ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-06-30 03:37:38 +00:00
Code Issues Packages Projects Releases Wiki Activity

remove some unnecessary hacky; enable 256x256x256 tilesize

Browse Source
This commit is contained in:
aska-0096
2025-05-09 07:54:28 +00:00
parent b2efb06315
commit bb043a3202
2 changed files with 42 additions and 40 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
example/67_gemm_microscaling/gemm_mx_fp4.cpp
View File

@@ -23,12 +23,15 @@ using BElementOp = PassThrough; // elementwise transformation for B matrix
using CElementOp = PassThrough; // elementwise transformation for C matrix
constexpr ck::index_t ScaleBlockSize = 32; // scaling block size
constexpr ck::index_t KPerBlock = 128;
constexpr ck::index_t KPerBlock = 256;
constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
constexpr auto BlkGemmPSched = ck::BlockGemmPipelineScheduler::Intrawave;
constexpr auto BlkGemmPVer = ck::BlockGemmPipelineVersion::v1;
// AB DataType: f4x2_pk_t
// Mathmatically, all numbers are represented as f4.
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMX_Xdl_CShuffleV3<
ALayout, // ALayout
BLayout, // BLayout
@@ -45,29 +48,29 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMX_Xdl_CShuffle
CElementOp, // CElementwiseOperation
GemmSpec, // GemmSpec
ScaleBlockSize, // ScaleBlockSize: Scaling block size
256, // BlockSize: Thread block size
128, // MPerBlock
128, // NPerBlock
256, // BlockSize: Thread block size
256, // MPerBlock
256, // NPerBlock
KPerBlock, // KPerBlock
16, // AK1
16, // BK1
32, // AK1
32, // BK1
16, // MPerXDL
16, // NPerXDL
4, // MXdlPerWave
4, // NXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
8, // MXdlPerWave
8, // NXdlPerWave
S<8, 32, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
16, // ABlockTransferSrcScalarPerVector
16, // ABlockTransferDstScalarPerVector_AK1
32, // ABlockTransferSrcScalarPerVector
32, // ABlockTransferDstScalarPerVector_AK1
false, // ABlockLdsExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
S<8, 32, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
16, // BBlockTransferSrcScalarPerVector
16, // BBlockTransferDstScalarPerVector_BK1
32, // BBlockTransferSrcScalarPerVector
32, // BBlockTransferDstScalarPerVector_BK1
false, // BBlockLdsExtraN
1, // CShuffleMXdlPerWavePerShuffle
2, // CShuffleNXdlPerWavePerShuffle