提交 f503f129 编写于 作者: T tensor-tang

enable dynamic load mklml lib on fluid

上级 25241e9e
...@@ -114,7 +114,12 @@ INCLUDE_DIRECTORIES(${CBLAS_INC_DIR}) ...@@ -114,7 +114,12 @@ INCLUDE_DIRECTORIES(${CBLAS_INC_DIR})
SET(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/cblas_dummy.c) SET(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/cblas_dummy.c)
FILE(WRITE ${dummyfile} "const char *dummy_cblas = \"${dummyfile}\";") FILE(WRITE ${dummyfile} "const char *dummy_cblas = \"${dummyfile}\";")
ADD_LIBRARY(cblas STATIC ${dummyfile}) ADD_LIBRARY(cblas STATIC ${dummyfile})
TARGET_LINK_LIBRARIES(cblas ${CBLAS_LIBRARIES})
IF("${CBLAS_PROVIDER}" STREQUAL "MKLML")
TARGET_LINK_LIBRARIES(cblas dynload_mklml)
ELSE()
TARGET_LINK_LIBRARIES(cblas ${CBLAS_LIBRARIES})
ENDIF("${CBLAS_PROVIDER}" STREQUAL "MKLML")
IF(NOT ${CBLAS_FOUND}) IF(NOT ${CBLAS_FOUND})
ADD_DEPENDENCIES(cblas extern_openblas) ADD_DEPENDENCIES(cblas extern_openblas)
......
...@@ -19,8 +19,8 @@ limitations under the License. */ ...@@ -19,8 +19,8 @@ limitations under the License. */
#include "gflags/gflags.h" #include "gflags/gflags.h"
#include "gtest/gtest.h" #include "gtest/gtest.h"
#include "paddle/fluid/inference/tests/test_helper.h" #include "paddle/fluid/inference/tests/test_helper.h"
#include "paddle/fluid/operators/math/blas.h"
#ifdef PADDLE_WITH_MKLML #ifdef PADDLE_WITH_MKLML
#include <mkl_service.h>
#include <omp.h> #include <omp.h>
#endif #endif
...@@ -164,7 +164,7 @@ TEST(inference, nlp) { ...@@ -164,7 +164,7 @@ TEST(inference, nlp) {
// only use 1 thread number per std::thread // only use 1 thread number per std::thread
omp_set_dynamic(0); omp_set_dynamic(0);
omp_set_num_threads(1); omp_set_num_threads(1);
mkl_set_num_threads(1); paddle::operators::math::SetNumThreads(1);
#endif #endif
double start_ms = 0, stop_ms = 0; double start_ms = 0, stop_ms = 0;
......
...@@ -18,10 +18,7 @@ ...@@ -18,10 +18,7 @@
#include "paddle/fluid/framework/tensor.h" #include "paddle/fluid/framework/tensor.h"
#ifdef PADDLE_WITH_MKLML #ifdef PADDLE_WITH_MKLML
#include <mkl_cblas.h> #include "paddle/fluid/platform/dynload/mklml.h"
#include <mkl_lapacke.h>
#include <mkl_service.h>
#include <mkl_vml_functions.h>
#endif #endif
#ifdef PADDLE_USE_OPENBLAS #ifdef PADDLE_USE_OPENBLAS
...@@ -55,7 +52,7 @@ static void SetNumThreads(int num_threads) { ...@@ -55,7 +52,7 @@ static void SetNumThreads(int num_threads) {
openblas_set_num_threads(real_num_threads); openblas_set_num_threads(real_num_threads);
#elif defined(PADDLE_WITH_MKLML) #elif defined(PADDLE_WITH_MKLML)
int real_num_threads = num_threads > 1 ? num_threads : 1; int real_num_threads = num_threads > 1 ? num_threads : 1;
mkl_set_num_threads(real_num_threads); platform::dynload::MKL_Set_Num_Threads(real_num_threads);
#else #else
PADDLE_ENFORCE(false, "To be implemented."); PADDLE_ENFORCE(false, "To be implemented.");
#endif #endif
......
...@@ -22,61 +22,109 @@ namespace math { ...@@ -22,61 +22,109 @@ namespace math {
template <typename T> template <typename T>
struct CBlas; struct CBlas;
#ifdef PADDLE_WITH_MKLML
template <> template <>
struct CBlas<float> { struct CBlas<float> {
template <typename... ARGS> template <typename... ARGS>
static void GEMM(ARGS... args) { static void GEMM(ARGS... args) {
cblas_sgemm(args...); platform::dynload::cblas_sgemm(args...);
} }
template <typename... ARGS> template <typename... ARGS>
static void AXPY(ARGS... args) { static void AXPY(ARGS... args) {
cblas_saxpy(args...); platform::dynload::cblas_saxpy(args...);
}
template <typename... ARGS>
static void VCOPY(ARGS... args) {
platform::dynload::cblas_scopy(args...);
}
template <typename... ARGS>
static void GEMV(ARGS... args) {
platform::dynload::cblas_sgemv(args...);
}
template <typename... ARGS>
static void GEMM_BATCH(ARGS... args) {
platform::dynload::cblas_sgemm_batch(args...);
} }
#ifdef PADDLE_WITH_MKLML
template <typename... ARGS> template <typename... ARGS>
static void VADD(ARGS... args) { static void VADD(ARGS... args) {
vsAdd(args...); platform::dynload::vsAdd(args...);
}
};
template <>
struct CBlas<double> {
template <typename... ARGS>
static void GEMM(ARGS... args) {
platform::dynload::cblas_dgemm(args...);
}
template <typename... ARGS>
static void AXPY(ARGS... args) {
platform::dynload::cblas_daxpy(args...);
} }
#endif
template <typename... ARGS> template <typename... ARGS>
static void VCOPY(ARGS... args) { static void VCOPY(ARGS... args) {
cblas_scopy(args...); platform::dynload::cblas_dcopy(args...);
} }
template <typename... ARGS> template <typename... ARGS>
static void GEMV(ARGS... args) { static void GEMV(ARGS... args) {
cblas_sgemv(args...); platform::dynload::cblas_dgemv(args...);
} }
#ifdef PADDLE_WITH_MKLML
template <typename... ARGS> template <typename... ARGS>
static void GEMM_BATCH(ARGS... args) { static void GEMM_BATCH(ARGS... args) {
cblas_sgemm_batch(args...); platform::dynload::cblas_dgemm_batch(args...);
}
template <typename... ARGS>
static void VADD(ARGS... args) {
platform::dynload::vdAdd(args...);
} }
#endif
}; };
#else
template <> template <>
struct CBlas<double> { struct CBlas<float> {
template <typename... ARGS> template <typename... ARGS>
static void GEMM(ARGS... args) { static void GEMM(ARGS... args) {
cblas_dgemm(args...); cblas_sgemm(args...);
} }
template <typename... ARGS> template <typename... ARGS>
static void AXPY(ARGS... args) { static void AXPY(ARGS... args) {
cblas_daxpy(args...); cblas_saxpy(args...);
} }
#ifdef PADDLE_WITH_MKLML
template <typename... ARGS> template <typename... ARGS>
static void VADD(ARGS... args) { static void VCOPY(ARGS... args) {
vdAdd(args...); cblas_scopy(args...);
}
template <typename... ARGS>
static void GEMV(ARGS... args) {
cblas_sgemv(args...);
}
};
template <>
struct CBlas<double> {
template <typename... ARGS>
static void GEMM(ARGS... args) {
cblas_dgemm(args...);
}
template <typename... ARGS>
static void AXPY(ARGS... args) {
cblas_daxpy(args...);
} }
#endif
template <typename... ARGS> template <typename... ARGS>
static void VCOPY(ARGS... args) { static void VCOPY(ARGS... args) {
...@@ -87,15 +135,8 @@ struct CBlas<double> { ...@@ -87,15 +135,8 @@ struct CBlas<double> {
static void GEMV(ARGS... args) { static void GEMV(ARGS... args) {
cblas_dgemv(args...); cblas_dgemv(args...);
} }
#ifdef PADDLE_WITH_MKLML
template <typename... ARGS>
static void GEMM_BATCH(ARGS... args) {
cblas_dgemm_batch(args...);
}
#endif
}; };
#endif
template <> template <>
struct CBlas<platform::float16> { struct CBlas<platform::float16> {
static void GEMM(...) { PADDLE_THROW("float16 GEMM not supported on CPU"); } static void GEMM(...) { PADDLE_THROW("float16 GEMM not supported on CPU"); }
......
...@@ -14,9 +14,7 @@ limitations under the License. */ ...@@ -14,9 +14,7 @@ limitations under the License. */
#pragma once #pragma once
#ifdef PADDLE_WITH_MKLML #ifdef PADDLE_WITH_MKLML
#include <mkl_cblas.h> #include "paddle/fluid/platform/dynload/mklml.h"
#include <mkl_lapacke.h>
#include <mkl_vml_functions.h>
#endif #endif
#ifdef PADDLE_USE_OPENBLAS #ifdef PADDLE_USE_OPENBLAS
......
...@@ -12,3 +12,7 @@ if (CUPTI_FOUND) ...@@ -12,3 +12,7 @@ if (CUPTI_FOUND)
endif(CUPTI_FOUND) endif(CUPTI_FOUND)
nv_library(dynload_cuda SRCS ${CUDA_SRCS} DEPS dynamic_loader) nv_library(dynload_cuda SRCS ${CUDA_SRCS} DEPS dynamic_loader)
cc_library(dynload_warpctc SRCS warpctc.cc DEPS dynamic_loader warpctc) cc_library(dynload_warpctc SRCS warpctc.cc DEPS dynamic_loader warpctc)
if (WITH_MKLML)
cc_library(dynload_mklml SRCS mklml.cc DEPS dynamic_loader mklml)
endif()
# TODO(TJ): add iomp, mkldnn?
...@@ -49,6 +49,8 @@ DEFINE_string( ...@@ -49,6 +49,8 @@ DEFINE_string(
tensorrt_dir, "", tensorrt_dir, "",
"Specify path for loading tensorrt library, such as libnvinfer.so."); "Specify path for loading tensorrt library, such as libnvinfer.so.");
DEFINE_string(mklml_dir, "", "Specify path for loading libmklml_intel.so.");
namespace paddle { namespace paddle {
namespace platform { namespace platform {
namespace dynload { namespace dynload {
...@@ -206,6 +208,14 @@ void* GetTensorRtDsoHandle() { ...@@ -206,6 +208,14 @@ void* GetTensorRtDsoHandle() {
#endif #endif
} }
void* GetMKLMLDsoHandle() {
#if defined(__APPLE__) || defined(__OSX__)
return GetDsoHandleFromSearchPath(FLAGS_mklml_dir, "libmklml_intel.dylib");
#else
return GetDsoHandleFromSearchPath(FLAGS_mklml_dir, "libmklml_intel.so");
#endif
}
} // namespace dynload } // namespace dynload
} // namespace platform } // namespace platform
} // namespace paddle } // namespace paddle
...@@ -26,6 +26,7 @@ void* GetWarpCTCDsoHandle(); ...@@ -26,6 +26,7 @@ void* GetWarpCTCDsoHandle();
void* GetLapackDsoHandle(); void* GetLapackDsoHandle();
void* GetNCCLDsoHandle(); void* GetNCCLDsoHandle();
void* GetTensorRtDsoHandle(); void* GetTensorRtDsoHandle();
void* GetMKLMLDsoHandle();
} // namespace dynload } // namespace dynload
} // namespace platform } // namespace platform
......
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/platform/dynload/mklml.h"
namespace paddle {
namespace platform {
namespace dynload {
std::once_flag mklml_dso_flag;
void* mklml_dso_handle = nullptr;
#define DEFINE_WRAP(__name) DynLoad__##__name __name
MKLML_ROUTINE_EACH(DEFINE_WRAP);
} // namespace dynload
} // namespace platform
} // namespace paddle
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <dlfcn.h>
#include <mkl.h>
#include <mutex> // NOLINT
#include "paddle/fluid/platform/dynload/dynamic_loader.h"
namespace paddle {
namespace platform {
namespace dynload {
extern std::once_flag mklml_dso_flag;
extern void* mklml_dso_handle;
/**
* The following macro definition can generate structs
* (for each function) to dynamic load mklml routine
* via operator overloading.
*/
#define DYNAMIC_LOAD_MKLML_WRAP(__name) \
struct DynLoad__##__name { \
template <typename... Args> \
auto operator()(Args... args) -> decltype(__name(args...)) { \
using mklmlFunc = decltype(&::__name); \
std::call_once(mklml_dso_flag, []() { \
mklml_dso_handle = paddle::platform::dynload::GetMKLMLDsoHandle(); \
}); \
static void* p_##_name = dlsym(mklml_dso_handle, #__name); \
return reinterpret_cast<mklmlFunc>(p_##_name)(args...); \
} \
}; \
extern DynLoad__##__name __name
#define DECLARE_DYNAMIC_LOAD_MKLML_WRAP(__name) DYNAMIC_LOAD_MKLML_WRAP(__name)
#define MKLML_ROUTINE_EACH(__macro) \
__macro(cblas_sgemm); \
__macro(cblas_saxpy); \
__macro(cblas_scopy); \
__macro(cblas_sgemv); \
__macro(cblas_sgemm_batch); \
__macro(cblas_dgemm); \
__macro(cblas_daxpy); \
__macro(cblas_dcopy); \
__macro(cblas_dgemv); \
__macro(cblas_dgemm_batch); \
__macro(vsAdd); \
__macro(vdAdd); \
__macro(MKL_Set_Num_Threads)
MKLML_ROUTINE_EACH(DECLARE_DYNAMIC_LOAD_MKLML_WRAP);
#undef DYNAMIC_LOAD_MKLML_WRAP
} // namespace dynload
} // namespace platform
} // namespace paddle
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册