expand_v2_op.h

/* Copyright (c) 2016 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 <algorithm>
#include <vector>

#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/operators/eigen/eigen_function.h"

#define MAX_RANK_SUPPORTED 6

namespace paddle {
namespace operators {
inline std::vector<int> get_expand_shape(
    const framework::ExecutionContext& ctx) {
  if (ctx.HasInput("Shape")) {
    auto* shape_tensor = ctx.Input<framework::LoDTensor>("Shape");
    auto* shape_data = shape_tensor->data<int>();
    framework::Tensor cpu_shape_tensor;
    if (platform::is_gpu_place(shape_tensor->place())) {
      TensorCopySync(*shape_tensor, platform::CPUPlace(), &cpu_shape_tensor);
      shape_data = cpu_shape_tensor.data<int>();
    }
#ifdef PADDLE_WITH_ASCEND_CL
    if (platform::is_npu_place(shape_tensor->place())) {
      TensorCopySync(*shape_tensor, platform::CPUPlace(), &cpu_shape_tensor);
      shape_data = cpu_shape_tensor.data<int>();
    }
#endif
    auto vec_shape =
        std::vector<int>(shape_data, shape_data + shape_tensor->numel());
    return vec_shape;
  }

  auto list_expand_shapes_tensor =
      ctx.MultiInput<framework::Tensor>("expand_shapes_tensor");
  if (list_expand_shapes_tensor.size() > 0) {
    // get tensor from
    std::vector<int> vec_epxand_shape;
    for (size_t i = 0; i < list_expand_shapes_tensor.size(); ++i) {
      auto tensor = list_expand_shapes_tensor[i];
      if (platform::is_gpu_place(tensor->place())) {
        framework::Tensor temp;
        TensorCopySync(*tensor, platform::CPUPlace(), &temp);
        vec_epxand_shape.push_back(*temp.data<int32_t>());
      }
#ifdef PADDLE_WITH_ASCEND_CL
      else if (platform::is_npu_place(tensor->place())) {  // NOLINT
        framework::Tensor temp;
        TensorCopySync(*tensor, platform::CPUPlace(), &temp);
        vec_epxand_shape.push_back(*temp.data<int32_t>());
      }
#endif
      else {  // NOLINT
        vec_epxand_shape.push_back(*tensor->data<int32_t>());
      }
    }
    return vec_epxand_shape;
  } else {
    return ctx.Attr<std::vector<int>>("shape");
  }
}

using Tensor = framework::Tensor;
template <typename T, int MajorType = Eigen::RowMajor,
          typename IndexType = Eigen::DenseIndex>
using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
template <typename T, size_t D, int MajorType = Eigen::RowMajor,
          typename IndexType = Eigen::DenseIndex>
using EigenTensor = framework::EigenTensor<T, D, MajorType, IndexType>;
using framework::To32BitIndex;

template <typename DeviceContext, typename T>
class ExpandV2Kernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& context) const override {
    auto rank = context.Input<Tensor>("X")->dims().size();
    PADDLE_ENFORCE_GE(
        rank, 1,
        platform::errors::InvalidArgument(
            "The rank of the input 'X' for expand_v2 op must be positive, "
            "but the value received is %d.",
            rank));
    PADDLE_ENFORCE_LE(
        rank, MAX_RANK_SUPPORTED,
        platform::errors::InvalidArgument(
            "The rank of the input 'X' for expand_v2 op must be less than "
            "or equal to %d, but the value received is %d.",
            MAX_RANK_SUPPORTED, rank));
    auto expand_shape = get_expand_shape(context);
    auto shape_size = expand_shape.size();
    PADDLE_ENFORCE_GE(
        shape_size, rank,
        platform::errors::InvalidArgument(
            "The number (%d) of elements of 'shape' for expand_v2 op must be "
            "greater than or equal to the rank (%d) of the input 'X'.",
            shape_size, rank));
    PADDLE_ENFORCE_LE(
        shape_size, MAX_RANK_SUPPORTED,
        platform::errors::InvalidArgument(
            "The number (%d) of elements of 'shape' for expand_v2 op must be "
            "less than or equal to %d.",
            shape_size, MAX_RANK_SUPPORTED));
    rank = std::max(rank, static_cast<int>(shape_size));
    switch (rank) {
      case 1:
        Expand<1>(context);
        break;
      case 2:
        Expand<2>(context);
        break;
      case 3:
        Expand<3>(context);
        break;
      case 4:
        Expand<4>(context);
        break;
      case 5:
        Expand<5>(context);
        break;
      case 6:
        Expand<6>(context);
        break;
    }
  }

 protected:
  template <int Rank>
  void Expand(const framework::ExecutionContext& context) const {
    auto* in0 = context.Input<Tensor>("X");

    auto in_dims = in0->dims();
    auto expand_shape = get_expand_shape(context);
    auto vec_in_dims = framework::vectorize<int>(in_dims);
    auto diff = expand_shape.size() - vec_in_dims.size();
    vec_in_dims.insert(vec_in_dims.begin(), diff, 1);
    std::vector<int> repeat_times(vec_in_dims.size());
    for (size_t i = 0; i < vec_in_dims.size(); ++i) {
      PADDLE_ENFORCE_NE(expand_shape[i], 0,
                        platform::errors::InvalidArgument(
                            "The expanded size cannot be zero."));
      if (i < diff) {
        PADDLE_ENFORCE_GT(
            expand_shape[i], 0,
            platform::errors::InvalidArgument(
                "The expanded size (%d) for non-existing dimensions must be "
                "positive for expand_v2 op.",
                expand_shape[i]));
        repeat_times[i] = expand_shape[i];
      } else if (expand_shape[i] > 0) {
        if (vec_in_dims[i] != 1) {
          PADDLE_ENFORCE_EQ(
              vec_in_dims[i], expand_shape[i],
              platform::errors::InvalidArgument(
                  "The value (%d) of the non-singleton dimension does not match"
                  " the corresponding value (%d) in shape for expand_v2 op.",
                  vec_in_dims[i], expand_shape[i]));
          repeat_times[i] = 1;
        } else {
          repeat_times[i] = expand_shape[i];
        }
      } else {
        PADDLE_ENFORCE_EQ(
            expand_shape[i], -1,
            platform::errors::InvalidArgument(
                "When the value in shape is negative for expand_v2 op, "
                "only -1 is supported, but the value received is %d.",
                expand_shape[i]));
        repeat_times[i] = 1;
      }
    }

    auto* out0 = context.Output<Tensor>("Out");
    Eigen::DSizes<Eigen::DenseIndex, Rank> bcast_dims;
    for (size_t i = 0; i < repeat_times.size(); ++i) {
      bcast_dims[i] = repeat_times[i];
    }

    framework::DDim new_in_dims = framework::make_ddim(vec_in_dims);
    framework::DDim out_dims(new_in_dims);
    for (size_t i = 0; i < repeat_times.size(); ++i) {
      out_dims[i] *= repeat_times[i];
    }

    out0->Resize(out_dims);
    auto x = EigenTensor<T, Rank>::From(*in0, new_in_dims);
    out0->mutable_data<T>(context.GetPlace());
    auto y = EigenTensor<T, Rank>::From(*out0, out_dims);
    auto& place =
        *context.template device_context<DeviceContext>().eigen_device();
    // use 32-bit index to speed up
    bool use_32bit_index = y.size() < Eigen::NumTraits<int>::highest();
    if (use_32bit_index) {
      EigenBroadcast<std::decay_t<decltype(place)>, T, Rank>::Eval(
          place, To32BitIndex(y), To32BitIndex(x), bcast_dims);
    } else {
      EigenBroadcast<std::decay_t<decltype(place)>, T, Rank>::Eval(place, y, x,
                                                                   bcast_dims);
    }
  }
};

template <typename DeviceContext, typename T>
class ExpandV2GradKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& context) const override {
    auto* in0 = context.Input<Tensor>("X");
    auto expand_shape = get_expand_shape(context);
    auto x_dims = in0->dims();
    auto vec_in_dims = framework::vectorize<int>(x_dims);
    auto diff = expand_shape.size() - vec_in_dims.size();
    vec_in_dims.insert(vec_in_dims.begin(), diff, 1);
    // 1. reshape_dims_vec is the broadcast parameter.
    // 2. reduce_dims_vec is the dimension parameter to compute gradients. For
    //    each dimension expanded, the gradients should be summed to original
    //    size.
    std::vector<int> repeat_times(vec_in_dims.size());
    for (size_t i = 0; i < vec_in_dims.size(); ++i) {
      if (expand_shape[i] < 0) {
        repeat_times[i] = 1;
      } else {
        repeat_times[i] = expand_shape[i] / vec_in_dims[i];
      }
    }
    std::vector<int> reshape_dims_vec;
    std::vector<int> reduce_dims_vec;
    for (size_t i = 0; i < repeat_times.size(); ++i) {
      reduce_dims_vec.push_back(reshape_dims_vec.size());
      reshape_dims_vec.push_back(repeat_times[i]);
      reshape_dims_vec.push_back(vec_in_dims[i]);
    }

    int dims = reduce_dims_vec.size();

    bool just_copy = true;
    for (size_t i = 0; i < repeat_times.size(); i++) {
      if (repeat_times[i] != 1) {
        just_copy = false;
        break;
      }
    }
    // no need reduce, just copy
    if (just_copy) {
      auto* in0 = context.Input<Tensor>(framework::GradVarName("Out"));
      auto* out0 = context.Output<Tensor>(framework::GradVarName("X"));
      out0->mutable_data<T>(context.GetPlace());
      framework::TensorCopy(*in0, context.GetPlace(), context.device_context(),
                            out0);
    } else {
      PADDLE_ENFORCE_GE(dims, 1,
                        platform::errors::InvalidArgument(
                            "The rank of the input 'Out@GRAD' for "
                            "expand_v2_grad op must be greater than or "
                            "equal to 1, but the value received is %d.",
                            dims));
      PADDLE_ENFORCE_LE(dims, MAX_RANK_SUPPORTED,
                        platform::errors::InvalidArgument(
                            "The rank of the input 'Out@GRAD' for "
                            "expand_v2_grad op must be less than or equal "
                            "to %d, but the value received is %d.",
                            MAX_RANK_SUPPORTED, dims));
      switch (dims) {
        case 1:
          ExpandBackward<1>(context, reshape_dims_vec, reduce_dims_vec);
          break;
        case 2:
          ExpandBackward<2>(context, reshape_dims_vec, reduce_dims_vec);
          break;
        case 3:
          ExpandBackward<3>(context, reshape_dims_vec, reduce_dims_vec);
          break;
        case 4:
          ExpandBackward<4>(context, reshape_dims_vec, reduce_dims_vec);
          break;
        case 5:
          ExpandBackward<5>(context, reshape_dims_vec, reduce_dims_vec);
          break;
        case 6:
          ExpandBackward<6>(context, reshape_dims_vec, reduce_dims_vec);
          break;
        default:
          PADDLE_THROW(platform::errors::InvalidArgument(
              "Only support tensor with rank being between 1 and 6. But "
              "received tensor's rank = %d.",
              dims));
      }
    }
  }

 protected:
  template <int Dims>
  void ExpandBackward(const framework::ExecutionContext& context,
                      const std::vector<int>& reshape_dims_vec,
                      const std::vector<int>& reduce_dims_vec) const {
    size_t reshape_size = reshape_dims_vec.size();
    size_t reduce_size = reduce_dims_vec.size();
    auto* in0 = context.Input<Tensor>(framework::GradVarName("Out"));
    auto* out0 = context.Output<Tensor>(framework::GradVarName("X"));
    out0->mutable_data<T>(context.GetPlace());
    auto x_grad = EigenVector<T>::Flatten(*out0);
    Eigen::DSizes<Eigen::DenseIndex, Dims * 2> reshape_dims;
    for (size_t i = 0; i < reshape_size; ++i) {
      reshape_dims[i] = reshape_dims_vec[i];
    }
    Eigen::DSizes<Eigen::DenseIndex, Dims> reduce_dims;
    for (size_t i = 0; i < reduce_size; ++i) {
      reduce_dims[i] = reduce_dims_vec[i];
    }
    auto out_grad = EigenVector<T>::Flatten(*in0);
    auto& place =
        *context.template device_context<DeviceContext>().eigen_device();
    EigenBroadcastGrad<std::decay_t<decltype(place)>, T, Dims>::Eval(
        place, x_grad, out_grad, reduce_dims, reshape_dims);
  }
};

}  // namespace operators
}  // namespace paddle