ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-07-18 09:38:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

Update profiling doc.

Browse Source
This commit is contained in:
Ville Pietilä
2026-02-10 04:45:54 -05:00
parent d32bdb1412
commit c9504f2c27
2 changed files with 100 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files

88
example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_example.inc
View File

@@ -58,10 +58,12 @@ using DeviceConvFwdInstance =
8>; // Num groups to merge
*/
// WIP
/*
// Current best instance for Phantom case
template <ck::index_t NDimSpatial>
using DeviceConvFwdInstance =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<
NDimSpatial,
InputLayout<NDimSpatial>,
WeightLayout<NDimSpatial>,
@@ -76,43 +78,97 @@ using DeviceConvFwdInstance =
InElementOp,
WeiElementOp,
OutElementOp,
//ConvSpec, // ConvForwardSpecialization
ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter3x3Stride1Pad1Dilation1_32_4_4_200x200,
ConvSpec, // ConvForwardSpecialization
GemmSpec, // GemmSpecialization
1,
256, // BlockSize
128, // MPerBlock
32, // NPerBlock
256, // MPerBlock
128, // NPerBlock
32, // KPerBlock
8, // AK1
8, // BK1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
1, // NXdlPerWave
4, // MXdlPerWave
2, // NXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
4, // ABlockTransferSrcScalarPerVector
8, // ABlockTransferSrcScalarPerVector
8, // ABlockTransferDstScalarPerVector_AK1
1, // ABlockLdsExtraM
S<4, 32, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
4, // BBlockTransferSrcScalarPerVector
8, // BBlockTransferSrcScalarPerVector
8, // BBlockTransferDstScalarPerVector_BK1
1, // BBlockLdsExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 4>, // CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, must match with num merged groups
1, // Vector load/store size for output tensor = CDEBlockTransferScalarPerVector_NPerBlock
ck::BlockGemmPipelineScheduler::Interwave,
ck::BlockGemmPipelineVersion::v1,
S<1, 32, 1, 8>, // CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, must match with num merged groups
4>; // Vector load/store size for output tensor = CDEBlockTransferScalarPerVector_NPerBlock
*/
// WIP
template <ck::index_t NDimSpatial>
using DeviceConvFwdInstance =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<
NDimSpatial,
InputLayout<NDimSpatial>,
WeightLayout<NDimSpatial>,
ck::Tuple<>,
OutputLayout<NDimSpatial>,
InKernelDataType,
WeiKernelDataType,
AccDataType,
CShuffleDataType,
ck::Tuple<>,
OutKernelDataType,
InElementOp,
WeiElementOp,
OutElementOp,
ConvSpec, // ConvForwardSpecialization
GemmSpec, // GemmSpecialization
1,
256, // BlockSize
256, // MPerBlock
128, // NPerBlock
32, // KPerBlock
8, // AK1
8, // BK1
32, // MPerXdl
32, // NPerXdl
4, // MXdlPerWave
2, // NXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
8, // ABlockTransferDstScalarPerVector_AK1
1, // ABlockLdsExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
8, // BBlockTransferDstScalarPerVector_BK1
1, // BBlockLdsExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, must match with num merged groups
4>; // Vector load/store size for output tensor = CDEBlockTransferScalarPerVector_NPerBlock
/*
ck::BlockGemmPipelineScheduler::Intrawave,
ck::BlockGemmPipelineVersion::v3,
InKernelDataType,
WeiKernelDataType,
false, // No direct load
8>; // Num groups to merge
1>; // Num groups to merge
*/
// Best instance for 2nd most difficult fwd case, increase the B block src vector size?
/*