Merge pull request #717 from qnqinan/develop

update added FPGA kernels to support quantization #closed 716

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();