Skip to content

GitLab

"magic_pdf/vscode:/vscode.git/clone" did not exist on "ad9abc322de72f5db3534844f97781728aeca6cf"
Commit 5e8b5703 authored by chenjun's avatar chenjun
Browse files

add multiple multiple add profiler

parent c1f8d53c
Hide whitespace changes
Inline Side-by-side
Showing with 341 additions and 143 deletions
profiler/src/CMakeLists.txt
View file @ 5e8b5703
...@@ -2,86 +2,87 @@ ...@@ -2,86 +2,87 @@
set(PROFILER_SOURCES set(PROFILER_SOURCES
profiler.cpp profiler.cpp
profile_gemm.cpp profile_gemm.cpp
profile_reduce.cpp # profile_reduce.cpp
profile_groupnorm_bwd_data.cpp # profile_groupnorm_bwd_data.cpp
profile_groupnorm_fwd.cpp # profile_groupnorm_fwd.cpp
profile_layernorm_bwd_data.cpp # profile_layernorm_bwd_data.cpp
profile_layernorm_bwd_gamma_beta.cpp # profile_layernorm_bwd_gamma_beta.cpp
profile_groupnorm_bwd_gamma_beta.cpp # profile_groupnorm_bwd_gamma_beta.cpp
profile_layernorm_fwd.cpp # profile_layernorm_fwd.cpp
profile_max_pool2d_fwd.cpp # profile_max_pool2d_fwd.cpp
profile_pool3d_fwd.cpp # profile_pool3d_fwd.cpp
profile_avg_pool3d_bwd.cpp # profile_avg_pool3d_bwd.cpp
profile_max_pool3d_bwd.cpp # profile_max_pool3d_bwd.cpp
profile_avg_pool2d_bwd.cpp # profile_avg_pool2d_bwd.cpp
profile_max_pool2d_bwd.cpp # profile_max_pool2d_bwd.cpp
profile_softmax.cpp # profile_softmax.cpp
profile_batchnorm_fwd.cpp # profile_batchnorm_fwd.cpp
profile_batchnorm_bwd.cpp # profile_batchnorm_bwd.cpp
profile_batchnorm_infer.cpp # profile_batchnorm_infer.cpp
profile_conv_tensor_rearrange.cpp # profile_conv_tensor_rearrange.cpp
profile_transpose.cpp # profile_transpose.cpp
profile_permute_scale.cpp # profile_permute_scale.cpp
) )
if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES) # if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
list(APPEND PROFILER_SOURCES profile_contraction_bilinear.cpp) # list(APPEND PROFILER_SOURCES profile_contraction_bilinear.cpp)
list(APPEND PROFILER_SOURCES profile_contraction_scale.cpp) # list(APPEND PROFILER_SOURCES profile_contraction_scale.cpp)
endif() # endif()
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) # if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND PROFILER_SOURCES profile_gemm_reduce.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_reduce.cpp)
list(APPEND PROFILER_SOURCES profile_batched_gemm_gemm.cpp) # list(APPEND PROFILER_SOURCES profile_batched_gemm_gemm.cpp)
list(APPEND PROFILER_SOURCES profile_batched_gemm_add_relu_gemm_add.cpp) # list(APPEND PROFILER_SOURCES profile_batched_gemm_add_relu_gemm_add.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_add.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add_add_fastgelu.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_add_add_fastgelu.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add_fastgelu.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_add_fastgelu.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_gemm.cpp) # list(APPEND PROFILER_SOURCES profile_grouped_gemm.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_streamk.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_streamk.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_fastgelu.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_fastgelu.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add_relu.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_add_relu.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add_silu.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_add_silu.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add_relu_add_layernorm.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_add_relu_add_layernorm.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_gemm_fixed_nk.cpp) # list(APPEND PROFILER_SOURCES profile_grouped_gemm_fixed_nk.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_gemm_two_stage.cpp) # list(APPEND PROFILER_SOURCES profile_grouped_gemm_two_stage.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp) # list(APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_gemm_tile_loop.cpp) # list(APPEND PROFILER_SOURCES profile_grouped_gemm_tile_loop.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_gemm_multiply_tile_loop.cpp) # list(APPEND PROFILER_SOURCES profile_grouped_gemm_multiply_tile_loop.cpp)
endif() # endif()
list(APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp)
if(SUPPORTED_GPU_TARGETS MATCHES "gfx94") if(SUPPORTED_GPU_TARGETS MATCHES "gfx94")
list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp) list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp) list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply_add.cpp)
# list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp)
endif() endif()
list(APPEND PROFILER_SOURCES profile_batched_gemm.cpp) # list(APPEND PROFILER_SOURCES profile_batched_gemm.cpp)
list(APPEND PROFILER_SOURCES profile_batched_gemm_reduce.cpp) # list(APPEND PROFILER_SOURCES profile_batched_gemm_reduce.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add_multiply.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_add_multiply.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_bias_add_reduce.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_bias_add_reduce.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_splitk.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_splitk.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_universal.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_universal.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_universal_reduce.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_universal_reduce.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_universal_streamk.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_universal_streamk.cpp)
list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu.cpp) # list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu.cpp)
list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu_add.cpp) # list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu_add.cpp)
list(APPEND PROFILER_SOURCES profile_conv_bwd_data.cpp) # list(APPEND PROFILER_SOURCES profile_conv_bwd_data.cpp)
list(APPEND PROFILER_SOURCES profile_conv_fwd.cpp) # list(APPEND PROFILER_SOURCES profile_conv_fwd.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd_outelementop.cpp) # list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd_outelementop.cpp)
endif() endif()
if(SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12" OR SUPPORTED_GPU_TARGETS MATCHES "gfx9") # if(SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12" OR SUPPORTED_GPU_TARGETS MATCHES "gfx9")
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) # if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp) # list(APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp)
endif() # endif()
list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd.cpp) # list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_data.cpp) # list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_data.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp) # list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp)
endif() # endif()
if(DL_KERNELS) # if(DL_KERNELS)
list(APPEND PROFILER_SOURCES profile_batched_gemm_multi_d.cpp) # list(APPEND PROFILER_SOURCES profile_batched_gemm_multi_d.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp) # list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp)
endif() # endif()
set(PROFILER_EXECUTABLE ckProfiler) set(PROFILER_EXECUTABLE ckProfiler)
...@@ -95,86 +96,87 @@ endif() ...@@ -95,86 +96,87 @@ endif()
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE utility getopt::getopt) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE utility getopt::getopt)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_fwd_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_fwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_data_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_data_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_gamma_beta_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_gamma_beta_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool2d_fwd_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool2d_fwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool3d_fwd_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool3d_fwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool2d_bwd_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool2d_bwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool3d_bwd_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool3d_bwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_max_pool_bwd_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_max_pool_bwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_image_to_column_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_image_to_column_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_column_to_image_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_column_to_image_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_transpose_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_transpose_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_permute_scale_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_permute_scale_instance)
if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES) # if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance)
endif() # endif()
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) # if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_add_fastgelu_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_add_fastgelu_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_fastgelu_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_fastgelu_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_gemm_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_gemm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_add_relu_gemm_add_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_add_relu_gemm_add_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_streamk_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_streamk_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_fastgelu_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_fastgelu_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_silu_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_silu_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_add_layernorm_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_add_layernorm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fixed_nk_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fixed_nk_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_tile_loop_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_tile_loop_instance)
endif() # endif()
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance)
if(SUPPORTED_GPU_TARGETS MATCHES "gfx94") if(SUPPORTED_GPU_TARGETS MATCHES "gfx94")
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_add_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance)
endif() endif()
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_reduce_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_streamk_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_streamk_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_reduce_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bias_add_reduce_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bias_add_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_add_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_add_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_fwd_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_fwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv1d_bwd_data_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv1d_bwd_data_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv3d_bwd_data_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv3d_bwd_data_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_bwd_data_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_bwd_data_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convscale_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convscale_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convinvscale_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convinvscale_instance)
endif() endif()
if(SUPPORTED_GPU_TARGETS MATCHES "gfx9" OR SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12") # if(SUPPORTED_GPU_TARGETS MATCHES "gfx9" OR SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12")
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) # if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance)
endif() # endif()
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_data_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_data_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_data_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_data_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_fwd_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_fwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance)
endif() # endif()
if(DL_KERNELS) # if(DL_KERNELS)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_multi_d_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_multi_d_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance) # target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance)
endif() # endif()
rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler) rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler)
profiler/src/profile_gemm_multiply_multiply_add.cpp 0 → 100644
View file @ 5e8b5703
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "profiler/profile_gemm_multiply_multiply_add_impl.hpp"
#include "profiler_operation_registry.hpp"
enum struct GemmMatrixLayout
{
MK_KN_MN, // 0
MK_NK_MN, // 1
KM_KN_MN, // 2
KM_NK_MN, // 3
};
enum struct GemmDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
BF16_BF16_BF16, // 2
INT8_INT8_INT8, // 3
F8_F16_F16, // 4
F16_F8_F16, // 5
F16_F16_F16_F8, // 6
F8_F8_BF16, // 7
INT8_INT8_BF16, // 8
};
#define OP_NAME "gemm_multiply_multiply_add"
#define OP_DESC "GEMM_Multiply_Multiply_ADD"
int profile_gemm_multiply_multiply_add(int argc, char* argv[])
{
if(argc != 16 && argc != 20)
{
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: f8@f16; 5: f16@f8; 6: "
"f16->f8; 7: f8->bf16, "
"comp f8; 8: int8->bf16)\n");
printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
printf(" 1: A[m, k] * B[n, k] = C[m, n];\n");
printf(" 2: A[k, m] * B[k, n] = C[m, n];\n");
printf(" 3: A[k, m] * B[n, k] = C[m, n])\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=no, 1=yes)\n");
printf("arg8 to 15: M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE\n");
printf("optional:\n");
printf("arg16: number of kbatch (default 1)\n");
printf("arg17: number of warm-up cycles (default 1)\n");
printf("arg18: number of iterations (default 10)\n");
printf("arg19: memory for rotating buffer (default 0, size in MB)\n");
exit(1);
}
const auto data_type = static_cast<GemmDataType>(std::stoi(argv[2]));
const auto layout = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
const bool do_verification = std::stoi(argv[4]);
const int init_method = std::stoi(argv[5]);
const bool do_log = std::stoi(argv[6]);
const bool time_kernel = std::stoi(argv[7]);
const int M = std::stoi(argv[8]);
const int N = std::stoi(argv[9]);
const int K = std::stoi(argv[10]);
const int StrideA = std::stoi(argv[11]);
const int StrideB = std::stoi(argv[12]);
const int StrideD0 = std::stoi(argv[13]);
const int StrideD1 = std::stoi(argv[14]);
const int StrideE = std::stoi(argv[15]);
int n_warmup = 1;
int n_iter = 10;
uint64_t rotating = 0;
int KBatch = 1;
if(argc == 20)
{
KBatch = std::stoi(argv[16]);
n_warmup = std::stoi(argv[17]);
n_iter = std::stoi(argv[18]);
rotating = std::stoull(argv[19]) * 1024 * 1024;
}
using F32 = float;
using BF16 = ck::bhalf_t;
// using F8 = ck::f8_t;
using I8 = int8_t;
using I32 = int;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
auto profile = [&](auto a_type,
auto b_type,
auto comp_type,
auto acc_type,
auto d0_type,
auto d1_type,
auto d2_type,
auto c_type,
auto a_layout,
auto b_layout,
auto d0_layout,
auto d1_layout,
auto d2_layout,
auto e_layout) {
using ADataType = decltype(a_type);
using BDataType = decltype(b_type);
using ComputeDataType = decltype(comp_type);
using D0DataType = decltype(d0_type);
using D1DataType = decltype(d1_type);
using D2DataType = decltype(d2_type);
using AccDataType = decltype(acc_type);
using EDataType = decltype(c_type);
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using D0Layout = decltype(d0_layout);
using D1Layout = decltype(d1_layout);
using D2Layout = decltype(d2_layout);
using ELayout = decltype(e_layout);
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
const int DefaultStrideD0 = ck::is_same_v<D0Layout, Row> ? N : M;
const int DefaultStrideD1 = ck::is_same_v<D1Layout, Row> ? N : M;
const int DefaultStrideD2 = ck::is_same_v<D2Layout, Row> ? N : M;
const int DefaultStrideE = ck::is_same_v<ELayout, Row> ? N : M;
bool pass = ck::profiler::profile_gemm_multiply_multiply_add_impl<ADataType,
BDataType,
ComputeDataType,
AccDataType,
D0DataType,
D1DataType,
D2DataType,
EDataType,
ALayout,
BLayout,
D0Layout,
D1Layout,
D2Layout,
ELayout>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? DefaultStrideA : StrideA,
(StrideB < 0) ? DefaultStrideB : StrideB,
(StrideD0 < 0) ? DefaultStrideD0 : StrideD0,
(StrideD1 < 0) ? DefaultStrideD1 : StrideD1,
(StrideD1 < 0) ? DefaultStrideD2 : StrideD1,
(StrideE < 0) ? DefaultStrideE : StrideE,
KBatch,
n_warmup,
n_iter,
rotating);
return pass ? 0 : 1;
};
if(data_type == GemmDataType::INT8_INT8_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
{
return profile(I8{},
I8{},
I8{},
I32{},
F32{},
F32{},
BF16{},
BF16{},
Row{},
Col{},
Row{},
Col{},
Row{},
Row{});
}
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
return 1;
}
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_multiply_multiply_add);
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