From 14303717dcddae43cdc55beb0685dae86f566fd8 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Sat, 25 Oct 2014 18:57:26 +0000
Subject: [PATCH] Fast, needs to be faster
---
src/convolutional_layer.c | 70 +++++++++++++++++++++++++++++------
1 files changed, 58 insertions(+), 12 deletions(-)
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index bdbfbfd..1587ae8 100644
--- a/src/convolutional_layer.c
+++ b/src/convolutional_layer.c
@@ -2,6 +2,7 @@
#include "utils.h"
#include "mini_blas.h"
#include <stdio.h>
+#include <time.h>
int convolutional_out_height(convolutional_layer layer)
{
@@ -195,13 +196,14 @@
b = layer.delta;
c = layer.col_image;
- memset(delta, 0, layer.batch*layer.h*layer.w*layer.c*sizeof(float));
-
for(i = 0; i < layer.batch; ++i){
gemm(1,0,m,n,k,1,a,m,b,n,0,c,n);
b += k*n;
c += m*n;
}
+
+ memset(delta, 0, layer.batch*layer.h*layer.w*layer.c*sizeof(float));
+
col2im_cpu(layer.col_image, layer.batch, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, delta);
}
}
@@ -210,7 +212,7 @@
{
int size = layer.size*layer.size*layer.c*layer.n;
axpy_cpu(layer.n, layer.learning_rate, layer.bias_updates, 1, layer.biases, 1);
- scal_cpu(layer.n,layer.momentum, layer.bias_updates, 1);
+ scal_cpu(layer.n, layer.momentum, layer.bias_updates, 1);
scal_cpu(size, 1.-layer.learning_rate*layer.decay, layer.filters, 1);
axpy_cpu(size, layer.learning_rate, layer.filter_updates, 1, layer.filters, 1);
@@ -340,6 +342,8 @@
check_error(cl);
}
+//#define TIMEIT
+
void forward_convolutional_layer_gpu(convolutional_layer layer, cl_mem in)
{
int i;
@@ -348,10 +352,21 @@
int n = convolutional_out_height(layer)*
convolutional_out_width(layer);
- //cl_write_array(layer.filters_cl, layer.filters, m*k);
- //cl_write_array(layer.biases_cl, layer.biases, m);
bias_output_gpu(layer);
+
+ #ifdef TIMEIT
+ clock_t time = clock();
+ printf("Forward\n");
+ #endif
+
im2col_ongpu(in, layer.batch, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, layer.col_image_cl);
+
+ #ifdef TIMEIT
+ clFinish(cl.queue);
+ printf("Im2col %f\n", sec(clock()-time));
+ time = clock();
+ #endif
+
for(i = 0; i < layer.batch; ++i){
cl_mem a = layer.filters_cl;
cl_mem b = cl_sub_array(layer.col_image_cl, i*k*n, k*n);
@@ -360,8 +375,14 @@
clReleaseMemObject(b);
clReleaseMemObject(c);
}
+ #ifdef TIMEIT
+ clFinish(cl.queue);
+ printf("Gemm %f\n", sec(clock()-time));
+ #endif
activate_array_ongpu(layer.output_cl, m*n*layer.batch, layer.activation);
+ #ifdef TIMEIT
cl_read_array(layer.output_cl, layer.output, m*n*layer.batch);
+ #endif
}
void backward_convolutional_layer_gpu(convolutional_layer layer, cl_mem delta_cl)
@@ -384,9 +405,7 @@
clReleaseMemObject(a);
clReleaseMemObject(b);
}
- cl_read_array(layer.filter_updates_cl, layer.filter_updates, m*n);
- cl_read_array(layer.bias_updates_cl, layer.bias_updates, m);
-
+ //cl_read_array(layer.delta_cl, layer.delta, m*k*layer.batch);
if(delta_cl){
m = layer.size*layer.size*layer.c;
@@ -395,17 +414,44 @@
convolutional_out_width(layer);
for(i = 0; i < layer.batch; ++i){
- a = layer.filters_cl;
- b = cl_sub_array(layer.delta_cl, i*k*n, k*n);
- c = cl_sub_array(layer.col_image_cl, i*m*n, m*n);
+ cl_mem a = layer.filters_cl;
+ cl_mem b = cl_sub_array(layer.delta_cl, i*k*n, k*n);
+ cl_mem c = cl_sub_array(layer.col_image_cl, i*m*n, m*n);
gemm_ongpu(1,0,m,n,k,1,a,m,b,n,0,c,n);
clReleaseMemObject(b);
clReleaseMemObject(c);
}
- col2im_gpu(layer.col_image_cl, layer.batch, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, delta_cl);
+
+ scal_ongpu(layer.batch*layer.h*layer.w*layer.c,0,delta_cl, 1);
+ col2im_ongpu(layer.col_image_cl, layer.batch, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, delta_cl);
}
}
+void pull_convolutional_layer(convolutional_layer layer)
+{
+ cl_read_array(layer.filters_cl, layer.filters, layer.c*layer.n*layer.size*layer.size);
+ cl_read_array(layer.biases_cl, layer.biases, layer.n);
+}
+
+void push_convolutional_layer(convolutional_layer layer)
+{
+ cl_write_array(layer.filters_cl, layer.filters, layer.c*layer.n*layer.size*layer.size);
+ cl_write_array(layer.biases_cl, layer.biases, layer.n);
+}
+
+void update_convolutional_layer_gpu(convolutional_layer layer)
+{
+ int size = layer.size*layer.size*layer.c*layer.n;
+ axpy_ongpu(layer.n, layer.learning_rate, layer.bias_updates_cl, 1, layer.biases_cl, 1);
+ scal_ongpu(layer.n,layer.momentum, layer.bias_updates_cl, 1);
+
+ scal_ongpu(size, 1.-layer.learning_rate*layer.decay, layer.filters_cl, 1);
+ axpy_ongpu(size, layer.learning_rate, layer.filter_updates_cl, 1, layer.filters_cl, 1);
+ scal_ongpu(size, layer.momentum, layer.filter_updates_cl, 1);
+ pull_convolutional_layer(layer);
+}
+
+
#endif
--
Gitblit v1.10.0