Add alignment to cuda_malloc_async_memory_resource. #4923
Merged
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
Signed-off-by: Michał Zientkiewicz <[email protected]>
|
CI MESSAGE: [8730704]: BUILD STARTED |
JanuszL
reviewed
Jun 23, 2023
|
|
||
| class aligned_alloc_helper { | ||
| public: | ||
| static constexpr size_t kCudaMallocAlignment = 256; |
There was a problem hiding this comment.
Is it given once and for all or may change depending on the version?
There was a problem hiding this comment.
I think it's given - I asked about it on CUDA channel and all functions that allocate global memory return data aligned to at least 256B.
JanuszL
reviewed
Jun 23, 2023
JanuszL
reviewed
Jun 23, 2023
JanuszL
reviewed
Jun 23, 2023
mzient
force-pushed
the
cudaMallocAsync_align
branch
from
51b93c5 to
f99fdc5
Compare
Signed-off-by: Michał Zientkiewicz <[email protected]>
|
CI MESSAGE: [8731034]: BUILD STARTED |
JanuszL
approved these changes
Jun 23, 2023
|
CI MESSAGE: [8731034]: BUILD FAILED |
|
CI MESSAGE: [8731034]: BUILD PASSED |
banasraf
approved these changes
Jun 27, 2023
Closed
JanuszL
pushed a commit
to JanuszL/DALI
that referenced
this pull request
Oct 13, 2023
* Overallocate and align allocations with alignment > 256B * Store a mapping from aligned to original addresses in a global map * Update tests Signed-off-by: Michał Zientkiewicz <[email protected]>
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Outdated suggestions cannot be applied.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.
Category:
Bug fix
New feature
Description:
cudaMallocAsyncallocates memory aligned to a multiple of 256 bytes, however, DALI's memory_resource interface can be used to request overaligned memory. Before this PR, DALI would neither return properly aligned memory nor raise an error, should cudaMallocAsync return insufficiently aligned pointer.This PR adds overalignment support by requesting larger blocks for allocations with alignment >256B and aligning the pointer accoringly. The mapping aligned->original pointer is kept (in a global state) and upon deletion, the pointer being deleted is looked up in the map and, if found, replaced with the original pointer.
Additionally, a tiny performance bug is fixed in AsyncPool tests.
Additional information:
Affected modules and functionalities:
cuda_malloc_async_memory_resourceAsyncPool tests.
Key points relevant for the review:
Tests:
New tests addedTest adjustedChecklist
Documentation
DALI team only
Requirements
REQ IDs: N/A
JIRA TASK: N/A