[Metal] Faster logsumexp for short rows via a simdgroup-per-row kernel#3913
Open
qx555 wants to merge 1 commit into
Open
[Metal] Faster logsumexp for short rows via a simdgroup-per-row kernel#3913qx555 wants to merge 1 commit into
qx555 wants to merge 1 commit into
Conversation
One simdgroup per row (8 rows per 256-thread threadgroup) with the online max/normalizer rescaling already used by logsumexp_looped: reductions never leave the simdgroup, so the kernel needs no threadgroup memory and no barriers (the block kernel uses 5). Dispatched for axis_size <= 2048; the block kernel is unchanged and still serves 2048 < axis_size <= 4096. On M4 Max: fp16 up to 2.8x, bf16 up to 2.6x, fp32 up to 1.5x on affected sizes; parity at the 2048 boundary and no change on block/looped paths. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
One simdgroup per row (8 rows per 256-thread threadgroup) with the online max/normalizer rescaling already used by logsumexp_looped: reductions never leave the simdgroup, so the kernel needs no threadgroup memory and no barriers (the block kernel uses 5). Dispatched for axis_size <= 2048; the block kernel is unchanged and still serves 2048 < axis_size <= 4096.
On M4 Max: fp16 up to 2.8x, bf16 up to 2.6x, fp32 up to 1.5x on affected sizes; parity at the 2048 boundary and no change on block/looped paths.