Optimize group_index_select_or_add_2d_kernel on ROCm by adding a separate codepath for small embedding dimensions #5233
Conversation
This file contains hidden or 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
Contributor
…p_index_select_or_add_2d_kernel
…zed small embedding dims path
…isable optimized smallEmbD path
q10
reviewed
Jan 20, 2026
| return GROUP_INDEX_SELECT_COLS_PER_WARP; | ||
| } | ||
|
|
||
| int get_group_index_select_unroll_factor() { |
Contributor
Author
There was a problem hiding this comment.
Not trivially, unfortunately. We would need to move the function definition to sparse_ops.h where it is declared, and move the GROUP_INDEX_SELECT_UNROLL_FACTOR variable there as well.
Happy to do this refactor if you would suggest.
| if (num_cols < COLS_PER_WARP && num_cols >= UNROLL_FACTOR) { | ||
| // Need to ensure that [member_id] and [member_warp_id] are calculated correctly | ||
| // for the small embedding dimension path below | ||
| int rows_per_warp = COLS_PER_WARP / num_cols; |
Contributor
Author
There was a problem hiding this comment.
Done in bbdc17d
I left input, output, indices and idx as they were previously defined
q10
pushed a commit
to q10/FBGEMM
that referenced
this pull request
Jan 28, 2026
…rate codepath for small embedding dimensions (pytorch#5233) Summary: This PR optimizes the performance on ROCm of the `group_index_select_or_add_2d_kernel` kernel on tables with small embedding dimensions (i.e., `num_cols`). For tables with small embedding dimensions, the code is refactored to process multiple rows within the same warp. Two files are changed: 1. `fbgemm_gpu/src/sparse_ops/sparse_ops_gpu.cpp` - The calculation of the `warp_offsets` is changed in the host-side code. 2. `fbgemm_gpu/src/sparse_ops/sparse_group_index.cu` - The `group_index_select_or_add_2d_kernel` kernel is modified to process multiple rows within a warp for small embedding dimensions. Test Plan: ``` cd ~/fbsource/fbcode/ai_codesign/nonprod/bensonma415/scripts/D89316371 bash run_benchmark.sh amd/mi300 2>&1 | pastry ``` https://docs.google.com/document/d/12ywiAQhA3eZqcIwUyc8_CQinwZJy_WDD8B4_43G8Gr8/edit?tab=t.0#heading=h.5row1qfol66k Reviewed By: echen4096 Differential Revision: D89316371 Pulled By: q10
cyyever
pushed a commit
to cyyever/FBGEMM
that referenced
this pull request
Jan 29, 2026
…rate codepath for small embedding dimensions (pytorch#5233) Summary: X-link: https://github.com/facebookresearch/FBGEMM/pull/2340 This PR optimizes the performance on ROCm of the `group_index_select_or_add_2d_kernel` kernel on tables with small embedding dimensions (i.e., `num_cols`). For tables with small embedding dimensions, the code is refactored to process multiple rows within the same warp. Two files are changed: 1. `fbgemm_gpu/src/sparse_ops/sparse_ops_gpu.cpp` - The calculation of the `warp_offsets` is changed in the host-side code. 2. `fbgemm_gpu/src/sparse_ops/sparse_group_index.cu` - The `group_index_select_or_add_2d_kernel` kernel is modified to process multiple rows within a warp for small embedding dimensions. Pull Request resolved: pytorch#5233 Test Plan: ``` cd ~/fbsource/fbcode/ai_codesign/nonprod/bensonma415/scripts/D89316371 bash run_benchmark.sh amd/mi300 2>&1 | pastry ``` https://docs.google.com/document/d/12ywiAQhA3eZqcIwUyc8_CQinwZJy_WDD8B4_43G8Gr8/edit?tab=t.0#heading=h.5row1qfol66k Reviewed By: echen4096 Differential Revision: D89316371 Pulled By: q10 fbshipit-source-id: 2742965773c92ff96419fa5978c93ca6d23dbed4
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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.
This PR optimizes the performance on ROCm of the
group_index_select_or_add_2d_kernelkernel on tables with small embedding dimensions (i.e.,num_cols).For tables with small embedding dimensions, the code is refactored to process multiple rows within the same warp. Two files are changed:
fbgemm_gpu/src/sparse_ops/sparse_ops_gpu.cpp- The calculation of thewarp_offsetsis changed in the host-side code.fbgemm_gpu/src/sparse_ops/sparse_group_index.cu- Thegroup_index_select_or_add_2d_kernelkernel is modified to process multiple rows within a warp for small embedding dimensions.