using caffe's im2col, it's so much better\!

Makefile

@@ -8,7 +8,7 @@ OBJDIR=./obj/
 
 CC=gcc
 NVCC=nvcc
-OPTS=-O0
+OPTS=-O3
 LDFLAGS=`pkg-config --libs opencv` -lm -pthread -lstdc++
 COMMON=`pkg-config --cflags opencv` -I/usr/local/cuda/include/
 CFLAGS=-Wall -Wfatal-errors

src/col2im_kernels.cu

@@ -3,60 +3,112 @@ extern "C" {
 #include "cuda.h"
 }
 
-__global__ void col2im_kernel(float *data_col,
-        int channels, int height, int width,
-        int ksize, int stride, int pad, float *data_im)
-{
+// src: https://github.com/BVLC/caffe/blob/master/src/caffe/util/im2col.cu
+// You may also want to read: https://github.com/BVLC/caffe/blob/master/LICENSE
 
-    int height_col = (height - ksize) / stride + 1;
-    int width_col = (width - ksize) / stride + 1;
-    if (pad){
-        height_col = 1 + (height-1) / stride;
-        width_col = 1 + (width-1) / stride;
-        pad = ksize/2;
-    }
-
-    int id = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
-    if(id >= channels*height*width) return;
-
-    int index = id;
-    int w = id%width + pad;
-    id /= width;
-    int h = id%height + pad;
-    id /= height;
-    int c = id%channels;
-
-    int w_start = (w-ksize+stride)/stride;
-    int w_end = w/stride + 1;
-
-    int h_start = (h-ksize+stride)/stride;
-    int h_end = h/stride + 1;
-
-    // int rows = channels * ksize * ksize;
-    // int cols = height_col*width_col;
-    int col_offset = (c*ksize*ksize + h * ksize + w)*height_col*width_col;
-    int h_coeff = (1-stride*ksize*height_col)*width_col;
-    int w_coeff = 1-stride*height_col*width_col;
-    float val = 0;
-    int h_col, w_col;
-    for(h_col = h_start; h_col < h_end; ++h_col){
-        for(w_col = w_start; w_col < w_end; ++w_col){
-            int col_index = col_offset +h_col*h_coeff + w_col*w_coeff;
-            float part = (w_col < 0 || h_col < 0 || h_col >= height_col || w_col >= width_col) ? 0 : data_col[col_index];
-            val += part;
+__global__ void col2im_gpu_kernel(const int n, const float* data_col,
+        const int height, const int width, const int ksize,
+        const int pad,
+        const int stride,
+        const int height_col, const int width_col,
+        float *data_im) {
+    int index = blockIdx.x*blockDim.x+threadIdx.x;
+    for(; index < n; index += blockDim.x*gridDim.x){
+        float val = 0;
+        int w = index % width + pad;
+        int h = (index / width) % height + pad;
+        int c = index / (width * height);
+        // compute the start and end of the output
+        int w_col_start = (w < ksize) ? 0 : (w - ksize) / stride + 1;
+        int w_col_end = min(w / stride + 1, width_col);
+        int h_col_start = (h < ksize) ? 0 : (h - ksize) / stride + 1;
+        int h_col_end = min(h / stride + 1, height_col);
+        // equivalent implementation
+        int offset =
+            (c * ksize * ksize + h * ksize + w) * height_col * width_col;
+        int coeff_h_col = (1 - stride * ksize * height_col) * width_col;
+        int coeff_w_col = (1 - stride * 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;
     }
-    data_im[index] = val;
+}
+
+void col2im_ongpu(float *im,
+        int channels, int height, int width,
+        int ksize, int stride, int pad, float *data_col){
+    // We are going to launch channels * height_col * width_col kernels, each
+    // kernel responsible for copying a single-channel grid.
+    pad = pad ? ksize/2 : 0;
+    int height_col = (height + 2 * pad - ksize) / stride + 1;
+    int width_col = (width + 2 * pad - ksize) / stride + 1;
+    int num_kernels = channels * height * width;
+    col2im_gpu_kernel<<<(num_kernels+BLOCK-1)/BLOCK,
+        BLOCK>>>(
+                num_kernels, data_col, height, width, ksize, pad,
+                stride, height_col,
+                width_col, im);
+}
+
+/*
+   __global__ void col2im_kernel(float *data_col,
+   int channels, int height, int width,
+   int ksize, int stride, int pad, float *data_im)
+   {
+
+   int height_col = (height - ksize) / stride + 1;
+   int width_col = (width - ksize) / stride + 1;
+   if (pad){
+   height_col = 1 + (height-1) / stride;
+   width_col = 1 + (width-1) / stride;
+   pad = ksize/2;
+   }
+
+   int id = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
+   if(id >= channels*height*width) return;
+
+   int index = id;
+   int w = id%width + pad;
+   id /= width;
+   int h = id%height + pad;
+   id /= height;
+   int c = id%channels;
+
+   int w_start = (w-ksize+stride)/stride;
+   int w_end = w/stride + 1;
+
+   int h_start = (h-ksize+stride)/stride;
+   int h_end = h/stride + 1;
+
+// int rows = channels * ksize * ksize;
+// int cols = height_col*width_col;
+int col_offset = (c*ksize*ksize + h * ksize + w)*height_col*width_col;
+int h_coeff = (1-stride*ksize*height_col)*width_col;
+int w_coeff = 1-stride*height_col*width_col;
+float val = 0;
+int h_col, w_col;
+for(h_col = h_start; h_col < h_end; ++h_col){
+for(w_col = w_start; w_col < w_end; ++w_col){
+int col_index = col_offset +h_col*h_coeff + w_col*w_coeff;
+float part = (w_col < 0 || h_col < 0 || h_col >= height_col || w_col >= width_col) ? 0 : data_col[col_index];
+val += part;
+}
+}
+data_im[index] = val;
 }
 
 
 extern "C" void col2im_ongpu(float *data_col,
-        int channels,  int height,  int width,
-        int ksize,  int stride,  int pad, float *data_im)
+int channels,  int height,  int width,
+int ksize,  int stride,  int pad, float *data_im)
 {
 
-    size_t n = channels*height*width;
+size_t n = channels*height*width;
 
-    col2im_kernel<<<cuda_gridsize(n), BLOCK>>>(data_col, channels, height, width, ksize, stride, pad, data_im);
-    check_error(cudaPeekAtLastError());
+col2im_kernel<<<cuda_gridsize(n), BLOCK>>>(data_col, channels, height, width, ksize, stride, pad, data_im);
+check_error(cudaPeekAtLastError());
 }
+ */

src/convolutional_kernels.cu

@@ -56,7 +56,7 @@ extern "C" void backward_bias_gpu(float *bias_updates, float *delta, int batch,
 
 extern "C" void forward_convolutional_layer_gpu(convolutional_layer layer, network_state state)
 {
-clock_t time = clock();
+//clock_t time = clock();
     int i;
     int m = layer.n;
     int k = layer.size*layer.size*layer.c;
@@ -64,31 +64,31 @@ clock_t time = clock();
         convolutional_out_width(layer);
 
     bias_output_gpu(layer.output_gpu, layer.biases_gpu, layer.batch, layer.n, n);
-cudaDeviceSynchronize();
-printf("bias %f\n", sec(clock() - time));
-time = clock();
+//cudaDeviceSynchronize();
+//printf("bias %f\n", sec(clock() - time));
+//time = clock();
 
-float imt=0;
-float gemt = 0;
+//float imt=0;
+//float gemt = 0;
     for(i = 0; i < layer.batch; ++i){
-time = clock();
+//time = clock();
         im2col_ongpu(state.input + i*layer.c*layer.h*layer.w, layer.c,  layer.h,  layer.w,  layer.size,  layer.stride, layer.pad, layer.col_image_gpu);
-cudaDeviceSynchronize();
-imt += sec(clock()-time);
-time = clock();
+//cudaDeviceSynchronize();
+//imt += sec(clock()-time);
+//time = clock();
         float * a = layer.filters_gpu;
         float * b = layer.col_image_gpu;
         float * c = layer.output_gpu;
         gemm_ongpu(0,0,m,n,k,1.,a,k,b,n,1.,c+i*m*n,n);
-cudaDeviceSynchronize();
-gemt += sec(clock()-time);
-time = clock();
+//cudaDeviceSynchronize();
+//gemt += sec(clock()-time);
+//time = clock();
     }
     activate_array_ongpu(layer.output_gpu, m*n*layer.batch, layer.activation);
-cudaDeviceSynchronize();
-printf("activate %f\n", sec(clock() - time));
-printf("im2col %f\n", imt);
-printf("gemm %f\n", gemt);
+//cudaDeviceSynchronize();
+//printf("activate %f\n", sec(clock() - time));
+//printf("im2col %f\n", imt);
+//printf("gemm %f\n", gemt);
 }
 
 extern "C" void backward_convolutional_layer_gpu(convolutional_layer layer, network_state state)

src/im2col_kernels.cu

@@ -3,77 +3,127 @@ extern "C" {
 #include "cuda.h"
 }
 
-__global__ void im2col_pad_kernel(float *im,
-     int channels,  int height,  int width,
-     int ksize,  int stride, float *data_col)
-{
-    int c,h,w;
-    int height_col = 1 + (height-1) / stride;
-    int width_col = 1 + (width-1) / stride;
-    int channels_col = channels * ksize * ksize;
-
-    int pad = ksize/2;
-
-    int id = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
-    int col_size = height_col*width_col*channels_col;
-    if (id >= col_size) return;
-
-    int col_index = id;
-    w = id % width_col;
-    id /= width_col;
-    h = id % height_col;
-    id /= height_col;
-    c = id % channels_col;
-    id /= channels_col;
-
-    int w_offset = c % ksize;
-    int h_offset = (c / ksize) % ksize;
-    int im_channel = c / ksize / ksize;
-    int im_row = h_offset + h * stride - pad;
-    int im_col = w_offset + w * stride - pad;
-
-    int im_index = im_col + width*(im_row + height*im_channel);
-    float val = (im_row < 0 || im_col < 0 || im_row >= height || im_col >= width) ? 0 : im[im_index];
-
-    data_col[col_index] = val;
+// src: https://github.com/BVLC/caffe/blob/master/src/caffe/util/im2col.cu
+// You may also want to read: https://github.com/BVLC/caffe/blob/master/LICENSE
+
+__global__ void im2col_gpu_kernel(const int n, const float* data_im,
+        const int height, const int width, const int ksize,
+        const int pad,
+        const int stride,
+        const int height_col, const int width_col,
+        float *data_col) {
+    int index = blockIdx.x*blockDim.x+threadIdx.x;
+    for(; index < n; index += blockDim.x*gridDim.x){
+        int w_out = index % width_col;
+        int h_index = index / width_col;
+        int h_out = h_index % height_col;
+        int channel_in = h_index / height_col;
+        int channel_out = channel_in * ksize * ksize;
+        int h_in = h_out * stride - pad;
+        int w_in = w_out * stride - pad;
+        float* data_col_ptr = data_col;
+        data_col_ptr += (channel_out * height_col + h_out) * width_col + w_out;
+        const float* data_im_ptr = data_im;
+        data_im_ptr += (channel_in * height + h_in) * width + w_in;
+        for (int i = 0; i < ksize; ++i) {
+            for (int j = 0; j < ksize; ++j) {
+                int h = h_in + i;
+                int w = w_in + j;
+                *data_col_ptr = (h >= 0 && w >= 0 && h < height && w < width) ?
+                    data_im_ptr[i * width + j] : 0;
+                data_col_ptr += height_col * width_col;
+            }
+        }
+    }
 }
 
-__global__ void im2col_nopad_kernel(float *im,
-        int channels,  int height,  int width,
-        int ksize,  int stride, float *data_col)
-{
-    int c,h,w;
-    int height_col = (height - ksize) / stride + 1;
-    int width_col = (width - ksize) / stride + 1;
-    int channels_col = channels * ksize * ksize;
-
-    int id = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
-    int col_size = height_col*width_col*channels_col;
-    if (id >= col_size) return;
-
-    int col_index = id;
-    w = id % width_col;
-    id /= width_col;
-    h = id % height_col;
-    id /= height_col;
-    c = id % channels_col;
-    id /= channels_col;
-
-    int w_offset = c % ksize;
-    int h_offset = (c / ksize) % ksize;
-    int im_channel = c / ksize / ksize;
-    int im_row = h_offset + h * stride;
-    int im_col = w_offset + w * stride;
-
-    int im_index = im_col + width*(im_row + height*im_channel);
-    float val = (im_row < 0 || im_col < 0 || im_row >= height || im_col >= width) ? 0 : im[im_index];
-
-    data_col[col_index] = val;
+void im2col_ongpu(float *im,
+         int channels, int height, int width,
+         int ksize, int stride, int pad, float *data_col){
+    // We are going to launch channels * height_col * width_col kernels, each
+    // kernel responsible for copying a single-channel grid.
+    pad = pad ? ksize/2 : 0;
+    int height_col = (height + 2 * pad - ksize) / stride + 1;
+    int width_col = (width + 2 * pad - ksize) / stride + 1;
+    int num_kernels = channels * height_col * width_col;
+    im2col_gpu_kernel<<<(num_kernels+BLOCK-1)/BLOCK,
+        BLOCK>>>(
+                num_kernels, im, height, width, ksize, pad,
+                stride, height_col,
+                width_col, data_col);
 }
-
-extern "C" void im2col_ongpu(float *im,
-        int channels,  int height,  int width,
-        int ksize,  int stride,  int pad, float *data_col)
+/*
+   __global__ void im2col_pad_kernel(float *im,
+   int channels,  int height,  int width,
+   int ksize,  int stride, float *data_col)
+   {
+   int c,h,w;
+   int height_col = 1 + (height-1) / stride;
+   int width_col = 1 + (width-1) / stride;
+   int channels_col = channels * ksize * ksize;
+
+   int pad = ksize/2;
+
+   int id = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
+   int col_size = height_col*width_col*channels_col;
+   if (id >= col_size) return;
+
+   int col_index = id;
+   w = id % width_col;
+   id /= width_col;
+   h = id % height_col;
+   id /= height_col;
+   c = id % channels_col;
+   id /= channels_col;
+
+   int w_offset = c % ksize;
+   int h_offset = (c / ksize) % ksize;
+   int im_channel = c / ksize / ksize;
+   int im_row = h_offset + h * stride - pad;
+   int im_col = w_offset + w * stride - pad;
+
+   int im_index = im_col + width*(im_row + height*im_channel);
+   float val = (im_row < 0 || im_col < 0 || im_row >= height || im_col >= width) ? 0 : im[im_index];
+
+   data_col[col_index] = val;
+   }
+
+   __global__ void im2col_nopad_kernel(float *im,
+   int channels,  int height,  int width,
+   int ksize,  int stride, float *data_col)
+   {
+   int c,h,w;
+   int height_col = (height - ksize) / stride + 1;
+   int width_col = (width - ksize) / stride + 1;
+   int channels_col = channels * ksize * ksize;
+
+   int id = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
+   int col_size = height_col*width_col*channels_col;
+   if (id >= col_size) return;
+
+   int col_index = id;
+   w = id % width_col;
+   id /= width_col;
+   h = id % height_col;
+   id /= height_col;
+   c = id % channels_col;
+   id /= channels_col;
+
+   int w_offset = c % ksize;
+   int h_offset = (c / ksize) % ksize;
+   int im_channel = c / ksize / ksize;
+   int im_row = h_offset + h * stride;
+   int im_col = w_offset + w * stride;
+
+   int im_index = im_col + width*(im_row + height*im_channel);
+   float val = (im_row < 0 || im_col < 0 || im_row >= height || im_col >= width) ? 0 : im[im_index];
+
+   data_col[col_index] = val;
+   }
+
+   extern "C" void im2col_ongpu(float *im,
+   int channels,  int height,  int width,
+int ksize,  int stride,  int pad, float *data_col)
 {
 
     int height_col = (height - ksize) / stride + 1;
@@ -91,3 +141,4 @@ extern "C" void im2col_ongpu(float *im,
     else im2col_nopad_kernel<<<cuda_gridsize(n),BLOCK>>>(im,  channels, height, width, ksize, stride, data_col);
     check_error(cudaPeekAtLastError());
 }
+*/

src/imagenet.c

@@ -13,7 +13,7 @@ void train_imagenet(char *cfgfile, char *weightfile)
         load_weights(&net, weightfile);
     }
     printf("Learning Rate: %g, Momentum: %g, Decay: %g\n", net.learning_rate, net.momentum, net.decay);
-    int imgs = 128;
+    int imgs = 1024;
     int i = net.seen/imgs;
     char **labels = get_labels("/home/pjreddie/data/imagenet/cls.labels.list");
     list *plist = get_paths("/data/imagenet/cls.train.list");