From e92f7d301c971b4d27aa3dcd1e4047e94f04b3fc Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Wed, 25 Mar 2015 01:27:12 +0000
Subject: [PATCH] smaller gridsize in bias
---
src/maxpool_layer.c | 105 +++++++++++-----------------------------------------
1 files changed, 22 insertions(+), 83 deletions(-)
diff --git a/src/maxpool_layer.c b/src/maxpool_layer.c
index c05e939..790cb28 100644
--- a/src/maxpool_layer.c
+++ b/src/maxpool_layer.c
@@ -1,4 +1,5 @@
#include "maxpool_layer.h"
+#include "cuda.h"
#include <stdio.h>
image get_maxpool_image(maxpool_layer layer)
@@ -32,23 +33,32 @@
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);
+ layer->indexes_gpu = cuda_make_int_array(output_size);
+ layer->output_gpu = cuda_make_array(layer->output, output_size);
+ layer->delta_gpu = cuda_make_array(layer->delta, output_size);
#endif
return layer;
}
-void resize_maxpool_layer(maxpool_layer *layer, int h, int w, int c)
+void resize_maxpool_layer(maxpool_layer *layer, int h, int w)
{
layer->h = h;
layer->w = w;
- layer->c = c;
- layer->output = realloc(layer->output, ((h-1)/layer->stride+1) * ((w-1)/layer->stride+1) * c * layer->batch* sizeof(float));
- layer->delta = realloc(layer->delta, ((h-1)/layer->stride+1) * ((w-1)/layer->stride+1) * c * layer->batch*sizeof(float));
+ int output_size = ((h-1)/layer->stride+1) * ((w-1)/layer->stride+1) * layer->c * layer->batch;
+ layer->output = realloc(layer->output, output_size * sizeof(float));
+ layer->delta = realloc(layer->delta, output_size * sizeof(float));
+
+ #ifdef GPU
+ cuda_free((float *)layer->indexes_gpu);
+ cuda_free(layer->output_gpu);
+ cuda_free(layer->delta_gpu);
+ layer->indexes_gpu = cuda_make_int_array(output_size);
+ layer->output_gpu = cuda_make_array(layer->output, output_size);
+ layer->delta_gpu = cuda_make_array(layer->delta, output_size);
+ #endif
}
-void forward_maxpool_layer(const maxpool_layer layer, float *input)
+void forward_maxpool_layer(const maxpool_layer layer, network_state state)
{
int b,i,j,k,l,m;
int w_offset = (-layer.size-1)/2 + 1;
@@ -72,7 +82,7 @@
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;
+ float val = (valid != 0) ? state.input[index] : -FLT_MAX;
max_i = (val > max) ? index : max_i;
max = (val > max) ? val : max;
}
@@ -85,87 +95,16 @@
}
}
-void backward_maxpool_layer(const maxpool_layer layer, float *delta)
+void backward_maxpool_layer(const maxpool_layer layer, network_state state)
{
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));
+ memset(state.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];
+ state.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_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_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