intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-03 06:49:52 +08:00
Code Issues Packages Projects Releases Wiki Activity
21,349 Commits 43 Branches 655 Tags
a52f526929c4fa3cef40fa9709dc4eba99fe04ff
Commit Graph

6173 Commits

Author SHA1 Message Date
Zbigniew Zdanowicz
42d2c32795 fix: disable state compute mode thread arbitration on xe3 platforms 
Related-To: NEO-16330

Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>
 2025-10-07 15:48:42 +02:00
Mateusz Jablonski
f3e4ba9ff6 refactor: remove not needed code 
Signed-off-by: Mateusz Jablonski <mateusz.jablonski@intel.com>
 2025-10-07 13:44:30 +02:00
Dominik Dabek
17199556ca performance: lazy start usm reuse cleaner thread 
Delay starting usm reuse cleaner thread to first allocation reuse.

Related-To: HSD-16028754950

Signed-off-by: Dominik Dabek <dominik.dabek@intel.com>
 2025-10-07 13:29:10 +02:00
Zbigniew Zdanowicz
24db09e3e9 fix: make pipeline thread arbitration one-time switch property 
- detect switch happens by making "is dirty" property
- make sure to call the property setter
- setting to always true will make "is dirty" true only once

Related-To: NEO-16330

Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>
 2025-10-07 11:27:22 +02:00
Damian Tomczak
bb742a88a4 feature: fail stateful kernel when stateless is required 
Resolves: NEO-11859

Signed-off-by: Damian Tomczak <damian.tomczak@intel.com>
 2025-10-07 10:49:26 +02:00
Mateusz Hoppe
f411dc2e65 fix: debug key for using temp memory when writing to aub 
Related-To: NEO-16276

Signed-off-by: Mateusz Hoppe <mateusz.hoppe@intel.com>
 2025-10-07 09:43:02 +02:00
Compute-Runtime-Validation
bafb847c73 Revert "fix: use condition variables instead of busy waits in worker threads" 
This reverts commit 9d1da44e08.

Signed-off-by: Compute-Runtime-Validation <compute-runtime-validation@intel.com>
 2025-10-06 20:38:46 +02:00
Szymon Morek
4df9783448 performance: use 2MB heaps for 2MB-aligned allocations 
Related-To: NEO-16165

Signed-off-by: Szymon Morek <szymon.morek@intel.com>
 2025-10-06 18:11:01 +02:00
Mateusz Jablonski
9ca3a9b5d8 build: remove not needed test kernels 
Signed-off-by: Mateusz Jablonski <mateusz.jablonski@intel.com>
 2025-10-06 17:06:59 +02:00
Radoslaw Jablonski
54de14a9dc refactor: apply clang-format's InsertBraces rule 
Formats code before InsertBraces rule is enabled.

Signed-off-by: Radoslaw Jablonski <radoslaw.jablonski@intel.com>
 2025-10-06 15:32:46 +02:00
Igor Venevtsev
9d1da44e08 fix: use condition variables instead of busy waits in worker threads 
Resolves: NEO-16085, GSD-11678, HSD-14025819208

Signed-off-by: Igor Venevtsev <igor.venevtsev@intel.com>
 2025-10-06 13:27:34 +02:00
Dominik Dabek
8575573dbd fix: l0, bindless image from pooled alloc 
Make sure pool offset is applied when growing pools are used.

Related-To: NEO-16317, GSD-11804

Signed-off-by: Dominik Dabek <dominik.dabek@intel.com>
 2025-10-06 10:21:05 +02:00
Mateusz Jablonski
6c0c229c82 refactor: remove not needed flags to force OCL version support 
Signed-off-by: Mateusz Jablonski <mateusz.jablonski@intel.com>
 2025-10-03 16:57:08 +02:00
Katarzyna Cencelewska
0f866e24e6 fix: enable setting timestamp via submission on tbx mode 
Resolves: NEO-16293
Signed-off-by: Katarzyna Cencelewska <katarzyna.cencelewska@intel.com>
 2025-10-03 08:43:18 +02:00
Mateusz Jablonski
69eeaee0eb refactor: remove not needed function - isDotIntegerProductExtensionSupported 
the function always returns true

Signed-off-by: Mateusz Jablonski <mateusz.jablonski@intel.com>
 2025-10-02 16:20:59 +02:00
Szymon Morek
ee032982a6 refactor: print faulted address when waiting for tag 
getDeviceState on Windows prints faulted GPU VA
if OOB access happened.

Signed-off-by: Szymon Morek <szymon.morek@intel.com>
 2025-10-02 14:59:02 +02:00
Slawomir Milczarek
98b6259129 fix: Switch compute-engine idle check to try-lock 
Related-To: NEO-13225

Replace blocking obtainUniqueOwnership with non-blocking
tryObtainUniqueOwnership in isDirectSubmissionIdle.
Treat compute engine as not idle when lock is contended
(conservative; prevents false stops).
Avoid deadlock scenarios caused by holding
directSubmissionsMutex while waiting on a CSR lock.

Signed-off-by: Slawomir Milczarek <slawomir.milczarek@intel.com>
 2025-10-02 14:41:24 +02:00
Mateusz Jablonski
710d70a787 refactor: remove not needed code 
Related-To: NEO-16154

Signed-off-by: Mateusz Jablonski <mateusz.jablonski@intel.com>
 2025-10-02 11:41:08 +02:00
Mateusz Jablonski
165305090b refactor: remove always true variable - ftrSupportsInteger64BitAtomics 
Signed-off-by: Mateusz Jablonski <mateusz.jablonski@intel.com>
 2025-10-01 16:45:16 +02:00
Jack Myers
7b18a66710 fix: ensure proper heap alignment 
Noticed an issue on machine with following `sycl-ls`:

When running `sycl-ls` with a recent debug build of NEO, it exposed a
heap alignment problem:

```shell
Assert was called at 324 line in file:
../../src/neo/shared/source/memory_manager/gfx_partition.cpp
```

Related-To: NEO-16287
Signed-off-by: Jack Myers <jack.myers@intel.com>
 2025-10-01 14:46:52 +02:00
Mateusz Jablonski
5b8e8b8b1a refactor: unify GmmMemory definition 
Related-To: NEO-11080
Signed-off-by: Mateusz Jablonski <mateusz.jablonski@intel.com>
 2025-10-01 14:23:35 +02:00
Mateusz Jablonski
0caa8d5297 refactor: remove always true variable - supportsFloatAtomics 
Signed-off-by: Mateusz Jablonski <mateusz.jablonski@intel.com>
 2025-10-01 14:06:44 +02:00
Slawomir Milczarek
393c4d0985 Revert "fix: Fix DirectSubmissionController deadlock with try_lock pattern" 
Related-To: NEO-13325

This reverts commit 40ce5eb55e.

Signed-off-by: Slawomir Milczarek <slawomir.milczarek@intel.com>
 2025-10-01 11:23:16 +02:00
Bellekallu Rajkiran
286f973d0c feature(sysman): Add support for Memory properties & state for iGPU's in Linux 
Related-To: NEO-14198, NEO-15464

Signed-off-by: Bellekallu Rajkiran <bellekallu.rajkiran@intel.com>
 2025-10-01 09:46:43 +02:00
Filip Hazubski
728bdc2178 fix: Correct dummy blit pitch size for xe lpg and xe hpg 
Resolves: HSD-18043456303

Signed-off-by: Filip Hazubski <filip.hazubski@intel.com>
 2025-10-01 07:41:43 +02:00
Compute-Runtime-Validation
eed28d6552 Revert "fix: remove spir support from xe2 and later" 
This reverts commit a744ce6fbb.

Signed-off-by: Compute-Runtime-Validation <compute-runtime-validation@intel.com>
 2025-10-01 04:41:59 +02:00
Compute-Runtime-Validation
a0d55768a0 Revert "feature: fail stateful kernel when stateless is required" 
This reverts commit 2a8c0d867f.

Signed-off-by: Compute-Runtime-Validation <compute-runtime-validation@intel.com>
 2025-09-30 18:54:40 +02:00
Grochowski, Stanislaw
a744ce6fbb fix: remove spir support from xe2 and later 
Related-To: NEO-15215

Signed-off-by: Grochowski, Stanislaw <stanislaw.grochowski@intel.com>
 2025-09-30 18:23:24 +02:00
Marcel Skierkowski
047f6851f7 refactor: correct programming of SLM size 
Unify reading hw avaialble slm size:
Pre-Xe2 SLMSizeInKb stores total SLM size across all DSS
Xe2+ SLMSizeInKb stores SLM size per SS

apply restrictions for preferred/programmable SLM size

Related-To: NEO-12949
Signed-off-by: Marcel Skierkowski <marcel.skierkowski@intel.com>
 2025-09-30 17:57:18 +02:00
Filip Hazubski
b73e85de0d refactor: Refactor direct submission logic 
Remove ReleaseHelper::isDirectSubmissionSupported function.
Update ProductHelper::isDirectSubmissionSupported to return correct values.

Signed-off-by: Filip Hazubski <filip.hazubski@intel.com>
 2025-09-30 13:41:03 +02:00
Damian Tomczak
2a8c0d867f feature: fail stateful kernel when stateless is required 
Resolves: NEO-11859

Signed-off-by: Damian Tomczak <damian.tomczak@intel.com>
 2025-09-30 12:48:48 +02:00
Slawomir Milczarek
7334aee8a8 fix: Switch copy-engine idle check to try-lock 
Related-To: NEO-13325

Replace blocking obtainUniqueOwnership with non-blocking
tryObtainUniqueOwnership in isCopyEngineOnDeviceIdle.
Treat copy engine as not idle when lock is contended
(conservative; prevents false stops).
Avoid deadlock scenarios caused by holding
directSubmissionsMutex while waiting on a CSR lock.

Signed-off-by: Slawomir Milczarek <slawomir.milczarek@intel.com>
 2025-09-29 17:28:32 +02:00
Mateusz Jablonski
15d0feeda8 fix: improve propagating external functions info to kernel 
When relocation points to symbol that is not defined within module mark
it as optional. When symbol is available at dynamic linking time then
info from the function is retrieved but when the symbol is not available
then ignore the dependency.

Any unresolved symbol needed for module linking is already handled
in a separate place.

Related-To: NEO-16243, NEO-16263, NEO-16262, NEO-16268
Signed-off-by: Mateusz Jablonski <mateusz.jablonski@intel.com>
 2025-09-29 15:34:41 +02:00
Mateusz Hoppe
4b5f206058 feature: add ForceCompatibilityMode debug flag 
- allows using device binary without validating device against target
device

Related-To: HSD-13013893426

Signed-off-by: Mateusz Hoppe <mateusz.hoppe@intel.com>
 2025-09-29 15:22:06 +02:00
Slawomir Milczarek
40ce5eb55e fix: Fix DirectSubmissionController deadlock with try_lock pattern 
Related-To: NEO-13325

Replace blocking obtainUniqueOwnership() with tryObtainUniqueOwnership()
in critical paths to prevent deadlock between controller thread holding
directSubmissionsMutex and submission threads holding CSR locks.

Changes:
- Add CommandStreamReceiver::tryObtainUniqueOwnership() method
- Use try_lock in checkNewSubmissions() and context group idle detection
- Skip contended CSRs instead of blocking (conservative approach)
- Add comprehensive unit tests for try_lock functionality

Fixes deadlock scenario where registerDirectSubmission() waits for
directSubmissionsMutex while checkNewSubmissions() holds it and waits
for CSR ownership locks.

Signed-off-by: Slawomir Milczarek <slawomir.milczarek@intel.com>
 2025-09-29 11:38:26 +02:00
Radoslaw Jablonski
44645a3ed6 test: remove flawed blit test 
This test naively compares estimated blit command stream sizes instead
of actually checking if a command stream is correctly programmed.

The tested functionality is already covered by
givenMultiRootDeviceSyncNodeWhenFlushBcsTaskThenMiFlushAdded.

Signed-off-by: Radoslaw Jablonski <radoslaw.jablonski@intel.com>
 2025-09-29 11:06:11 +02:00
Mateusz Jablonski
39f10f9c5e refactor: stop using DEFAULT_TEST_FAMILY_NAME 
Signed-off-by: Mateusz Jablonski <mateusz.jablonski@intel.com>
 2025-09-29 10:15:13 +02:00
Compute-Runtime-Validation
aaa9342141 Revert "fix: improve parsing relocations in external usage" 
This reverts commit 65d883bc9d.

Signed-off-by: Compute-Runtime-Validation <compute-runtime-validation@intel.com>
 2025-09-29 10:05:58 +02:00
Mateusz Hoppe
98fcaf4ceb feature: support for context group on Windows 
Related-To: NEO-14289


Signed-off-by: Mateusz Hoppe <mateusz.hoppe@intel.com>
 2025-09-28 13:38:53 +02:00
Mateusz Jablonski
65d883bc9d fix: improve parsing relocations in external usage 
ignore relocations for symbols with undefined section

Related-To: NEO-16243
Signed-off-by: Mateusz Jablonski <mateusz.jablonski@intel.com>
 2025-09-26 16:33:11 +02:00
Compute-Runtime-Validation
ce65d34a8f Revert "fix: set proper gmm usage type for misaligned user ptr" 
This reverts commit a665f27676.

Signed-off-by: Compute-Runtime-Validation <compute-runtime-validation@intel.com>
 2025-09-26 16:24:45 +02:00
Szymon Morek
d429f0d8a0 fix: respect uncacheable flag for USM host allocations 
Related-To: NEO-14360

Signed-off-by: Szymon Morek <szymon.morek@intel.com>
 2025-09-25 16:31:19 +02:00
Bartosz Dunajski
72a5bcbfa1 feature: extend copy offload hint support 
Related-To: NEO-14557

Signed-off-by: Bartosz Dunajski <bartosz.dunajski@intel.com>
 2025-09-25 12:56:40 +02:00
Szymon Morek
a665f27676 fix: set proper gmm usage type for misaligned user ptr 
Related-To: NEO-15295

Signed-off-by: Szymon Morek <szymon.morek@intel.com>
 2025-09-25 12:06:43 +02:00
Mateusz Hoppe
cfb3dade30 refactor: replace isContextGroup with contextGroupCount in osContext 
Related-To: NEO-14289

Signed-off-by: Mateusz Hoppe <mateusz.hoppe@intel.com>
 2025-09-25 11:33:38 +02:00
Compute-Runtime-Validation
79cdf69efe Revert "fix: disable support for deprecated patchtokens format" 
This reverts commit 506d5837df.

Signed-off-by: Compute-Runtime-Validation <compute-runtime-validation@intel.com>
 2025-09-25 03:09:03 +02:00
Bellekallu Rajkiran
9cccbcabe1 fix: Reset kernelWithAssertAppended flag 
On new append calls, reset flag if previous submissions
are completed.

Related-To: NEO-16184

Signed-off-by: Bellekallu Rajkiran <bellekallu.rajkiran@intel.com>
 2025-09-24 18:30:51 +02:00
Maciej Plewka
cf35f8b40e fix: wa for multi ccs execution on bmg 
Related-To: NEO-15771
Signed-off-by: Maciej Plewka <maciej.plewka@intel.com>
 2025-09-24 18:15:07 +02:00
Kamil Kopryk
4133827e6f fix: set host function allocation as uncached 
The usage of host function allocation is similar to tagBuffer.
Ensure it's uncached and it follows tag buffer's cache policy.
Host function allocation will be readable and writable
by both: cpu and gpu.

Related-To: NEO-14577
Signed-off-by: Kamil Kopryk <kamil.kopryk@intel.com>
 2025-09-24 18:00:50 +02:00
Szymon Morek
804b8f5db6 fix: do not cache tag buffer allocation on GPU 
Related-To: NEO-14360

Signed-off-by: Szymon Morek <szymon.morek@intel.com>
 2025-09-24 16:28:35 +02:00