From af4e4f92dc9e5da160eb6c6870a7b38b863f1c6c Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Tue, 28 Oct 2014 02:45:06 +0000
Subject: [PATCH] getting rid of sub_arrays, nvidia driver memory leak
---
src/convolutional_layer.c | 66 +++++++++++++++++++++++----------
1 files changed, 46 insertions(+), 20 deletions(-)
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index 00de153..42f4f21 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)
{
@@ -211,7 +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);
@@ -341,6 +342,8 @@
check_error(cl);
}
+//#define TIMEIT
+
void forward_convolutional_layer_gpu(convolutional_layer layer, cl_mem in)
{
int i;
@@ -349,20 +352,35 @@
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);
- cl_mem c = cl_sub_array(layer.output_cl, i*m*n, m*n);
- gemm_ongpu(0,0,m,n,k,1.,a,k,b,n,1.,c,n);
- clReleaseMemObject(b);
- clReleaseMemObject(c);
+ cl_mem b = layer.col_image_cl;
+ cl_mem c = layer.output_cl;
+ gemm_ongpu_offset(0,0,m,n,k,1.,a,0,k,b,i*k*n,n,1.,c,i*m*n,n);
}
+ #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);
- //cl_read_array(layer.output_cl, layer.output, m*n*layer.batch);
+ #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)
@@ -376,14 +394,11 @@
learn_bias_convolutional_layer_ongpu(layer);
for(i = 0; i < layer.batch; ++i){
- cl_mem a = cl_sub_array(layer.delta_cl,i*m*k, m*k);
- cl_mem b = cl_sub_array(layer.col_image_cl,i*k*n, k*n);
+ cl_mem a = layer.delta_cl;
+ cl_mem b = layer.col_image_cl;
cl_mem c = layer.filter_updates_cl;
- gemm_ongpu(0,1,m,n,k,1,a,k,b,k,1,c,n);
-
- clReleaseMemObject(a);
- clReleaseMemObject(b);
+ gemm_ongpu_offset(0,1,m,n,k,1,a,i*m*k,k,b,i*k*n,k,1,c,0,n);
}
//cl_read_array(layer.delta_cl, layer.delta, m*k*layer.batch);
@@ -395,12 +410,10 @@
for(i = 0; i < layer.batch; ++i){
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);
+ cl_mem b = layer.delta_cl;
+ cl_mem c = layer.col_image_cl;
- gemm_ongpu(1,0,m,n,k,1,a,m,b,n,0,c,n);
- clReleaseMemObject(b);
- clReleaseMemObject(c);
+ gemm_ongpu_offset(1,0,m,n,k,1,a,0,m,b,i*k*n,n,0,c,i*m*n,n);
}
scal_ongpu(layer.batch*layer.h*layer.w*layer.c,0,delta_cl, 1);
@@ -408,6 +421,18 @@
}
}
+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;
@@ -417,6 +442,7 @@
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);
}
--
Gitblit v1.10.0