Merge branch 'develop' into develop

src/common/enforce.h

@@ -61,7 +61,14 @@ struct PaddleMobileException : public std::exception {
   }
 #else
 #define PADDLE_MOBILE_THROW_EXCEPTION(...)
-#define PADDLE_MOBILE_ENFORCE(stat, ...)
+
+#define PADDLE_MOBILE_ENFORCE(stat, ...) \
+  {                                      \
+    if (stat) {                          \
+    } else {                             \
+    }                                    \
+  }
+
 #endif
 
 }  // namespace paddle_mobile

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/feed_op.h

@@ -74,4 +74,5 @@ USE_OP_CPU(feed);
 USE_OP_MALI_GPU(feed);
 #endif
 #ifdef PADDLE_MOBILE_FPGA
+USE_OP_FPGA(feed);
 #endif

src/operators/fetch_op.cpp

@@ -25,4 +25,5 @@ REGISTER_OPERATOR_CPU(fetch, ops::FetchOp);
 REGISTER_OPERATOR_MALI_GPU(fetch, ops::FetchOp);
 #endif
 #ifdef PADDLE_MOBILE_FPGA
+REGISTER_OPERATOR_FPGA(fetch, ops::FetchOp);
 #endif

src/operators/fetch_op.h

@@ -54,4 +54,5 @@ USE_OP_CPU(fetch);
 USE_OP_MALI_GPU(fetch);
 #endif
 #ifdef PADDLE_MOBILE_FPGA
+USE_OP_FPGA(fetch);
 #endif

src/operators/kernel/central-arm-func/conv_add_relu_arm_func.h

@@ -32,12 +32,7 @@ void ConvAddReluCompute(const FusionConvAddReluParam &param) {
   Tensor bias = *param.Bias();
   int axis = param.Axis();
   Tensor *output = param.Output();
-  //  math::expand_bias(bias, axis, output->dims());
-  float *output_data = output->data<float>();
   float *biase_data = bias.data<float>();
-  //  for (int k = 0; k < output->numel(); ++k) {
-  //    output_data[k] = biase_data[k];
-  //  }
 
   int groups = param.Groups();
   std::vector<int> strides = param.Strides();

src/operators/kernel/central-arm-func/conv_arm_func.h

@@ -30,7 +30,6 @@ inline void ConvBasic(const ConvParam &param) {
   Tensor filter = *param.Filter();
   Tensor *output = param.Output();
   output->mutable_data<float>();
-  float *bias_data = output->mutable_data<float>();
   int groups = param.Groups();
   std::vector<int> strides = param.Strides();
   std::vector<int> paddings = param.Paddings();
@@ -107,7 +106,7 @@ inline void ConvBasic(const ConvParam &param) {
       Tensor filter_slice = filter.Slice(g * out_step, (g + 1) * out_step);
       math::matmul<float>(filter_slice, false, col_matrix, false,
                           static_cast<float>(1), &out_slice,
-                          static_cast<float>(0), false, bias_data);
+                          static_cast<float>(0));
     }
   }
 }

src/operators/kernel/central-arm-func/mul_arm_func.h

@@ -59,7 +59,6 @@ void MulCompute(const MulParam &param) {
   const Tensor *input_y = param.InputY();
   Tensor *out = param.Out();
   out->mutable_data<float>();
-  float *bias_data = out->mutable_data<float>();
   const Tensor x_matrix =
       input_x->dims().size() > 2
           ? framework::ReshapeToMatrix(*input_x, param.XNumColDims())
@@ -73,7 +72,7 @@ void MulCompute(const MulParam &param) {
     out->Resize({x_matrix.dims()[0], y_matrix.dims()[1]});
   }
   math::matmul<float>(x_matrix, false, y_matrix, false, static_cast<float>(1),
-                      out, static_cast<float>(0), false, bias_data);
+                      out, static_cast<float>(0));
   if (out_dim.size() != 2) {
     out->Resize(out_dim);
   }

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.cpp

@@ -373,9 +373,9 @@ void InnerKernel(int mc, int nc, float alpha, const float *a, const float *b,
 #endif
     }
   }
+
   if (alpha != 1) {
     WriteWithAlphaBeta(mc, nc, c, C, ldc);
-
     return;
   }
   if (beta == 0) {
@@ -2244,6 +2244,27 @@ void AddDot4x4(int k, const float *a, const float *b, float *c, int ldc) {
   }
 }
 
+void AddDot4x8(int k, const float *a, const float *b, float *c, int ldc) {}
+
+void WriteBasic(int mc, int nc, float *c, float *C, int ldc) {}
+
+void WriteWithAlphaBeta(int mc, int nc, float *c, float *C, int ldc) {}
+
+void WriteWithAdd(int mc, int nc, float *c, float *C, int ldc) {}
+
+void WriteWithAddV1(int mc, int nc, float *c, float *C, int ldc, float *bias) {}
+
+void WriteWithAddRelu(int mc, int nc, float *c, float *C, int ldc) {}
+
+void WriteWithAddReluV1(int mc, int nc, float *c, float *C, int ldc,
+                        float *bias) {}
+
+void WriteWithBn(int mc, int nc, float *c, float *C, int ldc, float *new_scale,
+                 float *new_bias) {}
+
+void WriteWithBnRelu(int mc, int nc, float *c, float *C, int ldc,
+                     float *new_scale, float *new_bias) {}
+
 #endif  // __ARM_NEON
 
 // 32位 float 矩阵乘法