fusion_conv_add.h 2.9 KB
Newer Older
L
liuruilong 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
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. */
W
wangliu 已提交
14
#define FUSION_CONVADD_OP
W
wangliu 已提交
15
#ifdef FUSION_CONVADD_OP
L
liuruilong 已提交
16

L
liuruilong 已提交
17 18 19 20 21 22
#pragma once

#include <string>
#include <vector>
#include "framework/operator.h"
#include "framework/program/program-optimize/fusion_op_register.h"
W
wangliu 已提交
23
#include "op_param.h"
W
wangliu 已提交
24
#include "operators/kernel/conv_add_kernel.h"
L
liuruilong 已提交
25 26 27 28 29 30 31 32 33 34 35 36

namespace paddle_mobile {
namespace operators {
using std::string;
using std::vector;
class FusionConvAddMatcher : public framework::FusionOpMatcher {
 public:
  FusionConvAddMatcher() {
    node_ = framework::Node(G_OP_TYPE_CONV);
    node_ > std::make_shared<framework::Node>(G_OP_TYPE_ELEMENTWISE_ADD);
  }

L
liuruilong 已提交
37 38 39
  void FolderNodes(
      framework::Node *node,
      std::vector<std::shared_ptr<framework::Node>> *removed_nodes) {
L
liuruilong 已提交
40
    vector<std::shared_ptr<framework::OpDesc>> origin_descs =
L
liuruilong 已提交
41 42 43
        node->OpDescs(node_.Depth());
    node->Folder(node_.Depth(), Type(),
                 {{G_OP_TYPE_ELEMENTWISE_ADD, {"Y", "Y"}}}, removed_nodes);
L
liuruilong 已提交
44 45 46 47 48 49
  }

  std::string Type() { return G_OP_TYPE_CONV_ADD; }
};

template <typename DeviceType, typename T>
L
liuruilong 已提交
50
class FushionConvAddOp : public framework::OperatorWithKernel<DeviceType, FushionConvAddParam, operators::ConvAddKernel<DeviceType, T>> {
L
liuruilong 已提交
51 52
 public:
  FushionConvAddOp(const string &type, const VariableNameMap &inputs,
L
liuruilong 已提交
53
                   const VariableNameMap &outputs,
L
liuruilong 已提交
54
                   const framework::AttributeMap &attrs,
L
liuruilong 已提交
55
                   std::shared_ptr<framework::Scope> scope)
L
liuruilong 已提交
56 57
      : framework::OperatorWithKernel<DeviceType, FushionConvAddParam, operators::ConvAddKernel<DeviceType, T>>(type, inputs, outputs, attrs,
                                                  scope) {}
L
liuruilong 已提交
58

L
liuruilong 已提交
59
  using framework::OperatorWithKernel<DeviceType, FushionConvAddParam, operators::ConvAddKernel<DeviceType, T>>::OperatorWithKernel;
L
liuruilong 已提交
60 61 62 63 64
  void InferShape() const override;

 protected:
};

W
wangliu 已提交
65 66 67 68 69 70 71
inline int ConvOutputSize(int input_size, int filter_size, int dilation,
                          int padding, int stride) {
  const int dkernel = dilation * (filter_size - 1) + 1;
  int output_size = (input_size + 2 * padding - dkernel) / stride + 1;
  return output_size;
}

L
for  
liuruilong 已提交
72
#ifdef PADDLE_MOBILE_CPU
W
wangliu 已提交
73 74
static framework::FusionOpRegistrar convadd_registrar(
    new FusionConvAddMatcher());
L
for  
liuruilong 已提交
75 76 77 78 79
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
L
liuruilong 已提交
80 81 82

}  // namespace operators
}  // namespace paddle_mobile
L
liuruilong 已提交
83 84

#endif