ReflectionPad.cpp

#include <ATen/ATen.h>
#include <ATen/NativeFunctions.h>
#include <ATen/Parallel.h>

namespace at {
namespace native {

namespace {

template <typename scalar_t>
static void reflection_pad1d_out_frame(
    scalar_t *input_p, scalar_t *output_p,
    int64_t nplane,
    int64_t input_w, int64_t output_w,
    int64_t pad_l) {
  int64_t i_start_x = std::max(int64_t(0), -pad_l);
  int64_t o_start_x = std::max(int64_t(0), pad_l);

  at::parallel_for(0, nplane, 0, [&](int64_t start, int64_t end) {
    int64_t ip_x;
    for (auto k = start; k < end; k++) {
      for (int64_t j = 0; j < output_w; j++) {
        if (j < pad_l) {
          ip_x = pad_l * 2 - j;
        } else if (j >= pad_l && j < input_w + pad_l) {
          ip_x = j;
        } else {
          ip_x = (input_w + pad_l - 1) * 2 - j;
        }
        ip_x = ip_x - o_start_x + i_start_x;

        scalar_t *dest_p = output_p + k*output_w + j;
        scalar_t *src_p = input_p + k*input_w + ip_x;
        *dest_p = *src_p;
      }
    }
  });
}

template <typename scalar_t>
inline void reflection_pad1d_out_loop(
    scalar_t *input_p, scalar_t *output_p,
    int64_t nbatch, int64_t nplane,
    int64_t input_w, int64_t output_w,
    int64_t pad_l) {
  at::parallel_for(0, nbatch, 0, [&](int64_t start, int64_t end) {
    for (auto p = start; p < end; p++) {
      reflection_pad1d_out_frame<scalar_t>(
        input_p + p * nplane * input_w,
        output_p + p * nplane * output_w,
        nplane,
        input_w, output_w,
        pad_l);
    }
  });
}

void reflection_pad1d_out_template(
    Tensor& output, const Tensor& input_, IntArrayRef padding) {
  int64_t dim_plane = 0;
  int64_t dim_w = 1;
  int64_t nbatch = 1;
  // allow dim=0 only in the batch dimension.
  TORCH_CHECK(
      (input_.ndimension() == 2 && input_.size(1) != 0) ||
      (input_.ndimension() == 3 && input_.size(1) != 0 && input_.size(2) != 0),
      "2D or 3D (batch mode) tensor expected for input, but got: ", input_);

  if (input_.ndimension() == 3) {
    nbatch = input_.size(0);
    dim_w++;
    dim_plane++;
  }

  /* sizes */
  auto pad_l = padding[0];
  auto pad_r = padding[1];

  int64_t nplane = input_.size(dim_plane);
  int64_t input_w = input_.size(dim_w);
  int64_t output_w  = input_w + pad_l + pad_r;

  TORCH_CHECK(pad_l < input_w && pad_r < input_w, "Argument #4: Padding size "
    "should be less than the corresponding input dimension, but got: padding (",
    pad_l, ", ", pad_r, ") at dimension ", dim_w, " of input ", input_.sizes());

  TORCH_CHECK(output_w >= 1 , 2,
    "input (W: ", input_w, ")is too small. Calculated output W: ", output_w);

  /* get contiguous input */
  Tensor input = input_.contiguous();

  /* resize output */
  if (input.ndimension() == 2) {
    output.resize_({nplane, output_w});
    if (input.is_quantized()) {
      AT_DISPATCH_QINT_TYPES(input.scalar_type(), "qreflection_pad1d", [&]() {
        reflection_pad1d_out_frame<scalar_t>(
          input.data_ptr<scalar_t>(), output.data_ptr<scalar_t>(),
          nplane,
          input_w, output_w,
          pad_l);
      });
    } else {
      AT_DISPATCH_FLOATING_TYPES(input.scalar_type(), "reflection_pad1d", [&] {
        reflection_pad1d_out_frame<scalar_t>(
          input.data_ptr<scalar_t>(), output.data_ptr<scalar_t>(),
          nplane,
          input_w, output_w,
          pad_l);
      });
    }
  } else {
    output.resize_({nbatch, nplane, output_w});
    if (input.is_quantized()) {
      AT_DISPATCH_QINT_TYPES(input.scalar_type(), "qreflection_pad1d", [&]() {
        reflection_pad1d_out_loop<scalar_t>(
          input.data_ptr<scalar_t>(), output.data_ptr<scalar_t>(),
          nbatch, nplane,
          input_w, output_w,
          pad_l);
      });
    } else {
      AT_DISPATCH_FLOATING_TYPES(input.scalar_type(), "reflection_pad1d", [&] {
        reflection_pad1d_out_loop<scalar_t>(
          input.data_ptr<scalar_t>(), output.data_ptr<scalar_t>(),
          nbatch, nplane,
          input_w, output_w,
          pad_l);
      });
    }
  }
}

template <typename scalar_t>
static void reflection_pad1d_backward_out_frame(
    scalar_t * grad_input, scalar_t * grad_output,
    int64_t nplane,
    int64_t input_w, int64_t output_w,
    int64_t pad_l) {
  int64_t i_start_x = std::max(int64_t(0), -pad_l);
  int64_t o_start_x = std::max(int64_t(0), pad_l);

  at::parallel_for(0, nplane, 0, [&](int64_t start, int64_t end) {
    int64_t ip_x;
    for (auto k = start; k < end; k++) {
      for (int64_t j = 0; j < output_w; j++) {
        if (j < pad_l) {
          ip_x = pad_l * 2 - j;
        } else if (j >= pad_l && j < input_w + pad_l) {
          ip_x = j;
        } else {
          ip_x = (input_w + pad_l - 1) * 2 - j;
        }
        ip_x = ip_x - o_start_x + i_start_x;

        scalar_t *src_p = grad_output + k*output_w + j;
        scalar_t *dest_p = grad_input + k*input_w + ip_x;
        *dest_p += *src_p;
      }
    }
  });
}

template <typename scalar_t>
inline void reflection_pad1d_backward_out_loop(
    scalar_t *grad_input, scalar_t *grad_output,
    int64_t nbatch, int64_t nplane,
    int64_t input_w, int64_t output_w,
    int64_t pad_l) {
  at::parallel_for(0, nbatch, 0, [&](int64_t start, int64_t end) {
    for (auto p = start; p < end; p++) {
      reflection_pad1d_backward_out_frame<scalar_t>(
        grad_input + p * nplane * input_w,
        grad_output + p * nplane * output_w,
        nplane,
        input_w, output_w,
        pad_l);
    }
  });
}

void reflection_pad1d_backward_out_template(
    Tensor& grad_input, const Tensor& grad_output_, const Tensor& input,
    IntArrayRef padding) {
  int64_t dim_plane = 0;
  int64_t dim_w = 1;
  int64_t nbatch = 1;

  if (input.ndimension() == 3) {
    nbatch = input.size(0);
    dim_w++;
    dim_plane++;
  }

  /* sizes */
  auto pad_l = padding[0];
  auto pad_r = padding[1];
  int64_t nplane = input.size(dim_plane);
  int64_t input_w = input.size(dim_w);
  int64_t output_w  = input_w + pad_l + pad_r;

  TORCH_CHECK(output_w == grad_output_.size(dim_w), "grad_output width unexpected."
    " Expected: ", output_w, ", Got: ", grad_output_.size(dim_w));

  /* get contiguous grad_output */
  Tensor grad_output = grad_output_.contiguous();

  /* backprop */
  if (input.ndimension() == 2) {
    AT_DISPATCH_FLOATING_TYPES(
      grad_input.scalar_type(), "reflection_pad1d_backward", [&] {
        reflection_pad1d_backward_out_frame(
          grad_input.data_ptr<scalar_t>(), grad_output.data_ptr<scalar_t>(),
          nplane,
          input_w, output_w,
          pad_l);
        }
    );
  } else {
    AT_DISPATCH_FLOATING_TYPES(
      grad_input.scalar_type(), "reflection_pad1d_backward", [&] {
        reflection_pad1d_backward_out_loop(
          grad_input.data_ptr<scalar_t>(),
          grad_output.data_ptr<scalar_t>(),
          nbatch, nplane,
          input_w, output_w,
          pad_l);
      }
    );
  }
}

template <typename scalar_t>
static void reflection_pad2d_out_frame(
    scalar_t * input_p, scalar_t * output_p,
    int64_t nplane,
    int64_t input_w, int64_t input_h,
    int64_t output_w, int64_t output_h,
    int64_t pad_l, int64_t pad_t) {
  auto i_start_x = std::max(int64_t(0), -pad_l);
  auto i_start_y = std::max(int64_t(0), -pad_t);
  auto o_start_x = std::max(int64_t(0), pad_l);
  auto o_start_y = std::max(int64_t(0), pad_t);

  at::parallel_for(0, nplane, 0, [&](int64_t start, int64_t end) {
    int64_t ip_x, ip_y;
    for (auto k = start; k < end; k++) {
      for (int64_t i = 0; i < output_h; i++) {
        for (int64_t j = 0; j < output_w; j++) {
          if (j < pad_l) {
            ip_x = pad_l * 2 - j;
          } else if (j >= pad_l && j < input_w + pad_l) {
            ip_x = j;
          } else {
            ip_x = (input_w + pad_l - 1) * 2 - j;
          }
          ip_x = ip_x - o_start_x + i_start_x;

          if (i < pad_t) {
            ip_y = pad_t * 2 - i;
          } else if (i >= pad_t && i < input_h + pad_t) {
            ip_y = i;
          } else {
            ip_y = (input_h + pad_t - 1) * 2 - i;
          }
          ip_y = ip_y - o_start_y + i_start_y;

          scalar_t *dest_p = output_p + k*output_w*output_h + i * output_w + j;
          scalar_t *src_p = input_p + k*input_w*input_h + ip_y * input_w + ip_x;
          *dest_p = *src_p;
        }
      }
    }
  });
}

template <typename scalar_t>
inline void reflection_pad2d_out_loop(
    scalar_t * input_p, scalar_t * output_p,
    int64_t nbatch, int64_t nplane,
    int64_t input_w, int64_t input_h,
    int64_t output_w, int64_t output_h,
    int64_t pad_l, int64_t pad_t) {
  at::parallel_for(0, nbatch, 0, [&](int64_t start, int64_t end) {
    for (auto p = start; p < end; p++) {
      reflection_pad2d_out_frame(
        input_p + p * nplane * input_w * input_h,
        output_p + p * nplane * output_w * output_h,
        nplane,
        input_w, input_h, output_w, output_h,
        pad_l, pad_t);
    }
  });
}

void reflection_pad2d_out_template(
    Tensor &output, const Tensor &input_, IntArrayRef padding) {
  int dim_w = 2;
  int dim_h = 1;
  int dim_slices = 0;
  int64_t nbatch = 1;

  bool valid_dims = input_.size(1) != 0 && input_.size(2) != 0;
  TORCH_CHECK(
      (input_.ndimension() == 3 && valid_dims) ||
      (input_.ndimension() == 4 && valid_dims && input_.size(3) != 0),
      "3D or 4D (batch mode) tensor expected for input, but got: ", input_);

  if (input_.ndimension() == 4) {
    nbatch = input_.size(0);
    dim_w++;
    dim_h++;
    dim_slices++;
  }

  /* sizes */
  int64_t pad_l = padding[0];
  int64_t pad_r = padding[1];
  int64_t pad_t = padding[2];
  int64_t pad_b = padding[3];

  int64_t nplane = input_.size(dim_slices);
  int64_t input_h = input_.size(dim_h);
  int64_t input_w = input_.size(dim_w);
  int64_t output_h = input_h + pad_t + pad_b;
  int64_t output_w  = input_w + pad_l + pad_r;

  TORCH_CHECK(pad_l < input_w && pad_r < input_w,
    "Argument #4: Padding size should be less than the corresponding "
    "input dimension, but got: padding (", pad_l, ", ", pad_r,
    ") at dimension ", dim_w, " of input ", input_.ndimension());

  TORCH_CHECK(pad_t < input_h && pad_b < input_h,
    "Argument #6: Padding size should be less than the corresponding "
    "input dimension, but got: padding (", pad_t, ", ", pad_b,
    ") at dimension ", dim_h, " of input ", input_.ndimension());

  TORCH_CHECK(output_w >= 1 || output_h >= 1,
    "input (H: ", input_h, ", W: ", input_w, ")is too small. Calculated "
    "output H: ", output_h, " W: ", output_w);

  /* get contiguous input */
  Tensor input = input_.contiguous();

  if (input.ndimension() == 3) {
    /* resize output */
    output.resize_({nplane, output_h, output_w});
    if (input.is_quantized()) {
      AT_DISPATCH_QINT_TYPES(input.scalar_type(), "qreflection_pad2d", [&] {
        reflection_pad2d_out_frame(
          input.data_ptr<scalar_t>(), output.data_ptr<scalar_t>(),
          nplane,
          input_w, input_h, output_w, output_h,
          pad_l, pad_t);
      });
    } else {
      AT_DISPATCH_FLOATING_TYPES(input.scalar_type(), "reflection_pad2d", [&] {
        reflection_pad2d_out_frame(
          input.data_ptr<scalar_t>(), output.data_ptr<scalar_t>(),
          nplane,
          input_w, input_h, output_w, output_h,
          pad_l, pad_t);
      });
    }
  } else {
    /* resize output */
    output.resize_({nbatch, nplane, output_h, output_w});
    if (input.is_quantized()) {
      AT_DISPATCH_QINT_TYPES(input.scalar_type(), "qreflection_pad2d", [&] {
        reflection_pad2d_out_loop(
          input.data_ptr<scalar_t>(), output.data_ptr<scalar_t>(),
          nbatch, nplane,
          input_w, input_h, output_w, output_h,
          pad_l, pad_t);
      });
    } else {
      AT_DISPATCH_FLOATING_TYPES(input.scalar_type(), "reflection_pad2d", [&] {
        reflection_pad2d_out_loop(
          input.data_ptr<scalar_t>(), output.data_ptr<scalar_t>(),
          nbatch, nplane,
          input_w, input_h, output_w, output_h,
          pad_l, pad_t);
      });
    }
  }
}

template <typename scalar_t>
static void reflection_pad2d_backward_out_frame(
    scalar_t *grad_input, scalar_t *grad_output,
    int64_t nplane,
    int64_t input_w, int64_t input_h,
    int64_t output_w, int64_t output_h,
    int64_t pad_l, int64_t pad_t) {
  auto i_start_x = std::max(int64_t(0), -pad_l);
  auto i_start_y = std::max(int64_t(0), -pad_t);
  auto o_start_x = std::max(int64_t(0), pad_l);
  auto o_start_y = std::max(int64_t(0), pad_t);

  at::parallel_for(0, nplane, 0, [&](int64_t start, int64_t end) {
    int64_t ip_x, ip_y;
    for (auto k = start; k < end; k++) {
      for (int64_t i = 0; i < output_h; i++) {
        for (int64_t j = 0; j < output_w; j++) {
          if (j < pad_l) {
            ip_x = pad_l * 2 - j;
          } else if (j >= pad_l && j < input_w + pad_l) {
            ip_x = j;
          } else {
            ip_x = (input_w + pad_l - 1) * 2 - j;
          }
          ip_x = ip_x - o_start_x + i_start_x;

          if (i < pad_t) {
            ip_y = pad_t * 2 - i;
          } else if (i >= pad_t && i < input_h + pad_t) {
            ip_y = i;
          } else {
            ip_y = (input_h + pad_t - 1) * 2 - i;
          }
          ip_y = ip_y - o_start_y + i_start_y;

          scalar_t *src_p =
            grad_output + k * output_w * output_h + i * output_w + j;
          scalar_t *dest_p =
            grad_input + k * input_w * input_h + ip_y * input_w + ip_x;
          *dest_p += *src_p;
        }
      }
    }
  });
}

template <typename scalar_t>
inline void reflection_pad2d_backward_out_loop(
    scalar_t *grad_input, scalar_t *grad_output,
    int64_t nbatch, int64_t nplane,
    int64_t input_w, int64_t input_h,
    int64_t output_w, int64_t output_h,
    int64_t pad_l, int64_t pad_t) {
  at::parallel_for(0, nbatch, 0, [&](int64_t start, int64_t end) {
    for (auto p = start; p < end; p++) {
      reflection_pad2d_backward_out_frame(
        grad_input + p * nplane * input_h * input_w,
        grad_output + p * nplane * output_h * output_w,
        nplane,
        input_w, input_h, output_w, output_h,
        pad_l, pad_t);
    }
  });
}

void reflection_pad2d_backward_out_template(
    Tensor &grad_input, const Tensor &grad_output_,
    const Tensor &input, IntArrayRef padding) {
  int dim_w = 2;
  int dim_h = 1;
  int dim_plane = 0;
  int64_t nbatch = 1;

  if (input.ndimension() == 4) {
    nbatch = input.size(0);
    dim_w++;
    dim_h++;
    dim_plane++;
  }

  /* sizes */
  int64_t pad_l = padding[0];
  int64_t pad_r = padding[1];
  int64_t pad_t = padding[2];
  int64_t pad_b = padding[3];

  int64_t nplane = input.size(dim_plane);
  int64_t input_h = input.size(dim_h);
  int64_t input_w = input.size(dim_w);
  int64_t output_h = input_h + pad_t + pad_b;
  int64_t output_w  = input_w + pad_l + pad_r;

  TORCH_CHECK(output_w == grad_output_.size(dim_w),
    "gradOutput width unexpected. Expected: ", output_w, ", Got: ",
    grad_output_.size(dim_w));

  TORCH_CHECK(output_h == grad_output_.size(dim_h),
    "gradOutput height unexpected. Expected: ", output_h, ", Got: ",
    grad_output_.size(dim_h));

  /* get contiguous gradOutput */
  Tensor grad_output = grad_output_.contiguous();

  /* backprop */
  if (input.ndimension() == 3) {
    AT_DISPATCH_FLOATING_TYPES(
      grad_output.scalar_type(), "reflection_pad2d_backward", [&] {
        reflection_pad2d_backward_out_frame(
          grad_input.data_ptr<scalar_t>(), grad_output.data_ptr<scalar_t>(),
          nplane,
          input_w, input_h, output_w, output_h,
          pad_l, pad_t);
      }
    );
  } else {
    AT_DISPATCH_FLOATING_TYPES(
      grad_output.scalar_type(), "reflection_pad2d_backward", [&] {
        reflection_pad2d_backward_out_loop(
          grad_input.data_ptr<scalar_t>(), grad_output.data_ptr<scalar_t>(),
          nbatch, nplane,
          input_w, input_h, output_w, output_h,
          pad_l, pad_t);
      }
    );
  }
}

} // namespace

Tensor& reflection_pad1d_out_cpu(
    Tensor& output, const Tensor& input, IntArrayRef padding) {
  reflection_pad1d_out_template(output, input, padding);
  return output;
}

Tensor reflection_pad1d_cpu(const Tensor& input, IntArrayRef padding) {
  Tensor output;
  if (input.is_quantized()) {
    if (input.qscheme() == kPerTensorAffine) {
      output = at::_empty_affine_quantized({0}, input.options(),
                                           input.q_scale(),
                                           input.q_zero_point());
    } else {
      TORCH_CHECK(false, "Only per tensor quantization is supported");
    }
  } else {
    output = at::empty({0}, input.options());
  }
  reflection_pad1d_out_template(output, input, padding);
  return output;
}

Tensor& reflection_pad1d_backward_out_cpu(
    Tensor& grad_input,
    const Tensor& grad_output,
    const Tensor& input,
    IntArrayRef padding) {
  grad_input.resize_as_(input);
  grad_input.zero_();
  reflection_pad1d_backward_out_template(
    grad_input, grad_output, input, padding);
  return grad_input;
}

Tensor reflection_pad1d_backward_cpu(
    const Tensor& grad_output,
    const Tensor& input,
    IntArrayRef padding) {
  auto grad_input = at::zeros_like(input, LEGACY_CONTIGUOUS_MEMORY_FORMAT);
  reflection_pad1d_backward_out_template(
    grad_input, grad_output, input, padding);
  return grad_input;
}

Tensor& reflection_pad2d_out_cpu(
    Tensor& output, const Tensor& input, IntArrayRef padding) {
  reflection_pad2d_out_template(output, input, padding);
  return output;
}

Tensor reflection_pad2d_cpu(const Tensor& input, IntArrayRef padding) {
  Tensor output;
  if (input.is_quantized()) {
    if (input.qscheme() == kPerTensorAffine) {
      output = at::_empty_affine_quantized({0}, input.options(),
                                           input.q_scale(),
                                           input.q_zero_point());
    } else {
      TORCH_CHECK(false, "Only per tensor quantization is supported");
    }
  } else {
    output = at::empty({0}, input.options());
  }
  reflection_pad2d_out_template(output, input, padding);
  return output;
}

Tensor& reflection_pad2d_backward_out_cpu(
    Tensor& grad_input,
    const Tensor& grad_output,
    const Tensor& input,
    IntArrayRef padding) {
  grad_input.resize_as_(input);
  grad_input.zero_();
  reflection_pad2d_backward_out_template(
    grad_input, grad_output, input, padding);
  return grad_input;
}

Tensor reflection_pad2d_backward_cpu(
    const Tensor& grad_output,
    const Tensor& input,
    IntArrayRef padding) {
  auto grad_input = at::zeros_like(input, LEGACY_CONTIGUOUS_MEMORY_FORMAT);
  reflection_pad2d_backward_out_template(
    grad_input, grad_output, input, padding);
  return grad_input;
}

} // namespace native
} // namespace at