Merge pull request #553 from cocodark/develop

accelerate with openmp

CMakeLists.txt

@@ -2,7 +2,7 @@ cmake_minimum_required(VERSION 3.0)
 project(paddle-mobile)
 
 option(DEBUGING "enable debug mode" ON)
-option(USE_OPENMP "openmp support" OFF)
+option(USE_OPENMP "openmp support" ON)
 option(USE_EXCEPTION "use std exception" ON)
 option(LOG_PROFILE "log profile" ON)
 # select the platform to build

src/io/executor.cpp

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "io/executor.h"
+#include <operators/math/gemm.h>
 #include <algorithm>
 #include <vector>
 #include "common/enforce.h"
@@ -25,6 +26,9 @@ 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>
@@ -403,6 +407,14 @@ 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

@@ -58,6 +58,8 @@ 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/operators/kernel/lrn_kernel.h

@@ -13,7 +13,9 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #ifdef LRN_OP
-
+#ifdef _OPENMP
+#include <omp.h>
+#endif
 #include "framework/operator.h"
 #include "operators/op_param.h"
 
@@ -47,6 +49,7 @@ struct LRNFunctor {
     std::fill(sqr_buffer_ptr, sqr_buffer_ptr + sqr_buffer.numel(), 0.0);
 
     for (int a = 0; a < N; a++) {
+#pragma parallel for
       for (int b = 0; b < C; b++) {
         for (int index = start; index < end; index++) {
           int channel = b + index;

src/operators/math/pool_3x3.h

@@ -15,10 +15,13 @@ limitations under the License. */
 #ifdef POOL_OP
 
 #pragma once
+#ifdef _OPENMP
+#include <omp.h>
+#endif
 #include <algorithm>
 #include <vector>
 #include "framework/tensor.h"
-#ifdef __ARM_NEON
+#if __ARM_NEON
 #include <arm_neon.h>
 #endif  // __ARM_NEON
 

src/operators/math/pooling.cpp

@@ -14,10 +14,11 @@ limitations under the License. */
 
 #ifdef POOL_OP
 
-#include "operators/math/pooling.h"
-#include <algorithm>
-#include <vector>
+#include "pooling.h"
 #include "common/types.h"
+#ifdef _OPENMP
+#include <omp.h>
+#endif
 
 namespace paddle_mobile {
 namespace operators {
@@ -59,7 +60,7 @@ class PoolFunctor<CPU, PoolProcess, T> {
     T *output_data = output->mutable_data<T>();
 
     for (int i = 0; i < batch_size; i++) {
-      //  #pragma omp parallel for
+#pragma omp parallel for
       for (int c = 0; c < output_channels; ++c) {
         for (int ph = 0; ph < output_height; ++ph) {
           int hstart = ph * stride_height - padding_height;

test/net/test_googlenet.cpp

@@ -26,16 +26,17 @@ int main() {
   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);
   std::vector<float> input;
   std::vector<int64_t> dims{1, 3, 224, 224};
   GetInput<float>(g_test_image_1x3x224x224, &input, dims);
   auto time3 = time();
-
-  for (int i = 0; i < 10; ++i) {
+  int count = 1;
+  for (int i = 0; i < count; ++i) {
     executor.Predict(input, dims);
   }
 
   auto time4 = time();
-  DLOG << "predict cost :" << time_diff(time3, time4) << "ms\n";
+  DLOG << "predict cost :" << time_diff(time3, time4) / count << "ms\n";
   return 0;
 }