im2col.h

#ifndef THCUNN_IM2COL_H
#define THCUNN_IM2COL_H

#include "common.h"

// Kernel for fast unfold+copy
// (borrowed from Caffe: https://github.com/BVLC/caffe/blob/master/src/caffe/layers/conv_layer.cu)
template <typename Dtype>
__global__ void im2col_kernel(const int n, const Dtype* data_im,
    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 height_col, const int width_col,
    Dtype* data_col) {
  CUDA_KERNEL_LOOP(index, n) {
    int w_out = index % width_col;
    index /= width_col;
    int h_out = index % height_col;
    int channel_in = index / height_col;
    int channel_out = channel_in * ksize_h * ksize_w;
    int h_in = h_out * stride_h - pad_h;
    int w_in = w_out * stride_w - pad_w;
    data_col += (channel_out * height_col + h_out) * width_col + w_out;
    data_im += (channel_in * height + h_in) * width + w_in;
    for (int i = 0; i < ksize_h; ++i) {
      for (int j = 0; j < ksize_w; ++j) {
        int h = h_in + i;
        int w = w_in + j;
        *data_col = (h >= 0 && w >= 0 && h < height && w < width) ?
          data_im[i * width + j] : 0;
        data_col += height_col * width_col;
      }
    }
  }
}

template <typename Dtype>
void im2col(cudaStream_t stream, const Dtype* data_im, 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, Dtype* data_col) {
  // We are going to launch channels * height_col * width_col kernels, each
  // kernel responsible for copying a single-channel grid.
  int height_col = (height + 2 * pad_h - ksize_h) / stride_h + 1;
  int width_col = (width + 2 * pad_w - ksize_w) / stride_w + 1;
  int num_kernels = channels * height_col * width_col;
  // Launch
  im2col_kernel <<<GET_BLOCKS(num_kernels), CUDA_NUM_THREADS, 0, stream>>> (
      num_kernels, data_im, height, width, ksize_h, ksize_w,
      pad_h, pad_w, stride_h, stride_w,
      height_col, width_col, data_col
  );
}

template <typename Dtype>
__global__ void col2im_kernel(const int n, const Dtype* data_col,
    const int height, const int width, const int channels, const int patch_h, const int patch_w,
    const int pad_h, const int pad_w, const int stride_h, const int stride_w, const int height_col, const int width_col,
    Dtype* data_im) {
  CUDA_KERNEL_LOOP(index, n) {
    Dtype val = 0;
    int w = index % width + pad_w;
    int h = (index / width) % height + pad_h;
    int c = index / (width * height);
    // compute the start and end of the output
    int w_col_start = (w < patch_w) ? 0 : (w - patch_w) / stride_w + 1;
    int w_col_end = min(w / stride_w + 1, width_col);
    int h_col_start = (h < patch_h) ? 0 : (h - patch_h) / stride_h + 1;
    int h_col_end = min(h / stride_h + 1, height_col);
    /*
       for (int h_col = h_col_start; h_col < h_col_end; ++h_col) {
       for (int w_col = w_col_start; w_col < w_col_end; ++w_col) {
    // the col location: [c * width * height + h_out, w_out]
    int c_col = c * patch_h * patch_w + (h - h_col * stride_h) * ksize + (w - w_col * stride_w);
    val += data_col[(c_col * height_col + h_col) * width_col + w_col];
    }
    }
     */
    // equivalent implementation
    int offset = (c * patch_h * patch_w + h * patch_w + w) * height_col * width_col;
    int coeff_h_col = (1 - stride_h * patch_w * height_col) * width_col;
    int coeff_w_col = (1 - stride_w * height_col * width_col);
    for (int h_col = h_col_start; h_col < h_col_end; ++h_col) {
      for (int w_col = w_col_start; w_col < w_col_end; ++w_col) {
        val += data_col[offset + h_col * coeff_h_col + w_col * coeff_w_col];
      }
    }
    data_im[index] = val;
  }
}

template <typename Dtype>
void col2im(cudaStream_t stream, const Dtype* data_col, const int channels,
    const int height, const int width, const int patch_h, const int patch_w, const int pad_h,
    const int pad_w, const int stride_h, const int stride_w, Dtype* data_im) {
  int height_col = (height + 2 * pad_h - patch_h) / stride_h + 1;
  int width_col = (width + 2 * pad_w - patch_w) / stride_w + 1;
  int num_kernels = channels * height * width;
  // To avoid involving atomic operations, we will launch one kernel per
  // bottom dimension, and then in the kernel add up the top dimensions.
  col2im_kernel <<<GET_BLOCKS(num_kernels), CUDA_NUM_THREADS, 0, stream>>> (
      num_kernels, data_col, height, width, channels,
      patch_h, patch_w, pad_h, pad_w, stride_h, stride_w,
      height_col, width_col, data_im
  );
}

#endif