ridcp

basicsr/ops/dcn/src/deform_conv_cuda.cpp

+// modify from
+// https://github.com/chengdazhi/Deformable-Convolution-V2-PyTorch/blob/mmdetection/mmdet/ops/dcn/src/deform_conv_cuda.c
+
+#include <torch/extension.h>
+#include <ATen/DeviceGuard.h>
+
+#include <cmath>
+#include <vector>
+
+void deformable_im2col(const at::Tensor data_im, const at::Tensor data_offset,
+                       const int channels, const int height, const int width,
+                       const int ksize_h, const int ksize_w, const int pad_h,
+                       const int pad_w, const int stride_h, const int stride_w,
+                       const int dilation_h, const int dilation_w,
+                       const int parallel_imgs, const int deformable_group,
+                       at::Tensor data_col);
+
+void deformable_col2im(const at::Tensor data_col, const at::Tensor data_offset,
+                       const int channels, const int height, const int width,
+                       const int ksize_h, const int ksize_w, const int pad_h,
+                       const int pad_w, const int stride_h, const int stride_w,
+                       const int dilation_h, const int dilation_w,
+                       const int parallel_imgs, const int deformable_group,
+                       at::Tensor grad_im);
+
+void deformable_col2im_coord(
+    const at::Tensor data_col, const at::Tensor data_im,
+    const at::Tensor data_offset, const int channels, const int height,
+    const int width, const int ksize_h, const int ksize_w, const int pad_h,
+    const int pad_w, const int stride_h, const int stride_w,
+    const int dilation_h, const int dilation_w, const int parallel_imgs,
+    const int deformable_group, at::Tensor grad_offset);
+
+void modulated_deformable_im2col_cuda(
+    const at::Tensor data_im, const at::Tensor data_offset,
+    const at::Tensor data_mask, const int batch_size, const int channels,
+    const int height_im, const int width_im, const int height_col,
+    const int width_col, const int kernel_h, const int kenerl_w,
+    const int pad_h, const int pad_w, const int stride_h, const int stride_w,
+    const int dilation_h, const int dilation_w, const int deformable_group,
+    at::Tensor data_col);
+
+void modulated_deformable_col2im_cuda(
+    const at::Tensor data_col, const at::Tensor data_offset,
+    const at::Tensor data_mask, const int batch_size, const int channels,
+    const int height_im, const int width_im, const int height_col,
+    const int width_col, const int kernel_h, const int kenerl_w,
+    const int pad_h, const int pad_w, const int stride_h, const int stride_w,
+    const int dilation_h, const int dilation_w, const int deformable_group,
+    at::Tensor grad_im);
+
+void modulated_deformable_col2im_coord_cuda(
+    const at::Tensor data_col, const at::Tensor data_im,
+    const at::Tensor data_offset, const at::Tensor data_mask,
+    const int batch_size, const int channels, const int height_im,
+    const int width_im, const int height_col, const int width_col,
+    const int kernel_h, const int kenerl_w, const int pad_h, const int pad_w,
+    const int stride_h, const int stride_w, const int dilation_h,
+    const int dilation_w, const int deformable_group, at::Tensor grad_offset,
+    at::Tensor grad_mask);
+
+void shape_check(at::Tensor input, at::Tensor offset, at::Tensor *gradOutput,
+                 at::Tensor weight, int kH, int kW, int dH, int dW, int padH,
+                 int padW, int dilationH, int dilationW, int group,
+                 int deformable_group) {
+  TORCH_CHECK(weight.ndimension() == 4,
+           "4D weight tensor (nOutputPlane,nInputPlane,kH,kW) expected, "
+           "but got: %s",
+           weight.ndimension());
+
+  TORCH_CHECK(weight.is_contiguous(), "weight tensor has to be contiguous");
+
+  TORCH_CHECK(kW > 0 && kH > 0,
+           "kernel size should be greater than zero, but got kH: %d kW: %d", kH,
+           kW);
+
+  TORCH_CHECK((weight.size(2) == kH && weight.size(3) == kW),
+           "kernel size should be consistent with weight, ",
+           "but got kH: %d kW: %d weight.size(2): %d, weight.size(3): %d", kH,
+           kW, weight.size(2), weight.size(3));
+
+  TORCH_CHECK(dW > 0 && dH > 0,
+           "stride should be greater than zero, but got dH: %d dW: %d", dH, dW);
+
+  TORCH_CHECK(
+      dilationW > 0 && dilationH > 0,
+      "dilation should be greater than 0, but got dilationH: %d dilationW: %d",
+      dilationH, dilationW);
+
+  int ndim = input.ndimension();
+  int dimf = 0;
+  int dimh = 1;
+  int dimw = 2;
+
+  if (ndim == 4) {
+    dimf++;
+    dimh++;
+    dimw++;
+  }
+
+  TORCH_CHECK(ndim == 3 || ndim == 4, "3D or 4D input tensor expected but got: %s",
+           ndim);
+
+  long nInputPlane = weight.size(1) * group;
+  long inputHeight = input.size(dimh);
+  long inputWidth = input.size(dimw);
+  long nOutputPlane = weight.size(0);
+  long outputHeight =
+      (inputHeight + 2 * padH - (dilationH * (kH - 1) + 1)) / dH + 1;
+  long outputWidth =
+      (inputWidth + 2 * padW - (dilationW * (kW - 1) + 1)) / dW + 1;
+
+  TORCH_CHECK(nInputPlane % deformable_group == 0,
+           "input channels must divide deformable group size");
+
+  if (outputWidth < 1 || outputHeight < 1)
+    AT_ERROR(
+        "Given input size: (%ld x %ld x %ld). "
+        "Calculated output size: (%ld x %ld x %ld). Output size is too small",
+        nInputPlane, inputHeight, inputWidth, nOutputPlane, outputHeight,
+        outputWidth);
+
+  TORCH_CHECK(input.size(1) == nInputPlane,
+           "invalid number of input planes, expected: %d, but got: %d",
+           nInputPlane, input.size(1));
+
+  TORCH_CHECK((inputHeight >= kH && inputWidth >= kW),
+           "input image is smaller than kernel");
+
+  TORCH_CHECK((offset.size(2) == outputHeight && offset.size(3) == outputWidth),
+           "invalid spatial size of offset, expected height: %d width: %d, but "
+           "got height: %d width: %d",
+           outputHeight, outputWidth, offset.size(2), offset.size(3));
+
+  TORCH_CHECK((offset.size(1) == deformable_group * 2 * kH * kW),
+           "invalid number of channels of offset");
+
+  if (gradOutput != NULL) {
+    TORCH_CHECK(gradOutput->size(dimf) == nOutputPlane,
+             "invalid number of gradOutput planes, expected: %d, but got: %d",
+             nOutputPlane, gradOutput->size(dimf));
+
+    TORCH_CHECK((gradOutput->size(dimh) == outputHeight &&
+              gradOutput->size(dimw) == outputWidth),
+             "invalid size of gradOutput, expected height: %d width: %d , but "
+             "got height: %d width: %d",
+             outputHeight, outputWidth, gradOutput->size(dimh),
+             gradOutput->size(dimw));
+  }
+}
+
+int deform_conv_forward_cuda(at::Tensor input, at::Tensor weight,
+                             at::Tensor offset, at::Tensor output,
+                             at::Tensor columns, at::Tensor ones, int kW,
+                             int kH, int dW, int dH, int padW, int padH,
+                             int dilationW, int dilationH, int group,
+                             int deformable_group, int im2col_step) {
+  // todo: resize columns to include im2col: done
+  // todo: add im2col_step as input
+  // todo: add new output buffer and transpose it to output (or directly
+  // transpose output) todo: possibly change data indexing because of
+  // parallel_imgs
+
+  shape_check(input, offset, NULL, weight, kH, kW, dH, dW, padH, padW,
+              dilationH, dilationW, group, deformable_group);
+  at::DeviceGuard guard(input.device());
+
+  input = input.contiguous();
+  offset = offset.contiguous();
+  weight = weight.contiguous();
+
+  int batch = 1;
+  if (input.ndimension() == 3) {
+    // Force batch
+    batch = 0;
+    input.unsqueeze_(0);
+    offset.unsqueeze_(0);
+  }
+
+  // todo: assert batchsize dividable by im2col_step
+
+  long batchSize = input.size(0);
+  long nInputPlane = input.size(1);
+  long inputHeight = input.size(2);
+  long inputWidth = input.size(3);
+
+  long nOutputPlane = weight.size(0);
+
+  long outputWidth =
+      (inputWidth + 2 * padW - (dilationW * (kW - 1) + 1)) / dW + 1;
+  long outputHeight =
+      (inputHeight + 2 * padH - (dilationH * (kH - 1) + 1)) / dH + 1;
+
+  TORCH_CHECK((offset.size(0) == batchSize), "invalid batch size of offset");
+
+  output = output.view({batchSize / im2col_step, im2col_step, nOutputPlane,
+                        outputHeight, outputWidth});
+  columns = at::zeros(
+      {nInputPlane * kW * kH, im2col_step * outputHeight * outputWidth},
+      input.options());
+
+  if (ones.ndimension() != 2 ||
+      ones.size(0) * ones.size(1) < outputHeight * outputWidth) {
+    ones = at::ones({outputHeight, outputWidth}, input.options());
+  }
+
+  input = input.view({batchSize / im2col_step, im2col_step, nInputPlane,
+                      inputHeight, inputWidth});
+  offset =
+      offset.view({batchSize / im2col_step, im2col_step,
+                   deformable_group * 2 * kH * kW, outputHeight, outputWidth});
+
+  at::Tensor output_buffer =
+      at::zeros({batchSize / im2col_step, nOutputPlane,
+                 im2col_step * outputHeight, outputWidth},
+                output.options());
+
+  output_buffer = output_buffer.view(
+      {output_buffer.size(0), group, output_buffer.size(1) / group,
+       output_buffer.size(2), output_buffer.size(3)});
+
+  for (int elt = 0; elt < batchSize / im2col_step; elt++) {
+    deformable_im2col(input[elt], offset[elt], nInputPlane, inputHeight,
+                      inputWidth, kH, kW, padH, padW, dH, dW, dilationH,
+                      dilationW, im2col_step, deformable_group, columns);
+
+    columns = columns.view({group, columns.size(0) / group, columns.size(1)});
+    weight = weight.view({group, weight.size(0) / group, weight.size(1),
+                          weight.size(2), weight.size(3)});
+
+    for (int g = 0; g < group; g++) {
+      output_buffer[elt][g] = output_buffer[elt][g]
+                                  .flatten(1)
+                                  .addmm_(weight[g].flatten(1), columns[g])
+                                  .view_as(output_buffer[elt][g]);
+    }
+  }
+
+  output_buffer = output_buffer.view(
+      {output_buffer.size(0), output_buffer.size(1) * output_buffer.size(2),
+       output_buffer.size(3), output_buffer.size(4)});
+
+  output_buffer = output_buffer.view({batchSize / im2col_step, nOutputPlane,
+                                      im2col_step, outputHeight, outputWidth});
+  output_buffer.transpose_(1, 2);
+  output.copy_(output_buffer);
+  output = output.view({batchSize, nOutputPlane, outputHeight, outputWidth});
+
+  input = input.view({batchSize, nInputPlane, inputHeight, inputWidth});
+  offset = offset.view(
+      {batchSize, deformable_group * 2 * kH * kW, outputHeight, outputWidth});
+
+  if (batch == 0) {
+    output = output.view({nOutputPlane, outputHeight, outputWidth});
+    input = input.view({nInputPlane, inputHeight, inputWidth});
+    offset = offset.view({offset.size(1), offset.size(2), offset.size(3)});
+  }
+
+  return 1;
+}
+
+int deform_conv_backward_input_cuda(at::Tensor input, at::Tensor offset,
+                                    at::Tensor gradOutput, at::Tensor gradInput,
+                                    at::Tensor gradOffset, at::Tensor weight,
+                                    at::Tensor columns, int kW, int kH, int dW,
+                                    int dH, int padW, int padH, int dilationW,
+                                    int dilationH, int group,
+                                    int deformable_group, int im2col_step) {
+  shape_check(input, offset, &gradOutput, weight, kH, kW, dH, dW, padH, padW,
+              dilationH, dilationW, group, deformable_group);
+  at::DeviceGuard guard(input.device());
+
+  input = input.contiguous();
+  offset = offset.contiguous();
+  gradOutput = gradOutput.contiguous();
+  weight = weight.contiguous();
+
+  int batch = 1;
+
+  if (input.ndimension() == 3) {
+    // Force batch
+    batch = 0;
+    input = input.view({1, input.size(0), input.size(1), input.size(2)});
+    offset = offset.view({1, offset.size(0), offset.size(1), offset.size(2)});
+    gradOutput = gradOutput.view(
+        {1, gradOutput.size(0), gradOutput.size(1), gradOutput.size(2)});
+  }
+
+  long batchSize = input.size(0);
+  long nInputPlane = input.size(1);
+  long inputHeight = input.size(2);
+  long inputWidth = input.size(3);
+
+  long nOutputPlane = weight.size(0);
+
+  long outputWidth =
+      (inputWidth + 2 * padW - (dilationW * (kW - 1) + 1)) / dW + 1;
+  long outputHeight =
+      (inputHeight + 2 * padH - (dilationH * (kH - 1) + 1)) / dH + 1;
+
+  TORCH_CHECK((offset.size(0) == batchSize), 3, "invalid batch size of offset");
+  gradInput = gradInput.view({batchSize, nInputPlane, inputHeight, inputWidth});
+  columns = at::zeros(
+      {nInputPlane * kW * kH, im2col_step * outputHeight * outputWidth},
+      input.options());
+
+  // change order of grad output
+  gradOutput = gradOutput.view({batchSize / im2col_step, im2col_step,
+                                nOutputPlane, outputHeight, outputWidth});
+  gradOutput.transpose_(1, 2);
+
+  gradInput = gradInput.view({batchSize / im2col_step, im2col_step, nInputPlane,
+                              inputHeight, inputWidth});
+  input = input.view({batchSize / im2col_step, im2col_step, nInputPlane,
+                      inputHeight, inputWidth});
+  gradOffset = gradOffset.view({batchSize / im2col_step, im2col_step,
+                                deformable_group * 2 * kH * kW, outputHeight,
+                                outputWidth});
+  offset =
+      offset.view({batchSize / im2col_step, im2col_step,
+                   deformable_group * 2 * kH * kW, outputHeight, outputWidth});
+
+  for (int elt = 0; elt < batchSize / im2col_step; elt++) {
+    // divide into groups
+    columns = columns.view({group, columns.size(0) / group, columns.size(1)});
+    weight = weight.view({group, weight.size(0) / group, weight.size(1),
+                          weight.size(2), weight.size(3)});
+    gradOutput = gradOutput.view(
+        {gradOutput.size(0), group, gradOutput.size(1) / group,
+         gradOutput.size(2), gradOutput.size(3), gradOutput.size(4)});
+
+    for (int g = 0; g < group; g++) {
+      columns[g] = columns[g].addmm_(weight[g].flatten(1).transpose(0, 1),
+                                     gradOutput[elt][g].flatten(1), 0.0f, 1.0f);
+    }
+
+    columns =
+        columns.view({columns.size(0) * columns.size(1), columns.size(2)});
+    gradOutput = gradOutput.view(
+        {gradOutput.size(0), gradOutput.size(1) * gradOutput.size(2),
+         gradOutput.size(3), gradOutput.size(4), gradOutput.size(5)});
+
+    deformable_col2im_coord(columns, input[elt], offset[elt], nInputPlane,
+                            inputHeight, inputWidth, kH, kW, padH, padW, dH, dW,
+                            dilationH, dilationW, im2col_step, deformable_group,
+                            gradOffset[elt]);
+
+    deformable_col2im(columns, offset[elt], nInputPlane, inputHeight,
+                      inputWidth, kH, kW, padH, padW, dH, dW, dilationH,
+                      dilationW, im2col_step, deformable_group, gradInput[elt]);
+  }
+
+  gradOutput.transpose_(1, 2);
+  gradOutput =
+      gradOutput.view({batchSize, nOutputPlane, outputHeight, outputWidth});
+
+  gradInput = gradInput.view({batchSize, nInputPlane, inputHeight, inputWidth});
+  input = input.view({batchSize, nInputPlane, inputHeight, inputWidth});
+  gradOffset = gradOffset.view(
+      {batchSize, deformable_group * 2 * kH * kW, outputHeight, outputWidth});
+  offset = offset.view(
+      {batchSize, deformable_group * 2 * kH * kW, outputHeight, outputWidth});
+
+  if (batch == 0) {
+    gradOutput = gradOutput.view({nOutputPlane, outputHeight, outputWidth});
+    input = input.view({nInputPlane, inputHeight, inputWidth});
+    gradInput = gradInput.view({nInputPlane, inputHeight, inputWidth});
+    offset = offset.view({offset.size(1), offset.size(2), offset.size(3)});
+    gradOffset =
+        gradOffset.view({offset.size(1), offset.size(2), offset.size(3)});
+  }
+
+  return 1;
+}
+
+int deform_conv_backward_parameters_cuda(
+    at::Tensor input, at::Tensor offset, at::Tensor gradOutput,
+    at::Tensor gradWeight,  // at::Tensor gradBias,
+    at::Tensor columns, at::Tensor ones, int kW, int kH, int dW, int dH,
+    int padW, int padH, int dilationW, int dilationH, int group,
+    int deformable_group, float scale, int im2col_step) {
+  // todo: transpose and reshape outGrad
+  // todo: reshape columns
+  // todo: add im2col_step as input
+
+  shape_check(input, offset, &gradOutput, gradWeight, kH, kW, dH, dW, padH,
+              padW, dilationH, dilationW, group, deformable_group);
+  at::DeviceGuard guard(input.device());
+
+  input = input.contiguous();
+  offset = offset.contiguous();
+  gradOutput = gradOutput.contiguous();
+
+  int batch = 1;
+
+  if (input.ndimension() == 3) {
+    // Force batch
+    batch = 0;
+    input = input.view(
+        at::IntList({1, input.size(0), input.size(1), input.size(2)}));
+    gradOutput = gradOutput.view(
+        {1, gradOutput.size(0), gradOutput.size(1), gradOutput.size(2)});
+  }
+
+  long batchSize = input.size(0);
+  long nInputPlane = input.size(1);
+  long inputHeight = input.size(2);
+  long inputWidth = input.size(3);
+
+  long nOutputPlane = gradWeight.size(0);
+
+  long outputWidth =
+      (inputWidth + 2 * padW - (dilationW * (kW - 1) + 1)) / dW + 1;
+  long outputHeight =
+      (inputHeight + 2 * padH - (dilationH * (kH - 1) + 1)) / dH + 1;
+
+  TORCH_CHECK((offset.size(0) == batchSize), "invalid batch size of offset");
+
+  columns = at::zeros(
+      {nInputPlane * kW * kH, im2col_step * outputHeight * outputWidth},
+      input.options());
+
+  gradOutput = gradOutput.view({batchSize / im2col_step, im2col_step,
+                                nOutputPlane, outputHeight, outputWidth});
+  gradOutput.transpose_(1, 2);
+
+  at::Tensor gradOutputBuffer = at::zeros_like(gradOutput);
+  gradOutputBuffer =
+      gradOutputBuffer.view({batchSize / im2col_step, nOutputPlane, im2col_step,
+                             outputHeight, outputWidth});
+  gradOutputBuffer.copy_(gradOutput);
+  gradOutputBuffer =
+      gradOutputBuffer.view({batchSize / im2col_step, nOutputPlane,
+                             im2col_step * outputHeight, outputWidth});
+
+  gradOutput.transpose_(1, 2);
+  gradOutput =
+      gradOutput.view({batchSize, nOutputPlane, outputHeight, outputWidth});
+
+  input = input.view({batchSize / im2col_step, im2col_step, nInputPlane,
+                      inputHeight, inputWidth});
+  offset =
+      offset.view({batchSize / im2col_step, im2col_step,
+                   deformable_group * 2 * kH * kW, outputHeight, outputWidth});
+
+  for (int elt = 0; elt < batchSize / im2col_step; elt++) {
+    deformable_im2col(input[elt], offset[elt], nInputPlane, inputHeight,
+                      inputWidth, kH, kW, padH, padW, dH, dW, dilationH,
+                      dilationW, im2col_step, deformable_group, columns);
+
+    // divide into group
+    gradOutputBuffer = gradOutputBuffer.view(
+        {gradOutputBuffer.size(0), group, gradOutputBuffer.size(1) / group,
+         gradOutputBuffer.size(2), gradOutputBuffer.size(3)});
+    columns = columns.view({group, columns.size(0) / group, columns.size(1)});
+    gradWeight =
+        gradWeight.view({group, gradWeight.size(0) / group, gradWeight.size(1),
+                         gradWeight.size(2), gradWeight.size(3)});
+
+    for (int g = 0; g < group; g++) {
+      gradWeight[g] = gradWeight[g]
+                          .flatten(1)
+                          .addmm_(gradOutputBuffer[elt][g].flatten(1),
+                                  columns[g].transpose(1, 0), 1.0, scale)
+                          .view_as(gradWeight[g]);
+    }
+    gradOutputBuffer = gradOutputBuffer.view(
+        {gradOutputBuffer.size(0),
+         gradOutputBuffer.size(1) * gradOutputBuffer.size(2),
+         gradOutputBuffer.size(3), gradOutputBuffer.size(4)});
+    columns =
+        columns.view({columns.size(0) * columns.size(1), columns.size(2)});
+    gradWeight = gradWeight.view({gradWeight.size(0) * gradWeight.size(1),
+                                  gradWeight.size(2), gradWeight.size(3),
+                                  gradWeight.size(4)});
+  }
+
+  input = input.view({batchSize, nInputPlane, inputHeight, inputWidth});
+  offset = offset.view(
+      {batchSize, deformable_group * 2 * kH * kW, outputHeight, outputWidth});
+
+  if (batch == 0) {
+    gradOutput = gradOutput.view({nOutputPlane, outputHeight, outputWidth});
+    input = input.view({nInputPlane, inputHeight, inputWidth});
+  }
+
+  return 1;
+}
+
+void modulated_deform_conv_cuda_forward(
+    at::Tensor input, at::Tensor weight, at::Tensor bias, at::Tensor ones,
+    at::Tensor offset, at::Tensor mask, at::Tensor output, at::Tensor columns,
+    int kernel_h, int kernel_w, const int stride_h, const int stride_w,
+    const int pad_h, const int pad_w, const int dilation_h,
+    const int dilation_w, const int group, const int deformable_group,
+    const bool with_bias) {
+  TORCH_CHECK(input.is_contiguous(), "input tensor has to be contiguous");
+  TORCH_CHECK(weight.is_contiguous(), "weight tensor has to be contiguous");
+  at::DeviceGuard guard(input.device());
+
+  const int batch = input.size(0);
+  const int channels = input.size(1);
+  const int height = input.size(2);
+  const int width = input.size(3);
+
+  const int channels_out = weight.size(0);
+  const int channels_kernel = weight.size(1);
+  const int kernel_h_ = weight.size(2);
+  const int kernel_w_ = weight.size(3);
+
+  if (kernel_h_ != kernel_h || kernel_w_ != kernel_w)
+    AT_ERROR("Input shape and kernel shape won't match: (%d x %d vs %d x %d).",
+             kernel_h_, kernel_w, kernel_h_, kernel_w_);
+  if (channels != channels_kernel * group)
+    AT_ERROR("Input shape and kernel channels won't match: (%d vs %d).",
+             channels, channels_kernel * group);
+
+  const int height_out =
+      (height + 2 * pad_h - (dilation_h * (kernel_h - 1) + 1)) / stride_h + 1;
+  const int width_out =
+      (width + 2 * pad_w - (dilation_w * (kernel_w - 1) + 1)) / stride_w + 1;
+
+  if (ones.ndimension() != 2 ||
+      ones.size(0) * ones.size(1) < height_out * width_out) {
+    // Resize plane and fill with ones...
+    ones = at::ones({height_out, width_out}, input.options());
+  }
+
+  // resize output
+  output = output.view({batch, channels_out, height_out, width_out}).zero_();
+  // resize temporary columns
+  columns =
+      at::zeros({channels * kernel_h * kernel_w, 1 * height_out * width_out},
+                input.options());
+
+  output = output.view({output.size(0), group, output.size(1) / group,
+                        output.size(2), output.size(3)});
+
+  for (int b = 0; b < batch; b++) {
+    modulated_deformable_im2col_cuda(
+        input[b], offset[b], mask[b], 1, channels, height, width, height_out,
+        width_out, kernel_h, kernel_w, pad_h, pad_w, stride_h, stride_w,
+        dilation_h, dilation_w, deformable_group, columns);
+
+    // divide into group
+    weight = weight.view({group, weight.size(0) / group, weight.size(1),
+                          weight.size(2), weight.size(3)});
+    columns = columns.view({group, columns.size(0) / group, columns.size(1)});
+
+    for (int g = 0; g < group; g++) {
+      output[b][g] = output[b][g]
+                         .flatten(1)
+                         .addmm_(weight[g].flatten(1), columns[g])
+                         .view_as(output[b][g]);
+    }
+
+    weight = weight.view({weight.size(0) * weight.size(1), weight.size(2),
+                          weight.size(3), weight.size(4)});
+    columns =
+        columns.view({columns.size(0) * columns.size(1), columns.size(2)});
+  }
+
+  output = output.view({output.size(0), output.size(1) * output.size(2),
+                        output.size(3), output.size(4)});
+
+  if (with_bias) {
+    output += bias.view({1, bias.size(0), 1, 1});
+  }
+}
+
+void modulated_deform_conv_cuda_backward(
+    at::Tensor input, at::Tensor weight, at::Tensor bias, at::Tensor ones,
+    at::Tensor offset, at::Tensor mask, at::Tensor columns,
+    at::Tensor grad_input, at::Tensor grad_weight, at::Tensor grad_bias,
+    at::Tensor grad_offset, at::Tensor grad_mask, at::Tensor grad_output,
+    int kernel_h, int kernel_w, int stride_h, int stride_w, int pad_h,
+    int pad_w, int dilation_h, int dilation_w, int group, int deformable_group,
+    const bool with_bias) {
+  TORCH_CHECK(input.is_contiguous(), "input tensor has to be contiguous");
+  TORCH_CHECK(weight.is_contiguous(), "weight tensor has to be contiguous");
+  at::DeviceGuard guard(input.device());
+
+  const int batch = input.size(0);
+  const int channels = input.size(1);
+  const int height = input.size(2);
+  const int width = input.size(3);
+
+  const int channels_kernel = weight.size(1);
+  const int kernel_h_ = weight.size(2);
+  const int kernel_w_ = weight.size(3);
+  if (kernel_h_ != kernel_h || kernel_w_ != kernel_w)
+    AT_ERROR("Input shape and kernel shape won't match: (%d x %d vs %d x %d).",
+             kernel_h_, kernel_w, kernel_h_, kernel_w_);
+  if (channels != channels_kernel * group)
+    AT_ERROR("Input shape and kernel channels won't match: (%d vs %d).",
+             channels, channels_kernel * group);
+
+  const int height_out =
+      (height + 2 * pad_h - (dilation_h * (kernel_h - 1) + 1)) / stride_h + 1;
+  const int width_out =
+      (width + 2 * pad_w - (dilation_w * (kernel_w - 1) + 1)) / stride_w + 1;
+
+  if (ones.ndimension() != 2 ||
+      ones.size(0) * ones.size(1) < height_out * width_out) {
+    // Resize plane and fill with ones...
+    ones = at::ones({height_out, width_out}, input.options());
+  }
+
+  grad_input = grad_input.view({batch, channels, height, width});
+  columns = at::zeros({channels * kernel_h * kernel_w, height_out * width_out},
+                      input.options());
+
+  grad_output =
+      grad_output.view({grad_output.size(0), group, grad_output.size(1) / group,
+                        grad_output.size(2), grad_output.size(3)});
+
+  for (int b = 0; b < batch; b++) {
+    // divide int group
+    columns = columns.view({group, columns.size(0) / group, columns.size(1)});
+    weight = weight.view({group, weight.size(0) / group, weight.size(1),
+                          weight.size(2), weight.size(3)});
+
+    for (int g = 0; g < group; g++) {
+      columns[g].addmm_(weight[g].flatten(1).transpose(0, 1),
+                        grad_output[b][g].flatten(1), 0.0f, 1.0f);
+    }
+
+    columns =
+        columns.view({columns.size(0) * columns.size(1), columns.size(2)});
+    weight = weight.view({weight.size(0) * weight.size(1), weight.size(2),
+                          weight.size(3), weight.size(4)});
+
+    // gradient w.r.t. input coordinate data
+    modulated_deformable_col2im_coord_cuda(
+        columns, input[b], offset[b], mask[b], 1, channels, height, width,
+        height_out, width_out, kernel_h, kernel_w, pad_h, pad_w, stride_h,
+        stride_w, dilation_h, dilation_w, deformable_group, grad_offset[b],
+        grad_mask[b]);
+    // gradient w.r.t. input data
+    modulated_deformable_col2im_cuda(
+        columns, offset[b], mask[b], 1, channels, height, width, height_out,
+        width_out, kernel_h, kernel_w, pad_h, pad_w, stride_h, stride_w,
+        dilation_h, dilation_w, deformable_group, grad_input[b]);
+
+    // gradient w.r.t. weight, dWeight should accumulate across the batch and
+    // group
+    modulated_deformable_im2col_cuda(
+        input[b], offset[b], mask[b], 1, channels, height, width, height_out,
+        width_out, kernel_h, kernel_w, pad_h, pad_w, stride_h, stride_w,
+        dilation_h, dilation_w, deformable_group, columns);
+
+    columns = columns.view({group, columns.size(0) / group, columns.size(1)});
+    grad_weight = grad_weight.view({group, grad_weight.size(0) / group,
+                                    grad_weight.size(1), grad_weight.size(2),
+                                    grad_weight.size(3)});
+    if (with_bias)
+      grad_bias = grad_bias.view({group, grad_bias.size(0) / group});
+
+    for (int g = 0; g < group; g++) {
+      grad_weight[g] =
+          grad_weight[g]
+              .flatten(1)
+              .addmm_(grad_output[b][g].flatten(1), columns[g].transpose(0, 1))
+              .view_as(grad_weight[g]);
+      if (with_bias) {
+        grad_bias[g] =
+            grad_bias[g]
+                .view({-1, 1})
+                .addmm_(grad_output[b][g].flatten(1), ones.view({-1, 1}))
+                .view(-1);
+      }
+    }
+
+    columns =
+        columns.view({columns.size(0) * columns.size(1), columns.size(2)});
+    grad_weight = grad_weight.view({grad_weight.size(0) * grad_weight.size(1),
+                                    grad_weight.size(2), grad_weight.size(3),
+                                    grad_weight.size(4)});
+    if (with_bias)
+      grad_bias = grad_bias.view({grad_bias.size(0) * grad_bias.size(1)});
+  }
+  grad_output = grad_output.view({grad_output.size(0) * grad_output.size(1),
+                                  grad_output.size(2), grad_output.size(3),
+                                  grad_output.size(4)});
+}

basicsr/ops/dcn/src/deform_conv_cuda_kernel.cu

+/*!
+ ******************* BEGIN Caffe Copyright Notice and Disclaimer ****************
+ *
+ * COPYRIGHT
+ *
+ * All contributions by the University of California:
+ * Copyright (c) 2014-2017 The Regents of the University of California (Regents)
+ * All rights reserved.
+ *
+ * All other contributions:
+ * Copyright (c) 2014-2017, the respective contributors
+ * All rights reserved.
+ *
+ * Caffe uses a shared copyright model: each contributor holds copyright over
+ * their contributions to Caffe. The project versioning records all such
+ * contribution and copyright details. If a contributor wants to further mark
+ * their specific copyright on a particular contribution, they should indicate
+ * their copyright solely in the commit message of the change when it is
+ * committed.
+ *
+ * LICENSE
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are met:
+ *
+ * 1. Redistributions of source code must retain the above copyright notice, this
+ * list of conditions and the following disclaimer.
+ * 2. Redistributions in binary form must reproduce the above copyright notice,
+ * this list of conditions and the following disclaimer in the documentation
+ * and/or other materials provided with the distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
+ * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ * CONTRIBUTION AGREEMENT
+ *
+ * By contributing to the BVLC/caffe repository through pull-request, comment,
+ * or otherwise, the contributor releases their content to the
+ * license and copyright terms herein.
+ *
+ ***************** END Caffe Copyright Notice and Disclaimer ********************
+ *
+ * Copyright (c) 2018 Microsoft
+ * Licensed under The MIT License [see LICENSE for details]
+ * \file modulated_deformable_im2col.cuh
+ * \brief Function definitions of converting an image to
+ * column matrix based on kernel, padding, dilation, and offset.
+ * These functions are mainly used in deformable convolution operators.
+ * \ref: https://arxiv.org/abs/1703.06211
+ * \author Yuwen Xiong, Haozhi Qi, Jifeng Dai, Xizhou Zhu, Han Hu, Dazhi Cheng
+ */
+
+// modified from https://github.com/chengdazhi/Deformable-Convolution-V2-PyTorch/blob/mmdetection/mmdet/ops/dcn/src/deform_conv_cuda_kernel.cu
+
+#include <ATen/ATen.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <THC/THCAtomics.cuh>
+#include <stdio.h>
+#include <math.h>
+#include <float.h>
+
+using namespace at;
+
+#define CUDA_KERNEL_LOOP(i, n)                                 \
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < (n); \
+       i += blockDim.x * gridDim.x)
+
+const int CUDA_NUM_THREADS = 1024;
+const int kMaxGridNum = 65535;
+
+inline int GET_BLOCKS(const int N)
+{
+  return std::min(kMaxGridNum, (N + CUDA_NUM_THREADS - 1) / CUDA_NUM_THREADS);
+}
+
+template <typename scalar_t>
+__device__ scalar_t deformable_im2col_bilinear(const scalar_t *bottom_data, const int data_width,
+                                               const int height, const int width, scalar_t h, scalar_t w)
+{
+
+  int h_low = floor(h);
+  int w_low = floor(w);
+  int h_high = h_low + 1;
+  int w_high = w_low + 1;
+
+  scalar_t lh = h - h_low;
+  scalar_t lw = w - w_low;
+  scalar_t hh = 1 - lh, hw = 1 - lw;
+
+  scalar_t v1 = 0;
+  if (h_low >= 0 && w_low >= 0)
+    v1 = bottom_data[h_low * data_width + w_low];
+  scalar_t v2 = 0;
+  if (h_low >= 0 && w_high <= width - 1)
+    v2 = bottom_data[h_low * data_width + w_high];
+  scalar_t v3 = 0;
+  if (h_high <= height - 1 && w_low >= 0)
+    v3 = bottom_data[h_high * data_width + w_low];
+  scalar_t v4 = 0;
+  if (h_high <= height - 1 && w_high <= width - 1)
+    v4 = bottom_data[h_high * data_width + w_high];
+
+  scalar_t w1 = hh * hw, w2 = hh * lw, w3 = lh * hw, w4 = lh * lw;
+
+  scalar_t val = (w1 * v1 + w2 * v2 + w3 * v3 + w4 * v4);
+  return val;
+}
+
+template <typename scalar_t>
+__device__ scalar_t get_gradient_weight(scalar_t argmax_h, scalar_t argmax_w,
+                                        const int h, const int w, const int height, const int width)
+{
+
+  if (argmax_h <= -1 || argmax_h >= height || argmax_w <= -1 || argmax_w >= width)
+  {
+    //empty
+    return 0;
+  }
+
+  int argmax_h_low = floor(argmax_h);
+  int argmax_w_low = floor(argmax_w);
+  int argmax_h_high = argmax_h_low + 1;
+  int argmax_w_high = argmax_w_low + 1;
+
+  scalar_t weight = 0;
+  if (h == argmax_h_low && w == argmax_w_low)
+    weight = (h + 1 - argmax_h) * (w + 1 - argmax_w);
+  if (h == argmax_h_low && w == argmax_w_high)
+    weight = (h + 1 - argmax_h) * (argmax_w + 1 - w);
+  if (h == argmax_h_high && w == argmax_w_low)
+    weight = (argmax_h + 1 - h) * (w + 1 - argmax_w);
+  if (h == argmax_h_high && w == argmax_w_high)
+    weight = (argmax_h + 1 - h) * (argmax_w + 1 - w);
+  return weight;
+}
+
+template <typename scalar_t>
+__device__ scalar_t get_coordinate_weight(scalar_t argmax_h, scalar_t argmax_w,
+                                          const int height, const int width, const scalar_t *im_data,
+                                          const int data_width, const int bp_dir)
+{
+
+  if (argmax_h <= -1 || argmax_h >= height || argmax_w <= -1 || argmax_w >= width)
+  {
+    //empty
+    return 0;
+  }
+
+  int argmax_h_low = floor(argmax_h);
+  int argmax_w_low = floor(argmax_w);
+  int argmax_h_high = argmax_h_low + 1;
+  int argmax_w_high = argmax_w_low + 1;
+
+  scalar_t weight = 0;
+
+  if (bp_dir == 0)
+  {
+    if (argmax_h_low >= 0 && argmax_w_low >= 0)
+      weight += -1 * (argmax_w_low + 1 - argmax_w) * im_data[argmax_h_low * data_width + argmax_w_low];
+    if (argmax_h_low >= 0 && argmax_w_high <= width - 1)
+      weight += -1 * (argmax_w - argmax_w_low) * im_data[argmax_h_low * data_width + argmax_w_high];
+    if (argmax_h_high <= height - 1 && argmax_w_low >= 0)
+      weight += (argmax_w_low + 1 - argmax_w) * im_data[argmax_h_high * data_width + argmax_w_low];
+    if (argmax_h_high <= height - 1 && argmax_w_high <= width - 1)
+      weight += (argmax_w - argmax_w_low) * im_data[argmax_h_high * data_width + argmax_w_high];
+  }
+  else if (bp_dir == 1)
+  {
+    if (argmax_h_low >= 0 && argmax_w_low >= 0)
+      weight += -1 * (argmax_h_low + 1 - argmax_h) * im_data[argmax_h_low * data_width + argmax_w_low];
+    if (argmax_h_low >= 0 && argmax_w_high <= width - 1)
+      weight += (argmax_h_low + 1 - argmax_h) * im_data[argmax_h_low * data_width + argmax_w_high];
+    if (argmax_h_high <= height - 1 && argmax_w_low >= 0)
+      weight += -1 * (argmax_h - argmax_h_low) * im_data[argmax_h_high * data_width + argmax_w_low];
+    if (argmax_h_high <= height - 1 && argmax_w_high <= width - 1)
+      weight += (argmax_h - argmax_h_low) * im_data[argmax_h_high * data_width + argmax_w_high];
+  }
+
+  return weight;
+}
+
+template <typename scalar_t>
+__global__ void deformable_im2col_gpu_kernel(const int n, const scalar_t *data_im, const scalar_t *data_offset,
+                                             const int height, const int width, const int kernel_h, const int kernel_w,
+                                             const int pad_h, const int pad_w, const int stride_h, const int stride_w,
+                                             const int dilation_h, const int dilation_w, const int channel_per_deformable_group,
+                                             const int batch_size, const int num_channels, const int deformable_group,
+                                             const int height_col, const int width_col,
+                                             scalar_t *data_col)
+{
+  CUDA_KERNEL_LOOP(index, n)
+  {
+    // index index of output matrix
+    const int w_col = index % width_col;
+    const int h_col = (index / width_col) % height_col;
+    const int b_col = (index / width_col / height_col) % batch_size;
+    const int c_im = (index / width_col / height_col) / batch_size;
+    const int c_col = c_im * kernel_h * kernel_w;
+
+    // compute deformable group index
+    const int deformable_group_index = c_im / channel_per_deformable_group;
+
+    const int h_in = h_col * stride_h - pad_h;
+    const int w_in = w_col * stride_w - pad_w;
+    scalar_t *data_col_ptr = data_col + ((c_col * batch_size + b_col) * height_col + h_col) * width_col + w_col;
+    //const scalar_t* data_im_ptr = data_im + ((b_col * num_channels + c_im) * height + h_in) * width + w_in;
+    const scalar_t *data_im_ptr = data_im + (b_col * num_channels + c_im) * height * width;
+    const scalar_t *data_offset_ptr = data_offset + (b_col * deformable_group + deformable_group_index) * 2 * kernel_h * kernel_w * height_col * width_col;
+
+    for (int i = 0; i < kernel_h; ++i)
+    {
+      for (int j = 0; j < kernel_w; ++j)
+      {
+        const int data_offset_h_ptr = ((2 * (i * kernel_w + j)) * height_col + h_col) * width_col + w_col;
+        const int data_offset_w_ptr = ((2 * (i * kernel_w + j) + 1) * height_col + h_col) * width_col + w_col;
+        const scalar_t offset_h = data_offset_ptr[data_offset_h_ptr];
+        const scalar_t offset_w = data_offset_ptr[data_offset_w_ptr];
+        scalar_t val = static_cast<scalar_t>(0);
+        const scalar_t h_im = h_in + i * dilation_h + offset_h;
+        const scalar_t w_im = w_in + j * dilation_w + offset_w;
+        if (h_im > -1 && w_im > -1 && h_im < height && w_im < width)
+        {
+          //const scalar_t map_h = i * dilation_h + offset_h;
+          //const scalar_t map_w = j * dilation_w + offset_w;
+          //const int cur_height = height - h_in;
+          //const int cur_width = width - w_in;
+          //val = deformable_im2col_bilinear(data_im_ptr, width, cur_height, cur_width, map_h, map_w);
+          val = deformable_im2col_bilinear(data_im_ptr, width, height, width, h_im, w_im);
+        }
+        *data_col_ptr = val;
+        data_col_ptr += batch_size * height_col * width_col;
+      }
+    }
+  }
+}
+
+void deformable_im2col(
+    const at::Tensor data_im, const at::Tensor data_offset, const int channels,
+    const int height, const int width, const int ksize_h, const int ksize_w,
+    const int pad_h, const int pad_w, const int stride_h, const int stride_w,
+    const int dilation_h, const int dilation_w, const int parallel_imgs,
+    const int deformable_group, at::Tensor data_col)
+{
+  // num_axes should be smaller than block size
+  // todo: check parallel_imgs is correctly passed in
+  int height_col = (height + 2 * pad_h - (dilation_h * (ksize_h - 1) + 1)) / stride_h + 1;
+  int width_col = (width + 2 * pad_w - (dilation_w * (ksize_w - 1) + 1)) / stride_w + 1;
+  int num_kernels = channels * height_col * width_col * parallel_imgs;
+  int channel_per_deformable_group = channels / deformable_group;
+
+  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
+      data_im.scalar_type(), "deformable_im2col_gpu", ([&] {
+        const scalar_t *data_im_ = data_im.data_ptr<scalar_t>();
+        const scalar_t *data_offset_ = data_offset.data_ptr<scalar_t>();
+        scalar_t *data_col_ = data_col.data_ptr<scalar_t>();
+
+        deformable_im2col_gpu_kernel<<<GET_BLOCKS(num_kernels), CUDA_NUM_THREADS, 0, at::cuda::getCurrentCUDAStream()>>>(
+            num_kernels, data_im_, data_offset_, height, width, ksize_h, ksize_w,
+            pad_h, pad_w, stride_h, stride_w, dilation_h, dilation_w,
+            channel_per_deformable_group, parallel_imgs, channels, deformable_group,
+            height_col, width_col, data_col_);
+      }));
+
+  cudaError_t err = cudaGetLastError();
+  if (err != cudaSuccess)
+  {
+    printf("error in deformable_im2col: %s\n", cudaGetErrorString(err));
+  }
+}
+
+template <typename scalar_t>
+__global__ void deformable_col2im_gpu_kernel(
+    const int n, const scalar_t *data_col, const scalar_t *data_offset,
+    const int channels, const int height, const int width,
+    const int kernel_h, const int kernel_w,
+    const int pad_h, const int pad_w,
+    const int stride_h, const int stride_w,
+    const int dilation_h, const int dilation_w,
+    const int channel_per_deformable_group,
+    const int batch_size, const int deformable_group,
+    const int height_col, const int width_col,
+    scalar_t *grad_im)
+{
+  CUDA_KERNEL_LOOP(index, n)
+  {
+    const int j = (index / width_col / height_col / batch_size) % kernel_w;
+    const int i = (index / width_col / height_col / batch_size / kernel_w) % kernel_h;
+    const int c = index / width_col / height_col / batch_size / kernel_w / kernel_h;
+    // compute the start and end of the output
+
+    const int deformable_group_index = c / channel_per_deformable_group;
+
+    int w_out = index % width_col;
+    int h_out = (index / width_col) % height_col;
+    int b = (index / width_col / height_col) % batch_size;
+    int w_in = w_out * stride_w - pad_w;
+    int h_in = h_out * stride_h - pad_h;
+
+    const scalar_t *data_offset_ptr = data_offset + (b * deformable_group + deformable_group_index) *
+                                                        2 * kernel_h * kernel_w * height_col * width_col;
+    const int data_offset_h_ptr = ((2 * (i * kernel_w + j)) * height_col + h_out) * width_col + w_out;
+    const int data_offset_w_ptr = ((2 * (i * kernel_w + j) + 1) * height_col + h_out) * width_col + w_out;
+    const scalar_t offset_h = data_offset_ptr[data_offset_h_ptr];
+    const scalar_t offset_w = data_offset_ptr[data_offset_w_ptr];
+    const scalar_t cur_inv_h_data = h_in + i * dilation_h + offset_h;
+    const scalar_t cur_inv_w_data = w_in + j * dilation_w + offset_w;
+
+    const scalar_t cur_top_grad = data_col[index];
+    const int cur_h = (int)cur_inv_h_data;
+    const int cur_w = (int)cur_inv_w_data;
+    for (int dy = -2; dy <= 2; dy++)
+    {
+      for (int dx = -2; dx <= 2; dx++)
+      {
+        if (cur_h + dy >= 0 && cur_h + dy < height &&
+            cur_w + dx >= 0 && cur_w + dx < width &&
+            abs(cur_inv_h_data - (cur_h + dy)) < 1 &&
+            abs(cur_inv_w_data - (cur_w + dx)) < 1)
+        {
+          int cur_bottom_grad_pos = ((b * channels + c) * height + cur_h + dy) * width + cur_w + dx;
+          scalar_t weight = get_gradient_weight(cur_inv_h_data, cur_inv_w_data, cur_h + dy, cur_w + dx, height, width);
+          atomicAdd(grad_im + cur_bottom_grad_pos, weight * cur_top_grad);
+        }
+      }
+    }
+  }
+}
+
+void deformable_col2im(
+    const at::Tensor data_col, const at::Tensor data_offset, const int channels,
+    const int height, const int width, const int ksize_h,
+    const int ksize_w, const int pad_h, const int pad_w,
+    const int stride_h, const int stride_w,
+    const int dilation_h, const int dilation_w,
+    const int parallel_imgs, const int deformable_group,
+    at::Tensor grad_im)
+{
+
+  // todo: make sure parallel_imgs is passed in correctly
+  int height_col = (height + 2 * pad_h - (dilation_h * (ksize_h - 1) + 1)) / stride_h + 1;
+  int width_col = (width + 2 * pad_w - (dilation_w * (ksize_w - 1) + 1)) / stride_w + 1;
+  int num_kernels = channels * ksize_h * ksize_w * height_col * width_col * parallel_imgs;
+  int channel_per_deformable_group = channels / deformable_group;
+
+  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
+      data_col.scalar_type(), "deformable_col2im_gpu", ([&] {
+        const scalar_t *data_col_ = data_col.data_ptr<scalar_t>();
+        const scalar_t *data_offset_ = data_offset.data_ptr<scalar_t>();
+        scalar_t *grad_im_ = grad_im.data_ptr<scalar_t>();
+
+        deformable_col2im_gpu_kernel<<<GET_BLOCKS(num_kernels), CUDA_NUM_THREADS, 0, at::cuda::getCurrentCUDAStream()>>>(
+            num_kernels, data_col_, data_offset_, channels, height, width, ksize_h,
+            ksize_w, pad_h, pad_w, stride_h, stride_w,
+            dilation_h, dilation_w, channel_per_deformable_group,
+            parallel_imgs, deformable_group, height_col, width_col, grad_im_);
+      }));
+
+  cudaError_t err = cudaGetLastError();
+  if (err != cudaSuccess)
+  {
+    printf("error in deformable_col2im: %s\n", cudaGetErrorString(err));
+  }
+}
+
+template <typename scalar_t>
+__global__ void deformable_col2im_coord_gpu_kernel(const int n, const scalar_t *data_col,
+                                                   const scalar_t *data_im, const scalar_t *data_offset,
+                                                   const int channels, const int height, const int width,
+                                                   const int kernel_h, const int kernel_w,
+                                                   const int pad_h, const int pad_w,
+                                                   const int stride_h, const int stride_w,
+                                                   const int dilation_h, const int dilation_w,
+                                                   const int channel_per_deformable_group,
+                                                   const int batch_size, const int offset_channels, const int deformable_group,
+                                                   const int height_col, const int width_col, scalar_t *grad_offset)
+{
+  CUDA_KERNEL_LOOP(index, n)
+  {
+    scalar_t val = 0;
+    int w = index % width_col;
+    int h = (index / width_col) % height_col;
+    int c = (index / width_col / height_col) % offset_channels;
+    int b = (index / width_col / height_col) / offset_channels;
+    // compute the start and end of the output
+
+    const int deformable_group_index = c / (2 * kernel_h * kernel_w);
+    const int col_step = kernel_h * kernel_w;
+    int cnt = 0;
+    const scalar_t *data_col_ptr = data_col + deformable_group_index * channel_per_deformable_group *
+                                                  batch_size * width_col * height_col;
+    const scalar_t *data_im_ptr = data_im + (b * deformable_group + deformable_group_index) *
+                                                channel_per_deformable_group / kernel_h / kernel_w * height * width;
+    const scalar_t *data_offset_ptr = data_offset + (b * deformable_group + deformable_group_index) * 2 *
+                                                        kernel_h * kernel_w * height_col * width_col;
+
+    const int offset_c = c - deformable_group_index * 2 * kernel_h * kernel_w;
+
+    for (int col_c = (offset_c / 2); col_c < channel_per_deformable_group; col_c += col_step)
+    {
+      const int col_pos = (((col_c * batch_size + b) * height_col) + h) * width_col + w;
+      const int bp_dir = offset_c % 2;
+
+      int j = (col_pos / width_col / height_col / batch_size) % kernel_w;
+      int i = (col_pos / width_col / height_col / batch_size / kernel_w) % kernel_h;
+      int w_out = col_pos % width_col;
+      int h_out = (col_pos / width_col) % height_col;
+      int w_in = w_out * stride_w - pad_w;
+      int h_in = h_out * stride_h - pad_h;
+      const int data_offset_h_ptr = (((2 * (i * kernel_w + j)) * height_col + h_out) * width_col + w_out);
+      const int data_offset_w_ptr = (((2 * (i * kernel_w + j) + 1) * height_col + h_out) * width_col + w_out);
+      const scalar_t offset_h = data_offset_ptr[data_offset_h_ptr];
+      const scalar_t offset_w = data_offset_ptr[data_offset_w_ptr];
+      scalar_t inv_h = h_in + i * dilation_h + offset_h;
+      scalar_t inv_w = w_in + j * dilation_w + offset_w;
+      if (inv_h <= -1 || inv_w <= -1 || inv_h >= height || inv_w >= width)
+      {
+        inv_h = inv_w = -2;
+      }
+      const scalar_t weight = get_coordinate_weight(
+          inv_h, inv_w,
+          height, width, data_im_ptr + cnt * height * width, width, bp_dir);
+      val += weight * data_col_ptr[col_pos];
+      cnt += 1;
+    }
+
+    grad_offset[index] = val;
+  }
+}
+
+void deformable_col2im_coord(
+    const at::Tensor data_col, const at::Tensor data_im, const at::Tensor data_offset,
+    const int channels, const int height, const int width, const int ksize_h,
+    const int ksize_w, const int pad_h, const int pad_w, const int stride_h,
+    const int stride_w, const int dilation_h, const int dilation_w,
+    const int parallel_imgs, const int deformable_group, at::Tensor grad_offset)
+{
+
+  int height_col = (height + 2 * pad_h - (dilation_h * (ksize_h - 1) + 1)) / stride_h + 1;
+  int width_col = (width + 2 * pad_w - (dilation_w * (ksize_w - 1) + 1)) / stride_w + 1;
+  int num_kernels = height_col * width_col * 2 * ksize_h * ksize_w * deformable_group * parallel_imgs;
+  int channel_per_deformable_group = channels * ksize_h * ksize_w / deformable_group;
+
+  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
+      data_col.scalar_type(), "deformable_col2im_coord_gpu", ([&] {
+        const scalar_t *data_col_ = data_col.data_ptr<scalar_t>();
+        const scalar_t *data_im_ = data_im.data_ptr<scalar_t>();
+        const scalar_t *data_offset_ = data_offset.data_ptr<scalar_t>();
+        scalar_t *grad_offset_ = grad_offset.data_ptr<scalar_t>();
+
+        deformable_col2im_coord_gpu_kernel<<<GET_BLOCKS(num_kernels), CUDA_NUM_THREADS, 0, at::cuda::getCurrentCUDAStream()>>>(
+            num_kernels, data_col_, data_im_, data_offset_, channels, height, width,
+            ksize_h, ksize_w, pad_h, pad_w, stride_h, stride_w,
+            dilation_h, dilation_w, channel_per_deformable_group,
+            parallel_imgs, 2 * ksize_h * ksize_w * deformable_group, deformable_group,
+            height_col, width_col, grad_offset_);
+      }));
+}
+
+template <typename scalar_t>
+__device__ scalar_t dmcn_im2col_bilinear(const scalar_t *bottom_data, const int data_width,
+                                         const int height, const int width, scalar_t h, scalar_t w)
+{
+  int h_low = floor(h);
+  int w_low = floor(w);
+  int h_high = h_low + 1;
+  int w_high = w_low + 1;
+
+  scalar_t lh = h - h_low;
+  scalar_t lw = w - w_low;
+  scalar_t hh = 1 - lh, hw = 1 - lw;
+
+  scalar_t v1 = 0;
+  if (h_low >= 0 && w_low >= 0)
+    v1 = bottom_data[h_low * data_width + w_low];
+  scalar_t v2 = 0;
+  if (h_low >= 0 && w_high <= width - 1)
+    v2 = bottom_data[h_low * data_width + w_high];
+  scalar_t v3 = 0;
+  if (h_high <= height - 1 && w_low >= 0)
+    v3 = bottom_data[h_high * data_width + w_low];
+  scalar_t v4 = 0;
+  if (h_high <= height - 1 && w_high <= width - 1)
+    v4 = bottom_data[h_high * data_width + w_high];
+
+  scalar_t w1 = hh * hw, w2 = hh * lw, w3 = lh * hw, w4 = lh * lw;
+
+  scalar_t val = (w1 * v1 + w2 * v2 + w3 * v3 + w4 * v4);
+  return val;
+}
+
+template <typename scalar_t>
+__device__ scalar_t dmcn_get_gradient_weight(scalar_t argmax_h, scalar_t argmax_w,
+                                             const int h, const int w, const int height, const int width)
+{
+  if (argmax_h <= -1 || argmax_h >= height || argmax_w <= -1 || argmax_w >= width)
+  {
+    //empty
+    return 0;
+  }
+
+  int argmax_h_low = floor(argmax_h);
+  int argmax_w_low = floor(argmax_w);
+  int argmax_h_high = argmax_h_low + 1;
+  int argmax_w_high = argmax_w_low + 1;
+
+  scalar_t weight = 0;
+  if (h == argmax_h_low && w == argmax_w_low)
+    weight = (h + 1 - argmax_h) * (w + 1 - argmax_w);
+  if (h == argmax_h_low && w == argmax_w_high)
+    weight = (h + 1 - argmax_h) * (argmax_w + 1 - w);
+  if (h == argmax_h_high && w == argmax_w_low)
+    weight = (argmax_h + 1 - h) * (w + 1 - argmax_w);
+  if (h == argmax_h_high && w == argmax_w_high)
+    weight = (argmax_h + 1 - h) * (argmax_w + 1 - w);
+  return weight;
+}
+
+template <typename scalar_t>
+__device__ scalar_t dmcn_get_coordinate_weight(scalar_t argmax_h, scalar_t argmax_w,
+                                               const int height, const int width, const scalar_t *im_data,
+                                               const int data_width, const int bp_dir)
+{
+  if (argmax_h <= -1 || argmax_h >= height || argmax_w <= -1 || argmax_w >= width)
+  {
+    //empty
+    return 0;
+  }
+
+  int argmax_h_low = floor(argmax_h);
+  int argmax_w_low = floor(argmax_w);
+  int argmax_h_high = argmax_h_low + 1;
+  int argmax_w_high = argmax_w_low + 1;
+
+  scalar_t weight = 0;
+
+  if (bp_dir == 0)
+  {
+    if (argmax_h_low >= 0 && argmax_w_low >= 0)
+      weight += -1 * (argmax_w_low + 1 - argmax_w) * im_data[argmax_h_low * data_width + argmax_w_low];
+    if (argmax_h_low >= 0 && argmax_w_high <= width - 1)
+      weight += -1 * (argmax_w - argmax_w_low) * im_data[argmax_h_low * data_width + argmax_w_high];
+    if (argmax_h_high <= height - 1 && argmax_w_low >= 0)
+      weight += (argmax_w_low + 1 - argmax_w) * im_data[argmax_h_high * data_width + argmax_w_low];
+    if (argmax_h_high <= height - 1 && argmax_w_high <= width - 1)
+      weight += (argmax_w - argmax_w_low) * im_data[argmax_h_high * data_width + argmax_w_high];
+  }
+  else if (bp_dir == 1)
+  {
+    if (argmax_h_low >= 0 && argmax_w_low >= 0)
+      weight += -1 * (argmax_h_low + 1 - argmax_h) * im_data[argmax_h_low * data_width + argmax_w_low];
+    if (argmax_h_low >= 0 && argmax_w_high <= width - 1)
+      weight += (argmax_h_low + 1 - argmax_h) * im_data[argmax_h_low * data_width + argmax_w_high];
+    if (argmax_h_high <= height - 1 && argmax_w_low >= 0)
+      weight += -1 * (argmax_h - argmax_h_low) * im_data[argmax_h_high * data_width + argmax_w_low];
+    if (argmax_h_high <= height - 1 && argmax_w_high <= width - 1)
+      weight += (argmax_h - argmax_h_low) * im_data[argmax_h_high * data_width + argmax_w_high];
+  }
+
+  return weight;
+}
+
+template <typename scalar_t>
+__global__ void modulated_deformable_im2col_gpu_kernel(const int n,
+                                                       const scalar_t *data_im, const scalar_t *data_offset, const scalar_t *data_mask,
+                                                       const int height, const int width, const int kernel_h, const int kernel_w,
+                                                       const int pad_h, const int pad_w,
+                                                       const int stride_h, const int stride_w,
+                                                       const int dilation_h, const int dilation_w,
+                                                       const int channel_per_deformable_group,
+                                                       const int batch_size, const int num_channels, const int deformable_group,
+                                                       const int height_col, const int width_col,
+                                                       scalar_t *data_col)
+{
+  CUDA_KERNEL_LOOP(index, n)
+  {
+    // index index of output matrix
+    const int w_col = index % width_col;
+    const int h_col = (index / width_col) % height_col;
+    const int b_col = (index / width_col / height_col) % batch_size;
+    const int c_im = (index / width_col / height_col) / batch_size;
+    const int c_col = c_im * kernel_h * kernel_w;
+
+    // compute deformable group index
+    const int deformable_group_index = c_im / channel_per_deformable_group;
+
+    const int h_in = h_col * stride_h - pad_h;
+    const int w_in = w_col * stride_w - pad_w;
+
+    scalar_t *data_col_ptr = data_col + ((c_col * batch_size + b_col) * height_col + h_col) * width_col + w_col;
+    //const float* data_im_ptr = data_im + ((b_col * num_channels + c_im) * height + h_in) * width + w_in;
+    const scalar_t *data_im_ptr = data_im + (b_col * num_channels + c_im) * height * width;
+    const scalar_t *data_offset_ptr = data_offset + (b_col * deformable_group + deformable_group_index) * 2 * kernel_h * kernel_w * height_col * width_col;
+
+    const scalar_t *data_mask_ptr = data_mask + (b_col * deformable_group + deformable_group_index) * kernel_h * kernel_w * height_col * width_col;
+
+    for (int i = 0; i < kernel_h; ++i)
+    {
+      for (int j = 0; j < kernel_w; ++j)
+      {
+        const int data_offset_h_ptr = ((2 * (i * kernel_w + j)) * height_col + h_col) * width_col + w_col;
+        const int data_offset_w_ptr = ((2 * (i * kernel_w + j) + 1) * height_col + h_col) * width_col + w_col;
+        const int data_mask_hw_ptr = ((i * kernel_w + j) * height_col + h_col) * width_col + w_col;
+        const scalar_t offset_h = data_offset_ptr[data_offset_h_ptr];
+        const scalar_t offset_w = data_offset_ptr[data_offset_w_ptr];
+        const scalar_t mask = data_mask_ptr[data_mask_hw_ptr];
+        scalar_t val = static_cast<scalar_t>(0);
+        const scalar_t h_im = h_in + i * dilation_h + offset_h;
+        const scalar_t w_im = w_in + j * dilation_w + offset_w;
+        //if (h_im >= 0 && w_im >= 0 && h_im < height && w_im < width) {
+        if (h_im > -1 && w_im > -1 && h_im < height && w_im < width)
+        {
+          //const float map_h = i * dilation_h + offset_h;
+          //const float map_w = j * dilation_w + offset_w;
+          //const int cur_height = height - h_in;
+          //const int cur_width = width - w_in;
+          //val = dmcn_im2col_bilinear(data_im_ptr, width, cur_height, cur_width, map_h, map_w);
+          val = dmcn_im2col_bilinear(data_im_ptr, width, height, width, h_im, w_im);
+        }
+        *data_col_ptr = val * mask;
+        data_col_ptr += batch_size * height_col * width_col;
+        //data_col_ptr += height_col * width_col;
+      }
+    }
+  }
+}
+
+template <typename scalar_t>
+__global__ void modulated_deformable_col2im_gpu_kernel(const int n,
+                                                       const scalar_t *data_col, const scalar_t *data_offset, const scalar_t *data_mask,
+                                                       const int channels, const int height, const int width,
+                                                       const int kernel_h, const int kernel_w,
+                                                       const int pad_h, const int pad_w,
+                                                       const int stride_h, const int stride_w,
+                                                       const int dilation_h, const int dilation_w,
+                                                       const int channel_per_deformable_group,
+                                                       const int batch_size, const int deformable_group,
+                                                       const int height_col, const int width_col,
+                                                       scalar_t *grad_im)
+{
+  CUDA_KERNEL_LOOP(index, n)
+  {
+    const int j = (index / width_col / height_col / batch_size) % kernel_w;
+    const int i = (index / width_col / height_col / batch_size / kernel_w) % kernel_h;
+    const int c = index / width_col / height_col / batch_size / kernel_w / kernel_h;
+    // compute the start and end of the output
+
+    const int deformable_group_index = c / channel_per_deformable_group;
+
+    int w_out = index % width_col;
+    int h_out = (index / width_col) % height_col;
+    int b = (index / width_col / height_col) % batch_size;
+    int w_in = w_out * stride_w - pad_w;
+    int h_in = h_out * stride_h - pad_h;
+
+    const scalar_t *data_offset_ptr = data_offset + (b * deformable_group + deformable_group_index) * 2 * kernel_h * kernel_w * height_col * width_col;
+    const scalar_t *data_mask_ptr = data_mask + (b * deformable_group + deformable_group_index) * kernel_h * kernel_w * height_col * width_col;
+    const int data_offset_h_ptr = ((2 * (i * kernel_w + j)) * height_col + h_out) * width_col + w_out;
+    const int data_offset_w_ptr = ((2 * (i * kernel_w + j) + 1) * height_col + h_out) * width_col + w_out;
+    const int data_mask_hw_ptr = ((i * kernel_w + j) * height_col + h_out) * width_col + w_out;
+    const scalar_t offset_h = data_offset_ptr[data_offset_h_ptr];
+    const scalar_t offset_w = data_offset_ptr[data_offset_w_ptr];
+    const scalar_t mask = data_mask_ptr[data_mask_hw_ptr];
+    const scalar_t cur_inv_h_data = h_in + i * dilation_h + offset_h;
+    const scalar_t cur_inv_w_data = w_in + j * dilation_w + offset_w;
+
+    const scalar_t cur_top_grad = data_col[index] * mask;
+    const int cur_h = (int)cur_inv_h_data;
+    const int cur_w = (int)cur_inv_w_data;
+    for (int dy = -2; dy <= 2; dy++)
+    {
+      for (int dx = -2; dx <= 2; dx++)
+      {
+        if (cur_h + dy >= 0 && cur_h + dy < height &&
+            cur_w + dx >= 0 && cur_w + dx < width &&
+            abs(cur_inv_h_data - (cur_h + dy)) < 1 &&
+            abs(cur_inv_w_data - (cur_w + dx)) < 1)
+        {
+          int cur_bottom_grad_pos = ((b * channels + c) * height + cur_h + dy) * width + cur_w + dx;
+          scalar_t weight = dmcn_get_gradient_weight(cur_inv_h_data, cur_inv_w_data, cur_h + dy, cur_w + dx, height, width);
+          atomicAdd(grad_im + cur_bottom_grad_pos, weight * cur_top_grad);
+        }
+      }
+    }
+  }
+}
+
+template <typename scalar_t>
+__global__ void modulated_deformable_col2im_coord_gpu_kernel(const int n,
+                                                             const scalar_t *data_col, const scalar_t *data_im,
+                                                             const scalar_t *data_offset, const scalar_t *data_mask,
+                                                             const int channels, const int height, const int width,
+                                                             const int kernel_h, const int kernel_w,
+                                                             const int pad_h, const int pad_w,
+                                                             const int stride_h, const int stride_w,
+                                                             const int dilation_h, const int dilation_w,
+                                                             const int channel_per_deformable_group,
+                                                             const int batch_size, const int offset_channels, const int deformable_group,
+                                                             const int height_col, const int width_col,
+                                                             scalar_t *grad_offset, scalar_t *grad_mask)
+{
+  CUDA_KERNEL_LOOP(index, n)
+  {
+    scalar_t val = 0, mval = 0;
+    int w = index % width_col;
+    int h = (index / width_col) % height_col;
+    int c = (index / width_col / height_col) % offset_channels;
+    int b = (index / width_col / height_col) / offset_channels;
+    // compute the start and end of the output
+
+    const int deformable_group_index = c / (2 * kernel_h * kernel_w);
+    const int col_step = kernel_h * kernel_w;
+    int cnt = 0;
+    const scalar_t *data_col_ptr = data_col + deformable_group_index * channel_per_deformable_group * batch_size * width_col * height_col;
+    const scalar_t *data_im_ptr = data_im + (b * deformable_group + deformable_group_index) * channel_per_deformable_group / kernel_h / kernel_w * height * width;
+    const scalar_t *data_offset_ptr = data_offset + (b * deformable_group + deformable_group_index) * 2 * kernel_h * kernel_w * height_col * width_col;
+    const scalar_t *data_mask_ptr = data_mask + (b * deformable_group + deformable_group_index) * kernel_h * kernel_w * height_col * width_col;
+
+    const int offset_c = c - deformable_group_index * 2 * kernel_h * kernel_w;
+
+    for (int col_c = (offset_c / 2); col_c < channel_per_deformable_group; col_c += col_step)
+    {
+      const int col_pos = (((col_c * batch_size + b) * height_col) + h) * width_col + w;
+      const int bp_dir = offset_c % 2;
+
+      int j = (col_pos / width_col / height_col / batch_size) % kernel_w;
+      int i = (col_pos / width_col / height_col / batch_size / kernel_w) % kernel_h;
+      int w_out = col_pos % width_col;
+      int h_out = (col_pos / width_col) % height_col;
+      int w_in = w_out * stride_w - pad_w;
+      int h_in = h_out * stride_h - pad_h;
+      const int data_offset_h_ptr = (((2 * (i * kernel_w + j)) * height_col + h_out) * width_col + w_out);
+      const int data_offset_w_ptr = (((2 * (i * kernel_w + j) + 1) * height_col + h_out) * width_col + w_out);
+      const int data_mask_hw_ptr = (((i * kernel_w + j) * height_col + h_out) * width_col + w_out);
+      const scalar_t offset_h = data_offset_ptr[data_offset_h_ptr];
+      const scalar_t offset_w = data_offset_ptr[data_offset_w_ptr];
+      const scalar_t mask = data_mask_ptr[data_mask_hw_ptr];
+      scalar_t inv_h = h_in + i * dilation_h + offset_h;
+      scalar_t inv_w = w_in + j * dilation_w + offset_w;
+      if (inv_h <= -1 || inv_w <= -1 || inv_h >= height || inv_w >= width)
+      {
+        inv_h = inv_w = -2;
+      }
+      else
+      {
+        mval += data_col_ptr[col_pos] * dmcn_im2col_bilinear(data_im_ptr + cnt * height * width, width, height, width, inv_h, inv_w);
+      }
+      const scalar_t weight = dmcn_get_coordinate_weight(
+          inv_h, inv_w,
+          height, width, data_im_ptr + cnt * height * width, width, bp_dir);
+      val += weight * data_col_ptr[col_pos] * mask;
+      cnt += 1;
+    }
+    // KERNEL_ASSIGN(grad_offset[index], offset_req, val);
+    grad_offset[index] = val;
+    if (offset_c % 2 == 0)
+      // KERNEL_ASSIGN(grad_mask[(((b * deformable_group + deformable_group_index) * kernel_h * kernel_w + offset_c / 2) * height_col + h) * width_col + w], mask_req, mval);
+      grad_mask[(((b * deformable_group + deformable_group_index) * kernel_h * kernel_w + offset_c / 2) * height_col + h) * width_col + w] = mval;
+  }
+}
+
+void modulated_deformable_im2col_cuda(
+    const at::Tensor data_im, const at::Tensor data_offset, const at::Tensor data_mask,
+    const int batch_size, const int channels, const int height_im, const int width_im,
+    const int height_col, const int width_col, const int kernel_h, const int kenerl_w,
+    const int pad_h, const int pad_w, const int stride_h, const int stride_w,
+    const int dilation_h, const int dilation_w,
+    const int deformable_group, at::Tensor data_col)
+{
+  // num_axes should be smaller than block size
+  const int channel_per_deformable_group = channels / deformable_group;
+  const int num_kernels = channels * batch_size * height_col * width_col;
+
+  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
+      data_im.scalar_type(), "modulated_deformable_im2col_gpu", ([&] {
+        const scalar_t *data_im_ = data_im.data_ptr<scalar_t>();
+        const scalar_t *data_offset_ = data_offset.data_ptr<scalar_t>();
+        const scalar_t *data_mask_ = data_mask.data_ptr<scalar_t>();
+        scalar_t *data_col_ = data_col.data_ptr<scalar_t>();
+
+        modulated_deformable_im2col_gpu_kernel<<<GET_BLOCKS(num_kernels), CUDA_NUM_THREADS, 0, at::cuda::getCurrentCUDAStream()>>>(
+            num_kernels, data_im_, data_offset_, data_mask_, height_im, width_im, kernel_h, kenerl_w,
+            pad_h, pad_w, stride_h, stride_w, dilation_h, dilation_w, channel_per_deformable_group,
+            batch_size, channels, deformable_group, height_col, width_col, data_col_);
+      }));
+
+  cudaError_t err = cudaGetLastError();
+  if (err != cudaSuccess)
+  {
+    printf("error in modulated_deformable_im2col_cuda: %s\n", cudaGetErrorString(err));
+  }
+}
+
+void modulated_deformable_col2im_cuda(
+    const at::Tensor data_col, const at::Tensor data_offset, const at::Tensor data_mask,
+    const int batch_size, const int channels, const int height_im, const int width_im,
+    const int height_col, const int width_col, const int kernel_h, const int kernel_w,
+    const int pad_h, const int pad_w, const int stride_h, const int stride_w,
+    const int dilation_h, const int dilation_w,
+    const int deformable_group, at::Tensor grad_im)
+{
+
+  const int channel_per_deformable_group = channels / deformable_group;
+  const int num_kernels = channels * kernel_h * kernel_w * batch_size * height_col * width_col;
+
+  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
+      data_col.scalar_type(), "modulated_deformable_col2im_gpu", ([&] {
+        const scalar_t *data_col_ = data_col.data_ptr<scalar_t>();
+        const scalar_t *data_offset_ = data_offset.data_ptr<scalar_t>();
+        const scalar_t *data_mask_ = data_mask.data_ptr<scalar_t>();
+        scalar_t *grad_im_ = grad_im.data_ptr<scalar_t>();
+
+        modulated_deformable_col2im_gpu_kernel<<<GET_BLOCKS(num_kernels), CUDA_NUM_THREADS, 0, at::cuda::getCurrentCUDAStream()>>>(
+            num_kernels, data_col_, data_offset_, data_mask_, channels, height_im, width_im,
+            kernel_h, kernel_w, pad_h, pad_w, stride_h, stride_w,
+            dilation_h, dilation_w, channel_per_deformable_group,
+            batch_size, deformable_group, height_col, width_col, grad_im_);
+      }));
+
+  cudaError_t err = cudaGetLastError();
+  if (err != cudaSuccess)
+  {
+    printf("error in modulated_deformable_col2im_cuda: %s\n", cudaGetErrorString(err));
+  }
+}
+
+void modulated_deformable_col2im_coord_cuda(
+    const at::Tensor data_col, const at::Tensor data_im, const at::Tensor data_offset, const at::Tensor data_mask,
+    const int batch_size, const int channels, const int height_im, const int width_im,
+    const int height_col, const int width_col, const int kernel_h, const int kernel_w,
+    const int pad_h, const int pad_w, const int stride_h, const int stride_w,
+    const int dilation_h, const int dilation_w,
+    const int deformable_group,
+    at::Tensor grad_offset, at::Tensor grad_mask)
+{
+  const int num_kernels = batch_size * height_col * width_col * 2 * kernel_h * kernel_w * deformable_group;
+  const int channel_per_deformable_group = channels * kernel_h * kernel_w / deformable_group;
+
+  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
+      data_col.scalar_type(), "modulated_deformable_col2im_coord_gpu", ([&] {
+        const scalar_t *data_col_ = data_col.data_ptr<scalar_t>();
+        const scalar_t *data_im_ = data_im.data_ptr<scalar_t>();
+        const scalar_t *data_offset_ = data_offset.data_ptr<scalar_t>();
+        const scalar_t *data_mask_ = data_mask.data_ptr<scalar_t>();
+        scalar_t *grad_offset_ = grad_offset.data_ptr<scalar_t>();
+        scalar_t *grad_mask_ = grad_mask.data_ptr<scalar_t>();
+
+        modulated_deformable_col2im_coord_gpu_kernel<<<GET_BLOCKS(num_kernels), CUDA_NUM_THREADS, 0, at::cuda::getCurrentCUDAStream()>>>(
+            num_kernels, data_col_, data_im_, data_offset_, data_mask_, channels, height_im, width_im,
+            kernel_h, kernel_w, pad_h, pad_w, stride_h, stride_w,
+            dilation_h, dilation_w, channel_per_deformable_group,
+            batch_size, 2 * kernel_h * kernel_w * deformable_group, deformable_group, height_col, width_col,
+            grad_offset_, grad_mask_);
+      }));
+  cudaError_t err = cudaGetLastError();
+  if (err != cudaSuccess)
+  {
+    printf("error in modulated_deformable_col2im_coord_cuda: %s\n", cudaGetErrorString(err));
+  }
+}

basicsr/ops/dcn/src/deform_conv_ext.cpp

+// modify from
+// https://github.com/chengdazhi/Deformable-Convolution-V2-PyTorch/blob/mmdetection/mmdet/ops/dcn/src/deform_conv_cuda.c
+
+#include <torch/extension.h>
+#include <ATen/DeviceGuard.h>
+
+#include <cmath>
+#include <vector>
+
+#define WITH_CUDA  // always use cuda
+#ifdef WITH_CUDA
+int deform_conv_forward_cuda(at::Tensor input, at::Tensor weight,
+                             at::Tensor offset, at::Tensor output,
+                             at::Tensor columns, at::Tensor ones, int kW,
+                             int kH, int dW, int dH, int padW, int padH,
+                             int dilationW, int dilationH, int group,
+                             int deformable_group, int im2col_step);
+
+int deform_conv_backward_input_cuda(at::Tensor input, at::Tensor offset,
+                                    at::Tensor gradOutput, at::Tensor gradInput,
+                                    at::Tensor gradOffset, at::Tensor weight,
+                                    at::Tensor columns, int kW, int kH, int dW,
+                                    int dH, int padW, int padH, int dilationW,
+                                    int dilationH, int group,
+                                    int deformable_group, int im2col_step);
+
+int deform_conv_backward_parameters_cuda(
+    at::Tensor input, at::Tensor offset, at::Tensor gradOutput,
+    at::Tensor gradWeight,  // at::Tensor gradBias,
+    at::Tensor columns, at::Tensor ones, int kW, int kH, int dW, int dH,
+    int padW, int padH, int dilationW, int dilationH, int group,
+    int deformable_group, float scale, int im2col_step);
+
+void modulated_deform_conv_cuda_forward(
+    at::Tensor input, at::Tensor weight, at::Tensor bias, at::Tensor ones,
+    at::Tensor offset, at::Tensor mask, at::Tensor output, at::Tensor columns,
+    int kernel_h, int kernel_w, const int stride_h, const int stride_w,
+    const int pad_h, const int pad_w, const int dilation_h,
+    const int dilation_w, const int group, const int deformable_group,
+    const bool with_bias);
+
+void modulated_deform_conv_cuda_backward(
+    at::Tensor input, at::Tensor weight, at::Tensor bias, at::Tensor ones,
+    at::Tensor offset, at::Tensor mask, at::Tensor columns,
+    at::Tensor grad_input, at::Tensor grad_weight, at::Tensor grad_bias,
+    at::Tensor grad_offset, at::Tensor grad_mask, at::Tensor grad_output,
+    int kernel_h, int kernel_w, int stride_h, int stride_w, int pad_h,
+    int pad_w, int dilation_h, int dilation_w, int group, int deformable_group,
+    const bool with_bias);
+#endif
+
+int deform_conv_forward(at::Tensor input, at::Tensor weight,
+                             at::Tensor offset, at::Tensor output,
+                             at::Tensor columns, at::Tensor ones, int kW,
+                             int kH, int dW, int dH, int padW, int padH,
+                             int dilationW, int dilationH, int group,
+                             int deformable_group, int im2col_step) {
+  if (input.device().is_cuda()) {
+#ifdef WITH_CUDA
+    return deform_conv_forward_cuda(input, weight, offset, output, columns,
+        ones, kW, kH, dW, dH, padW, padH, dilationW, dilationH, group,
+        deformable_group, im2col_step);
+#else
+    AT_ERROR("deform conv is not compiled with GPU support");
+#endif
+  }
+  AT_ERROR("deform conv is not implemented on CPU");
+}
+
+int deform_conv_backward_input(at::Tensor input, at::Tensor offset,
+                                    at::Tensor gradOutput, at::Tensor gradInput,
+                                    at::Tensor gradOffset, at::Tensor weight,
+                                    at::Tensor columns, int kW, int kH, int dW,
+                                    int dH, int padW, int padH, int dilationW,
+                                    int dilationH, int group,
+                                    int deformable_group, int im2col_step) {
+  if (input.device().is_cuda()) {
+#ifdef WITH_CUDA
+    return deform_conv_backward_input_cuda(input, offset, gradOutput,
+        gradInput, gradOffset, weight, columns, kW, kH, dW, dH, padW, padH,
+        dilationW, dilationH, group, deformable_group, im2col_step);
+#else
+    AT_ERROR("deform conv is not compiled with GPU support");
+#endif
+  }
+  AT_ERROR("deform conv is not implemented on CPU");
+}
+
+int deform_conv_backward_parameters(
+    at::Tensor input, at::Tensor offset, at::Tensor gradOutput,
+    at::Tensor gradWeight,  // at::Tensor gradBias,
+    at::Tensor columns, at::Tensor ones, int kW, int kH, int dW, int dH,
+    int padW, int padH, int dilationW, int dilationH, int group,
+    int deformable_group, float scale, int im2col_step) {
+  if (input.device().is_cuda()) {
+#ifdef WITH_CUDA
+    return deform_conv_backward_parameters_cuda(input, offset, gradOutput,
+        gradWeight, columns, ones, kW, kH, dW, dH, padW, padH, dilationW,
+        dilationH, group, deformable_group, scale, im2col_step);
+#else
+    AT_ERROR("deform conv is not compiled with GPU support");
+#endif
+  }
+  AT_ERROR("deform conv is not implemented on CPU");
+}
+
+void modulated_deform_conv_forward(
+    at::Tensor input, at::Tensor weight, at::Tensor bias, at::Tensor ones,
+    at::Tensor offset, at::Tensor mask, at::Tensor output, at::Tensor columns,
+    int kernel_h, int kernel_w, const int stride_h, const int stride_w,
+    const int pad_h, const int pad_w, const int dilation_h,
+    const int dilation_w, const int group, const int deformable_group,
+    const bool with_bias) {
+  if (input.device().is_cuda()) {
+#ifdef WITH_CUDA
+    return modulated_deform_conv_cuda_forward(input, weight, bias, ones,
+        offset, mask, output, columns, kernel_h, kernel_w, stride_h,
+        stride_w, pad_h, pad_w, dilation_h, dilation_w, group,
+        deformable_group, with_bias);
+#else
+    AT_ERROR("modulated deform conv is not compiled with GPU support");
+#endif
+  }
+  AT_ERROR("modulated deform conv is not implemented on CPU");
+}
+
+void modulated_deform_conv_backward(
+    at::Tensor input, at::Tensor weight, at::Tensor bias, at::Tensor ones,
+    at::Tensor offset, at::Tensor mask, at::Tensor columns,
+    at::Tensor grad_input, at::Tensor grad_weight, at::Tensor grad_bias,
+    at::Tensor grad_offset, at::Tensor grad_mask, at::Tensor grad_output,
+    int kernel_h, int kernel_w, int stride_h, int stride_w, int pad_h,
+    int pad_w, int dilation_h, int dilation_w, int group, int deformable_group,
+    const bool with_bias) {
+  if (input.device().is_cuda()) {
+#ifdef WITH_CUDA
+    return modulated_deform_conv_cuda_backward(input, weight, bias, ones,
+        offset, mask, columns, grad_input, grad_weight, grad_bias, grad_offset,
+        grad_mask, grad_output, kernel_h, kernel_w, stride_h, stride_w,
+        pad_h, pad_w, dilation_h, dilation_w, group, deformable_group,
+        with_bias);
+#else
+    AT_ERROR("modulated deform conv is not compiled with GPU support");
+#endif
+  }
+  AT_ERROR("modulated deform conv is not implemented on CPU");
+}
+
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  m.def("deform_conv_forward", &deform_conv_forward,
+        "deform forward");
+  m.def("deform_conv_backward_input", &deform_conv_backward_input,
+        "deform_conv_backward_input");
+  m.def("deform_conv_backward_parameters",
+        &deform_conv_backward_parameters,
+        "deform_conv_backward_parameters");
+  m.def("modulated_deform_conv_forward",
+        &modulated_deform_conv_forward,
+        "modulated deform conv forward");
+  m.def("modulated_deform_conv_backward",
+        &modulated_deform_conv_backward,
+        "modulated deform conv backward");
+}

basicsr/ops/dcn/src/deform_conv_hip_kernel.hip

+// !!! This is a file automatically generated by hipify!!!
+#include "hip/hip_runtime.h"
+/*!
+ ******************* BEGIN Caffe Copyright Notice and Disclaimer ****************
+ *
+ * COPYRIGHT
+ *
+ * All contributions by the University of California:
+ * Copyright (c) 2014-2017 The Regents of the University of California (Regents)
+ * All rights reserved.
+ *
+ * All other contributions:
+ * Copyright (c) 2014-2017, the respective contributors
+ * All rights reserved.
+ *
+ * Caffe uses a shared copyright model: each contributor holds copyright over
+ * their contributions to Caffe. The project versioning records all such
+ * contribution and copyright details. If a contributor wants to further mark
+ * their specific copyright on a particular contribution, they should indicate
+ * their copyright solely in the commit message of the change when it is
+ * committed.
+ *
+ * LICENSE
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are met:
+ *
+ * 1. Redistributions of source code must retain the above copyright notice, this
+ * list of conditions and the following disclaimer.
+ * 2. Redistributions in binary form must reproduce the above copyright notice,
+ * this list of conditions and the following disclaimer in the documentation
+ * and/or other materials provided with the distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
+ * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ * CONTRIBUTION AGREEMENT
+ *
+ * By contributing to the BVLC/caffe repository through pull-request, comment,
+ * or otherwise, the contributor releases their content to the
+ * license and copyright terms herein.
+ *
+ ***************** END Caffe Copyright Notice and Disclaimer ********************
+ *
+ * Copyright (c) 2018 Microsoft
+ * Licensed under The MIT License [see LICENSE for details]
+ * \file modulated_deformable_im2col.cuh
+ * \brief Function definitions of converting an image to
+ * column matrix based on kernel, padding, dilation, and offset.
+ * These functions are mainly used in deformable convolution operators.
+ * \ref: https://arxiv.org/abs/1703.06211
+ * \author Yuwen Xiong, Haozhi Qi, Jifeng Dai, Xizhou Zhu, Han Hu, Dazhi Cheng
+ */
+
+// modified from https://github.com/chengdazhi/Deformable-Convolution-V2-PyTorch/blob/mmdetection/mmdet/ops/dcn/src/deform_conv_cuda_kernel.cu
+
+#include <ATen/ATen.h>
+#include <ATen/hip/HIPContext.h>
+#include <THH/THHAtomics.cuh>
+#include <stdio.h>
+#include <math.h>
+#include <float.h>
+
+using namespace at;
+
+#define CUDA_KERNEL_LOOP(i, n)                                 \
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < (n); \
+       i += blockDim.x * gridDim.x)
+
+const int CUDA_NUM_THREADS = 1024;
+const int kMaxGridNum = 65535;
+
+inline int GET_BLOCKS(const int N)
+{
+  return ::min(kMaxGridNum, (N + CUDA_NUM_THREADS - 1) / CUDA_NUM_THREADS);
+}
+
+template <typename scalar_t>
+__device__ scalar_t deformable_im2col_bilinear(const scalar_t *bottom_data, const int data_width,
+                                               const int height, const int width, scalar_t h, scalar_t w)
+{
+
+  int h_low = floor(h);
+  int w_low = floor(w);
+  int h_high = h_low + 1;
+  int w_high = w_low + 1;
+
+  scalar_t lh = h - h_low;
+  scalar_t lw = w - w_low;
+  scalar_t hh = 1 - lh, hw = 1 - lw;
+
+  scalar_t v1 = 0;
+  if (h_low >= 0 && w_low >= 0)
+    v1 = bottom_data[h_low * data_width + w_low];
+  scalar_t v2 = 0;
+  if (h_low >= 0 && w_high <= width - 1)
+    v2 = bottom_data[h_low * data_width + w_high];
+  scalar_t v3 = 0;
+  if (h_high <= height - 1 && w_low >= 0)
+    v3 = bottom_data[h_high * data_width + w_low];
+  scalar_t v4 = 0;
+  if (h_high <= height - 1 && w_high <= width - 1)
+    v4 = bottom_data[h_high * data_width + w_high];
+
+  scalar_t w1 = hh * hw, w2 = hh * lw, w3 = lh * hw, w4 = lh * lw;
+
+  scalar_t val = (w1 * v1 + w2 * v2 + w3 * v3 + w4 * v4);
+  return val;
+}
+
+template <typename scalar_t>
+__device__ scalar_t get_gradient_weight(scalar_t argmax_h, scalar_t argmax_w,
+                                        const int h, const int w, const int height, const int width)
+{
+
+  if (argmax_h <= -1 || argmax_h >= height || argmax_w <= -1 || argmax_w >= width)
+  {
+    //empty
+    return 0;
+  }
+
+  int argmax_h_low = floor(argmax_h);
+  int argmax_w_low = floor(argmax_w);
+  int argmax_h_high = argmax_h_low + 1;
+  int argmax_w_high = argmax_w_low + 1;
+
+  scalar_t weight = 0;
+  if (h == argmax_h_low && w == argmax_w_low)
+    weight = (h + 1 - argmax_h) * (w + 1 - argmax_w);
+  if (h == argmax_h_low && w == argmax_w_high)
+    weight = (h + 1 - argmax_h) * (argmax_w + 1 - w);
+  if (h == argmax_h_high && w == argmax_w_low)
+    weight = (argmax_h + 1 - h) * (w + 1 - argmax_w);
+  if (h == argmax_h_high && w == argmax_w_high)
+    weight = (argmax_h + 1 - h) * (argmax_w + 1 - w);
+  return weight;
+}
+
+template <typename scalar_t>
+__device__ scalar_t get_coordinate_weight(scalar_t argmax_h, scalar_t argmax_w,
+                                          const int height, const int width, const scalar_t *im_data,
+                                          const int data_width, const int bp_dir)
+{
+
+  if (argmax_h <= -1 || argmax_h >= height || argmax_w <= -1 || argmax_w >= width)
+  {
+    //empty
+    return 0;
+  }
+
+  int argmax_h_low = floor(argmax_h);
+  int argmax_w_low = floor(argmax_w);
+  int argmax_h_high = argmax_h_low + 1;
+  int argmax_w_high = argmax_w_low + 1;
+
+  scalar_t weight = 0;
+
+  if (bp_dir == 0)
+  {
+    if (argmax_h_low >= 0 && argmax_w_low >= 0)
+      weight += -1 * (argmax_w_low + 1 - argmax_w) * im_data[argmax_h_low * data_width + argmax_w_low];
+    if (argmax_h_low >= 0 && argmax_w_high <= width - 1)
+      weight += -1 * (argmax_w - argmax_w_low) * im_data[argmax_h_low * data_width + argmax_w_high];
+    if (argmax_h_high <= height - 1 && argmax_w_low >= 0)
+      weight += (argmax_w_low + 1 - argmax_w) * im_data[argmax_h_high * data_width + argmax_w_low];
+    if (argmax_h_high <= height - 1 && argmax_w_high <= width - 1)
+      weight += (argmax_w - argmax_w_low) * im_data[argmax_h_high * data_width + argmax_w_high];
+  }
+  else if (bp_dir == 1)
+  {
+    if (argmax_h_low >= 0 && argmax_w_low >= 0)
+      weight += -1 * (argmax_h_low + 1 - argmax_h) * im_data[argmax_h_low * data_width + argmax_w_low];
+    if (argmax_h_low >= 0 && argmax_w_high <= width - 1)
+      weight += (argmax_h_low + 1 - argmax_h) * im_data[argmax_h_low * data_width + argmax_w_high];
+    if (argmax_h_high <= height - 1 && argmax_w_low >= 0)
+      weight += -1 * (argmax_h - argmax_h_low) * im_data[argmax_h_high * data_width + argmax_w_low];
+    if (argmax_h_high <= height - 1 && argmax_w_high <= width - 1)
+      weight += (argmax_h - argmax_h_low) * im_data[argmax_h_high * data_width + argmax_w_high];
+  }
+
+  return weight;
+}
+
+template <typename scalar_t>
+__global__ void deformable_im2col_gpu_kernel(const int n, const scalar_t *data_im, const scalar_t *data_offset,
+                                             const int height, const int width, const int kernel_h, const int kernel_w,
+                                             const int pad_h, const int pad_w, const int stride_h, const int stride_w,
+                                             const int dilation_h, const int dilation_w, const int channel_per_deformable_group,
+                                             const int batch_size, const int num_channels, const int deformable_group,
+                                             const int height_col, const int width_col,
+                                             scalar_t *data_col)
+{
+  CUDA_KERNEL_LOOP(index, n)
+  {
+    // index index of output matrix
+    const int w_col = index % width_col;
+    const int h_col = (index / width_col) % height_col;
+    const int b_col = (index / width_col / height_col) % batch_size;
+    const int c_im = (index / width_col / height_col) / batch_size;
+    const int c_col = c_im * kernel_h * kernel_w;
+
+    // compute deformable group index
+    const int deformable_group_index = c_im / channel_per_deformable_group;
+
+    const int h_in = h_col * stride_h - pad_h;
+    const int w_in = w_col * stride_w - pad_w;
+    scalar_t *data_col_ptr = data_col + ((c_col * batch_size + b_col) * height_col + h_col) * width_col + w_col;
+    //const scalar_t* data_im_ptr = data_im + ((b_col * num_channels + c_im) * height + h_in) * width + w_in;
+    const scalar_t *data_im_ptr = data_im + (b_col * num_channels + c_im) * height * width;
+    const scalar_t *data_offset_ptr = data_offset + (b_col * deformable_group + deformable_group_index) * 2 * kernel_h * kernel_w * height_col * width_col;
+
+    for (int i = 0; i < kernel_h; ++i)
+    {
+      for (int j = 0; j < kernel_w; ++j)
+      {
+        const int data_offset_h_ptr = ((2 * (i * kernel_w + j)) * height_col + h_col) * width_col + w_col;
+        const int data_offset_w_ptr = ((2 * (i * kernel_w + j) + 1) * height_col + h_col) * width_col + w_col;
+        const scalar_t offset_h = data_offset_ptr[data_offset_h_ptr];
+        const scalar_t offset_w = data_offset_ptr[data_offset_w_ptr];
+        scalar_t val = static_cast<scalar_t>(0);
+        const scalar_t h_im = h_in + i * dilation_h + offset_h;
+        const scalar_t w_im = w_in + j * dilation_w + offset_w;
+        if (h_im > -1 && w_im > -1 && h_im < height && w_im < width)
+        {
+          //const scalar_t map_h = i * dilation_h + offset_h;
+          //const scalar_t map_w = j * dilation_w + offset_w;
+          //const int cur_height = height - h_in;
+          //const int cur_width = width - w_in;
+          //val = deformable_im2col_bilinear(data_im_ptr, width, cur_height, cur_width, map_h, map_w);
+          val = deformable_im2col_bilinear(data_im_ptr, width, height, width, h_im, w_im);
+        }
+        *data_col_ptr = val;
+        data_col_ptr += batch_size * height_col * width_col;
+      }
+    }
+  }
+}
+
+void deformable_im2col(
+    const at::Tensor data_im, const at::Tensor data_offset, const int channels,
+    const int height, const int width, const int ksize_h, const int ksize_w,
+    const int pad_h, const int pad_w, const int stride_h, const int stride_w,
+    const int dilation_h, const int dilation_w, const int parallel_imgs,
+    const int deformable_group, at::Tensor data_col)
+{
+  // num_axes should be smaller than block size
+  // todo: check parallel_imgs is correctly passed in
+  int height_col = (height + 2 * pad_h - (dilation_h * (ksize_h - 1) + 1)) / stride_h + 1;
+  int width_col = (width + 2 * pad_w - (dilation_w * (ksize_w - 1) + 1)) / stride_w + 1;
+  int num_kernels = channels * height_col * width_col * parallel_imgs;
+  int channel_per_deformable_group = channels / deformable_group;
+
+  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
+      data_im.scalar_type(), "deformable_im2col_gpu", ([&] {
+        const scalar_t *data_im_ = data_im.data_ptr<scalar_t>();
+        const scalar_t *data_offset_ = data_offset.data_ptr<scalar_t>();
+        scalar_t *data_col_ = data_col.data_ptr<scalar_t>();
+
+       hipLaunchKernelGGL(( deformable_im2col_gpu_kernel), dim3(GET_BLOCKS(num_kernels)), dim3(CUDA_NUM_THREADS), 0, at::hip::getCurrentHIPStreamMasqueradingAsCUDA(), 
+            num_kernels, data_im_, data_offset_, height, width, ksize_h, ksize_w,
+            pad_h, pad_w, stride_h, stride_w, dilation_h, dilation_w,
+            channel_per_deformable_group, parallel_imgs, channels, deformable_group,
+            height_col, width_col, data_col_);
+      }));
+
+  hipError_t err = hipGetLastError();
+  if (err != hipSuccess)
+  {
+    printf("error in deformable_im2col: %s\n", hipGetErrorString(err));
+  }
+}
+
+template <typename scalar_t>
+__global__ void deformable_col2im_gpu_kernel(
+    const int n, const scalar_t *data_col, const scalar_t *data_offset,
+    const int channels, const int height, const int width,
+    const int kernel_h, const int kernel_w,
+    const int pad_h, const int pad_w,
+    const int stride_h, const int stride_w,
+    const int dilation_h, const int dilation_w,
+    const int channel_per_deformable_group,
+    const int batch_size, const int deformable_group,
+    const int height_col, const int width_col,
+    scalar_t *grad_im)
+{
+  CUDA_KERNEL_LOOP(index, n)
+  {
+    const int j = (index / width_col / height_col / batch_size) % kernel_w;
+    const int i = (index / width_col / height_col / batch_size / kernel_w) % kernel_h;
+    const int c = index / width_col / height_col / batch_size / kernel_w / kernel_h;
+    // compute the start and end of the output
+
+    const int deformable_group_index = c / channel_per_deformable_group;
+
+    int w_out = index % width_col;
+    int h_out = (index / width_col) % height_col;
+    int b = (index / width_col / height_col) % batch_size;
+    int w_in = w_out * stride_w - pad_w;
+    int h_in = h_out * stride_h - pad_h;
+
+    const scalar_t *data_offset_ptr = data_offset + (b * deformable_group + deformable_group_index) *
+                                                        2 * kernel_h * kernel_w * height_col * width_col;
+    const int data_offset_h_ptr = ((2 * (i * kernel_w + j)) * height_col + h_out) * width_col + w_out;
+    const int data_offset_w_ptr = ((2 * (i * kernel_w + j) + 1) * height_col + h_out) * width_col + w_out;
+    const scalar_t offset_h = data_offset_ptr[data_offset_h_ptr];
+    const scalar_t offset_w = data_offset_ptr[data_offset_w_ptr];
+    const scalar_t cur_inv_h_data = h_in + i * dilation_h + offset_h;
+    const scalar_t cur_inv_w_data = w_in + j * dilation_w + offset_w;
+
+    const scalar_t cur_top_grad = data_col[index];
+    const int cur_h = (int)cur_inv_h_data;
+    const int cur_w = (int)cur_inv_w_data;
+    for (int dy = -2; dy <= 2; dy++)
+    {
+      for (int dx = -2; dx <= 2; dx++)
+      {
+        if (cur_h + dy >= 0 && cur_h + dy < height &&
+            cur_w + dx >= 0 && cur_w + dx < width &&
+            abs(cur_inv_h_data - (cur_h + dy)) < 1 &&
+            abs(cur_inv_w_data - (cur_w + dx)) < 1)
+        {
+          int cur_bottom_grad_pos = ((b * channels + c) * height + cur_h + dy) * width + cur_w + dx;
+          scalar_t weight = get_gradient_weight(cur_inv_h_data, cur_inv_w_data, cur_h + dy, cur_w + dx, height, width);
+          atomicAdd(grad_im + cur_bottom_grad_pos, weight * cur_top_grad);
+        }
+      }
+    }
+  }
+}
+
+void deformable_col2im(
+    const at::Tensor data_col, const at::Tensor data_offset, const int channels,
+    const int height, const int width, const int ksize_h,
+    const int ksize_w, const int pad_h, const int pad_w,
+    const int stride_h, const int stride_w,
+    const int dilation_h, const int dilation_w,
+    const int parallel_imgs, const int deformable_group,
+    at::Tensor grad_im)
+{
+
+  // todo: make sure parallel_imgs is passed in correctly
+  int height_col = (height + 2 * pad_h - (dilation_h * (ksize_h - 1) + 1)) / stride_h + 1;
+  int width_col = (width + 2 * pad_w - (dilation_w * (ksize_w - 1) + 1)) / stride_w + 1;
+  int num_kernels = channels * ksize_h * ksize_w * height_col * width_col * parallel_imgs;
+  int channel_per_deformable_group = channels / deformable_group;
+
+  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
+      data_col.scalar_type(), "deformable_col2im_gpu", ([&] {
+        const scalar_t *data_col_ = data_col.data_ptr<scalar_t>();
+        const scalar_t *data_offset_ = data_offset.data_ptr<scalar_t>();
+        scalar_t *grad_im_ = grad_im.data_ptr<scalar_t>();
+
+       hipLaunchKernelGGL(( deformable_col2im_gpu_kernel), dim3(GET_BLOCKS(num_kernels)), dim3(CUDA_NUM_THREADS), 0, at::hip::getCurrentHIPStreamMasqueradingAsCUDA(), 
+            num_kernels, data_col_, data_offset_, channels, height, width, ksize_h,
+            ksize_w, pad_h, pad_w, stride_h, stride_w,
+            dilation_h, dilation_w, channel_per_deformable_group,
+            parallel_imgs, deformable_group, height_col, width_col, grad_im_);
+      }));
+
+  hipError_t err = hipGetLastError();
+  if (err != hipSuccess)
+  {
+    printf("error in deformable_col2im: %s\n", hipGetErrorString(err));
+  }
+}
+
+template <typename scalar_t>
+__global__ void deformable_col2im_coord_gpu_kernel(const int n, const scalar_t *data_col,
+                                                   const scalar_t *data_im, const scalar_t *data_offset,
+                                                   const int channels, const int height, const int width,
+                                                   const int kernel_h, const int kernel_w,
+                                                   const int pad_h, const int pad_w,
+                                                   const int stride_h, const int stride_w,
+                                                   const int dilation_h, const int dilation_w,
+                                                   const int channel_per_deformable_group,
+                                                   const int batch_size, const int offset_channels, const int deformable_group,
+                                                   const int height_col, const int width_col, scalar_t *grad_offset)
+{
+  CUDA_KERNEL_LOOP(index, n)
+  {
+    scalar_t val = 0;
+    int w = index % width_col;
+    int h = (index / width_col) % height_col;
+    int c = (index / width_col / height_col) % offset_channels;
+    int b = (index / width_col / height_col) / offset_channels;
+    // compute the start and end of the output
+
+    const int deformable_group_index = c / (2 * kernel_h * kernel_w);
+    const int col_step = kernel_h * kernel_w;
+    int cnt = 0;
+    const scalar_t *data_col_ptr = data_col + deformable_group_index * channel_per_deformable_group *
+                                                  batch_size * width_col * height_col;
+    const scalar_t *data_im_ptr = data_im + (b * deformable_group + deformable_group_index) *
+                                                channel_per_deformable_group / kernel_h / kernel_w * height * width;
+    const scalar_t *data_offset_ptr = data_offset + (b * deformable_group + deformable_group_index) * 2 *
+                                                        kernel_h * kernel_w * height_col * width_col;
+
+    const int offset_c = c - deformable_group_index * 2 * kernel_h * kernel_w;
+
+    for (int col_c = (offset_c / 2); col_c < channel_per_deformable_group; col_c += col_step)
+    {
+      const int col_pos = (((col_c * batch_size + b) * height_col) + h) * width_col + w;
+      const int bp_dir = offset_c % 2;
+
+      int j = (col_pos / width_col / height_col / batch_size) % kernel_w;
+      int i = (col_pos / width_col / height_col / batch_size / kernel_w) % kernel_h;
+      int w_out = col_pos % width_col;
+      int h_out = (col_pos / width_col) % height_col;
+      int w_in = w_out * stride_w - pad_w;
+      int h_in = h_out * stride_h - pad_h;
+      const int data_offset_h_ptr = (((2 * (i * kernel_w + j)) * height_col + h_out) * width_col + w_out);
+      const int data_offset_w_ptr = (((2 * (i * kernel_w + j) + 1) * height_col + h_out) * width_col + w_out);
+      const scalar_t offset_h = data_offset_ptr[data_offset_h_ptr];
+      const scalar_t offset_w = data_offset_ptr[data_offset_w_ptr];
+      scalar_t inv_h = h_in + i * dilation_h + offset_h;
+      scalar_t inv_w = w_in + j * dilation_w + offset_w;
+      if (inv_h <= -1 || inv_w <= -1 || inv_h >= height || inv_w >= width)
+      {
+        inv_h = inv_w = -2;
+      }
+      const scalar_t weight = get_coordinate_weight(
+          inv_h, inv_w,
+          height, width, data_im_ptr + cnt * height * width, width, bp_dir);
+      val += weight * data_col_ptr[col_pos];
+      cnt += 1;
+    }
+
+    grad_offset[index] = val;
+  }
+}
+
+void deformable_col2im_coord(
+    const at::Tensor data_col, const at::Tensor data_im, const at::Tensor data_offset,
+    const int channels, const int height, const int width, const int ksize_h,
+    const int ksize_w, const int pad_h, const int pad_w, const int stride_h,
+    const int stride_w, const int dilation_h, const int dilation_w,
+    const int parallel_imgs, const int deformable_group, at::Tensor grad_offset)
+{
+
+  int height_col = (height + 2 * pad_h - (dilation_h * (ksize_h - 1) + 1)) / stride_h + 1;
+  int width_col = (width + 2 * pad_w - (dilation_w * (ksize_w - 1) + 1)) / stride_w + 1;
+  int num_kernels = height_col * width_col * 2 * ksize_h * ksize_w * deformable_group * parallel_imgs;
+  int channel_per_deformable_group = channels * ksize_h * ksize_w / deformable_group;
+
+  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
+      data_col.scalar_type(), "deformable_col2im_coord_gpu", ([&] {
+        const scalar_t *data_col_ = data_col.data_ptr<scalar_t>();
+        const scalar_t *data_im_ = data_im.data_ptr<scalar_t>();
+        const scalar_t *data_offset_ = data_offset.data_ptr<scalar_t>();
+        scalar_t *grad_offset_ = grad_offset.data_ptr<scalar_t>();
+
+       hipLaunchKernelGGL(( deformable_col2im_coord_gpu_kernel), dim3(GET_BLOCKS(num_kernels)), dim3(CUDA_NUM_THREADS), 0, at::hip::getCurrentHIPStreamMasqueradingAsCUDA(), 
+            num_kernels, data_col_, data_im_, data_offset_, channels, height, width,
+            ksize_h, ksize_w, pad_h, pad_w, stride_h, stride_w,
+            dilation_h, dilation_w, channel_per_deformable_group,
+            parallel_imgs, 2 * ksize_h * ksize_w * deformable_group, deformable_group,
+            height_col, width_col, grad_offset_);
+      }));
+}
+
+template <typename scalar_t>
+__device__ scalar_t dmcn_im2col_bilinear(const scalar_t *bottom_data, const int data_width,
+                                         const int height, const int width, scalar_t h, scalar_t w)
+{
+  int h_low = floor(h);
+  int w_low = floor(w);
+  int h_high = h_low + 1;
+  int w_high = w_low + 1;
+
+  scalar_t lh = h - h_low;
+  scalar_t lw = w - w_low;
+  scalar_t hh = 1 - lh, hw = 1 - lw;
+
+  scalar_t v1 = 0;
+  if (h_low >= 0 && w_low >= 0)
+    v1 = bottom_data[h_low * data_width + w_low];
+  scalar_t v2 = 0;
+  if (h_low >= 0 && w_high <= width - 1)
+    v2 = bottom_data[h_low * data_width + w_high];
+  scalar_t v3 = 0;
+  if (h_high <= height - 1 && w_low >= 0)
+    v3 = bottom_data[h_high * data_width + w_low];
+  scalar_t v4 = 0;
+  if (h_high <= height - 1 && w_high <= width - 1)
+    v4 = bottom_data[h_high * data_width + w_high];
+
+  scalar_t w1 = hh * hw, w2 = hh * lw, w3 = lh * hw, w4 = lh * lw;
+
+  scalar_t val = (w1 * v1 + w2 * v2 + w3 * v3 + w4 * v4);
+  return val;
+}
+
+template <typename scalar_t>
+__device__ scalar_t dmcn_get_gradient_weight(scalar_t argmax_h, scalar_t argmax_w,
+                                             const int h, const int w, const int height, const int width)
+{
+  if (argmax_h <= -1 || argmax_h >= height || argmax_w <= -1 || argmax_w >= width)
+  {
+    //empty
+    return 0;
+  }
+
+  int argmax_h_low = floor(argmax_h);
+  int argmax_w_low = floor(argmax_w);
+  int argmax_h_high = argmax_h_low + 1;
+  int argmax_w_high = argmax_w_low + 1;
+
+  scalar_t weight = 0;
+  if (h == argmax_h_low && w == argmax_w_low)
+    weight = (h + 1 - argmax_h) * (w + 1 - argmax_w);
+  if (h == argmax_h_low && w == argmax_w_high)
+    weight = (h + 1 - argmax_h) * (argmax_w + 1 - w);
+  if (h == argmax_h_high && w == argmax_w_low)
+    weight = (argmax_h + 1 - h) * (w + 1 - argmax_w);
+  if (h == argmax_h_high && w == argmax_w_high)
+    weight = (argmax_h + 1 - h) * (argmax_w + 1 - w);
+  return weight;
+}
+
+template <typename scalar_t>
+__device__ scalar_t dmcn_get_coordinate_weight(scalar_t argmax_h, scalar_t argmax_w,
+                                               const int height, const int width, const scalar_t *im_data,
+                                               const int data_width, const int bp_dir)
+{
+  if (argmax_h <= -1 || argmax_h >= height || argmax_w <= -1 || argmax_w >= width)
+  {
+    //empty
+    return 0;
+  }
+
+  int argmax_h_low = floor(argmax_h);
+  int argmax_w_low = floor(argmax_w);
+  int argmax_h_high = argmax_h_low + 1;
+  int argmax_w_high = argmax_w_low + 1;
+
+  scalar_t weight = 0;
+
+  if (bp_dir == 0)
+  {
+    if (argmax_h_low >= 0 && argmax_w_low >= 0)
+      weight += -1 * (argmax_w_low + 1 - argmax_w) * im_data[argmax_h_low * data_width + argmax_w_low];
+    if (argmax_h_low >= 0 && argmax_w_high <= width - 1)
+      weight += -1 * (argmax_w - argmax_w_low) * im_data[argmax_h_low * data_width + argmax_w_high];
+    if (argmax_h_high <= height - 1 && argmax_w_low >= 0)
+      weight += (argmax_w_low + 1 - argmax_w) * im_data[argmax_h_high * data_width + argmax_w_low];
+    if (argmax_h_high <= height - 1 && argmax_w_high <= width - 1)
+      weight += (argmax_w - argmax_w_low) * im_data[argmax_h_high * data_width + argmax_w_high];
+  }
+  else if (bp_dir == 1)
+  {
+    if (argmax_h_low >= 0 && argmax_w_low >= 0)
+      weight += -1 * (argmax_h_low + 1 - argmax_h) * im_data[argmax_h_low * data_width + argmax_w_low];
+    if (argmax_h_low >= 0 && argmax_w_high <= width - 1)
+      weight += (argmax_h_low + 1 - argmax_h) * im_data[argmax_h_low * data_width + argmax_w_high];
+    if (argmax_h_high <= height - 1 && argmax_w_low >= 0)
+      weight += -1 * (argmax_h - argmax_h_low) * im_data[argmax_h_high * data_width + argmax_w_low];
+    if (argmax_h_high <= height - 1 && argmax_w_high <= width - 1)
+      weight += (argmax_h - argmax_h_low) * im_data[argmax_h_high * data_width + argmax_w_high];
+  }
+
+  return weight;
+}
+
+template <typename scalar_t>
+__global__ void modulated_deformable_im2col_gpu_kernel(const int n,
+                                                       const scalar_t *data_im, const scalar_t *data_offset, const scalar_t *data_mask,
+                                                       const int height, const int width, const int kernel_h, const int kernel_w,
+                                                       const int pad_h, const int pad_w,
+                                                       const int stride_h, const int stride_w,
+                                                       const int dilation_h, const int dilation_w,
+                                                       const int channel_per_deformable_group,
+                                                       const int batch_size, const int num_channels, const int deformable_group,
+                                                       const int height_col, const int width_col,
+                                                       scalar_t *data_col)
+{
+  CUDA_KERNEL_LOOP(index, n)
+  {
+    // index index of output matrix
+    const int w_col = index % width_col;
+    const int h_col = (index / width_col) % height_col;
+    const int b_col = (index / width_col / height_col) % batch_size;
+    const int c_im = (index / width_col / height_col) / batch_size;
+    const int c_col = c_im * kernel_h * kernel_w;
+
+    // compute deformable group index
+    const int deformable_group_index = c_im / channel_per_deformable_group;
+
+    const int h_in = h_col * stride_h - pad_h;
+    const int w_in = w_col * stride_w - pad_w;
+
+    scalar_t *data_col_ptr = data_col + ((c_col * batch_size + b_col) * height_col + h_col) * width_col + w_col;
+    //const float* data_im_ptr = data_im + ((b_col * num_channels + c_im) * height + h_in) * width + w_in;
+    const scalar_t *data_im_ptr = data_im + (b_col * num_channels + c_im) * height * width;
+    const scalar_t *data_offset_ptr = data_offset + (b_col * deformable_group + deformable_group_index) * 2 * kernel_h * kernel_w * height_col * width_col;
+
+    const scalar_t *data_mask_ptr = data_mask + (b_col * deformable_group + deformable_group_index) * kernel_h * kernel_w * height_col * width_col;
+
+    for (int i = 0; i < kernel_h; ++i)
+    {
+      for (int j = 0; j < kernel_w; ++j)
+      {
+        const int data_offset_h_ptr = ((2 * (i * kernel_w + j)) * height_col + h_col) * width_col + w_col;
+        const int data_offset_w_ptr = ((2 * (i * kernel_w + j) + 1) * height_col + h_col) * width_col + w_col;
+        const int data_mask_hw_ptr = ((i * kernel_w + j) * height_col + h_col) * width_col + w_col;
+        const scalar_t offset_h = data_offset_ptr[data_offset_h_ptr];
+        const scalar_t offset_w = data_offset_ptr[data_offset_w_ptr];
+        const scalar_t mask = data_mask_ptr[data_mask_hw_ptr];
+        scalar_t val = static_cast<scalar_t>(0);
+        const scalar_t h_im = h_in + i * dilation_h + offset_h;
+        const scalar_t w_im = w_in + j * dilation_w + offset_w;
+        //if (h_im >= 0 && w_im >= 0 && h_im < height && w_im < width) {
+        if (h_im > -1 && w_im > -1 && h_im < height && w_im < width)
+        {
+          //const float map_h = i * dilation_h + offset_h;
+          //const float map_w = j * dilation_w + offset_w;
+          //const int cur_height = height - h_in;
+          //const int cur_width = width - w_in;
+          //val = dmcn_im2col_bilinear(data_im_ptr, width, cur_height, cur_width, map_h, map_w);
+          val = dmcn_im2col_bilinear(data_im_ptr, width, height, width, h_im, w_im);
+        }
+        *data_col_ptr = val * mask;
+        data_col_ptr += batch_size * height_col * width_col;
+        //data_col_ptr += height_col * width_col;
+      }
+    }
+  }
+}
+
+template <typename scalar_t>
+__global__ void modulated_deformable_col2im_gpu_kernel(const int n,
+                                                       const scalar_t *data_col, const scalar_t *data_offset, const scalar_t *data_mask,
+                                                       const int channels, const int height, const int width,
+                                                       const int kernel_h, const int kernel_w,
+                                                       const int pad_h, const int pad_w,
+                                                       const int stride_h, const int stride_w,
+                                                       const int dilation_h, const int dilation_w,
+                                                       const int channel_per_deformable_group,
+                                                       const int batch_size, const int deformable_group,
+                                                       const int height_col, const int width_col,
+                                                       scalar_t *grad_im)
+{
+  CUDA_KERNEL_LOOP(index, n)
+  {
+    const int j = (index / width_col / height_col / batch_size) % kernel_w;
+    const int i = (index / width_col / height_col / batch_size / kernel_w) % kernel_h;
+    const int c = index / width_col / height_col / batch_size / kernel_w / kernel_h;
+    // compute the start and end of the output
+
+    const int deformable_group_index = c / channel_per_deformable_group;
+
+    int w_out = index % width_col;
+    int h_out = (index / width_col) % height_col;
+    int b = (index / width_col / height_col) % batch_size;
+    int w_in = w_out * stride_w - pad_w;
+    int h_in = h_out * stride_h - pad_h;
+
+    const scalar_t *data_offset_ptr = data_offset + (b * deformable_group + deformable_group_index) * 2 * kernel_h * kernel_w * height_col * width_col;
+    const scalar_t *data_mask_ptr = data_mask + (b * deformable_group + deformable_group_index) * kernel_h * kernel_w * height_col * width_col;
+    const int data_offset_h_ptr = ((2 * (i * kernel_w + j)) * height_col + h_out) * width_col + w_out;
+    const int data_offset_w_ptr = ((2 * (i * kernel_w + j) + 1) * height_col + h_out) * width_col + w_out;
+    const int data_mask_hw_ptr = ((i * kernel_w + j) * height_col + h_out) * width_col + w_out;
+    const scalar_t offset_h = data_offset_ptr[data_offset_h_ptr];
+    const scalar_t offset_w = data_offset_ptr[data_offset_w_ptr];
+    const scalar_t mask = data_mask_ptr[data_mask_hw_ptr];
+    const scalar_t cur_inv_h_data = h_in + i * dilation_h + offset_h;
+    const scalar_t cur_inv_w_data = w_in + j * dilation_w + offset_w;
+
+    const scalar_t cur_top_grad = data_col[index] * mask;
+    const int cur_h = (int)cur_inv_h_data;
+    const int cur_w = (int)cur_inv_w_data;
+    for (int dy = -2; dy <= 2; dy++)
+    {
+      for (int dx = -2; dx <= 2; dx++)
+      {
+        if (cur_h + dy >= 0 && cur_h + dy < height &&
+            cur_w + dx >= 0 && cur_w + dx < width &&
+            abs(cur_inv_h_data - (cur_h + dy)) < 1 &&
+            abs(cur_inv_w_data - (cur_w + dx)) < 1)
+        {
+          int cur_bottom_grad_pos = ((b * channels + c) * height + cur_h + dy) * width + cur_w + dx;
+          scalar_t weight = dmcn_get_gradient_weight(cur_inv_h_data, cur_inv_w_data, cur_h + dy, cur_w + dx, height, width);
+          atomicAdd(grad_im + cur_bottom_grad_pos, weight * cur_top_grad);
+        }
+      }
+    }
+  }
+}
+
+template <typename scalar_t>
+__global__ void modulated_deformable_col2im_coord_gpu_kernel(const int n,
+                                                             const scalar_t *data_col, const scalar_t *data_im,
+                                                             const scalar_t *data_offset, const scalar_t *data_mask,
+                                                             const int channels, const int height, const int width,
+                                                             const int kernel_h, const int kernel_w,
+                                                             const int pad_h, const int pad_w,
+                                                             const int stride_h, const int stride_w,
+                                                             const int dilation_h, const int dilation_w,
+                                                             const int channel_per_deformable_group,
+                                                             const int batch_size, const int offset_channels, const int deformable_group,
+                                                             const int height_col, const int width_col,
+                                                             scalar_t *grad_offset, scalar_t *grad_mask)
+{
+  CUDA_KERNEL_LOOP(index, n)
+  {
+    scalar_t val = 0, mval = 0;
+    int w = index % width_col;
+    int h = (index / width_col) % height_col;
+    int c = (index / width_col / height_col) % offset_channels;
+    int b = (index / width_col / height_col) / offset_channels;
+    // compute the start and end of the output
+
+    const int deformable_group_index = c / (2 * kernel_h * kernel_w);
+    const int col_step = kernel_h * kernel_w;
+    int cnt = 0;
+    const scalar_t *data_col_ptr = data_col + deformable_group_index * channel_per_deformable_group * batch_size * width_col * height_col;
+    const scalar_t *data_im_ptr = data_im + (b * deformable_group + deformable_group_index) * channel_per_deformable_group / kernel_h / kernel_w * height * width;
+    const scalar_t *data_offset_ptr = data_offset + (b * deformable_group + deformable_group_index) * 2 * kernel_h * kernel_w * height_col * width_col;
+    const scalar_t *data_mask_ptr = data_mask + (b * deformable_group + deformable_group_index) * kernel_h * kernel_w * height_col * width_col;
+
+    const int offset_c = c - deformable_group_index * 2 * kernel_h * kernel_w;
+
+    for (int col_c = (offset_c / 2); col_c < channel_per_deformable_group; col_c += col_step)
+    {
+      const int col_pos = (((col_c * batch_size + b) * height_col) + h) * width_col + w;
+      const int bp_dir = offset_c % 2;
+
+      int j = (col_pos / width_col / height_col / batch_size) % kernel_w;
+      int i = (col_pos / width_col / height_col / batch_size / kernel_w) % kernel_h;
+      int w_out = col_pos % width_col;
+      int h_out = (col_pos / width_col) % height_col;
+      int w_in = w_out * stride_w - pad_w;
+      int h_in = h_out * stride_h - pad_h;
+      const int data_offset_h_ptr = (((2 * (i * kernel_w + j)) * height_col + h_out) * width_col + w_out);
+      const int data_offset_w_ptr = (((2 * (i * kernel_w + j) + 1) * height_col + h_out) * width_col + w_out);
+      const int data_mask_hw_ptr = (((i * kernel_w + j) * height_col + h_out) * width_col + w_out);
+      const scalar_t offset_h = data_offset_ptr[data_offset_h_ptr];
+      const scalar_t offset_w = data_offset_ptr[data_offset_w_ptr];
+      const scalar_t mask = data_mask_ptr[data_mask_hw_ptr];
+      scalar_t inv_h = h_in + i * dilation_h + offset_h;
+      scalar_t inv_w = w_in + j * dilation_w + offset_w;
+      if (inv_h <= -1 || inv_w <= -1 || inv_h >= height || inv_w >= width)
+      {
+        inv_h = inv_w = -2;
+      }
+      else
+      {
+        mval += data_col_ptr[col_pos] * dmcn_im2col_bilinear(data_im_ptr + cnt * height * width, width, height, width, inv_h, inv_w);
+      }
+      const scalar_t weight = dmcn_get_coordinate_weight(
+          inv_h, inv_w,
+          height, width, data_im_ptr + cnt * height * width, width, bp_dir);
+      val += weight * data_col_ptr[col_pos] * mask;
+      cnt += 1;
+    }
+    // KERNEL_ASSIGN(grad_offset[index], offset_req, val);
+    grad_offset[index] = val;
+    if (offset_c % 2 == 0)
+      // KERNEL_ASSIGN(grad_mask[(((b * deformable_group + deformable_group_index) * kernel_h * kernel_w + offset_c / 2) * height_col + h) * width_col + w], mask_req, mval);
+      grad_mask[(((b * deformable_group + deformable_group_index) * kernel_h * kernel_w + offset_c / 2) * height_col + h) * width_col + w] = mval;
+  }
+}
+
+void modulated_deformable_im2col_cuda(
+    const at::Tensor data_im, const at::Tensor data_offset, const at::Tensor data_mask,
+    const int batch_size, const int channels, const int height_im, const int width_im,
+    const int height_col, const int width_col, const int kernel_h, const int kenerl_w,
+    const int pad_h, const int pad_w, const int stride_h, const int stride_w,
+    const int dilation_h, const int dilation_w,
+    const int deformable_group, at::Tensor data_col)
+{
+  // num_axes should be smaller than block size
+  const int channel_per_deformable_group = channels / deformable_group;
+  const int num_kernels = channels * batch_size * height_col * width_col;
+
+  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
+      data_im.scalar_type(), "modulated_deformable_im2col_gpu", ([&] {
+        const scalar_t *data_im_ = data_im.data_ptr<scalar_t>();
+        const scalar_t *data_offset_ = data_offset.data_ptr<scalar_t>();
+        const scalar_t *data_mask_ = data_mask.data_ptr<scalar_t>();
+        scalar_t *data_col_ = data_col.data_ptr<scalar_t>();
+
+       hipLaunchKernelGGL(( modulated_deformable_im2col_gpu_kernel), dim3(GET_BLOCKS(num_kernels)), dim3(CUDA_NUM_THREADS), 0, at::hip::getCurrentHIPStreamMasqueradingAsCUDA(), 
+            num_kernels, data_im_, data_offset_, data_mask_, height_im, width_im, kernel_h, kenerl_w,
+            pad_h, pad_w, stride_h, stride_w, dilation_h, dilation_w, channel_per_deformable_group,
+            batch_size, channels, deformable_group, height_col, width_col, data_col_);
+      }));
+
+  hipError_t err = hipGetLastError();
+  if (err != hipSuccess)
+  {
+    printf("error in modulated_deformable_im2col_cuda: %s\n", hipGetErrorString(err));
+  }
+}
+
+void modulated_deformable_col2im_cuda(
+    const at::Tensor data_col, const at::Tensor data_offset, const at::Tensor data_mask,
+    const int batch_size, const int channels, const int height_im, const int width_im,
+    const int height_col, const int width_col, const int kernel_h, const int kernel_w,
+    const int pad_h, const int pad_w, const int stride_h, const int stride_w,
+    const int dilation_h, const int dilation_w,
+    const int deformable_group, at::Tensor grad_im)
+{
+
+  const int channel_per_deformable_group = channels / deformable_group;
+  const int num_kernels = channels * kernel_h * kernel_w * batch_size * height_col * width_col;
+
+  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
+      data_col.scalar_type(), "modulated_deformable_col2im_gpu", ([&] {
+        const scalar_t *data_col_ = data_col.data_ptr<scalar_t>();
+        const scalar_t *data_offset_ = data_offset.data_ptr<scalar_t>();
+        const scalar_t *data_mask_ = data_mask.data_ptr<scalar_t>();
+        scalar_t *grad_im_ = grad_im.data_ptr<scalar_t>();
+
+       hipLaunchKernelGGL(( modulated_deformable_col2im_gpu_kernel), dim3(GET_BLOCKS(num_kernels)), dim3(CUDA_NUM_THREADS), 0, at::hip::getCurrentHIPStreamMasqueradingAsCUDA(), 
+            num_kernels, data_col_, data_offset_, data_mask_, channels, height_im, width_im,
+            kernel_h, kernel_w, pad_h, pad_w, stride_h, stride_w,
+            dilation_h, dilation_w, channel_per_deformable_group,
+            batch_size, deformable_group, height_col, width_col, grad_im_);
+      }));
+
+  hipError_t err = hipGetLastError();
+  if (err != hipSuccess)
+  {
+    printf("error in modulated_deformable_col2im_cuda: %s\n", hipGetErrorString(err));
+  }
+}
+
+void modulated_deformable_col2im_coord_cuda(
+    const at::Tensor data_col, const at::Tensor data_im, const at::Tensor data_offset, const at::Tensor data_mask,
+    const int batch_size, const int channels, const int height_im, const int width_im,
+    const int height_col, const int width_col, const int kernel_h, const int kernel_w,
+    const int pad_h, const int pad_w, const int stride_h, const int stride_w,
+    const int dilation_h, const int dilation_w,
+    const int deformable_group,
+    at::Tensor grad_offset, at::Tensor grad_mask)
+{
+  const int num_kernels = batch_size * height_col * width_col * 2 * kernel_h * kernel_w * deformable_group;
+  const int channel_per_deformable_group = channels * kernel_h * kernel_w / deformable_group;
+
+  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
+      data_col.scalar_type(), "modulated_deformable_col2im_coord_gpu", ([&] {
+        const scalar_t *data_col_ = data_col.data_ptr<scalar_t>();
+        const scalar_t *data_im_ = data_im.data_ptr<scalar_t>();
+        const scalar_t *data_offset_ = data_offset.data_ptr<scalar_t>();
+        const scalar_t *data_mask_ = data_mask.data_ptr<scalar_t>();
+        scalar_t *grad_offset_ = grad_offset.data_ptr<scalar_t>();
+        scalar_t *grad_mask_ = grad_mask.data_ptr<scalar_t>();
+
+       hipLaunchKernelGGL(( modulated_deformable_col2im_coord_gpu_kernel), dim3(GET_BLOCKS(num_kernels)), dim3(CUDA_NUM_THREADS), 0, at::hip::getCurrentHIPStreamMasqueradingAsCUDA(), 
+            num_kernels, data_col_, data_im_, data_offset_, data_mask_, channels, height_im, width_im,
+            kernel_h, kernel_w, pad_h, pad_w, stride_h, stride_w,
+            dilation_h, dilation_w, channel_per_deformable_group,
+            batch_size, 2 * kernel_h * kernel_w * deformable_group, deformable_group, height_col, width_col,
+            grad_offset_, grad_mask_);
+      }));
+  hipError_t err = hipGetLastError();
+  if (err != hipSuccess)
+  {
+    printf("error in modulated_deformable_col2im_coord_cuda: %s\n", hipGetErrorString(err));
+  }
+}

basicsr/test.py

+import logging
+import torch
+from os import path as osp
+
+from basicsr.data import build_dataloader, build_dataset
+from basicsr.models import build_model
+from basicsr.utils import get_env_info, get_root_logger, get_time_str, make_exp_dirs
+from basicsr.utils.options import dict2str, parse_options
+
+
+def test_pipeline(root_path):
+    # parse options, set distributed setting, set ramdom seed
+    opt, _ = parse_options(root_path, is_train=False)
+
+    torch.backends.cudnn.benchmark = True
+    # torch.backends.cudnn.deterministic = True
+
+    # mkdir and initialize loggers
+    make_exp_dirs(opt)
+    log_file = osp.join(opt['path']['log'], f"test_{opt['name']}_{get_time_str()}.log")
+    logger = get_root_logger(logger_name='basicsr', log_level=logging.INFO, log_file=log_file)
+    logger.info(get_env_info())
+    logger.info(dict2str(opt))
+
+    # create test dataset and dataloader
+    test_loaders = []
+    for _, dataset_opt in sorted(opt['datasets'].items()):
+        test_set = build_dataset(dataset_opt)
+        test_loader = build_dataloader(
+            test_set, dataset_opt, num_gpu=opt['num_gpu'], dist=opt['dist'], sampler=None, seed=opt['manual_seed'])
+        logger.info(f"Number of test images in {dataset_opt['name']}: {len(test_set)}")
+        test_loaders.append(test_loader)
+
+    # create model
+    model = build_model(opt)
+
+    for test_loader in test_loaders:
+        test_set_name = test_loader.dataset.opt['name']
+        logger.info(f'Testing {test_set_name}...')
+        model.validation(test_loader, current_iter=opt['name'], tb_logger=None, save_img=opt['val']['save_img'])
+
+
+if __name__ == '__main__':
+    root_path = osp.abspath(osp.join(__file__, osp.pardir, osp.pardir))
+    test_pipeline(root_path)

basicsr/train.py

+import datetime
+import logging
+import math
+import time
+import torch
+import os
+import shutil
+from os import path as osp
+
+from basicsr.data import build_dataloader, build_dataset
+from basicsr.data.data_sampler import EnlargedSampler
+from basicsr.data.prefetch_dataloader import CPUPrefetcher, CUDAPrefetcher
+from basicsr.models import build_model
+from basicsr.utils import (AvgTimer, MessageLogger, check_resume, get_env_info, get_root_logger, get_time_str,
+                           init_tb_logger, init_wandb_logger, make_exp_dirs, mkdir_and_rename, scandir)
+from basicsr.utils.options import copy_opt_file, dict2str, parse_options
+torch.backends.cudnn.benchmark = True
+
+def mkdir_and_rename(path):
+    """mkdirs. If path exists, rename it with timestamp and create a new one.
+
+    Args:
+        path (str): Folder path.
+    """
+    if osp.exists(path):
+        new_name = path + '_archived_' + get_time_str()
+        new_name = new_name.replace('tb_logger', 'tb_logger_archived')
+        print(f'Path already exists. Rename it to {new_name}', flush=True)
+        shutil.move(path, new_name)
+    os.makedirs(path, exist_ok=True)
+
+
+def init_tb_loggers(opt):
+    # initialize wandb logger before tensorboard logger to allow proper sync
+    if (opt['logger'].get('wandb') is not None) and (opt['logger']['wandb'].get('project')
+                                                     is not None) and ('debug' not in opt['name']):
+        assert opt['logger'].get('use_tb_logger') is True, ('should turn on tensorboard when using wandb')
+        init_wandb_logger(opt)
+    tb_logger = None
+    if opt['logger'].get('use_tb_logger') and 'debug' not in opt['name']:
+        tb_logger = init_tb_logger(log_dir=osp.join(opt['root_path'], 'tb_logger', opt['name']))
+    return tb_logger
+
+
+def create_train_val_dataloader(opt, logger):
+    # create train and val dataloaders
+    train_loader, val_loaders = None, []
+    for phase, dataset_opt in opt['datasets'].items():
+        if phase == 'train':
+            dataset_enlarge_ratio = dataset_opt.get('dataset_enlarge_ratio', 1)
+            train_set = build_dataset(dataset_opt)
+            train_sampler = EnlargedSampler(train_set, opt['world_size'], opt['rank'], dataset_enlarge_ratio)
+            train_loader = build_dataloader(
+                train_set,
+                dataset_opt,
+                num_gpu=opt['num_gpu'],
+                dist=opt['dist'],
+                sampler=train_sampler,
+                seed=opt['manual_seed'])
+
+            num_iter_per_epoch = math.ceil(
+                len(train_set) * dataset_enlarge_ratio / (dataset_opt['batch_size_per_gpu'] * opt['world_size']))
+            total_iters = int(opt['train']['total_iter'])
+            total_epochs = math.ceil(total_iters / (num_iter_per_epoch))
+            logger.info('Training statistics:'
+                        f'\n\tNumber of train images: {len(train_set)}'
+                        f'\n\tDataset enlarge ratio: {dataset_enlarge_ratio}'
+                        f'\n\tBatch size per gpu: {dataset_opt["batch_size_per_gpu"]}'
+                        f'\n\tWorld size (gpu number): {opt["world_size"]}'
+                        f'\n\tRequire iter number per epoch: {num_iter_per_epoch}'
+                        f'\n\tTotal epochs: {total_epochs}; iters: {total_iters}.')
+        elif phase.split('_')[0] == 'val':
+            val_set = build_dataset(dataset_opt)
+            val_loader = build_dataloader(
+                val_set, dataset_opt, num_gpu=opt['num_gpu'], dist=opt['dist'], sampler=None, seed=opt['manual_seed'])
+            logger.info(f'Number of val images/folders in {dataset_opt["name"]}: {len(val_set)}')
+            val_loaders.append(val_loader)
+        else:
+            raise ValueError(f'Dataset phase {phase} is not recognized.')
+
+    return train_loader, train_sampler, val_loaders, total_epochs, total_iters
+
+
+def load_resume_state(opt):
+    resume_state_path = None
+    if opt['auto_resume']:
+        state_path = osp.join('experiments', opt['name'], 'training_states')
+        if osp.isdir(state_path):
+            states = list(scandir(state_path, suffix='state', recursive=False, full_path=False))
+            if len(states) != 0:
+                states = [float(v.split('.state')[0]) for v in states]
+                resume_state_path = osp.join(state_path, f'{max(states):.0f}.state')
+                opt['path']['resume_state'] = resume_state_path
+    else:
+        if opt['path'].get('resume_state'):
+            resume_state_path = opt['path']['resume_state']
+
+    if resume_state_path is None:
+        resume_state = None
+    else:
+        device_id = torch.cuda.current_device()
+        resume_state = torch.load(resume_state_path, map_location=lambda storage, loc: storage.cuda(device_id))
+        check_resume(opt, resume_state['iter'])
+    return resume_state
+
+
+def train_pipeline(root_path):
+    # parse options, set distributed setting, set ramdom seed
+    opt, args = parse_options(root_path, is_train=True)
+    opt['root_path'] = root_path
+
+    torch.backends.cudnn.benchmark = True
+    # torch.backends.cudnn.deterministic = True
+
+    # load resume states if necessary
+    resume_state = load_resume_state(opt)
+    # mkdir for experiments and logger
+    if resume_state is None:
+        make_exp_dirs(opt)
+        if opt['logger'].get('use_tb_logger') and 'debug' not in opt['name'] and opt['rank'] == 0:
+            os.makedirs(osp.join(opt['root_path'], 'tb_logger_archived'), exist_ok=True)
+            mkdir_and_rename(osp.join(opt['root_path'], 'tb_logger', opt['name']))
+
+    # copy the yml file to the experiment root
+    copy_opt_file(args.opt, opt['path']['experiments_root'])
+
+    # WARNING: should not use get_root_logger in the above codes, including the called functions
+    # Otherwise the logger will not be properly initialized
+    log_file = osp.join(opt['path']['log'], f"train_{opt['name']}_{get_time_str()}.log")
+    logger = get_root_logger(logger_name='basicsr', log_level=logging.INFO, log_file=log_file)
+    logger.info(get_env_info())
+    logger.info(dict2str(opt))
+    # initialize wandb and tb loggers
+    tb_logger = init_tb_loggers(opt)
+
+    # create train and validation dataloaders
+    result = create_train_val_dataloader(opt, logger)
+    train_loader, train_sampler, val_loaders, total_epochs, total_iters = result
+
+    # create model
+    model = build_model(opt)
+    if resume_state:  # resume training
+        model.resume_training(resume_state)  # handle optimizers and schedulers
+        logger.info(f"Resuming training from epoch: {resume_state['epoch']}, " f"iter: {resume_state['iter']}.")
+        start_epoch = resume_state['epoch']
+        current_iter = resume_state['iter']
+    else:
+        start_epoch = 0
+        current_iter = 0
+
+    # create message logger (formatted outputs)
+    msg_logger = MessageLogger(opt, current_iter, tb_logger)
+
+    # dataloader prefetcher
+    prefetch_mode = opt['datasets']['train'].get('prefetch_mode')
+    if prefetch_mode is None or prefetch_mode == 'cpu':
+        prefetcher = CPUPrefetcher(train_loader)
+    elif prefetch_mode == 'cuda':
+        prefetcher = CUDAPrefetcher(train_loader, opt)
+        logger.info(f'Use {prefetch_mode} prefetch dataloader')
+        if opt['datasets']['train'].get('pin_memory') is not True:
+            raise ValueError('Please set pin_memory=True for CUDAPrefetcher.')
+    else:
+        raise ValueError(f'Wrong prefetch_mode {prefetch_mode}.' "Supported ones are: None, 'cuda', 'cpu'.")
+
+    # training
+    logger.info(f'Start training from epoch: {start_epoch}, iter: {current_iter}')
+    data_timer, iter_timer = AvgTimer(), AvgTimer()
+    start_time = time.time()
+
+    for epoch in range(start_epoch, total_epochs + 1):
+        train_sampler.set_epoch(epoch)
+        prefetcher.reset()
+        train_data = prefetcher.next()
+
+        while train_data is not None:
+            data_timer.record()
+
+            current_iter += 1
+            if current_iter > total_iters:
+                break
+            # update learning rate
+            model.update_learning_rate(current_iter, warmup_iter=opt['train'].get('warmup_iter', -1))
+            # training
+            model.feed_data(train_data)
+            model.optimize_parameters(current_iter)
+            iter_timer.record()
+            if current_iter == 1:
+                # reset start time in msg_logger for more accurate eta_time
+                # not work in resume mode
+                msg_logger.reset_start_time()
+            # log
+            if current_iter % opt['logger']['print_freq'] == 0:
+                log_vars = {'epoch': epoch, 'iter': current_iter}
+                log_vars.update({'lrs': model.get_current_learning_rate()})
+                log_vars.update({'time': iter_timer.get_avg_time(), 'data_time': data_timer.get_avg_time()})
+                log_vars.update(model.get_current_log())
+                msg_logger(log_vars)
+
+            # save images
+            if current_iter % (opt['logger']['show_tf_imgs_freq']) == 0:
+                visual_imgs = model.get_current_visuals()
+                if tb_logger:
+                    for k, v in visual_imgs.items(): 
+                        tb_logger.add_images(f'ckpt_imgs/{k}', v.clamp(0, 1), current_iter)
+
+            # save models and training states
+            if current_iter % opt['logger']['save_checkpoint_freq'] == 0:
+                logger.info('Saving models and training states.')
+                model.save(epoch, current_iter)
+
+            if current_iter % opt['logger']['save_latest_freq'] == 0:
+                logger.info('Saving models and training states.')
+                model.save(epoch, -1)
+
+            # validation
+            if opt.get('val') is not None and (current_iter % opt['val']['val_freq'] == 0):
+                if len(val_loaders) > 1:
+                    logger.warning('Multiple validation datasets are *only* supported by SRModel.')
+                for val_loader in val_loaders:
+                    model.validation(val_loader, current_iter, tb_logger, opt['val']['save_img'])
+
+            data_timer.start()
+            iter_timer.start()
+            train_data = prefetcher.next()
+        # end of iter
+
+    # end of epoch
+
+    consumed_time = str(datetime.timedelta(seconds=int(time.time() - start_time)))
+    logger.info(f'End of training. Time consumed: {consumed_time}')
+    logger.info('Save the latest model.')
+    model.save(epoch=-1, current_iter=-1)  # -1 stands for the latest
+    if opt.get('val') is not None:
+        for val_loader in val_loaders:
+            model.validation(val_loader, current_iter, tb_logger, opt['val']['save_img'])
+    if tb_logger:
+        tb_logger.close()
+
+
+if __name__ == '__main__':
+    root_path = osp.abspath(osp.join(__file__, osp.pardir, osp.pardir))
+    train_pipeline(root_path)

basicsr/utils/__init__.py

+from .diffjpeg import DiffJPEG
+from .file_client import FileClient
+from .img_process_util import USMSharp, usm_sharp
+from .img_util import crop_border, imfrombytes, img2tensor, imwrite, tensor2img
+from .logger import AvgTimer, MessageLogger, get_env_info, get_root_logger, init_tb_logger, init_wandb_logger
+from .misc import check_resume, get_time_str, make_exp_dirs, mkdir_and_rename, scandir, set_random_seed, sizeof_fmt
+
+__all__ = [
+    # file_client.py
+    'FileClient',
+    # img_util.py
+    'img2tensor',
+    'tensor2img',
+    'imfrombytes',
+    'imwrite',
+    'crop_border',
+    # logger.py
+    'MessageLogger',
+    'AvgTimer',
+    'init_tb_logger',
+    'init_wandb_logger',
+    'get_root_logger',
+    'get_env_info',
+    # misc.py
+    'set_random_seed',
+    'get_time_str',
+    'mkdir_and_rename',
+    'make_exp_dirs',
+    'scandir',
+    'check_resume',
+    'sizeof_fmt',
+    # diffjpeg
+    'DiffJPEG',
+    # img_process_util
+    'USMSharp',
+    'usm_sharp'
+]

basicsr/utils/diffjpeg.py

+"""
+Modified from https://github.com/mlomnitz/DiffJPEG
+
+For images not divisible by 8
+https://dsp.stackexchange.com/questions/35339/jpeg-dct-padding/35343#35343
+"""
+import itertools
+import numpy as np
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+
+# ------------------------ utils ------------------------#
+y_table = np.array(
+    [[16, 11, 10, 16, 24, 40, 51, 61], [12, 12, 14, 19, 26, 58, 60, 55], [14, 13, 16, 24, 40, 57, 69, 56],
+     [14, 17, 22, 29, 51, 87, 80, 62], [18, 22, 37, 56, 68, 109, 103, 77], [24, 35, 55, 64, 81, 104, 113, 92],
+     [49, 64, 78, 87, 103, 121, 120, 101], [72, 92, 95, 98, 112, 100, 103, 99]],
+    dtype=np.float32).T
+y_table = nn.Parameter(torch.from_numpy(y_table))
+c_table = np.empty((8, 8), dtype=np.float32)
+c_table.fill(99)
+c_table[:4, :4] = np.array([[17, 18, 24, 47], [18, 21, 26, 66], [24, 26, 56, 99], [47, 66, 99, 99]]).T
+c_table = nn.Parameter(torch.from_numpy(c_table))
+
+
+def diff_round(x):
+    """ Differentiable rounding function
+    """
+    return torch.round(x) + (x - torch.round(x))**3
+
+
+def quality_to_factor(quality):
+    """ Calculate factor corresponding to quality
+
+    Args:
+        quality(float): Quality for jpeg compression.
+
+    Returns:
+        float: Compression factor.
+    """
+    if quality < 50:
+        quality = 5000. / quality
+    else:
+        quality = 200. - quality * 2
+    return quality / 100.
+
+
+# ------------------------ compression ------------------------#
+class RGB2YCbCrJpeg(nn.Module):
+    """ Converts RGB image to YCbCr
+    """
+
+    def __init__(self):
+        super(RGB2YCbCrJpeg, self).__init__()
+        matrix = np.array([[0.299, 0.587, 0.114], [-0.168736, -0.331264, 0.5], [0.5, -0.418688, -0.081312]],
+                          dtype=np.float32).T
+        self.shift = nn.Parameter(torch.tensor([0., 128., 128.]))
+        self.matrix = nn.Parameter(torch.from_numpy(matrix))
+
+    def forward(self, image):
+        """
+        Args:
+            image(Tensor): batch x 3 x height x width
+
+        Returns:
+            Tensor: batch x height x width x 3
+        """
+        image = image.permute(0, 2, 3, 1)
+        result = torch.tensordot(image, self.matrix, dims=1) + self.shift
+        return result.view(image.shape)
+
+
+class ChromaSubsampling(nn.Module):
+    """ Chroma subsampling on CbCr channels
+    """
+
+    def __init__(self):
+        super(ChromaSubsampling, self).__init__()
+
+    def forward(self, image):
+        """
+        Args:
+            image(tensor): batch x height x width x 3
+
+        Returns:
+            y(tensor): batch x height x width
+            cb(tensor): batch x height/2 x width/2
+            cr(tensor): batch x height/2 x width/2
+        """
+        image_2 = image.permute(0, 3, 1, 2).clone()
+        cb = F.avg_pool2d(image_2[:, 1, :, :].unsqueeze(1), kernel_size=2, stride=(2, 2), count_include_pad=False)
+        cr = F.avg_pool2d(image_2[:, 2, :, :].unsqueeze(1), kernel_size=2, stride=(2, 2), count_include_pad=False)
+        cb = cb.permute(0, 2, 3, 1)
+        cr = cr.permute(0, 2, 3, 1)
+        return image[:, :, :, 0], cb.squeeze(3), cr.squeeze(3)
+
+
+class BlockSplitting(nn.Module):
+    """ Splitting image into patches
+    """
+
+    def __init__(self):
+        super(BlockSplitting, self).__init__()
+        self.k = 8
+
+    def forward(self, image):
+        """
+        Args:
+            image(tensor): batch x height x width
+
+        Returns:
+            Tensor:  batch x h*w/64 x h x w
+        """
+        height, _ = image.shape[1:3]
+        batch_size = image.shape[0]
+        image_reshaped = image.view(batch_size, height // self.k, self.k, -1, self.k)
+        image_transposed = image_reshaped.permute(0, 1, 3, 2, 4)
+        return image_transposed.contiguous().view(batch_size, -1, self.k, self.k)
+
+
+class DCT8x8(nn.Module):
+    """ Discrete Cosine Transformation
+    """
+
+    def __init__(self):
+        super(DCT8x8, self).__init__()
+        tensor = np.zeros((8, 8, 8, 8), dtype=np.float32)
+        for x, y, u, v in itertools.product(range(8), repeat=4):
+            tensor[x, y, u, v] = np.cos((2 * x + 1) * u * np.pi / 16) * np.cos((2 * y + 1) * v * np.pi / 16)
+        alpha = np.array([1. / np.sqrt(2)] + [1] * 7)
+        self.tensor = nn.Parameter(torch.from_numpy(tensor).float())
+        self.scale = nn.Parameter(torch.from_numpy(np.outer(alpha, alpha) * 0.25).float())
+
+    def forward(self, image):
+        """
+        Args:
+            image(tensor): batch x height x width
+
+        Returns:
+            Tensor: batch x height x width
+        """
+        image = image - 128
+        result = self.scale * torch.tensordot(image, self.tensor, dims=2)
+        result.view(image.shape)
+        return result
+
+
+class YQuantize(nn.Module):
+    """ JPEG Quantization for Y channel
+
+    Args:
+        rounding(function): rounding function to use
+    """
+
+    def __init__(self, rounding):
+        super(YQuantize, self).__init__()
+        self.rounding = rounding
+        self.y_table = y_table
+
+    def forward(self, image, factor=1):
+        """
+        Args:
+            image(tensor): batch x height x width
+
+        Returns:
+            Tensor: batch x height x width
+        """
+        if isinstance(factor, (int, float)):
+            image = image.float() / (self.y_table * factor)
+        else:
+            b = factor.size(0)
+            table = self.y_table.expand(b, 1, 8, 8) * factor.view(b, 1, 1, 1)
+            image = image.float() / table
+        image = self.rounding(image)
+        return image
+
+
+class CQuantize(nn.Module):
+    """ JPEG Quantization for CbCr channels
+
+    Args:
+        rounding(function): rounding function to use
+    """
+
+    def __init__(self, rounding):
+        super(CQuantize, self).__init__()
+        self.rounding = rounding
+        self.c_table = c_table
+
+    def forward(self, image, factor=1):
+        """
+        Args:
+            image(tensor): batch x height x width
+
+        Returns:
+            Tensor: batch x height x width
+        """
+        if isinstance(factor, (int, float)):
+            image = image.float() / (self.c_table * factor)
+        else:
+            b = factor.size(0)
+            table = self.c_table.expand(b, 1, 8, 8) * factor.view(b, 1, 1, 1)
+            image = image.float() / table
+        image = self.rounding(image)
+        return image
+
+
+class CompressJpeg(nn.Module):
+    """Full JPEG compression algorithm
+
+    Args:
+        rounding(function): rounding function to use
+    """
+
+    def __init__(self, rounding=torch.round):
+        super(CompressJpeg, self).__init__()
+        self.l1 = nn.Sequential(RGB2YCbCrJpeg(), ChromaSubsampling())
+        self.l2 = nn.Sequential(BlockSplitting(), DCT8x8())
+        self.c_quantize = CQuantize(rounding=rounding)
+        self.y_quantize = YQuantize(rounding=rounding)
+
+    def forward(self, image, factor=1):
+        """
+        Args:
+            image(tensor): batch x 3 x height x width
+
+        Returns:
+            dict(tensor): Compressed tensor with batch x h*w/64 x 8 x 8.
+        """
+        y, cb, cr = self.l1(image * 255)
+        components = {'y': y, 'cb': cb, 'cr': cr}
+        for k in components.keys():
+            comp = self.l2(components[k])
+            if k in ('cb', 'cr'):
+                comp = self.c_quantize(comp, factor=factor)
+            else:
+                comp = self.y_quantize(comp, factor=factor)
+
+            components[k] = comp
+
+        return components['y'], components['cb'], components['cr']
+
+
+# ------------------------ decompression ------------------------#
+
+
+class YDequantize(nn.Module):
+    """Dequantize Y channel
+    """
+
+    def __init__(self):
+        super(YDequantize, self).__init__()
+        self.y_table = y_table
+
+    def forward(self, image, factor=1):
+        """
+        Args:
+            image(tensor): batch x height x width
+
+        Returns:
+            Tensor: batch x height x width
+        """
+        if isinstance(factor, (int, float)):
+            out = image * (self.y_table * factor)
+        else:
+            b = factor.size(0)
+            table = self.y_table.expand(b, 1, 8, 8) * factor.view(b, 1, 1, 1)
+            out = image * table
+        return out
+
+
+class CDequantize(nn.Module):
+    """Dequantize CbCr channel
+    """
+
+    def __init__(self):
+        super(CDequantize, self).__init__()
+        self.c_table = c_table
+
+    def forward(self, image, factor=1):
+        """
+        Args:
+            image(tensor): batch x height x width
+
+        Returns:
+            Tensor: batch x height x width
+        """
+        if isinstance(factor, (int, float)):
+            out = image * (self.c_table * factor)
+        else:
+            b = factor.size(0)
+            table = self.c_table.expand(b, 1, 8, 8) * factor.view(b, 1, 1, 1)
+            out = image * table
+        return out
+
+
+class iDCT8x8(nn.Module):
+    """Inverse discrete Cosine Transformation
+    """
+
+    def __init__(self):
+        super(iDCT8x8, self).__init__()
+        alpha = np.array([1. / np.sqrt(2)] + [1] * 7)
+        self.alpha = nn.Parameter(torch.from_numpy(np.outer(alpha, alpha)).float())
+        tensor = np.zeros((8, 8, 8, 8), dtype=np.float32)
+        for x, y, u, v in itertools.product(range(8), repeat=4):
+            tensor[x, y, u, v] = np.cos((2 * u + 1) * x * np.pi / 16) * np.cos((2 * v + 1) * y * np.pi / 16)
+        self.tensor = nn.Parameter(torch.from_numpy(tensor).float())
+
+    def forward(self, image):
+        """
+        Args:
+            image(tensor): batch x height x width
+
+        Returns:
+            Tensor: batch x height x width
+        """
+        image = image * self.alpha
+        result = 0.25 * torch.tensordot(image, self.tensor, dims=2) + 128
+        result.view(image.shape)
+        return result
+
+
+class BlockMerging(nn.Module):
+    """Merge patches into image
+    """
+
+    def __init__(self):
+        super(BlockMerging, self).__init__()
+
+    def forward(self, patches, height, width):
+        """
+        Args:
+            patches(tensor) batch x height*width/64, height x width
+            height(int)
+            width(int)
+
+        Returns:
+            Tensor: batch x height x width
+        """
+        k = 8
+        batch_size = patches.shape[0]
+        image_reshaped = patches.view(batch_size, height // k, width // k, k, k)
+        image_transposed = image_reshaped.permute(0, 1, 3, 2, 4)
+        return image_transposed.contiguous().view(batch_size, height, width)
+
+
+class ChromaUpsampling(nn.Module):
+    """Upsample chroma layers
+    """
+
+    def __init__(self):
+        super(ChromaUpsampling, self).__init__()
+
+    def forward(self, y, cb, cr):
+        """
+        Args:
+            y(tensor): y channel image
+            cb(tensor): cb channel
+            cr(tensor): cr channel
+
+        Returns:
+            Tensor: batch x height x width x 3
+        """
+
+        def repeat(x, k=2):
+            height, width = x.shape[1:3]
+            x = x.unsqueeze(-1)
+            x = x.repeat(1, 1, k, k)
+            x = x.view(-1, height * k, width * k)
+            return x
+
+        cb = repeat(cb)
+        cr = repeat(cr)
+        return torch.cat([y.unsqueeze(3), cb.unsqueeze(3), cr.unsqueeze(3)], dim=3)
+
+
+class YCbCr2RGBJpeg(nn.Module):
+    """Converts YCbCr image to RGB JPEG
+    """
+
+    def __init__(self):
+        super(YCbCr2RGBJpeg, self).__init__()
+
+        matrix = np.array([[1., 0., 1.402], [1, -0.344136, -0.714136], [1, 1.772, 0]], dtype=np.float32).T
+        self.shift = nn.Parameter(torch.tensor([0, -128., -128.]))
+        self.matrix = nn.Parameter(torch.from_numpy(matrix))
+
+    def forward(self, image):
+        """
+        Args:
+            image(tensor): batch x height x width x 3
+
+        Returns:
+            Tensor: batch x 3 x height x width
+        """
+        result = torch.tensordot(image + self.shift, self.matrix, dims=1)
+        return result.view(image.shape).permute(0, 3, 1, 2)
+
+
+class DeCompressJpeg(nn.Module):
+    """Full JPEG decompression algorithm
+
+    Args:
+        rounding(function): rounding function to use
+    """
+
+    def __init__(self, rounding=torch.round):
+        super(DeCompressJpeg, self).__init__()
+        self.c_dequantize = CDequantize()
+        self.y_dequantize = YDequantize()
+        self.idct = iDCT8x8()
+        self.merging = BlockMerging()
+        self.chroma = ChromaUpsampling()
+        self.colors = YCbCr2RGBJpeg()
+
+    def forward(self, y, cb, cr, imgh, imgw, factor=1):
+        """
+        Args:
+            compressed(dict(tensor)): batch x h*w/64 x 8 x 8
+            imgh(int)
+            imgw(int)
+            factor(float)
+
+        Returns:
+            Tensor: batch x 3 x height x width
+        """
+        components = {'y': y, 'cb': cb, 'cr': cr}
+        for k in components.keys():
+            if k in ('cb', 'cr'):
+                comp = self.c_dequantize(components[k], factor=factor)
+                height, width = int(imgh / 2), int(imgw / 2)
+            else:
+                comp = self.y_dequantize(components[k], factor=factor)
+                height, width = imgh, imgw
+            comp = self.idct(comp)
+            components[k] = self.merging(comp, height, width)
+            #
+        image = self.chroma(components['y'], components['cb'], components['cr'])
+        image = self.colors(image)
+
+        image = torch.min(255 * torch.ones_like(image), torch.max(torch.zeros_like(image), image))
+        return image / 255
+
+
+# ------------------------ main DiffJPEG ------------------------ #
+
+
+class DiffJPEG(nn.Module):
+    """This JPEG algorithm result is slightly different from cv2.
+    DiffJPEG supports batch processing.
+
+    Args:
+        differentiable(bool): If True, uses custom differentiable rounding function, if False, uses standard torch.round
+    """
+
+    def __init__(self, differentiable=True):
+        super(DiffJPEG, self).__init__()
+        if differentiable:
+            rounding = diff_round
+        else:
+            rounding = torch.round
+
+        self.compress = CompressJpeg(rounding=rounding)
+        self.decompress = DeCompressJpeg(rounding=rounding)
+
+    def forward(self, x, quality):
+        """
+        Args:
+            x (Tensor): Input image, bchw, rgb, [0, 1]
+            quality(float): Quality factor for jpeg compression scheme.
+        """
+        factor = quality
+        if isinstance(factor, (int, float)):
+            factor = quality_to_factor(factor)
+        else:
+            for i in range(factor.size(0)):
+                factor[i] = quality_to_factor(factor[i])
+        h, w = x.size()[-2:]
+        h_pad, w_pad = 0, 0
+        # why should use 16
+        if h % 16 != 0:
+            h_pad = 16 - h % 16
+        if w % 16 != 0:
+            w_pad = 16 - w % 16
+        x = F.pad(x, (0, w_pad, 0, h_pad), mode='constant', value=0)
+
+        y, cb, cr = self.compress(x, factor=factor)
+        recovered = self.decompress(y, cb, cr, (h + h_pad), (w + w_pad), factor=factor)
+        recovered = recovered[:, :, 0:h, 0:w]
+        return recovered
+
+
+if __name__ == '__main__':
+    import cv2
+
+    from basicsr.utils import img2tensor, tensor2img
+
+    img_gt = cv2.imread('test.png') / 255.
+
+    # -------------- cv2 -------------- #
+    encode_param = [int(cv2.IMWRITE_JPEG_QUALITY), 20]
+    _, encimg = cv2.imencode('.jpg', img_gt * 255., encode_param)
+    img_lq = np.float32(cv2.imdecode(encimg, 1))
+    cv2.imwrite('cv2_JPEG_20.png', img_lq)
+
+    # -------------- DiffJPEG -------------- #
+    jpeger = DiffJPEG(differentiable=False).cuda()
+    img_gt = img2tensor(img_gt)
+    img_gt = torch.stack([img_gt, img_gt]).cuda()
+    quality = img_gt.new_tensor([20, 40])
+    out = jpeger(img_gt, quality=quality)
+
+    cv2.imwrite('pt_JPEG_20.png', tensor2img(out[0]))
+    cv2.imwrite('pt_JPEG_40.png', tensor2img(out[1]))

basicsr/utils/dist_util.py

+# Modified from https://github.com/open-mmlab/mmcv/blob/master/mmcv/runner/dist_utils.py  # noqa: E501
+import functools
+import os
+import subprocess
+import torch
+import torch.distributed as dist
+import torch.multiprocessing as mp
+
+
+def init_dist(launcher, backend='nccl', **kwargs):
+    if mp.get_start_method(allow_none=True) is None:
+        mp.set_start_method('spawn')
+    if launcher == 'pytorch':
+        _init_dist_pytorch(backend, **kwargs)
+    elif launcher == 'slurm':
+        _init_dist_slurm(backend, **kwargs)
+    else:
+        raise ValueError(f'Invalid launcher type: {launcher}')
+
+
+def _init_dist_pytorch(backend, **kwargs):
+    rank = int(os.environ['RANK'])
+    num_gpus = torch.cuda.device_count()
+    torch.cuda.set_device(rank % num_gpus)
+    dist.init_process_group(backend=backend, **kwargs)
+
+
+def _init_dist_slurm(backend, port=None):
+    """Initialize slurm distributed training environment.
+
+    If argument ``port`` is not specified, then the master port will be system
+    environment variable ``MASTER_PORT``. If ``MASTER_PORT`` is not in system
+    environment variable, then a default port ``29500`` will be used.
+
+    Args:
+        backend (str): Backend of torch.distributed.
+        port (int, optional): Master port. Defaults to None.
+    """
+    proc_id = int(os.environ['SLURM_PROCID'])
+    ntasks = int(os.environ['SLURM_NTASKS'])
+    node_list = os.environ['SLURM_NODELIST']
+    num_gpus = torch.cuda.device_count()
+    torch.cuda.set_device(proc_id % num_gpus)
+    addr = subprocess.getoutput(f'scontrol show hostname {node_list} | head -n1')
+    # specify master port
+    if port is not None:
+        os.environ['MASTER_PORT'] = str(port)
+    elif 'MASTER_PORT' in os.environ:
+        pass  # use MASTER_PORT in the environment variable
+    else:
+        # 29500 is torch.distributed default port
+        os.environ['MASTER_PORT'] = '29500'
+    os.environ['MASTER_ADDR'] = addr
+    os.environ['WORLD_SIZE'] = str(ntasks)
+    os.environ['LOCAL_RANK'] = str(proc_id % num_gpus)
+    os.environ['RANK'] = str(proc_id)
+    dist.init_process_group(backend=backend)
+
+
+def get_dist_info():
+    if dist.is_available():
+        initialized = dist.is_initialized()
+    else:
+        initialized = False
+    if initialized:
+        rank = dist.get_rank()
+        world_size = dist.get_world_size()
+    else:
+        rank = 0
+        world_size = 1
+    return rank, world_size
+
+
+def master_only(func):
+
+    @functools.wraps(func)
+    def wrapper(*args, **kwargs):
+        rank, _ = get_dist_info()
+        if rank == 0:
+            return func(*args, **kwargs)
+
+    return wrapper