ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-05 06:01:23 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fusion Conv+Bias+ReLU(+Add) (#62)

Browse Source 
* fix relu

* clean up

* clean up

* adding 1x1 conv

* adding 1x1 conv

* added 1x1 conv

* refactor

* refactor

* refactor

* added profiler for conv+bias+relu+add

* clean up

* adding conv+bias+relu

* adding conv+bias+relu

* added conv+bias+relu

* Update README.md

* update cpu verification

* adding c shuffle

* update static_tensor for dealing with invalid element

* adding c shuffle

* debugging

* fix bug

* convert to fp16 before shuffle

* shuffle more than one M/NRepeat

* clean up

* remove coordinate step hack from GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1

* clean up

* remove coordinate step hack from all gridwise gemm xdl

* clean up coordinate step hack

* clean up coordinate step hack

* ThreadwiseTensorSliceTransfer_v3r2 support pointwise op on both src and dst

* adding output shuffle in conv+bias+relu+add

* update

* added conv+bias+relu+add with c shuffle

* added conv+bias+relu+add with c shuffle

* fix forward_sweep bugs in threadwise copy

* clean up

* refactor

* clean up

* clean up

* added conv_c_shuffle+bias_relu

* clean up

* added conv+bias+relu+atomic_add

* clean up

* clean up

* clean up

* clean up

* clean up

* clean up

* misc fixes; add 1x1 specialization

* clean up

* delete unused device op

* clean up

* add support for odd C value
This commit is contained in:
Chao Liu
2021-12-26 08:43:42 -06:00
committed by GitHub
parent a4f24233e5
commit acbd7bd7c5
90 changed files with 13347 additions and 2639 deletions
Download Patch File Download Diff File Expand all files Collapse all files

100
device_operation/include/device_gemm_xdl.hpp
View File

@@ -2,6 +2,7 @@
#define DEVICE_GEMM_XDL_HPP
#include <iostream>
#include <sstream>
#include "device.hpp"
#include "device_base.hpp"
#include "device_gemm.hpp"
@@ -34,24 +35,22 @@ template <typename ADataType,
ck::index_t NPerXDL,
ck::index_t MXdlPerWave,
ck::index_t NXdlPerWave,
typename ABlockTransferThreadSliceLengths_K0_M_K1,
typename ABlockTransferThreadClusterLengths_K0_M_K1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
ck::index_t ABlockTransferSrcVectorDim,
ck::index_t ABlockTransferSrcScalarPerVector,
ck::index_t ABlockTransferDstScalarPerVector_K1,
typename BBlockTransferThreadSliceLengths_K0_N_K1,
bool ABlockLdsAddExtraM,
typename BBlockTransferThreadClusterLengths_K0_N_K1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
ck::index_t BBlockTransferSrcVectorDim,
ck::index_t BBlockTransferSrcScalarPerVector,
ck::index_t BBlockTransferDstScalarPerVector_K1,
bool BBlockLdsAddExtraN,
ck::index_t CThreadTransferSrcDstVectorDim,
ck::index_t CThreadTransferDstScalarPerVector,
bool ABlockLdsAddExtraM,
bool BBlockLdsAddExtraN>
ck::index_t CThreadTransferDstScalarPerVector>
struct DeviceGemmXdl
: public DeviceGemm<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>
{
@@ -131,45 +130,6 @@ struct DeviceGemmXdl
using BGridDesc_K0_N_K1 = decltype(MakeBGridDescriptor_K0_N_K1(1, 1, 1));
using CGridDesc_M_N = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
// TODO remove these hacks
static constexpr auto a_k0_m_k1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0>{}, // 0+: K0
Sequence<0, 0, 0>{}, // 1+: M
Sequence<0, 0, 0>{}), // 2+: K1
make_tuple(Sequence<0, 0, 0>{}, // 0-: K0
Sequence<0, 0, 0>{}, // 1-: M
Sequence<0, 0, 0>{})); // 2-: K1
static constexpr auto b_k0_n_k1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0>{}, // 0+: K0
Sequence<0, 0, 0>{}, // 1+: N
Sequence<0, 0, 0>{}), // 2+: K1
make_tuple(Sequence<0, 0, 0>{}, // 0-: K0
Sequence<0, 0, 0>{}, // 1-: N
Sequence<0, 0, 0>{})); // 2-: K1
static constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1+: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2+: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3+: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4+: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5+: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6+: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 7+: N2
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2-: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3-: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4-: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5-: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6-: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 7-: N2
static constexpr auto a_k0_m_k1_grid_move_slice_window_step_hacks = Sequence<0, 0, 0>{};
static constexpr auto b_k0_n_k1_grid_move_slice_window_step_hacks = Sequence<0, 0, 0>{};
// GridwiseGemm
using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<
BlockSize,
@@ -191,7 +151,6 @@ struct DeviceGemmXdl
K1,
MXdlPerWave,
NXdlPerWave,
ABlockTransferThreadSliceLengths_K0_M_K1,
ABlockTransferThreadClusterLengths_K0_M_K1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
@@ -199,30 +158,18 @@ struct DeviceGemmXdl
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_K1,
false, // AThreadTransferSrcResetCoordinateAfterRun,
BBlockTransferThreadSliceLengths_K0_N_K1,
ABlockLdsAddExtraM,
BBlockTransferThreadClusterLengths_K0_N_K1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_K1,
false, // BThreadTransferSrcResetCoordinateAfterRun,
false, // BThreadTransferSrcResetCoordinateAfterRun,
BBlockLdsAddExtraN,
Sequence<0, 2, 4, 5, 6, 1, 3, 7>, // CThreadTransferSrcDstAccessOrder,
CThreadTransferSrcDstVectorDim,
CThreadTransferDstScalarPerVector,
decltype(a_k0_m_k1_grid_step_hacks), // AGridStepHacks,
decltype(b_k0_n_k1_grid_step_hacks), // BGridStepHacks,
decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks), // CGridStepHacks,
decltype(a_k0_m_k1_grid_move_slice_window_step_hacks), // AGridMoveSliceWindowStepHacks,
decltype(b_k0_n_k1_grid_move_slice_window_step_hacks), // BGridMoveSliceWindowStepHacks,
false, // CAccessOrderMRepeatNRepeat,
ABlockLdsAddExtraM,
BBlockLdsAddExtraN>;
using CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 =
decltype(GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(CGridDesc_M_N{}));
using Block2CTileMap = decltype(GridwiseGemm::MakeBlock2CTileMap(CGridDesc_M_N{}, 1, 1));
CThreadTransferDstScalarPerVector>;
// Argument
struct Argument : public BaseArgument
@@ -276,8 +223,9 @@ struct DeviceGemmXdl
AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
CGridDesc_M_N c_grid_desc_m_n_;
CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_;
Block2CTileMap block_2_ctile_map_;
typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_;
typename GridwiseGemm::Block2CTileMap block_2_ctile_map_;
index_t M01_;
index_t N01_;
AElementwiseOperation a_element_op_;
@@ -331,11 +279,11 @@ struct DeviceGemmXdl
CDataType,
remove_reference_t<DeviceGemmXdl::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceGemmXdl::BGridDesc_K0_N_K1>,
remove_reference_t<DeviceGemmXdl::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
remove_reference_t<DeviceGemmXdl::Block2CTileMap>,
remove_reference_t<typename GridwiseGemm::Block2CTileMap>,
true>;
ave_time = launch_and_time_kernel(kernel,
@@ -362,11 +310,11 @@ struct DeviceGemmXdl
CDataType,
remove_reference_t<DeviceGemmXdl::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceGemmXdl::BGridDesc_K0_N_K1>,
remove_reference_t<DeviceGemmXdl::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
remove_reference_t<DeviceGemmXdl::Block2CTileMap>,
remove_reference_t<typename GridwiseGemm::Block2CTileMap>,
false>;
ave_time = launch_and_time_kernel(kernel,
@@ -483,6 +431,24 @@ struct DeviceGemmXdl
{
return std::make_unique<Invoker>(Invoker{});
}
// polymorphic
std::string GetTypeString() const override
{
auto str = std::stringstream();
// clang-format off
str << "DeviceGemmXdl"
<< "<"
<< BlockSize << ", "
<< MPerBlock << ", "
<< NPerBlock << ", "
<< K0PerBlock
<< ">";
// clang-format on
return str.str();
}
};
} // namespace device