From 48586c8d4db5c00d3d4b9dabcc9a5d2294c5b15d Mon Sep 17 00:00:00 2001
From: AlexeyAB <alexeyab84@gmail.com>
Date: Sun, 04 Mar 2018 15:15:56 +0000
Subject: [PATCH] Compile fix
---
src/convolutional_layer.c | 8 ++++++--
1 files changed, 6 insertions(+), 2 deletions(-)
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index d35246e..377b898 100644
--- a/src/convolutional_layer.c
+++ b/src/convolutional_layer.c
@@ -141,7 +141,8 @@
{
#ifdef CUDNN_HALF
- // TRUE_HALF_CONFIG is only supported on architectures with true fp16 support (compute capability 5.3 and 6.0): Tegra X1, Jetson TX1, DRIVE CX, DRIVE PX, Quadro GP100, Tesla P100
+ // TRUE_HALF_CONFIG is only supported on architectures with true fp16 support (compute capability 5.3 and 6.0):
+ // Tegra X1, Jetson TX1, DRIVE CX, DRIVE PX, Quadro GP100, Tesla P100
// PSEUDO_HALF_CONFIG is required for Tensor Cores - our case!
const cudnnDataType_t data_type = CUDNN_DATA_HALF;
#else
@@ -164,10 +165,12 @@
// on architectures with DP4A support (compute capability 6.1 and later).
//cudnnDataType_t data_type = CUDNN_DATA_INT8;
+ // backward delta
cudnnSetTensor4dDescriptor(l->dsrcTensorDesc, CUDNN_TENSOR_NCHW, data_type, l->batch, l->c, l->h, l->w);
cudnnSetTensor4dDescriptor(l->ddstTensorDesc, CUDNN_TENSOR_NCHW, data_type, l->batch, l->out_c, l->out_h, l->out_w);
cudnnSetFilter4dDescriptor(l->dweightDesc, data_type, CUDNN_TENSOR_NCHW, l->n, l->c, l->size, l->size);
+ // forward
cudnnSetTensor4dDescriptor(l->srcTensorDesc, CUDNN_TENSOR_NCHW, data_type, l->batch, l->c, l->h, l->w);
cudnnSetTensor4dDescriptor(l->dstTensorDesc, CUDNN_TENSOR_NCHW, data_type, l->batch, l->out_c, l->out_h, l->out_w);
cudnnSetFilter4dDescriptor(l->weightDesc, data_type, CUDNN_TENSOR_NCHW, l->n, l->c, l->size, l->size);
@@ -302,7 +305,8 @@
l.weights_gpu = cuda_make_array(l.weights, c*n*size*size);
#ifdef CUDNN_HALF
- l.weights_gpu16 = cuda_make_array(l.weights, c*n*size*size/2);
+ l.weights_gpu16 = cuda_make_array(l.weights, c*n*size*size / 2);
+ l.weight_updates_gpu16 = cuda_make_array(l.weight_updates, c*n*size*size / 2);
#endif
l.weight_updates_gpu = cuda_make_array(l.weight_updates, c*n*size*size);
--
Gitblit v1.10.0