
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
&PyOpView::buildGeneric, nb::arg("cls"),
|
|
|
|
|
nb::arg("results").none() = nb::none(),
|
|
|
|
|
nb::arg("operands").none() = nb::none(),
|
|
|
|
|
nb::arg("attributes").none() = nb::none(),
|
|
|
|
|
nb::arg("successors").none() = nb::none(),
|
|
|
|
|
nb::arg("regions").none() = nb::none(),
|
|
|
|
|
nb::arg("loc").none() = nb::none(), nb::arg("ip").none() = nb::none(),
|
2021-01-18 11:27:19 -08:00
|
|
|
"Builds a specific, generated OpView based on class level attributes.");
|
2022-11-08 16:55:06 -05:00
|
|
|
opViewClass.attr("parse") = classmethod(
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
[](const nb::object &cls, const std::string &sourceStr,
|
2022-11-08 16:55:06 -05:00
|
|
|
const std::string &sourceName, DefaultingPyMlirContext context) {
|
|
|
|
|
PyOperationRef parsed =
|
|
|
|
|
PyOperation::parse(context->getRef(), sourceStr, sourceName);
|
|
|
|
|
|
|
|
|
|
// Check if the expected operation was parsed, and cast to to the
|
|
|
|
|
// appropriate `OpView` subclass if successful.
|
|
|
|
|
// NOTE: This accesses attributes that have been automatically added to
|
|
|
|
|
// `OpView` subclasses, and is not intended to be used on `OpView`
|
|
|
|
|
// directly.
|
|
|
|
|
std::string clsOpName =
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::cast<std::string>(cls.attr("OPERATION_NAME"));
|
[mlir][python] Capture error diagnostics in exceptions
This updates most (all?) error-diagnostic-emitting python APIs to
capture error diagnostics and include them in the raised exception's
message:
```
>>> Operation.parse('"arith.addi"() : () -> ()'))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
mlir._mlir_libs.MLIRError: Unable to parse operation assembly:
error: "-":1:1: 'arith.addi' op requires one result
note: "-":1:1: see current operation: "arith.addi"() : () -> ()
```
The diagnostic information is available on the exception for users who
may want to customize the error message:
```
>>> try:
... Operation.parse('"arith.addi"() : () -> ()')
... except MLIRError as e:
... print(e.message)
... print(e.error_diagnostics)
... print(e.error_diagnostics[0].message)
...
Unable to parse operation assembly
[<mlir._mlir_libs._mlir.ir.DiagnosticInfo object at 0x7fed32bd6b70>]
'arith.addi' op requires one result
```
Error diagnostics captured in exceptions aren't propagated to diagnostic
handlers, to avoid double-reporting of errors. The context-level
`emit_error_diagnostics` option can be used to revert to the old
behaviour, causing error diagnostics to be reported to handlers instead
of as part of exceptions.
API changes:
- `Operation.verify` now raises an exception on verification failure,
instead of returning `false`
- The exception raised by the following methods has been changed to
`MLIRError`:
- `PassManager.run`
- `{Module,Operation,Type,Attribute}.parse`
- `{RankedTensorType,UnrankedTensorType}.get`
- `{MemRefType,UnrankedMemRefType}.get`
- `VectorType.get`
- `FloatAttr.get`
closes #60595
depends on D144804, D143830
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D143869
2023-02-07 16:07:50 -05:00
|
|
|
MlirStringRef identifier =
|
2022-11-08 16:55:06 -05:00
|
|
|
mlirIdentifierStr(mlirOperationGetName(*parsed.get()));
|
[mlir][python] Capture error diagnostics in exceptions
This updates most (all?) error-diagnostic-emitting python APIs to
capture error diagnostics and include them in the raised exception's
message:
```
>>> Operation.parse('"arith.addi"() : () -> ()'))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
mlir._mlir_libs.MLIRError: Unable to parse operation assembly:
error: "-":1:1: 'arith.addi' op requires one result
note: "-":1:1: see current operation: "arith.addi"() : () -> ()
```
The diagnostic information is available on the exception for users who
may want to customize the error message:
```
>>> try:
... Operation.parse('"arith.addi"() : () -> ()')
... except MLIRError as e:
... print(e.message)
... print(e.error_diagnostics)
... print(e.error_diagnostics[0].message)
...
Unable to parse operation assembly
[<mlir._mlir_libs._mlir.ir.DiagnosticInfo object at 0x7fed32bd6b70>]
'arith.addi' op requires one result
```
Error diagnostics captured in exceptions aren't propagated to diagnostic
handlers, to avoid double-reporting of errors. The context-level
`emit_error_diagnostics` option can be used to revert to the old
behaviour, causing error diagnostics to be reported to handlers instead
of as part of exceptions.
API changes:
- `Operation.verify` now raises an exception on verification failure,
instead of returning `false`
- The exception raised by the following methods has been changed to
`MLIRError`:
- `PassManager.run`
- `{Module,Operation,Type,Attribute}.parse`
- `{RankedTensorType,UnrankedTensorType}.get`
- `{MemRefType,UnrankedMemRefType}.get`
- `VectorType.get`
- `FloatAttr.get`
closes #60595
depends on D144804, D143830
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D143869
2023-02-07 16:07:50 -05:00
|
|
|
std::string_view parsedOpName(identifier.data, identifier.length);
|
|
|
|
|
if (clsOpName != parsedOpName)
|
|
|
|
|
throw MLIRError(Twine("Expected a '") + clsOpName + "' op, got: '" +
|
|
|
|
|
parsedOpName + "'");
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
return PyOpView::constructDerived(cls, parsed.getObject());
|
2022-11-08 16:55:06 -05:00
|
|
|
},
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("cls"), nb::arg("source"), nb::kw_only(),
|
|
|
|
|
nb::arg("source_name") = "", nb::arg("context").none() = nb::none(),
|
2022-11-08 16:55:06 -05:00
|
|
|
"Parses a specific, generated OpView based on class level attributes");
|
2020-10-21 23:34:01 -07:00
|
|
|
|
|
|
|
|
//----------------------------------------------------------------------------
|
2020-08-26 23:48:42 -07:00
|
|
|
// Mapping of PyRegion.
|
2020-10-21 23:34:01 -07:00
|
|
|
//----------------------------------------------------------------------------
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::class_<PyRegion>(m, "Region")
|
|
|
|
|
.def_prop_ro(
|
2020-09-19 22:02:32 -07:00
|
|
|
"blocks",
|
|
|
|
|
[](PyRegion &self) {
|
|
|
|
|
return PyBlockList(self.getParentOperation(), self.get());
|
2020-08-26 23:48:42 -07:00
|
|
|
},
|
2020-09-19 22:02:32 -07:00
|
|
|
"Returns a forward-optimized sequence of blocks.")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro(
|
2021-10-11 18:24:48 +02:00
|
|
|
"owner",
|
|
|
|
|
[](PyRegion &self) {
|
|
|
|
|
return self.getParentOperation()->createOpView();
|
|
|
|
|
},
|
|
|
|
|
"Returns the operation owning this region.")
|
2020-08-26 23:48:42 -07:00
|
|
|
.def(
|
2020-09-19 22:02:32 -07:00
|
|
|
"__iter__",
|
2020-08-26 23:48:42 -07:00
|
|
|
[](PyRegion &self) {
|
2020-09-19 22:02:32 -07:00
|
|
|
self.checkValid();
|
|
|
|
|
MlirBlock firstBlock = mlirRegionGetFirstBlock(self.get());
|
|
|
|
|
return PyBlockIterator(self.getParentOperation(), firstBlock);
|
2020-08-26 23:48:42 -07:00
|
|
|
},
|
2020-09-19 22:02:32 -07:00
|
|
|
"Iterates over blocks in the region.")
|
2020-11-01 23:05:36 -08:00
|
|
|
.def("__eq__",
|
|
|
|
|
[](PyRegion &self, PyRegion &other) {
|
|
|
|
|
return self.get().ptr == other.get().ptr;
|
|
|
|
|
})
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def("__eq__", [](PyRegion &self, nb::object &other) { return false; });
|
2020-08-26 23:48:42 -07:00
|
|
|
|
2020-10-21 23:34:01 -07:00
|
|
|
//----------------------------------------------------------------------------
|
2020-08-26 23:48:42 -07:00
|
|
|
// Mapping of PyBlock.
|
2020-10-21 23:34:01 -07:00
|
|
|
//----------------------------------------------------------------------------
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::class_<PyBlock>(m, "Block")
|
|
|
|
|
.def_prop_ro(MLIR_PYTHON_CAPI_PTR_ATTR, &PyBlock::getCapsule)
|
|
|
|
|
.def_prop_ro(
|
2021-08-19 00:02:09 -07:00
|
|
|
"owner",
|
|
|
|
|
[](PyBlock &self) {
|
|
|
|
|
return self.getParentOperation()->createOpView();
|
|
|
|
|
},
|
|
|
|
|
"Returns the owning operation of this block.")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro(
|
2021-08-28 20:15:51 -07:00
|
|
|
"region",
|
|
|
|
|
[](PyBlock &self) {
|
|
|
|
|
MlirRegion region = mlirBlockGetParentRegion(self.get());
|
|
|
|
|
return PyRegion(self.getParentOperation(), region);
|
|
|
|
|
},
|
|
|
|
|
"Returns the owning region of this block.")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro(
|
2020-10-20 11:21:05 +02:00
|
|
|
"arguments",
|
|
|
|
|
[](PyBlock &self) {
|
|
|
|
|
return PyBlockArgumentList(self.getParentOperation(), self.get());
|
|
|
|
|
},
|
|
|
|
|
"Returns a list of block arguments.")
|
2024-06-05 22:10:55 +00:00
|
|
|
.def(
|
|
|
|
|
"add_argument",
|
|
|
|
|
[](PyBlock &self, const PyType &type, const PyLocation &loc) {
|
|
|
|
|
return mlirBlockAddArgument(self.get(), type, loc);
|
|
|
|
|
},
|
|
|
|
|
"Append an argument of the specified type to the block and returns "
|
|
|
|
|
"the newly added argument.")
|
|
|
|
|
.def(
|
|
|
|
|
"erase_argument",
|
|
|
|
|
[](PyBlock &self, unsigned index) {
|
|
|
|
|
return mlirBlockEraseArgument(self.get(), index);
|
|
|
|
|
},
|
|
|
|
|
"Erase the argument at 'index' and remove it from the argument list.")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro(
|
2020-09-19 22:02:32 -07:00
|
|
|
"operations",
|
2020-08-26 23:48:42 -07:00
|
|
|
[](PyBlock &self) {
|
2020-09-19 22:02:32 -07:00
|
|
|
return PyOperationList(self.getParentOperation(), self.get());
|
2020-08-26 23:48:42 -07:00
|
|
|
},
|
2020-09-19 22:02:32 -07:00
|
|
|
"Returns a forward-optimized sequence of operations.")
|
2021-10-11 18:24:48 +02:00
|
|
|
.def_static(
|
|
|
|
|
"create_at_start",
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
[](PyRegion &parent, const nb::sequence &pyArgTypes,
|
|
|
|
|
const std::optional<nb::sequence> &pyArgLocs) {
|
2021-10-11 18:24:48 +02:00
|
|
|
parent.checkValid();
|
2023-05-24 21:51:36 -04:00
|
|
|
MlirBlock block = createBlock(pyArgTypes, pyArgLocs);
|
2021-10-11 18:24:48 +02:00
|
|
|
mlirRegionInsertOwnedBlock(parent, 0, block);
|
|
|
|
|
return PyBlock(parent.getParentOperation(), block);
|
|
|
|
|
},
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("parent"), nb::arg("arg_types") = nb::list(),
|
|
|
|
|
nb::arg("arg_locs") = std::nullopt,
|
2021-10-11 18:24:48 +02:00
|
|
|
"Creates and returns a new Block at the beginning of the given "
|
2023-05-24 21:51:36 -04:00
|
|
|
"region (with given argument types and locations).")
|
2022-04-06 10:06:30 -07:00
|
|
|
.def(
|
|
|
|
|
"append_to",
|
|
|
|
|
[](PyBlock &self, PyRegion ®ion) {
|
|
|
|
|
MlirBlock b = self.get();
|
|
|
|
|
if (!mlirRegionIsNull(mlirBlockGetParentRegion(b)))
|
|
|
|
|
mlirBlockDetach(b);
|
|
|
|
|
mlirRegionAppendOwnedBlock(region.get(), b);
|
|
|
|
|
},
|
|
|
|
|
"Append this block to a region, transferring ownership if necessary")
|
2021-08-28 20:15:51 -07:00
|
|
|
.def(
|
|
|
|
|
"create_before",
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
[](PyBlock &self, const nb::args &pyArgTypes,
|
|
|
|
|
const std::optional<nb::sequence> &pyArgLocs) {
|
2021-08-28 20:15:51 -07:00
|
|
|
self.checkValid();
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
MlirBlock block =
|
|
|
|
|
createBlock(nb::cast<nb::sequence>(pyArgTypes), pyArgLocs);
|
2021-08-28 20:15:51 -07:00
|
|
|
MlirRegion region = mlirBlockGetParentRegion(self.get());
|
|
|
|
|
mlirRegionInsertOwnedBlockBefore(region, self.get(), block);
|
|
|
|
|
return PyBlock(self.getParentOperation(), block);
|
|
|
|
|
},
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("arg_types"), nb::kw_only(),
|
|
|
|
|
nb::arg("arg_locs") = std::nullopt,
|
2021-08-28 20:15:51 -07:00
|
|
|
"Creates and returns a new Block before this block "
|
2023-05-24 21:51:36 -04:00
|
|
|
"(with given argument types and locations).")
|
2021-08-28 20:15:51 -07:00
|
|
|
.def(
|
|
|
|
|
"create_after",
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
[](PyBlock &self, const nb::args &pyArgTypes,
|
|
|
|
|
const std::optional<nb::sequence> &pyArgLocs) {
|
2021-08-28 20:15:51 -07:00
|
|
|
self.checkValid();
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
MlirBlock block =
|
|
|
|
|
createBlock(nb::cast<nb::sequence>(pyArgTypes), pyArgLocs);
|
2021-08-28 20:15:51 -07:00
|
|
|
MlirRegion region = mlirBlockGetParentRegion(self.get());
|
|
|
|
|
mlirRegionInsertOwnedBlockAfter(region, self.get(), block);
|
|
|
|
|
return PyBlock(self.getParentOperation(), block);
|
|
|
|
|
},
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("arg_types"), nb::kw_only(),
|
|
|
|
|
nb::arg("arg_locs") = std::nullopt,
|
2021-08-28 20:15:51 -07:00
|
|
|
"Creates and returns a new Block after this block "
|
2023-05-24 21:51:36 -04:00
|
|
|
"(with given argument types and locations).")
|
2020-09-19 22:02:32 -07:00
|
|
|
.def(
|
|
|
|
|
"__iter__",
|
|
|
|
|
[](PyBlock &self) {
|
|
|
|
|
self.checkValid();
|
|
|
|
|
MlirOperation firstOperation =
|
|
|
|
|
mlirBlockGetFirstOperation(self.get());
|
|
|
|
|
return PyOperationIterator(self.getParentOperation(),
|
|
|
|
|
firstOperation);
|
|
|
|
|
},
|
|
|
|
|
"Iterates over operations in the block.")
|
|
|
|
|
.def("__eq__",
|
2020-11-01 23:05:36 -08:00
|
|
|
[](PyBlock &self, PyBlock &other) {
|
|
|
|
|
return self.get().ptr == other.get().ptr;
|
2020-09-19 22:02:32 -07:00
|
|
|
})
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def("__eq__", [](PyBlock &self, nb::object &other) { return false; })
|
2022-12-07 13:03:24 -07:00
|
|
|
.def("__hash__",
|
|
|
|
|
[](PyBlock &self) {
|
|
|
|
|
return static_cast<size_t>(llvm::hash_value(self.get().ptr));
|
|
|
|
|
})
|
2020-08-26 23:48:42 -07:00
|
|
|
.def(
|
|
|
|
|
"__str__",
|
|
|
|
|
[](PyBlock &self) {
|
2020-09-19 22:02:32 -07:00
|
|
|
self.checkValid();
|
2020-08-26 23:48:42 -07:00
|
|
|
PyPrintAccumulator printAccum;
|
2020-09-19 22:02:32 -07:00
|
|
|
mlirBlockPrint(self.get(), printAccum.getCallback(),
|
2020-08-26 23:48:42 -07:00
|
|
|
printAccum.getUserData());
|
|
|
|
|
return printAccum.join();
|
|
|
|
|
},
|
2021-10-31 09:37:20 +01:00
|
|
|
"Returns the assembly form of the block.")
|
|
|
|
|
.def(
|
|
|
|
|
"append",
|
|
|
|
|
[](PyBlock &self, PyOperationBase &operation) {
|
|
|
|
|
if (operation.getOperation().isAttached())
|
|
|
|
|
operation.getOperation().detachFromParent();
|
|
|
|
|
|
|
|
|
|
MlirOperation mlirOperation = operation.getOperation().get();
|
|
|
|
|
mlirBlockAppendOwnedOperation(self.get(), mlirOperation);
|
|
|
|
|
operation.getOperation().setAttached(
|
|
|
|
|
self.getParentOperation().getObject());
|
|
|
|
|
},
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("operation"),
|
2021-10-31 09:37:20 +01:00
|
|
|
"Appends an operation to this block. If the operation is currently "
|
|
|
|
|
"in another block, it will be moved.");
|
2020-08-26 23:48:42 -07:00
|
|
|
|
2020-10-28 23:16:36 -07:00
|
|
|
//----------------------------------------------------------------------------
|
|
|
|
|
// Mapping of PyInsertionPoint.
|
|
|
|
|
//----------------------------------------------------------------------------
|
|
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::class_<PyInsertionPoint>(m, "InsertionPoint")
|
|
|
|
|
.def(nb::init<PyBlock &>(), nb::arg("block"),
|
2020-10-28 23:16:36 -07:00
|
|
|
"Inserts after the last operation but still inside the block.")
|
|
|
|
|
.def("__enter__", &PyInsertionPoint::contextEnter)
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def("__exit__", &PyInsertionPoint::contextExit,
|
|
|
|
|
nb::arg("exc_type").none(), nb::arg("exc_value").none(),
|
|
|
|
|
nb::arg("traceback").none())
|
|
|
|
|
.def_prop_ro_static(
|
2020-10-31 23:40:25 -07:00
|
|
|
"current",
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
[](nb::object & /*class*/) {
|
2020-10-31 23:40:25 -07:00
|
|
|
auto *ip = PyThreadContextEntry::getDefaultInsertionPoint();
|
|
|
|
|
if (!ip)
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
throw nb::value_error("No current InsertionPoint");
|
2020-10-31 23:40:25 -07:00
|
|
|
return ip;
|
|
|
|
|
},
|
|
|
|
|
"Gets the InsertionPoint bound to the current thread or raises "
|
|
|
|
|
"ValueError if none has been set")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def(nb::init<PyOperationBase &>(), nb::arg("beforeOperation"),
|
2020-10-28 23:16:36 -07:00
|
|
|
"Inserts before a referenced operation.")
|
|
|
|
|
.def_static("at_block_begin", &PyInsertionPoint::atBlockBegin,
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("block"), "Inserts at the beginning of the block.")
|
2020-10-28 23:16:36 -07:00
|
|
|
.def_static("at_block_terminator", &PyInsertionPoint::atBlockTerminator,
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("block"), "Inserts before the block terminator.")
|
|
|
|
|
.def("insert", &PyInsertionPoint::insert, nb::arg("operation"),
|
2021-08-28 20:15:51 -07:00
|
|
|
"Inserts an operation.")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro(
|
2021-08-28 20:15:51 -07:00
|
|
|
"block", [](PyInsertionPoint &self) { return self.getBlock(); },
|
2023-10-18 16:53:18 +02:00
|
|
|
"Returns the block that this InsertionPoint points to.")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro(
|
2023-10-18 16:53:18 +02:00
|
|
|
"ref_operation",
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
[](PyInsertionPoint &self) -> nb::object {
|
2023-10-20 01:35:56 -07:00
|
|
|
auto refOperation = self.getRefOperation();
|
|
|
|
|
if (refOperation)
|
|
|
|
|
return refOperation->getObject();
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
return nb::none();
|
2023-10-18 16:53:18 +02:00
|
|
|
},
|
|
|
|
|
"The reference operation before which new operations are "
|
|
|
|
|
"inserted, or None if the insertion point is at the end of "
|
|
|
|
|
"the block");
|
2020-10-28 23:16:36 -07:00
|
|
|
|
2020-10-21 23:34:01 -07:00
|
|
|
//----------------------------------------------------------------------------
|
2020-10-20 11:22:29 +02:00
|
|
|
// Mapping of PyAttribute.
|
2020-10-21 23:34:01 -07:00
|
|
|
//----------------------------------------------------------------------------
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::class_<PyAttribute>(m, "Attribute")
|
2021-05-02 15:15:21 -07:00
|
|
|
// Delegate to the PyAttribute copy constructor, which will also lifetime
|
|
|
|
|
// extend the backing context which owns the MlirAttribute.
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def(nb::init<PyAttribute &>(), nb::arg("cast_from_type"),
|
2021-05-02 15:15:21 -07:00
|
|
|
"Casts the passed attribute to the generic Attribute")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro(MLIR_PYTHON_CAPI_PTR_ATTR, &PyAttribute::getCapsule)
|
2020-11-29 13:30:23 -08:00
|
|
|
.def(MLIR_PYTHON_CAPI_FACTORY_ATTR, &PyAttribute::createFromCapsule)
|
2020-10-31 23:40:25 -07:00
|
|
|
.def_static(
|
|
|
|
|
"parse",
|
2023-07-05 15:02:59 -05:00
|
|
|
[](const std::string &attrSpec, DefaultingPyMlirContext context) {
|
[mlir][python] Capture error diagnostics in exceptions
This updates most (all?) error-diagnostic-emitting python APIs to
capture error diagnostics and include them in the raised exception's
message:
```
>>> Operation.parse('"arith.addi"() : () -> ()'))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
mlir._mlir_libs.MLIRError: Unable to parse operation assembly:
error: "-":1:1: 'arith.addi' op requires one result
note: "-":1:1: see current operation: "arith.addi"() : () -> ()
```
The diagnostic information is available on the exception for users who
may want to customize the error message:
```
>>> try:
... Operation.parse('"arith.addi"() : () -> ()')
... except MLIRError as e:
... print(e.message)
... print(e.error_diagnostics)
... print(e.error_diagnostics[0].message)
...
Unable to parse operation assembly
[<mlir._mlir_libs._mlir.ir.DiagnosticInfo object at 0x7fed32bd6b70>]
'arith.addi' op requires one result
```
Error diagnostics captured in exceptions aren't propagated to diagnostic
handlers, to avoid double-reporting of errors. The context-level
`emit_error_diagnostics` option can be used to revert to the old
behaviour, causing error diagnostics to be reported to handlers instead
of as part of exceptions.
API changes:
- `Operation.verify` now raises an exception on verification failure,
instead of returning `false`
- The exception raised by the following methods has been changed to
`MLIRError`:
- `PassManager.run`
- `{Module,Operation,Type,Attribute}.parse`
- `{RankedTensorType,UnrankedTensorType}.get`
- `{MemRefType,UnrankedMemRefType}.get`
- `VectorType.get`
- `FloatAttr.get`
closes #60595
depends on D144804, D143830
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D143869
2023-02-07 16:07:50 -05:00
|
|
|
PyMlirContext::ErrorCapture errors(context->getRef());
|
2023-07-05 15:02:59 -05:00
|
|
|
MlirAttribute attr = mlirAttributeParseGet(
|
2020-11-24 18:35:22 +00:00
|
|
|
context->get(), toMlirStringRef(attrSpec));
|
2023-07-05 15:02:59 -05:00
|
|
|
if (mlirAttributeIsNull(attr))
|
[mlir][python] Capture error diagnostics in exceptions
This updates most (all?) error-diagnostic-emitting python APIs to
capture error diagnostics and include them in the raised exception's
message:
```
>>> Operation.parse('"arith.addi"() : () -> ()'))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
mlir._mlir_libs.MLIRError: Unable to parse operation assembly:
error: "-":1:1: 'arith.addi' op requires one result
note: "-":1:1: see current operation: "arith.addi"() : () -> ()
```
The diagnostic information is available on the exception for users who
may want to customize the error message:
```
>>> try:
... Operation.parse('"arith.addi"() : () -> ()')
... except MLIRError as e:
... print(e.message)
... print(e.error_diagnostics)
... print(e.error_diagnostics[0].message)
...
Unable to parse operation assembly
[<mlir._mlir_libs._mlir.ir.DiagnosticInfo object at 0x7fed32bd6b70>]
'arith.addi' op requires one result
```
Error diagnostics captured in exceptions aren't propagated to diagnostic
handlers, to avoid double-reporting of errors. The context-level
`emit_error_diagnostics` option can be used to revert to the old
behaviour, causing error diagnostics to be reported to handlers instead
of as part of exceptions.
API changes:
- `Operation.verify` now raises an exception on verification failure,
instead of returning `false`
- The exception raised by the following methods has been changed to
`MLIRError`:
- `PassManager.run`
- `{Module,Operation,Type,Attribute}.parse`
- `{RankedTensorType,UnrankedTensorType}.get`
- `{MemRefType,UnrankedMemRefType}.get`
- `VectorType.get`
- `FloatAttr.get`
closes #60595
depends on D144804, D143830
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D143869
2023-02-07 16:07:50 -05:00
|
|
|
throw MLIRError("Unable to parse attribute", errors.take());
|
2023-07-05 15:02:59 -05:00
|
|
|
return attr;
|
2020-10-31 23:40:25 -07:00
|
|
|
},
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("asm"), nb::arg("context").none() = nb::none(),
|
[mlir][python] Capture error diagnostics in exceptions
This updates most (all?) error-diagnostic-emitting python APIs to
capture error diagnostics and include them in the raised exception's
message:
```
>>> Operation.parse('"arith.addi"() : () -> ()'))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
mlir._mlir_libs.MLIRError: Unable to parse operation assembly:
error: "-":1:1: 'arith.addi' op requires one result
note: "-":1:1: see current operation: "arith.addi"() : () -> ()
```
The diagnostic information is available on the exception for users who
may want to customize the error message:
```
>>> try:
... Operation.parse('"arith.addi"() : () -> ()')
... except MLIRError as e:
... print(e.message)
... print(e.error_diagnostics)
... print(e.error_diagnostics[0].message)
...
Unable to parse operation assembly
[<mlir._mlir_libs._mlir.ir.DiagnosticInfo object at 0x7fed32bd6b70>]
'arith.addi' op requires one result
```
Error diagnostics captured in exceptions aren't propagated to diagnostic
handlers, to avoid double-reporting of errors. The context-level
`emit_error_diagnostics` option can be used to revert to the old
behaviour, causing error diagnostics to be reported to handlers instead
of as part of exceptions.
API changes:
- `Operation.verify` now raises an exception on verification failure,
instead of returning `false`
- The exception raised by the following methods has been changed to
`MLIRError`:
- `PassManager.run`
- `{Module,Operation,Type,Attribute}.parse`
- `{RankedTensorType,UnrankedTensorType}.get`
- `{MemRefType,UnrankedMemRefType}.get`
- `VectorType.get`
- `FloatAttr.get`
closes #60595
depends on D144804, D143830
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D143869
2023-02-07 16:07:50 -05:00
|
|
|
"Parses an attribute from an assembly form. Raises an MLIRError on "
|
|
|
|
|
"failure.")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro(
|
2020-10-12 21:19:13 -07:00
|
|
|
"context",
|
|
|
|
|
[](PyAttribute &self) { return self.getContext().getObject(); },
|
|
|
|
|
"Context that owns the Attribute")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro("type",
|
|
|
|
|
[](PyAttribute &self) { return mlirAttributeGetType(self); })
|
2020-08-19 15:33:02 -07:00
|
|
|
.def(
|
|
|
|
|
"get_named",
|
|
|
|
|
[](PyAttribute &self, std::string name) {
|
2020-11-29 13:30:23 -08:00
|
|
|
return PyNamedAttribute(self, std::move(name));
|
2020-08-19 15:33:02 -07:00
|
|
|
},
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::keep_alive<0, 1>(), "Binds a name to the attribute")
|
2020-08-19 15:33:02 -07:00
|
|
|
.def("__eq__",
|
2020-11-01 23:05:36 -08:00
|
|
|
[](PyAttribute &self, PyAttribute &other) { return self == other; })
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def("__eq__", [](PyAttribute &self, nb::object &other) { return false; })
|
2021-11-02 17:04:42 +01:00
|
|
|
.def("__hash__",
|
|
|
|
|
[](PyAttribute &self) {
|
|
|
|
|
return static_cast<size_t>(llvm::hash_value(self.get().ptr));
|
|
|
|
|
})
|
2020-08-19 15:33:02 -07:00
|
|
|
.def(
|
2020-11-29 13:30:23 -08:00
|
|
|
"dump", [](PyAttribute &self) { mlirAttributeDump(self); },
|
2020-08-19 15:33:02 -07:00
|
|
|
kDumpDocstring)
|
|
|
|
|
.def(
|
|
|
|
|
"__str__",
|
|
|
|
|
[](PyAttribute &self) {
|
|
|
|
|
PyPrintAccumulator printAccum;
|
2020-11-29 13:30:23 -08:00
|
|
|
mlirAttributePrint(self, printAccum.getCallback(),
|
2020-08-19 15:33:02 -07:00
|
|
|
printAccum.getUserData());
|
|
|
|
|
return printAccum.join();
|
|
|
|
|
},
|
2020-10-20 11:22:29 +02:00
|
|
|
"Returns the assembly form of the Attribute.")
|
2023-05-31 15:52:46 -05:00
|
|
|
.def("__repr__",
|
|
|
|
|
[](PyAttribute &self) {
|
|
|
|
|
// Generally, assembly formats are not printed for __repr__ because
|
|
|
|
|
// this can cause exceptionally long debug output and exceptions.
|
|
|
|
|
// However, attribute values are generally considered useful and
|
|
|
|
|
// are printed. This may need to be re-evaluated if debug dumps end
|
|
|
|
|
// up being excessive.
|
|
|
|
|
PyPrintAccumulator printAccum;
|
|
|
|
|
printAccum.parts.append("Attribute(");
|
|
|
|
|
mlirAttributePrint(self, printAccum.getCallback(),
|
|
|
|
|
printAccum.getUserData());
|
|
|
|
|
printAccum.parts.append(")");
|
|
|
|
|
return printAccum.join();
|
|
|
|
|
})
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro("typeid",
|
|
|
|
|
[](PyAttribute &self) -> MlirTypeID {
|
|
|
|
|
MlirTypeID mlirTypeID = mlirAttributeGetTypeID(self);
|
|
|
|
|
assert(!mlirTypeIDIsNull(mlirTypeID) &&
|
|
|
|
|
"mlirTypeID was expected to be non-null.");
|
|
|
|
|
return mlirTypeID;
|
|
|
|
|
})
|
2023-05-31 15:52:46 -05:00
|
|
|
.def(MLIR_PYTHON_MAYBE_DOWNCAST_ATTR, [](PyAttribute &self) {
|
|
|
|
|
MlirTypeID mlirTypeID = mlirAttributeGetTypeID(self);
|
|
|
|
|
assert(!mlirTypeIDIsNull(mlirTypeID) &&
|
|
|
|
|
"mlirTypeID was expected to be non-null.");
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
std::optional<nb::callable> typeCaster =
|
2023-05-31 15:52:46 -05:00
|
|
|
PyGlobals::get().lookupTypeCaster(mlirTypeID,
|
|
|
|
|
mlirAttributeGetDialect(self));
|
|
|
|
|
if (!typeCaster)
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
return nb::cast(self);
|
2023-05-31 15:52:46 -05:00
|
|
|
return typeCaster.value()(self);
|
2020-08-19 15:33:02 -07:00
|
|
|
});
|
|
|
|
|
|
2020-10-21 23:34:01 -07:00
|
|
|
//----------------------------------------------------------------------------
|
|
|
|
|
// Mapping of PyNamedAttribute
|
|
|
|
|
//----------------------------------------------------------------------------
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::class_<PyNamedAttribute>(m, "NamedAttribute")
|
2020-08-19 15:33:02 -07:00
|
|
|
.def("__repr__",
|
|
|
|
|
[](PyNamedAttribute &self) {
|
|
|
|
|
PyPrintAccumulator printAccum;
|
|
|
|
|
printAccum.parts.append("NamedAttribute(");
|
2020-12-11 18:50:04 +00:00
|
|
|
printAccum.parts.append(
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::str(mlirIdentifierStr(self.namedAttr.name).data,
|
2021-11-11 01:44:58 +00:00
|
|
|
mlirIdentifierStr(self.namedAttr.name).length));
|
2020-08-19 15:33:02 -07:00
|
|
|
printAccum.parts.append("=");
|
|
|
|
|
mlirAttributePrint(self.namedAttr.attribute,
|
|
|
|
|
printAccum.getCallback(),
|
|
|
|
|
printAccum.getUserData());
|
|
|
|
|
printAccum.parts.append(")");
|
|
|
|
|
return printAccum.join();
|
|
|
|
|
})
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro(
|
2020-08-19 15:33:02 -07:00
|
|
|
"name",
|
|
|
|
|
[](PyNamedAttribute &self) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
return nb::str(mlirIdentifierStr(self.namedAttr.name).data,
|
2020-12-11 18:50:04 +00:00
|
|
|
mlirIdentifierStr(self.namedAttr.name).length);
|
2020-08-19 15:33:02 -07:00
|
|
|
},
|
|
|
|
|
"The name of the NamedAttribute binding")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro(
|
2020-08-19 15:33:02 -07:00
|
|
|
"attr",
|
2023-05-31 15:52:46 -05:00
|
|
|
[](PyNamedAttribute &self) { return self.namedAttr.attribute; },
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::keep_alive<0, 1>(),
|
2020-08-19 15:33:02 -07:00
|
|
|
"The underlying generic attribute of the NamedAttribute binding");
|
|
|
|
|
|
2020-10-21 23:34:01 -07:00
|
|
|
//----------------------------------------------------------------------------
|
2020-10-20 11:22:29 +02:00
|
|
|
// Mapping of PyType.
|
2020-10-21 23:34:01 -07:00
|
|
|
//----------------------------------------------------------------------------
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::class_<PyType>(m, "Type")
|
2021-05-02 15:15:21 -07:00
|
|
|
// Delegate to the PyType copy constructor, which will also lifetime
|
|
|
|
|
// extend the backing context which owns the MlirType.
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def(nb::init<PyType &>(), nb::arg("cast_from_type"),
|
2021-05-02 15:15:21 -07:00
|
|
|
"Casts the passed type to the generic Type")
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro(MLIR_PYTHON_CAPI_PTR_ATTR, &PyType::getCapsule)
|
2020-11-29 13:30:23 -08:00
|
|
|
.def(MLIR_PYTHON_CAPI_FACTORY_ATTR, &PyType::createFromCapsule)
|
2020-10-31 23:40:25 -07:00
|
|
|
.def_static(
|
|
|
|
|
"parse",
|
|
|
|
|
[](std::string typeSpec, DefaultingPyMlirContext context) {
|
[mlir][python] Capture error diagnostics in exceptions
This updates most (all?) error-diagnostic-emitting python APIs to
capture error diagnostics and include them in the raised exception's
message:
```
>>> Operation.parse('"arith.addi"() : () -> ()'))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
mlir._mlir_libs.MLIRError: Unable to parse operation assembly:
error: "-":1:1: 'arith.addi' op requires one result
note: "-":1:1: see current operation: "arith.addi"() : () -> ()
```
The diagnostic information is available on the exception for users who
may want to customize the error message:
```
>>> try:
... Operation.parse('"arith.addi"() : () -> ()')
... except MLIRError as e:
... print(e.message)
... print(e.error_diagnostics)
... print(e.error_diagnostics[0].message)
...
Unable to parse operation assembly
[<mlir._mlir_libs._mlir.ir.DiagnosticInfo object at 0x7fed32bd6b70>]
'arith.addi' op requires one result
```
Error diagnostics captured in exceptions aren't propagated to diagnostic
handlers, to avoid double-reporting of errors. The context-level
`emit_error_diagnostics` option can be used to revert to the old
behaviour, causing error diagnostics to be reported to handlers instead
of as part of exceptions.
API changes:
- `Operation.verify` now raises an exception on verification failure,
instead of returning `false`
- The exception raised by the following methods has been changed to
`MLIRError`:
- `PassManager.run`
- `{Module,Operation,Type,Attribute}.parse`
- `{RankedTensorType,UnrankedTensorType}.get`
- `{MemRefType,UnrankedMemRefType}.get`
- `VectorType.get`
- `FloatAttr.get`
closes #60595
depends on D144804, D143830
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D143869
2023-02-07 16:07:50 -05:00
|
|
|
PyMlirContext::ErrorCapture errors(context->getRef());
|
2020-11-24 18:35:22 +00:00
|
|
|
MlirType type =
|
|
|
|
|
mlirTypeParseGet(context->get(), toMlirStringRef(typeSpec));
|
[mlir][python] Capture error diagnostics in exceptions
This updates most (all?) error-diagnostic-emitting python APIs to
capture error diagnostics and include them in the raised exception's
message:
```
>>> Operation.parse('"arith.addi"() : () -> ()'))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
mlir._mlir_libs.MLIRError: Unable to parse operation assembly:
error: "-":1:1: 'arith.addi' op requires one result
note: "-":1:1: see current operation: "arith.addi"() : () -> ()
```
The diagnostic information is available on the exception for users who
may want to customize the error message:
```
>>> try:
... Operation.parse('"arith.addi"() : () -> ()')
... except MLIRError as e:
... print(e.message)
... print(e.error_diagnostics)
... print(e.error_diagnostics[0].message)
...
Unable to parse operation assembly
[<mlir._mlir_libs._mlir.ir.DiagnosticInfo object at 0x7fed32bd6b70>]
'arith.addi' op requires one result
```
Error diagnostics captured in exceptions aren't propagated to diagnostic
handlers, to avoid double-reporting of errors. The context-level
`emit_error_diagnostics` option can be used to revert to the old
behaviour, causing error diagnostics to be reported to handlers instead
of as part of exceptions.
API changes:
- `Operation.verify` now raises an exception on verification failure,
instead of returning `false`
- The exception raised by the following methods has been changed to
`MLIRError`:
- `PassManager.run`
- `{Module,Operation,Type,Attribute}.parse`
- `{RankedTensorType,UnrankedTensorType}.get`
- `{MemRefType,UnrankedMemRefType}.get`
- `VectorType.get`
- `FloatAttr.get`
closes #60595
depends on D144804, D143830
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D143869
2023-02-07 16:07:50 -05:00
|
|
|
if (mlirTypeIsNull(type))
|
|
|
|
|
throw MLIRError("Unable to parse type", errors.take());
|
2023-05-26 10:23:17 -05:00
|
|
|
return type;
|
2020-10-31 23:40:25 -07:00
|
|
|
},
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("asm"), nb::arg("context").none() = nb::none(),
|
2020-10-31 23:40:25 -07:00
|
|
|
kContextParseTypeDocstring)
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro(
|
2020-10-12 21:19:13 -07:00
|
|
|
"context", [](PyType &self) { return self.getContext().getObject(); },
|
|
|
|
|
"Context that owns the Type")
|
2020-11-01 23:05:36 -08:00
|
|
|
.def("__eq__", [](PyType &self, PyType &other) { return self == other; })
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def(
|
|
|
|
|
"__eq__", [](PyType &self, nb::object &other) { return false; },
|
|
|
|
|
nb::arg("other").none())
|
2021-11-02 17:04:42 +01:00
|
|
|
.def("__hash__",
|
|
|
|
|
[](PyType &self) {
|
|
|
|
|
return static_cast<size_t>(llvm::hash_value(self.get().ptr));
|
|
|
|
|
})
|
2020-08-18 17:23:46 -07:00
|
|
|
.def(
|
2020-11-29 13:30:23 -08:00
|
|
|
"dump", [](PyType &self) { mlirTypeDump(self); }, kDumpDocstring)
|
2020-08-18 17:23:46 -07:00
|
|
|
.def(
|
|
|
|
|
"__str__",
|
|
|
|
|
[](PyType &self) {
|
|
|
|
|
PyPrintAccumulator printAccum;
|
2020-11-29 13:30:23 -08:00
|
|
|
mlirTypePrint(self, printAccum.getCallback(),
|
2020-08-18 17:23:46 -07:00
|
|
|
printAccum.getUserData());
|
|
|
|
|
return printAccum.join();
|
|
|
|
|
},
|
2020-10-20 11:22:29 +02:00
|
|
|
"Returns the assembly form of the type.")
|
2023-05-22 11:12:53 -05:00
|
|
|
.def("__repr__",
|
|
|
|
|
[](PyType &self) {
|
|
|
|
|
// Generally, assembly formats are not printed for __repr__ because
|
|
|
|
|
// this can cause exceptionally long debug output and exceptions.
|
|
|
|
|
// However, types are an exception as they typically have compact
|
|
|
|
|
// assembly forms and printing them is useful.
|
|
|
|
|
PyPrintAccumulator printAccum;
|
|
|
|
|
printAccum.parts.append("Type(");
|
|
|
|
|
mlirTypePrint(self, printAccum.getCallback(),
|
|
|
|
|
printAccum.getUserData());
|
|
|
|
|
printAccum.parts.append(")");
|
|
|
|
|
return printAccum.join();
|
|
|
|
|
})
|
2023-05-26 10:23:17 -05:00
|
|
|
.def(MLIR_PYTHON_MAYBE_DOWNCAST_ATTR,
|
|
|
|
|
[](PyType &self) {
|
|
|
|
|
MlirTypeID mlirTypeID = mlirTypeGetTypeID(self);
|
|
|
|
|
assert(!mlirTypeIDIsNull(mlirTypeID) &&
|
|
|
|
|
"mlirTypeID was expected to be non-null.");
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
std::optional<nb::callable> typeCaster =
|
2023-05-26 10:23:17 -05:00
|
|
|
PyGlobals::get().lookupTypeCaster(mlirTypeID,
|
|
|
|
|
mlirTypeGetDialect(self));
|
|
|
|
|
if (!typeCaster)
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
return nb::cast(self);
|
2023-05-26 10:23:17 -05:00
|
|
|
return typeCaster.value()(self);
|
|
|
|
|
})
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro("typeid", [](PyType &self) -> MlirTypeID {
|
2023-05-22 11:12:53 -05:00
|
|
|
MlirTypeID mlirTypeID = mlirTypeGetTypeID(self);
|
|
|
|
|
if (!mlirTypeIDIsNull(mlirTypeID))
|
|
|
|
|
return mlirTypeID;
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
auto origRepr = nb::cast<std::string>(nb::repr(nb::cast(self)));
|
|
|
|
|
throw nb::value_error(
|
|
|
|
|
(origRepr + llvm::Twine(" has no typeid.")).str().c_str());
|
2023-05-22 11:12:53 -05:00
|
|
|
});
|
|
|
|
|
|
|
|
|
|
//----------------------------------------------------------------------------
|
|
|
|
|
// Mapping of PyTypeID.
|
|
|
|
|
//----------------------------------------------------------------------------
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::class_<PyTypeID>(m, "TypeID")
|
|
|
|
|
.def_prop_ro(MLIR_PYTHON_CAPI_PTR_ATTR, &PyTypeID::getCapsule)
|
2023-05-22 11:12:53 -05:00
|
|
|
.def(MLIR_PYTHON_CAPI_FACTORY_ATTR, &PyTypeID::createFromCapsule)
|
|
|
|
|
// Note, this tests whether the underlying TypeIDs are the same,
|
|
|
|
|
// not whether the wrapper MlirTypeIDs are the same, nor whether
|
|
|
|
|
// the Python objects are the same (i.e., PyTypeID is a value type).
|
|
|
|
|
.def("__eq__",
|
|
|
|
|
[](PyTypeID &self, PyTypeID &other) { return self == other; })
|
|
|
|
|
.def("__eq__",
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
[](PyTypeID &self, const nb::object &other) { return false; })
|
2023-05-22 11:12:53 -05:00
|
|
|
// Note, this gives the hash value of the underlying TypeID, not the
|
|
|
|
|
// hash value of the Python object, nor the hash value of the
|
|
|
|
|
// MlirTypeID wrapper.
|
|
|
|
|
.def("__hash__", [](PyTypeID &self) {
|
|
|
|
|
return static_cast<size_t>(mlirTypeIDHashValue(self));
|
2020-08-18 17:23:46 -07:00
|
|
|
});
|
|
|
|
|
|
2020-10-21 23:34:01 -07:00
|
|
|
//----------------------------------------------------------------------------
|
2020-10-20 11:21:05 +02:00
|
|
|
// Mapping of Value.
|
2020-10-21 23:34:01 -07:00
|
|
|
//----------------------------------------------------------------------------
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::class_<PyValue>(m, "Value")
|
|
|
|
|
.def(nb::init<PyValue &>(), nb::keep_alive<0, 1>(), nb::arg("value"))
|
|
|
|
|
.def_prop_ro(MLIR_PYTHON_CAPI_PTR_ATTR, &PyValue::getCapsule)
|
2021-04-22 00:07:30 -06:00
|
|
|
.def(MLIR_PYTHON_CAPI_FACTORY_ATTR, &PyValue::createFromCapsule)
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro(
|
2020-10-20 11:21:05 +02:00
|
|
|
"context",
|
|
|
|
|
[](PyValue &self) { return self.getParentOperation()->getContext(); },
|
|
|
|
|
"Context in which the value lives.")
|
|
|
|
|
.def(
|
|
|
|
|
"dump", [](PyValue &self) { mlirValueDump(self.get()); },
|
|
|
|
|
kDumpDocstring)
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro(
|
2021-07-14 20:19:27 -07:00
|
|
|
"owner",
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
[](PyValue &self) -> nb::object {
|
2022-08-09 19:37:04 -07:00
|
|
|
MlirValue v = self.get();
|
|
|
|
|
if (mlirValueIsAOpResult(v)) {
|
|
|
|
|
assert(
|
|
|
|
|
mlirOperationEqual(self.getParentOperation()->get(),
|
|
|
|
|
mlirOpResultGetOwner(self.get())) &&
|
|
|
|
|
"expected the owner of the value in Python to match that in "
|
|
|
|
|
"the IR");
|
|
|
|
|
return self.getParentOperation().getObject();
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (mlirValueIsABlockArgument(v)) {
|
|
|
|
|
MlirBlock block = mlirBlockArgumentGetOwner(self.get());
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
return nb::cast(PyBlock(self.getParentOperation(), block));
|
2022-08-09 19:37:04 -07:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
assert(false && "Value must be a block argument or an op result");
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
return nb::none();
|
2021-07-14 20:19:27 -07:00
|
|
|
})
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def_prop_ro("uses",
|
|
|
|
|
[](PyValue &self) {
|
|
|
|
|
return PyOpOperandIterator(
|
|
|
|
|
mlirValueGetFirstUse(self.get()));
|
|
|
|
|
})
|
2020-11-06 15:09:54 +01:00
|
|
|
.def("__eq__",
|
|
|
|
|
[](PyValue &self, PyValue &other) {
|
|
|
|
|
return self.get().ptr == other.get().ptr;
|
|
|
|
|
})
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def("__eq__", [](PyValue &self, nb::object other) { return false; })
|
2021-11-02 17:04:42 +01:00
|
|
|
.def("__hash__",
|
|
|
|
|
[](PyValue &self) {
|
|
|
|
|
return static_cast<size_t>(llvm::hash_value(self.get().ptr));
|
|
|
|
|
})
|
2020-10-20 11:21:05 +02:00
|
|
|
.def(
|
|
|
|
|
"__str__",
|
|
|
|
|
[](PyValue &self) {
|
|
|
|
|
PyPrintAccumulator printAccum;
|
|
|
|
|
printAccum.parts.append("Value(");
|
|
|
|
|
mlirValuePrint(self.get(), printAccum.getCallback(),
|
|
|
|
|
printAccum.getUserData());
|
|
|
|
|
printAccum.parts.append(")");
|
|
|
|
|
return printAccum.join();
|
|
|
|
|
},
|
|
|
|
|
kValueDunderStrDocstring)
|
2023-05-07 18:19:46 -05:00
|
|
|
.def(
|
|
|
|
|
"get_name",
|
2023-10-02 21:17:49 -07:00
|
|
|
[](PyValue &self, bool useLocalScope) {
|
2023-05-07 18:19:46 -05:00
|
|
|
PyPrintAccumulator printAccum;
|
2023-10-02 21:17:49 -07:00
|
|
|
MlirOpPrintingFlags flags = mlirOpPrintingFlagsCreate();
|
|
|
|
|
if (useLocalScope)
|
|
|
|
|
mlirOpPrintingFlagsUseLocalScope(flags);
|
|
|
|
|
MlirAsmState valueState =
|
|
|
|
|
mlirAsmStateCreateForValue(self.get(), flags);
|
|
|
|
|
mlirValuePrintAsOperand(self.get(), valueState,
|
|
|
|
|
printAccum.getCallback(),
|
|
|
|
|
printAccum.getUserData());
|
|
|
|
|
mlirOpPrintingFlagsDestroy(flags);
|
|
|
|
|
mlirAsmStateDestroy(valueState);
|
|
|
|
|
return printAccum.join();
|
|
|
|
|
},
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("use_local_scope") = false)
|
2023-10-02 21:17:49 -07:00
|
|
|
.def(
|
|
|
|
|
"get_name",
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
[](PyValue &self, PyAsmState &state) {
|
2023-10-02 21:17:49 -07:00
|
|
|
PyPrintAccumulator printAccum;
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
MlirAsmState valueState = state.get();
|
2023-10-02 15:37:25 -05:00
|
|
|
mlirValuePrintAsOperand(self.get(), valueState,
|
|
|
|
|
printAccum.getCallback(),
|
2023-05-07 18:19:46 -05:00
|
|
|
printAccum.getUserData());
|
|
|
|
|
return printAccum.join();
|
|
|
|
|
},
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("state"), kGetNameAsOperand)
|
|
|
|
|
.def_prop_ro("type",
|
|
|
|
|
[](PyValue &self) { return mlirValueGetType(self.get()); })
|
2023-07-25 23:17:00 -05:00
|
|
|
.def(
|
|
|
|
|
"set_type",
|
|
|
|
|
[](PyValue &self, const PyType &type) {
|
|
|
|
|
return mlirValueSetType(self.get(), type);
|
|
|
|
|
},
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("type"))
|
2023-04-26 09:55:27 -05:00
|
|
|
.def(
|
|
|
|
|
"replace_all_uses_with",
|
|
|
|
|
[](PyValue &self, PyValue &with) {
|
|
|
|
|
mlirValueReplaceAllUsesOfWith(self.get(), with.get());
|
|
|
|
|
},
|
[mlir][python] value casting (#69644)
This PR adds "value casting", i.e., a mechanism to wrap `ir.Value` in a
proxy class that overloads dunders such as `__add__`, `__sub__`, and
`__mul__` for fun and great profit.
This is thematically similar to
https://github.com/llvm/llvm-project/commit/bfb1ba752655bf09b35c486f6cc9817dbedfb1bb
and
https://github.com/llvm/llvm-project/commit/9566ee280607d91fa2e5eca730a6765ac84dfd0f.
The example in the test demonstrates the value of the feature (no pun
intended):
```python
@register_value_caster(F16Type.static_typeid)
@register_value_caster(F32Type.static_typeid)
@register_value_caster(F64Type.static_typeid)
@register_value_caster(IntegerType.static_typeid)
class ArithValue(Value):
__add__ = partialmethod(_binary_op, op="add")
__sub__ = partialmethod(_binary_op, op="sub")
__mul__ = partialmethod(_binary_op, op="mul")
a = arith.constant(value=FloatAttr.get(f16_t, 42.42))
b = a + a
# CHECK: ArithValue(%0 = arith.addf %cst, %cst : f16)
print(b)
a = arith.constant(value=FloatAttr.get(f32_t, 42.42))
b = a - a
# CHECK: ArithValue(%1 = arith.subf %cst_0, %cst_0 : f32)
print(b)
a = arith.constant(value=FloatAttr.get(f64_t, 42.42))
b = a * a
# CHECK: ArithValue(%2 = arith.mulf %cst_1, %cst_1 : f64)
print(b)
```
**EDIT**: this now goes through the bindings and thus supports automatic
casting of `OpResult` (including as an element of `OpResultList`),
`BlockArgument` (including as an element of `BlockArgumentList`), as
well as `Value`.
2023-11-07 10:49:41 -06:00
|
|
|
kValueReplaceAllUsesWithDocstring)
|
[mlir,python] Expose replaceAllUsesExcept to Python bindings (#115850)
Problem originally described in [the forums
here](https://discourse.llvm.org/t/mlir-python-expose-replaceallusesexcept/83068/1).
Using the MLIR Python bindings, the method
[`replaceAllUsesWith`](https://mlir.llvm.org/doxygen/classmlir_1_1Value.html#ac56b0fdb6246bcf7fa1805ba0eb71aa2)
for `Value` is exposed, e.g.,
```python
orig_value.replace_all_uses_with(
new_value
)
```
However, in my use-case I am separating a block into multiple blocks, so
thus want to exclude certain Operations from having their Values
replaced (since I want them to diverge).
Within Value, we have
[`replaceAllUsesExcept`](https://mlir.llvm.org/doxygen/classmlir_1_1Value.html#a9ec8d5c61f8a6aada4062f609372cce4),
where we can pass the Operations which should be skipped.
This is not currently exposed in the Python bindings: this PR fixes
this. Adds `replace_all_uses_except`, which works with individual
Operations, and lists of Operations.
2024-11-20 00:00:57 +00:00
|
|
|
.def(
|
|
|
|
|
"replace_all_uses_except",
|
|
|
|
|
[](MlirValue self, MlirValue with, PyOperation &exception) {
|
|
|
|
|
MlirOperation exceptedUser = exception.get();
|
|
|
|
|
mlirValueReplaceAllUsesExcept(self, with, 1, &exceptedUser);
|
|
|
|
|
},
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("with"), nb::arg("exceptions"),
|
[mlir,python] Expose replaceAllUsesExcept to Python bindings (#115850)
Problem originally described in [the forums
here](https://discourse.llvm.org/t/mlir-python-expose-replaceallusesexcept/83068/1).
Using the MLIR Python bindings, the method
[`replaceAllUsesWith`](https://mlir.llvm.org/doxygen/classmlir_1_1Value.html#ac56b0fdb6246bcf7fa1805ba0eb71aa2)
for `Value` is exposed, e.g.,
```python
orig_value.replace_all_uses_with(
new_value
)
```
However, in my use-case I am separating a block into multiple blocks, so
thus want to exclude certain Operations from having their Values
replaced (since I want them to diverge).
Within Value, we have
[`replaceAllUsesExcept`](https://mlir.llvm.org/doxygen/classmlir_1_1Value.html#a9ec8d5c61f8a6aada4062f609372cce4),
where we can pass the Operations which should be skipped.
This is not currently exposed in the Python bindings: this PR fixes
this. Adds `replace_all_uses_except`, which works with individual
Operations, and lists of Operations.
2024-11-20 00:00:57 +00:00
|
|
|
kValueReplaceAllUsesExceptDocstring)
|
|
|
|
|
.def(
|
|
|
|
|
"replace_all_uses_except",
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
[](MlirValue self, MlirValue with, nb::list exceptions) {
|
[mlir,python] Expose replaceAllUsesExcept to Python bindings (#115850)
Problem originally described in [the forums
here](https://discourse.llvm.org/t/mlir-python-expose-replaceallusesexcept/83068/1).
Using the MLIR Python bindings, the method
[`replaceAllUsesWith`](https://mlir.llvm.org/doxygen/classmlir_1_1Value.html#ac56b0fdb6246bcf7fa1805ba0eb71aa2)
for `Value` is exposed, e.g.,
```python
orig_value.replace_all_uses_with(
new_value
)
```
However, in my use-case I am separating a block into multiple blocks, so
thus want to exclude certain Operations from having their Values
replaced (since I want them to diverge).
Within Value, we have
[`replaceAllUsesExcept`](https://mlir.llvm.org/doxygen/classmlir_1_1Value.html#a9ec8d5c61f8a6aada4062f609372cce4),
where we can pass the Operations which should be skipped.
This is not currently exposed in the Python bindings: this PR fixes
this. Adds `replace_all_uses_except`, which works with individual
Operations, and lists of Operations.
2024-11-20 00:00:57 +00:00
|
|
|
// Convert Python list to a SmallVector of MlirOperations
|
|
|
|
|
llvm::SmallVector<MlirOperation> exceptionOps;
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
for (nb::handle exception : exceptions) {
|
|
|
|
|
exceptionOps.push_back(nb::cast<PyOperation &>(exception).get());
|
[mlir,python] Expose replaceAllUsesExcept to Python bindings (#115850)
Problem originally described in [the forums
here](https://discourse.llvm.org/t/mlir-python-expose-replaceallusesexcept/83068/1).
Using the MLIR Python bindings, the method
[`replaceAllUsesWith`](https://mlir.llvm.org/doxygen/classmlir_1_1Value.html#ac56b0fdb6246bcf7fa1805ba0eb71aa2)
for `Value` is exposed, e.g.,
```python
orig_value.replace_all_uses_with(
new_value
)
```
However, in my use-case I am separating a block into multiple blocks, so
thus want to exclude certain Operations from having their Values
replaced (since I want them to diverge).
Within Value, we have
[`replaceAllUsesExcept`](https://mlir.llvm.org/doxygen/classmlir_1_1Value.html#a9ec8d5c61f8a6aada4062f609372cce4),
where we can pass the Operations which should be skipped.
This is not currently exposed in the Python bindings: this PR fixes
this. Adds `replace_all_uses_except`, which works with individual
Operations, and lists of Operations.
2024-11-20 00:00:57 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
mlirValueReplaceAllUsesExcept(
|
|
|
|
|
self, with, static_cast<intptr_t>(exceptionOps.size()),
|
|
|
|
|
exceptionOps.data());
|
|
|
|
|
},
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("with"), nb::arg("exceptions"),
|
[mlir,python] Expose replaceAllUsesExcept to Python bindings (#115850)
Problem originally described in [the forums
here](https://discourse.llvm.org/t/mlir-python-expose-replaceallusesexcept/83068/1).
Using the MLIR Python bindings, the method
[`replaceAllUsesWith`](https://mlir.llvm.org/doxygen/classmlir_1_1Value.html#ac56b0fdb6246bcf7fa1805ba0eb71aa2)
for `Value` is exposed, e.g.,
```python
orig_value.replace_all_uses_with(
new_value
)
```
However, in my use-case I am separating a block into multiple blocks, so
thus want to exclude certain Operations from having their Values
replaced (since I want them to diverge).
Within Value, we have
[`replaceAllUsesExcept`](https://mlir.llvm.org/doxygen/classmlir_1_1Value.html#a9ec8d5c61f8a6aada4062f609372cce4),
where we can pass the Operations which should be skipped.
This is not currently exposed in the Python bindings: this PR fixes
this. Adds `replace_all_uses_except`, which works with individual
Operations, and lists of Operations.
2024-11-20 00:00:57 +00:00
|
|
|
kValueReplaceAllUsesExceptDocstring)
|
[mlir][python] value casting (#69644)
This PR adds "value casting", i.e., a mechanism to wrap `ir.Value` in a
proxy class that overloads dunders such as `__add__`, `__sub__`, and
`__mul__` for fun and great profit.
This is thematically similar to
https://github.com/llvm/llvm-project/commit/bfb1ba752655bf09b35c486f6cc9817dbedfb1bb
and
https://github.com/llvm/llvm-project/commit/9566ee280607d91fa2e5eca730a6765ac84dfd0f.
The example in the test demonstrates the value of the feature (no pun
intended):
```python
@register_value_caster(F16Type.static_typeid)
@register_value_caster(F32Type.static_typeid)
@register_value_caster(F64Type.static_typeid)
@register_value_caster(IntegerType.static_typeid)
class ArithValue(Value):
__add__ = partialmethod(_binary_op, op="add")
__sub__ = partialmethod(_binary_op, op="sub")
__mul__ = partialmethod(_binary_op, op="mul")
a = arith.constant(value=FloatAttr.get(f16_t, 42.42))
b = a + a
# CHECK: ArithValue(%0 = arith.addf %cst, %cst : f16)
print(b)
a = arith.constant(value=FloatAttr.get(f32_t, 42.42))
b = a - a
# CHECK: ArithValue(%1 = arith.subf %cst_0, %cst_0 : f32)
print(b)
a = arith.constant(value=FloatAttr.get(f64_t, 42.42))
b = a * a
# CHECK: ArithValue(%2 = arith.mulf %cst_1, %cst_1 : f64)
print(b)
```
**EDIT**: this now goes through the bindings and thus supports automatic
casting of `OpResult` (including as an element of `OpResultList`),
`BlockArgument` (including as an element of `BlockArgumentList`), as
well as `Value`.
2023-11-07 10:49:41 -06:00
|
|
|
.def(MLIR_PYTHON_MAYBE_DOWNCAST_ATTR,
|
|
|
|
|
[](PyValue &self) { return self.maybeDownCast(); });
|
2020-10-20 11:21:05 +02:00
|
|
|
PyBlockArgument::bind(m);
|
|
|
|
|
PyOpResult::bind(m);
|
2022-12-06 19:30:56 -07:00
|
|
|
PyOpOperand::bind(m);
|
2020-10-20 11:21:05 +02:00
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::class_<PyAsmState>(m, "AsmState")
|
|
|
|
|
.def(nb::init<PyValue &, bool>(), nb::arg("value"),
|
|
|
|
|
nb::arg("use_local_scope") = false)
|
|
|
|
|
.def(nb::init<PyOperationBase &, bool>(), nb::arg("op"),
|
|
|
|
|
nb::arg("use_local_scope") = false);
|
2023-09-20 15:12:06 -07:00
|
|
|
|
2021-11-02 12:39:36 +01:00
|
|
|
//----------------------------------------------------------------------------
|
|
|
|
|
// Mapping of SymbolTable.
|
|
|
|
|
//----------------------------------------------------------------------------
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::class_<PySymbolTable>(m, "SymbolTable")
|
|
|
|
|
.def(nb::init<PyOperationBase &>())
|
2021-11-02 12:39:36 +01:00
|
|
|
.def("__getitem__", &PySymbolTable::dunderGetItem)
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
.def("insert", &PySymbolTable::insert, nb::arg("operation"))
|
|
|
|
|
.def("erase", &PySymbolTable::erase, nb::arg("operation"))
|
2021-11-02 12:39:36 +01:00
|
|
|
.def("__delitem__", &PySymbolTable::dunderDel)
|
2021-11-28 20:30:18 -08:00
|
|
|
.def("__contains__",
|
|
|
|
|
[](PySymbolTable &table, const std::string &name) {
|
|
|
|
|
return !mlirOperationIsNull(mlirSymbolTableLookup(
|
|
|
|
|
table, mlirStringRefCreate(name.data(), name.length())));
|
|
|
|
|
})
|
|
|
|
|
// Static helpers.
|
|
|
|
|
.def_static("set_symbol_name", &PySymbolTable::setSymbolName,
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("symbol"), nb::arg("name"))
|
2021-11-28 20:30:18 -08:00
|
|
|
.def_static("get_symbol_name", &PySymbolTable::getSymbolName,
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("symbol"))
|
2021-11-28 20:30:18 -08:00
|
|
|
.def_static("get_visibility", &PySymbolTable::getVisibility,
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("symbol"))
|
2021-11-28 20:30:18 -08:00
|
|
|
.def_static("set_visibility", &PySymbolTable::setVisibility,
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("symbol"), nb::arg("visibility"))
|
2021-11-28 20:30:18 -08:00
|
|
|
.def_static("replace_all_symbol_uses",
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
&PySymbolTable::replaceAllSymbolUses, nb::arg("old_symbol"),
|
|
|
|
|
nb::arg("new_symbol"), nb::arg("from_op"))
|
2021-11-28 20:30:18 -08:00
|
|
|
.def_static("walk_symbol_tables", &PySymbolTable::walkSymbolTables,
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::arg("from_op"), nb::arg("all_sym_uses_visible"),
|
|
|
|
|
nb::arg("callback"));
|
2021-11-02 12:39:36 +01:00
|
|
|
|
2020-09-19 22:02:32 -07:00
|
|
|
// Container bindings.
|
2020-10-20 11:21:05 +02:00
|
|
|
PyBlockArgumentList::bind(m);
|
2020-09-19 22:02:32 -07:00
|
|
|
PyBlockIterator::bind(m);
|
|
|
|
|
PyBlockList::bind(m);
|
|
|
|
|
PyOperationIterator::bind(m);
|
|
|
|
|
PyOperationList::bind(m);
|
2020-11-06 11:25:41 +01:00
|
|
|
PyOpAttributeMap::bind(m);
|
2022-12-06 19:30:56 -07:00
|
|
|
PyOpOperandIterator::bind(m);
|
2020-10-21 23:34:01 -07:00
|
|
|
PyOpOperandList::bind(m);
|
2020-10-20 11:21:05 +02:00
|
|
|
PyOpResultList::bind(m);
|
2023-10-02 15:37:25 -05:00
|
|
|
PyOpSuccessors::bind(m);
|
2020-09-19 22:02:32 -07:00
|
|
|
PyRegionIterator::bind(m);
|
|
|
|
|
PyRegionList::bind(m);
|
2021-04-19 13:37:01 +02:00
|
|
|
|
|
|
|
|
// Debug bindings.
|
|
|
|
|
PyGlobalDebugFlag::bind(m);
|
2022-12-21 16:22:39 -08:00
|
|
|
|
|
|
|
|
// Attribute builder getter.
|
|
|
|
|
PyAttrBuilderMap::bind(m);
|
[mlir][python] Capture error diagnostics in exceptions
This updates most (all?) error-diagnostic-emitting python APIs to
capture error diagnostics and include them in the raised exception's
message:
```
>>> Operation.parse('"arith.addi"() : () -> ()'))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
mlir._mlir_libs.MLIRError: Unable to parse operation assembly:
error: "-":1:1: 'arith.addi' op requires one result
note: "-":1:1: see current operation: "arith.addi"() : () -> ()
```
The diagnostic information is available on the exception for users who
may want to customize the error message:
```
>>> try:
... Operation.parse('"arith.addi"() : () -> ()')
... except MLIRError as e:
... print(e.message)
... print(e.error_diagnostics)
... print(e.error_diagnostics[0].message)
...
Unable to parse operation assembly
[<mlir._mlir_libs._mlir.ir.DiagnosticInfo object at 0x7fed32bd6b70>]
'arith.addi' op requires one result
```
Error diagnostics captured in exceptions aren't propagated to diagnostic
handlers, to avoid double-reporting of errors. The context-level
`emit_error_diagnostics` option can be used to revert to the old
behaviour, causing error diagnostics to be reported to handlers instead
of as part of exceptions.
API changes:
- `Operation.verify` now raises an exception on verification failure,
instead of returning `false`
- The exception raised by the following methods has been changed to
`MLIRError`:
- `PassManager.run`
- `{Module,Operation,Type,Attribute}.parse`
- `{RankedTensorType,UnrankedTensorType}.get`
- `{MemRefType,UnrankedMemRefType}.get`
- `VectorType.get`
- `FloatAttr.get`
closes #60595
depends on D144804, D143830
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D143869
2023-02-07 16:07:50 -05:00
|
|
|
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::register_exception_translator([](const std::exception_ptr &p,
|
|
|
|
|
void *payload) {
|
[mlir][python] Capture error diagnostics in exceptions
This updates most (all?) error-diagnostic-emitting python APIs to
capture error diagnostics and include them in the raised exception's
message:
```
>>> Operation.parse('"arith.addi"() : () -> ()'))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
mlir._mlir_libs.MLIRError: Unable to parse operation assembly:
error: "-":1:1: 'arith.addi' op requires one result
note: "-":1:1: see current operation: "arith.addi"() : () -> ()
```
The diagnostic information is available on the exception for users who
may want to customize the error message:
```
>>> try:
... Operation.parse('"arith.addi"() : () -> ()')
... except MLIRError as e:
... print(e.message)
... print(e.error_diagnostics)
... print(e.error_diagnostics[0].message)
...
Unable to parse operation assembly
[<mlir._mlir_libs._mlir.ir.DiagnosticInfo object at 0x7fed32bd6b70>]
'arith.addi' op requires one result
```
Error diagnostics captured in exceptions aren't propagated to diagnostic
handlers, to avoid double-reporting of errors. The context-level
`emit_error_diagnostics` option can be used to revert to the old
behaviour, causing error diagnostics to be reported to handlers instead
of as part of exceptions.
API changes:
- `Operation.verify` now raises an exception on verification failure,
instead of returning `false`
- The exception raised by the following methods has been changed to
`MLIRError`:
- `PassManager.run`
- `{Module,Operation,Type,Attribute}.parse`
- `{RankedTensorType,UnrankedTensorType}.get`
- `{MemRefType,UnrankedMemRefType}.get`
- `VectorType.get`
- `FloatAttr.get`
closes #60595
depends on D144804, D143830
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D143869
2023-02-07 16:07:50 -05:00
|
|
|
// We can't define exceptions with custom fields through pybind, so instead
|
|
|
|
|
// the exception class is defined in python and imported here.
|
|
|
|
|
try {
|
|
|
|
|
if (p)
|
|
|
|
|
std::rethrow_exception(p);
|
|
|
|
|
} catch (const MLIRError &e) {
|
[mlir python] Port Python core code to nanobind. (#120473)
Relands #118583, with a fix for Python 3.8 compatibility. It was not
possible to set the buffer protocol accessers via slots in Python 3.8.
Why? https://nanobind.readthedocs.io/en/latest/why.html says it better
than I can, but my primary motivation for this change is to improve MLIR
IR construction time from JAX.
For a complicated Google-internal LLM model in JAX, this change improves
the MLIR
lowering time by around 5s (out of around 30s), which is a significant
speedup for simply switching binding frameworks.
To a large extent, this is a mechanical change, for instance changing
`pybind11::` to `nanobind::`.
Notes:
* this PR needs Nanobind 2.4.0, because it needs a bug fix
(https://github.com/wjakob/nanobind/pull/806) that landed in that
release.
* this PR does not port the in-tree dialect extension modules. They can
be ported in a future PR.
* I removed the py::sibling() annotations from def_static and def_class
in `PybindAdapters.h`. These ask pybind11 to try to form an overload
with an existing method, but it's not possible to form mixed
pybind11/nanobind overloads this ways and the parent class is now
defined in nanobind. Better solutions may be possible here.
* nanobind does not contain an exact equivalent of pybind11's buffer
protocol support. It was not hard to add a nanobind implementation of a
similar API.
* nanobind is pickier about casting to std::vector<bool>, expecting that
the input is a sequence of bool types, not truthy values. In a couple of
places I added code to support truthy values during casting.
* nanobind distinguishes bytes (`nb::bytes`) from strings (e.g.,
`std::string`). This required nb::bytes overloads in a few places.
2024-12-18 21:55:42 -05:00
|
|
|
nb::object obj = nb::module_::import_(MAKE_MLIR_PYTHON_QUALNAME("ir"))
|
[mlir][python] Capture error diagnostics in exceptions
This updates most (all?) error-diagnostic-emitting python APIs to
capture error diagnostics and include them in the raised exception's
message:
```
>>> Operation.parse('"arith.addi"() : () -> ()'))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
mlir._mlir_libs.MLIRError: Unable to parse operation assembly:
error: "-":1:1: 'arith.addi' op requires one result
note: "-":1:1: see current operation: "arith.addi"() : () -> ()
```
The diagnostic information is available on the exception for users who
may want to customize the error message:
```
>>> try:
... Operation.parse('"arith.addi"() : () -> ()')
... except MLIRError as e:
... print(e.message)
... print(e.error_diagnostics)
... print(e.error_diagnostics[0].message)
...
Unable to parse operation assembly
[<mlir._mlir_libs._mlir.ir.DiagnosticInfo object at 0x7fed32bd6b70>]
'arith.addi' op requires one result
```
Error diagnostics captured in exceptions aren't propagated to diagnostic
handlers, to avoid double-reporting of errors. The context-level
`emit_error_diagnostics` option can be used to revert to the old
behaviour, causing error diagnostics to be reported to handlers instead
of as part of exceptions.
API changes:
- `Operation.verify` now raises an exception on verification failure,
instead of returning `false`
- The exception raised by the following methods has been changed to
`MLIRError`:
- `PassManager.run`
- `{Module,Operation,Type,Attribute}.parse`
- `{RankedTensorType,UnrankedTensorType}.get`
- `{MemRefType,UnrankedMemRefType}.get`
- `VectorType.get`
- `FloatAttr.get`
closes #60595
depends on D144804, D143830
Reviewed By: stellaraccident
Differential Revision: https://reviews.llvm.org/D143869
2023-02-07 16:07:50 -05:00
|
|
|
.attr("MLIRError")(e.message, e.errorDiagnostics);
|
|
|
|
|
PyErr_SetObject(PyExc_Exception, obj.ptr());
|
|
|
|
|
}
|
|
|
|
|
});
|
2020-08-16 18:49:28 -07:00
|
|
|
}
|