Merge pull request #10789 from Xreki/core_fix_openblas_threads

Add an interface to set the number of threads for math function, and set the default value to 1 for inference.

cmake/external/openblas.cmake

@@ -29,6 +29,8 @@ IF(NOT ${CBLAS_FOUND})
         "${CBLAS_INSTALL_DIR}/lib/${CMAKE_STATIC_LIBRARY_PREFIX}openblas${CMAKE_STATIC_LIBRARY_SUFFIX}"
         CACHE FILEPATH "openblas library." FORCE)
 
+    ADD_DEFINITIONS(-DPADDLE_USE_OPENBLAS)
+
     SET(OPENBLAS_CC "${CMAKE_C_COMPILER} -Wno-unused-but-set-variable -Wno-unused-variable")
     SET(OPENBLAS_COMMIT "v0.2.20")
 

paddle/fluid/inference/io.cc

@@ -20,16 +20,20 @@ limitations under the License. */
 #include "paddle/fluid/framework/block_desc.h"
 #include "paddle/fluid/framework/feed_fetch_type.h"
 #include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/math/blas.h"
 #include "paddle/fluid/pybind/pybind.h"
 
 DEFINE_string(devices, "", "The devices to be used which is joined by comma.");
 DEFINE_bool(init_p2p, false, "Whether to init p2p.");
+DEFINE_int32(math_num_threads, 1,
+             "Number of threads used to run math functions.");
 
 namespace paddle {
 namespace inference {
 
 void Init(const std::vector<std::string> argv) {
   framework::InitGflags(argv);
+  operators::math::SetNumThreads(FLAGS_math_num_threads);
   // init devices
   std::vector<int> devices;
   std::string token;

paddle/fluid/operators/math/blas.h

@@ -20,13 +20,16 @@
 #ifdef PADDLE_WITH_MKLML
 #include <mkl_cblas.h>
 #include <mkl_lapacke.h>
+#include <mkl_service.h>
 #include <mkl_vml_functions.h>
 #endif
 
 #ifdef PADDLE_USE_OPENBLAS
 #include <cblas.h>
+#ifdef LAPACK_FOUND
 #include <lapacke.h>
 #endif
+#endif
 
 #ifndef LAPACK_FOUND
 extern "C" {
@@ -46,6 +49,18 @@ namespace paddle {
 namespace operators {
 namespace math {
 
+static void SetNumThreads(int num_threads) {
+#ifdef PADDLE_USE_OPENBLAS
+  int real_num_threads = num_threads > 1 ? num_threads : 1;
+  openblas_set_num_threads(real_num_threads);
+#elif defined(PADDLE_WITH_MKLML)
+  int real_num_threads = num_threads > 1 ? num_threads : 1;
+  mkl_set_num_threads(real_num_threads);
+#else
+  PADDLE_ENFORCE(false, "To be implemented.");
+#endif
+}
+
 /**
  * Matrix Descriptor of a memory buffer.
  *

paddle/fluid/operators/math/math_function.h

@@ -21,8 +21,10 @@ limitations under the License. */
 
 #ifdef PADDLE_USE_OPENBLAS
 #include <cblas.h>
+#ifdef LAPACK_FOUND
 #include <lapacke.h>
 #endif
+#endif
 
 #ifndef LAPACK_FOUND
 extern "C" {

paddle/math/MathFunctions.h

@@ -12,8 +12,7 @@ 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. */
 
-#ifndef MATHFUNCTIONS_H_
-#define MATHFUNCTIONS_H_
+#pragma once
 
 #ifdef PADDLE_WITH_MKLML
 #include <mkl_cblas.h>
@@ -21,7 +20,7 @@ limitations under the License. */
 #include <mkl_vml_functions.h>
 #endif
 
-#if defined(PADDLE_USE_VECLIB)
+#ifdef PADDLE_USE_VECLIB
 extern "C" {
 #include <cblas.h>
 #include <clapack.h>
@@ -30,8 +29,10 @@ extern "C" {
 
 #ifdef PADDLE_USE_OPENBLAS
 #include <cblas.h>
+#ifdef LAPACK_FOUND
 #include <lapacke.h>
 #endif
+#endif
 
 #ifndef LAPACK_FOUND
 extern "C" {
@@ -126,5 +127,3 @@ template <class T>
 void vTanh(const int n, const T* a, T* r);
 
 }  // namespace paddle
-
-#endif  // MATHFUNCTIONS_H_