Merge branch 'develop' into develop

src/operators/feed_op.h

@@ -30,9 +30,6 @@ class FeedOp : public framework::OperatorBase<DeviceType> {
       : framework::OperatorBase<DeviceType>(type, inputs, outputs, attrs,
                                             scope),
         param_(inputs, outputs, attrs, scope.get()) {}
-  void RunImpl() const { param_.Out()->ShareDataWith(*param_.InputX()); }
-
-  void Init() {}
 
   void InferShape() const {
     auto out_dims = param_.Out()->dims();
@@ -40,6 +37,29 @@ class FeedOp : public framework::OperatorBase<DeviceType> {
     param_.Out()->Resize(out_dims);
   }
 
+#ifdef PADDLE_MOBILE_FPGA
+  void RunImpl() const { fpga::PerformBypass(param_.FpgaArgs()); }
+  void Init() {
+    const Tensor *input = param_.InputX();
+    auto input_ptr = input->data<float>();
+    Tensor *output = param_.Out();
+    auto output_ptr = output->mutable_data<half>();
+    fpga::BypassArgs args;
+    args.convert_type = fpga::DATA_FP32_TO_FP16;
+    args.layout_type = fpga::LAYOUT_CHW_TO_HWC;
+    args.image.address = (void *)input_ptr;
+    args.image.channels = input->dims()[1];
+    args.image.height = input->dims()[2];
+    args.image.width = input->dims()[3];
+    args.output.address = output_ptr;
+    param_.SetFpgaArgs(args);
+  }
+
+#else
+  void RunImpl() const { param_.Out()->ShareDataWith(*param_.InputX()); }
+  void Init() {}
+#endif
+
  protected:
   FeedParam param_;
 };

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

@@ -16,6 +16,8 @@ limitations under the License. */
 #pragma once
 
 #include <cmath>
+#include "framework/tensor.h"
+#include "operators/op_param.h"
 
 namespace paddle_mobile {
 namespace operators {
@@ -122,7 +124,7 @@ void BoxCoderCompute(const BoxCoderParam& param) {
   auto col = input_priorbox->dims()[0];
   auto len = input_priorbox->dims()[1];
 
-  Tensor* output_box = param.OutputBox();
+  framework::Tensor* output_box = param.OutputBox();
   auto* output_box_dataptr = output_box->mutable_data<float>({row, col, len});
 
   if (code_type == "encode_center_size") {

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

@@ -18,6 +18,9 @@ limitations under the License. */
 
 #include <vector>
 #include "operators/math/depthwise_conv_3x3.h"
+#include "operators/math/im2col.h"
+#include "operators/math/math_function.h"
+#include "operators/math/vol2col.h"
 #include "operators/op_param.h"
 
 namespace paddle_mobile {

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

@@ -16,6 +16,10 @@ limitations under the License. */
 
 #pragma once
 #include <vector>
+#include "operators/math/conv_func.h"
+#include "operators/math/im2col.h"
+#include "operators/math/math_function.h"
+#include "operators/math/vol2col.h"
 #include "operators/op_param.h"
 
 namespace paddle_mobile {

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

@@ -17,6 +17,9 @@ limitations under the License. */
 #pragma once
 #include <vector>
 #include "operators/math/depthwise_conv_3x3.h"
+#include "operators/math/im2col.h"
+#include "operators/math/math_function.h"
+#include "operators/math/vol2col.h"
 #include "operators/op_param.h"
 namespace paddle_mobile {
 namespace operators {

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

@@ -17,6 +17,9 @@ limitations under the License. */
 #pragma once
 #include <vector>
 #include "operators/math/depthwise_conv_3x3.h"
+#include "operators/math/im2col.h"
+#include "operators/math/math_function.h"
+#include "operators/math/vol2col.h"
 #include "operators/op_param.h"
 namespace paddle_mobile {
 namespace operators {

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

@@ -15,6 +15,8 @@ limitations under the License. */
 #ifdef ELEMENTWISEADD_OP
 
 #pragma once
+#include "operators/math/elementwise_op_function.h"
+#include "operators/op_param.h"
 
 namespace paddle_mobile {
 namespace operators {

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

@@ -15,6 +15,8 @@ limitations under the License. */
 #ifdef FUSION_FC_OP
 
 #pragma once
+#include "operators/math/math_function.h"
+#include "operators/op_param.h"
 
 namespace paddle_mobile {
 namespace operators {

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

@@ -15,7 +15,7 @@ limitations under the License. */
 #ifdef LRN_OP
 
 #pragma once
-
+#include "operators/op_param.h"
 namespace paddle_mobile {
 namespace operators {
 

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

@@ -19,6 +19,8 @@ limitations under the License. */
 #include <map>
 #include <utility>
 #include <vector>
+#include "framework/tensor.h"
+#include "operators/op_param.h"
 
 namespace paddle_mobile {
 namespace operators {
@@ -89,7 +91,8 @@ static inline T JaccardOverlap(const T* box1, const T* box2,
 }
 
 template <typename T>
-static inline void NMSFast(const Tensor& bbox, const Tensor& scores,
+static inline void NMSFast(const framework::Tensor& bbox,
+                           const framework::Tensor& scores,
                            const T score_threshold, const T nms_threshold,
                            const T eta, const int64_t top_k,
                            std::vector<int>* selected_indices) {
@@ -131,7 +134,8 @@ static inline void NMSFast(const Tensor& bbox, const Tensor& scores,
 }
 
 template <typename T>
-void MultiClassNMS(const Tensor& scores, const Tensor& bboxes,
+void MultiClassNMS(const framework::Tensor& scores,
+                   const framework::Tensor& bboxes,
                    std::map<int, std::vector<int>>* indices, int* num_nmsed_out,
                    const int& background_label, const int& nms_top_k,
                    const int& keep_top_k, const T& nms_threshold,
@@ -141,7 +145,7 @@ void MultiClassNMS(const Tensor& scores, const Tensor& bboxes,
   int num_det = 0;
   for (int64_t c = 0; c < class_num; ++c) {
     if (c == background_label) continue;
-    Tensor score = scores.Slice(c, c + 1);
+    framework::Tensor score = scores.Slice(c, c + 1);
     /// [c] is key
     NMSFast<float>(bboxes, score, score_threshold, nms_threshold, nms_eta,
                    nms_top_k, &((*indices)[c]));
@@ -181,9 +185,10 @@ void MultiClassNMS(const Tensor& scores, const Tensor& bboxes,
 }
 
 template <typename T>
-void MultiClassOutput(const Tensor& scores, const Tensor& bboxes,
+void MultiClassOutput(const framework::Tensor& scores,
+                      const framework::Tensor& bboxes,
                       const std::map<int, std::vector<int>>& selected_indices,
-                      Tensor* outs) {
+                      framework::Tensor* outs) {
   int predict_dim = scores.dims()[1];
   auto* scores_data = scores.data<T>();
   auto* bboxes_data = bboxes.data<T>();
@@ -231,10 +236,10 @@ void MultiClassNMSCompute(const MultiClassNMSParam& param) {
   std::vector<std::map<int, std::vector<int>>> all_indices;
   std::vector<size_t> batch_starts = {0};
   for (int64_t i = 0; i < batch_size; ++i) {
-    Tensor ins_score = input_scores->Slice(i, i + 1);
+    framework::Tensor ins_score = input_scores->Slice(i, i + 1);
     ins_score.Resize({class_num, predict_dim});
 
-    Tensor ins_boxes = input_bboxes->Slice(i, i + 1);
+    framework::Tensor ins_boxes = input_bboxes->Slice(i, i + 1);
     ins_boxes.Resize({predict_dim, box_dim});
 
     std::map<int, std::vector<int>> indices;
@@ -253,16 +258,16 @@ void MultiClassNMSCompute(const MultiClassNMSParam& param) {
   } else {
     outs->mutable_data<float>({num_kept, kOutputDim});
     for (int64_t i = 0; i < batch_size; ++i) {
-      Tensor ins_score = input_scores->Slice(i, i + 1);
+      framework::Tensor ins_score = input_scores->Slice(i, i + 1);
       ins_score.Resize({class_num, predict_dim});
 
-      Tensor ins_boxes = input_bboxes->Slice(i, i + 1);
+      framework::Tensor ins_boxes = input_bboxes->Slice(i, i + 1);
       ins_boxes.Resize({predict_dim, box_dim});
 
       int64_t s = batch_starts[i];
       int64_t e = batch_starts[i + 1];
       if (e > s) {
-        Tensor out = outs->Slice(s, e);
+        framework::Tensor out = outs->Slice(s, e);
         MultiClassOutput<float>(ins_score, ins_boxes, all_indices[i], &out);
       }
     }

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

@@ -16,6 +16,7 @@ limitations under the License. */
 #pragma once
 
 #include <operators/math/transform.h>
+#include "operators/op_param.h"
 
 namespace paddle_mobile {
 namespace operators {

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

@@ -16,6 +16,8 @@ limitations under the License. */
 #pragma once
 
 #include <vector>
+#include "operators/kernel/reshape_kernel.h"
+#include "operators/op_param.h"
 
 namespace paddle_mobile {
 namespace operators {

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

@@ -15,6 +15,7 @@ limitations under the License. */
 #ifdef SOFTMAX_OP
 #pragma once
 #include "../../math/softmax.h"
+#include "operators/op_param.h"
 namespace paddle_mobile {
 namespace operators {
 template <typename P>

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

@@ -16,6 +16,7 @@ limitations under the License. */
 #pragma once
 
 #include <vector>
+#include "operators/op_param.h"
 
 namespace paddle_mobile {
 namespace operators {

src/operators/kernel/fpga/conv_add_bn_kernel.cpp

@@ -24,13 +24,13 @@ template <>
 bool ConvAddBNKernel<FPGA, float>::Init(FusionConvAddBNParam *param) {
   bool relu_enabled = false;
   const Tensor *input = param->Input();
-  auto input_ptr = input->data<float>();
+  auto input_ptr = input->data<half>();
   const Tensor *bias = param->Bias();
   auto bias_ptr = bias->data<float>();
   const Tensor *filter = param->Filter();
   auto filter_ptr = filter->data<float>();
   Tensor *out = param->Output();
-  auto out_ptr = out->mutable_data<float>();
+  auto out_ptr = out->mutable_data<half>();
   auto bn_mean_ptr = param->InputMean()->data<float>();
   auto bn_var_ptr = param->InputVariance()->data<float>();
   auto bn_scale_ptr = param->InputScale()->data<float>();

src/operators/kernel/fpga/conv_add_bn_relu_kernel.cpp

@@ -24,13 +24,13 @@ template <>
 bool ConvAddBNReluKernel<FPGA, float>::Init(FusionConvAddBNReluParam *param) {
   bool relu_enabled = true;
   const Tensor *input = param->Input();
-  auto input_ptr = input->data<float>();
+  auto input_ptr = input->data<half>();
   const Tensor *bias = param->Bias();
   auto bias_ptr = bias->data<float>();
   const Tensor *filter = param->Filter();
   auto filter_ptr = filter->data<float>();
   Tensor *out = param->Output();
-  auto out_ptr = out->mutable_data<float>();
+  auto out_ptr = out->mutable_data<half>();
   auto bn_mean_ptr = param->InputMean()->data<float>();
   auto bn_var_ptr = param->InputVariance()->data<float>();
   auto bn_scale_ptr = param->InputScale()->data<float>();

src/operators/kernel/fpga/conv_add_relu_kernel.cpp

@@ -24,13 +24,13 @@ template <>
 bool ConvAddReluKernel<FPGA, float>::Init(FusionConvAddReluParam *param) {
   bool relu_enabled = true;
   const Tensor *input = param->Input();
-  auto input_ptr = input->data<float>();
+  auto input_ptr = input->data<half>();
   const Tensor *bias = param->Bias();
   auto bias_ptr = bias->data<float>();
   const Tensor *filter = param->Filter();
   auto filter_ptr = filter->data<float>();
   Tensor *out = param->Output();
-  auto out_ptr = out->mutable_data<float>();
+  auto out_ptr = out->mutable_data<half>();
 
   PADDLE_MOBILE_ENFORCE(input->dims()[1] == bias->dims()[0],
                         "Image channel should be equal to bias number");

src/operators/op_param.h

@@ -665,6 +665,16 @@ class FeedParam : public OpParam {
   Tensor *input_x_;
   Tensor *out_;
   int batch_size;
+
+#ifdef PADDLE_MOBILE_FPGA
+
+ private:
+  fpga::BypassArgs fpga_bypass_args;
+
+ public:
+  const fpga::BypassArgs &FpgaArgs() const { return fpga_bypass_args; }
+  void SetFpgaArgs(const fpga::BypassArgs &args) { fpga_bypass_args = args; }
+#endif
 };
 
 class FetchParam : public OpParam {