github
/
darknet
mirror of https://github.com/AlexeyAB/darknet.git
1
0
Fork 0
Code Issues Releases Wiki Activity

Disabled CUDNN_TENSOR_OP_MATH_ALLOW_CONVERSION, and only if groups==1 use CUDNN_TENSOR_OP_MATH

Browse Source
pull/4560/head
AlexeyAB 5 years ago
parent dc7277f152
commit e66d3b1bdb
1 changed files with 2 additions and 2 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Unified View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 4
      src/convolutional_layer.c

4
src/convolutional_layer.c
Unescape View File

@ -231,10 +231,10 @@ void cudnn_convolutional_setup(layer *l, int cudnn_preference, size_t workspace_
// 2. Loss Scaling - required only for: activation gradients. We do not use. // 2. Loss Scaling - required only for: activation gradients. We do not use.
// 3. FP32 Master Copy of Weights // 3. FP32 Master Copy of Weights
// More: http://docs.nvidia.com/deeplearning/sdk/cudnn-developer-guide/index.html#tensor_ops // More: http://docs.nvidia.com/deeplearning/sdk/cudnn-developer-guide/index.html#tensor_ops
CHECK_CUDNN(cudnnSetConvolutionMathType(l->convDesc, CUDNN_TENSOR_OP_MATH));
CHECK_CUDNN(cudnnSetConvolutionGroupCount(l->convDesc, l->groups)); CHECK_CUDNN(cudnnSetConvolutionGroupCount(l->convDesc, l->groups));
if (l->groups == 1) CHECK_CUDNN(cudnnSetConvolutionMathType(l->convDesc, CUDNN_TENSOR_OP_MATH));
#if((CUDNN_MAJOR*10 + CUDNN_MINOR) >= 72) // cuDNN >= 7.2 #if((CUDNN_MAJOR*10 + CUDNN_MINOR) >= 72) // cuDNN >= 7.2
CHECK_CUDNN(cudnnSetConvolutionMathType(l->convDesc, CUDNN_TENSOR_OP_MATH_ALLOW_CONVERSION)); //CHECK_CUDNN(cudnnSetConvolutionMathType(l->convDesc, CUDNN_TENSOR_OP_MATH_ALLOW_CONVERSION)); // reduces the speed of regular and group convolution
#endif #endif
#else //if(CUDNN_MAJOR >= 7) #else //if(CUDNN_MAJOR >= 7)
if (l->groups > 1) { if (l->groups > 1) {

Write Preview
Loading…
Cancel
Save