add_position_encoding_op.h 4.3 KB
Newer Older
G
gmcather 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41
/* 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
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"

namespace paddle {
namespace operators {

template <typename DeviceContext, typename T>
class AddPositionEncodingKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& context) const override {
    auto* X = context.Input<framework::LoDTensor>("X");
    auto& x_lod = X->lod();
    auto* src_ptr = X->data<T>();

    auto* Out = context.Output<framework::LoDTensor>("Out");
    auto* dst_ptr = Out->mutable_data<T>(context.GetPlace());

    float alpha = context.Attr<float>("alpha");
    float beta = context.Attr<float>("beta");

    auto x_dim = X->dims();
    int batch_size = 0;
    int max_seq_len = 0;
    int enc_size = 0;

    if (x_lod.empty()) {
42 43 44 45 46 47
      PADDLE_ENFORCE_EQ(x_dim.size(), 3,
                        platform::errors::InvalidArgument(
                            "The input(X)'s dimension of AddPositionEncodingOp "
                            "should be equal to "
                            "3, but received %d. ",
                            x_dim.size()));
G
gmcather 已提交
48 49 50 51
      batch_size = x_dim[0];
      max_seq_len = x_dim[1];
      enc_size = x_dim[2];
    } else {
52 53 54 55 56 57 58 59 60 61 62 63 64
      PADDLE_ENFORCE_EQ(x_dim.size(), 2,
                        platform::errors::InvalidArgument(
                            "The input(X)'s dimension of AddPositionEncodingOp "
                            "should be equal to "
                            "2, but received %d. ",
                            x_dim.size()));
      PADDLE_ENFORCE_EQ(x_lod.size(), 1,
                        platform::errors::InvalidArgument(
                            "The input(X)'s lod level of AddPositionEncodingOp "
                            "should be equal to "
                            "1, but received %d. ",
                            x_lod.size()));

G
gmcather 已提交
65 66 67 68 69
      batch_size = x_lod[0].size() - 1;
      max_seq_len = -1;
      enc_size = x_dim[1];
    }

70 71 72 73 74 75
    PADDLE_ENFORCE_EQ(enc_size % 2, 0,
                      platform::errors::InvalidArgument(
                          "The input(X)'s feature size of "
                          "AddPositionEncodingOp only support even, "
                          "but received an odd number: %d. ",
                          enc_size));
G
gmcather 已提交
76 77 78 79 80 81 82

    const int half_size = enc_size / 2;
    for (int i = 0; i < batch_size; ++i) {
      const int max_length =
          x_lod.empty() ? max_seq_len : x_lod[0][i + 1] - x_lod[0][i];
      for (int j = 0; j < max_length; ++j) {
        for (int k = 0; k < half_size; ++k) {
83 84 85 86
          const double val =
              (half_size > 1)
                  ? j / pow(10000.0, static_cast<double>(k) / (half_size - 1))
                  : j / 10000.0;
G
gmcather 已提交
87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120
          dst_ptr[k] = src_ptr[k] * alpha + sin(val) * beta;
          dst_ptr[half_size + k] =
              src_ptr[half_size + k] * alpha + cos(val) * beta;
        }
        src_ptr += enc_size;
        dst_ptr += enc_size;
      }
    }
  }
};

template <typename DeviceContext, typename T>
class AddPositionEncodingGradKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& context) const override {
    auto* dOut =
        context.Input<framework::LoDTensor>(framework::GradVarName("Out"));
    auto dout = framework::EigenVector<T>::Flatten(*dOut);

    auto* dX =
        context.Output<framework::LoDTensor>(framework::GradVarName("X"));
    dX->mutable_data<T>(context.GetPlace());
    auto dx = framework::EigenVector<T>::Flatten(*dX);

    float alpha = context.Attr<float>("alpha");

    auto* place =
        context.template device_context<DeviceContext>().eigen_device();
    dx.device(*place) = dout * static_cast<T>(alpha);
  }
};

}  // namespace operators
}  // namespace paddle