reduce_sum_op.h 3.2 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
// 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.

#pragma once

Q
qiaolongfei 已提交
17 18
#include <vector>

19 20 21 22 23
#include "paddle/fluid/operators/reduce_op.h"

namespace paddle {
namespace operators {

Q
qiaolongfei 已提交
24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79
// use for loop to speed up Eigen broadcast. 4 timer faster then broadcast
template <typename DeviceContext, typename T, typename Functor>
class ReduceSumGradKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& context) const override {
    auto dims = context.Attr<std::vector<int>>("dim");
    if (context.GetPlace().type() == typeid(platform::CPUPlace) &&
        dims.size() == 1) {
      auto* input0 = context.Input<Tensor>("X");
      auto* input2 = context.Input<Tensor>(framework::GradVarName("Out"));
      auto* output = context.Output<Tensor>(framework::GradVarName("X"));
      output->mutable_data<T>(context.GetPlace());
      const auto* input2_d = input2->data<T>();
      auto* output_d = output->data<T>();

      // handle reduce_all
      if (input2->dims().size() == 1 && input2->dims()[0] == 1) {
        for (int64_t i = 0; i < framework::product(input0->dims()); ++i) {
          output_d[i] = input2_d[0];
        }
        return;
      }

      // handle reduce by one dimension
      int reduce_dim_index = dims[0];
      if (reduce_dim_index < 0) {
        reduce_dim_index += input0->dims().size();
      }

      auto& input_dim = input0->dims();
      int64_t before_dim = 1;
      for (int i = 0; i < reduce_dim_index; ++i) {
        before_dim *= input_dim[i];
      }
      int64_t reduce_dim = input_dim[reduce_dim_index];
      int64_t after_dim = 1;
      for (int i = reduce_dim_index + 1; i < input_dim.size(); ++i) {
        after_dim *= input_dim[i];
      }
      for (int64_t i = 0; i < before_dim; ++i) {
        for (int64_t j = 0; j < reduce_dim; ++j) {
          for (int64_t k = 0; k < after_dim; ++k) {
            output_d[i * reduce_dim * after_dim + j * after_dim + k] =
                input2_d[i * after_dim + k];
          }
        }
      }
      return;
    }

    // default use Eigen broadcast
    ReduceGradKernel<DeviceContext, T, Functor> kernel;
    kernel.Compute(context);
  }
};

80 81 82 83 84 85 86 87 88 89 90 91
struct SumFunctor {
  template <typename DeviceContext, typename X, typename Y, typename Dim>
  void operator()(const DeviceContext& place, X* x, Y* y, const Dim& dim) {
    y->device(place) = x->sum(dim);
  }
};

struct SumGradFunctor {
  template <typename DeviceContext, typename X, typename Y, typename DX,
            typename DY, typename Dim>
  void operator()(const DeviceContext& place, X* x, Y* y, DX* dx, DY* dy,
                  const Dim& dim, int size) {
Q
qiaolongfei 已提交
92
    dx->device(place) = dy->eval().broadcast(dim);
93 94 95 96 97
  }
};

}  // namespace operators
}  // namespace paddle