/* Copyright (c) 2022 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. */

#include "paddle/phi/kernels/funcs/cross_entropy.h"

#include "paddle/phi/backends/cpu/cpu_context.h"
#include "paddle/phi/core/utils/data_type.h"

namespace phi {
namespace funcs {

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

template <typename T>
struct HardLabelCrossEntropyCPUFunctorImpl {
  HardLabelCrossEntropyCPUFunctorImpl(DenseTensor* out,
                                      const DenseTensor* prob,
                                      const DenseTensor* labels,
                                      const int ignore_index,
                                      const int64_t axis_dim)
      : out_(out),
        prob_(prob),
        labels_(labels),
        ignore_index_(ignore_index),
        axis_dim_(axis_dim) {}

  template <typename U>
  void apply() const {
    int64_t batch_size = prob_->dims()[0];

    int64_t num_classes = prob_->dims()[1];

    const int64_t num_remain = num_classes / axis_dim_;

    const T* prob_data = prob_->template data<T>();
    T* loss_data = out_->template data<T>();

    const auto* label_data = labels_->template data<U>();
    for (int64_t i = 0; i < batch_size; ++i) {
      for (int64_t j = 0; j < num_remain; j++) {
        int64_t lbl = static_cast<int64_t>(label_data[i * num_remain + j]);
        if (lbl != ignore_index_) {
          PADDLE_ENFORCE_GE(lbl,
                            0,
                            common::errors::OutOfRange(
                                "label value should >= 0 when label "
                                "value(%f) not equal to ignore_index(%f)",
                                lbl,
                                ignore_index_));
          PADDLE_ENFORCE_LT(
              lbl,
              axis_dim_,
              common::errors::OutOfRange(
                  "label value should less than the shape of axis dimension "
                  "when label value(%f) not equal to ignore_index(%f), But "
                  "received label value as %ld and shape of axis dimension "
                  "is %ld",
                  lbl,
                  ignore_index_,
                  lbl,
                  axis_dim_));
        }
        int64_t index = i * num_classes + lbl * num_remain + j;
        int64_t loss_idx = i * num_remain + j;
        loss_data[loss_idx] =
            lbl == ignore_index_
                ? 0
                : -funcs::TolerableValue<T>()(std::log(prob_data[index]));
      }
    }
  }

 private:
  DenseTensor* out_;
  const DenseTensor* prob_;
  const DenseTensor* labels_;
  const int ignore_index_;
  const int64_t axis_dim_;
};

template <typename DeviceContext, typename T>
void CrossEntropyFunctor<DeviceContext, T>::operator()(
    const DeviceContext& dev_ctx,
    DenseTensor* out,
    const DenseTensor* prob,
    const DenseTensor* labels,
    const bool softLabel,
    const int ignore_index,
    const int64_t axis_dim) {
  if (softLabel) {
    // TODO(large-tensor): Eigen::DSizes not support int64
    PADDLE_ENFORCE_LE_INT_MAX(prob->dims()[0], "prob->dims()[0]");
    PADDLE_ENFORCE_LE_INT_MAX(prob->dims()[1], "prob->dims()[1]");
    PADDLE_ENFORCE_LE_INT_MAX(axis_dim, "axis_dim");
    const int batch_size = static_cast<const int>(prob->dims()[0]);
    const int num_classes = static_cast<const int>(prob->dims()[1]);
    const int axis_dim_int = static_cast<int>(axis_dim);
    const int num_remain = num_classes / axis_dim_int;

    Eigen::DSizes<int, 3> batch_axis_remain(
        batch_size, axis_dim_int, num_remain);
    auto in = EigenMatrix<T>::From(*prob);
    auto lbl = EigenMatrix<T>::From(*labels);
    auto loss = EigenMatrix<T>::From(*out);

    loss.device(*dev_ctx.eigen_device()) =
        -((lbl * in.log().unaryExpr(funcs::TolerableValue<T>()))
              .reshape(batch_axis_remain)
              .sum(Eigen::DSizes<int, 1>(1)));
  } else {
    HardLabelCrossEntropyCPUFunctorImpl<T> functor_impl(
        out, prob, labels, ignore_index, axis_dim);
    phi::VisitDataType(labels->dtype(), functor_impl);
  }
}

template class CrossEntropyFunctor<CPUContext, float>;
template class CrossEntropyFunctor<CPUContext, double>;

}  // namespace funcs
}  // namespace phi