locally_connected_op_impl.h

// locally_connected_impl.h is the templated implementation of the
// locally_connected.h file.

#ifndef CAFFE2_OPERATORS_LOCALLY_CONNECTED_OP_IMPL_H_
#define CAFFE2_OPERATORS_LOCALLY_CONNECTED_OP_IMPL_H_

#include <vector>

#include "caffe2/core/context.h"
#include "caffe2/core/flags.h"
#include "caffe2/core/logging.h"
#include "caffe2/core/operator.h"
#include "caffe2/operators/conv_pool_op_base.h"
#include "caffe2/operators/locally_connected_op.h"
#include "caffe2/utils/math.h"

namespace caffe2 {

template <typename T, class Context>
bool LocallyConnectedOp<T, Context>::RunOnDeviceWithOrderNCHW() {
  const auto& X = Input(INPUT);
  const auto& filter = Input(FILTER);
  auto* Y = Output(0);
  const int image_ndim = X.dim() - 2;
  CAFFE_ENFORCE_EQ(X.dim() + image_ndim, filter.dim());
  lc_op_util::ShapeParams shape;
  shape.N = X.dim32(0);
  shape.C = X.dim32(1);
  shape.M = filter.dim32(image_ndim);
  CAFFE_ENFORCE(
      shape.C == filter.dim32(image_ndim + 1) * group_,
      "Locally Connected op: input channels does not match: "
      "# of input channels ",
      shape.C,
      " is not equal to kernel channels * group:",
      filter.dim32(image_ndim + 1),
      "*",
      group_);
  CAFFE_ENFORCE_EQ(
      shape.M % group_,
      0,
      "The number of output channels is not divisible by group.");

  ConvPoolOpBase<Context>::SetOutputSize(X, Y, shape.M);
  shape.input_image_size = GetDimsSize(X);
  shape.output_image_size = GetDimsSize(*Y);
  const std::vector<int> output_image_dims = GetDims(*Y);
  for (int i = 0; i < image_ndim; ++i) {
    CAFFE_ENFORCE_EQ(output_image_dims[i], filter.dim32(i));
  }

  int kernel_dims_size = 1;
  for (std::size_t i = 0; i < kernel_.size(); ++i) {
    CAFFE_ENFORCE_EQ(filter.dim32(i + image_ndim + 2), kernel_[i]);
    kernel_dims_size *= kernel_[i];
  }

  shape.X_dims.assign(X.sizes().cbegin() + 1, X.sizes().cend());
  shape.kernel_size = shape.C / group_ * kernel_dims_size;
  lc_op_util::SetColumnBufferShape(
      shape.N,
      shape.kernel_size,
      shape.output_image_size,
      output_image_dims,
      order_,
      &shape.column_slice_dims,
      &shape.column_dims,
      &shape.column_transposed_dims,
      &shape.column_axes);
  lc_op_util::SetYBufferShape(
      shape.N,
      shape.M,
      shape.output_image_size,
      order_,
      &shape.Y_dims,
      &shape.Y_transposed_dims,
      &shape.Y_axes);

  const T* X_data = X.template data<T>();
  const T* filter_data = filter.template data<T>();
  const T* bias_data = nullptr;
  if (InputSize() == 3) {
    const auto& bias = Input(BIAS);
    CAFFE_ENFORCE_EQ(bias.dim(), image_ndim + 1);
    for (int i = 0; i < image_ndim; ++i) {
      CAFFE_ENFORCE_EQ(bias.dim32(i), output_image_dims[i]);
    }
    CAFFE_ENFORCE_EQ(bias.dim32(image_ndim), shape.M);
    bias_data = bias.template data<T>();
    ConvPoolOpBase<Context>::template SetBiasMultiplier<T>(
        shape.N, &bias_multiplier_);
  }
  T* Y_data = Y->template mutable_data<T>();

  RunOnDeviceWithOrderNCHWImpl(
      shape,
      X_data,
      filter_data,
      bias_data,
      Y_data,
      &column_buffer_,
      &column_transposed_buffer_,
      &Y_transposed_buffer_);

  return true;
}

template <typename T, class Context>
bool LocallyConnectedOp<T, Context>::RunOnDeviceWithOrderNHWC() {
  const auto& X = Input(INPUT);
  const auto& filter = Input(FILTER);
  auto* Y = Output(0);
  CAFFE_ENFORCE_EQ(
      kernel_.size(),
      2,
      "Only 2d locally connected op is supported for NHWC storage type.");
  const int image_ndim = X.dim() - 2;
  CAFFE_ENFORCE_EQ(X.dim() + image_ndim, filter.dim());
  lc_op_util::ShapeParams shape;
  shape.N = X.dim32(0);
  shape.C = X.dim32(3);
  shape.X_dims = {X.dim32(1), X.dim32(2), X.dim32(3)};
  shape.M = filter.dim32(image_ndim);
  CAFFE_ENFORCE_EQ(filter.dim32(image_ndim + 1), kernel_h());
  CAFFE_ENFORCE_EQ(filter.dim32(image_ndim + 2), kernel_w());
  CAFFE_ENFORCE_EQ(filter.dim32(image_ndim + 3), shape.C);
  ConvPoolOpBase<Context>::SetOutputSize(X, Y, shape.M);

  shape.input_image_size = GetDimsSize(X);
  shape.output_image_size = GetDimsSize(*Y);
  const std::vector<int> output_image_dims = GetDims(*Y);
  for (int i = 0; i < image_ndim; ++i) {
    CAFFE_ENFORCE_EQ(output_image_dims[i], filter.dim32(i));
  }

  shape.kernel_size = kernel_h() * kernel_w() * shape.C;
  lc_op_util::SetColumnBufferShape(
      shape.N,
      shape.kernel_size,
      shape.output_image_size,
      output_image_dims,
      order_,
      &shape.column_slice_dims,
      &shape.column_dims,
      &shape.column_transposed_dims,
      &shape.column_axes);
  lc_op_util::SetYBufferShape(
      shape.N,
      shape.M,
      shape.output_image_size,
      order_,
      &shape.Y_dims,
      &shape.Y_transposed_dims,
      &shape.Y_axes);

  const T* X_data = X.template data<T>();
  const T* filter_data = filter.template data<T>();
  const T* bias_data = nullptr;
  if (InputSize() == 3) {
    const auto& bias = Input(BIAS);
    CAFFE_ENFORCE_EQ(bias.dim(), image_ndim + 1);
    for (int i = 0; i < image_ndim; ++i) {
      CAFFE_ENFORCE_EQ(bias.dim32(i), output_image_dims[i]);
    }
    CAFFE_ENFORCE_EQ(bias.dim32(image_ndim), shape.M);
    bias_data = bias.template data<T>();
    ConvPoolOpBase<Context>::template SetBiasMultiplier<T>(
        shape.N, &bias_multiplier_);
  }
  T* Y_data = Y->template mutable_data<T>();

  RunOnDeviceWithOrderNHWCImpl(
      shape,
      X_data,
      filter_data,
      bias_data,
      Y_data,
      &column_buffer_,
      &column_transposed_buffer_,
      &Y_transposed_buffer_);

  return true;
}

template <typename T, class Context>
void LocallyConnectedOp<T, Context>::RunOnDeviceWithOrderNCHWImpl(
    const lc_op_util::ShapeParams& shape,
    const T* X_data,
    const T* filter_data,
    const T* bias_data,
    T* Y_data,
    Tensor* column_buffer,
    Tensor* column_transposed_buffer,
    Tensor* Y_transposed_buffer) {
  const int input_stride = shape.C / group_ * shape.input_image_size;
  const int column_stride = shape.kernel_size * shape.output_image_size;
  column_buffer->Resize(shape.column_dims);
  column_transposed_buffer->Resize(shape.column_transposed_dims);
  Y_transposed_buffer->Resize(shape.Y_transposed_dims);
  T* column_buffer_data = column_buffer->template mutable_data<T>();
  T* Y_transposed_buffer_data = Y_transposed_buffer->template mutable_data<T>();

  for (int image_id = 0; image_id < shape.N; ++image_id) {
    for (int group_id = 0; group_id < group_; ++group_id) {
      if (kernel_.size() == 2) {
        math::Im2Col<T, Context, StorageOrder::NCHW>(
            shape.C / group_,
            shape.X_dims[1],
            shape.X_dims[2],
            kernel_h(),
            kernel_w(),
            dilation_h(),
            dilation_w(),
            pad_t(),
            pad_l(),
            pad_b(),
            pad_r(),
            stride_h(),
            stride_w(),
            X_data + group_id * input_stride,
            column_buffer_data + group_id * column_stride,
            &context_);
      } else {
        math::Im2ColNd<T, Context, StorageOrder::NCHW>(
            kernel_.size(),
            shape.C * shape.input_image_size,
            column_stride,
            shape.X_dims.data(),
            shape.column_slice_dims.data(),
            kernel_.data(),
            stride_.data(),
            dilation_.data(),
            pads_.data(),
            X_data + group_id * input_stride,
            column_buffer_data + group_id * column_stride,
            &context_);
      }
    }
    X_data += input_stride * group_;
    column_buffer_data += column_stride * group_;
  }
  math::Transpose(
      shape.column_dims.size(),
      shape.column_dims.data(),
      shape.column_axes.data(),
      column_buffer->template data<T>(),
      column_transposed_buffer->template mutable_data<T>(),
      &context_);
  math::GemmStridedBatched(
      CblasNoTrans,
      CblasNoTrans,
      shape.output_image_size * group_,
      shape.M / group_,
      shape.N,
      shape.kernel_size,
      1.0f,
      filter_data,
      shape.M / group_ * shape.kernel_size,
      column_transposed_buffer->template data<T>(),
      shape.kernel_size * shape.N,
      0.0f,
      Y_transposed_buffer_data,
      shape.M / group_ * shape.N,
      &context_);
  if (bias_data != nullptr) {
    math::Gemm<T, Context>(
        CblasNoTrans,
        CblasNoTrans,
        shape.output_image_size * shape.M,
        shape.N,
        1,
        1.0,
        bias_data,
        bias_multiplier_.template data<T>(),
        1.0,
        Y_transposed_buffer_data,
        &context_);
  }
  math::Transpose(
      shape.Y_transposed_dims.size(),
      shape.Y_transposed_dims.data(),
      shape.Y_axes.data(),
      Y_transposed_buffer_data,
      Y_data,
      &context_);
}

template <typename T, class Context>
void LocallyConnectedOp<T, Context>::RunOnDeviceWithOrderNHWCImpl(
    const lc_op_util::ShapeParams& shape,
    const T* X_data,
    const T* filter_data,
    const T* bias_data,
    T* Y_data,
    Tensor* column_buffer,
    Tensor* column_transposed_buffer,
    Tensor* Y_transposed_buffer) {
  const int input_stride = shape.C * shape.input_image_size;
  const int column_stride = shape.kernel_size * shape.output_image_size;
  column_buffer->Resize(shape.column_dims);
  column_transposed_buffer->Resize(shape.column_transposed_dims);
  Y_transposed_buffer->Resize(shape.Y_transposed_dims);
  T* column_buffer_data = column_buffer->template mutable_data<T>();
  T* Y_transposed_buffer_data = Y_transposed_buffer->template mutable_data<T>();
  for (int image_id = 0; image_id < shape.N; ++image_id) {
    math::Im2Col<T, Context, StorageOrder::NHWC>(
        shape.C,
        shape.X_dims[0],
        shape.X_dims[1],
        kernel_h(),
        kernel_w(),
        dilation_h(),
        dilation_w(),
        pad_t(),
        pad_l(),
        pad_b(),
        pad_r(),
        stride_h(),
        stride_w(),
        X_data + image_id * input_stride,
        column_buffer_data + image_id * column_stride,
        &context_);
  }
  math::Transpose(
      shape.column_dims.size(),
      shape.column_dims.data(),
      shape.column_axes.data(),
      column_buffer->template data<T>(),
      column_transposed_buffer->template mutable_data<T>(),
      &context_);
  math::GemmStridedBatched(
      CblasNoTrans,
      CblasTrans,
      shape.output_image_size,
      shape.N,
      shape.M,
      shape.kernel_size,
      1.0f,
      column_transposed_buffer->template data<T>(),
      shape.N * shape.kernel_size,
      filter_data,
      shape.kernel_size * shape.M,
      0.0f,
      Y_transposed_buffer_data,
      shape.N * shape.M,
      &context_);
  math::Transpose(
      shape.Y_transposed_dims.size(),
      shape.Y_transposed_dims.data(),
      shape.Y_axes.data(),
      Y_transposed_buffer_data,
      Y_data,
      &context_);
  if (bias_data != nullptr) {
    math::Gemm<T, Context>(
        CblasNoTrans,
        CblasNoTrans,
        shape.N,
        shape.output_image_size * shape.M,
        1,
        1.0f,
        bias_multiplier_.template data<T>(),
        bias_data,
        1.0f,
        Y_data,
        &context_);
  }
}

template <typename T, class Context>
bool LocallyConnectedGradientOp<T, Context>::RunOnDeviceWithOrderNCHW() {
  const auto& X = Input(INPUT);
  const auto& filter = Input(FILTER);
  const auto& dY = Input(OUTPUT_GRAD);

  const int image_ndim = X.dim() - 2;
  CAFFE_ENFORCE_EQ(X.dim() + image_ndim, filter.dim());

  lc_op_util::ShapeParams shape;
  shape.N = X.dim32(0);
  shape.C = X.dim32(1);
  shape.M = filter.dim32(image_ndim);
  CAFFE_ENFORCE_EQ(filter.dim32(image_ndim + 1) * group_, shape.C);
  CAFFE_ENFORCE_EQ(shape.M % group_, 0);

  const std::vector<int> input_image_dims = GetDims(X);
  shape.input_image_size = GetDimsSize(X);
  const std::vector<int> output_image_dims = GetDims(dY);
  shape.output_image_size = GetDimsSize(dY);
  for (int i = 0; i < image_ndim; ++i) {
    CAFFE_ENFORCE_EQ(output_image_dims[i], filter.dim32(i));
  }
  ConvPoolOpBase<Context>::ComputePads(input_image_dims);

  int kernel_dims_size = 1;
  for (std::size_t i = 0; i < kernel_.size(); ++i) {
    CAFFE_ENFORCE_EQ(filter.dim32(i + image_ndim + 2), kernel_[i]);
    kernel_dims_size *= kernel_[i];
  }

  shape.X_dims.assign(X.sizes().cbegin() + 1, X.sizes().cend());
  shape.kernel_size = shape.C / group_ * kernel_dims_size;
  lc_op_util::SetColumnBufferShape(
      shape.N,
      shape.kernel_size,
      shape.output_image_size,
      output_image_dims,
      order_,
      &shape.column_slice_dims,
      &shape.column_dims,
      &shape.column_transposed_dims,
      &shape.column_axes);
  lc_op_util::SetYBufferShape(
      shape.N,
      shape.M,
      shape.output_image_size,
      order_,
      &shape.Y_dims,
      &shape.Y_transposed_dims,
      &shape.Y_axes);

  auto* dfilter = Output(FILTER_GRAD, filter.sizes(), at::dtype<T>());
  const T* X_data = X.template data<T>();
  const T* filter_data = filter.template data<T>();
  const T* dY_data = dY.template data<T>();
  T* dfilter_data = dfilter->template mutable_data<T>();
  T* dX_data = nullptr;
  T* dbias_data = nullptr;
  if (OutputSize() == 3 || (no_bias_ && OutputSize() == 2)) {
    auto* dX = Output(
        no_bias_ ? BIAS_OR_INPUT_GRAD : INPUT_GRAD, X.sizes(), at::dtype<T>());
    dX_data = dX->template mutable_data<T>();
  }
  if (!no_bias_) {
    std::vector<int64_t> dbias_dims;
    std::copy(
        output_image_dims.cbegin(),
        output_image_dims.cend(),
        std::back_inserter(dbias_dims));
    dbias_dims.push_back(shape.M);
    auto* dbias = Output(BIAS_OR_INPUT_GRAD, dbias_dims, at::dtype<T>());
    ConvPoolOpBase<Context>::template SetBiasMultiplier<T>(
        shape.N, &bias_multiplier_);
    dbias_data = dbias->template mutable_data<T>();
  }
  RunOnDeviceWithOrderNCHWImpl(
      shape,
      X_data,
      filter_data,
      dY_data,
      dfilter_data,
      dX_data,
      dbias_data,
      &column_buffer_,
      &column_transposed_buffer_,
      &dY_transposed_buffer_);

  return true;
}

template <typename T, class Context>
bool LocallyConnectedGradientOp<T, Context>::RunOnDeviceWithOrderNHWC() {
  const auto& X = Input(INPUT);
  const auto& filter = Input(FILTER);
  const auto& dY = Input(OUTPUT_GRAD);

  CAFFE_ENFORCE_EQ(
      kernel_.size(),
      2,
      "Only 2d locally connected op is supported for NHWC storage type.");
  const int image_ndim = X.dim() - 2;
  CAFFE_ENFORCE_EQ(X.dim() + image_ndim, filter.dim());
  lc_op_util::ShapeParams shape;
  shape.N = X.dim32(0);
  shape.C = X.dim32(3);
  shape.X_dims = {X.dim32(1), X.dim32(2), X.dim32(3)};
  shape.M = filter.dim32(image_ndim);
  CAFFE_ENFORCE_EQ(filter.dim32(image_ndim + 1), kernel_h());
  CAFFE_ENFORCE_EQ(filter.dim32(image_ndim + 2), kernel_w());
  CAFFE_ENFORCE_EQ(filter.dim32(image_ndim + 3), shape.C);
  const std::vector<int> input_image_dims = {X.dim32(1), X.dim32(2)};
  ConvPoolOpBase<Context>::ComputePads(input_image_dims);

  shape.input_image_size = GetDimsSize(X);
  shape.output_image_size = GetDimsSize(dY);
  const std::vector<int> output_image_dims = GetDims(dY);
  for (int i = 0; i < image_ndim; ++i) {
    CAFFE_ENFORCE_EQ(output_image_dims[i], filter.dim32(i));
  }

  shape.kernel_size = kernel_h() * kernel_w() * shape.C;
  lc_op_util::SetColumnBufferShape(
      shape.N,
      shape.kernel_size,
      shape.output_image_size,
      output_image_dims,
      order_,
      &shape.column_slice_dims,
      &shape.column_dims,
      &shape.column_transposed_dims,
      &shape.column_axes);
  lc_op_util::SetYBufferShape(
      shape.N,
      shape.M,
      shape.output_image_size,
      order_,
      &shape.Y_dims,
      &shape.Y_transposed_dims,
      &shape.Y_axes);

  auto* dfilter = Output(FILTER_GRAD, filter.sizes(), at::dtype<T>());
  const T* X_data = X.template data<T>();
  const T* filter_data = filter.template data<T>();
  const T* dY_data = dY.template data<T>();
  T* dfilter_data = dfilter->template mutable_data<T>();
  T* dX_data = nullptr;
  T* dbias_data = nullptr;
  if (OutputSize() == 3 || (no_bias_ && OutputSize() == 2)) {
    auto* dX = Output(
        no_bias_ ? BIAS_OR_INPUT_GRAD : INPUT_GRAD, X.sizes(), at::dtype<T>());
    dX_data = dX->template mutable_data<T>();
  }
  if (!no_bias_) {
    std::vector<int64_t> dbias_dims;
    std::copy(
        output_image_dims.cbegin(),
        output_image_dims.cend(),
        std::back_inserter(dbias_dims));
    dbias_dims.push_back(shape.M);
    auto* dbias = Output(BIAS_OR_INPUT_GRAD, dbias_dims, at::dtype<T>());
    ConvPoolOpBase<Context>::template SetBiasMultiplier<T>(
        shape.N, &bias_multiplier_);
    dbias_data = dbias->template mutable_data<T>();
  }
  RunOnDeviceWithOrderNHWCImpl(
      shape,
      X_data,
      filter_data,
      dY_data,
      dfilter_data,
      dX_data,
      dbias_data,
      &column_buffer_,
      &column_transposed_buffer_,
      &dY_transposed_buffer_);

  return true;
}

template <typename T, class Context>
void LocallyConnectedGradientOp<T, Context>::RunOnDeviceWithOrderNCHWImpl(
    const lc_op_util::ShapeParams& shape,
    const T* X_data,
    const T* filter_data,
    const T* dY_data,
    T* dfilter_data,
    T* dX_data,
    T* dbias_data,
    Tensor* column_buffer,
    Tensor* column_transposed_buffer,
    Tensor* dY_transposed_buffer) {
  const int input_stride = shape.C * shape.input_image_size;
  const int column_stride = shape.kernel_size * shape.output_image_size;
  column_buffer->Resize(shape.column_dims);
  column_transposed_buffer->Resize(shape.column_transposed_dims);
  dY_transposed_buffer->Resize(shape.Y_transposed_dims);
  T* column_buffer_data = column_buffer->template mutable_data<T>();
  T* dY_transposed_buffer_data =
      dY_transposed_buffer->template mutable_data<T>();

  for (int image_id = 0; image_id < shape.N; ++image_id) {
    for (int group_id = 0; group_id < group_; ++group_id) {
      if (kernel_.size() == 2) {
        math::Im2Col<T, Context, StorageOrder::NCHW>(
            shape.C / group_,
            shape.X_dims[1],
            shape.X_dims[2],
            kernel_h(),
            kernel_w(),
            dilation_h(),
            dilation_w(),
            pad_t(),
            pad_l(),
            pad_b(),
            pad_r(),
            stride_h(),
            stride_w(),
            X_data + group_id * input_stride,
            column_buffer_data + group_id * column_stride,
            &context_);
      } else {
        math::Im2ColNd<T, Context, StorageOrder::NCHW>(
            kernel_.size(),
            shape.C * shape.input_image_size,
            column_stride,
            shape.X_dims.data(),
            shape.column_slice_dims.data(),
            kernel_.data(),
            stride_.data(),
            dilation_.data(),
            pads_.data(),
            X_data + group_id * input_stride,
            column_buffer_data + group_id * column_stride,
            &context_);
      }
    }
    X_data += input_stride * group_;
    column_buffer_data += column_stride * group_;
  }
  math::Transpose(
      shape.column_dims.size(),
      shape.column_dims.data(),
      shape.column_axes.data(),
      column_buffer->template data<T>(),
      column_transposed_buffer->template mutable_data<T>(),
      &context_);

  math::Transpose(
      shape.Y_dims.size(),
      shape.Y_dims.data(),
      shape.Y_axes.data(),
      dY_data,
      dY_transposed_buffer_data,
      &context_);

  // Gradient respect to filter.
  math::GemmStridedBatched(
      CblasNoTrans,
      CblasTrans,
      shape.output_image_size * group_,
      shape.M / group_,
      shape.kernel_size,
      shape.N,
      1.0f,
      dY_transposed_buffer_data,
      shape.M / group_ * shape.N,
      column_transposed_buffer->template data<T>(),
      shape.N * shape.kernel_size,
      0.0f,
      dfilter_data,
      shape.M / group_ * shape.kernel_size,
      &context_);

  if (dbias_data != nullptr) {
    // Gradient respect to bias.
    math::Gemv<T, Context>(
        CblasNoTrans,
        shape.output_image_size * shape.M,
        shape.N,
        1.0f,
        dY_transposed_buffer_data,
        bias_multiplier_.template data<T>(),
        0.0f,
        dbias_data,
        &context_);
  }

  if (dX_data != nullptr) {
    // Gradient respect to X.
    math::GemmStridedBatched(
        CblasTrans,
        CblasNoTrans,
        shape.output_image_size * group_,
        shape.kernel_size,
        shape.N,
        shape.M / group_,
        1.0f,
        filter_data,
        shape.kernel_size * shape.M / group_,
        dY_transposed_buffer_data,
        shape.M / group_ * shape.N,
        0.0f,
        column_transposed_buffer->template mutable_data<T>(),
        shape.kernel_size * shape.N,
        &context_);
    math::Transpose(
        shape.column_transposed_dims.size(),
        shape.column_transposed_dims.data(),
        shape.column_axes.data(),
        column_transposed_buffer->template data<T>(),
        column_buffer->template mutable_data<T>(),
        &context_);
    const T* const_column_buffer_data = column_buffer->template data<T>();
    for (int image_id = 0; image_id < shape.N; ++image_id) {
      for (int group_id = 0; group_id < group_; ++group_id) {
        if (kernel_.size() == 2) {
          math::Col2Im<T, Context, StorageOrder::NCHW>(
              shape.C / group_,
              shape.X_dims[1],
              shape.X_dims[2],
              kernel_h(),
              kernel_w(),
              dilation_h(),
              dilation_w(),
              pad_t(),
              pad_l(),
              pad_b(),
              pad_r(),
              stride_h(),
              stride_w(),
              const_column_buffer_data + group_id * column_stride,
              dX_data + group_id * input_stride,
              &context_);
        } else {
          math::Col2ImNd<T, Context, StorageOrder::NCHW>(
              kernel_.size(),
              shape.C * shape.input_image_size,
              column_stride,
              shape.X_dims.data(),
              shape.column_slice_dims.data(),
              kernel_.data(),
              stride_.data(),
              dilation_.data(),
              pads_.data(),
              const_column_buffer_data + group_id * column_stride,
              dX_data + group_id * input_stride,
              &context_);
        }
      }
      dX_data += input_stride * group_;
      const_column_buffer_data += column_stride * group_;
    }
  }
}

template <typename T, class Context>
void LocallyConnectedGradientOp<T, Context>::RunOnDeviceWithOrderNHWCImpl(
    const lc_op_util::ShapeParams& shape,
    const T* X_data,
    const T* filter_data,
    const T* dY_data,
    T* dfilter_data,
    T* dX_data,
    T* dbias_data,
    Tensor* column_buffer,
    Tensor* column_transposed_buffer,
    Tensor* dY_transposed_buffer) {
  const int input_stride = shape.C * shape.input_image_size;
  const int column_stride = shape.kernel_size * shape.output_image_size;
  column_buffer->Resize(shape.column_dims);
  column_transposed_buffer->Resize(shape.column_transposed_dims);
  dY_transposed_buffer->Resize(shape.Y_transposed_dims);
  T* column_buffer_data = column_buffer->template mutable_data<T>();
  T* dY_transposed_buffer_data =
      dY_transposed_buffer->template mutable_data<T>();
  for (int image_id = 0; image_id < shape.N; ++image_id) {
    math::Im2Col<T, Context, StorageOrder::NHWC>(
        shape.C,
        shape.X_dims[0],
        shape.X_dims[1],
        kernel_h(),
        kernel_w(),
        dilation_h(),
        dilation_w(),
        pad_t(),
        pad_l(),
        pad_b(),
        pad_r(),
        stride_h(),
        stride_w(),
        X_data + image_id * input_stride,
        column_buffer_data + image_id * column_stride,
        &context_);
  }
  math::Transpose(
      shape.column_dims.size(),
      shape.column_dims.data(),
      shape.column_axes.data(),
      column_buffer->template data<T>(),
      column_transposed_buffer->template mutable_data<T>(),
      &context_);
  math::Transpose(
      shape.Y_dims.size(),
      shape.Y_dims.data(),
      shape.Y_axes.data(),
      dY_data,
      dY_transposed_buffer_data,
      &context_);

  // Gradient respect to filter.
  math::GemmStridedBatched(
      CblasTrans,
      CblasNoTrans,
      shape.output_image_size,
      shape.M,
      shape.kernel_size,
      shape.N,
      1.0f,
      dY_transposed_buffer_data,
      shape.M * shape.N,
      column_transposed_buffer->template data<T>(),
      shape.N * shape.kernel_size,
      0.0f,
      dfilter_data,
      shape.M * shape.kernel_size,
      &context_);

  if (dbias_data != nullptr) {
    // Gradient respect to bias.
    math::Gemv<T, Context>(
        CblasTrans,
        shape.N,
        shape.output_image_size * shape.M,
        1.0f,
        dY_data,
        bias_multiplier_.template data<T>(),
        0.0f,
        dbias_data,
        &context_);
  }

  if (dX_data != nullptr) {
    // Gradient respect to X.
    math::GemmStridedBatched(
        CblasNoTrans,
        CblasNoTrans,
        shape.output_image_size,
        shape.N,
        shape.kernel_size,
        shape.M,
        1.0f,
        dY_transposed_buffer_data,
        shape.N * shape.M,
        filter_data,
        shape.M * shape.kernel_size,
        0.0f,
        column_transposed_buffer->template mutable_data<T>(),
        shape.N * shape.kernel_size,
        &context_);
    math::Transpose(
        shape.column_transposed_dims.size(),
        shape.column_transposed_dims.data(),
        shape.column_axes.data(),
        column_transposed_buffer->template data<T>(),
        column_buffer->template mutable_data<T>(),
        &context_);
    const T* const_column_buffer_data = column_buffer->template data<T>();
    for (int image_id = 0; image_id < shape.N; ++image_id) {
      math::Col2Im<T, Context, StorageOrder::NHWC>(
          shape.C,
          shape.X_dims[0],
          shape.X_dims[1],
          kernel_h(),
          kernel_w(),
          dilation_h(),
          dilation_w(),
          pad_t(),
          pad_l(),
          pad_b(),
          pad_r(),
          stride_h(),
          stride_w(),
          const_column_buffer_data,
          dX_data,
          &context_);
      dX_data += input_stride;
      const_column_buffer_data += column_stride;
    }
  }
}

} // namespace caffe2

#endif // CAFFE2_OPERATORS_LOCALLY_CONNECTED_OP_IMPL_H_