From ff67f0347653c35c67ddbafad8dc76bbd868047e Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Wed, 03 Dec 2014 16:48:07 +0000
Subject: [PATCH] Starting on server/client
---
src/maxpool_layer.c | 204 +++++++++++++++++++++++++++++---------------------
1 files changed, 118 insertions(+), 86 deletions(-)
diff --git a/src/maxpool_layer.c b/src/maxpool_layer.c
index 08c9f2f..df19040 100644
--- a/src/maxpool_layer.c
+++ b/src/maxpool_layer.c
@@ -27,8 +27,15 @@
layer->c = c;
layer->size = size;
layer->stride = stride;
- layer->output = calloc(((h-1)/stride+1) * ((w-1)/stride+1) * c*batch, sizeof(float));
- layer->delta = calloc(((h-1)/stride+1) * ((w-1)/stride+1) * c*batch, sizeof(float));
+ int output_size = ((h-1)/stride+1) * ((w-1)/stride+1) * c * batch;
+ layer->indexes = calloc(output_size, sizeof(int));
+ layer->output = calloc(output_size, sizeof(float));
+ layer->delta = calloc(output_size, sizeof(float));
+ #ifdef GPU
+ layer->indexes_cl = cl_make_int_array(layer->indexes, output_size);
+ layer->output_cl = cl_make_array(layer->output, output_size);
+ layer->delta_cl = cl_make_array(layer->delta, output_size);
+ #endif
return layer;
}
@@ -41,101 +48,126 @@
layer->delta = realloc(layer->delta, ((h-1)/layer->stride+1) * ((w-1)/layer->stride+1) * c * layer->batch*sizeof(float));
}
-float get_max_region(image im, int h, int w, int c, int size)
+void forward_maxpool_layer(const maxpool_layer layer, float *input)
{
- int i,j;
- int lower = (-size-1)/2 + 1;
- int upper = size/2 + 1;
-
- int lh = (h-lower < 0) ? 0 : h-lower;
- int uh = (h+upper > im.h) ? im.h : h+upper;
+ int b,i,j,k,l,m;
+ int w_offset = (-layer.size-1)/2 + 1;
+ int h_offset = (-layer.size-1)/2 + 1;
- int lw = (w-lower < 0) ? 0 : w-lower;
- int uw = (w+upper > im.w) ? im.w : w+upper;
-
- //printf("%d\n", -3/2);
- //printf("%d %d\n", lower, upper);
- //printf("%d %d %d %d\n", lh, uh, lw, uw);
-
- float max = -FLT_MAX;
- for(i = lh; i < uh; ++i){
- for(j = lw; j < uw; ++j){
- float val = get_pixel(im, i, j, c);
- if (val > max) max = val;
- }
- }
- return max;
-}
+ int h = (layer.h-1)/layer.stride + 1;
+ int w = (layer.w-1)/layer.stride + 1;
+ int c = layer.c;
-void forward_maxpool_layer(const maxpool_layer layer, float *in)
-{
- int b;
for(b = 0; b < layer.batch; ++b){
- image input = float_to_image(layer.h, layer.w, layer.c, in+b*layer.h*layer.w*layer.c);
-
- int h = (layer.h-1)/layer.stride + 1;
- int w = (layer.w-1)/layer.stride + 1;
- int c = layer.c;
- image output = float_to_image(h,w,c,layer.output+b*h*w*c);
-
- int i,j,k;
- for(k = 0; k < input.c; ++k){
- for(i = 0; i < input.h; i += layer.stride){
- for(j = 0; j < input.w; j += layer.stride){
- float max = get_max_region(input, i, j, k, layer.size);
- set_pixel(output, i/layer.stride, j/layer.stride, k, max);
+ for(k = 0; k < c; ++k){
+ for(i = 0; i < h; ++i){
+ for(j = 0; j < w; ++j){
+ int out_index = j + w*(i + h*(k + c*b));
+ float max = -FLT_MAX;
+ int max_i = -1;
+ for(l = 0; l < layer.size; ++l){
+ for(m = 0; m < layer.size; ++m){
+ int cur_h = h_offset + i*layer.stride + l;
+ int cur_w = w_offset + j*layer.stride + m;
+ int index = cur_w + layer.w*(cur_h + layer.h*(k + b*layer.c));
+ int valid = (cur_h >= 0 && cur_h < layer.h &&
+ cur_w >= 0 && cur_w < layer.w);
+ float val = (valid != 0) ? input[index] : -FLT_MAX;
+ max_i = (val > max) ? index : max_i;
+ max = (val > max) ? val : max;
+ }
+ }
+ layer.output[out_index] = max;
+ layer.indexes[out_index] = max_i;
}
}
}
}
}
-float set_max_region_delta(image im, image delta, int h, int w, int c, int size, float max, float error)
+void backward_maxpool_layer(const maxpool_layer layer, float *delta)
{
- int i,j;
- int lower = (-size-1)/2 + 1;
- int upper = size/2 + 1;
-
- int lh = (h-lower < 0) ? 0 : h-lower;
- int uh = (h+upper > im.h) ? im.h : h+upper;
-
- int lw = (w-lower < 0) ? 0 : w-lower;
- int uw = (w+upper > im.w) ? im.w : w+upper;
-
- for(i = lh; i < uh; ++i){
- for(j = lw; j < uw; ++j){
- float val = get_pixel(im, i, j, c);
- if (val == max){
- add_pixel(delta, i, j, c, error);
- }
- }
- }
- return max;
-}
-
-void backward_maxpool_layer(const maxpool_layer layer, float *in, float *delta)
-{
- int b;
- for(b = 0; b < layer.batch; ++b){
- image input = float_to_image(layer.h, layer.w, layer.c, in+b*layer.h*layer.w*layer.c);
- image input_delta = float_to_image(layer.h, layer.w, layer.c, delta+b*layer.h*layer.w*layer.c);
- int h = (layer.h-1)/layer.stride + 1;
- int w = (layer.w-1)/layer.stride + 1;
- int c = layer.c;
- image output = float_to_image(h,w,c,layer.output+b*h*w*c);
- image output_delta = float_to_image(h,w,c,layer.delta+b*h*w*c);
- zero_image(input_delta);
-
- int i,j,k;
- for(k = 0; k < input.c; ++k){
- for(i = 0; i < input.h; i += layer.stride){
- for(j = 0; j < input.w; j += layer.stride){
- float max = get_pixel(output, i/layer.stride, j/layer.stride, k);
- float error = get_pixel(output_delta, i/layer.stride, j/layer.stride, k);
- set_max_region_delta(input, input_delta, i, j, k, layer.size, max, error);
- }
- }
- }
+ int i;
+ int h = (layer.h-1)/layer.stride + 1;
+ int w = (layer.w-1)/layer.stride + 1;
+ int c = layer.c;
+ memset(delta, 0, layer.batch*layer.h*layer.w*layer.c*sizeof(float));
+ for(i = 0; i < h*w*c*layer.batch; ++i){
+ int index = layer.indexes[i];
+ delta[index] += layer.delta[i];
}
}
+#ifdef GPU
+cl_kernel get_forward_kernel()
+{
+ static int init = 0;
+ static cl_kernel kernel;
+ if(!init){
+ kernel = get_kernel("src/maxpool_layer.cl", "forward", 0);
+ init = 1;
+ }
+ return kernel;
+}
+
+void forward_maxpool_layer_gpu(maxpool_layer layer, cl_mem input)
+{
+ int h = (layer.h-1)/layer.stride + 1;
+ int w = (layer.w-1)/layer.stride + 1;
+ int c = layer.c;
+ cl_setup();
+ cl_kernel kernel = get_forward_kernel();
+ cl_command_queue queue = cl.queue;
+
+ cl_uint i = 0;
+ cl.error = clSetKernelArg(kernel, i++, sizeof(layer.h), (void*) &layer.h);
+ cl.error = clSetKernelArg(kernel, i++, sizeof(layer.w), (void*) &layer.w);
+ cl.error = clSetKernelArg(kernel, i++, sizeof(layer.c), (void*) &layer.c);
+ cl.error = clSetKernelArg(kernel, i++, sizeof(layer.stride), (void*) &layer.stride);
+ cl.error = clSetKernelArg(kernel, i++, sizeof(layer.size), (void*) &layer.size);
+ cl.error = clSetKernelArg(kernel, i++, sizeof(input), (void*) &input);
+ cl.error = clSetKernelArg(kernel, i++, sizeof(layer.output_cl), (void*) &layer.output_cl);
+ cl.error = clSetKernelArg(kernel, i++, sizeof(layer.indexes_cl), (void*) &layer.indexes_cl);
+ check_error(cl);
+
+ const size_t global_size[] = {h*w*c*layer.batch};
+
+ cl.error = clEnqueueNDRangeKernel(queue, kernel, 1, 0, global_size, 0, 0, 0, 0);
+ check_error(cl);
+}
+
+cl_kernel get_backward_kernel()
+{
+ static int init = 0;
+ static cl_kernel kernel;
+ if(!init){
+ kernel = get_kernel("src/maxpool_layer.cl", "backward", 0);
+ init = 1;
+ }
+ return kernel;
+}
+
+void backward_maxpool_layer_gpu(maxpool_layer layer, cl_mem delta)
+{
+ cl_setup();
+ cl_kernel kernel = get_backward_kernel();
+ cl_command_queue queue = cl.queue;
+
+ cl_uint i = 0;
+ cl.error = clSetKernelArg(kernel, i++, sizeof(layer.h), (void*) &layer.h);
+ cl.error = clSetKernelArg(kernel, i++, sizeof(layer.w), (void*) &layer.w);
+ cl.error = clSetKernelArg(kernel, i++, sizeof(layer.c), (void*) &layer.c);
+ cl.error = clSetKernelArg(kernel, i++, sizeof(layer.stride), (void*) &layer.stride);
+ cl.error = clSetKernelArg(kernel, i++, sizeof(layer.size), (void*) &layer.size);
+ cl.error = clSetKernelArg(kernel, i++, sizeof(layer.delta_cl), (void*) &layer.delta_cl);
+ cl.error = clSetKernelArg(kernel, i++, sizeof(delta), (void*) &delta);
+ cl.error = clSetKernelArg(kernel, i++, sizeof(layer.indexes_cl), (void*) &layer.indexes_cl);
+ check_error(cl);
+
+ const size_t global_size[] = {layer.h*layer.w*layer.c*layer.batch};
+
+ cl.error = clEnqueueNDRangeKernel(queue, kernel, 1, 0, global_size, 0, 0, 0, 0);
+ check_error(cl);
+}
+
+#endif
--
Gitblit v1.10.0