void col2im_cpu(const Dtype* data_col, 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,
Dtype* data_im) {
caffe_set(height * width * channels, Dtype(0), data_im);
const int output_h = (height + 2 * pad_h -
(dilation_h * (kernel_h - 1) + 1)) / stride_h + 1;
const int output_w = (width + 2 * pad_w -
(dilation_w * (kernel_w - 1) + 1)) / stride_w + 1;
const int channel_size = height * width;
for (int channel = channels; channel--; data_im += channel_size) {
for (int kernel_row = 0; kernel_row < kernel_h; kernel_row++) {
for (int kernel_col = 0; kernel_col < kernel_w; kernel_col++) {
int input_row = -pad_h + kernel_row * dilation_h;
for (int output_rows = output_h; output_rows; output_rows--) {
if (!is_a_ge_zero_and_a_lt_b(input_row, height)) {
data_col += output_w;
} else {
int input_col = -pad_w + kernel_col * dilation_w;
for (int output_col = output_w; output_col; output_col--) {
if (is_a_ge_zero_and_a_lt_b(input_col, width)) {
data_im[input_row * width + input_col] += *data_col;
}
data_col++;
input_col += stride_w;
}
}
input_row += stride_h;
}
}
}
}
}
void im2col_cpu_big(const Dtype* data_im, 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,
Dtype* data_col, int bs) {
const int output_h = (height + 2 * pad_h -
(dilation_h * (kernel_h - 1) + 1)) / stride_h + 1;
const int output_w = (width + 2 * pad_w -
(dilation_w * (kernel_w - 1) + 1)) / stride_w + 1;
const int channel_size = height * width;
for (int channel = 0; channel < channels; ++ channel) {
for (int kernel_row = 0; kernel_row < kernel_h; kernel_row++) {
for (int kernel_col = 0; kernel_col < kernel_w; kernel_col++) {
for (int b = 0; b < bs; ++b) {
int input_row = -pad_h + kernel_row * dilation_h;
const Dtype * data = data_im + (b * channels + channel) * channel_size;
for (int output_rows = output_h; output_rows; output_rows--) {
if (!is_a_ge_zero_and_a_lt_b(input_row, height)) {
for (int output_cols = output_w; output_cols; output_cols--) {
*(data_col++) = 0;
} // end for ow?
} else {
int input_col = -pad_w + kernel_col * dilation_w;
for (int output_col = output_w; output_col; output_col--) {
if (is_a_ge_zero_and_a_lt_b(input_col, width)) {
*(data_col++) = data[input_row * width + input_col];
} else {
*(data_col++) = 0;
} // end if
input_col += stride_w;
} // end for ow
}
input_row += stride_h;
} // end for oh
} // end for batch size
} // end for kenel_col
} // end for kernel_row
} // end for channel
}
void im2col_cpu(const Dtype* data_im, const int_tp channels,
const int_tp height, const int_tp width, const int_tp kernel_h,
const int_tp kernel_w, const int_tp pad_h, const int_tp pad_w,
const int_tp stride_h, const int_tp stride_w,
const int_tp dilation_h, const int_tp dilation_w,
Dtype* data_col, const QuantizerValues* const data_quant) {
const Dtype zero = data_quant ? data_quant->template get_zero<Dtype>()
: Dtype(0);
const int_tp output_h = (height + 2 * pad_h
- (dilation_h * (kernel_h - 1) + 1)) / stride_h + 1;
const int_tp output_w =
(width + 2 * pad_w - (dilation_w * (kernel_w - 1) + 1)) / stride_w + 1;
const int_tp channel_size = height * width;
for (int_tp channel = channels; channel--; data_im += channel_size) {
for (int_tp kernel_row = 0; kernel_row < kernel_h; kernel_row++) {
for (int_tp kernel_col = 0; kernel_col < kernel_w; kernel_col++) {
int_tp input_row = -pad_h + kernel_row * dilation_h;
for (int_tp output_rows = output_h; output_rows; output_rows--) {
if (!is_a_ge_zero_and_a_lt_b(input_row, height)) {
for (int_tp output_cols = output_w; output_cols; output_cols--) {
*(data_col++) = zero;
}
} else {
int_tp input_col = -pad_w + kernel_col * dilation_w;
for (int_tp output_col = output_w; output_col; output_col--) {
if (is_a_ge_zero_and_a_lt_b(input_col, width)) {
*(data_col++) = data_im[input_row * width + input_col];
} else {
*(data_col++) = zero;
}
input_col += stride_w;
}
}
input_row += stride_h;
}
}
}
}
}
void im2col_cpu(const Dtype* data_im, 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,
Dtype* data_col)
{
const int output_h = (height + 2 * pad_h -
(dilation_h * (kernel_h - 1) + 1)) / stride_h + 1;
const int output_w = (width + 2 * pad_w -
(dilation_w * (kernel_w - 1) + 1)) / stride_w + 1;
const int channel_size = height * width;
// 遍历一个kernel大小的图像
for (int channel = channels; channel--; data_im += channel_size)
{
// 下面三行是计算在kernel大小的图像上面的位置
// c_im h_offset w_offset
// 遍历卷积之后的图像的上面的每一个像素
for (int kernel_row = 0; kernel_row < kernel_h; kernel_row++)
{
for (int kernel_col = 0; kernel_col < kernel_w; kernel_col++)
{
// 计算卷积之后的图像与卷积之前的图像的位置
// 卷积之后的图像与卷积之前的图像像素所对应的位置
// 卷积之后的像素为h和w那么所对应的原图像的位置为 [h * stride_h - pad_h, h * stride_h - pad_h+kernel_h]以及
// [w * stride_w - pad_w, w * stride_w - pad_w+kernel_w]
int input_row = -pad_h + kernel_row * dilation_h;
for (int output_rows = output_h; output_rows; output_rows--)
{
// 如果符合input_row>=height运行循环里面的代码,然后去掉大于height的部分
if (!is_a_ge_zero_and_a_lt_b(input_row, height))
{
for (int output_cols = output_w; output_cols; output_cols--)
{
*(data_col++) = 0;
}
}
else
{
int input_col = -pad_w + kernel_col * dilation_w;
for (int output_col = output_w; output_col; output_col--)
{
// 如果符合input_col<width运行循环里面的代码,然后将这个位置的kernel的地址给了data_col,否则去掉大于width的部分
if (is_a_ge_zero_and_a_lt_b(input_col, width))
{
*(data_col++) = data_im[input_row * width + input_col];
}
else
{
*(data_col++) = 0;
}
input_col += stride_w;
}
}
input_row += stride_h;
}
}
}
}
}
void im2col_cpu(const Dtype* data_im, 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,
Dtype* data_col) {
#if 0
const int output_h = (height + 2 * pad_h -
(dilation_h * (kernel_h - 1) + 1)) / stride_h + 1;
const int output_w = (width + 2 * pad_w -
(dilation_w * (kernel_w - 1) + 1)) / stride_w + 1;
const int channel_size = height * width;
for (int channel = channels; channel--; data_im += channel_size) {
for (int kernel_row = 0; kernel_row < kernel_h; kernel_row++) {
for (int kernel_col = 0; kernel_col < kernel_w; kernel_col++) {
int input_row = -pad_h + kernel_row * dilation_h;
for (int output_rows = output_h; output_rows; output_rows--) {
if (!is_a_ge_zero_and_a_lt_b(input_row, height)) {
for (int output_cols = output_w; output_cols; output_cols--) {
*(data_col++) = 0;
}
} else {
int input_col = -pad_w + kernel_col * dilation_w;
for (int output_col = output_w; output_col; output_col--) {
if (is_a_ge_zero_and_a_lt_b(input_col, width)) {
*(data_col++) = data_im[input_row * width + input_col];
} else {
*(data_col++) = 0;
}
input_col += stride_w;
}
}
input_row += stride_h;
}
}
}
}
#else
int dil_kernel_h = (kernel_h - 1) * dilation_h + 1;
int dil_kernel_w = (kernel_w - 1) * dilation_w + 1;
int height_col = (height + 2 * pad_h - dil_kernel_h) / stride_h + 1;
int width_col = (width + 2 * pad_w - dil_kernel_w) / stride_w + 1;
int channels_col = channels * kernel_h * kernel_w;
#ifdef _OPENMP
#pragma omp parallel for
#endif
for (int c = 0; c < channels_col; ++c) {
int w_offset = c % kernel_w;
int h_offset = (c / kernel_w) % kernel_h;
int c_im = c / kernel_h / kernel_w;
const int hc0 = h_offset * dilation_h - pad_h;
const int wc0 = w_offset * dilation_w - pad_w;
for (int h = 0; h < height_col; ++h) {
int h_pad = h * stride_h + hc0;
const int row_offset = (c * height_col + h) * width_col;
const int srow_offset = (c_im * height + h_pad) * width;
for (int w = 0; w < width_col; ++w) {
int w_pad = w * stride_w + wc0;
if ((((unsigned)h_pad) < ((unsigned)height)) && (((unsigned)w_pad) < ((unsigned)width)))
data_col[row_offset + w] = data_im[srow_offset + w_pad];
else {
data_col[row_offset + w] = 0.;
}
}
}
}
#endif
}
void col2im_cpu(const Dtype* data_col, 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,
Dtype* data_im) {
#if 0
caffe_set(height * width * channels, Dtype(0), data_im);
const int output_h = (height + 2 * pad_h -
(dilation_h * (kernel_h - 1) + 1)) / stride_h + 1;
const int output_w = (width + 2 * pad_w -
(dilation_w * (kernel_w - 1) + 1)) / stride_w + 1;
const int channel_size = height * width;
for (int channel = channels; channel--; data_im += channel_size) {
for (int kernel_row = 0; kernel_row < kernel_h; kernel_row++) {
for (int kernel_col = 0; kernel_col < kernel_w; kernel_col++) {
int input_row = -pad_h + kernel_row * dilation_h;
for (int output_rows = output_h; output_rows; output_rows--) {
if (!is_a_ge_zero_and_a_lt_b(input_row, height)) {
data_col += output_w;
} else {
int input_col = -pad_w + kernel_col * dilation_w;
for (int output_col = output_w; output_col; output_col--) {
if (is_a_ge_zero_and_a_lt_b(input_col, width)) {
data_im[input_row * width + input_col] += *data_col;
}
data_col++;
input_col += stride_w;
}
}
input_row += stride_h;
}
}
}
}
#else
int dil_patch_h = (kernel_h - 1) * dilation_h + 1;
int dil_patch_w = (kernel_w - 1) * dilation_w + 1;
int height_col = (height + 2 * pad_h - dil_patch_h) / stride_h + 1;
int width_col = (width + 2 * pad_w - dil_patch_w) / stride_w + 1;
long chunk_len = kernel_h * kernel_w;
caffe_set(height * width * channels, Dtype(0), data_im);
#ifdef _OPENMP
#pragma omp parallel for if (channels > 1)
#endif
for (int idx = 0; idx < channels; ++idx) {
for (int inner_idx = 0; inner_idx < chunk_len; ++inner_idx) {
int c = idx * chunk_len + inner_idx;
int w_offset = c % kernel_w;
int h_offset = (c / kernel_w) % kernel_h;
int c_im = c / kernel_h / kernel_w;
const int hc0 = h_offset * dilation_h - pad_h;
const int wc0 = w_offset * dilation_w - pad_w;
for (int h = 0; h < height_col; ++h) {
for (int w = 0; w < width_col; ++w) {
int h_pad = h * stride_h + hc0;
const int srow_offset = (c_im * height + h_pad) * width;
const int row_offset = (c * height_col + h) * width_col;
int w_pad = w * stride_w + wc0;
if ((((unsigned)h_pad) < ((unsigned)height)) && (((unsigned)w_pad) < ((unsigned)width))) {
data_im[srow_offset + w_pad] += data_col[row_offset + w];
}
}
}
}
}
#endif
}
