[MIGraphX EP] Fix CopyTensorAsync and add guards for stream sync CopyTensors #16787
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…Tensors() Add compile guards to gate functionality based on MIGRAPHX_STREAM_SYNC for adding the following - remove excess hipStreamSyncronize to nullstream on CopyTensor calls - Add proper call for stream synchronized CopyTensorAsync for DeviceToHost case Without this change subsequent CopyTensorAsync() calls will fail for cards that don't use pinned memory thus causing hipMemcpy() calls to occur before certain kernel operations occur.
|
ping @cloudhan @PeixuanZuo . This is related to the issue I found earlier. #16774 |
|
/azp run Windows ARM64 QNN CI Pipeline,Windows CPU CI Pipeline,Windows GPU CI Pipeline,Windows GPU TensorRT CI Pipeline,onnxruntime-binary-size-checks-ci-pipeline,orttraining-linux-ci-pipeline,orttraining-linux-gpu-ci-pipeline,orttraining-ortmodule-distributed |
|
/azp run Linux CPU CI Pipeline,Linux CPU Minimal Build E2E CI Pipeline,Linux GPU CI Pipeline,Linux GPU TensorRT CI Pipeline,Linux OpenVINO CI Pipeline,Linux QNN CI Pipeline,MacOS CI Pipeline,ONNX Runtime Web CI Pipeline |
|
Azure Pipelines successfully started running 8 pipeline(s). |
1 similar comment
|
Azure Pipelines successfully started running 8 pipeline(s). |
cloudhan
reviewed
Jul 21, 2023
This is already handled in the EP as end of run performs: OnRunEnd()->hipStreamQuery()->hipStreamSyncronize() as well as Sync()->hipStreamSync() Also after every hipMemCpy() we perform a hipStreamSyncronize(stream)
|
let me know if you need anything further for this or there are any additional concerns. |
|
/azp run Windows ARM64 QNN CI Pipeline,Windows CPU CI Pipeline,Windows GPU CI Pipeline,Windows GPU TensorRT CI Pipeline,onnxruntime-binary-size-checks-ci-pipeline,orttraining-linux-ci-pipeline,orttraining-linux-gpu-ci-pipeline,orttraining-ortmodule-distributed |
|
/azp run Linux CPU CI Pipeline,Linux CPU Minimal Build E2E CI Pipeline,Linux GPU CI Pipeline,Linux GPU TensorRT CI Pipeline,Linux OpenVINO CI Pipeline,Linux QNN CI Pipeline,MacOS CI Pipeline,ONNX Runtime Web CI Pipeline |
|
Azure Pipelines successfully started running 8 pipeline(s). |
1 similar comment
|
Azure Pipelines successfully started running 8 pipeline(s). |
|
@cloudhan, looks like this test isn't building for some reason |
cloudhan
approved these changes
Jul 22, 2023
TedThemistokleous
added a commit
to TedThemistokleous/onnxruntime
that referenced
this pull request
Jul 25, 2023
…Tensors (microsoft#16787) Add compile guards to gate functionality based on MIGRAPHX_STREAM_SYNC for adding the following - remove excess hipStreamSyncronize to nullstream on CopyTensor calls - Add proper call for stream synchronized CopyTensorAsync for DeviceToHost case Without this change subsequent CopyTensorAsync() calls will fail for cards that don't use pinned memory thus causing hipMemcpy() calls to occur before certain kernel operations occur.  becomes  --------- Co-authored-by: Ted Themistokleous <tthemist@amd.com>
This was referenced Jul 25, 2023
TedThemistokleous
added a commit
to TedThemistokleous/onnxruntime
that referenced
this pull request
Jul 25, 2023
…Tensors (microsoft#16787) Add compile guards to gate functionality based on MIGRAPHX_STREAM_SYNC for adding the following - remove excess hipStreamSyncronize to nullstream on CopyTensor calls - Add proper call for stream synchronized CopyTensorAsync for DeviceToHost case Without this change subsequent CopyTensorAsync() calls will fail for cards that don't use pinned memory thus causing hipMemcpy() calls to occur before certain kernel operations occur.  becomes  --------- Co-authored-by: Ted Themistokleous <tthemist@amd.com>
jeffdaily
pushed a commit
to ROCm/onnxruntime
that referenced
this pull request
Jul 25, 2023
…Tensors (microsoft#16787) (#13) Add compile guards to gate functionality based on MIGRAPHX_STREAM_SYNC for adding the following - remove excess hipStreamSyncronize to nullstream on CopyTensor calls - Add proper call for stream synchronized CopyTensorAsync for DeviceToHost case Without this change subsequent CopyTensorAsync() calls will fail for cards that don't use pinned memory thus causing hipMemcpy() calls to occur before certain kernel operations occur.  becomes  --------- Co-authored-by: Ted Themistokleous <tthemist@amd.com>
TedThemistokleous
added a commit
to TedThemistokleous/onnxruntime
that referenced
this pull request
Jul 26, 2023
…Tensors (microsoft#16787) Add compile guards to gate functionality based on MIGRAPHX_STREAM_SYNC for adding the following - remove excess hipStreamSyncronize to nullstream on CopyTensor calls - Add proper call for stream synchronized CopyTensorAsync for DeviceToHost case Without this change subsequent CopyTensorAsync() calls will fail for cards that don't use pinned memory thus causing hipMemcpy() calls to occur before certain kernel operations occur.  becomes  --------- Co-authored-by: Ted Themistokleous <tthemist@amd.com>
jchen351
pushed a commit
that referenced
this pull request
Aug 12, 2023
…Tensors (#16787) Add compile guards to gate functionality based on MIGRAPHX_STREAM_SYNC for adding the following - remove excess hipStreamSyncronize to nullstream on CopyTensor calls - Add proper call for stream synchronized CopyTensorAsync for DeviceToHost case Without this change subsequent CopyTensorAsync() calls will fail for cards that don't use pinned memory thus causing hipMemcpy() calls to occur before certain kernel operations occur.  becomes  --------- Co-authored-by: Ted Themistokleous <tthemist@amd.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Add compile guards to gate functionality based on MIGRAPHX_STREAM_SYNC for adding the following
Without this change subsequent CopyTensorAsync() calls will fail for cards that don't use pinned memory thus causing hipMemcpy() calls to occur before certain kernel operations occur.
Description
Remove excess syncronization when stream sync is enabled but also block the DeviceTohost CopyTensorAsync() to the desired GPU stream.
Motivation and Context
Without this change we fail to properly wait for a kernel to compute and synchronize correctly if memory isn't specifically pinned for the task. This was observed when doing another run with test_parity_gelu and test_parity_layernorm tests.
becomes