mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-04-19 22:39:03 +00:00
37554c31e8
* [fix] align v3 gufusion pipeline * fix device kernel selection. * Add .co direct asm support by CK_USE_ASM_MOE_STAGE2_BLOCKSCALE * experimental optimization for scale load in blkscale gemm * Add asm for no-loop v3_128x128x128 * fix bugs * tune fp8 example * Update v1_128x128x128 to 2x2 instead of 4x1 * wip * add warmup to asm launch * wip2 * 16x16 function merged to moe * temp save, a performant version. * wip3 * Update .co binary to 16x16 * 16x16x128 correct; 64x64x128 failed * update * use mem_op::set when topk=1 * add mx fp8 b_preshuffle support, function not yet tested. * Spilt the fp4 target. Fix the known bugs. 128x128x128 sanity checked; remove prints * some fixes * fix update * remove some unnecessary hacky; enable 256x256x256 tilesize * update for function debug * Add pipeline v3. Have some runtime issue and register spill * Fix pipe v3 correctness issue * remove unnecessary hacky * clang format * fix a bug * fix the bug, functional test passed * tempsave; buggy at passed 4 e8m0 to scaled mfma * added fp4_bpreshuffle example, build failures * fixed some bugs * implement shuffled scale mxfp4gemm, blocker: opsel not effect * hotfix * fix bugs, build passed * (M, N, K)=(128, 128, 128) function failed. * temp save for gemm1. Function not ready * fix compile error. Gemm2 pass. Gemm1 WIP * fix bug for a lds read * update moe * Compile pass. Gemm1 function WIP * update moe * fix fp8; fix even/odd * tempsave * update moe * Revert "update" This reverts commit960b2bce1c. * Revert "use mem_op::set when topk=1" This reverts commitdef952a178. * Add v3 128x128x128_4x4_16x16.co for gfx950 * temp cmake flag suppression for aiter test * add code for mxfp4 gemm, blockscale not supported yet * gemm1 up-only pass. GU WIP * function pass with inline asm hacky * revert unexpected file change * updated and build passed * update CE elementOP * added code for debug * Gemm1 GUFusion function pass. Perf WIP * Fix fp8/bf8; remove duplicated code * disable the scheduler in v3; bring it back when compiler feature ready. * update moe v1 pipeline * Add gemm1 v1 32x128x128 * remove schedule barrier * updated * Fix fp8/bf8 B-row * mfma using asm, device result correct, host result need to check * gemm1 v3 64x128x128 debug * fix cpu ref * a/b thread_desc stride fix * Use random scale for init1 * 16x16x128 input size blockscale function passed * fix blockscale gemm bug * tempsave. Almost all instances passed. * v1 fix for mi350. * temp save * debug save * update debug * fix the bug, 128x128x256 tile function passed * v3 * rename moe block selector and pipeline * Add gemm1 v1 * Add gemm1 v1 to selector * added mx moe block v3 support, function passed * compile error fix * Improve the pipeline * Pack e8m0 as int32_t * v1 compile pass. Function not ready * debug synchronize issue over different GPU/ROCm * minor fix * Add profiler filter * Add f4 ckProfiler * Fix example compile error * Add f4 profiler examples * tempsave * v1 function pass. * v3 function pass * align file and function name * mx_moe_fp4 ready for aiter with clang-format. * modify the way we represent fp4 * generalize the pipeline scheduling. * init moe mx f4 scale shuffle * Cmakelist diable compiler-bound flags * mx_fp4 default parameter change * Moe blockscale gemm1&gemm2 asm support for aiter. Suppression cmkae flag til new compler. * update code * tempsave; modify the way we represent fp4 * generalize the pipeline scheduling. * Add gemm1 gfx942 .co support * updated code, build passed. * Update gemm2 asm with latest compiler flag * Fix mx f4 ckProfiler * Fix blockwise gemm mx v1 * lds conflict free + buffer load lds * Add gemm2 v3 64x128x128 * fix a, b scale loading bugs, a, b scale loading now correctly * Add gemm2 v3 64x128x128 * commit with debug info * fix fp4 profiler * Add mx fp4 pileline v1 instances * Fix v2 topk_weight cal. Add silu asm. * v2 tok_weight WIP * init mx fp4 B no preshuffle version * tempsave. compile pass, function wrong * enable fp4 moe no weigth preshuffle, function pass * update the TFlops calculation in the example * Add gemm2 64x128x128 asm. Fix BF16 ref. * fix 2 typos in fp4_preshuffle * Better kernel selection in device classes * correct preShuffleBuffer we should used packed k to do shuffle. * lds conflict free + buffer load lds * optimize offset math in dma * Fix fp4 ckProfiler * Fix MX MFMA tests * fix f4 pipeline issues * gemm1 func pass * update mx moe gemm1_bns tile size to 64x128x256 * update mx moe gemm1 gemm2 TF and BW calculation * fix typo * temp save * Fix example_gemm_mx build * rename the block pipeline * correct a typo in tail * Add rotating to mx examples * fix the correctness issue * Fix v1; use M padding * Add NT flag to B/BScale buffer * Merge gemm_mx_common.hpp * temp save, 4.4~4.5 * Fix 'Merge gemm_mx_common.hpp' * refactor the pipeline * Pad the M for scale buffer unconditionaly * update MX moe GEMM1 hotloopscheduling * change the gemm1 tile from 64x128x128 to 128x64x128 * Unconditional Ascale padding * Pad shuffled a scale only * pad ascale * add vmcnt guard for async copy * Profiler add f4 wp * Merge preshuffle device * Add more fp4 wp instances * Fix do_weight in gemm1. Fix cshuffle_datatype. Clang-format * Clang-format after 2 merges * Remove rocm6.3 workaround flags and macro * Fix fp8 config * Fix bf8 config * flag and barrier fix for copmiler branch MainOpSelV3 * Add fp8 profiler instances * Remove debug infos; Enable flags for blockscale f8 * No asm ver. for merging moe blocksale fp8 into mainline * update the flag name for f8blockscale * recover example * fix performance bug of bpreshuffle f8 gemm * clang format, remove single rate mfma restriction for f8 * remove single rate mfma restriction for f8 blockscale gemm * Fix moe blockscale gemm1 barrier 0x800 for new compiler * add pipeline v1 for MOE Gemm2 * Use v1 pipeline for example_moe_gemm2_xdl_mx_fp4_bns * Fix OOB; add MB96 instances * remove unnecessary files * fix the cmake issue * Enable splitk for mxfp4; clang format; * Generate random tensor values with multiple threads * Use packed_size_v for A/BPackedSize * Fix warning * Fix target_compile_options for disabled target on gfx942 * fix moe pki4 on gfx950 * doc the kGroup definition * Fix ThreadwiseTensorSliceTransfer_v4::Run (Fuse scale) * Refactor thread_copy_lds_direct_load; fix gfx942 direct lds load example; fix f16_pki4 example * Fix unknown compiler flag * fix two failed examples. * fix some failure tile size in gfx950 universal gemm. fix test_gemm_fp16 * workaround fix for test_gemm_f32; * We have very limited support for lds direct load if input matrix is not K major * fix test_gemm_splitk; * Fix compile for mx_mfma_op * add mfma selection logic for multipled_v3 * Clean up * Fix device gemm mx link error * improve the global atomic pattern * Revert unnecessary copyright updates * restore minimum_occupancy logic * Avoid data race in moe gemm2 ref * Build fp8 gemm_multiply_multiply and moe only on gfx94/95 * update the instance in device_mx_gemm * Resolve comments * Copyright 2025 * Remove unused code * fix library linking issue --------- Co-authored-by: OscarXu <huaiguxu@amd.com> Co-authored-by: lalala-sh <Jiaxing.Wen@amd.com> Co-authored-by: mtgu0705 <mtgu@amd.com> Co-authored-by: aska-0096 <haocwang@amd.com> Co-authored-by: Your Name <you@example.com> Co-authored-by: valarLip <340077269@qq.com> Co-authored-by: feifei14119 <feiw@amd.com> Co-authored-by: Lin, Qun <qlin@amd.com> Co-authored-by: Andriy Roshchenko <andriy.roshchenko@amd.com> Co-authored-by: joye <joye@amd.com> Co-authored-by: asleepzzz <hanwen.chang@amd.com>
Instructions for example_gemm_xdl
Run example_gemm_xdl
#arg1: verification (0=no, 1=yes)
#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
#arg3: run kernel # of times (>1)
./bin/example_gemm_xdl 0 1 5
Instructions for example_gemm_xdl_fp16_streamk_v3
Run example_gemm_xdl_fp16_streamk_v3
arg1: verification (0=no, 1=yes)
arg2: initialization (0=no init, 1=integer value, 2=decimal value)
arg3: time kernel (0=no, 1=yes)
arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC
arg10: stream-k select (-1: default config, 0: all DP, 1: 1-tile SK, 2: 2-tile SK)
arg11: Grid_size(-1 for max occupancy)
bin/example_gemm_xdl_fp16_streamk_v3 1 2 1 3840 4096 4096 4096 4096 4096 1 -1
a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
b_k_n: dim 2, lengths {4096, 4096}, strides {4096, 1}
c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
problem {M:3840, N:4096, K:4096, SA:4096, SB:4096, SC:4096, MP:4032, NP:4096, KRead:4096, KP:4096, AK0:512, BK0:2048, MBlock: 18, NBlock: 16, Stream-K Selection:1, Grid size:-1}
Perf: 0.292022 ms, 441.23 TFlops, 330.348 GB/s, DeviceGemmXdlUniversal<MNPadding, RRR> BlkSize: 256, BlkTile: 224x256x64, WaveTile: 16x16, WaveMap: 7x8, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v3, BlkGemmPipelinePrefetchStages: 2