backward.cc

/* 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/infermeta/backward.h"

#include "paddle/phi/common/type_traits.h"
#include "paddle/phi/kernels/funcs/axis_utils.h"

namespace phi {

void BilinearTensorProductGradInferMeta(const MetaTensor& x,
                                        const MetaTensor& y,
                                        const MetaTensor& weight,
                                        const MetaTensor& dout,
                                        MetaTensor* dx,
                                        MetaTensor* dy,
                                        MetaTensor* dweight,
                                        MetaTensor* dbias) {
  auto x_dims = x.dims();
  auto y_dims = y.dims();
  auto weight_dims = weight.dims();
  auto out_dims = dout.dims();

  PADDLE_ENFORCE_EQ(
      out_dims.size(),
      2UL,
      errors::InvalidArgument("The input(Out@GRAD) must be a 2D Tensor."));
  PADDLE_ENFORCE_EQ(
      x_dims[0],
      out_dims[0],
      errors::InvalidArgument(
          "The first dimension(batch_size) of input(Out@GRAD) must be "
          "equal to the first dimension of the Input(X)."));
  PADDLE_ENFORCE_EQ(
      weight_dims[0],
      out_dims[1],
      errors::InvalidArgument(
          "The second dimension of input(Out@GRAD) must be equal to "
          "the third dimension of the Input(Weight)."));

  if (dx) {
    dx->set_dims(x_dims);
    dx->set_dtype(x.dtype());
  }
  if (dy) {
    dy->set_dims(y_dims);
    dy->set_dtype(y.dtype());
  }
  if (dweight) {
    dweight->set_dims(weight_dims);
    dweight->set_dtype(weight.dtype());
  }
  if (dbias) {
    dbias->set_dims({1, out_dims[1]});
    dbias->set_dtype(dout.dtype());
  }
}

void ChannelShuffleGradInferMeta(const MetaTensor& out_grad,
                                 int groups,
                                 const std::string& data_format,
                                 MetaTensor* x_grad) {
  auto do_dims = out_grad.dims();
  PADDLE_ENFORCE_EQ(do_dims.size(),
                    4,
                    phi::errors::InvalidArgument(
                        "Input should be a 4-D tensor of format [N, C, H, W] "
                        "or [N, H, W, C], but got %u.",
                        do_dims.size()));
  auto dx_dims = do_dims;
  x_grad->set_dims(dx_dims);
  x_grad->set_dtype(out_grad.dtype());
}

void ConvTransposeGradInferMeta(const MetaTensor& x,
                                const MetaTensor& filter,
                                const MetaTensor& dout,
                                const std::vector<int>& strides,
                                const std::vector<int>& paddings,
                                const std::vector<int>& output_padding,
                                const std::vector<int>& output_size,
                                const std::string& padding_algorithm,
                                int groups,
                                const std::vector<int>& dilations,
                                const std::string& data_format,
                                MetaTensor* dx,
                                MetaTensor* dfilter) {
  GeneralBinaryGradInferMeta(x, filter, dx, dfilter);
}

void Conv2dTransposeDoubleGradInferMeta(const MetaTensor& x,
                                        const MetaTensor& filter,
                                        const MetaTensor& dout,
                                        const MetaTensor& ddx,
                                        const MetaTensor& ddfilter,
                                        const std::vector<int>& strides,
                                        const std::vector<int>& paddings,
                                        const std::vector<int>& output_padding,
                                        const std::vector<int>& output_size,
                                        const std::string& padding_algorithm,
                                        int groups,
                                        const std::vector<int>& dilations,
                                        const std::string& data_format,
                                        MetaTensor* dx,
                                        MetaTensor* dfilter,
                                        MetaTensor* ddout) {
  GeneralBinaryGradInferMeta(x, filter, dx, dfilter);

  if (ddout) {
    ddout->share_meta(dout);
  }
}

void CrossEntropyWithSoftmaxGradInferMeta(const MetaTensor& label,
                                          const MetaTensor& softmax,
                                          const MetaTensor& loss_grad,
                                          bool soft_label,
                                          bool use_softmax,
                                          bool numeric_stable_mode,
                                          int ignore_index,
                                          int axis,
                                          MetaTensor* logits_grad,
                                          MetaConfig config) {
  auto softmax_dims = softmax.dims();
  auto labels_dims = label.dims();
  auto softmax_rank = softmax_dims.size();
  PADDLE_ENFORCE_GE(axis,
                    -softmax_rank,
                    phi::errors::InvalidArgument(
                        "Attr(axis) value should be in range [-R, R-1], "
                        "R is the rank of Input(Logits)."));
  PADDLE_ENFORCE_LT(axis,
                    softmax_rank,
                    phi::errors::InvalidArgument(
                        "Attr(axis) value should be in range [-R, R-1], "
                        "R is the rank of Input(Logits)."));

  axis = phi::funcs::CanonicalAxis(axis, softmax_rank);
  for (int i = 0; i < softmax_rank; i++) {
    if (i != axis) {
      if (config.is_runtime || (softmax_dims[i] > 0 && labels_dims[i] > 0)) {
        PADDLE_ENFORCE_EQ(
            softmax_dims[i],
            labels_dims[i],
            phi::errors::InvalidArgument(
                "Input(Logits) and Input(Label) should in same shape in "
                "dimensions except axis."));
      }
    }
  }

  if (soft_label) {
    if (config.is_runtime ||
        (softmax_dims[axis] > 0 && labels_dims[axis] > 0)) {
      PADDLE_ENFORCE_EQ(softmax_dims[axis],
                        labels_dims[axis],
                        phi::errors::InvalidArgument(
                            "If Attr(soft_label) == true, "
                            "the axis dimension of "
                            "Input(X) and Input(Label) should be equal."));
    }
  } else {
    if (config.is_runtime || labels_dims[axis] > 0) {
      PADDLE_ENFORCE_EQ(
          labels_dims[axis],
          1UL,
          phi::errors::InvalidArgument("If Attr(soft_label) == false, "
                                       "the axis dimension of "
                                       "Input(Label) should be 1."));
    }
  }

  logits_grad->set_dims(softmax.dims());
  logits_grad->set_dtype(softmax.dtype());
}

void DeformableConvGradInferMeta(const MetaTensor& x,
                                 const MetaTensor& offset,
                                 const MetaTensor& filter,
                                 paddle::optional<const MetaTensor&> mask,
                                 const MetaTensor& out_grad,
                                 const std::vector<int>& strides,
                                 const std::vector<int>& paddings,
                                 const std::vector<int>& dilations,
                                 int deformable_groups,
                                 int groups,
                                 int im2col_step,
                                 MetaTensor* dx,
                                 MetaTensor* offset_grad,
                                 MetaTensor* filter_grad,
                                 MetaTensor* mask_grad) {
  GeneralTernaryGradInferMeta(x, offset, filter, dx, offset_grad, filter_grad);
  if (mask) {
    UnchangedInferMeta(mask.get(), mask_grad);
  }
}

void GatherNdGradInferMeta(const MetaTensor& x,
                           const MetaTensor& index,
                           const MetaTensor& out_grad,
                           MetaTensor* x_grad) {
  const auto& dtype = out_grad.dtype();
  x_grad->set_dims(x.dims());
  x_grad->share_lod(x);
  x_grad->set_dtype(dtype);
}

void GeneralBinaryGradInferMeta(const MetaTensor& x,
                                const MetaTensor& y,
                                MetaTensor* dx,
                                MetaTensor* dy) {
  if (dx) {
    dx->share_meta(x);
  }
  if (dy) {
    dy->share_meta(y);
  }
}

void GeneralTernaryGradInferMeta(const MetaTensor& x,
                                 const MetaTensor& y,
                                 const MetaTensor& z,
                                 MetaTensor* dx,
                                 MetaTensor* dy,
                                 MetaTensor* dz) {
  if (dx) {
    dx->share_meta(x);
  }
  if (dy) {
    dy->share_meta(y);
  }
  if (dz) {
    dz->share_meta(z);
  }
}
void GeneralQuaternaryGradInferMeta(const MetaTensor& x,
                                    const MetaTensor& y,
                                    const MetaTensor& z,
                                    const MetaTensor& k,
                                    MetaTensor* dx,
                                    MetaTensor* dy,
                                    MetaTensor* dz,
                                    MetaTensor* dk) {
  if (dx) {
    dx->share_meta(x);
  }
  if (dy) {
    dy->share_meta(y);
  }
  if (dz) {
    dz->share_meta(z);
  }
  if (dk) {
    dk->share_meta(k);
  }
}

void GeneralQuinaryGradInferMeta(const MetaTensor& x,
                                 const MetaTensor& y,
                                 const MetaTensor& z,
                                 const MetaTensor& k,
                                 const MetaTensor& l,
                                 MetaTensor* dx,
                                 MetaTensor* dy,
                                 MetaTensor* dz,
                                 MetaTensor* dk,
                                 MetaTensor* dl) {
  if (dx) {
    dx->share_meta(x);
  }
  if (dy) {
    dy->share_meta(y);
  }
  if (dz) {
    dz->share_meta(z);
  }
  if (dk) {
    dk->share_meta(k);
  }
  if (dl) {
    dl->share_meta(l);
  }
}

void GeneralUnaryGradInferMeta(const MetaTensor& x, MetaTensor* dx) {
  if (dx) {
    dx->share_meta(x);
  }
}

void GumbelSoftmaxGradInferMeta(const MetaTensor& out,
                                const MetaTensor& dout,
                                int axis,
                                MetaTensor* dx) {
  PADDLE_ENFORCE_EQ(
      out.dims(),
      dout.dims(),
      errors::InvalidArgument(
          "Input(Out) and its gradients should have the same shape."));

  dx->share_meta(dout);
}

void KernelWithXShapeInferMeta(const MetaTensor& xshape, MetaTensor* dx) {
  auto xshape_dims = xshape.dims();
  auto x_dims = phi::slice_ddim(xshape_dims, 1, xshape_dims.size());
  dx->set_dims(x_dims);
  dx->share_lod(xshape);
}

void MaxPoolWithIndexGradInferMeta(const MetaTensor& x,
                                   const MetaTensor& mask,
                                   const MetaTensor& dout,
                                   const std::vector<int>& kernel_size,
                                   const std::vector<int>& strides,
                                   const std::vector<int>& paddings,
                                   bool global_pooling,
                                   bool adaptive,
                                   MetaTensor* dx) {
  dx->share_meta(x);
}

void MeshgridGradInferMeta(const std::vector<const MetaTensor*>& inputs,
                           const std::vector<const MetaTensor*>& outputs_grad,
                           std::vector<MetaTensor*> inputs_grad) {
  PADDLE_ENFORCE_GT(outputs_grad.size(),
                    1,
                    errors::InvalidArgument(
                        "Number of Inputs(Out@Grad) should be larger than 1."
                        "But received Inputs(Out@Grad)' size = %d .",
                        outputs_grad.size()));
  for (size_t i = 0; i < inputs.size(); i++) {
    inputs_grad[i]->share_meta(*inputs[i]);
  }
}

void MultiDotGradInferMeta(const std::vector<const MetaTensor*>& x,
                           const MetaTensor& out_grad,
                           std::vector<MetaTensor*> x_grad) {
  PADDLE_ENFORCE_EQ(
      x.size(),
      x_grad.size(),
      errors::InvalidArgument(
          "Number of Inputs(X) should be equal with Outputs(X@Grad)."
          "But received Inputs(X)' size = %d , Outputs(X@Grad)' size = %d.",
          x.size(),
          x_grad.size()));
  for (size_t i = 0; i < x.size(); i++) {
    if (x_grad[i] != nullptr) {
      x_grad[i]->set_dims(x[i]->dims());
      x_grad[i]->share_lod(*x[i]);
    }
  }
}

void MultiplexGradInferMeta(const MetaTensor& ids,
                            const MetaTensor& out_grad,
                            std::vector<MetaTensor*> ins_grad) {
  PADDLE_ENFORCE_NE(
      ins_grad.empty(),
      true,
      errors::InvalidArgument("Output(X@Grad) should not be null."));
  auto dout_dim = out_grad.dims();
  for (auto in_grad : ins_grad) {
    in_grad->set_dims(dout_dim);
  }
}

void NllLossGradInferMeta(const MetaTensor& x,
                          const MetaTensor& label,
                          paddle::optional<const MetaTensor&> weight,
                          const MetaTensor& total_weight,
                          const MetaTensor& out_grad,
                          int64_t ignore_index,
                          const std::string& reduction,
                          MetaTensor* dx,
                          MetaConfig config) {
  const auto& x_dims = x.dims();
  const auto& label_dims = label.dims();
  const auto& dout_dims = out_grad.dims();
  bool contain_unknown_dim =
      phi::contain_unknown_dim(x_dims) || phi::contain_unknown_dim(dout_dims);
  bool check = config.is_runtime || !contain_unknown_dim;

  if (check) {
    auto batch_size = x_dims[0];
    if (x_dims.size() == 2) {
      PADDLE_ENFORCE_EQ(dout_dims.size(),
                        1,
                        phi::errors::InvalidArgument(
                            "The dimensions of Input(Out@Grad) must be 1"));
      if (reduction == "none") {
        PADDLE_ENFORCE_EQ(
            dout_dims[0],
            batch_size,
            phi::errors::InvalidArgument(
                "The unreduced size ofInput(Out@Grad) must be the "
                "same as batch_size."));
      } else {
        PADDLE_ENFORCE_EQ(dout_dims[0],
                          1,
                          phi::errors::InvalidArgument(
                              "The reduced size of Input(Out@Grad) must be 1"));
      }
    } else if (x_dims.size() == 4) {
      if (reduction == "none") {
        PADDLE_ENFORCE_EQ(
            dout_dims.size(),
            3,
            phi::errors::InvalidArgument(
                "The dimensions of Input(Out@Grad) must be 3,But got [%s].",
                dout_dims.size()));
        PADDLE_ENFORCE_EQ(dout_dims[0] == label_dims[0] &&
                              dout_dims[1] == label_dims[1] &&
                              dout_dims[2] == label_dims[2],
                          true,
                          phi::errors::InvalidArgument(
                              "The dimensions of Input(Out@Grad) must be match "
                              "to Input(Label) dimensions."));
      } else {
        PADDLE_ENFORCE_EQ(dout_dims[0],
                          1,
                          phi::errors::InvalidArgument(
                              "The reduced size of Input(Out@Grad) must be 1"));
      }
    }
  }

  if (dx) {
    dx->set_dims(x_dims);
    dx->set_dtype(x.dtype());
  }
}

void PixelUnshuffleGradInferMeta(const MetaTensor& out_grad,
                                 int downscale_factor,
                                 const std::string& data_format,
                                 MetaTensor* x_grad) {
  auto do_dims = out_grad.dims();
  PADDLE_ENFORCE_EQ(do_dims.size(),
                    4,
                    phi::errors::InvalidArgument(
                        "Input should be a 4-D tensor of format [N, C, H, W] "
                        "or [N, H, W, C], but got %u.",
                        do_dims.size()));

  const bool channel_last = (data_format == "NHWC");

  auto dx_dims = do_dims;
  dx_dims[0] = do_dims[0];

  if (!channel_last) {
    dx_dims[1] = do_dims[1] / (downscale_factor * downscale_factor);
    dx_dims[2] = do_dims[2] * downscale_factor;
    dx_dims[3] = do_dims[3] * downscale_factor;
  } else {
    dx_dims[1] = do_dims[1] * downscale_factor;
    dx_dims[2] = do_dims[2] * downscale_factor;
    dx_dims[3] = do_dims[3] / (downscale_factor * downscale_factor);
  }
  x_grad->set_dims(dx_dims);
  x_grad->set_dtype(out_grad.dtype());
}

void PoolGradInferMeta(const MetaTensor& x,
                       const MetaTensor& out,
                       const MetaTensor& dout,
                       const std::vector<int>& kernel_size,
                       const std::vector<int>& strides,
                       const std::vector<int>& paddings,
                       bool ceil_mode,
                       bool exclusive,
                       const std::string& data_format,
                       const std::string& pooling_type,
                       bool global_pooling,
                       bool adaptive,
                       const std::string& padding_algorithm,
                       MetaTensor* dx) {
  dx->share_meta(x);
}

void PsroiPoolGradInferMeta(const MetaTensor& x,
                            const MetaTensor& rois,
                            paddle::optional<const MetaTensor&> rois_num,
                            const MetaTensor& dout,
                            int pooled_height,
                            int pooled_width,
                            int output_channels,
                            float spatial_scale,
                            MetaTensor* dx) {
  dx->share_meta(x);
}

void RealAndImagGradInferMeta(const MetaTensor& out_grad, MetaTensor* dx) {
  dx->set_dims(out_grad.dims());
  dx->set_dtype(dtype::ToComplex(out_grad.dtype()));
  dx->set_layout(out_grad.layout());
}

void ReshapeDoubleGradInferMeta(const MetaTensor& out_grad,
                                const MetaTensor& x_grad_grad,
                                MetaTensor* out_grad_grad) {
  if (out_grad_grad != nullptr) {
    out_grad_grad->share_dims(out_grad);
  }
}

void ScatterGradInferMeta(const MetaTensor& index,
                          const MetaTensor& updates,
                          const MetaTensor& out_grad,
                          bool overwrite,
                          MetaTensor* x_grad,
                          MetaTensor* updates_grad) {
  const auto& dtype = out_grad.dtype();
  if (updates_grad) {
    updates_grad->set_dims(updates.dims());
    updates_grad->set_dtype(dtype);
  }

  if (x_grad) {
    x_grad->set_dims(out_grad.dims());
    x_grad->set_dtype(dtype);
  }
}

void ScatterNdAddGradInferMeta(const MetaTensor& index,
                               const MetaTensor& updates,
                               const MetaTensor& out_grad,
                               MetaTensor* x_grad,
                               MetaTensor* updates_grad) {
  const auto& dtype = out_grad.dtype();
  if (updates_grad) {
    updates_grad->set_dims(updates.dims());
    updates_grad->set_dtype(dtype);
  }

  if (x_grad) {
    x_grad->set_dims(out_grad.dims());
    x_grad->set_dtype(dtype);
  }
}

void StackGradInferMeta(const MetaTensor& out_grad,
                        int axis,
                        std::vector<MetaTensor*> x_grad) {
  auto dy_dim = out_grad.dims();
  int rank = dy_dim.size();
  PADDLE_ENFORCE_GE(
      axis,
      -rank,
      phi::errors::InvalidArgument(
          "Attr(axis) must be inside [-rank, rank), where rank = %d, "
          "but received axis is:%d.",
          rank,
          axis));
  PADDLE_ENFORCE_LT(
      axis,
      rank,
      phi::errors::InvalidArgument(
          "Attr(axis) must be inside [-rank, rank), where rank = %d, "
          "but received axis is:%d.",
          rank,
          axis));

  if (axis < 0) axis += rank;
  PADDLE_ENFORCE_LE(
      x_grad.size(),
      static_cast<size_t>(dy_dim[axis]),
      phi::errors::InvalidArgument(
          "Number of Outputs(X@Grad) should be less than or equal to dy dim "
          "at axis, but received outputs size is:%d, dy dims is:%d.",
          x_grad.size(),
          static_cast<size_t>(dy_dim[axis])));

  auto vec = phi::vectorize<int>(dy_dim);
  vec.erase(vec.begin() + axis);

  for (size_t i = 0; i < x_grad.size(); ++i) {
    if (x_grad[i]) {
      x_grad[i]->set_dims(phi::make_ddim(vec));
      x_grad[i]->set_dtype(out_grad.dtype());
    }
  }
}

}  // namespace phi