mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-06-10 16:28:38 +00:00
7ecbf82708
[CK_TILE][FMHA] Optimize long-context decoding on gfx11/12 (#7500) ## Motivation Relevant issue: ROCM-22065 FMHA has less-than-optimal performance of long-context decoding (i.e. when seqlen_q = 1) on gfx11/12. This PR optimizes the splitkv pipeline and configs for such scenarios. ## Technical Details Optimizations applied in this PR: 1. use tiles with smaller M0 (16 vs 64), these tiles are used when seqlen_q <= 16 2. adapt qr_nwarp_sshuffle pipeline for gfx11, it allows to use more warps even for M0 = 16 (the qr pipeline parallelizes work between warps in M dim so with M0 = 16 it allows to use only 1 warp) 3. enable kMergeNumHeadGroupsSeqLenQ (an optimization that merges one group of heads in GQA) for all hdim values, not only 128 4. increase the number of splits (multiply by the number of head groups) if (3) is used 5. increase the number of splits for RDNAs (`multiProcessorCount` is the number of WGPs on RDNAs, not CUs, so it should be doubled to have meaning similar to CDNAs) Performance on gfx1151: | Case | develop (GB/s) | This PR (GB/s) | |:-------|-------:|-------:| | [fp16\|group\|bshd] b:1, h:32/32, s:1/45056, d:64/64 | 127.58 | 183.11 | | [fp16\|group\|bhsd] b:1, h:32/32, s:1/45056, d:64/64 | 153.64 | 215.02 | | [fp16\|group\|bshd] b:1, h:16/8, s:1/77184, d:128/128 | 120.51 | 225.76 | | [fp16\|group\|bhsd] b:1, h:16/8, s:1/77184, d:128/128 | 130.62 | 223.84 | | [fp16\|group\|bshd] b:1, h:32/32, s:1/9600, d:128/128 | 82.65 | 138.44 | | [fp16\|group\|bhsd] b:1, h:32/32, s:1/9600, d:128/128 | 105.75 | 220.45 | | [fp16\|group\|bshd] b:1, h:8/1, s:1/401024, d:256/256 | 16.27 | 187.89 | | [fp16\|group\|bhsd] b:1, h:8/1, s:1/401024, d:256/256 | 16.28 | 188.19 | ## Test Plan An additional test case is added to the exiting test. It uses seqlen_q = 1, GQA, no mask to trigger the changes ``` ninja test_ck_tile_fmha_fwd_fp16 && bin/test_ck_tile_fmha_fwd_fp16 --gtest_filter="*SplitKV* ninja test_ck_tile_fmha_fwd_bf16 && bin/test_ck_tile_fmha_fwd_bf16 --gtest_filter="*SplitKV* ``` Manual testing can be done with these commands: ``` bin/tile_example_fmha_fwd -prec=fp16 -mode=1 -page_block_size=128 -b=1 -h=32 -h_k=32 -d=64 -s=1 -s_k=$((352 * 128)) -lse=1 -mask=0 -num_splits=0 -kname=1 -v=1 bin/tile_example_fmha_fwd -prec=fp16 -mode=1 -page_block_size=128 -b=1 -h=16 -h_k=8 -d=128 -s=1 -s_k=$((603 * 128)) -lse=1 -mask=0 -num_splits=0 -kname=1 -v=1 bin/tile_example_fmha_fwd -prec=fp16 -mode=1 -page_block_size=128 -b=1 -h=32 -h_k=32 -d=128 -s=1 -s_k=$((75 * 128)) -lse=1 -mask=0 -num_splits=0 -kname=1 -v=1 bin/tile_example_fmha_fwd -prec=fp16 -mode=1 -page_block_size=128 -b=1 -h=8 -h_k=1 -d=256 -s=1 -s_k=$((3133 * 128)) -lse=1 -mask=0 -num_splits=0 -kname=1 -v=1 ``` ## Test Result All the tests must pass. ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.