Skip to content

GitLab

Commit 68886f7d authored by raman jana's avatar raman jana
Browse files

merging with latest develop branch

parents a9ee2960 1677cf70
Hide whitespace changes
Inline Side-by-side
Showing with 261 additions and 647 deletions
library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_partial_reduce_f16_f16_f16.cpp deleted 100644 → 0
View file @ a9ee2960
#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 @ a9ee2960
#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 @ a9ee2960
#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 @ a9ee2960
#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 @ a9ee2960
#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 @ a9ee2960
#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 @ 68886f7d
include_directories(BEFORE
${PROJECT_SOURCE_DIR}/include/ck
${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
${PROJECT_SOURCE_DIR}/include/ck/problem_transform
......
profiler/include/profile_batched_gemm_impl.hpp
View file @ 68886f7d
......@@ -63,7 +63,7 @@ template <typename ADataType,
bool profile_batched_gemm_impl(int do_verification,
int init_method,
bool do_log,
int nrepeat,
bool time_kernel,
int M,
int N,
int K,
......@@ -356,11 +356,12 @@ bool profile_batched_gemm_impl(int do_verification,
{
std::string gemm_name = gemm_ptr->GetTypeString();
float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * BatchCount * M * N * K;
std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(BDataType) * K * M +
std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(CDataType) * M * N) *
BatchCount;
......
profiler/include/profile_batched_gemm_reduce_impl.hpp
View file @ 68886f7d
......@@ -17,11 +17,21 @@ namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F32 = float;
using F16 = ck::half_t;
using DPtrsGlobal = ck::Tuple<F32*, F32*>;
using Identity = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Identity, Identity>;
using DeviceGemmReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmReducePtr<
DPtrsGlobal,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::UnarySquare<float, float, false>>;
DInElementOps,
DOutElementOps>;
void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances(
std::vector<DeviceGemmReduceNoOpPtr>&);
......@@ -53,7 +63,7 @@ template <typename ADataType,
bool profile_batched_gemm_reduce_impl(int do_verification,
int init_method,
bool do_log,
int nrepeat,
bool time_kernel,
int M,
int N,
int K,
......@@ -119,19 +129,25 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
}
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::reduce::Add<float>;
using D1ReduceOp = ck::reduce::Add<float>;
using D1ElementOp = ck::tensor_operation::element_wise::UnarySquare<float, float, false>;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
const auto d0_reduce_op = D0ReduceOp{};
const auto d1_reduce_op = D1ReduceOp{};
const auto d1_element_op = D1ElementOp{};
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::reduce::Add<float>;
using D1ReduceOp = ck::reduce::Add<float>;
using UnaryIdenticElementOp =
ck::tensor_operation::element_wise::UnaryIdentic<float, float, false>;
using UnarySquareElementOp =
ck::tensor_operation::element_wise::UnarySquare<float, float, false>;
using DxsInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOps = ck::Tuple<UnaryIdenticElementOp, UnaryIdenticElementOp>;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
const auto dxs_in_element_op = DxsInElementOps{};
const auto dxs_out_element_op = DxsOutElementOps{};
const auto d0_reduce_op = D0ReduceOp{};
const auto d1_reduce_op = D1ReduceOp{};
if(do_verification)
{
......@@ -155,15 +171,15 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
{
for(int m = 0; m < M; ++m)
{
float d0_acc = d0_reduce_op.GetReductionZeroVal();
float d1_acc = d1_reduce_op.GetReductionZeroVal();
float d0_acc = d0_reduce_op.GetIdentityValue();
float d1_acc = d1_reduce_op.GetIdentityValue();
for(int n = 0; n < N; ++n)
{
float d0_val = ck::type_convert<float>(c_g_m_n_host_result(batch, m, n));
float d1_val;
d1_element_op(d1_val, d0_val);
UnarySquareElementOp{}(d1_val, d0_val);
d0_reduce_op(d0_acc, d0_val);
d1_reduce_op(d1_acc, d1_val);
}
......@@ -180,6 +196,9 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
DeviceMem d0_device_buf(sizeof(DDataType) * d0_g_m_device_result.mDesc.GetElementSpace());
DeviceMem d1_device_buf(sizeof(DDataType) * d1_g_m_device_result.mDesc.GetElementSpace());
auto dxs_global = ck::make_tuple(static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()));
a_device_buf.ToDevice(a_g_m_k.mData.data());
b_device_buf.ToDevice(b_g_k_n.mData.data());
......@@ -241,8 +260,7 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
gemm_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()),
dxs_global,
M,
N,
K,
......@@ -252,37 +270,20 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
a_element_op,
b_element_op,
c_element_op,
d1_element_op,
dxs_in_element_op,
dxs_out_element_op,
BatchCount);
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
// warm up
invoker_ptr->Run(argument_ptr.get());
// timing
float total_time = 0;
for(int i = 0; i < nrepeat; ++i)
{
// init DO, D1 to 0
d0_device_buf.SetZero();
d1_device_buf.SetZero();
KernelTimer timer;
timer.Start();
invoker_ptr->Run(argument_ptr.get());
timer.End();
total_time += timer.GetElapsedTime();
}
// init DO, D1 to 0
d0_device_buf.SetZero();
d1_device_buf.SetZero();
float ave_time = total_time / nrepeat;
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::string gemm_name = gemm_ptr->GetTypeString();
......
profiler/include/profile_conv_bwd_data_impl.hpp deleted 100644 → 0
View file @ a9ee2960
#pragma once
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "device_conv_bwd_data.hpp"
#include "element_wise_operation.hpp"
#include "reference_conv_bwd_data.hpp"
using F16 = ck::half_t;
using F32 = float;
using BF16 = ck::bhalf_t;
using INT8 = int8_t;
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_conv2d_bwd_data_instance {
using DeviceConvBwdDataNoOpPtr =
DeviceConvBwdDataPtr<ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
} // namespace device_conv2d_bwd_data_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace ck {
namespace profiler {
template <int NDimSpatial,
typename InDataType,
typename WeiDataType,
typename OutDataType,
typename AccDataType,
typename InLayout,
typename WeiLayout,
typename OutLayout>
void profile_conv_bwd_data_impl(int do_verification,
int init_method,
bool do_log,
int nrepeat,
ck::index_t N,
ck::index_t K,
ck::index_t C,
std::vector<ck::index_t> input_spatial_lengths,
std::vector<ck::index_t> filter_spatial_lengths,
std::vector<ck::index_t> output_spatial_lengths,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads)
{
const ck::index_t Y = filter_spatial_lengths[0];
const ck::index_t X = filter_spatial_lengths[1];
const ck::index_t Hi = input_spatial_lengths[0];
const ck::index_t Wi = input_spatial_lengths[1];
const ck::index_t Ho = output_spatial_lengths[0];
const ck::index_t Wo = output_spatial_lengths[1];
auto f_host_tensor_descriptor =
[](std::size_t N_, std::size_t C_, std::size_t H, std::size_t W, auto layout) {
if constexpr(is_same<decltype(layout), ck::tensor_layout::convolution::NCHW>::value ||
is_same<decltype(layout), ck::tensor_layout::convolution::KCYX>::value ||
is_same<decltype(layout), ck::tensor_layout::convolution::NKHW>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
std::vector<std::size_t>({C_ * H * W, H * W, W, 1}));
}
else if constexpr(is_same<decltype(layout), tensor_layout::convolution::NHWC>::value ||
is_same<decltype(layout), tensor_layout::convolution::KYXC>::value ||
is_same<decltype(layout), tensor_layout::convolution::NHWK>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
std::vector<std::size_t>({C_ * H * W, 1, W * C_, C_}));
}
};
Tensor<InDataType> in_n_c_hi_wi_host_result(f_host_tensor_descriptor(N, C, Hi, Wi, InLayout{}));
Tensor<InDataType> in_n_c_hi_wi_device_result(
f_host_tensor_descriptor(N, C, Hi, Wi, InLayout{}));
Tensor<WeiDataType> wei_k_c_y_x(f_host_tensor_descriptor(K, C, Y, X, WeiLayout{}));
Tensor<OutDataType> out_n_k_ho_wo(f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi_host_result.mDesc << std::endl;
std::cout << "wei_k_c_y_x: " << wei_k_c_y_x.mDesc << std::endl;
std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
break;
default:
out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
}
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
if(do_verification)
{
using ReferenceConvBwdDataInstance =
ck::tensor_operation::host::ReferenceConvBwdData<InDataType,
WeiDataType,
OutDataType,
AccDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
auto ref_conv = ReferenceConvBwdDataInstance{};
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi_host_result,
wei_k_c_y_x,
out_n_k_ho_wo,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op);
ref_invoker.Run(ref_argument);
}
DeviceMem in_device_buf(sizeof(InDataType) *
in_n_c_hi_wi_device_result.mDesc.GetElementSpace());
DeviceMem wei_device_buf(sizeof(WeiDataType) * wei_k_c_y_x.mDesc.GetElementSpace());
DeviceMem out_device_buf(sizeof(OutDataType) * out_n_k_ho_wo.mDesc.GetElementSpace());
out_device_buf.ToDevice(out_n_k_ho_wo.mData.data());
wei_device_buf.ToDevice(wei_k_c_y_x.mData.data());
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceConvBwdDataNoOpPtr =
ck::tensor_operation::device::DeviceConvBwdDataPtr<PassThrough, PassThrough, PassThrough>;
// add device Conv instances
std::vector<DeviceConvBwdDataNoOpPtr> conv_ptrs;
if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, float> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, float> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, float>)
{
ck::tensor_operation::device::device_conv2d_bwd_data_instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(conv_ptrs);
}
else if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, ck::half_t> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, ck::half_t> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, ck::half_t>)
{
ck::tensor_operation::device::device_conv2d_bwd_data_instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(conv_ptrs);
}
else if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, ck::bhalf_t> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, ck::bhalf_t> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, ck::bhalf_t>)
{
ck::tensor_operation::device::device_conv2d_bwd_data_instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(conv_ptrs);
}
else if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, int8_t> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, int8_t> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, int8_t>)
{
ck::tensor_operation::device::device_conv2d_bwd_data_instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(conv_ptrs);
}
if(conv_ptrs.size() <= 0)
{
throw std::runtime_error("wrong! no device Conv instance found");
}
std::string best_conv_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device Conv instances
for(auto& conv_ptr : conv_ptrs)
{
auto argument_ptr = conv_ptr->MakeArgumentPointer(
static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op);
auto invoker_ptr = conv_ptr->MakeInvokerPointer();
if(conv_ptr->IsSupportedArgument(argument_ptr.get()))
{
std::string conv_name = conv_ptr->GetTypeString();
float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
std::size_t num_btype = sizeof(InDataType) * (N * C * Hi * Wi) +
sizeof(WeiDataType) * (K * C * Y * X) +
sizeof(OutDataType) * (N * K * Ho * Wo);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s, " << conv_name << std::endl;
if(tflops > best_tflops)
{
best_conv_name = conv_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
in_device_buf.FromDevice(in_n_c_hi_wi_device_result.mData.data());
ck::utils::check_err(in_n_c_hi_wi_device_result.mData,
in_n_c_hi_wi_host_result.mData);
if(do_log)
{
LogRangeAsType<float>(std::cout << "in : ", out_n_k_ho_wo.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "wei: ", wei_k_c_y_x.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "out_host : ", in_n_c_hi_wi_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "out_device: ", in_n_c_hi_wi_device_result.mData, ",")
<< std::endl;
}
}
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_conv_name << std::endl;
}
} // namespace profiler
} // namespace ck
profiler/include/profile_conv_bwd_weight_impl.hpp
View file @ 68886f7d
#pragma once
#include "stream_config.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
......@@ -43,7 +45,7 @@ template <int NDimSpatial,
bool profile_conv_bwd_weight_impl(int do_verification,
int init_method,
bool do_log,
int nrepeat,
bool time_kernel,
ck::index_t N,
ck::index_t K,
ck::index_t C,
......@@ -182,6 +184,7 @@ bool profile_conv_bwd_weight_impl(int do_verification,
// profile device Conv instances
bool pass = true;
for(auto& conv_ptr : conv_ptrs)
{
// using atomic, so need to reset input
......@@ -189,6 +192,7 @@ bool profile_conv_bwd_weight_impl(int do_verification,
{
wei_device_buf.SetZero();
}
auto argument_ptr = conv_ptr->MakeArgumentPointer(
static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
......@@ -214,7 +218,8 @@ bool profile_conv_bwd_weight_impl(int do_verification,
{
std::string conv_name = conv_ptr->GetTypeString();
float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
......@@ -242,6 +247,7 @@ bool profile_conv_bwd_weight_impl(int do_verification,
wei_device_buf.FromDevice(wei_k_c_y_x_device_result.mData.data());
float max_error = check_error(wei_k_c_y_x_host_result, wei_k_c_y_x_device_result);
if(max_error > 8)
{
pass = false;
......
profiler/include/profile_conv_fwd_bias_relu_add_impl.hpp
View file @ 68886f7d
......@@ -42,7 +42,7 @@ template <int NDimSpatial,
void profile_conv_fwd_bias_relu_add_impl(int do_verification,
int init_method,
bool do_log,
int nrepeat,
bool time_kernel,
ck::index_t N,
ck::index_t K,
ck::index_t C,
......@@ -219,7 +219,8 @@ void profile_conv_fwd_bias_relu_add_impl(int do_verification,
{
std::string conv_name = op_ptr->GetTypeString();
float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
......
profiler/include/profile_conv_fwd_bias_relu_atomic_add_impl.hpp
View file @ 68886f7d
......@@ -119,7 +119,7 @@ template <int NDimSpatial,
void profile_conv_fwd_bias_relu_atomic_add_impl(int do_verification,
int init_method,
bool do_log,
int nrepeat,
bool time_kernel,
ck::index_t N,
ck::index_t K,
ck::index_t C,
......@@ -275,7 +275,8 @@ void profile_conv_fwd_bias_relu_atomic_add_impl(int do_verification,
{
std::string conv_name = op_ptr->GetTypeString();
float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
......
profiler/include/profile_conv_fwd_bias_relu_impl.hpp
View file @ 68886f7d
......@@ -41,7 +41,7 @@ template <int NDimSpatial,
void profile_conv_fwd_bias_relu_impl(int do_verification,
int init_method,
bool do_log,
int nrepeat,
bool time_kernel,
ck::index_t N,
ck::index_t K,
ck::index_t C,
......@@ -207,7 +207,8 @@ void profile_conv_fwd_bias_relu_impl(int do_verification,
{
std::string conv_name = op_ptr->GetTypeString();
float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
......
profiler/include/profile_convnd_bwd_data_impl.hpp
View file @ 68886f7d
......@@ -269,7 +269,7 @@ template <int NDimSpatial,
bool profile_convnd_bwd_data_impl(int do_verification,
int init_method,
bool do_log,
int nrepeat,
bool time_kernel,
ck::index_t N,
ck::index_t K,
ck::index_t C,
......@@ -410,7 +410,8 @@ bool profile_convnd_bwd_data_impl(int do_verification,
{
std::string conv_name = conv_ptr->GetTypeString();
float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop =
ck::utils::conv::get_flops(N, C, K, filter_spatial_lengths, output_spatial_lengths);
......
profiler/include/profile_gemm_bias_2d_impl.hpp
View file @ 68886f7d
......@@ -65,7 +65,7 @@ template <typename ADataType,
void profile_gemm_bias_2d_impl(int do_verification,
int init_method,
bool do_log,
int nrepeat,
bool time_kernel,
int M,
int N,
int K,
......@@ -259,7 +259,8 @@ void profile_gemm_bias_2d_impl(int do_verification,
{
std::string gemm_name = gemm_ptr->GetTypeString();
float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
......
profiler/include/profile_gemm_bias_relu_add_impl.hpp
View file @ 68886f7d
......@@ -48,7 +48,7 @@ template <typename ADataType,
void profile_gemm_bias_relu_add_impl(int do_verification,
int init_method,
bool do_log,
int nrepeat,
bool time_kernel,
int M,
int N,
int K,
......@@ -232,7 +232,8 @@ void profile_gemm_bias_relu_add_impl(int do_verification,
{
std::string gemm_name = gemm_ptr->GetTypeString();
float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
......
profiler/include/profile_gemm_bias_relu_impl.hpp
View file @ 68886f7d
......@@ -48,7 +48,7 @@ template <typename ADataType,
void profile_gemm_bias_relu_impl(int do_verification,
int init_method,
bool do_log,
int nrepeat,
bool time_kernel,
int M,
int N,
int K,
......@@ -212,7 +212,8 @@ void profile_gemm_bias_relu_impl(int do_verification,
{
std::string gemm_name = gemm_ptr->GetTypeString();
float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
......
profiler/include/profile_gemm_impl.hpp
View file @ 68886f7d
#pragma once
#include <iomanip>
#include <iostream>
#include <typeinfo>
#include "check_err.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
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_int8_int8_int8_mk_kn_mn_instances(
std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_int8_int8_int8_mk_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_i8_i8_i8_mk_kn_mn_instances(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_i8_i8_i8_km_kn_mn_instances(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_2_stage_f16_f16_f16_mk_nk_mn_instances(
std::vector<DeviceGemmNoOpPtr>&);
......@@ -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_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
} // namespace tensor_operation
......@@ -85,13 +98,14 @@ namespace profiler {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename AccDataType,
typename ALayout,
typename BLayout,
typename CLayout>
void profile_gemm_impl(int do_verification,
int init_method,
bool do_log,
int nrepeat,
bool time_kernel,
int M,
int N,
int K,
......@@ -125,7 +139,11 @@ void profile_gemm_impl(int do_verification,
std::size_t num_thread = 1;
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:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-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,
ck::tensor_operation::device::device_gemm_instance::
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::
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,
ck::tensor_operation::device::device_gemm_instance::
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::
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,
ck::tensor_operation::device::device_gemm_instance::
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::
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,
ck::tensor_operation::device::device_gemm_instance::
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::
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,
ck::tensor_operation::device::device_gemm_instance::
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::
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,
ck::tensor_operation::device::device_gemm_instance::
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::
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,
ck::tensor_operation::device::device_gemm_instance::
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::
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,
ck::tensor_operation::device::device_gemm_instance::
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::
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,
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
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 &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
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 &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
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 &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
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);
}
}
......@@ -416,7 +470,8 @@ void profile_gemm_impl(int do_verification,
std::string gemm_name = gemm_ptr->GetTypeString();
float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
......@@ -457,8 +512,14 @@ void profile_gemm_impl(int do_verification,
bf16_to_f32_(b_k_n, b_f32_k_n);
bf16_to_f32_(c_m_n_device_result, c_m_n_device_f32_result);
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<float, float, float, AElementOp, BElementOp, CElementOp>;
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceGemm<float,
float,
float,
float,
AElementOp,
BElementOp,
CElementOp>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
......@@ -490,6 +551,7 @@ void profile_gemm_impl(int do_verification,
ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
AccDataType,
AElementOp,
BElementOp,
CElementOp>;
......@@ -522,12 +584,50 @@ void profile_gemm_impl(int do_verification,
}
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, "
<< best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
if constexpr(is_same<CDataType, float>::value)
{
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
......
profiler/include/profile_gemm_reduce_impl.hpp
View file @ 68886f7d
......@@ -16,11 +16,22 @@ namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F32 = float;
using F16 = ck::half_t;
using DPtrsGlobal = ck::Tuple<F32*, F32*>;
using Div = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, true>;
using Identity = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Div, Div>;
using DeviceGemmReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmReducePtr<
DPtrsGlobal,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::UnarySquare<float, float, false>>;
DInElementOps,
DOutElementOps>;
void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances(
std::vector<DeviceGemmReduceNoOpPtr>&);
......@@ -52,7 +63,7 @@ template <typename ADataType,
bool profile_gemm_reduce_impl(int do_verification,
int init_method,
bool do_log,
int nrepeat,
bool time_kernel,
int M,
int N,
int K,
......@@ -112,24 +123,37 @@ bool profile_gemm_reduce_impl(int do_verification,
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
}
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::reduce::Add<float>;
using D1ReduceOp = ck::reduce::Add<float>;
using D1ElementOp = ck::tensor_operation::element_wise::UnarySquare<float, float, false>;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
const auto d0_reduce_op = D0ReduceOp{};
const auto d1_reduce_op = D1ReduceOp{};
const auto d1_element_op = D1ElementOp{};
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::reduce::Add<float>;
using D1ReduceOp = ck::reduce::Add<float>;
using UnaryDivElementOp = ck::tensor_operation::element_wise::UnaryIdentic<float, float, true>;
using UnaryIdenticElementOp =
ck::tensor_operation::element_wise::UnaryIdentic<float, float, false>;
using UnarySquareElementOp =
ck::tensor_operation::element_wise::UnarySquare<float, float, false>;
using DxsInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOps = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
const auto d0_reduce_op = D0ReduceOp{};
const auto d1_reduce_op = D1ReduceOp{};
auto dxs_in_element_op = DxsInElementOps{};
auto dxs_out_element_op = DxsOutElementOps{M, M};
if(do_verification)
{
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, AElementOp, BElementOp, CElementOp>;
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
DDataType,
AElementOp,
BElementOp,
CElementOp>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
......@@ -141,19 +165,23 @@ bool profile_gemm_reduce_impl(int do_verification,
for(int m = 0; m < M; ++m)
{
float d0_acc = d0_reduce_op.GetReductionZeroVal();
float d1_acc = d1_reduce_op.GetReductionZeroVal();
float d0_acc = d0_reduce_op.GetIdentityValue();
float d1_acc = d1_reduce_op.GetIdentityValue();
for(int n = 0; n < N; ++n)
{
float d0_val = ck::type_convert<float>(c_m_n_host_result(m, n));
float d1_val;
float c_val = ck::type_convert<float>(c_m_n_host_result(m, n));
float d0_val = 0;
float d1_val = 0;
d1_element_op(d1_val, d0_val);
dxs_in_element_op(ck::Number<0>{})(d0_val, c_val);
dxs_in_element_op(ck::Number<1>{})(d1_val, c_val);
d0_reduce_op(d0_acc, d0_val);
d1_reduce_op(d1_acc, d1_val);
}
dxs_out_element_op(ck::Number<0>{})(d0_acc, d0_acc);
dxs_out_element_op(ck::Number<1>{})(d1_acc, d1_acc);
d0_m_host_result(m) = ck::type_convert<DDataType>(d0_acc);
d1_m_host_result(m) = ck::type_convert<DDataType>(d1_acc);
}
......@@ -165,6 +193,9 @@ bool profile_gemm_reduce_impl(int do_verification,
DeviceMem d0_device_buf(sizeof(DDataType) * d0_m_device_result.mDesc.GetElementSpace());
DeviceMem d1_device_buf(sizeof(DDataType) * d1_m_device_result.mDesc.GetElementSpace());
auto dxs_global = ck::make_tuple(static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()));
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
......@@ -226,8 +257,7 @@ bool profile_gemm_reduce_impl(int do_verification,
gemm_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()),
dxs_global,
M,
N,
K,
......@@ -237,42 +267,25 @@ bool profile_gemm_reduce_impl(int do_verification,
a_element_op,
b_element_op,
c_element_op,
d1_element_op);
dxs_in_element_op,
dxs_out_element_op);
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
// warm up
invoker_ptr->Run(argument_ptr.get());
// timing
float total_time = 0;
for(int i = 0; i < nrepeat; ++i)
{
// init DO, D1 to 0
d0_device_buf.SetZero();
d1_device_buf.SetZero();
KernelTimer timer;
timer.Start();
invoker_ptr->Run(argument_ptr.get());
timer.End();
total_time += timer.GetElapsedTime();
}
// init DO, D1 to 0
d0_device_buf.SetZero();
d1_device_buf.SetZero();
float ave_time = total_time / nrepeat;
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::string gemm_name = gemm_ptr->GetTypeString();
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * M +
std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(CDataType) * M * N + sizeof(CDataType) * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
......
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