Skip to content

GitLab

Commit bccc6d8b authored by wangshaojie6's avatar wangshaojie6
Browse files

merge develop and resolve conflict

parents c6b52884 91d8b7d6
Hide whitespace changes
Inline Side-by-side
Showing with 272 additions and 569 deletions
library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_partial_reduce_f16_f16_f16.cpp deleted 100644 → 0
View file @ c6b52884
#include "device_reduce_instance_multiblock_partial_reduce.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_reduce_instance {
// clang-format off
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 0, 4, 3); // for MIN
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 0, 4, 3); // for MAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 0, 4, 3); // for AMAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 1, 4, 3); // for MIN
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 1, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 1, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 1, 4, 3); // for MAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 1, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 1, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 1, 4, 3); // for AMAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 1, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 1, 2, 1);
// clang-format on
} // namespace device_reduce_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_partial_reduce_f16_f32_f16.cpp deleted 100644 → 0
View file @ c6b52884
#include "device_reduce_instance_multiblock_partial_reduce.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_reduce_instance {
// clang-format off
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, float, half_t, 0, 0, 0, 4, 3); // for ADD
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, float, half_t, 0, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, float, half_t, 0, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, float, half_t, 0, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, float, half_t, 5, 0, 0, 4, 3); // for AVG
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, float, half_t, 5, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, float, half_t, 5, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, float, half_t, 5, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, float, half_t, 7, 0, 0, 4, 3); // for NORM2
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, float, half_t, 7, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, float, half_t, 7, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, float, half_t, 7, 0, 0, 2, 1);
// clang-format on
} // namespace device_reduce_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_partial_reduce_f32_f32_f32.cpp deleted 100644 → 0
View file @ c6b52884
#include "device_reduce_instance_multiblock_partial_reduce.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_reduce_instance {
// clang-format off
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 2, 0, 0, 4, 3); // for MIN
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 2, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 2, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 2, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 3, 0, 0, 4, 3); // for MAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 3, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 3, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 3, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 4, 0, 0, 4, 3); // for AMAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 4, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 4, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 4, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 2, 0, 1, 4, 3); // for MIN
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 2, 0, 1, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 2, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 2, 0, 1, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 3, 0, 1, 4, 3); // for MAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 3, 0, 1, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 3, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 3, 0, 1, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 4, 0, 1, 4, 3); // for AMAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 4, 0, 1, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 4, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 4, 0, 1, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 7, 0, 0, 4, 3); // for NORM2
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 7, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 7, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 7, 0, 0, 2, 1);
// clang-format on
} // namespace device_reduce_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_partial_reduce_f64_f64_f64.cpp deleted 100644 → 0
View file @ c6b52884
#include "device_reduce_instance_multiblock_partial_reduce.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_reduce_instance {
// clang-format off
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 0, 4, 3); // for MIN
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 0, 4, 3); // for MAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 0, 4, 3); // for AMAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 1, 4, 3); // for MIN
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 1, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 1, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 1, 4, 3); // for MAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 1, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 1, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 1, 4, 3); // for AMAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 1, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 1, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 7, 0, 0, 4, 3); // for NORM2
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 7, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 7, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 7, 0, 0, 2, 1);
// Will be moved to use MultiBlockAtomicAdd
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 0, 0, 0, 4, 3); // for ADD
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 0, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 0, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 0, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 5, 0, 0, 4, 3); // for AVG
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 5, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 5, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 5, 0, 0, 2, 1);
// clang-format on
} // namespace device_reduce_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_partial_reduce_i8_i32_i8.cpp deleted 100644 → 0
View file @ c6b52884
#include "device_reduce_instance_multiblock_partial_reduce.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_reduce_instance {
// clang-format off
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int32_t, int8_t, 0, 0, 0, 4, 3); // for ADD
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int32_t, int8_t, 0, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int32_t, int8_t, 0, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int32_t, int8_t, 0, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int32_t, int8_t, 5, 0, 0, 4, 3); // for AVG
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int32_t, int8_t, 5, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int32_t, int8_t, 5, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int32_t, int8_t, 5, 0, 0, 2, 1);
// clang-format on
} // namespace device_reduce_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_partial_reduce_i8_i8_i8.cpp deleted 100644 → 0
View file @ c6b52884
#include "device_reduce_instance_multiblock_partial_reduce.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_reduce_instance {
// clang-format off
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 4, 3); // for MIN
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 4, 3); // for MAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 4, 3); // for AMAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 4, 3); // for MIN
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 4, 3); // for MAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 4, 3); // for AMAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 2, 1);
// clang-format on
} // namespace device_reduce_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
profiler/CMakeLists.txt
View file @ bccc6d8b
include_directories(BEFORE include_directories(BEFORE
${PROJECT_SOURCE_DIR}/include/ck ${PROJECT_SOURCE_DIR}/include/ck
${PROJECT_SOURCE_DIR}/include/ck/utility ${PROJECT_SOURCE_DIR}/include/ck/utility
${PROJECT_SOURCE_DIR}/include/ck/host_utility
${PROJECT_SOURCE_DIR}/include/ck/tensor_description ${PROJECT_SOURCE_DIR}/include/ck/tensor_description
${PROJECT_SOURCE_DIR}/include/ck/tensor ${PROJECT_SOURCE_DIR}/include/ck/tensor
${PROJECT_SOURCE_DIR}/include/ck/problem_transform ${PROJECT_SOURCE_DIR}/include/ck/problem_transform
......
profiler/include/profile_gemm_impl.hpp
View file @ bccc6d8b
#pragma once #pragma once
#include <iomanip> #include <iomanip>
#include <iostream>
#include <typeinfo>
#include "check_err.hpp" #include "check_err.hpp"
#include "config.hpp" #include "config.hpp"
...@@ -42,14 +44,10 @@ void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(std::vector<De ...@@ -42,14 +44,10 @@ void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(std::vector<De
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_int8_int8_int8_mk_kn_mn_instances( void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_int8_int8_int8_mk_nk_mn_instances( void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_int8_int8_int8_km_kn_mn_instances(
std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_int8_int8_int8_km_nk_mn_instances(
std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances( void add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(
std::vector<DeviceGemmNoOpPtr>&); std::vector<DeviceGemmNoOpPtr>&);
...@@ -74,6 +72,21 @@ void add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(std::vector<Devic ...@@ -74,6 +72,21 @@ void add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(std::vector<Devic
void add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&); void add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_dl_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_dl_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_dl_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_dl_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_dl_i8_i8_i8_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_dl_i8_i8_i8_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_dl_i8_i8_i8_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_dl_i8_i8_i8_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
} // namespace device_gemm_instance } // namespace device_gemm_instance
} // namespace device } // namespace device
} // namespace tensor_operation } // namespace tensor_operation
...@@ -85,6 +98,7 @@ namespace profiler { ...@@ -85,6 +98,7 @@ namespace profiler {
template <typename ADataType, template <typename ADataType,
typename BDataType, typename BDataType,
typename CDataType, typename CDataType,
typename AccDataType,
typename ALayout, typename ALayout,
typename BLayout, typename BLayout,
typename CLayout> typename CLayout>
...@@ -125,7 +139,11 @@ void profile_gemm_impl(int do_verification, ...@@ -125,7 +139,11 @@ void profile_gemm_impl(int do_verification,
std::size_t num_thread = 1; std::size_t num_thread = 1;
switch(init_method) switch(init_method)
{ {
case 0: break; // case 0: break;
case 0:
a_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{}, num_thread);
b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{}, num_thread);
break;
case 1: case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread); a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread); b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
...@@ -174,6 +192,9 @@ void profile_gemm_impl(int do_verification, ...@@ -174,6 +192,9 @@ void profile_gemm_impl(int do_verification,
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(gemm_ptrs); add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_f32_f32_f32_mk_kn_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(gemm_ptrs); add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(gemm_ptrs);
} }
...@@ -192,6 +213,9 @@ void profile_gemm_impl(int do_verification, ...@@ -192,6 +213,9 @@ void profile_gemm_impl(int do_verification,
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(gemm_ptrs); add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(gemm_ptrs); add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(gemm_ptrs);
} }
...@@ -210,6 +234,9 @@ void profile_gemm_impl(int do_verification, ...@@ -210,6 +234,9 @@ void profile_gemm_impl(int do_verification,
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(gemm_ptrs); add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(gemm_ptrs); add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(gemm_ptrs);
} }
...@@ -228,6 +255,9 @@ void profile_gemm_impl(int do_verification, ...@@ -228,6 +255,9 @@ void profile_gemm_impl(int do_verification,
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(gemm_ptrs); add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(gemm_ptrs); add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(gemm_ptrs);
} }
...@@ -250,6 +280,9 @@ void profile_gemm_impl(int do_verification, ...@@ -250,6 +280,9 @@ void profile_gemm_impl(int do_verification,
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(gemm_ptrs); add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_f16_f16_f16_mk_kn_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(gemm_ptrs); add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(gemm_ptrs);
} }
...@@ -268,6 +301,9 @@ void profile_gemm_impl(int do_verification, ...@@ -268,6 +301,9 @@ void profile_gemm_impl(int do_verification,
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs); add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs); add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
...@@ -289,6 +325,9 @@ void profile_gemm_impl(int do_verification, ...@@ -289,6 +325,9 @@ void profile_gemm_impl(int do_verification,
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(gemm_ptrs); add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(gemm_ptrs); add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(gemm_ptrs);
} }
...@@ -307,6 +346,9 @@ void profile_gemm_impl(int do_verification, ...@@ -307,6 +346,9 @@ void profile_gemm_impl(int do_verification,
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(gemm_ptrs); add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(gemm_ptrs); add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
} }
...@@ -353,28 +395,40 @@ void profile_gemm_impl(int do_verification, ...@@ -353,28 +395,40 @@ void profile_gemm_impl(int do_verification,
is_same<CLayout, tensor_layout::gemm::RowMajor>::value) is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{ {
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_int8_int8_int8_mk_kn_mn_instances(gemm_ptrs); add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_i8_i8_i8_mk_kn_mn_instances(gemm_ptrs);
} }
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value && else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value && is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value) is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{ {
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_int8_int8_int8_mk_nk_mn_instances(gemm_ptrs); add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_i8_i8_i8_mk_nk_mn_instances(gemm_ptrs);
} }
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value && else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value && is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value) is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{ {
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_int8_int8_int8_km_kn_mn_instances(gemm_ptrs); add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_i8_i8_i8_km_kn_mn_instances(gemm_ptrs);
} }
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value && else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value && is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value) is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{ {
ck::tensor_operation::device::device_gemm_instance:: ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_int8_int8_int8_km_nk_mn_instances(gemm_ptrs); add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(gemm_ptrs);
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_dl_i8_i8_i8_km_nk_mn_instances(gemm_ptrs);
} }
} }
...@@ -458,8 +512,14 @@ void profile_gemm_impl(int do_verification, ...@@ -458,8 +512,14 @@ void profile_gemm_impl(int do_verification,
bf16_to_f32_(b_k_n, b_f32_k_n); bf16_to_f32_(b_k_n, b_f32_k_n);
bf16_to_f32_(c_m_n_device_result, c_m_n_device_f32_result); bf16_to_f32_(c_m_n_device_result, c_m_n_device_f32_result);
using ReferenceGemmInstance = ck::tensor_operation::host:: using ReferenceGemmInstance =
ReferenceGemm<float, float, float, AElementOp, BElementOp, CElementOp>; ck::tensor_operation::host::ReferenceGemm<float,
float,
float,
float,
AElementOp,
BElementOp,
CElementOp>;
auto ref_gemm = ReferenceGemmInstance{}; auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker(); auto ref_invoker = ref_gemm.MakeInvoker();
...@@ -491,6 +551,7 @@ void profile_gemm_impl(int do_verification, ...@@ -491,6 +551,7 @@ void profile_gemm_impl(int do_verification,
ck::tensor_operation::host::ReferenceGemm<ADataType, ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType, BDataType,
CDataType, CDataType,
AccDataType,
AElementOp, AElementOp,
BElementOp, BElementOp,
CElementOp>; CElementOp>;
...@@ -523,12 +584,50 @@ void profile_gemm_impl(int do_verification, ...@@ -523,12 +584,50 @@ void profile_gemm_impl(int do_verification,
} }
else else
{ {
std::cout << "does not support this GEMM problem" << std::endl; std::cout << gemm_ptr->GetTypeString() << " does not support this GEMM problem"
<< std::endl;
} }
} }
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, " if constexpr(is_same<CDataType, float>::value)
<< best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl; {
std::cout << "Best Perf for datatype = f32";
}
else if constexpr(is_same<CDataType, half_t>::value)
{
std::cout << "Best Perf for datatype = f16";
}
else if constexpr(is_same<CDataType, bhalf_t>::value)
{
std::cout << "Best Perf for datatype = bf16";
}
else if constexpr(is_same<CDataType, int8_t>::value)
{
std::cout << "Best Perf for datatype = int8";
}
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value)
{
std::cout << " ALayout = RowMajor";
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value)
{
std::cout << " ALayout = ColumnMajor";
}
if constexpr(is_same<BLayout, tensor_layout::gemm::RowMajor>::value)
{
std::cout << " BLayout = RowMajor";
}
else if constexpr(is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value)
{
std::cout << " BLayout = ColumnMajor";
}
std::cout << " M = " << M << " N = " << N << " K = " << K << " StrideA = " << StrideA
<< " StrideB = " << StrideB << " StrideC = " << StrideC << " : " << best_ave_time
<< " ms, " << best_tflops << " TFlops, " << best_gb_per_sec << " GB/s, "
<< best_gemm_name << std::endl;
} }
} // namespace profiler } // namespace profiler
......
profiler/include/profile_gemm_reduce_impl.hpp
View file @ bccc6d8b
...@@ -144,8 +144,13 @@ bool profile_gemm_reduce_impl(int do_verification, ...@@ -144,8 +144,13 @@ bool profile_gemm_reduce_impl(int do_verification,
if(do_verification) if(do_verification)
{ {
using ReferenceGemmInstance = ck::tensor_operation::host:: using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
ReferenceGemm<ADataType, BDataType, CDataType, AElementOp, BElementOp, CElementOp>; BDataType,
CDataType,
DDataType,
AElementOp,
BElementOp,
CElementOp>;
auto ref_gemm = ReferenceGemmInstance{}; auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker(); auto ref_invoker = ref_gemm.MakeInvoker();
......
profiler/include/profile_grouped_gemm_impl.hpp
View file @ bccc6d8b
...@@ -43,6 +43,7 @@ namespace profiler { ...@@ -43,6 +43,7 @@ namespace profiler {
template <typename ADataType, template <typename ADataType,
typename BDataType, typename BDataType,
typename CDataType, typename CDataType,
typename AccDataType,
typename ALayout, typename ALayout,
typename BLayout, typename BLayout,
typename CLayout> typename CLayout>
...@@ -271,6 +272,7 @@ void profile_grouped_gemm_impl(int do_verification, ...@@ -271,6 +272,7 @@ void profile_grouped_gemm_impl(int do_verification,
ck::tensor_operation::host::ReferenceGemm<ADataType, ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType, BDataType,
CDataType, CDataType,
AccDataType,
AElementOp, AElementOp,
BElementOp, BElementOp,
CElementOp>; CElementOp>;
......
profiler/include/profile_reduce_impl.hpp
View file @ bccc6d8b
...@@ -5,74 +5,77 @@ ...@@ -5,74 +5,77 @@
#include "device_reduce_instance.hpp" #include "device_reduce_instance.hpp"
#include "reduction_enums.hpp" #include "reduction_enums.hpp"
#include "host_reduction.hpp" #include "host_reduction.hpp"
#include "host_common_util.hpp"
#include "host_tensor_generator.hpp"
namespace ck { namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
namespace device_reduce_instance { namespace device_reduce_instance {
template <int Rank, int NumReduceDim, int ReduceOpId, int NanOpt, int IndicesOpt> template <int Rank, int NumReduceDim, int ReduceOpId, bool PropagateNan, bool UseIndex>
struct ReduceDescription struct ReduceDescription
{ {
static constexpr int Rank_ = Rank; static constexpr int Rank_ = Rank;
static constexpr int NumReduceDim_ = NumReduceDim; static constexpr int NumReduceDim_ = NumReduceDim;
static constexpr int ReduceOpId_ = ReduceOpId; static constexpr int ReduceOpId_ = ReduceOpId;
static constexpr int NanOpt_ = NanOpt; static constexpr int PropagateNan_ = PropagateNan;
static constexpr int IndicesOpt_ = IndicesOpt; static constexpr int UseIndex_ = UseIndex;
}; };
using reduce_description_instances = std::tuple<ReduceDescription<4, 3, 0, 0, 0>, // for ADD using reduce_description_instances =
ReduceDescription<4, 4, 0, 0, 0>, std::tuple<ReduceDescription<4, 3, 0, false, false>, // for ADD
ReduceDescription<4, 1, 0, 0, 0>, ReduceDescription<4, 4, 0, false, false>,
ReduceDescription<2, 1, 0, 0, 0>, ReduceDescription<4, 1, 0, false, false>,
ReduceDescription<2, 1, 0, false, false>,
ReduceDescription<4, 3, 5, 0, 0>, // for AVG
ReduceDescription<4, 4, 5, 0, 0>, ReduceDescription<4, 3, 5, false, false>, // for AVG
ReduceDescription<4, 1, 5, 0, 0>, ReduceDescription<4, 4, 5, false, false>,
ReduceDescription<2, 1, 5, 0, 0>, ReduceDescription<4, 1, 5, false, false>,
ReduceDescription<2, 1, 5, false, false>,
ReduceDescription<4, 3, 7, 0, 0>, // for NORM2
ReduceDescription<4, 4, 7, 0, 0>, ReduceDescription<4, 3, 7, false, false>, // for NORM2
ReduceDescription<4, 1, 7, 0, 0>, ReduceDescription<4, 4, 7, false, false>,
ReduceDescription<2, 1, 7, 0, 0>, ReduceDescription<4, 1, 7, false, false>,
ReduceDescription<2, 1, 7, false, false>,
ReduceDescription<4, 3, 2, 0, 0>, // for MIN
ReduceDescription<4, 4, 2, 0, 0>, ReduceDescription<4, 3, 2, false, false>, // for MIN
ReduceDescription<4, 1, 2, 0, 0>, ReduceDescription<4, 4, 2, false, false>,
ReduceDescription<2, 1, 2, 0, 0>, ReduceDescription<4, 1, 2, false, false>,
ReduceDescription<4, 3, 3, 0, 0>, // for MAX ReduceDescription<2, 1, 2, false, false>,
ReduceDescription<4, 4, 3, 0, 0>, ReduceDescription<4, 3, 3, false, false>, // for MAX
ReduceDescription<4, 1, 3, 0, 0>, ReduceDescription<4, 4, 3, false, false>,
ReduceDescription<2, 1, 3, 0, 0>, ReduceDescription<4, 1, 3, false, false>,
ReduceDescription<4, 3, 4, 0, 0>, // for AMAX ReduceDescription<2, 1, 3, false, false>,
ReduceDescription<4, 4, 4, 0, 0>, ReduceDescription<4, 3, 4, false, false>, // for AMAX
ReduceDescription<4, 1, 4, 0, 0>, ReduceDescription<4, 4, 4, false, false>,
ReduceDescription<2, 1, 4, 0, 0>, ReduceDescription<4, 1, 4, false, false>,
ReduceDescription<2, 1, 4, false, false>,
ReduceDescription<4, 3, 2, 0, 1>, // for MIN
ReduceDescription<4, 4, 2, 0, 1>, ReduceDescription<4, 3, 2, false, true>, // for MIN
ReduceDescription<4, 1, 2, 0, 1>, ReduceDescription<4, 4, 2, false, true>,
ReduceDescription<2, 1, 2, 0, 1>, ReduceDescription<4, 1, 2, false, true>,
ReduceDescription<4, 3, 3, 0, 1>, // for MAX ReduceDescription<2, 1, 2, false, true>,
ReduceDescription<4, 4, 3, 0, 1>, ReduceDescription<4, 3, 3, false, true>, // for MAX
ReduceDescription<4, 1, 3, 0, 1>, ReduceDescription<4, 4, 3, false, true>,
ReduceDescription<2, 1, 3, 0, 1>, ReduceDescription<4, 1, 3, false, true>,
ReduceDescription<4, 3, 4, 0, 1>, // for AMAX ReduceDescription<2, 1, 3, false, true>,
ReduceDescription<4, 4, 4, 0, 1>, ReduceDescription<4, 3, 4, false, true>, // for AMAX
ReduceDescription<4, 1, 4, 0, 1>, ReduceDescription<4, 4, 4, false, true>,
ReduceDescription<2, 1, 4, 0, 1>>; ReduceDescription<4, 1, 4, false, true>,
ReduceDescription<2, 1, 4, false, true>>;
template <typename DescriptionType> template <typename DescriptionType>
bool description_match(const DescriptionType& description, bool description_match(const DescriptionType& description,
int Rank, int Rank,
const std::vector<int>& reduceDims, const std::vector<int>& reduceDims,
ReduceTensorOp ReduceOpId, ReduceTensorOp ReduceOpId,
NanPropagation NanOpt, bool PropagateNan,
ReduceTensorIndices IndicesOpt) bool UseIndex)
{ {
if(description.Rank_ != Rank || description.ReduceOpId_ != static_cast<int>(ReduceOpId) || if(description.Rank_ != Rank || description.ReduceOpId_ != static_cast<int>(ReduceOpId) ||
description.NanOpt_ != static_cast<int>(NanOpt) || description.PropagateNan_ != static_cast<int>(PropagateNan) ||
description.IndicesOpt_ != static_cast<int>(IndicesOpt)) description.UseIndex_ != static_cast<int>(UseIndex))
return (false); return (false);
if(DescriptionType::NumReduceDim_ != reduceDims.size()) if(DescriptionType::NumReduceDim_ != reduceDims.size())
...@@ -116,46 +119,16 @@ static inline std::vector<int> get_invariant_dims(const std::vector<int>& reduce ...@@ -116,46 +119,16 @@ static inline std::vector<int> get_invariant_dims(const std::vector<int>& reduce
return invariantDims; return invariantDims;
}; };
template <typename T>
static void dumpBufferToFile(const char* fileName, T* data, size_t dataNumItems)
{
std::ofstream outFile(fileName, std::ios::binary);
if(outFile)
{
outFile.write(reinterpret_cast<char*>(data), dataNumItems * sizeof(T));
outFile.close();
std::cout << "Write output to file " << fileName << std::endl;
}
else
{
std::cout << "Could not open file " << fileName << " for writing" << std::endl;
}
};
// map the data type used by the GPU kernels to the corresponding type used by the host codes
template <typename InType>
struct type_mapping
{
using OutType = InType;
};
template <>
struct type_mapping<ck::half_t>
{
using OutType = half_float::half;
};
template <typename InDataType, template <typename InDataType,
typename AccDataType, typename AccDataType,
typename OutDataType, typename OutDataType,
int Rank, int Rank,
int NumReduceDim, int NumReduceDim,
ReduceTensorOp ReduceOpId, ReduceTensorOp ReduceOpId,
NanPropagation NanOpt, bool PropagateNan,
ReduceTensorIndices IndicesOpt> bool UseIndex>
void profile_reduce_impl_impl(bool do_verification, bool profile_reduce_impl_impl(bool do_verification,
int init_method, int init_method,
bool do_log,
bool do_dumpout, bool do_dumpout,
bool time_kernel, bool time_kernel,
const std::vector<size_t>& inLengths, const std::vector<size_t>& inLengths,
...@@ -166,15 +139,13 @@ void profile_reduce_impl_impl(bool do_verification, ...@@ -166,15 +139,13 @@ void profile_reduce_impl_impl(bool do_verification,
using namespace ck::tensor_operation::device; using namespace ck::tensor_operation::device;
using namespace ck::tensor_operation::device::device_reduce_instance; using namespace ck::tensor_operation::device::device_reduce_instance;
using namespace ck::host_reduce; using namespace ck::host_reduce;
using ck::host_common::dumpBufferToFile;
constexpr bool op_support_indices = constexpr bool op_support_indices =
(ReduceOpId == ReduceTensorOp::MIN || ReduceOpId == ReduceTensorOp::MAX || (ReduceOpId == ReduceTensorOp::MIN || ReduceOpId == ReduceTensorOp::MAX ||
ReduceOpId == ReduceTensorOp::AMAX); ReduceOpId == ReduceTensorOp::AMAX);
constexpr bool NeedIndices = constexpr bool OutputIndex = (op_support_indices && UseIndex);
(op_support_indices && (IndicesOpt != ReduceTensorIndices::NO_INDICES));
constexpr bool PropagateNan = (NanOpt == NanPropagation::PROPAGATE_NAN);
constexpr bool out_support_atomic_add = std::is_same<OutDataType, float>::value; constexpr bool out_support_atomic_add = std::is_same<OutDataType, float>::value;
constexpr bool op_support_atomic_add = constexpr bool op_support_atomic_add =
...@@ -195,8 +166,7 @@ void profile_reduce_impl_impl(bool do_verification, ...@@ -195,8 +166,7 @@ void profile_reduce_impl_impl(bool do_verification,
(op_support_indices && !std::is_same<AccDataType, float>::value); (op_support_indices && !std::is_same<AccDataType, float>::value);
// 1) The indices can only be used when the reduction operation is indexable // 1) The indices can only be used when the reduction operation is indexable
constexpr bool invalid_reduce_3 = constexpr bool invalid_reduce_3 = (!op_support_indices && UseIndex);
(!op_support_indices && IndicesOpt != ReduceTensorIndices::NO_INDICES);
// 1) If InDataType is int8_t, must use int8_t as AccDataType for indexable reduction operations // 1) If InDataType is int8_t, must use int8_t as AccDataType for indexable reduction operations
// 2) If InDataType is int8_t, must use int32_t as AccDataType for non-indexable reduction // 2) If InDataType is int8_t, must use int32_t as AccDataType for non-indexable reduction
...@@ -219,6 +189,8 @@ void profile_reduce_impl_impl(bool do_verification, ...@@ -219,6 +189,8 @@ void profile_reduce_impl_impl(bool do_verification,
constexpr bool invalid_reduce = (invalid_reduce_1 || invalid_reduce_2 || invalid_reduce_3 || constexpr bool invalid_reduce = (invalid_reduce_1 || invalid_reduce_2 || invalid_reduce_3 ||
invalid_reduce_4 || invalid_reduce_5 || invalid_reduce_6); invalid_reduce_4 || invalid_reduce_5 || invalid_reduce_6);
bool pass = true;
if constexpr(!invalid_reduce) if constexpr(!invalid_reduce)
{ {
Tensor<InDataType> in(inLengths); Tensor<InDataType> in(inLengths);
...@@ -282,7 +254,7 @@ void profile_reduce_impl_impl(bool do_verification, ...@@ -282,7 +254,7 @@ void profile_reduce_impl_impl(bool do_verification,
if(beta != 0.0f) if(beta != 0.0f)
out_dev.ToDevice(out.mData.data()); out_dev.ToDevice(out.mData.data());
size_t indicesSizeInBytes = NeedIndices ? out.mDesc.GetElementSize() * sizeof(int) : 0; size_t indicesSizeInBytes = OutputIndex ? out.mDesc.GetElementSize() * sizeof(int) : 0;
DeviceMem out_indices_dev(indicesSizeInBytes); DeviceMem out_indices_dev(indicesSizeInBytes);
...@@ -295,29 +267,11 @@ void profile_reduce_impl_impl(bool do_verification, ...@@ -295,29 +267,11 @@ void profile_reduce_impl_impl(bool do_verification,
using AccElementwiseOperation_0 = using AccElementwiseOperation_0 =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>:: typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::
AccElementwiseOperation; AccElementwiseOperation;
using InElementwiseOperation_1 =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, false>::
InElementwiseOperation;
using AccElementwiseOperation_1 =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, false>::
AccElementwiseOperation;
using InElementwiseOperation_2 =
typename reduce_unary_operator<AccDataType, ReduceOpId, false, true>::
InElementwiseOperation;
using AccElementwiseOperation_2 =
typename reduce_unary_operator<AccDataType, ReduceOpId, false, true>::
AccElementwiseOperation;
using DeviceReduceInstPtr0 = using DeviceReduceInstPtr0 =
DeviceReducePtr<InElementwiseOperation_0, AccElementwiseOperation_0>; DeviceReducePtr<InElementwiseOperation_0, AccElementwiseOperation_0>;
using DeviceReduceInstPtr1 =
DeviceReducePtr<InElementwiseOperation_1, AccElementwiseOperation_1>;
using DeviceReduceInstPtr2 =
DeviceReducePtr<InElementwiseOperation_2, AccElementwiseOperation_2>;
std::vector<DeviceReduceInstPtr0> reduce0_ptrs; std::vector<DeviceReduceInstPtr0> reduce0_ptrs;
std::vector<DeviceReduceInstPtr1> reduce1_ptrs;
std::vector<DeviceReduceInstPtr2> reduce2_ptrs;
add_device_reduce_instance_threadwise<InDataType, add_device_reduce_instance_threadwise<InDataType,
AccDataType, AccDataType,
...@@ -325,8 +279,8 @@ void profile_reduce_impl_impl(bool do_verification, ...@@ -325,8 +279,8 @@ void profile_reduce_impl_impl(bool do_verification,
Rank, Rank,
NumReduceDim, NumReduceDim,
ReduceOpId, ReduceOpId,
NanOpt, PropagateNan,
IndicesOpt>(reduce0_ptrs); UseIndex>(reduce0_ptrs);
add_device_reduce_instance_blockwise<InDataType, add_device_reduce_instance_blockwise<InDataType,
AccDataType, AccDataType,
...@@ -334,8 +288,8 @@ void profile_reduce_impl_impl(bool do_verification, ...@@ -334,8 +288,8 @@ void profile_reduce_impl_impl(bool do_verification,
Rank, Rank,
NumReduceDim, NumReduceDim,
ReduceOpId, ReduceOpId,
NanOpt, PropagateNan,
IndicesOpt>(reduce0_ptrs); UseIndex>(reduce0_ptrs);
if constexpr(use_atomic_add) if constexpr(use_atomic_add)
{ {
...@@ -345,35 +299,11 @@ void profile_reduce_impl_impl(bool do_verification, ...@@ -345,35 +299,11 @@ void profile_reduce_impl_impl(bool do_verification,
Rank, Rank,
NumReduceDim, NumReduceDim,
ReduceOpId, ReduceOpId,
NanOpt, PropagateNan,
IndicesOpt>(reduce0_ptrs); UseIndex>(reduce0_ptrs);
} }
else
{
add_device_reduce_instance_multiblock_partial_reduce<InDataType,
AccDataType,
OutDataType,
Rank,
NumReduceDim,
ReduceOpId,
NanOpt,
IndicesOpt>(reduce1_ptrs);
};
// used for secondary reduction if(reduce0_ptrs.empty())
if constexpr(!use_atomic_add)
{
add_device_reduce_instance_blockwise_second_call<AccDataType,
AccDataType,
OutDataType,
Rank,
NumReduceDim,
ReduceOpId,
NanOpt,
IndicesOpt>(reduce2_ptrs);
};
if(reduce0_ptrs.empty() && reduce1_ptrs.empty())
{ {
throw std::runtime_error("Wrong! No device REDUCE instance found"); throw std::runtime_error("Wrong! No device REDUCE instance found");
}; };
...@@ -387,23 +317,25 @@ void profile_reduce_impl_impl(bool do_verification, ...@@ -387,23 +317,25 @@ void profile_reduce_impl_impl(bool do_verification,
Rank, Rank,
NumReduceDim, NumReduceDim,
PropagateNan, PropagateNan,
NeedIndices> OutputIndex>
hostReduce(in.mDesc, out_ref.mDesc, invariantDims, reduceDims); hostReduce(in.mDesc, out_ref.mDesc, invariantDims, reduceDims);
hostReduce.Run( hostReduce.Run(
alpha, in.mData.data(), beta, out_ref.mData.data(), out_indices_ref.mData.data()); alpha, in.mData.data(), beta, out_ref.mData.data(), out_indices_ref.mData.data());
}; };
const auto i_inLengths = to_int_vector(inLengths); std::vector<ck::index_t> i_inLengths;
const auto i_inStrides = to_int_vector(inStrides); std::vector<ck::index_t> i_inStrides;
const auto i_outLengths = to_int_vector(outLengths); std::vector<ck::index_t> i_outLengths;
const auto i_outStrides = to_int_vector(outStrides); std::vector<ck::index_t> i_outStrides;
i_inLengths.assign(inLengths.begin(), inLengths.end());
i_inStrides.assign(inStrides.begin(), inStrides.end());
i_outLengths.assign(outLengths.begin(), outLengths.end());
i_outStrides.assign(outStrides.begin(), outStrides.end());
for(auto& reduce_ptr : reduce0_ptrs) for(auto& reduce_ptr : reduce0_ptrs)
{ {
auto wsSizeInBytes = reduce_ptr->GetWorkspaceSizeInBytes(i_inLengths, reduceDims);
DeviceMem ws_dev(wsSizeInBytes);
InElementwiseOperation_0 in_elementwise_op_0(static_cast<int32_t>(reduce_total_length)); InElementwiseOperation_0 in_elementwise_op_0(static_cast<int32_t>(reduce_total_length));
AccElementwiseOperation_0 acc_elementwise_op_0( AccElementwiseOperation_0 acc_elementwise_op_0(
...@@ -417,9 +349,9 @@ void profile_reduce_impl_impl(bool do_verification, ...@@ -417,9 +349,9 @@ void profile_reduce_impl_impl(bool do_verification,
alpha, alpha,
beta, beta,
in_dev.GetDeviceBuffer(), in_dev.GetDeviceBuffer(),
nullptr,
out_dev.GetDeviceBuffer(), out_dev.GetDeviceBuffer(),
out_indices_dev.GetDeviceBuffer(), out_indices_dev.GetDeviceBuffer(),
ws_dev.GetDeviceBuffer(),
in_elementwise_op_0, in_elementwise_op_0,
acc_elementwise_op_0); acc_elementwise_op_0);
...@@ -439,8 +371,9 @@ void profile_reduce_impl_impl(bool do_verification, ...@@ -439,8 +371,9 @@ void profile_reduce_impl_impl(bool do_verification,
float gb_per_sec = num_bytes / 1.E6 / avg_time; float gb_per_sec = num_bytes / 1.E6 / avg_time;
std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, " << reduce_name if(time_kernel)
<< std::endl; std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, "
<< reduce_name << std::endl;
if(gb_per_sec > best_gb_per_sec) if(gb_per_sec > best_gb_per_sec)
{ {
...@@ -450,22 +383,24 @@ void profile_reduce_impl_impl(bool do_verification, ...@@ -450,22 +383,24 @@ void profile_reduce_impl_impl(bool do_verification,
if(do_verification) if(do_verification)
{ {
bool single_pass;
out_dev.FromDevice(out.mData.data()); out_dev.FromDevice(out.mData.data());
ck::utils::check_err(out.mData, out_ref.mData); single_pass = ck::utils::check_err(out.mData, out_ref.mData);
if(NeedIndices) if(OutputIndex)
{ {
out_indices_dev.FromDevice(out_indices.mData.data()); out_indices_dev.FromDevice(out_indices.mData.data());
ck::utils::check_err(out_indices.mData, out_indices_ref.mData); single_pass = single_pass &&
; ck::utils::check_err(out_indices.mData, out_indices_ref.mData);
}; };
if(do_log) if(!single_pass)
{ {
LogRangeAsType<float>(std::cout << "out_host : ", out_ref.mData, ",") std::cout << "Fail Info: " << reduce_ptr->GetTypeString() << std::endl;
<< std::endl; }
LogRangeAsType<float>(std::cout << "out_device: ", out.mData, ",") << std::endl;
}; pass = pass && single_pass;
}; };
if(do_dumpout) if(do_dumpout)
...@@ -474,7 +409,7 @@ void profile_reduce_impl_impl(bool do_verification, ...@@ -474,7 +409,7 @@ void profile_reduce_impl_impl(bool do_verification,
dumpBufferToFile("dump_out.bin", out.mData.data(), out.mDesc.GetElementSize()); dumpBufferToFile("dump_out.bin", out.mData.data(), out.mDesc.GetElementSize());
dumpBufferToFile( dumpBufferToFile(
"dump_out_host.bin", out_ref.mData.data(), out_ref.mDesc.GetElementSize()); "dump_out_host.bin", out_ref.mData.data(), out_ref.mDesc.GetElementSize());
if(NeedIndices) if(OutputIndex)
{ {
dumpBufferToFile("dump_indices.bin", dumpBufferToFile("dump_indices.bin",
out_indices.mData.data(), out_indices.mData.data(),
...@@ -486,158 +421,34 @@ void profile_reduce_impl_impl(bool do_verification, ...@@ -486,158 +421,34 @@ void profile_reduce_impl_impl(bool do_verification,
}; };
}; };
for(auto& reduce_ptr : reduce1_ptrs) if(time_kernel)
{ std::cout << "Best Perf: " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s"
auto wsSizeInBytes = reduce_ptr->GetWorkspaceSizeInBytes(i_inLengths, reduceDims); << std::endl;
DeviceMem ws_dev(wsSizeInBytes);
InElementwiseOperation_1 in_elementwise_op_1(static_cast<int32_t>(reduce_total_length));
AccElementwiseOperation_1 acc_elementwise_op_1(
static_cast<int32_t>(reduce_total_length));
auto argument_ptr = reduce_ptr->MakeArgumentPointer(i_inLengths,
i_inStrides,
i_outLengths,
i_outStrides,
reduceDims,
alpha,
beta,
in_dev.GetDeviceBuffer(),
out_dev.GetDeviceBuffer(),
out_indices_dev.GetDeviceBuffer(),
ws_dev.GetDeviceBuffer(),
in_elementwise_op_1,
acc_elementwise_op_1);
if(!reduce_ptr->IsSupportedArgument(argument_ptr.get()))
continue;
std::string reduce_name = reduce_ptr->GetTypeString();
auto invoker_ptr = reduce_ptr->MakeInvokerPointer();
float avg_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t num_bytes =
invariant_total_length * reduce_total_length * sizeof(InDataType) +
invariant_total_length * sizeof(OutDataType);
std::vector<int> inLengths2 = reduce_ptr->GetWorkspace2dLengths(argument_ptr.get());
std::vector<int> inStrides2{inLengths2[1], 1};
for(auto& reduce2_ptr : reduce2_ptrs)
{
InElementwiseOperation_2 in_elementwise_op_2(
static_cast<int32_t>(reduce_total_length));
AccElementwiseOperation_2 acc_elementwise_op_2(
static_cast<int32_t>(reduce_total_length));
auto argument2_ptr =
reduce2_ptr->MakeArgumentPointer(inLengths2,
inStrides2,
i_outLengths,
i_outStrides,
reduceDims,
alpha,
beta,
ws_dev.GetDeviceBuffer(),
out_dev.GetDeviceBuffer(),
out_indices_dev.GetDeviceBuffer(),
ws_dev.GetDeviceBuffer(),
in_elementwise_op_2,
acc_elementwise_op_2);
if(!reduce2_ptr->IsSupportedArgument(argument2_ptr.get()))
continue;
std::string reduce2_name = reduce2_ptr->GetTypeString();
auto invoker2_ptr = reduce2_ptr->MakeInvokerPointer();
float avg_time_2 =
invoker2_ptr->Run(argument2_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t num_bytes_2 =
static_cast<size_t>(inLengths2[0]) * inLengths2[1] * sizeof(AccDataType);
float gb_per_sec = (num_bytes + num_bytes_2) / 1.E6 / (avg_time + avg_time_2);
std::cout << "Perf: " << (avg_time + avg_time_2) << " ms, " << gb_per_sec
<< " GB/s, " << reduce_name << " => " << reduce2_name << std::endl;
if(gb_per_sec > best_gb_per_sec)
{
best_avg_time = avg_time + avg_time_2;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
out_dev.FromDevice(out.mData.data());
ck::utils::check_err(out.mData, out_ref.mData);
if(NeedIndices)
{
out_indices_dev.FromDevice(out_indices.mData.data());
ck::utils::check_err(out_indices.mData, out_indices_ref.mData);
;
};
if(do_log)
{
LogRangeAsType<float>(std::cout << "out_host : ", out_ref.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "out_device: ", out.mData, ",")
<< std::endl;
}
}
if(do_dumpout)
{
dumpBufferToFile("dump_in.bin", in.mData.data(), in.mDesc.GetElementSize());
dumpBufferToFile("dump_out.bin", out.mData.data(), out.mDesc.GetElementSize());
dumpBufferToFile(
"dump_out_host.bin", out_ref.mData.data(), out_ref.mDesc.GetElementSize());
if(NeedIndices)
{
dumpBufferToFile("dump_indices.bin",
out_indices.mData.data(),
out_indices.mDesc.GetElementSize());
dumpBufferToFile("dump_indices_host.bin",
out_indices_ref.mData.data(),
out_indices_ref.mDesc.GetElementSize());
};
};
};
};
std::cout << "Best Perf: " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s"
<< std::endl;
} }
else else
{ {
std::cout << "The requested reduction operation is not supported, please check !!!" std::cout << "The requested reduction operation is not supported, please check !!!"
<< std::endl; << std::endl;
}; };
return pass;
}; };
template <typename InDataType, typename AccDataType, typename OutDataType> template <typename InDataType, typename AccDataType, typename OutDataType>
void profile_reduce_impl(bool do_verification, bool profile_reduce_impl(bool do_verification,
int init_method, int init_method,
bool do_log,
bool do_dumpout, bool do_dumpout,
bool time_kernel, bool time_kernel,
const std::vector<size_t>& inLengths, const std::vector<size_t>& inLengths,
const std::vector<int>& reduceDims, const std::vector<int>& reduceDims,
ReduceTensorOp ReduceOpId, ReduceTensorOp ReduceOpId,
NanPropagation NanOpt, bool PropagateNan,
ReduceTensorIndices IndicesOpt, bool UseIndex,
float alpha, float alpha,
float beta) float beta)
{ {
bool matched = false; bool matched = false;
bool pass = true;
using tuple_of_description_instances = using tuple_of_description_instances =
tensor_operation::device::device_reduce_instance::reduce_description_instances; tensor_operation::device::device_reduce_instance::reduce_description_instances;
...@@ -651,29 +462,30 @@ void profile_reduce_impl(bool do_verification, ...@@ -651,29 +462,30 @@ void profile_reduce_impl(bool do_verification,
using descType = remove_cvref_t<decltype(std::get<i>(tuple_object))>; using descType = remove_cvref_t<decltype(std::get<i>(tuple_object))>;
if(!description_match( if(!description_match(
descType{}, inLengths.size(), reduceDims, ReduceOpId, NanOpt, IndicesOpt)) descType{}, inLengths.size(), reduceDims, ReduceOpId, PropagateNan, UseIndex))
return; return;
profile_reduce_impl_impl<InDataType, pass = pass &&
AccDataType, profile_reduce_impl_impl<InDataType,
OutDataType, AccDataType,
descType::Rank_, OutDataType,
descType::NumReduceDim_, descType::Rank_,
static_cast<ReduceTensorOp>(descType::ReduceOpId_), descType::NumReduceDim_,
static_cast<NanPropagation>(descType::NanOpt_), static_cast<ReduceTensorOp>(descType::ReduceOpId_),
static_cast<ReduceTensorIndices>(descType::IndicesOpt_)>( static_cast<bool>(descType::PropagateNan_),
do_verification, static_cast<bool>(descType::UseIndex_)>(do_verification,
init_method, init_method,
do_log, do_dumpout,
do_dumpout, time_kernel,
time_kernel, inLengths,
inLengths, reduceDims,
reduceDims, alpha,
alpha, beta);
beta);
matched = true; matched = true;
}); });
return pass;
}; };
} // namespace profiler } // namespace profiler
......
profiler/src/profile_batched_gemm.cpp
View file @ bccc6d8b
...@@ -396,5 +396,5 @@ int profile_batched_gemm(int argc, char* argv[]) ...@@ -396,5 +396,5 @@ int profile_batched_gemm(int argc, char* argv[])
throw std::runtime_error("wrong! this GEMM data_type & layout is not implemented"); throw std::runtime_error("wrong! this GEMM data_type & layout is not implemented");
} }
return 1; return 0;
} }
profiler/src/profile_batched_gemm_reduce.cpp
View file @ bccc6d8b
...@@ -149,5 +149,5 @@ int profile_batched_gemm_reduce(int argc, char* argv[]) ...@@ -149,5 +149,5 @@ int profile_batched_gemm_reduce(int argc, char* argv[])
throw std::runtime_error("wrong! this data_type & layout is not implemented"); throw std::runtime_error("wrong! this data_type & layout is not implemented");
} }
return 1; return 0;
} }
profiler/src/profile_conv_bwd_weight.cpp
View file @ bccc6d8b
...@@ -142,5 +142,5 @@ int profile_conv_bwd_weight(int argc, char* argv[]) ...@@ -142,5 +142,5 @@ int profile_conv_bwd_weight(int argc, char* argv[])
throw std::runtime_error("wrong! this Conv data_type & layout is not implemented"); throw std::runtime_error("wrong! this Conv data_type & layout is not implemented");
} }
return 1; return 0;
} }
profiler/src/profile_conv_fwd_bias_relu.cpp
View file @ bccc6d8b
...@@ -110,5 +110,5 @@ int profile_conv_fwd_bias_relu(int argc, char* argv[]) ...@@ -110,5 +110,5 @@ int profile_conv_fwd_bias_relu(int argc, char* argv[])
throw std::runtime_error("wrong! data_type & layout for this operator is not implemented"); throw std::runtime_error("wrong! data_type & layout for this operator is not implemented");
} }
return 1; return 0;
} }
profiler/src/profile_conv_fwd_bias_relu_add.cpp
View file @ bccc6d8b
...@@ -111,5 +111,5 @@ int profile_conv_fwd_bias_relu_add(int argc, char* argv[]) ...@@ -111,5 +111,5 @@ int profile_conv_fwd_bias_relu_add(int argc, char* argv[])
throw std::runtime_error("wrong! data_type & layout for this operator is not implemented"); throw std::runtime_error("wrong! data_type & layout for this operator is not implemented");
} }
return 1; return 0;
} }
profiler/src/profile_conv_fwd_bias_relu_atomic_add.cpp
View file @ bccc6d8b
...@@ -112,5 +112,5 @@ int profile_conv_fwd_bias_relu_atomic_add(int argc, char* argv[]) ...@@ -112,5 +112,5 @@ int profile_conv_fwd_bias_relu_atomic_add(int argc, char* argv[])
throw std::runtime_error("wrong! data_type & layout for this operator is not implemented"); throw std::runtime_error("wrong! data_type & layout for this operator is not implemented");
} }
return 1; return 0;
} }
profiler/src/profile_convnd_fwd.cpp
View file @ bccc6d8b
...@@ -347,5 +347,5 @@ int ck::profiler::profile_convnd_fwd(int argc, char* argv[]) ...@@ -347,5 +347,5 @@ int ck::profiler::profile_convnd_fwd(int argc, char* argv[])
std::to_string(num_dim_spatial)); std::to_string(num_dim_spatial));
} }
return 1; return 0;
} }
profiler/src/profile_gemm.cpp
View file @ bccc6d8b
...@@ -68,6 +68,7 @@ int profile_gemm(int argc, char* argv[]) ...@@ -68,6 +68,7 @@ int profile_gemm(int argc, char* argv[])
ck::profiler::profile_gemm_impl<ck::half_t, ck::profiler::profile_gemm_impl<ck::half_t,
ck::half_t, ck::half_t,
ck::half_t, ck::half_t,
float,
ck::tensor_layout::gemm::RowMajor, ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor, ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>( ck::tensor_layout::gemm::RowMajor>(
...@@ -88,6 +89,7 @@ int profile_gemm(int argc, char* argv[]) ...@@ -88,6 +89,7 @@ int profile_gemm(int argc, char* argv[])
ck::profiler::profile_gemm_impl<ck::half_t, ck::profiler::profile_gemm_impl<ck::half_t,
ck::half_t, ck::half_t,
ck::half_t, ck::half_t,
float,
ck::tensor_layout::gemm::RowMajor, ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor, ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>( ck::tensor_layout::gemm::RowMajor>(
...@@ -108,6 +110,7 @@ int profile_gemm(int argc, char* argv[]) ...@@ -108,6 +110,7 @@ int profile_gemm(int argc, char* argv[])
ck::profiler::profile_gemm_impl<ck::half_t, ck::profiler::profile_gemm_impl<ck::half_t,
ck::half_t, ck::half_t,
ck::half_t, ck::half_t,
float,
ck::tensor_layout::gemm::ColumnMajor, ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor, ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>( ck::tensor_layout::gemm::RowMajor>(
...@@ -128,6 +131,7 @@ int profile_gemm(int argc, char* argv[]) ...@@ -128,6 +131,7 @@ int profile_gemm(int argc, char* argv[])
ck::profiler::profile_gemm_impl<ck::half_t, ck::profiler::profile_gemm_impl<ck::half_t,
ck::half_t, ck::half_t,
ck::half_t, ck::half_t,
float,
ck::tensor_layout::gemm::ColumnMajor, ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::ColumnMajor, ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>( ck::tensor_layout::gemm::RowMajor>(
...@@ -146,6 +150,7 @@ int profile_gemm(int argc, char* argv[]) ...@@ -146,6 +150,7 @@ int profile_gemm(int argc, char* argv[])
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN) else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
{ {
ck::profiler::profile_gemm_impl<float, ck::profiler::profile_gemm_impl<float,
float,
float, float,
float, float,
ck::tensor_layout::gemm::RowMajor, ck::tensor_layout::gemm::RowMajor,
...@@ -166,6 +171,7 @@ int profile_gemm(int argc, char* argv[]) ...@@ -166,6 +171,7 @@ int profile_gemm(int argc, char* argv[])
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_NK_MN) else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_NK_MN)
{ {
ck::profiler::profile_gemm_impl<float, ck::profiler::profile_gemm_impl<float,
float,
float, float,
float, float,
ck::tensor_layout::gemm::RowMajor, ck::tensor_layout::gemm::RowMajor,
...@@ -186,6 +192,7 @@ int profile_gemm(int argc, char* argv[]) ...@@ -186,6 +192,7 @@ int profile_gemm(int argc, char* argv[])
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_KN_MN) else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_KN_MN)
{ {
ck::profiler::profile_gemm_impl<float, ck::profiler::profile_gemm_impl<float,
float,
float, float,
float, float,
ck::tensor_layout::gemm::ColumnMajor, ck::tensor_layout::gemm::ColumnMajor,
...@@ -206,6 +213,7 @@ int profile_gemm(int argc, char* argv[]) ...@@ -206,6 +213,7 @@ int profile_gemm(int argc, char* argv[])
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_NK_MN) else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_NK_MN)
{ {
ck::profiler::profile_gemm_impl<float, ck::profiler::profile_gemm_impl<float,
float,
float, float,
float, float,
ck::tensor_layout::gemm::ColumnMajor, ck::tensor_layout::gemm::ColumnMajor,
...@@ -228,6 +236,7 @@ int profile_gemm(int argc, char* argv[]) ...@@ -228,6 +236,7 @@ int profile_gemm(int argc, char* argv[])
ck::profiler::profile_gemm_impl<int8_t, ck::profiler::profile_gemm_impl<int8_t,
int8_t, int8_t,
int8_t, int8_t,
int32_t,
ck::tensor_layout::gemm::RowMajor, ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor, ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>( ck::tensor_layout::gemm::RowMajor>(
...@@ -248,6 +257,7 @@ int profile_gemm(int argc, char* argv[]) ...@@ -248,6 +257,7 @@ int profile_gemm(int argc, char* argv[])
ck::profiler::profile_gemm_impl<int8_t, ck::profiler::profile_gemm_impl<int8_t,
int8_t, int8_t,
int8_t, int8_t,
int32_t,
ck::tensor_layout::gemm::RowMajor, ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor, ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>( ck::tensor_layout::gemm::RowMajor>(
...@@ -268,6 +278,7 @@ int profile_gemm(int argc, char* argv[]) ...@@ -268,6 +278,7 @@ int profile_gemm(int argc, char* argv[])
ck::profiler::profile_gemm_impl<int8_t, ck::profiler::profile_gemm_impl<int8_t,
int8_t, int8_t,
int8_t, int8_t,
int32_t,
ck::tensor_layout::gemm::ColumnMajor, ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor, ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>( ck::tensor_layout::gemm::RowMajor>(
...@@ -288,6 +299,7 @@ int profile_gemm(int argc, char* argv[]) ...@@ -288,6 +299,7 @@ int profile_gemm(int argc, char* argv[])
ck::profiler::profile_gemm_impl<int8_t, ck::profiler::profile_gemm_impl<int8_t,
int8_t, int8_t,
int8_t, int8_t,
int32_t,
ck::tensor_layout::gemm::ColumnMajor, ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::ColumnMajor, ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>( ck::tensor_layout::gemm::RowMajor>(
...@@ -308,6 +320,7 @@ int profile_gemm(int argc, char* argv[]) ...@@ -308,6 +320,7 @@ int profile_gemm(int argc, char* argv[])
ck::profiler::profile_gemm_impl<ck::bhalf_t, ck::profiler::profile_gemm_impl<ck::bhalf_t,
ck::bhalf_t, ck::bhalf_t,
ck::bhalf_t, ck::bhalf_t,
float,
ck::tensor_layout::gemm::RowMajor, ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor, ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>( ck::tensor_layout::gemm::RowMajor>(
...@@ -328,6 +341,7 @@ int profile_gemm(int argc, char* argv[]) ...@@ -328,6 +341,7 @@ int profile_gemm(int argc, char* argv[])
ck::profiler::profile_gemm_impl<ck::bhalf_t, ck::profiler::profile_gemm_impl<ck::bhalf_t,
ck::bhalf_t, ck::bhalf_t,
ck::bhalf_t, ck::bhalf_t,
float,
ck::tensor_layout::gemm::RowMajor, ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor, ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>( ck::tensor_layout::gemm::RowMajor>(
...@@ -348,6 +362,7 @@ int profile_gemm(int argc, char* argv[]) ...@@ -348,6 +362,7 @@ int profile_gemm(int argc, char* argv[])
ck::profiler::profile_gemm_impl<ck::bhalf_t, ck::profiler::profile_gemm_impl<ck::bhalf_t,
ck::bhalf_t, ck::bhalf_t,
ck::bhalf_t, ck::bhalf_t,
float,
ck::tensor_layout::gemm::ColumnMajor, ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor, ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>( ck::tensor_layout::gemm::RowMajor>(
...@@ -368,6 +383,7 @@ int profile_gemm(int argc, char* argv[]) ...@@ -368,6 +383,7 @@ int profile_gemm(int argc, char* argv[])
ck::profiler::profile_gemm_impl<ck::bhalf_t, ck::profiler::profile_gemm_impl<ck::bhalf_t,
ck::bhalf_t, ck::bhalf_t,
ck::bhalf_t, ck::bhalf_t,
float,
ck::tensor_layout::gemm::ColumnMajor, ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::ColumnMajor, ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>( ck::tensor_layout::gemm::RowMajor>(
...@@ -388,5 +404,5 @@ int profile_gemm(int argc, char* argv[]) ...@@ -388,5 +404,5 @@ int profile_gemm(int argc, char* argv[])
throw std::runtime_error("wrong! this GEMM data_type & layout is not implemented"); throw std::runtime_error("wrong! this GEMM data_type & layout is not implemented");
} }
return 1; return 0;
} }
profiler/src/profile_gemm_bias_2d.cpp
View file @ bccc6d8b
...@@ -252,5 +252,5 @@ int profile_gemm_bias_2d(int argc, char* argv[]) ...@@ -252,5 +252,5 @@ int profile_gemm_bias_2d(int argc, char* argv[])
throw std::runtime_error("wrong! this data_type & layout is not implemented"); throw std::runtime_error("wrong! this data_type & layout is not implemented");
} }
return 1; return 0;
} }
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