| | |
| | | #include "mini_blas.h" |
| | | #include <stdio.h> |
| | | |
| | | inline float im2col_get_pixel(float *im, int height, int width, int channels, |
| | | int row, int col, int channel, int pad) |
| | | int b, int row, int col, int channel, int pad) |
| | | { |
| | | row -= pad; |
| | | col -= pad; |
| | | |
| | | if (row < 0 || col < 0 || |
| | | row >= height || col >= width) return 0; |
| | | return im[col + width*(row + channel*height)]; |
| | | return im[col + width*(row + height*(channel+b*channels))]; |
| | | } |
| | | |
| | | //From Berkeley Vision's Caffe! |
| | | //https://github.com/BVLC/caffe/blob/master/LICENSE |
| | | void im2col_cpu(float* data_im, |
| | | const int batch, const int channels, const int height, const int width, |
| | | const int ksize, const int stride, int pad, float* data_col) |
| | | void im2col_cpu(float* data_im, int batch, |
| | | int channels, int height, int width, |
| | | int ksize, int stride, int pad, float* data_col) |
| | | { |
| | | int c,h,w,b; |
| | | int height_col = (height - ksize) / stride + 1; |
| | |
| | | pad = ksize/2; |
| | | } |
| | | int channels_col = channels * ksize * ksize; |
| | | int im_size = height*width*channels; |
| | | //int col_size = height_col*width_col*channels_col; |
| | | int col_size = height_col*width_col*channels_col; |
| | | for (b = 0; b < batch; ++b) { |
| | | for (c = 0; c < channels_col; ++c) { |
| | | int w_offset = c % ksize; |
| | |
| | | for (w = 0; w < width_col; ++w) { |
| | | int im_row = h_offset + h * stride; |
| | | int im_col = w_offset + w * stride; |
| | | int col_index = (c * height_col + h) * width_col + w + (batch-1) * c * height_col*width_col; |
| | | int col_index = (c * height_col + h) * width_col + w + b*col_size; |
| | | data_col[col_index] = im2col_get_pixel(data_im, height, width, channels, |
| | | im_row, im_col, c_im, pad); |
| | | b, im_row, im_col, c_im, pad); |
| | | } |
| | | } |
| | | } |
| | | data_im += im_size; |
| | | data_col+= channels_col; |
| | | } |
| | | } |
| | | |
| | |
| | | } |
| | | |
| | | |
| | | void im2col_ongpu(cl_mem data_im, const int batch, |
| | | const int channels, const int height, const int width, |
| | | const int ksize, const int stride, cl_mem data_col) |
| | | void im2col_ongpu(cl_mem data_im, int batch, |
| | | int channels, int height, int width, |
| | | int ksize, int stride, int pad, cl_mem data_col) |
| | | { |
| | | cl_setup(); |
| | | cl_kernel im2col_kernel = get_im2col_kernel(); |
| | |
| | | cl.error = clSetKernelArg(im2col_kernel, i++, sizeof(width), (void*) &width); |
| | | cl.error = clSetKernelArg(im2col_kernel, i++, sizeof(ksize), (void*) &ksize); |
| | | cl.error = clSetKernelArg(im2col_kernel, i++, sizeof(stride), (void*) &stride); |
| | | cl.error = clSetKernelArg(im2col_kernel, i++, sizeof(pad), (void*) &pad); |
| | | cl.error = clSetKernelArg(im2col_kernel, i++, sizeof(data_col), (void*) &data_col); |
| | | check_error(cl); |
| | | |
| | | int height_col = (height - ksize) / stride + 1; |
| | | int width_col = (width - ksize) / stride + 1; |
| | | int channels_col = channels * ksize * ksize; |
| | | if (pad){ |
| | | height_col = 1 + (height-1) / stride; |
| | | width_col = 1 + (width-1) / stride; |
| | | } |
| | | |
| | | size_t global_size[2]; |
| | | size_t local_size[2]; |
| | | global_size[0] = batch; |
| | | global_size[1] = channels_col; |
| | | local_size[0] = height_col; |
| | | local_size[1] = width_col; |
| | | size_t global_size = batch*channels_col*height_col*width_col; |
| | | |
| | | clEnqueueNDRangeKernel(queue, im2col_kernel, 2, 0, |
| | | global_size, local_size, 0, 0, 0); |
| | | clEnqueueNDRangeKernel(queue, im2col_kernel, 1, 0, |
| | | &global_size, 0, 0, 0, 0); |
| | | check_error(cl); |
| | | } |
| | | |
| | | void im2col_gpu(float *data_im, |
| | | const int batch, const int channels, const int height, const int width, |
| | | const int ksize, const int stride, |
| | | float *data_col) |
| | | void im2col_gpu(float *data_im, int batch, |
| | | int channels, int height, int width, |
| | | int ksize, int stride, int pad, float *data_col) |
| | | { |
| | | cl_setup(); |
| | | cl_context context = cl.context; |
| | |
| | | check_error(cl); |
| | | |
| | | im2col_ongpu(im_gpu, batch, channels, height, width, |
| | | ksize, stride, col_gpu); |
| | | ksize, stride, pad, col_gpu); |
| | | |
| | | clEnqueueReadBuffer(queue, col_gpu, CL_TRUE, 0, size, data_col, 0, 0, 0); |
| | | check_error(cl); |