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

Code refactor (#175)

Browse Source 
* format

* improving pipeline

* fix typo

* format

* adding thread group

* adding thread group

* adding thread group

* adding gemm pipeline

* tweak

* refactor

* refactor

* add missing type convert

* refactor

* refactor

* refactor

* clean

* fix build

* refactor

* format

* clean up

* use remove_cvref_t

* clean

* clean up

* clean up

* clean up
This commit is contained in:
Chao Liu
2022-05-09 14:57:59 -05:00
committed by GitHub
parent a3c910ac6c
commit ec7c2e912e
52 changed files with 1167 additions and 1912 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
include/ck/tensor_operation/gpu/device/device_batched_gemm_reduce_xdl_cshuffle.hpp
View File

@@ -720,11 +720,10 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<AElementwi
const index_t grid_size =
GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_) * arg.BatchCount_;
const auto K0 = arg.a_grid_desc_ak0_m_ak1_.GetLength(I0);
const auto K =
arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
if(has_main_k0_block_loop)
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
{
const auto kernel = kernel_batched_gemm_reduce_xdl_cshuffle_v1<
GridwiseGemm,

7
include/ck/tensor_operation/gpu/device/device_batched_gemm_xdl.hpp
View File

@@ -455,13 +455,12 @@ struct DeviceBatchedGemmXdl
const index_t grid_size =
GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_) * arg.BatchCount_;
const auto K0 = arg.a_grid_desc_k0_m_k1_.GetLength(I0);
const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
const auto K =
arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
float ave_time = 0;
if(has_main_k0_block_loop)
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
{
const auto kernel = kernel_batched_gemm_xdlops_v2r3<
GridwiseGemm,

7
include/ck/tensor_operation/gpu/device/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp
View File

@@ -582,11 +582,10 @@ struct DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
const index_t grid_size =
GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_container_[i]);
const auto K0 = arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I0);
const auto K = arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I0) *
arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I2);
const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
if(has_main_k0_block_loop)
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
{
const auto kernel = kernel_gemm_xdlops_v2r3<
GridwiseGemm,

7
include/ck/tensor_operation/gpu/device/device_conv2d_fwd_xdl_c_shuffle_bias_activation_add_nhwc_kyxc_nhwk.hpp
View File

@@ -698,13 +698,12 @@ struct
const index_t grid_size = GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_);
const auto K0 = arg.a_grid_desc_k0_m_k1_.GetLength(I0);
const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
const auto K =
arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
float ave_time = 0;
if(has_main_k0_block_loop)
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
{
const auto kernel = kernel_gemm_xdlops_v3r3<
GridwiseGemm,

12
include/ck/tensor_operation/gpu/device/device_conv2d_fwd_xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp
View File

@@ -1,6 +1,4 @@
#ifndef DEVICE_CONV2D_FWD_XDL_C_SHUFFLE_BIAS_ACTIVATION_NHWC_KYXC_NHWK_HPP
#define DEVICE_CONV2D_FWD_XDL_C_SHUFFLE_BIAS_ACTIVATION_NHWC_KYXC_NHWK_HPP
#pragma once
#include <iostream>
#include <sstream>
#include "device.hpp"
@@ -660,13 +658,12 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X
const index_t grid_size = GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_);
const auto K0 = arg.a_grid_desc_k0_m_k1_.GetLength(I0);
const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
const auto K =
arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
float ave_time = 0;
if(has_main_k0_block_loop)
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
{
const auto kernel = kernel_gemm_xdlops_v3r2<
GridwiseGemm,
@@ -919,4 +916,3 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X
} // namespace device
} // namespace tensor_operation
} // namespace ck
#endif

7
include/ck/tensor_operation/gpu/device/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp
View File

@@ -640,13 +640,12 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
const index_t grid_size = GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_);
const auto K0 = arg.a_grid_desc_k0_m_k1_.GetLength(I0);
const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
const auto K =
arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
float ave_time = 0;
if(has_main_k0_block_loop)
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
{
const auto kernel = kernel_gemm_xdlops_v3r1<
GridwiseGemm,

7
include/ck/tensor_operation/gpu/device/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp
View File

@@ -478,13 +478,12 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
const index_t grid_size = GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_);
const auto K0 = arg.a_grid_desc_k0_m_k1_.GetLength(I0);
const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
const auto K =
arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
float ave_time = 0;
if(has_main_k0_block_loop)
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
{
const auto kernel = kernel_gemm_xdlops_v2r3<
GridwiseGemm,

7
include/ck/tensor_operation/gpu/device/device_convnd_bwd_data_xdl_ndhwc_kzyxc_ndhwk.hpp
View File

@@ -1296,11 +1296,10 @@ struct DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho
const index_t grid_size =
GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_container_[i]);
const auto K0 = arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I0);
const auto K = arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I0) *
arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I2);
const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
if(has_main_k0_block_loop)
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
{
const auto kernel = kernel_gemm_xdlops_v2r3<
GridwiseGemm,

7
include/ck/tensor_operation/gpu/device/device_convnd_fwd_xdl_nhwc_kyxc_nhwk.hpp
View File

@@ -775,13 +775,12 @@ struct DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
const index_t grid_size = GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_);
const auto K0 = arg.a_grid_desc_k0_m_k1_.GetLength(I0);
const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
const auto K =
arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
float ave_time = 0;
if(has_main_k0_block_loop)
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
{
const auto kernel = kernel_gemm_xdlops_v2r3<
GridwiseGemm,

7
include/ck/tensor_operation/gpu/device/device_gemm_reduce_xdl_cshuffle.hpp
View File

@@ -528,11 +528,10 @@ struct DeviceGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<AElementwiseOpera
const index_t grid_size = GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_);
const auto K0 = arg.a_grid_desc_ak0_m_ak1_.GetLength(I0);
const auto K =
arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
if(has_main_k0_block_loop)
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
{
const auto kernel = kernel_gemm_reduce_xdl_cshuffle_v1<
GridwiseGemm,

9
include/ck/tensor_operation/gpu/device/device_gemm_xdl.hpp
View File

@@ -292,6 +292,7 @@ struct DeviceGemmXdl
float Run(const Argument& arg, int nrepeat = 1)
{
#if 0
{
std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
<< ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
@@ -304,6 +305,7 @@ struct DeviceGemmXdl
std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
<< arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
}
#endif
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
@@ -317,13 +319,12 @@ struct DeviceGemmXdl
const index_t grid_size = GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_);
const auto K0 = arg.a_grid_desc_k0_m_k1_.GetLength(I0);
const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
const auto K =
arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
float ave_time = 0;
if(has_main_k0_block_loop)
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
{
const auto kernel = kernel_gemm_xdlops_v2r3<
GridwiseGemm,

477
include/ck/tensor_operation/gpu/device/device_gemm_xdl_c_shuffle.hpp
View File

@@ -1,477 +0,0 @@
#ifndef DEVICE_GEMM_XDL_C_SHUFFLE_HPP
#define DEVICE_GEMM_XDL_C_SHUFFLE_HPP
#include <iostream>
#include <sstream>
#include "device.hpp"
#include "device_gemm.hpp"
#include "common_header.hpp"
#include "tensor_layout.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "gridwise_gemm_xdlops_v3r1.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
template <
typename ADataType,
typename BDataType,
typename CDataType,
typename AccDataType,
typename CShuffleDataType,
typename ALayout,
typename BLayout,
typename CLayout,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
ck::index_t BlockSize,
ck::index_t MPerBlock,
ck::index_t NPerBlock,
ck::index_t KPerBlock,
ck::index_t AK1,
ck::index_t BK1,
ck::index_t MPerXDL,
ck::index_t NPerXDL,
ck::index_t MXdlPerWave,
ck::index_t NXdlPerWave,
typename ABlockTransferThreadClusterLengths_K0_M_K1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
ck::index_t ABlockTransferSrcVectorDim,
ck::index_t ABlockTransferSrcScalarPerVector,
ck::index_t ABlockTransferDstScalarPerVector_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,
index_t CShuffleMXdlPerWavePerShuffle,
index_t CShuffleNXdlPerWavePerShuffle,
typename CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
index_t CBlockTransferScalarPerVector_NWaveNPerXdl,
index_t NumPrefetch = 1>
struct DeviceGemmXdl_C_Shuffle
: public DeviceGemm<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>
{
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static auto MakeAGridDescriptor_K0_M_K1(index_t M, index_t K, index_t StrideA)
{
assert(K % AK1 == 0);
const index_t K0 = K / AK1;
const auto a_grid_desc_m_k = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
}
}();
const auto a_grid_desc_k0_m_k1 = transform_tensor_descriptor(
a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(K0, AK1)), make_pass_through_transform(M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_k0_m_k1;
}
static auto MakeBGridDescriptor_K0_N_K1(index_t K, index_t N, index_t StrideB)
{
assert(K % BK1 == 0);
const index_t K0 = K / BK1;
const auto b_grid_desc_k_n = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
}
}();
const auto b_grid_desc_k0_n_k1 = transform_tensor_descriptor(
b_grid_desc_k_n,
make_tuple(make_unmerge_transform(make_tuple(K0, BK1)), make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_k0_n_k1;
}
static auto MakeCGridDescriptor_M_N(index_t M, index_t N, index_t StrideC)
{
if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
}
}
using AGridDesc_K0_M_K1 = decltype(MakeAGridDescriptor_K0_M_K1(1, 1, 1));
using BGridDesc_K0_N_K1 = decltype(MakeBGridDescriptor_K0_N_K1(1, 1, 1));
using CGridDesc_M_N = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
// GridwiseGemm
using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1<
BlockSize,
ADataType, // TODO: distinguish A/B datatype
AccDataType,
CShuffleDataType,
CDataType,
InMemoryDataOperationEnum::Set,
AGridDesc_K0_M_K1,
BGridDesc_K0_N_K1,
CGridDesc_M_N,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
MPerBlock,
NPerBlock,
KPerBlock,
AK1,
BK1,
MPerXDL,
NPerXDL,
MXdlPerWave,
NXdlPerWave,
ABlockTransferThreadClusterLengths_K0_M_K1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorDim,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_K1,
false,
ABlockLdsAddExtraM,
BBlockTransferThreadClusterLengths_K0_N_K1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_K1,
false,
BBlockLdsAddExtraN,
CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
CBlockTransferScalarPerVector_NWaveNPerXdl,
NumPrefetch>;
// Argument
struct Argument : public BaseArgument
{
Argument(const ADataType* p_a_grid,
const BDataType* p_b_grid,
CDataType* p_c_grid,
index_t M,
index_t N,
index_t K,
index_t StrideA,
index_t StrideB,
index_t StrideC,
index_t M01,
index_t N01,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
: p_a_grid_{p_a_grid},
p_b_grid_{p_b_grid},
p_c_grid_{p_c_grid},
a_grid_desc_k0_m_k1_{},
b_grid_desc_k0_n_k1_{},
c_grid_desc_m_n_{},
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_{},
block_2_ctile_map_{},
M01_{M01},
N01_{N01},
a_element_op_{a_element_op},
b_element_op_{b_element_op},
c_element_op_{c_element_op}
{
a_grid_desc_k0_m_k1_ =
DeviceGemmXdl_C_Shuffle::MakeAGridDescriptor_K0_M_K1(M, K, StrideA);
b_grid_desc_k0_n_k1_ =
DeviceGemmXdl_C_Shuffle::MakeBGridDescriptor_K0_N_K1(K, N, StrideB);
c_grid_desc_m_n_ = DeviceGemmXdl_C_Shuffle::MakeCGridDescriptor_M_N(M, N, StrideC);
if(GridwiseGemm::CheckValidity(
a_grid_desc_k0_m_k1_, b_grid_desc_k0_n_k1_, c_grid_desc_m_n_, M01_, N01_))
{
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_ =
GridwiseGemm::
MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
c_grid_desc_m_n_);
block_2_ctile_map_ =
GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_, M01, N01);
}
}
// private:
const ADataType* p_a_grid_;
const BDataType* p_b_grid_;
CDataType* p_c_grid_;
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_;
typename GridwiseGemm::
CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_;
typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
index_t M01_;
index_t N01_;
AElementwiseOperation a_element_op_;
BElementwiseOperation b_element_op_;
CElementwiseOperation c_element_op_;
};
// Invoker
struct Invoker : public BaseInvoker
{
using Argument = DeviceGemmXdl_C_Shuffle::Argument;
float Run(const Argument& arg, int nrepeat = 1)
{
{
std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
<< ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
<< arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
std::cout << "arg.b_grid_desc_k0_n_k1_{" << arg.b_grid_desc_k0_n_k1_.GetLength(I0)
<< ", " << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
<< arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
<< arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
}
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_m_n_,
arg.M01_,
arg.N01_))
{
throw std::runtime_error(
"wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v2r3 has invalid setting");
}
const index_t grid_size = GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_);
const auto K0 = arg.a_grid_desc_k0_m_k1_.GetLength(I0);
const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
float ave_time = 0;
if(has_main_k0_block_loop)
{
const auto kernel = kernel_gemm_xdlops_v3r1<
GridwiseGemm,
ADataType, // TODO: distiguish A/B datatype
CDataType,
remove_reference_t<DeviceGemmXdl_C_Shuffle::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceGemmXdl_C_Shuffle::BGridDesc_K0_N_K1>,
remove_reference_t<
typename GridwiseGemm::
CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl>,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
true>;
ave_time = launch_and_time_kernel(
kernel,
nrepeat,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_,
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_,
arg.block_2_ctile_map_);
}
else
{
const auto kernel = kernel_gemm_xdlops_v3r1<
GridwiseGemm,
ADataType, // TODO: distiguish A/B datatype
CDataType,
remove_reference_t<DeviceGemmXdl_C_Shuffle::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceGemmXdl_C_Shuffle::BGridDesc_K0_N_K1>,
remove_reference_t<
typename GridwiseGemm::
CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl>,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
false>;
ave_time = launch_and_time_kernel(
kernel,
nrepeat,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_,
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_,
arg.block_2_ctile_map_);
}
return ave_time;
}
// polymorphic
float Run(const BaseArgument* p_arg, int nrepeat = 1) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), nrepeat);
}
};
static constexpr bool IsValidCompilationParameter()
{
// TODO: properly implement this check
return true;
}
static bool IsSupportedArgument(const Argument& arg)
{
return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_m_n_,
arg.M01_,
arg.N01_);
}
// polymorphic
bool IsSupportedArgument(const BaseArgument* p_arg) override
{
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
}
static auto MakeArgument(const ADataType* p_a,
const BDataType* p_b,
CDataType* p_c,
index_t M,
index_t N,
index_t K,
index_t StrideA,
index_t StrideB,
index_t StrideC,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
{
return Argument{p_a,
p_b,
p_c,
M,
N,
K,
StrideA,
StrideB,
StrideC,
1,
1,
a_element_op,
b_element_op,
c_element_op};
}
static auto MakeInvoker() { return Invoker{}; }
// polymorphic
std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
const void* p_b,
void* p_c,
index_t M,
index_t N,
index_t K,
index_t StrideA,
index_t StrideB,
index_t StrideC,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
index_t /* KBatch */ = 1) override
{
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
static_cast<const BDataType*>(p_b),
static_cast<CDataType*>(p_c),
M,
N,
K,
StrideA,
StrideB,
StrideC,
1,
1,
a_element_op,
b_element_op,
c_element_op);
}
// polymorphic
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
{
return std::make_unique<Invoker>(Invoker{});
}
// polymorphic
std::string GetTypeString() const override
{
auto str = std::stringstream();
// clang-format off
str << "DeviceGemmXdl_C_Shuffle"
<< "<"
<< BlockSize << ", "
<< MPerBlock << ", "
<< NPerBlock << ", "
<< KPerBlock << ", "
<< AK1 << ", "
<< BK1
<< ">";
// clang-format on
return str.str();
}
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
#endif

14
include/ck/tensor_operation/gpu/device/device_gemm_xdl_c_shuffle_bias_2d.hpp
View File

@@ -1,6 +1,4 @@
#ifndef DEVICE_GEMM_XDL_C_SHUFFLE_BIAS_2D_HPP
#define DEVICE_GEMM_XDL_C_SHUFFLE_BIAS_2D_HPP
#pragma once
#include <iostream>
#include <sstream>
#include "device.hpp"
@@ -291,18 +289,17 @@ struct DeviceGemmXdl_C_Shuffle_Bias_2d
arg.N01_))
{
throw std::runtime_error(
"wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v2r3 has invalid setting");
"wrong! GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2 has invalid setting");
}
const index_t grid_size = GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_);
const auto K0 = arg.a_grid_desc_k0_m_k1_.GetLength(I0);
const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
const auto K =
arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
float ave_time = 0;
if(has_main_k0_block_loop)
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
{
const auto kernel = kernel_gemm_xdlops_v3r2<
GridwiseGemm,
@@ -505,4 +502,3 @@ struct DeviceGemmXdl_C_Shuffle_Bias_2d
} // namespace device
} // namespace tensor_operation
} // namespace ck
#endif

7
include/ck/tensor_operation/gpu/device/device_gemm_xdl_c_shuffle_bias_activation.hpp
View File

@@ -303,13 +303,12 @@ struct DeviceGemmXdl_C_Shuffle_Bias_Activation
const index_t grid_size = GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_);
const auto K0 = arg.a_grid_desc_k0_m_k1_.GetLength(I0);
const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
const auto K =
arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
float ave_time = 0;
if(has_main_k0_block_loop)
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
{
const auto kernel = kernel_gemm_xdlops_v3r2<
GridwiseGemm,

7
include/ck/tensor_operation/gpu/device/device_gemm_xdl_c_shuffle_bias_activation_add.hpp
View File

@@ -345,13 +345,12 @@ struct DeviceGemmXdl_C_Shuffle_Bias_Activation_Add
const index_t grid_size = GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_);
const auto K0 = arg.a_grid_desc_k0_m_k1_.GetLength(I0);
const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
const auto K =
arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
float ave_time = 0;
if(has_main_k0_block_loop)
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
{
const auto kernel = kernel_gemm_xdlops_v3r3<
GridwiseGemm,

7
include/ck/tensor_operation/gpu/device/device_gemm_xdl_cshuffle.hpp
View File

@@ -462,13 +462,12 @@ struct DeviceGemm_Xdl_CShuffle
const index_t grid_size = GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_);
const auto K0 = arg.a_grid_desc_ak0_m_ak1_.GetLength(I0);
const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
const auto K =
arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
float ave_time = 0;
if(has_main_k0_block_loop)
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
{
const auto kernel = kernel_gemm_xdl_cshuffle_v1<
GridwiseGemm,

134
include/ck/tensor_operation/gpu/device/device_grouped_gemm_xdl.hpp
View File

@@ -17,6 +17,88 @@ namespace ck {
namespace tensor_operation {
namespace device {
template <typename GridwiseGemm,
typename FloatAB,
typename FloatC,
typename GemmDesc,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
bool HasMainKBlockLoop,
index_t MaxGroupCount>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
kernel_grouped_gemm_xdlops_v2r3(
const StaticallyIndexedArray<GemmDesc, MaxGroupCount> gemm_descs,
const index_t group_count,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const CElementwiseOperation c_element_op)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
const index_t block_id = get_block_1d_id();
#if 1
static_for<0, MaxGroupCount, 1>{}([&](auto i) {
if(block_id >= gemm_descs[i].BlockStart_ && block_id < gemm_descs[i].BlockEnd_ &&
i < group_count)
{
auto group_id = i;
GridwiseGemm::template Run<HasMainKBlockLoop>(
gemm_descs[group_id].a_ptr,
gemm_descs[group_id].b_ptr,
gemm_descs[group_id].c_ptr,
p_shared,
gemm_descs[group_id].a_grid_desc_k0_m_k1_,
gemm_descs[group_id].b_grid_desc_k0_n_k1_,
gemm_descs[group_id].c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
a_element_op,
b_element_op,
c_element_op,
gemm_descs[group_id].grouped_gemm_block_2_ctile_map_);
}
});
#else
const auto gemm_desc_ptr = reinterpret_cast<const GemmDesc*>(&gemm_descs);
index_t group_id = 0;
static_for<0, MaxGroupCount, 1>{}([&](auto i) {
group_id = (block_id >= gemm_descs[i].BlockStart && block_id < gemm_descs[i].BlockEnd &&
i < group_count)
? i
: group_id;
});
const index_t block_id_grp = block_id - gemm_desc_ptr[group_id].BlockStart;
GridwiseGemm::template Run<HasMainKBlockLoop>(
gemm_desc_ptr[group_id].a_ptr,
gemm_desc_ptr[group_id].b_ptr,
gemm_desc_ptr[group_id].c_ptr,
p_shared,
gemm_desc_ptr[group_id].a_grid_desc_k0_m_k1_,
gemm_desc_ptr[group_id].b_grid_desc_k0_n_k1_,
gemm_desc_ptr[group_id].c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
a_element_op,
b_element_op,
c_element_op,
gemm_desc_ptr[group_id].block_2_ctile_map_,
block_id_grp);
#endif
#else
ignore = gemm_descs;
ignore = group_count;
ignore = a_element_op;
ignore = b_element_op;
ignore = c_element_op;
#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
}
template <typename ADataType,
typename BDataType,
typename CDataType,
@@ -368,59 +450,53 @@ struct DeviceGroupedGemmXdl
float Run(const Argument& arg, int nrepeat = 1)
{
StaticallyIndexedArray<GemmDescKernelArg, MaxGroupCount> gemm_desc_kernel_arg_arg;
StaticallyIndexedArray<GemmDescKernelArg, MaxGroupCount> gemm_desc_kernel_args;
bool has_main_k0_block_loop = true;
bool has_main_k_block_loop = true;
static_for<0, MaxGroupCount, 1>{}([&](auto i) {
if(i < arg.gemm_desc_kernel_arg_.size())
{
gemm_desc_kernel_arg_arg(i) = arg.gemm_desc_kernel_arg_[i];
gemm_desc_kernel_args(i) = arg.gemm_desc_kernel_arg_[i];
std::cout << "group: " << i << " arg.a_grid_desc_k0_m_k1_{"
<< gemm_desc_kernel_arg_arg[i].a_grid_desc_k0_m_k1_.GetLength(I0)
<< ", "
<< gemm_desc_kernel_arg_arg[i].a_grid_desc_k0_m_k1_.GetLength(I1)
<< ", "
<< gemm_desc_kernel_arg_arg[i].a_grid_desc_k0_m_k1_.GetLength(I2)
<< "}";
<< gemm_desc_kernel_args[i].a_grid_desc_k0_m_k1_.GetLength(I0) << ", "
<< gemm_desc_kernel_args[i].a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
<< gemm_desc_kernel_args[i].a_grid_desc_k0_m_k1_.GetLength(I2) << "}";
std::cout << ", arg.b_grid_desc_k0_n_k1_{"
<< gemm_desc_kernel_arg_arg[i].b_grid_desc_k0_n_k1_.GetLength(I0)
<< ", "
<< gemm_desc_kernel_arg_arg[i].b_grid_desc_k0_n_k1_.GetLength(I1)
<< ", "
<< gemm_desc_kernel_arg_arg[i].b_grid_desc_k0_n_k1_.GetLength(I2)
<< "}";
<< gemm_desc_kernel_args[i].b_grid_desc_k0_n_k1_.GetLength(I0) << ", "
<< gemm_desc_kernel_args[i].b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
<< gemm_desc_kernel_args[i].b_grid_desc_k0_n_k1_.GetLength(I2) << "}";
std::cout << ", arg.c_grid_desc_m_n_{ "
<< gemm_desc_kernel_arg_arg[i].c_grid_desc_m_n_.GetLength(I0) << ", "
<< gemm_desc_kernel_arg_arg[i].c_grid_desc_m_n_.GetLength(I1) << "}"
<< gemm_desc_kernel_args[i].c_grid_desc_m_n_.GetLength(I0) << ", "
<< gemm_desc_kernel_args[i].c_grid_desc_m_n_.GetLength(I1) << "}"
<< std::endl;
if(!GridwiseGemm::CheckValidity(
gemm_desc_kernel_arg_arg[i].a_grid_desc_k0_m_k1_,
gemm_desc_kernel_arg_arg[i].b_grid_desc_k0_n_k1_,
gemm_desc_kernel_arg_arg[i].c_grid_desc_m_n_,
arg.M01_,
arg.N01_))
if(!GridwiseGemm::CheckValidity(gemm_desc_kernel_args[i].a_grid_desc_k0_m_k1_,
gemm_desc_kernel_args[i].b_grid_desc_k0_n_k1_,
gemm_desc_kernel_args[i].c_grid_desc_m_n_,
arg.M01_,
arg.N01_))
{
throw std::runtime_error(
"wrong! GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3 has invalid setting");
}
const auto K0 = gemm_desc_kernel_arg_arg[i].a_grid_desc_k0_m_k1_.GetLength(I0);
const auto K = gemm_desc_kernel_args[i].a_grid_desc_k0_m_k1_.GetLength(I0) *
gemm_desc_kernel_args[i].a_grid_desc_k0_m_k1_.GetLength(I2);
if(GridwiseGemm::CalculateHasMainK0BlockLoop(K0) != has_main_k0_block_loop)
if(GridwiseGemm::CalculateHasMainKBlockLoop(K) != has_main_k_block_loop)
{
throw std::runtime_error("wrong! not all gemm has_main_k0_block_loop");
throw std::runtime_error("wrong! not all gemm has_main_k_block_loop");
}
}
});
float ave_time = 0;
if(has_main_k0_block_loop)
if(has_main_k_block_loop)
{
const auto kernel =
kernel_grouped_gemm_xdlops_v2r3<GridwiseGemm,
@@ -438,7 +514,7 @@ struct DeviceGroupedGemmXdl
dim3(arg.grid_size_),
dim3(BlockSize),
0,
gemm_desc_kernel_arg_arg,
gemm_desc_kernel_args,
arg.gemm_desc_kernel_arg_.size(),
arg.a_element_op_,
arg.b_element_op_,
@@ -462,7 +538,7 @@ struct DeviceGroupedGemmXdl
dim3(arg.grid_size_),
dim3(BlockSize),
0,
gemm_desc_kernel_arg_arg,
gemm_desc_kernel_args,
arg.gemm_desc_kernel_arg_.size(),
arg.a_element_op_,
arg.b_element_op_,