Merge remote-tracking branch 'upstream/develop' into develop

README.md

@@ -26,8 +26,15 @@ Paddle-Moible是PaddlePaddle组织下的项目，是一个致力于嵌入式平
 
 - **ARM CPU**
 
-
-![](http://mms-graph.bj.bcebos.com/paddle-mobile%2F2018_07_29.png)
+|mobilenet arm v7|1线程|2线程|4线程|
+|------------|----|-----|-----|
+|麒麟960(ms)|110.586|72.474|49.833|
+|||||
+|mobilenetssd arm v7|1线程|2线程|4线程|
+|麒麟960(ms)|224.464|142.544|96.068|
+|||||
+|googlenet(v1) arm v7|1线程|2线程|4线程|
+|麒麟960(ms)|348.018|242.689|169.998|
 
     arm cpu是paddle-mobile的主要支持方向，cpu的通用性一直是其优势。嵌入式深度学习，需要大量的cpu汇编实现。我们正在紧锣密鼓的编码，为的是能充分硬件的每一点加速能力。
     arm cpu的优化工作还在进行中，现在使用了常规的cpu优化。在arm a73上paddle-mobile arm-v7现在单核运行一次mobilenet1.0是110+ms，显然这不是我们的最终目标，我们正在用大量的汇编改写，后续性能仍会有巨大提升空间, 目前只支持armv7, 未来我们也会支持armv8。

src/io/api.cc

@@ -12,6 +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. */
 
+#include "cstring"
 #include "io/paddle_inference_api.h"
 
 namespace paddle_mobile {

src/operators/kernel/fpga/elementwise_add_relu_kernel.cpp

@@ -25,9 +25,9 @@ bool ElementwiseAddReluKernel<FPGA, float>::Init(
   const Tensor *input_x = param->InputX();
   const Tensor *input_y = param->InputY();
   Tensor *out = param->Out();
-  auto input_x_ptr = input_x->data<float>();
-  auto input_y_ptr = input_y->data<float>();
-  auto out_ptr = out->mutable_data<float>();
+  auto input_x_ptr = input_x->data<half>();
+  auto input_y_ptr = input_y->data<half>();
+  auto out_ptr = out->mutable_data<half>();
 
   fpga::EWAddArgs ewaddArgs;
   ewaddArgs.relu_enabled = relu_enabled;

src/operators/kernel/fpga/fc_relu_kernel.cpp

@@ -22,13 +22,13 @@ template <>
 bool FusionFcReluKernel<FPGA, float>::Init(FusionFcReluParam *param) {
   bool relu_enabled = true;
   const Tensor *input_x = param->InputX();
-  auto input_x_ptr = input_x->data<float>();
+  auto input_x_ptr = input_x->data<half>();
   const Tensor *input_y = param->InputY();
   auto input_y_ptr = input_y->data<float>();
   const Tensor *input_z = param->InputZ();
   auto input_z_ptr = input_z->data<float>();
   Tensor *out = param->Out();
-  auto out_ptr = out->mutable_data<float>();
+  auto out_ptr = out->mutable_data<half>();
 
   PADDLE_MOBILE_ENFORCE(input_x->dims()[1] == input_y->dims()[0],
                         "Image channel should be equal to weight number");

src/operators/kernel/fpga/fusion_fc_kernel.cpp

@@ -22,13 +22,13 @@ template <>
 bool FusionFcKernel<FPGA, float>::Init(FusionFcParam *param) {
   bool relu_enabled = false;
   const Tensor *input_x = param->InputX();
-  auto input_x_ptr = input_x->data<float>();
+  auto input_x_ptr = input_x->data<half>();
   const Tensor *input_y = param->InputY();
   auto input_y_ptr = input_y->data<float>();
   const Tensor *input_z = param->InputZ();
   auto input_z_ptr = input_z->data<float>();
   Tensor *out = param->Out();
-  auto out_ptr = out->mutable_data<float>();
+  auto out_ptr = out->mutable_data<half>();
 
   PADDLE_MOBILE_ENFORCE(input_x->dims()[1] == input_y->dims()[0],
                         "Image channel should be equal to weight number");

src/operators/kernel/fpga/pool_kernel.cpp

@@ -22,9 +22,9 @@ namespace operators {
 template <>
 bool PoolKernel<FPGA, float>::Init(PoolParam *param) {
   const Tensor *input = param->Input();
-  auto input_ptr = input->data<float>();
+  auto input_ptr = input->data<half>();
   Tensor *output = param->Output();
-  auto output_ptr = output->mutable_data<float>();
+  auto output_ptr = output->mutable_data<half>();
   vector<int> ksize = param->Ksize();
   vector<int> strides = param->Strides();
   vector<int> paddings = param->Paddings();

src/operators/math/gemm.h

@@ -50,6 +50,10 @@ void PackMatrixA_6r(int m, int k, int m_tail, const float *A, int lda,
                     float *buffer);
 void PackMatrixA_8r(int m, int k, int m_tail, const float *A, int lda,
                     float *buffer);
+void PackMatrixA_omp_6r(int m, int k, int m_tail, const float *A, int lda,
+                        float *buffer);
+void PackMatrixA_omp_8r(int m, int k, int m_tail, const float *A, int lda,
+                        float *buffer);
 
 // 将 B 矩阵分块复制到连续内存(RowMajor)
 void PackMatrixB_8c(int k, int n, int n_tail, const float *B, int ldb,
@@ -58,6 +62,12 @@ void PackMatrixB_12c(int k, int n, int n_tail, const float *B, int ldb,
                      float *buffer);
 void PackMatrixB_16c(int k, int n, int n_tail, const float *B, int ldb,
                      float *buffer);
+void PackMatrixB_omp_8c(int k, int n, int n_tail, const float *B, int ldb,
+                        float *buffer);
+void PackMatrixB_omp_12c(int k, int n, int n_tail, const float *B, int ldb,
+                         float *buffer);
+void PackMatrixB_omp_16c(int k, int n, int n_tail, const float *B, int ldb,
+                         float *buffer);
 
 // 分块矩阵乘法
 void InnerKernel(int mc, int nc, float alpha, const float *a, const float *b,
@@ -136,6 +146,16 @@ void SgemmWithBn(int m, int n, int k, float alpha, const float *A, int lda,
                  const float *B, int ldb, float beta, float *C, int ldc,
                  bool relu, float *new_scale, float *new_bias);
 
+// 32位 float 矩阵乘法（openmp 多线程版本）
+void Sgemm_omp(int m, int n, int k, float alpha, const float *A, int lda,
+               const float *B, int ldb, float beta, float *C, int ldc,
+               bool relu, float *bias);
+
+// 32位 float 矩阵乘法, 并对结果进行 batchnrom（openmp 多线程版本）
+void SgemmWithBn_omp(int m, int n, int k, float alpha, const float *A, int lda,
+                     const float *B, int ldb, float beta, float *C, int ldc,
+                     bool relu, float *new_scale, float *new_bias);
+
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/math/math_function.cpp

@@ -42,8 +42,13 @@ void matmul<float>(const framework::Tensor &matrix_a, bool trans_a,
   int N = dim_out[1];
   int K = (!trans_a) ? dim_a[1] : dim_a[0];
 
+#ifdef _OPENMP
+  Sgemm_omp(M, N, K, alpha, matrix_a.data<float>(), K, matrix_b.data<float>(),
+            N, beta, matrix_out->data<float>(), N, relu, bias);
+#else
   Sgemm(M, N, K, alpha, matrix_a.data<float>(), K, matrix_b.data<float>(), N,
         beta, matrix_out->data<float>(), N, relu, bias);
+#endif
 }
 
 template <>
@@ -70,10 +75,17 @@ void matmulWithBn<float>(const framework::Tensor &matrix_a, bool trans_a,
   int N = dim_out[1];
   int K = (!trans_a) ? dim_a[1] : dim_a[0];
 
+#ifdef _OPENMP
+  SgemmWithBn_omp(M, N, K, alpha, matrix_a.data<float>(), K,
+                  matrix_b.data<float>(), N, beta, matrix_out->data<float>(), N,
+                  relu, new_scale->data<float>() + group,
+                  new_bias->data<float>() + group);
+#else
   SgemmWithBn(M, N, K, alpha, matrix_a.data<float>(), K, matrix_b.data<float>(),
               N, beta, matrix_out->data<float>(), N, relu,
               new_scale->data<float>() + group,
               new_bias->data<float>() + group);
+#endif
 }
 
 }  // namespace math