Merge pull request #572 from cocodark/develop

add interface to dynamic set omp thread num

src/io/executor.cpp

@@ -26,9 +26,6 @@ limitations under the License. */
 #include "framework/program/var_desc.h"
 #include "framework/scope.h"
 #include "framework/tensor.h"
-#ifdef _OPENMP
-#include <omp.h>
-#endif  // _OPENMP
 #ifdef PADDLE_EXECUTOR_MULTITHREAD
 #include <queue>
 #include <utility>
@@ -407,14 +404,6 @@ std::vector<typename Executor<Dtype, P>::Ptype> Executor<Dtype, P>::Predict(
   return result_vector;
 }
 
-template <typename Dtype, Precision P>
-void Executor<Dtype, P>::SetThreadNum(int num) {
-#ifdef _OPENMP
-  //  omp_set_dynamic(0);
-  omp_set_num_threads(num);
-#endif
-}
-
 template class Executor<CPU, Precision::FP32>;
 template class Executor<FPGA, Precision::FP32>;
 template class Executor<GPU_MALI, Precision::FP32>;

src/io/executor.h

@@ -58,8 +58,6 @@ class Executor {
   std::vector<Ptype> Predict(const std::vector<Ptype> &input,
                              const std::vector<int64_t> &dims);
 
-  void SetThreadNum(int num);
-
  protected:
   Executor() = default;
   void InitMemory();

src/io/paddle_mobile.cpp

@@ -16,6 +16,14 @@ limitations under the License. */
 
 namespace paddle_mobile {
 
+template <typename Dtype, Precision P>
+void PaddleMobile<Dtype, P>::SetThreadNum(int num) {
+#ifdef _OPENMP
+  //  omp_set_dynamic(0);
+  omp_set_num_threads(num);
+#endif
+};
+
 template <typename Dtype, Precision P>
 bool PaddleMobile<Dtype, P>::Load(const std::string &dirname, bool optimize,
                                   int batch_size) {
@@ -81,7 +89,9 @@ PaddleMobile<Dtype, P>::~PaddleMobile() {
 }
 
 template class PaddleMobile<CPU, Precision::FP32>;
+
 template class PaddleMobile<FPGA, Precision::FP32>;
+
 template class PaddleMobile<GPU_MALI, Precision::FP32>;
 
 }  // namespace paddle_mobile

src/io/paddle_mobile.h

@@ -17,6 +17,9 @@ limitations under the License. */
 #include <memory>
 #include <string>
 #include <vector>
+#ifdef _OPENMP
+#include <omp.h>
+#endif  // _OPENMP
 
 #include "common/types.h"
 #include "framework/tensor.h"
@@ -44,6 +47,7 @@ class PaddleMobile {
    * */
   bool Load(const std::string &model_path, const std::string &para_path,
             bool optimize = false, int batch_size = 1);
+  void SetThreadNum(int num);
 
   /*
    * @b to predict

test/net/test_googlenet.cpp

@@ -17,26 +17,21 @@ limitations under the License. */
 #include "../test_include.h"
 
 int main() {
-  paddle_mobile::Loader<paddle_mobile::CPU> loader;
+  paddle_mobile::PaddleMobile<paddle_mobile::CPU> paddle_mobile;
+  paddle_mobile.SetThreadNum(4);
   bool optimize = true;
   auto time1 = time();
-  //  auto program = loader.Load(g_googlenet, optimize);
-  auto program = loader.Load(g_googlenet_combine + "/model",
-                             g_googlenet_combine + "/params", optimize);
+  if (paddle_mobile.Load(g_googlenet, optimize)) {
     auto time2 = time();
-  DLOG << "load cost :" << time_diff(time1, time2) << "ms\n";
-  paddle_mobile::Executor<paddle_mobile::CPU> executor(program, 1, optimize);
-  executor.SetThreadNum(4);
+    DLOG << "load cost :" << time_diff(time1, time1) << "ms";
     std::vector<float> input;
     std::vector<int64_t> dims{1, 3, 224, 224};
     GetInput<float>(g_test_image_1x3x224x224, &input, dims);
     auto time3 = time();
-  int count = 1;
-  for (int i = 0; i < count; ++i) {
-    executor.Predict(input, dims);
-  }
-
+    auto vec_result = paddle_mobile.Predict(input, dims);
     auto time4 = time();
-  DLOG << "predict cost :" << time_diff(time3, time4) / count << "ms\n";
+
+    DLOG << "predict cost :" << time_diff(time3, time4) << "ms";
+  }
   return 0;
 }