Skip to content

[release/2.6] MIOpen: Get current device from Torch rather than HIP in handle creation #154549 #2216

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Jun 6, 2025
Merged

[release/2.6] MIOpen: Get current device from Torch rather than HIP in handle creation #154549 #2216

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
51 changes: 28 additions & 23 deletions aten/src/ATen/miopen/Handle.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,40 +1,45 @@
#include <ATen/miopen/Exceptions.h>
#include <ATen/miopen/Handle.h>
#include <ATen/hip/detail/DeviceThreadHandles.h>
#include <ATen/miopen/Handle.h>
#include <c10/hip/HIPStream.h>

namespace at { namespace native {
#include <ATen/hip/Exceptions.h>
#include <ATen/miopen/Exceptions.h>

namespace at::native {
namespace {

void createMIOpenHandle(miopenHandle_t *handle) {
MIOPEN_CHECK(miopenCreate(handle));
}

void destroyMIOpenHandle(miopenHandle_t handle) {
// this is because of something dumb in the ordering of
// destruction. Sometimes atexit, the cuda context (or something)
// would already be destroyed by the time this gets destroyed. It
// happens in fbcode setting. @colesbury and I decided to not destroy
// the handle as a workaround.
// - @soumith
//
// Further note: this is now disabled globally, because we are seeing
// the same issue as mentioned above in CUDA 11 CI.
// - @zasdfgbnm
//
// #ifdef NO_MIOPEN_DESTROY_HANDLE
// #else
// miopenDestroy(handle);
// #endif
// this is because of something dumb in the ordering of
// destruction. Sometimes atexit, the cuda context (or something)
// would already be destroyed by the time this gets destroyed. It
// happens in fbcode setting. @colesbury and I decided to not destroy
// the handle as a workaround.
// - @soumith
//
// Further note: this is now disabled globally, because we are seeing
// the same issue as mentioned above in CUDA 11 CI.
// - @zasdfgbnm
//
// #ifdef NO_MIOPEN_DESTROY_HANDLE
// #else
// miopenDestroy(handle);
// #endif
}

using MIOpenPoolType = at::cuda::DeviceThreadHandlePool<miopenHandle_t, createMIOpenHandle, destroyMIOpenHandle>;
using MIOpenPoolType = at::cuda::DeviceThreadHandlePool<
miopenHandle_t,
createMIOpenHandle,
destroyMIOpenHandle>;

} // namespace

miopenHandle_t getMiopenHandle() {
int device;
HIP_CHECK(hipGetDevice(&device));
c10::DeviceIndex device = 0;
AT_CUDA_CHECK(c10::hip::GetDevice(&device));

// Thread local PoolWindows are lazily-initialized
// to avoid initialization issues that caused hangs on Windows.
Expand All @@ -46,8 +51,8 @@ miopenHandle_t getMiopenHandle() {
pool->newPoolWindow());

auto handle = myPoolWindow->reserve(device);
MIOPEN_CHECK(miopenSetStream(handle, at::hip::getCurrentHIPStream()));
MIOPEN_CHECK(miopenSetStream(handle, c10::hip::getCurrentHIPStream()));
return handle;
}

}} // namespace at::native
} // namespace at::native