Merge remote-tracking branch 'upstream/develop' into develop

src/common/variant.h

@@ -84,7 +84,7 @@ struct Variant {
     if (type_id == typeid(T).hash_code()) {
       return *const_cast<T *>(reinterpret_cast<const T *>(&data));
     } else {
-      PADDLE_MOBILE_THROW_EXCEPTION(" bad cast in variant ");
+      PADDLE_MOBILE_THROW_EXCEPTION(" bad cast in variant");
       exit(0);
     }
   }

src/framework/program/program-optimize/fusion_op_register.h

@@ -42,8 +42,17 @@ class FusionOpRegister {
     matchers_[matcher->Type()] = shared_matcher;
   }
 
-  const std::map<std::string, std::shared_ptr<FusionOpMatcher>> Matchers() {
-    return matchers_;
+  const std::vector<std::shared_ptr<FusionOpMatcher>> Matchers() {
+    std::vector<std::shared_ptr<FusionOpMatcher>> matchers;
+    for (const auto& match : matchers_) {
+      matchers.push_back(match.second);
+    }
+    std::sort(matchers.begin(), matchers.end(),
+              [](std::shared_ptr<FusionOpMatcher> first,
+                 std::shared_ptr<FusionOpMatcher> second) {
+                return first->BeginNode().Depth() > second->BeginNode().Depth();
+              });
+    return matchers;
   }
 
  private:

src/framework/program/program-optimize/node.cpp

@@ -44,23 +44,6 @@ bool Node::operator==(const Node &in) {
   return true;
 }
 
-std::vector<std::shared_ptr<framework::OpDesc>> Node::OpDescs(int size) {
-  std::vector<std::shared_ptr<framework::OpDesc>> op_descs;
-  OpDescs(size - 1, &op_descs);
-  return op_descs;
-}
-
-void Node::OpDescs(int index,
-                   std::vector<std::shared_ptr<framework::OpDesc>> *op_desc) {
-  if (index == 0) {
-    return;
-  }
-  op_desc->push_back(this->op_desc_);
-  for (auto &output : outputs_) {
-    output->OpDescs(index, op_desc);
-  }
-}
-
 std::shared_ptr<Node> Node::To(int size) {
   std::shared_ptr<Node> node = std::make_shared<Node>();
   this->To(size - 1, node);

src/framework/program/program-optimize/node.h

@@ -47,13 +47,10 @@ class Node {
       std::map<std::string, std::vector<std::pair<std::string, std::string>>>
           change,
       std::vector<std::shared_ptr<Node>> *removed_nodes);
-  std::vector<std::shared_ptr<framework::OpDesc>> OpDescs(int size);
   std::shared_ptr<framework::OpDesc> OpDescOfNode() { return op_desc_; }
   std::string Type() { return type_; }
 
  private:
-  void OpDescs(int size,
-               std::vector<std::shared_ptr<framework::OpDesc>> *op_desc);
   void To(int index, std::shared_ptr<Node>);
   void Folder(
       std::shared_ptr<framework::OpDesc> op_desc,

src/framework/program/program-optimize/program_optimize.cpp

@@ -78,9 +78,8 @@ std::shared_ptr<ProgramDesc> ProgramOptimize::FusionOptimize(
     }
 
     for (auto &registed : FusionOpRegister::Instance()->Matchers()) {
-      std::string fusion_type = registed.first;
-      std::shared_ptr<FusionOpMatcher> matcher = registed.second;
-      //      DLOG << " registed node \n " << matcher->BeginNode();
+      std::string fusion_type = registed->Type();
+      std::shared_ptr<FusionOpMatcher> matcher = registed;
 
       auto match_vector = type_map[matcher->BeginType()];
 

src/operators/batchnorm_op.cpp

@@ -26,7 +26,7 @@ void BatchNormOp<Dtype, T>::InferShape() const {
   auto x_dims = this->param_.InputX()->dims();
   this->param_.OutputY()->Resize(x_dims);
 }
-template class BatchNormOp<CPU, float>;
+
 }  // namespace operators
 }  // namespace paddle_mobile
 

src/operators/box_coder_op.cpp

@@ -47,7 +47,7 @@ void BoxCoderOp<Dtype, T>::InferShape() const {
   this->param_.OutputBox()->Resize(framework::make_ddim(
       {input_targetbox_dims[0], input_priorbox_dims[0], 4}));
 }
-template class BoxCoderOp<CPU, float>;
+
 }  // namespace operators
 }  // namespace paddle_mobile
 

src/operators/concat_op.cpp

@@ -56,7 +56,6 @@ void ConcatOp<Dtype, T>::InferShape() const {
 
   this->param_.Out()->Resize(out_dims);
 }
-template class ConcatOp<CPU, float>;
 
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/conv_op.cpp

@@ -48,8 +48,6 @@ void ConvOp<Dtype, T>::InferShape() const {
   this->param_.Output()->Resize(ddim);
 }
 
-template class ConvOp<CPU, float>;
-
 }  // namespace operators
 }  // namespace paddle_mobile
 

src/operators/depthwise_conv_op.cpp

@@ -49,8 +49,6 @@ void DepthwiseConvOp<Dtype, T>::InferShape() const {
   this->param_.Output()->Resize(ddim);
 }
 
-template class DepthwiseConvOp<CPU, float>;
-
 }  // namespace operators
 }  // namespace paddle_mobile
 

src/operators/dropout_op.cpp

@@ -22,7 +22,7 @@ void DropoutOp<Dtype, T>::InferShape() const {
   auto input_dims = this->param_.InputX()->dims();
   this->param_.Out()->Resize(input_dims);
 }
-template class DropoutOp<CPU, float>;
+
 }  // namespace operators
 }  // namespace paddle_mobile
 

src/operators/elementwise_add_op.cpp

@@ -24,7 +24,7 @@ void ElementwiseAddOp<Dtype, T>::InferShape() const {
   auto x_dim = this->param_.InputX()->dims();
   this->param_.Out()->Resize(x_dim);
 }
-template class ElementwiseAddOp<CPU, float>;
+
 }  // namespace operators
 }  // namespace paddle_mobile
 

src/operators/feed_op.cpp

@@ -14,10 +14,7 @@ limitations under the License. */
 
 #include "feed_op.h"
 namespace paddle_mobile {
-namespace operators {
-
-template class FeedOp<CPU, float>;
-}
+namespace operators {}
 }  // namespace paddle_mobile
 
 namespace ops = paddle_mobile::operators;

src/operators/fetch_op.cpp

@@ -14,10 +14,7 @@ limitations under the License. */
 
 #include "fetch_op.h"
 namespace paddle_mobile {
-namespace operators {
-
-template class FetchOp<CPU, float>;
-}
+namespace operators {}
 }  // namespace paddle_mobile
 
 namespace ops = paddle_mobile::operators;

src/operators/fusion_conv_add.cpp

@@ -45,7 +45,6 @@ void FusionConvAddOp<Dtype, T>::InferShape() const {
   this->param_.Output()->Resize(ddim);
 }
 
-template class FusionConvAddOp<CPU, float>;
 }  // namespace operators
 }  // namespace paddle_mobile
 

src/operators/fusion_conv_add.h

@@ -36,8 +36,6 @@ class FusionConvAddMatcher : public framework::FusionOpMatcher {
   void FolderNodes(
       framework::Node *node,
       std::vector<std::shared_ptr<framework::Node>> *removed_nodes) {
-    vector<std::shared_ptr<framework::OpDesc>> origin_descs =
-        node->OpDescs(node_.Depth());
     node->Folder(node_.Depth(), Type(),
                  {{G_OP_TYPE_ELEMENTWISE_ADD, {{"Y", "Y"}}}}, removed_nodes);
   }

src/operators/fusion_conv_add_bn_relu_op.cpp

@@ -44,7 +44,7 @@ void FusionConvAddBNReluOp<Dtype, T>::InferShape() const {
   framework::DDim ddim = framework::make_ddim(output_shape);
   this->param_.Output()->Resize(ddim);
 }
-template class FusionConvAddBNReluOp<CPU, float>;
+
 }  // namespace operators
 }  // namespace paddle_mobile
 

src/operators/fusion_conv_add_bn_relu_op.h

@@ -39,8 +39,6 @@ class FusionConvAddBNReluMatcher : public framework::FusionOpMatcher {
   void FolderNodes(
       framework::Node *node,
       std::vector<std::shared_ptr<framework::Node>> *removed_nodes) {
-    vector<std::shared_ptr<framework::OpDesc>> origin_descs =
-        node->OpDescs(node_.Depth());
     node->Folder(node_.Depth(), Type(),
                  {{G_OP_TYPE_ELEMENTWISE_ADD, {{"Y", "Y"}}},
                   {G_OP_TYPE_BATCHNORM,

src/operators/fusion_dwconv_bn_relu_op.cpp

@@ -44,7 +44,7 @@ void FusionDWConvBNReluOp<Dtype, T>::InferShape() const {
   framework::DDim ddim = framework::make_ddim(output_shape);
   this->param_.Output()->Resize(ddim);
 }
-template class FusionDWConvBNReluOp<CPU, float>;
+
 }  // namespace operators
 }  // namespace paddle_mobile
 

src/operators/fusion_dwconv_bn_relu_op.h

@@ -38,8 +38,6 @@ class FusionDWConvBNReluMatcher : public framework::FusionOpMatcher {
   void FolderNodes(
       framework::Node *node,
       std::vector<std::shared_ptr<framework::Node>> *removed_nodes) {
-    vector<std::shared_ptr<framework::OpDesc>> origin_descs =
-        node->OpDescs(node_.Depth());
     node->Folder(node_.Depth(), Type(),
                  {{G_OP_TYPE_BATCHNORM,
                    {{"Scale", "Scale"},

src/operators/fusion_fc_op.cpp

@@ -50,7 +50,6 @@ void FusionFcOp<Dtype, T>::InferShape() const {
   this->param_.Out()->Resize(ddim);
 }
 
-template class FusionFcOp<CPU, float>;
 }  // namespace operators
 }  // namespace paddle_mobile
 

src/operators/im2sequence_op.cpp

@@ -47,8 +47,6 @@ void Im2SequenceOp<Dtype, T>::InferShape() const {
   this->param_.Output()->Resize(ddim);
 }
 
-template class Im2SequenceOp<CPU, float>;
-
 }  // namespace operators
 }  // namespace paddle_mobile
 

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

@@ -104,7 +104,7 @@ void ConvAddBNReluBasic(const FusionConvAddBNReluParam &param) {
 
       math::matmulWithBn<float>(
           filter_slice, false, col_matrix, false, static_cast<float>(1),
-          &out_slice, static_cast<float>(0), true, &new_scale, &new_bias);
+          &out_slice, static_cast<float>(0), true, &new_scale, &new_bias, g);
     }
   }
 }

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

@@ -101,10 +101,9 @@ void DWConvBNReluBasic(const FusionDWConvBNReluParam &param) {
       // gemm
       Tensor out_slice = out_batch.Slice(g * out_step, (g + 1) * out_step);
       Tensor filter_slice = filter.Slice(g * out_step, (g + 1) * out_step);
-      std::cout << "***************" << std::endl;
       math::matmulWithBn<float>(
           filter_slice, false, col_matrix, false, static_cast<float>(1),
-          &out_slice, static_cast<float>(0), false, &new_scale, &new_bias);
+          &out_slice, static_cast<float>(0), true, &new_scale, &new_bias, g);
     }
   }
 }

src/operators/lrn_op.cpp

@@ -24,7 +24,7 @@ void LrnOp<Dtype, T>::InferShape() const {
   auto x_dims = this->param_.InputX()->dims();
   this->param_.Out()->Resize(x_dims);
 }
-template class LrnOp<CPU, float>;
+
 }  // namespace operators
 }  // namespace paddle_mobile
 

src/operators/math/math_function.cpp

@@ -50,7 +50,7 @@ void matmulWithBn<float>(const framework::Tensor &matrix_a, bool trans_a,
                          const framework::Tensor &matrix_b, bool trans_b,
                          float alpha, framework::Tensor *matrix_out, float beta,
                          bool relu, framework::Tensor *new_scale,
-                         framework::Tensor *new_bias) {
+                         framework::Tensor *new_bias, int group) {
   auto dim_a = matrix_a.dims();
   auto dim_b = matrix_b.dims();
   auto dim_out = matrix_out->dims();
@@ -71,7 +71,8 @@ void matmulWithBn<float>(const framework::Tensor &matrix_a, bool trans_a,
 
   SgemmWithBn(M, N, K, alpha, matrix_a.data<float>(), K, matrix_b.data<float>(),
               N, beta, matrix_out->data<float>(), N, relu,
-              new_scale->data<float>(), new_bias->data<float>());
+              new_scale->data<float>() + group,
+              new_bias->data<float>() + group);
 }
 
 }  // namespace math

src/operators/math/math_function.h

@@ -31,7 +31,8 @@ template <typename T>
 void matmulWithBn(const framework::Tensor &matrix_a, bool trans_a,
                   const framework::Tensor &matrix_b, bool trans_b, T alpha,
                   framework::Tensor *matrix_out, T beta, bool relu,
-                  framework::Tensor *new_scale, framework::Tensor *new_bias);
+                  framework::Tensor *new_scale, framework::Tensor *new_bias,
+                  int group);
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/mul_op.cpp

@@ -50,7 +50,7 @@ void MulOp<Dtype, T>::InferShape() const {
   framework::DDim ddim = framework::make_ddim(output_dims);
   this->param_.Out()->Resize(ddim);
 }
-template class MulOp<CPU, float>;
+
 }  // namespace operators
 }  // namespace paddle_mobile
 

src/operators/multiclass_nms_op.cpp

@@ -34,7 +34,7 @@ void MultiClassNMSOp<Dtype, T>::InferShape() const {
   // pre size, will change in Compute.
   this->param_.Out()->Resize(framework::make_ddim({input_bboxes_dims[1], 6}));
 }
-template class MultiClassNMSOp<CPU, float>;
+
 }  // namespace operators
 }  // namespace paddle_mobile
 

src/operators/op_param.h

@@ -371,7 +371,7 @@ class BatchNormParam : OpParam {
     input_variance_ = InputVarianceFrom<LoDTensor>(inputs, scope);
     epsilon_ = GetAttr<float>("epsilon", attrs);
     momentum_ = GetAttr<float>("momentum", attrs);
-    is_test_ = GetAttr<bool>("is_test", attrs);
+    //    is_test_ = GetAttr<bool>("is_test", attrs);
   }
 
   const Tensor *InputX() const { return input_x_; }

src/operators/pool_op.cpp

@@ -54,7 +54,7 @@ void PoolOp<DeviceType, T>::InferShape() const {
   }
   this->param_.Output()->Resize(framework::make_ddim(output_shape));
 }
-template class PoolOp<CPU, float>;
+
 }  // namespace operators
 }  // namespace paddle_mobile
 

src/operators/prelu_op.cpp

@@ -23,7 +23,7 @@ void PReluOp<Dtype, T>::InferShape() const {
   auto input_dims = this->param_.InputX()->dims();
   this->param_.Out()->Resize(input_dims);
 }
-template class PReluOp<CPU, float>;
+
 }  // namespace operators
 }  // namespace paddle_mobile
 

src/operators/prior_box_op.cpp

@@ -44,7 +44,7 @@ void PriorBoxOp<Dtype, T>::InferShape() const {
   this->param_.OutputBoxes()->Resize(framework::make_ddim(dim_vec));
   this->param_.OutputVariances()->Resize(framework::make_ddim(dim_vec));
 }
-template class PriorBoxOp<CPU, float>;
+
 }  // namespace operators
 }  // namespace paddle_mobile
 

src/operators/relu_op.cpp

@@ -23,7 +23,7 @@ void ReluOp<Dtype, T>::InferShape() const {
   auto input_dims = this->param_.InputX()->dims();
   this->param_.Out()->Resize(input_dims);
 }
-template class ReluOp<CPU, float>;
+
 }  // namespace operators
 }  // namespace paddle_mobile
 

src/operators/reshape_op.cpp

@@ -27,7 +27,7 @@ void ReshapeOp<Dtype, T>::InferShape() const {
   auto out_dims = ValidateShape(shape, input_x_dims);
   this->param_.Out()->Resize(out_dims);
 }
-template class ReshapeOp<CPU, float>;
+
 }  // namespace operators
 }  // namespace paddle_mobile
 

src/operators/resize_op.cpp

@@ -24,7 +24,7 @@ void ResizeOp<Dtype, T>::InferShape() const {
   auto out_dims = CalOutputShape(this->param_);
   this->param_.Out()->Resize(out_dims);
 }
-template class ResizeOp<CPU, float>;
+
 }  // namespace operators
 }  // namespace paddle_mobile
 

src/operators/scale_op.cpp

@@ -24,7 +24,7 @@ void ScaleOp<Dtype, T>::InferShape() const {
   auto input_dims = this->param_.InputX()->dims();
   this->param_.Out()->Resize(input_dims);
 }
-template class ScaleOp<CPU, float>;
+
 }  // namespace operators
 }  // namespace paddle_mobile
 

src/operators/sigmoid_op.cpp

@@ -22,7 +22,7 @@ template <typename DeviceType, typename T>
 void SigmoidOp<DeviceType, T>::InferShape() const {
   this->param_.Out()->Resize(this->param_.InputX()->dims());
 }
-template class SigmoidOp<CPU, float>;
+
 }  // namespace operators
 }  // namespace paddle_mobile
 

src/operators/slice_op.cpp

@@ -23,7 +23,7 @@ template <typename Dtype, typename T>
 void SliceOp<Dtype, T>::InferShape() const {
   /// todo: add InputShape() detection.
 }
-template class SliceOp<CPU, float>;
+
 }  // namespace operators
 }  // namespace paddle_mobile
 

src/operators/softmax_op.cpp

@@ -22,7 +22,7 @@ template <typename DeviceType, typename T>
 void SoftmaxOp<DeviceType, T>::InferShape() const {
   this->param_.Out()->Resize(this->param_.InputX()->dims());
 }
-template class SoftmaxOp<CPU, float>;
+
 }  // namespace operators
 }  // namespace paddle_mobile
 

src/operators/transpose_op.cpp

@@ -47,7 +47,7 @@ void TransposeOp<Dtype, T>::InferShape() const {
   }
   this->param_.Out()->Resize(out_dims);
 }
-template class TransposeOp<CPU, float>;
+
 }  // namespace operators
 }  // namespace paddle_mobile