From 880cf187d87c904f5fe574802ecff99118643f2d Mon Sep 17 00:00:00 2001
From: AlexeyAB <alexeyab84@gmail.com>
Date: Fri, 09 Mar 2018 16:44:46 +0000
Subject: [PATCH] Fixed multi-GPU training for Tensor Cores
---
src/network.c | 17 ++++++++++++++++-
1 files changed, 16 insertions(+), 1 deletions(-)
diff --git a/src/network.c b/src/network.c
index 10d1ead..964d3e8 100644
--- a/src/network.c
+++ b/src/network.c
@@ -140,6 +140,11 @@
#ifdef GPU
net.input_gpu = calloc(1, sizeof(float *));
net.truth_gpu = calloc(1, sizeof(float *));
+
+ net.input16_gpu = calloc(1, sizeof(float *));
+ net.output16_gpu = calloc(1, sizeof(float *));
+ net.max_input16_size = calloc(1, sizeof(size_t));
+ net.max_output16_size = calloc(1, sizeof(size_t));
#endif
return net;
}
@@ -316,6 +321,8 @@
net->layers[i].batch = b;
#ifdef CUDNN
if(net->layers[i].type == CONVOLUTIONAL){
+ cudnn_convolutional_setup(net->layers + i, cudnn_fastest);
+ /*
layer *l = net->layers + i;
cudnn_convolutional_setup(l, cudnn_fastest);
// check for excessive memory consumption
@@ -327,6 +334,7 @@
cudnn_convolutional_setup(l, cudnn_smallest);
l->workspace_size = get_workspace_size(*l);
}
+ */
}
#endif
}
@@ -356,7 +364,7 @@
//fflush(stderr);
for (i = 0; i < net->n; ++i){
layer l = net->layers[i];
- printf(" %d: layer = %d,", i, l.type);
+ //printf(" %d: layer = %d,", i, l.type);
if(l.type == CONVOLUTIONAL){
resize_convolutional_layer(&l, w, h);
}else if(l.type == CROP){
@@ -619,6 +627,13 @@
if (*net.truth_gpu) cuda_free(*net.truth_gpu);
if (net.input_gpu) free(net.input_gpu);
if (net.truth_gpu) free(net.truth_gpu);
+
+ if (*net.input16_gpu) cuda_free(*net.input16_gpu);
+ if (*net.output16_gpu) cuda_free(*net.output16_gpu);
+ if (net.input16_gpu) free(net.input16_gpu);
+ if (net.output16_gpu) free(net.output16_gpu);
+ if (net.max_input16_size) free(net.max_input16_size);
+ if (net.max_output16_size) free(net.max_output16_size);
#else
free(net.workspace);
#endif
--
Gitblit v1.10.0