dot_op.h

// Copyright (c) 2020 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/op_registry.h"
#include "paddle/fluid/framework/operator.h"

namespace paddle {
namespace operators {

using Tensor = framework::Tensor;

template <typename T, int MajorType = Eigen::RowMajor,
          typename IndexType = Eigen::DenseIndex>
using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;

template <typename DeviceContext, typename T>
void DotGradFunction(const Tensor* tensor_x, const Tensor* tensor_y,
                     const Tensor* tensor_dout, Tensor* tensor_dx,
                     Tensor* tensor_dy,
                     const paddle::framework::ExecutionContext& ctx) {
#ifdef __NVCC__
  if (1 == tensor_dout->dims().size()) {
    auto dout = framework::EigenVector<T>::Flatten(*tensor_dout);

    if (tensor_dx) {
      auto y = framework::EigenVector<T>::Flatten(*tensor_y);
      auto dx = framework::EigenVector<T>::Flatten(*tensor_dx);
      auto& dev = *ctx.template device_context<DeviceContext>().eigen_device();
      Eigen::DSizes<int, 1> size(tensor_dx->numel());
      dx.device(dev) = y * dout.broadcast(size);
    }

    if (tensor_dy) {
      auto x = framework::EigenVector<T>::Flatten(*tensor_x);
      auto dy = framework::EigenVector<T>::Flatten(*tensor_dy);
      auto& dev = *ctx.template device_context<DeviceContext>().eigen_device();
      Eigen::DSizes<int, 1> size(tensor_dy->numel());
      dy.device(dev) = x * dout.broadcast(size);
    }
  } else {
    auto dout = EigenMatrix<T>::From(*tensor_dout);

    if (tensor_dx) {
      tensor_dx->mutable_data<T>(ctx.GetPlace());
      auto y = EigenMatrix<T>::From(*tensor_y);
      auto dx = EigenMatrix<T>::From(*tensor_dx);
      auto& dev = *ctx.template device_context<DeviceContext>().eigen_device();
      Eigen::DSizes<int, 2> size(1, tensor_dx->dims()[1]);
      dx.device(dev) = y * dout.broadcast(size);
    }

    if (tensor_dy) {
      tensor_dy->mutable_data<T>(ctx.GetPlace());
      auto x = EigenMatrix<T>::From(*tensor_x);
      auto dy = EigenMatrix<T>::From(*tensor_dy);
      auto& dev = *ctx.template device_context<DeviceContext>().eigen_device();
      Eigen::DSizes<int, 2> size(1, tensor_dy->dims()[1]);
      dy.device(dev) = x * dout.broadcast(size);
    }
  }
#else
  const auto* data_dout = tensor_dout->data<T>();

  if (tensor_dx) {
    auto* data_dx = tensor_dx->mutable_data<T>(ctx.GetPlace());
    const auto* data_y = tensor_y->data<T>();
    const framework::DDim& dim = tensor_x->dims();
    size_t N = static_cast<size_t>(framework::product(dim));

    auto step = dim[dim.size() - 1];

    int s = -1;
    for (size_t i = 0; i < N; ++i) {
      if (0 == i % step) ++s;
      data_dx[i] = data_y[i] * data_dout[s];
    }
  }

  if (tensor_dy) {
    auto* data_dy = tensor_dy->mutable_data<T>(ctx.GetPlace());
    const auto* data_x = tensor_x->data<T>();
    const framework::DDim& dim = tensor_y->dims();
    size_t N = static_cast<size_t>(framework::product(dim));

    auto step = dim[dim.size() - 1];

    int s = -1;
    for (size_t i = 0; i < N; ++i) {
      if (0 == i % step) ++s;
      data_dy[i] = data_x[i] * data_dout[s];
    }
  }
#endif
}

template <typename DeviceContext, typename T>
class DotKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    auto* tensor_x = ctx.Input<Tensor>("X");
    auto* tensor_y = ctx.Input<Tensor>("Y");
    auto* tensor_out = ctx.Output<Tensor>("Out");
    tensor_out->mutable_data<T>(ctx.GetPlace());

#ifdef __NVCC__
    if (1 == tensor_out->dims().size()) {
      auto out = framework::EigenScalar<T>::From(*tensor_out);
      auto x = framework::EigenVector<T>::Flatten(*tensor_x);
      auto y = framework::EigenVector<T>::Flatten(*tensor_y);

      auto& dev = *ctx.template device_context<DeviceContext>().eigen_device();
      out.device(dev) = (x * y).sum();
    } else {
      auto out = EigenMatrix<T>::From(*tensor_out);
      auto x = EigenMatrix<T>::From(*tensor_x);
      auto y = EigenMatrix<T>::From(*tensor_y);

      auto& dev = *ctx.template device_context<DeviceContext>().eigen_device();
      out.device(dev) = (x * y).sum(Eigen::DSizes<int, 1>(1));
    }
#else
    const auto* data_x = tensor_x->data<T>();
    const auto* data_y = tensor_y->data<T>();
    auto* data_out = tensor_out->data<T>();

    auto x_dims = tensor_x->dims();
    auto step = x_dims[x_dims.size() - 1];
    int size = static_cast<int>(framework::product(x_dims));

    for (int ind = -1, j = 0; j < size; ++j) {
      if (j % step == 0) {
        ++ind;
        data_out[ind] = data_x[j] * data_y[j];
      } else {
        data_out[ind] += data_x[j] * data_y[j];
      }
    }
#endif
  }
};

template <typename DeviceContext, typename T>
class DotGradKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    auto* tensor_x = ctx.Input<Tensor>("X");
    auto* tensor_y = ctx.Input<Tensor>("Y");
    auto* tensor_dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
    auto* tensor_dx = ctx.Output<Tensor>(framework::GradVarName("X"));
    auto* tensor_dy = ctx.Output<Tensor>(framework::GradVarName("Y"));

    if (tensor_dx) tensor_dx->mutable_data<T>(ctx.GetPlace());
    if (tensor_dy) tensor_dy->mutable_data<T>(ctx.GetPlace());

    DotGradFunction<DeviceContext, T>(tensor_x, tensor_y, tensor_dout,
                                      tensor_dx, tensor_dy, ctx);
  }
};

}  // namespace operators
}  // namespace paddle