mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-17 03:19:48 +00:00
Add gridwise GEMM pipeline (#89)
* clean up
* add mutilple thread scratch to ThreadwiseTensorSliceTransfer_v3r1
* add 2 stage prefetch
* add more sanity check into transform_tensor_descriptor
* tweak
* enabling 2 stage prefetch to exsiting gridwise gemm; tweak
* enabling 2 stage prefetch to exsiting gridwise gemm
* move gridwise gemm pipeline in class; clean up
* add some irregular tile size
* update CalculateHasMainK0BlockLoop for multi-stage-prefetch
* refactor gridwise gemm pipeline class
[ROCm/composable_kernel commit: 22d438ae9e]
This commit is contained in:
Chao Liu
@@ -11,13 +11,23 @@
|
||||
#include "host_tensor_generator.hpp"
|
||||
#include "host_gemm.hpp"
|
||||
#include "device_tensor.hpp"
|
||||
#include "device_gemm_xdl.hpp"
|
||||
#include "device_gemm_xdl_c_shuffle.hpp"
|
||||
#include "element_wise_operation.hpp"
|
||||
#include "reference_gemm.hpp"
|
||||
#include "gemm_specialization.hpp"
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
using F16 = ck::half_t;
|
||||
using F32 = float;
|
||||
|
||||
using Row = ck::tensor_layout::gemm::RowMajor;
|
||||
using Col = ck::tensor_layout::gemm::ColumnMajor;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
using ADataType = ck::half_t;
|
||||
using BDataType = ck::half_t;
|
||||
using CDataType = ck::half_t;
|
||||
@@ -31,45 +41,56 @@ using AElementOp = ck::tensor_operation::element_wise::PassThrough;
|
||||
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
|
||||
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization_t::Default;
|
||||
static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization_t::MNPadding;
|
||||
|
||||
// clang-format off
|
||||
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl_C_Shuffle<
|
||||
ADataType, // ADataType
|
||||
BDataType, // BDataType
|
||||
CDataType, // CDataType
|
||||
AccDataType, // AccDataType
|
||||
ALayout, // ALayout
|
||||
BLayout, // BLayout
|
||||
CLayout, // CLayout
|
||||
AElementOp, // AElementwiseOperation
|
||||
BElementOp, // BElementwiseOperation
|
||||
CElementOp, // CElementwiseOperation
|
||||
256, // BlockSize
|
||||
256, // MPerBlock
|
||||
128, // NPerBlock
|
||||
4, // K0PerBlock
|
||||
8, // K1
|
||||
32, // MPerXDL
|
||||
32, // NPerXDL
|
||||
4, // MXdlPerWave
|
||||
2, // NXdlPerWave
|
||||
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_K0_M_K1
|
||||
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
|
||||
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
|
||||
2, // ABlockTransferSrcVectorDim
|
||||
8, // ABlockTransferSrcScalarPerVector
|
||||
8, // ABlockTransferDstScalarPerVector_K1
|
||||
true, // ABlockLdsAddExtraM
|
||||
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_K0_N_K1
|
||||
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
|
||||
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
|
||||
2, // BBlockTransferSrcVectorDim
|
||||
8, // BBlockTransferSrcScalarPerVector
|
||||
8, // BBlockTransferDstScalarPerVector_K1
|
||||
true, // BBlockLdsAddExtraN
|
||||
1, // CShuffleMXdlPerWavePerShuffle
|
||||
1, // CShuffleNXdlPerWavePerShuffle
|
||||
S<1, 1, 32, 1, 1, 8>, // CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
|
||||
8>; // CBlockTransferScalarPerVector_NWaveNPerXdl
|
||||
#if 1
|
||||
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl
|
||||
//######| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| Num|
|
||||
//######| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Spacialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar| Prefetch|
|
||||
//######| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| |
|
||||
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
// [256, 128, 4, 8], 1 stage, 2 occupancy
|
||||
< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 7, 1, 1>;
|
||||
#elif 0
|
||||
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl
|
||||
//######| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| Num|
|
||||
//######| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Spacialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar| Prefetch|
|
||||
//######| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| |
|
||||
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
// [128, 144, 8, 8], 1 stage, 1 occupancy, bounded by LDS size
|
||||
// 99 TFlops, 120 blocks (1024x2160x3840)
|
||||
// 99 TFlops, 960 blocks (4096x4320x3840)
|
||||
< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 144, 8, 8, 16, 16, 2, 9, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 8, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, true, 7, 1, 1>;
|
||||
// [128, 144, 4, 8], 1 stage, 2 occupancy,
|
||||
// 92 TFlops, 120 blocks (1024x2160x3840)
|
||||
// 120 TFlops, 240 blocks (1024x4320x3840)
|
||||
// 128 TFlops, 960 blocks (4096x4320x3840)
|
||||
// < F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 144, 4, 8, 16, 16, 2, 9, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, true, 7, 1, 1>;
|
||||
// [ 64, 144, 8, 8], 1 stage, 2 occupancy/
|
||||
// 96 TFlops, 240 blocks (1024x2160x3840)
|
||||
// 96 TFlops, 480 blocks (1024x4320x3840)
|
||||
// 99 TFlops,1920 blocks (4096x4320x3840)
|
||||
// < F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 144, 8, 8, 16, 16, 1, 9, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 8, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, true, 7, 1, 1>;
|
||||
// [ 64, 144, 8, 8], 2 stage, 2 occupancy
|
||||
// 93 TFlops
|
||||
// < F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 144, 8, 8, 16, 16, 1, 9, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 8, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, true, 7, 1, 2>;
|
||||
// [ 64, 144, 4, 8], 1 stage, 2 occupancy
|
||||
// 87 TFlops
|
||||
// < F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 144, 4, 8, 16, 16, 1, 9, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, true, 7, 1, 1>;
|
||||
// [ 64, 144, 4, 8], 2 stage, 2 occupancy
|
||||
// 85 TFlops
|
||||
// < F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 144, 4, 8, 16, 16, 1, 9, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, true, 7, 1, 2>;
|
||||
#elif 1
|
||||
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl_C_Shuffle
|
||||
//######| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Num|
|
||||
//######| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| 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| Prefetch|
|
||||
//######| | | | | | | | 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| |
|
||||
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
// [128, 144, 8, 8], 1 stage, 1 occupancy, bounded by LDS size
|
||||
< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 144, 8, 8, 16, 16, 2, 9, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 8, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, true, 1, 9, S<1, 1, 8, 1, 9, 2>, 8, 1>;
|
||||
#endif
|
||||
// clang-format on
|
||||
|
||||
using ReferenceGemmInstance = ck::tensor_operation::host::
|
||||
@@ -198,8 +219,8 @@ int main(int argc, char* argv[])
|
||||
|
||||
float gb_per_sec = num_btype / 1.E6 / ave_time;
|
||||
|
||||
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
|
||||
<< std::endl;
|
||||
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
|
||||
<< gemm.GetTypeString() << std::endl;
|
||||
|
||||
c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
|
||||
|
||||
|
||||
@@ -12,6 +12,7 @@
|
||||
#include "device_tensor.hpp"
|
||||
#include "tensor_layout.hpp"
|
||||
#include "element_wise_operation.hpp"
|
||||
#include "device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp"
|
||||
#include "device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp"
|
||||
#include "reference_conv_fwd.hpp"
|
||||
#include "convolution_utility.hpp"
|
||||
@@ -35,45 +36,41 @@ using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
|
||||
static constexpr auto ConvFwdDefault =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization_t::Default;
|
||||
|
||||
// clang-format off
|
||||
using DeviceConvFwdInstance = ck::tensor_operation::device::
|
||||
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<
|
||||
InDataType, // InDataType
|
||||
WeiDataType, // WeiDataType
|
||||
OutDataType, // OutDataType
|
||||
AccDataType, // AccDataType
|
||||
InElementOp, // InElementwiseOperation
|
||||
WeiElementOp, // WeiElementwiseOperation
|
||||
OutElementOp, // OutElementwiseOperation
|
||||
ConvFwdDefault, // ConvForwardSpecialization
|
||||
256, // BlockSize
|
||||
128, // MPerBlock
|
||||
256, // NPerBlock
|
||||
4, // K0PerBlock
|
||||
8, // K1
|
||||
32, // MPerXdl
|
||||
32, // NPerXdl
|
||||
2, // MXdlPerWave
|
||||
4, // NXdlPerWave
|
||||
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_K0_M_K1
|
||||
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
|
||||
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
|
||||
2, // ABlockTransferSrcVectorDim
|
||||
8, // ABlockTransferSrcScalarPerVector
|
||||
8, // ABlockTransferDstScalarPerVector_K1
|
||||
true, // ABlockLdsAddExtraM
|
||||
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_K0_N_K1
|
||||
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
|
||||
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
|
||||
2, // BBlockTransferSrcVectorDim
|
||||
8, // BBlockTransferSrcScalarPerVector
|
||||
8, // BBlockTransferDstScalarPerVector_K1
|
||||
true, // BBlockLdsAddExtraN
|
||||
1, // CShuffleMXdlPerWavePerShuffle
|
||||
1, // CShuffleNXdlPerWavePerShuffle
|
||||
S<1, 1, 32, 1, 1, 8>, // CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
|
||||
8>; // CBlockTransferScalarPerVector_NWaveNPerXdl
|
||||
// clang-format on
|
||||
DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<
|
||||
InDataType, // InDataType
|
||||
WeiDataType, // WeiDataType
|
||||
OutDataType, // OutDataType
|
||||
AccDataType, // AccDataType
|
||||
InElementOp, // InElementwiseOperation
|
||||
WeiElementOp, // WeiElementwiseOperation
|
||||
OutElementOp, // OutElementwiseOperation
|
||||
ConvFwdDefault, // ConvForwardSpecialization
|
||||
256, // BlockSize
|
||||
128, // MPerBlock
|
||||
256, // NPerBlock
|
||||
4, // K0PerBlock
|
||||
8, // K1
|
||||
32, // MPerXdl
|
||||
32, // NPerXdl
|
||||
2, // MXdlPerWave
|
||||
4, // NXdlPerWave
|
||||
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_K0_M_K1
|
||||
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
|
||||
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
|
||||
2, // ABlockTransferSrcVectorDim
|
||||
8, // ABlockTransferSrcScalarPerVector
|
||||
8, // ABlockTransferDstScalarPerVector_K1
|
||||
true, // ABlockLdsAddExtraM
|
||||
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_K0_N_K1
|
||||
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
|
||||
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
|
||||
2, // BBlockTransferSrcVectorDim
|
||||
8, // BBlockTransferSrcScalarPerVector
|
||||
8, // BBlockTransferDstScalarPerVector_K1
|
||||
true, // BBlockLdsAddExtraN
|
||||
7, // CThreadTransferSrcDstVectorDim
|
||||
1>; // CThreadTransferDstScalarPerVector
|
||||
|
||||
using ReferenceConvFwdInstance = ck::tensor_operation::host::
|
||||
ReferenceConvFwd<InDataType, WeiDataType, OutDataType, InElementOp, WeiElementOp, OutElementOp>;
|
||||
|
||||
Reference in New Issue
Block a user