Merge branch 'develop' into develop

src/fpga/fpga_quantilization.cpp

@@ -46,8 +46,8 @@ static Dtype find_max(Dtype* data, int num) {
   return max;
 }
 
-template <typename Dtype>
-framework::Tensor* quantilize_filter(framework::Tensor* filter) {
+// template <typename Dtype>
+framework::Tensor* quantify_filter(framework::Tensor* filter) {
   float scale = 0;
   float fix_range = static_cast<float>((1 << (8 - 1)) - 1);
 
@@ -57,7 +57,7 @@ framework::Tensor* quantilize_filter(framework::Tensor* filter) {
   const int width = filter->dims()[3];
 
   int8_t* int_data = nullptr;
-  int8_t* tmp_data = new int[filter->numel()];
+  int8_t* tmp_data = new int8_t[filter->numel()];
 
   // 32bit filter -> 8bit filter;
   if (filter->type() == typeid(float)) {

src/fpga/fpga_quantilization.h

@@ -24,7 +24,7 @@ template <typename Dtype>
 static void chw_to_hwc(Dtype* data_in, Dtype* data_out, int num, int channel,
                        int height, int width);
 
-template <typename Dtype>
-framework::Tensor* quantilize_filter(framework::Tensor* filter);
+// template <typename Dtype>
+framework::Tensor* quantify_filter(framework::Tensor* filter);
 }  // namespace fpga
 }  // namespace paddle_mobile

src/operators/kernel/fpga/conv_add_bn_kernel.cpp

@@ -28,7 +28,7 @@ bool ConvAddBNKernel<FPGA, float>::Init(FusionConvAddBNParam *param) {
   auto input_ptr = input->data<half>();
   const Tensor *bias = param->Bias();
   auto bias_ptr = bias->data<float>();
-  const Tensor *filter = param->Filter();
+  Tensor *filter = param->Filter();
 
   Tensor *out = param->Output();
   auto out_ptr = out->mutable_data<half>();
@@ -60,7 +60,7 @@ bool ConvAddBNKernel<FPGA, float>::Init(FusionConvAddBNParam *param) {
   param->SetNewScale(new_scale);
   param->SetNewBias(new_bias);
 
-  const Tensor *quant_filter = quantilize_filter(filter);
+  Tensor *quant_filter = fpga::quantify_filter(filter);
 
   // delete original filter?
   filter = quant_filter;
@@ -68,22 +68,22 @@ bool ConvAddBNKernel<FPGA, float>::Init(FusionConvAddBNParam *param) {
   auto filter_ptr = filter->data<float>();
   fpga::ConvArgs convArgs;
   convArgs.relu_enabled = relu_enabled;
-  convArgs.filter_address = reinterpret_cast<void *> filter_ptr;
+  convArgs.filter_address = (void *)filter_ptr;
   convArgs.filter_num = filter->dims()[0];
   convArgs.group_num = param->Groups();
-  convArgs.sb_address = reinterpret_cast<void *> bs_ptr;
+  convArgs.sb_address = (void *)bs_ptr;
   convArgs.kernel.stride_h = param->Strides()[0];
   convArgs.kernel.stride_w = param->Strides()[1];
   convArgs.kernel.height = filter->dims()[2];
   convArgs.kernel.width = filter->dims()[3];
-  convArgs.image.address = reinterpret_cast<void *> input_ptr;
+  convArgs.image.address = (void *)input_ptr;
   convArgs.image.channels = input->dims()[1];
   convArgs.image.height = input->dims()[2];
   convArgs.image.width = input->dims()[3];
   convArgs.image.pad_height = param->Paddings()[0];
   convArgs.image.pad_width = param->Paddings()[1];
   convArgs.image.scale_address = input->fpga_args().scale_pointer();
-  convArgs.output.address = reinterpret_cast<void *> out_ptr;
+  convArgs.output.address = (void *)out_ptr;
   convArgs.output.scale_address = out->fpga_args().scale_pointer();
   param->SetFpgaArgs(convArgs);
 

src/operators/op_param.h

@@ -920,7 +920,11 @@ class FusionFcParam : public OpParam {
   }
   const Tensor *InputX() const { return input_x_; }
 
+#ifdef PADDLE_MOBILE_FPGA
+  Tensor *InputY() const { return input_y_; }
+#else
   const Tensor *InputY() const { return input_y_; }
+#endif
 
   const Tensor *InputZ() const { return input_z_; }
 
@@ -976,7 +980,11 @@ class FusionConvAddParam : public OpParam {
 
   const Tensor *Input() const { return input_; }
 
+#ifdef PADDLE_MOBILE_FPGA
+  Tensor *Filter() const { return filter_; }
+#else
   const Tensor *Filter() const { return filter_; }
+#endif
 
   Tensor *Output() const { return output_; }
 
@@ -1050,7 +1058,11 @@ class FusionConvAddBNReluParam : public OpParam {
 
   const Tensor *Input() const { return input_; }
 
+#ifdef PADDLE_MOBILE_FPGA
+  Tensor *Filter() const { return filter_; }
+#else
   const Tensor *Filter() const { return filter_; }
+#endif
 
   Tensor *Output() const { return output_; }
 
@@ -1144,8 +1156,11 @@ class FusionConvAddBNParam : public OpParam {
 
   const Tensor *Input() const { return input_; }
 
+#ifdef PADDLE_MOBILE_FPGA
+  Tensor *Filter() const { return filter_; }
+#else
   const Tensor *Filter() const { return filter_; }
-
+#endif
   Tensor *Output() const { return output_y_; }
 
   const vector<int> &Strides() const { return strides_; }