ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-11 08:50:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

[CK] Fixed MPerBlock=32 build issue for MXFP4 GEMM decode (#2512)

Browse Source 
* added MPerBlock=32 for MXFP4 GEMM decode

* added two instance for M>128 scenario.

* added 1 instance

* format

---------

Co-authored-by: mtgu0705 <mtgu@amd.com>
Co-authored-by: felix <felix.li@amd.com>
This commit is contained in:
Mingtao Gu
2025-07-18 14:35:54 +08:00
committed by GitHub
parent f0a8c18017
commit 0198257d79
2 changed files with 186 additions and 65 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
library/src/tensor_operation_instance/gpu/gemm_mx/device_gemm_mx_xdl_f4_f4_f16/device_gemm_mx_xdl_f4_f4_f16_mk_mfma_mn.hpp
View File

@@ -46,10 +46,11 @@ using device_gemm_mx_xdl_f4_f4_f16_mk_mfma_mn_instances = std::tuple<
//#####################| | | | Type| Data| Type| Data| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//#####################| | | | | Type| | Type| | | | Operation| Operation| Operation| | | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// DeviceGemmMX_Xdl_CShuffleV3<Row, MFMA, Row, F4, E8M0PK, F4, E8M0PK, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 32, 128, 128, 16, 16, 16, 16, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, 2, 2, S<1, 16, 1, 16>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v3>,
// DeviceGemmMX_Xdl_CShuffleV3<Row, MFMA, Row, F4, E8M0PK, F4, E8M0PK, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 32, 256, 128, 16, 16, 16, 16, 2, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, 2, 4, S<1, 8, 1, 32>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v3>,
// DeviceGemmMX_Xdl_CShuffleV3<Row, MFMA, Row, F4, E8M0PK, F4, E8M0PK, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 32, 384, 128, 16, 16, 16, 16, 2, 6, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, 2, 2, S<1, 16, 1, 16>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v3>,
// DeviceGemmMX_Xdl_CShuffleV3<Row, MFMA, Row, F4, E8M0PK, F4, E8M0PK, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 32, 512, 128, 16, 16, 16, 16, 2, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, 2, 4, S<1, 8, 1, 32>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v3>,
DeviceGemmMX_Xdl_CShuffleV3<Row, MFMA, Row, F4, E8M0PK, F4, E8M0PK, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 32, 128, 128, 16, 16, 16, 16, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, 2, 2, S<1, 16, 1, 16>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v3>,
DeviceGemmMX_Xdl_CShuffleV3<Row, MFMA, Row, F4, E8M0PK, F4, E8M0PK, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 32, 256, 128, 16, 16, 16, 16, 2, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, 2, 4, S<1, 8, 1, 32>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v3>,
DeviceGemmMX_Xdl_CShuffleV3<Row, MFMA, Row, F4, E8M0PK, F4, E8M0PK, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 32, 384, 128, 16, 16, 16, 16, 2, 6, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, 2, 2, S<1, 16, 1, 16>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v3>,
DeviceGemmMX_Xdl_CShuffleV3<Row, MFMA, Row, F4, E8M0PK, F4, E8M0PK, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 32, 512, 128, 16, 16, 16, 16, 2, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, 2, 4, S<1, 8, 1, 32>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v3>,
DeviceGemmMX_Xdl_CShuffleV3<Row, MFMA, Row, F4, E8M0PK, F4, E8M0PK, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 64, 128, 128, 16, 16, 16, 16, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, 2, 2, S<1, 16, 1, 16>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v3>,
DeviceGemmMX_Xdl_CShuffleV3<Row, MFMA, Row, F4, E8M0PK, F4, E8M0PK, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 64, 256, 128, 16, 16, 16, 16, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, 2, 4, S<1, 8, 1, 32>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v3>,
@@ -65,6 +66,14 @@ using device_gemm_mx_xdl_f4_f4_f16_mk_mfma_mn_instances = std::tuple<
DeviceGemmMX_Xdl_CShuffleV3<Row, MFMA, Row, F4, E8M0PK, F4, E8M0PK, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 128, 256, 128, 16, 16, 16, 16, 8, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, 2, 4, S<1, 8, 1, 32>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v3>,
DeviceGemmMX_Xdl_CShuffleV3<Row, MFMA, Row, F4, E8M0PK, F4, E8M0PK, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 128, 384, 128, 16, 16, 16, 16, 8, 6, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, 2, 2, S<1, 16, 1, 16>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v3>,
DeviceGemmMX_Xdl_CShuffleV3<Row, MFMA, Row, F4, E8M0PK, F4, E8M0PK, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 128, 512, 128, 16, 16, 16, 16, 8, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, 2, 4, S<1, 8, 1, 32>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v3>,
//new instances for testing
DeviceGemmMX_Xdl_CShuffleV3<Row, MFMA, Row, F4, E8M0PK, F4, E8M0PK, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 256, 256, 128, 16, 16, 16, 16, 16, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, 2, 4, S<1, 8, 1, 32>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v3>,
DeviceGemmMX_Xdl_CShuffleV3<Row, MFMA, Row, F4, E8M0PK, F4, E8M0PK, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 256, 256, 128, 16, 16, 16, 16, 8, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, 2, 8, S<1, 8, 1, 32>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v3>,
DeviceGemmMX_Xdl_CShuffleV3<Row, MFMA, Row, F4, E8M0PK, F4, E8M0PK, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 256, 64, 128, 16, 16, 16, 16, 8, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, 4, 2, S<1, 32, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v3>,
// DeviceGemmMX_Xdl_CShuffleV3<Row, MFMA, Row, F4, E8M0PK, F4, E8M0PK, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 256, 128, 128, 16, 16, 16, 16, 16, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, 2, 4, S<1, 8, 1, 32>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v3>,
std::nullptr_t
// clang-format on
>;