fix some bugs of new added FPGA ops and kernels

CMakeLists.txt

@@ -6,9 +6,15 @@ option(USE_OPENMP "openmp support" OFF)
 option(USE_EXCEPTION "use std exception" ON)
 option(LOG_PROFILE "log profile" ON)
 # select the platform to build
-option(CPU "armv7 with neon" ON)
+option(CPU "armv7 with neon" OFF)
 option(MALI_GPU "mali gpu" OFF)
-option(FPGA "fpga" OFF)
+option(FPGA "fpga" ON)
+SET(FPGA ON)
+SET(FUSION_ELEMENTWISEADDRELU_OP ON)
+SET(FUSION_FC_OP ON)
+SET(FUSION_FCRELU_OP ON)
+SET(POOL_OP ON)
+SET(DROPOUT_OP ON)
 
 file(GLOB_RECURSE PADDLE_MOBILE_CC src/*.cc src/*.cpp src/*.c src/*.mm)
 file(GLOB_RECURSE PADDLE_MOBILE_H src/*.h)
@@ -139,7 +145,8 @@ set(CMAKE_LIBRARY_OUTPUT_DIRECTORY build)
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY build)
 
 # NET default
-set(NET "default" CACHE STRING "select net type")
+#set(NET "default" CACHE STRING "select net type")
+set(NET "FPGAnets" CACHE STRING "select net type")
 set_property(CACHE NET PROPERTY STRINGS "default" "googlenet" "mobilenet" "yolo" "squeezenet" "FPGAnets")
 include("${CMAKE_CURRENT_LIST_DIR}/tools/op.cmake")
 
@@ -151,7 +158,7 @@ if (ANDROID_NDK_TOOLCHAIN_INCLUDED)
 elseif(IS_IOS)
     add_library(paddle-mobile STATIC ${PADDLE_MOBILE_CC} ${PADDLE_MOBILE_H})
 else ()
-    add_library(paddle-mobile SHARED ${PADDLE_MOBILE_CC} ${PADDLE_MOBILE_H})
+    add_library(paddle-mobile SHARED ${PADDLE_MOBILE_CC} ${PADDLE_MOBILE_H} src/operators/kernel/fc_relu_kernel.h src/operators/kernel/fc_relu_kernel.h src/operators/kernel/fpga/fusion_fc_kernel.cpp src/operators/kernel/fpga/fc_relu_kernel.cpp src/operators/fusion_elementwise_add_relu_op.h src/operators/fusion_elementwise_add_relu_op.cpp src/operators/kernel/elementwise_add_relu_kernel.h src/operators/kernel/fpga/pool_kernel.cpp src/operators/kernel/fpga/elementwise_add_relu_kernel.cpp src/operators/kernel/fpga/dropout_kernel.cpp)
 endif ()
 
 # unit test

src/operators/fusion_elementwise_add_relu_op.cpp

@@ -30,12 +30,12 @@ void FusionElementwiseAddReluOp<Dtype, T>::InferShape() const {
 
 namespace ops = paddle_mobile::operators;
 #ifdef PADDLE_MOBILE_CPU
-REGISTER_OPERATOR_CPU(fusion_elementwise_add_relu,
-                      ops::FusionElementwiseAddReluOp);
+// REGISTER_OPERATOR_CPU(fusion_elementwise_add_relu,
+//                      ops::FusionElementwiseAddReluOp);
 #endif
 #ifdef PADDLE_MOBILE_MALI_GPU
-REGISTER_OPERATOR_MALI_GPU(fusion_elementwise_add_relu,
-                           ops::FusionElementwiseAddReluOp);
+// REGISTER_OPERATOR_MALI_GPU(fusion_elementwise_add_relu,
+//                           ops::FusionElementwiseAddReluOp);
 #endif
 #ifdef PADDLE_MOBILE_FPGA
 REGISTER_OPERATOR_FPGA(fusion_elementwise_add_relu,

src/operators/fusion_elementwise_add_relu_op.h

@@ -18,12 +18,29 @@ limitations under the License. */
 
 #include <string>
 #include "framework/operator.h"
-#include "kernel/elementwise_add_relu_kernel.h"
-#include "operators/op_param.h"
+#include "framework/program/program-optimize/fusion_op_register.h"
+#include "operators/kernel/elementwise_add_relu_kernel.h"
 
 namespace paddle_mobile {
 namespace operators {
 using std::string;
+using std::vector;
+class FusioneElementwiseAddReluMatcher : public framework::FusionOpMatcher {
+ public:
+  FusioneElementwiseAddReluMatcher() {
+    node_ = framework::Node(G_OP_TYPE_FUSION_ELEMENTWISE_ADD_RELU);
+    node_ > std::make_shared<framework::Node>(G_OP_TYPE_RELU);
+  }
+
+  void FolderNodes(
+      framework::Node *node,
+      std::vector<std::shared_ptr<framework::Node>> *removed_nodes) {
+    node->Folder(node_.Depth(), Type(), {}, removed_nodes);
+  }
+
+  std::string Type() { return G_OP_TYPE_FUSION_ELEMENTWISE_ADD_RELU; }
+};
+
 template <typename DeviceType, typename T>
 class FusionElementwiseAddReluOp
     : public framework::OperatorWithKernel<
@@ -39,13 +56,38 @@ class FusionElementwiseAddReluOp
             operators::ElementwiseAddReluKernel<DeviceType, T>>(
             type, inputs, outputs, attrs, scope) {}
 
-  using framework::OperatorWithKernel<
-      DeviceType, ElementwiseAddReluParam,
-      operators::ElementwiseAddReluKernel<DeviceType, T>>::OperatorWithKernel;
   void InferShape() const override;
 
  protected:
 };
+
+#ifdef PADDLE_MOBILE_CPU
+/*
+#ifndef FUSION_ELEMENTWISE_ADD_RELU_REGISTER
+        static framework::FusionOpRegistrar fusion_elementwise_relu_registrar(
+                new FusioneElementwiseAddReluMatcher());
+#define FUSION_ELEMENTWISE_ADD_RELU_REGISTER
+#endif
+*/
+#endif
+
+#ifdef PADDLE_MOBILE_MALI_GPU
+/*
+#ifndef FUSION_ELEMENTWISE_ADD_RELU_REGISTER
+        static framework::FusionOpRegistrar fusion_elementwise_relu_registrar(
+                new FusioneElementwiseAddReluMatcher());
+#define FUSION_ELEMENTWISE_ADD_RELU_REGISTER
+#endif
+*/
+#endif
+
+#ifdef PADDLE_MOBILE_FPGA
+#ifndef FUSION_ELEMENTWISE_ADD_RELU_REGISTER
+static framework::FusionOpRegistrar fusion_elementwise_relu_registrar(
+    new FusioneElementwiseAddReluMatcher());
+#define FUSION_ELEMENTWISE_ADD_RELU_REGISTER
+#endif
+
 }  // namespace operators
 }  // namespace paddle_mobile
 
@@ -53,10 +95,10 @@ class FusionElementwiseAddReluOp
 USE_OP_CPU(fusion_elementwise_add_relu);
 #endif
 #ifdef PADDLE_MOBILE_MALI_GPU
-USE_OP_MALI_GPU(fusion_elementwise_add_relu);
 #endif
 #ifdef PADDLE_MOBILE_FPGA
 USE_OP_FPGA(fusion_elementwise_add_relu);
 #endif
 
 #endif
+#endif

src/operators/fusion_fc_op.h

@@ -67,8 +67,8 @@ class FusionFcOp
 
 #ifdef PADDLE_MOBILE_CPU
 
-#ifndef CONV_CPU_REGISTER
-#define CONV_CPU_REGISTER
+#ifndef FUSION_FC_CPU_REGISTER
+#define FUSION_FC_CPU_REGISTER
 static framework::FusionOpRegistrar fc_registrar(new FusionFcMatcher());
 #endif
 
@@ -84,6 +84,10 @@ static framework::FusionOpRegistrar fc_registrar(new FusionFcMatcher());
 #endif
 
 #ifdef PADDLE_MOBILE_FPGA
+#ifndef FUSION_FC_CPU_REGISTER
+#define FUSION_FC_CPU_REGISTER
+static framework::FusionOpRegistrar fc_registrar(new FusionFcMatcher());
+#endif
 #endif
 
 }  // namespace operators

src/operators/fusion_fc_relu_op.h

@@ -18,7 +18,7 @@ limitations under the License. */
 
 #include "framework/operator.h"
 #include "framework/program/program-optimize/fusion_op_register.h"
-#include "operators/kernel/fusion_fc_relu_kernel.h"
+#include "operators/kernel/fc_relu_kernel.h"
 
 namespace paddle_mobile {
 namespace operators {

src/operators/kernel/elementwise_add_relu_kernel.h

@@ -17,7 +17,6 @@ limitations under the License. */
 #pragma once
 
 #include "framework/operator.h"
-#include "operators/math/elementwise_op_function.h"
 #include "operators/op_param.h"
 
 namespace paddle_mobile {

src/operators/kernel/fpga/elementwise_add_relu_kernel.cpp

@@ -14,7 +14,6 @@ limitations under the License. */
 #ifdef FUSION_ELEMENTWISEADDRELU_OP
 
 #include "operators/kernel/elementwise_add_relu_kernel.h"
-#include "fpga/api/fpga_api.h"
 
 namespace paddle_mobile {
 namespace operators {
@@ -28,7 +27,7 @@ bool ElementwiseAddReluKernel<FPGA, float>::Init(
   Tensor *out = param->Out();
   auto input_x_ptr = input_x->data<float>();
   auto input_y_ptr = input_y->data<float>();
-  auto out_ptr = out->data<float>();
+  auto out_ptr = out->mutable_data<float>();
 
   fpga::EWAddArgs ewaddArgs;
   ewaddArgs.relu_enabled = relu_enabled;
@@ -40,16 +39,16 @@ bool ElementwiseAddReluKernel<FPGA, float>::Init(
       input_x->fpga_args().scale_pointer();  // ew has scale attribute??
   ewaddArgs.image0.height = input_x->dims()[2];
   ewaddArgs.image0.width = input_x->dims()[3];
-  ewaddArgs.image0.pad_height = 1;
-  ewaddArgs.image0.pad_width = 1;
+  ewaddArgs.image0.pad_height = 0;
+  ewaddArgs.image0.pad_width = 0;
   ewaddArgs.image1.address = (void *)input_y_ptr;
   ewaddArgs.image1.channels = input_y->dims()[1];
   ewaddArgs.image1.scale_address =
       input_y->fpga_args().scale_pointer();  // ew has scale attribute??
   ewaddArgs.image1.height = input_y->dims()[2];
   ewaddArgs.image1.width = input_y->dims()[3];
-  ewaddArgs.image1.pad_height = 1;
-  ewaddArgs.image1.pad_width = 1;
+  ewaddArgs.image1.pad_height = 0;
+  ewaddArgs.image1.pad_width = 0;
   ewaddArgs.output.scale_address = out->fpga_args().scale_pointer();
   ewaddArgs.output.address = (void *)out_ptr;
   param->SetFpgaArgs(ewaddArgs);

src/operators/kernel/fpga/fc_relu_kernel.cpp

@@ -21,7 +21,6 @@ namespace operators {
 template <>
 bool FusionFcReluKernel<FPGA, float>::Init(FusionFcReluParam *param) {
   bool relu_enabled = true;
-  bool bn_enabled = false;
   const Tensor *input_x = param->InputX();
   auto input_x_ptr = input_x->data<float>();
   const Tensor *input_y = param->InputY();
@@ -31,8 +30,8 @@ bool FusionFcReluKernel<FPGA, float>::Init(FusionFcReluParam *param) {
   Tensor *out = param->Out();
   auto out_ptr = out->mutable_data<float>();
 
-  PADDLE_MOBILE_ENFORCE(input_x->dims()[1] == input_z->dims()[0],
-                        "Image channel should be equal to bias number");
+  PADDLE_MOBILE_ENFORCE(input_x->dims()[1] == input_y->dims()[0],
+                        "Image channel should be equal to weight number");
   int channel = input_x->dims()[1];
   float *bs_ptr = (float *)fpga::fpga_malloc(2 * channel * sizeof(float));
   for (int i = 0; i < channel; i++) {
@@ -55,8 +54,8 @@ bool FusionFcReluKernel<FPGA, float>::Init(FusionFcReluParam *param) {
   convArgs.image.channels = input_x->dims()[1];
   convArgs.image.height = input_x->dims()[2];
   convArgs.image.width = input_x->dims()[3];
-  convArgs.image.pad_height = 1;
-  convArgs.image.pad_width = 1;
+  convArgs.image.pad_height = 0;
+  convArgs.image.pad_width = 0;
   convArgs.image.scale_address =
       input_x->fpga_args().scale_pointer();  // fc input has scale attribute??
   convArgs.output.address = (void *)out_ptr;

src/operators/kernel/fpga/fusion_fc_kernel.cpp

@@ -21,7 +21,6 @@ namespace operators {
 template <>
 bool FusionFcKernel<FPGA, float>::Init(FusionFcParam *param) {
   bool relu_enabled = false;
-  bool bn_enabled = false;
   const Tensor *input_x = param->InputX();
   auto input_x_ptr = input_x->data<float>();
   const Tensor *input_y = param->InputY();
@@ -31,8 +30,8 @@ bool FusionFcKernel<FPGA, float>::Init(FusionFcParam *param) {
   Tensor *out = param->Out();
   auto out_ptr = out->mutable_data<float>();
 
-  PADDLE_MOBILE_ENFORCE(input_x->dims()[1] == input_z->dims()[0],
-                        "Image channel should be equal to bias number");
+  PADDLE_MOBILE_ENFORCE(input_x->dims()[1] == input_y->dims()[0],
+                        "Image channel should be equal to weight number");
   int channel = input_x->dims()[1];
   float *bs_ptr = (float *)fpga::fpga_malloc(2 * channel * sizeof(float));
   for (int i = 0; i < channel; i++) {
@@ -55,8 +54,8 @@ bool FusionFcKernel<FPGA, float>::Init(FusionFcParam *param) {
   convArgs.image.channels = input_x->dims()[1];
   convArgs.image.height = input_x->dims()[2];
   convArgs.image.width = input_x->dims()[3];
-  convArgs.image.pad_height = 1;
-  convArgs.image.pad_width = 1;
+  convArgs.image.pad_height = 0;
+  convArgs.image.pad_width = 0;
   convArgs.image.scale_address =
       input_x->fpga_args().scale_pointer();  // fc input has scale attribute??
   convArgs.output.address = (void *)out_ptr;

src/operators/kernel/fpga/pool_kernel.cpp

@@ -14,7 +14,6 @@ limitations under the License. */
 #ifdef POOL_OP
 
 #include "operators/kernel/pool_kernel.h"
-#include "fpga/api/fpga_api.h"
 
 class PoolingArgs;
 namespace paddle_mobile {
@@ -25,7 +24,7 @@ bool PoolKernel<FPGA, float>::Init(PoolParam *param) {
   const Tensor *input = param->Input();
   auto input_ptr = input->data<float>();
   Tensor *output = param->Output();
-  auto output_ptr = output->data<float>();
+  auto output_ptr = output->mutable_data<float>();
   vector<int> ksize = param->Ksize();
   vector<int> strides = param->Strides();
   vector<int> paddings = param->Paddings();