Skip to content

GitLab

Commit f0298581 authored by Harisankar Sadasivan's avatar Harisankar Sadasivan
Browse files

cmakelist changes to exclude navi cards for gemv splitk & merge changes from dev

parent 675aa69e
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 2706 additions and 196 deletions
example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt
View file @ f0298581
......@@ -3,44 +3,38 @@ list(APPEND gpu_list2 gfx1100 gfx1101 gfx1102)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list1 AND target EQUAL 0)
add_custom_target(example_grouped_conv_fwd_multiple_d)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_fp16 grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_fp16)
endif()
add_example_executable(example_grouped_conv_fwd_xdl_fp16 grouped_conv_fwd_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_xdl_fp16)
endif()
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_fp32 grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_fp32)
endif()
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_bf16 grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_bf16)
endif()
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_int8 grouped_conv_fwd_bias_relu_add_xdl_int8.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_int8)
endif()
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_int4 grouped_conv_fwd_bias_relu_add_xdl_int4.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_int4)
endif()
endif() # USE_BITINT_EXTENSION_INT4
set(target 1)
endif()
if(gpu IN_LIST gpu_list1 AND target EQUAL 0)
add_custom_target(example_grouped_conv_fwd_multiple_d)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_fp16 grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp)
add_example_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_fp16)
add_example_executable(example_grouped_conv_fwd_xdl_fp16 grouped_conv_fwd_xdl_fp16.cpp)
add_example_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_xdl_fp16)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_fp32 grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp)
add_example_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_fp32)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_bf16 grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp)
add_example_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_bf16)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_int8 grouped_conv_fwd_bias_relu_add_xdl_int8.cpp)
add_example_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_int8)
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_int4 grouped_conv_fwd_bias_relu_add_xdl_int4.cpp)
add_example_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_int4)
endif() # USE_BITINT_EXTENSION_INT4
set(target 1)
endif()
endforeach()
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_fp16 grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_int8 grouped_conv_fwd_bias_relu_add_wmma_int8.cpp)
set(target 1)
endif()
if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_fp16 grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_int8 grouped_conv_fwd_bias_relu_add_wmma_int8.cpp)
set(target 1)
endif()
endforeach()
example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt
View file @ f0298581
add_custom_target(example_gemm_scale_softmax_gemm)
add_example_executable(example_batched_gemm_scale_softmax_gemm_xdl_fp16 batched_gemm_scale_softmax_gemm_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_fp16)
endif()
add_example_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_fp16)
add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16)
endif()
add_example_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16)
add_example_executable(example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16 grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16)
endif()
add_example_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16)
add_example_executable(example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)
endif()
add_example_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)
add_example_executable(example_grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16 grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)
endif()
add_example_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)
add_example_executable(example_batched_gemm_scale_softmax_gemm_xdl_bf16 batched_gemm_scale_softmax_gemm_xdl_bf16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_bf16)
endif()
add_example_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_bf16)
add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_xdl_bf16 batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_bf16)
endif()
add_example_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_bf16)
example/35_splitK_gemm/CMakeLists.txt
View file @ f0298581
......@@ -4,28 +4,23 @@ foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_custom_target(example_splitK_gemm_xdl)
add_example_executable(example_splitK_gemm_xdl_fp32 splitK_gemm_xdl_fp32.cpp)
if(result EQUAL 0)
add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_fp32)
endif()
add_example_executable(example_splitK_gemm_xdl_fp16 splitK_gemm_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_fp16)
endif()
add_example_executable(example_splitK_gemm_xdl_bf16 splitK_gemm_xdl_bf16.cpp)
if(result EQUAL 0)
add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_bf16)
endif()
add_example_executable(example_splitK_gemm_xdl_int8 splitK_gemm_xdl_int8.cpp)
if(result EQUAL 0)
add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_int8)
endif()
add_example_executable(example_splitK_gemm_xdl_fp32 splitK_gemm_xdl_fp32.cpp)
add_example_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_fp32)
add_example_executable(example_splitK_gemm_xdl_fp16 splitK_gemm_xdl_fp16.cpp)
add_example_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_fp16)
add_example_executable(example_splitK_gemm_xdl_bf16 splitK_gemm_xdl_bf16.cpp)
add_example_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_bf16)
add_example_executable(example_splitK_gemm_xdl_int8 splitK_gemm_xdl_int8.cpp)
add_example_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_int8)
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_splitK_gemm_xdl_int4 splitK_gemm_xdl_int4.cpp)
if(result EQUAL 0)
add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_int4)
endif()
add_example_executable(example_splitK_gemm_xdl_int4 splitK_gemm_xdl_int4.cpp)
add_example_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_int4)
endif()
set(target 1)
endif()
endforeach()
example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt
View file @ f0298581
......@@ -2,27 +2,26 @@ list(APPEND gpu_list_xdl gfx908 gfx90a gfx940 gfx941 gfx942)
list(APPEND gpu_list_wmma gfx1100 gfx1101 gfx1102)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list_xdl AND target EQUAL 0)
add_custom_target(example_grouped_conv_bwd_data)
add_example_executable(example_grouped_conv_bwd_data_xdl_fp16 grouped_conv_bwd_data_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_xdl_fp16)
endif()
add_example_executable(example_grouped_conv_bwd_data_bias_relu_xdl_fp16 grouped_conv_bwd_data_bias_relu_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_bias_relu_xdl_fp16)
endif()
set(target 1)
endif()
if(gpu IN_LIST gpu_list_xdl AND target EQUAL 0)
add_custom_target(example_grouped_conv_bwd_data)
add_example_executable(example_grouped_conv_bwd_data_xdl_fp16 grouped_conv_bwd_data_xdl_fp16.cpp)
add_example_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_xdl_fp16)
add_example_executable(example_grouped_conv_bwd_data_bias_relu_xdl_fp16 grouped_conv_bwd_data_bias_relu_xdl_fp16.cpp)
add_example_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_bias_relu_xdl_fp16)
set(target 1)
endif()
endforeach()
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list_wmma AND target EQUAL 0)
add_custom_target(example_grouped_conv_bwd_data)
add_example_executable(example_grouped_conv_bwd_data_wmma_fp16 grouped_conv_bwd_data_wmma_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_wmma_fp16)
endif()
set(target 1)
endif()
if(gpu IN_LIST gpu_list_wmma AND target EQUAL 0)
add_custom_target(example_grouped_conv_bwd_data)
add_example_executable(example_grouped_conv_bwd_data_wmma_fp16 grouped_conv_bwd_data_wmma_fp16.cpp)
add_example_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_wmma_fp16)
set(target 1)
endif()
endforeach()
example/39_permute/CMakeLists.txt
View file @ f0298581
add_custom_target(example_permute)
add_example_executable(example_permute_1xHxW_fp16 permute_1xHxW_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_permute example_permute_1xHxW_fp16)
endif()
add_example_dependencies(example_permute example_permute_1xHxW_fp16)
add_example_executable(example_permute_NxHxW_fp16 permute_NxHxW_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_permute example_permute_NxHxW_fp16)
endif()
add_example_dependencies(example_permute example_permute_NxHxW_fp16)
add_example_executable(example_permute_HxWx4_fp16 permute_HxWx4_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_permute example_permute_HxWx4_fp16)
endif()
add_example_dependencies(example_permute example_permute_HxWx4_fp16)
example/52_im2col_col2im/CMakeLists.txt
View file @ f0298581
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_custom_target(example_im2col_col2im)
add_example_executable(example_image_to_column_f32 image_to_column_f32.cpp)
add_dependencies(example_im2col_col2im example_image_to_column_f32)
add_example_executable(example_column_to_image_f32 column_to_image_f32.cpp)
add_dependencies(example_im2col_col2im example_column_to_image_f32)
set(target 1)
endif()
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_custom_target(example_im2col_col2im)
add_example_executable(example_image_to_column_f32 image_to_column_f32.cpp)
add_example_dependencies(example_im2col_col2im example_image_to_column_f32)
add_example_executable(example_column_to_image_f32 column_to_image_f32.cpp)
add_example_dependencies(example_im2col_col2im example_column_to_image_f32)
set(target 1)
endif()
endforeach()
example/53_gemv_splitk/CMakeLists.txt
View file @ f0298581
add_custom_target(example_gemv_splitk)
add_example_executable(example_gemv_splitk_fp16 gemv_splitk_fp16.cpp)
add_dependencies(example_gemv_splitk
example_gemv_splitk_fp16)
\ No newline at end of file
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_custom_target(example_gemv_splitk)
add_example_executable(example_gemv_splitk_fp16 gemv_splitk_fp16.cpp)
add_dependencies(example_gemv_splitk
example_gemv_splitk_fp16)
set(target 1)
endif()
endforeach()
example/60_gemm_multi_ABD/CMakeLists.txt 0 → 100755
View file @ f0298581
list(APPEND gpu_list2 gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
add_example_executable(example_gemm_multi_ABD_xdl_fp16 gemm_multi_ABD_xdl_fp16.cpp)
set(target 1)
endif()
endforeach()
example/60_gemm_multi_ABD/gemm_multi_ABD_xdl_fp16.cpp 0 → 100755
View file @ f0298581
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_abd_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.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 = F16;
using BDataType = F16;
using AccDataType = F32;
using CShuffleDataType = F32;
using DDataType = F16;
using EDataType = F16;
using ALayout = Row;
using BLayout = Col;
using DLayout = Row;
using ELayout = Row;
struct AddScale
{
static constexpr auto I0 = ck::Number<0>{};
static constexpr auto I1 = ck::Number<1>{};
static constexpr auto I2 = ck::Number<2>{};
static constexpr auto I3 = ck::Number<3>{};
__host__ __device__ constexpr void
operator()(ck::half4_t& a, const ck::half4_t& a0, const ck::half4_t& a1) const
{
const auto a0_v_t = ck::vector_type<ck::half_t, 4>{a0};
const auto a1_v_t = ck::vector_type<ck::half_t, 4>{a1};
auto r_v_t = ck::vector_type<ck::half_t, 4>{};
r_v_t.AsType<ck::half_t>()(I0) =
scale * (a0_v_t.AsType<ck::half_t>()[I0] + a1_v_t.AsType<ck::half_t>()[I0]);
r_v_t.AsType<ck::half_t>()(I1) =
scale * (a0_v_t.AsType<ck::half_t>()[I1] + a1_v_t.AsType<ck::half_t>()[I1]);
r_v_t.AsType<ck::half_t>()(I2) =
scale * (a0_v_t.AsType<ck::half_t>()[I2] + a1_v_t.AsType<ck::half_t>()[I2]);
r_v_t.AsType<ck::half_t>()(I3) =
scale * (a0_v_t.AsType<ck::half_t>()[I3] + a1_v_t.AsType<ck::half_t>()[I3]);
a = r_v_t.AsType<ck::half4_t>()[I0];
}
__host__ __device__ constexpr void
operator()(ck::half_t& a, const ck::half_t& a0, const ck::half_t& a1) const
{
a = scale * (a0 + a1);
}
// this attribute controls the copy_function applying element_wise_op with
// pack4_data
constexpr const static bool is_pack4_invocable = true;
float scale = 1.0;
};
struct AlphaBetaAdd
{
AlphaBetaAdd(float alpha, float beta) : alpha_(alpha), beta_(beta){};
template <typename E, typename C, typename D>
__host__ __device__ constexpr void operator()(E& e, const C& c, const D& d) const;
template <>
__host__ __device__ constexpr void operator()<ck::half_t, float, ck::half_t>(
ck::half_t& e, const float& c, const ck::half_t& d) const
{
e = ck::type_convert<ck::half_t>(alpha_ * c + beta_ * ck::type_convert<float>(d));
};
float alpha_;
float beta_;
};
using AElementOp = AddScale;
using BElementOp = PassThrough;
using CDEElementOp = AlphaBetaAdd;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleABD_Xdl_CShuffle<
ck::Tuple<ALayout, ALayout>,
ck::Tuple<BLayout>,
ck::Tuple<DLayout>,
ELayout,
ck::Tuple<ADataType, ADataType>,
ck::Tuple<BDataType>,
AccDataType,
CShuffleDataType,
ck::Tuple<DDataType>,
EDataType,
AElementOp,
BElementOp,
CDEElementOp,
GemmSpec,
1,
256,
256,
128,
32,
8,
8,
32,
32,
4,
2,
S<4, 64, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
1,
S<4, 64, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
1,
1,
1,
S<1, 32, 1, 8>,
8>;
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 3840;
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t StrideA = 4096;
ck::index_t StrideB = 4096;
ck::index_t StrideD = 4096;
ck::index_t StrideE = 4096;
float alpha = 1.0f;
float beta = 1.0f;
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 6)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
alpha = std::stof(argv[4]);
beta = std::stof(argv[5]);
}
else if(argc == 13)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
StrideA = std::stoi(argv[7]);
StrideB = std::stoi(argv[8]);
StrideD = std::stoi(argv[9]);
StrideE = std::stoi(argv[10]);
alpha = std::stof(argv[11]);
beta = std::stof(argv[12]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE, alpha, "
"beta\n");
exit(0);
}
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
Tensor<ADataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<ADataType> a1_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<DDataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, DLayout{}));
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
std::cout << "a0_m_k: " << a0_m_k.mDesc << std::endl;
std::cout << "a1_m_k: " << a1_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a0_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
a1_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d_m_n.GenerateTensorValue(GeneratorTensor_2<DDataType>{-5, 5});
break;
default:
a0_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
a1_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d_m_n.GenerateTensorValue(GeneratorTensor_3<DDataType>{-0.5, 0.5});
}
DeviceMem a0_device_buf(sizeof(ADataType) * a0_m_k.mDesc.GetElementSpaceSize());
DeviceMem a1_device_buf(sizeof(ADataType) * a1_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem d_device_buf(sizeof(DDataType) * d_m_n.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
a0_device_buf.ToDevice(a0_m_k.mData.data());
a1_device_buf.ToDevice(a1_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
d_device_buf.ToDevice(d_m_n.mData.data());
e_device_buf.ToDevice(e_m_n_device_result.mData.data());
auto a_element_op = AElementOp{0.2};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{alpha, beta};
// do GEMM
auto device_op = DeviceOpInstance{};
auto invoker = device_op.MakeInvoker();
auto argument =
device_op.MakeArgument(std::array<const void*, 2>{a0_device_buf.GetDeviceBuffer(),
a1_device_buf.GetDeviceBuffer()},
std::array<const void*, 1>{b_device_buf.GetDeviceBuffer()},
std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
std::array<ck::index_t, 2>{StrideA, StrideA},
std::array<ck::index_t, 1>{StrideB},
std::array<ck::index_t, 1>{StrideD},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
if(!device_op.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
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;
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
if(do_verification)
{
Tensor<CShuffleDataType> c_m_n({M, N});
Tensor<ADataType> a_m_k({M, K});
for(int m = 0; m < M; ++m)
{
for(int k = 0; k < K; ++k)
{
a_element_op(a_m_k(m, k), a0_m_k(m, k), a1_m_k(m, k));
}
}
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CShuffleDataType,
AccDataType,
PassThrough,
BElementOp,
PassThrough>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument =
ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, PassThrough{}, b_element_op, PassThrough{});
ref_invoker.Run(ref_argument);
for(int m = 0; m < M; ++m)
{
for(int n = 0; n < N; ++n)
{
cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d_m_n(m, n));
}
}
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
}
return 0;
}
example/61_contraction_multi_ABD/CMakeLists.txt 0 → 100755
View file @ f0298581
list(APPEND gpu_list2 gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
add_example_executable(example_contraction_multi_ABD_xdl_fp16 contraction_multi_ABD_xdl_fp16.cpp)
set(target 1)
endif()
endforeach()
example/61_contraction_multi_ABD/contraction_multi_ABD_xdl_fp16.cpp 0 → 100755
View file @ f0298581
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_abd_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_contraction.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/numeric.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using A0DataType = F16;
using A1DataType = F32;
using BDataType = F16;
using AccDataType = F32;
using CShuffleDataType = F32;
using DDataType = F16;
using EDataType = F16;
static constexpr ck::index_t NumDimM = 2;
static constexpr ck::index_t NumDimN = 2;
static constexpr ck::index_t NumDimK = 2;
struct AlphaBetaAdd
{
AlphaBetaAdd(float alpha, float beta) : alpha_(alpha), beta_(beta){};
template <typename E, typename C, typename D>
__host__ __device__ constexpr void operator()(E& e, const C& c, const D& d) const;
template <>
__host__ __device__ constexpr void operator()<ck::half_t, float, ck::half_t>(
ck::half_t& e, const float& c, const ck::half_t& d) const
{
e = ck::type_convert<ck::half_t>(alpha_ * c + beta_ * ck::type_convert<float>(d));
};
float alpha_;
float beta_;
};
struct Multiply
{
__host__ __device__ constexpr void
operator()(ck::half_t& a, const ck::half_t& a0, const float& a1) const
{
a = ck::type_convert<ck::half_t>(ck::type_convert<float>(a0) * a1);
}
};
using AElementOp = Multiply;
using BElementOp = PassThrough;
using CDEElementOp = AlphaBetaAdd;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
using DeviceOpInstance = ck::tensor_operation::device::DeviceContractionMultipleABD_Xdl_CShuffle<
NumDimM,
NumDimN,
NumDimK,
ck::Tuple<A0DataType, A1DataType>,
ck::Tuple<BDataType>,
AccDataType,
CShuffleDataType,
ck::Tuple<DDataType>,
EDataType,
AElementOp,
BElementOp,
CDEElementOp,
GemmSpec,
1,
256,
256,
128,
32,
8,
8,
32,
32,
4,
2,
S<4, 64, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
1,
S<4, 64, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
1,
1,
1,
S<1, 32, 1, 8>,
8>;
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
float alpha = 1.0f;
float beta = 1.0f;
// A0[M0, M1, K0, K1]
std::vector<ck::index_t> a0_ms_ks_lengths{30, 128, 32, 64};
std::vector<ck::index_t> a0_ms_ks_strides{128 * 32 * 64, 32 * 64, 64, 1};
// A1[M1, K1] -> A1[M0, M1, K0, K1]
std::vector<ck::index_t> a1_ms_ks_lengths{30, 128, 32, 64};
std::vector<ck::index_t> a1_ms_ks_strides{0, 64, 0, 1};
// B[N0, N1, K0, K1]
std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
std::vector<ck::index_t> b_ns_ks_strides{64 * 32 * 64, 32 * 64, 64, 1};
// D[M0, M1, N0, N1]
std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
std::vector<ck::index_t> d_ms_ns_strides{128 * 32 * 64, 32 * 64, 64, 1};
// E[M0, M1, N0, N1]
std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
std::vector<ck::index_t> e_ms_ns_strides{128 * 32 * 64, 32 * 64, 64, 1};
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
exit(0);
}
Tensor<A0DataType> a0_ms_ks(a0_ms_ks_lengths, a0_ms_ks_strides);
Tensor<A1DataType> a1_ms_ks(a1_ms_ks_lengths, a1_ms_ks_strides);
Tensor<BDataType> b_ns_ks(b_ns_ks_lengths, b_ns_ks_strides);
Tensor<EDataType> d_ms_ns(d_ms_ns_lengths, d_ms_ns_strides);
Tensor<EDataType> e_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
Tensor<EDataType> e_ms_ns_device_result(e_ms_ns_lengths, e_ms_ns_strides);
std::cout << "a0_ms_ks: " << a0_ms_ks.mDesc << std::endl;
std::cout << "a1_ms_ks: " << a1_ms_ks.mDesc << std::endl;
std::cout << "b_ns_ks: " << b_ns_ks.mDesc << std::endl;
std::cout << "d_ms_ns: " << d_ms_ns.mDesc << std::endl;
std::cout << "e_ms_ns: " << e_ms_ns_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a0_ms_ks.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-5, 5});
a1_ms_ks.GenerateTensorValue(GeneratorTensor_2<A1DataType>{-5, 5});
b_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d_ms_ns.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
default:
a0_ms_ks.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
a1_ms_ks.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0.0, 1.0});
b_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d_ms_ns.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
}
DeviceMem a0_device_buf(sizeof(A0DataType) * a0_ms_ks.mDesc.GetElementSpaceSize());
DeviceMem a1_device_buf(sizeof(A1DataType) * a1_ms_ks.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_ns_ks.mDesc.GetElementSpaceSize());
DeviceMem d_device_buf(sizeof(DDataType) * d_ms_ns.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_ms_ns_device_result.mDesc.GetElementSpaceSize());
a0_device_buf.ToDevice(a0_ms_ks.mData.data());
a1_device_buf.ToDevice(a1_ms_ks.mData.data());
b_device_buf.ToDevice(b_ns_ks.mData.data());
d_device_buf.ToDevice(d_ms_ns.mData.data());
// set zero
e_device_buf.SetZero();
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{alpha, beta};
// do GEMM
auto device_op = DeviceOpInstance{};
auto invoker = device_op.MakeInvoker();
auto argument = device_op.MakeArgument(
std::array<const void*, 2>{a0_device_buf.GetDeviceBuffer(),
a1_device_buf.GetDeviceBuffer()},
std::array<const void*, 1>{b_device_buf.GetDeviceBuffer()},
std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
std::array<std::vector<ck::index_t>, 2>{a0_ms_ks_lengths, a1_ms_ks_lengths},
std::array<std::vector<ck::index_t>, 2>{a0_ms_ks_strides, a1_ms_ks_strides},
std::array<std::vector<ck::index_t>, 1>{b_ns_ks_lengths},
std::array<std::vector<ck::index_t>, 1>{b_ns_ks_strides},
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
e_ms_ns_lengths,
e_ms_ns_strides,
a_element_op,
b_element_op,
cde_element_op);
if(!device_op.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_contraction with the specified compilation parameters does "
"not support this problem");
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
if(time_kernel)
{
ck::index_t M =
ck::accumulate_n<ck::index_t>(e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
ck::index_t N = ck::accumulate_n<ck::index_t>(
e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
ck::index_t K = ck::accumulate_n<ck::index_t>(
a0_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(A0DataType) * M * K + sizeof(BDataType) * K * N + +sizeof(EDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
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;
}
if(do_verification)
{
Tensor<CShuffleDataType> c_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
Tensor<A0DataType> a_ms_ks(a0_ms_ks_lengths, a0_ms_ks_strides);
for(size_t m0 = 0; m0 < a_ms_ks.mDesc.GetLengths()[0]; ++m0)
{
for(size_t m1 = 0; m1 < a_ms_ks.mDesc.GetLengths()[1]; ++m1)
{
for(size_t k0 = 0; k0 < a_ms_ks.mDesc.GetLengths()[2]; ++k0)
{
for(size_t k1 = 0; k1 < a_ms_ks.mDesc.GetLengths()[3]; ++k1)
{
a_element_op(a_ms_ks(m0, m1, k0, k1),
a0_ms_ks(m0, m1, k0, k1),
a1_ms_ks(m0, m1, k0, k1));
}
}
}
}
using ReferenceOpInstance =
ck::tensor_operation::host::ReferenceContraction_M2_N2_K2<NumDimM,
NumDimN,
NumDimK,
A0DataType,
BDataType,
CShuffleDataType,
AccDataType,
PassThrough,
BElementOp>;
auto ref_op = ReferenceOpInstance{};
auto ref_invoker = ref_op.MakeInvoker();
Tensor<float> empty_tensor(std::vector<ck::index_t>{}, std::vector<ck::index_t>{});
auto ref_argument =
ref_op.MakeArgument(a_ms_ks, b_ns_ks, c_ms_ns_host_result, PassThrough{}, b_element_op);
ref_invoker.Run(ref_argument);
for(size_t m0 = 0; m0 < e_ms_ns_host_result.mDesc.GetLengths()[0]; ++m0)
{
for(size_t m1 = 0; m1 < e_ms_ns_host_result.mDesc.GetLengths()[1]; ++m1)
{
for(size_t n0 = 0; n0 < e_ms_ns_host_result.mDesc.GetLengths()[2]; ++n0)
{
for(size_t n1 = 0; n1 < e_ms_ns_host_result.mDesc.GetLengths()[3]; ++n1)
{
cde_element_op(e_ms_ns_host_result(m0, m1, n0, n1),
c_ms_ns_host_result(m0, m1, n0, n1),
d_ms_ns(m0, m1, n0, n1));
}
}
}
}
e_device_buf.FromDevice(e_ms_ns_device_result.mData.data());
return ck::utils::check_err(e_ms_ns_device_result, e_ms_ns_host_result) ? 0 : 1;
}
return 0;
}
example/CMakeLists.txt
View file @ f0298581
......@@ -62,6 +62,12 @@ function(add_example_executable EXAMPLE_NAME FILE_NAME)
set(result ${result} PARENT_SCOPE)
endfunction(add_example_executable EXAMPLE_NAME)
function(add_example_dependencies EXAMPLE_NAME FILE_NAME)
if(result EQUAL 0)
add_dependencies(${EXAMPLE_NAME} ${FILE_NAME})
endif()
endfunction(add_example_dependencies EXAMPLE_NAME)
function(add_example_executable_no_testing EXAMPLE_NAME FILE_NAME)
message("adding example ${EXAMPLE_NAME}")
set(result 1)
......
include/ck/tensor_operation/gpu/device/device_contraction_multiple_abd.hpp 0 → 100755
View file @ f0298581
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <array>
#include "ck/tensor_operation/gpu/device/device_base.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
// GEMM:
// input : A0[M0, M1, ... K0, K1, ...], ...
// input : B0[N0, N1, ... K0, K1, ...], ...
// input : D0[M0, M1, ... N0, N1, ...], D1[M0, M1, ... N0, N1, ...], ...
// output : E[M0, M1, ... N0, N1, ...]
// C = a_op(A) * b_op(B)
// E = cde_op(C, D0, D1, ...)
// Assume:
// D0, D1, ... and E have the same layout
template <index_t NumDimM,
index_t NumDimN,
index_t NumDimK,
typename AsDataType,
typename BsDataType,
typename DsDataType,
typename EDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation>
struct DeviceContractionMultipleABD : public BaseOperator
{
static constexpr index_t NumATensor = AsDataType::Size();
static constexpr index_t NumBTensor = BsDataType::Size();
static constexpr index_t NumDTensor = DsDataType::Size();
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(std::array<const void*, NumATensor> p_as,
std::array<const void*, NumBTensor> p_bs,
std::array<const void*, NumDTensor> p_ds,
void* p_e,
const std::array<std::vector<index_t>, NumATensor>& a_ms_ks_lengths,
const std::array<std::vector<index_t>, NumATensor>& a_ms_ks_strides,
const std::array<std::vector<index_t>, NumBTensor>& b_ns_ks_lengths,
const std::array<std::vector<index_t>, NumBTensor>& b_ns_ks_strides,
const std::array<std::vector<index_t>, NumDTensor>& d_ms_ns_lengths,
const std::array<std::vector<index_t>, NumDTensor>& d_ms_ns_strides,
const std::vector<index_t>& e_ms_ns_length,
const std::vector<index_t>& e_ms_ns_stride,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/device_gemm_splitk.hpp
View file @ f0298581
......@@ -20,7 +20,8 @@ template <typename ALayout,
typename CDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
typename CElementwiseOperation,
typename ComputeType = CDataType>
struct DeviceGemmSplitK : public BaseOperator
{
virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
......@@ -48,7 +49,8 @@ template <typename ALayout,
typename CDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
typename CElementwiseOperation,
typename ComputeType = CDataType>
using DeviceGemmSplitKPtr = std::unique_ptr<DeviceGemmSplitK<ALayout,
BLayout,
CLayout,
......@@ -57,7 +59,8 @@ using DeviceGemmSplitKPtr = std::unique_ptr<DeviceGemmSplitK<ALayout,
CDataType,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>>;
CElementwiseOperation,
ComputeType>>;
} // namespace device
} // namespace tensor_operation
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp
View file @ f0298581
......@@ -69,7 +69,8 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
CDataType,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>
CElementwiseOperation,
ComputeType>
{
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
......@@ -126,7 +127,50 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
PipelineVer,
ComputeType>;
using Argument = typename GridwiseGemm::Argument;
struct Argument : public GridwiseGemm::Argument
{
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 MPadded_,
index_t NPadded_,
index_t KPadded_,
index_t K0_,
index_t k_batch_,
AElementwiseOperation a_element_op_,
BElementwiseOperation b_element_op_,
CElementwiseOperation c_element_op_)
: GridwiseGemm::Argument(p_a_grid_,
p_b_grid_,
p_c_grid_,
M_,
N_,
K_,
StrideA_,
StrideB_,
StrideC_,
MPadded_,
NPadded_,
KPadded_,
K0_,
k_batch_),
a_element_op(a_element_op_),
b_element_op(b_element_op_),
c_element_op(c_element_op_)
{
}
AElementwiseOperation a_element_op;
BElementwiseOperation b_element_op;
CElementwiseOperation c_element_op;
};
using DefaultBlock2CTileMap = typename GridwiseGemm::DefaultBlock2CTileMap;
// Invoker
......@@ -167,8 +211,17 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
karg.M * karg.N * sizeof(CDataType),
stream_config.stream_id_));
ave_time = launch_and_time_kernel(
stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, karg, b2c_map);
ave_time =
launch_and_time_kernel(stream_config,
kernel,
dim3(gdx, gdy, gdz),
dim3(BlockSize),
0,
static_cast<typename GridwiseGemm::Argument>(karg),
b2c_map,
karg.a_element_op,
karg.b_element_op,
karg.c_element_op);
};
if(has_main_k0_block_loop)
......@@ -179,7 +232,10 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
kernel_gemm_xdlops_v2r4r2_simplified<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
DefaultBlock2CTileMap>;
DefaultBlock2CTileMap,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>;
Run(kernel);
}
......@@ -189,7 +245,10 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
kernel_gemm_xdlops_v2r4r2_simplified<GridwiseGemm,
true,
InMemoryDataOperationEnum::AtomicAdd,
DefaultBlock2CTileMap>;
DefaultBlock2CTileMap,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>;
Run(kernel);
}
......@@ -202,7 +261,10 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
kernel_gemm_xdlops_v2r4r2_simplified<GridwiseGemm,
false,
InMemoryDataOperationEnum::Set,
DefaultBlock2CTileMap>;
DefaultBlock2CTileMap,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>;
Run(kernel);
}
......@@ -212,7 +274,10 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
kernel_gemm_xdlops_v2r4r2_simplified<GridwiseGemm,
false,
InMemoryDataOperationEnum::AtomicAdd,
DefaultBlock2CTileMap>;
DefaultBlock2CTileMap,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>;
Run(kernel);
}
......@@ -260,12 +325,12 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
index_t StrideA,
index_t StrideB,
index_t StrideC,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
index_t KBatch)
{
return Argument{p_a,
return Argument(p_a,
p_b,
p_c,
M,
......@@ -278,7 +343,10 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
GridwiseGemm::CalculateNPadded(N),
GridwiseGemm::CalculateKPadded(K, KBatch),
GridwiseGemm::CalculateK0(K, KBatch),
KBatch};
KBatch,
a_element_op,
b_element_op,
c_element_op);
}
static auto MakeInvoker() { return Invoker{}; }
......@@ -293,9 +361,9 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
index_t StrideA,
index_t StrideB,
index_t StrideC,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
ck::index_t KBatch = 1) override
{
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
......@@ -311,7 +379,10 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
GridwiseGemm::CalculateNPadded(N),
GridwiseGemm::CalculateKPadded(K, KBatch),
GridwiseGemm::CalculateK0(K, KBatch),
KBatch);
KBatch,
a_element_op,
b_element_op,
c_element_op);
}
// polymorphic
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_wmma_cshuffle.hpp
View file @ f0298581
......@@ -565,7 +565,7 @@ struct DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle
auto launch_kernel = [&](auto has_main_k_block_loop) {
constexpr bool has_main_loop = has_main_k_block_loop.value;
const auto kernel = kernel_grouped_conv_fwd_multiple_d_wmma_cshuffle<
const auto kernel = kernel_grouped_conv_multiple_d_wmma_cshuffle<
GridwiseGemm,
ADataType,
BDataType,
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_dl.hpp
View file @ f0298581
......@@ -12,6 +12,7 @@
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_utils.hpp"
#include "ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_dl_v1r3.hpp"
#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
......@@ -22,32 +23,6 @@ namespace ck {
namespace tensor_operation {
namespace device {
namespace {
struct ComputePtrOffsetOfStridedBatch
{
__host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
{
return g_idx * static_cast<long_index_t>(BatchStrideA_);
}
__host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
{
return g_idx * static_cast<long_index_t>(BatchStrideB_);
}
__host__ __device__ constexpr long_index_t GetCPtrOffset(index_t g_idx) const
{
return g_idx * static_cast<long_index_t>(BatchStrideC_);
}
index_t BatchStrideA_;
index_t BatchStrideB_;
index_t BatchStrideC_;
};
} // namespace
template <typename GridwiseGemm,
typename FloatAB,
typename FloatC,
......@@ -952,7 +927,7 @@ struct DeviceGroupedConvBwdWeight_Dl : public DeviceGroupedConvBwdWeight<NDimSpa
Block2CTileMap block_2_ctile_map_;
// for computing batch offset
ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
ComputePtrOffsetOfStridedBatch<I0> compute_ptr_offset_of_batch_;
// element-wise op
OutElementwiseOperation a_element_op_;
......@@ -1024,7 +999,7 @@ struct DeviceGroupedConvBwdWeight_Dl : public DeviceGroupedConvBwdWeight<NDimSpa
remove_reference_t<DeviceOp::BGridDesc_B_K0_N0_N1_K1>,
remove_reference_t<DeviceOp::CGridDesc_M0_M10_M11_N0_N10_N11>,
remove_reference_t<DeviceOp::Block2CTileMap>,
ComputePtrOffsetOfStridedBatch,
ComputePtrOffsetOfStridedBatch<I0>,
has_main_loop,
has_double_loop>;
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp
View file @ f0298581
......@@ -14,6 +14,7 @@
#include "ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight.hpp"
#include "ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_utils.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
......@@ -21,32 +22,6 @@ namespace ck {
namespace tensor_operation {
namespace device {
namespace {
struct ComputePtrOffsetOfStridedBatch
{
__host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
{
return g_idx * static_cast<long_index_t>(BatchStrideA_);
}
__host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
{
return g_idx * static_cast<long_index_t>(BatchStrideB_);
}
__host__ __device__ constexpr long_index_t GetCPtrOffset(index_t g_idx) const
{
return g_idx * static_cast<long_index_t>(BatchStrideC_);
}
index_t BatchStrideA_;
index_t BatchStrideB_;
index_t BatchStrideC_;
};
} // namespace
template <typename GridwiseGemm,
typename FloatA,
typename FloatB,
......@@ -1222,7 +1197,7 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
Block2CTileMap block_2_ctile_map_;
// for computing batch offset
ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
ComputePtrOffsetOfStridedBatch<I0> compute_ptr_offset_of_batch_;
index_t M01_;
index_t N01_;
......@@ -1301,7 +1276,7 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
remove_reference_t<DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
remove_reference_t<DeviceOp::Block2CTileMap>,
ComputePtrOffsetOfStridedBatch,
ComputePtrOffsetOfStridedBatch<I0>,
has_main_loop>;
return launch_and_time_kernel(stream_config,
......@@ -1348,6 +1323,10 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
static bool IsSupportedArgument(const Argument& arg)
{
if(!ck::is_xdl_supported())
{
return false;
}
if constexpr(NDimSpatial == 1)
{
if constexpr(!is_GNWK_GKXC_GNWC)
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment