[XLA:GPU] Fuse shmem write loops for transposes in PackedTranspose #436
+41
−25
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
Replace per-transpose loops with a single unified loop that processes all transposes simultaneously, computing indices once and reusing them across all operations. Update packed_transpose_multiple_heroes.hlo test to verify the single-loop structure with multiple iter_args.
Collaborator
|
Could you check to see if this is applicable to the rocm-jaxlib-v0.8.0 and fix there as well, if needed? |
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Outdated suggestions cannot be applied.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.
Technical Details
Replace per-transpose loops with a single unified loop that processes all transposes simultaneously, computing indices once and reusing them across all operations.
Update
packed_transpose_multiple_heroes.hlotest to verify the single-loop structure with multiple iter_args.Motivation
This fixes the second performance regression (first #434) caused by the new implementation of fused transpose emitter for fp narrower than 32 (
PackedTranspose).Test Result
It reduces the execution time for
fused_convert_transpose_3.hloof Llama 3 8B FP8 by ~30%, bringing it almost back to v0.6.0 performance (~4% gap). Together with #434, the performance of the 4 top fused_convert_transpose kernels is improved by ~17%, resulting in an end-to-end model performance improvement of ~1% (tokens per second per gpu).