intel/llvm
1
0
Fork 0
mirror of https://github.com/intel/llvm.git synced 2026-01-28 01:04:49 +08:00
Code Issues Packages Projects Releases Wiki Activity

[flang][cuda] Defined allocator for unified data (#102189)

Browse Source 
CUDA unified variable where set to use the same allocator than managed
variable. This patch adds a specific allocator for the unified
variables. Currently it will call the managed allocator underneath but
we want to have the flexibility to change that in the future.
This commit is contained in:
Valentin Clement (バレンタイン クレメン)
2024-08-06 14:30:31 -07:00
committed by GitHub
parent 10d7805c4f
commit 388b63243c
4 changed files with 20 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
flang/include/flang/Runtime/CUDA/allocator.h
View File

@@ -36,5 +36,8 @@ void CUFFreeDevice(void *);
void *CUFAllocManaged(std::size_t);
void CUFFreeManaged(void *);
void *CUFAllocUnified(std::size_t);
void CUFFreeUnified(void *);
} // namespace Fortran::runtime::cuda
#endif // FORTRAN_RUNTIME_CUDA_ALLOCATOR_H_

3
flang/include/flang/Runtime/allocator-registry.h
View File

@@ -19,8 +19,9 @@ static constexpr unsigned kDefaultAllocator = 0;
static constexpr unsigned kPinnedAllocatorPos = 1;
static constexpr unsigned kDeviceAllocatorPos = 2;
static constexpr unsigned kManagedAllocatorPos = 3;
static constexpr unsigned kUnifiedAllocatorPos = 4;
#define MAX_ALLOCATOR 5
#define MAX_ALLOCATOR 7 // 3 bits are reserved in the descriptor.
namespace Fortran::runtime {

5
flang/lib/Lower/ConvertVariable.cpp
View File

@@ -1860,9 +1860,10 @@ static unsigned getAllocatorIdx(const Fortran::semantics::Symbol &sym) {
return kPinnedAllocatorPos;
if (*cudaAttr == Fortran::common::CUDADataAttr::Device)
return kDeviceAllocatorPos;
if (*cudaAttr == Fortran::common::CUDADataAttr::Managed ||
*cudaAttr == Fortran::common::CUDADataAttr::Unified)
if (*cudaAttr == Fortran::common::CUDADataAttr::Managed)
return kManagedAllocatorPos;
if (*cudaAttr == Fortran::common::CUDADataAttr::Unified)
return kUnifiedAllocatorPos;
}
return kDefaultAllocator;
}

12
flang/runtime/CUDA/allocator.cpp
View File

@@ -26,6 +26,8 @@ void CUFRegisterAllocator() {
kDeviceAllocatorPos, {&CUFAllocDevice, CUFFreeDevice});
allocatorRegistry.Register(
kManagedAllocatorPos, {&CUFAllocManaged, CUFFreeManaged});
allocatorRegistry.Register(
kUnifiedAllocatorPos, {&CUFAllocUnified, CUFFreeUnified});
}
void *CUFAllocPinned(std::size_t sizeInBytes) {
@@ -57,4 +59,14 @@ void CUFFreeManaged(void *p) {
CUDA_REPORT_IF_ERROR(cuMemFree(reinterpret_cast<CUdeviceptr>(p)));
}
void *CUFAllocUnified(std::size_t sizeInBytes) {
// Call alloc managed for the time being.
return CUFAllocManaged(sizeInBytes);
}
void CUFFreeUnified(void *p) {
// Call free managed for the time being.
CUFFreeManaged(p);
}
} // namespace Fortran::runtime::cuda