OrgLion-ML
/
dlib
mirror of https://github.com/davisking/dlib.git
1
0
Fork 0
Code Issues Releases Wiki Activity

Perform conv with bias for specific algo with cudnn (#2839)

This commit is contained in:
Facundo Galán 2023-07-27 19:11:32 -03:00 committed by GitHub
parent 7b6021eee7
commit c7b2917498
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 26 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
dlib/cuda/cudnn_dlibapi.cpp
View File

@ -928,7 +928,7 @@ namespace dlib
backward_filters_best_algo = CUDNN_CONVOLUTION_BWD_FILTER_ALGO_0;
}
#endif
backward_filters_algo = backward_filters_best_algo;
backward_filters_algo = backward_filters_best_algo;
// Save this algorithm selection in the cache
config_to_algo_cache[cache_key] = std::make_tuple(forward_algo, backward_data_algo, backward_filters_algo);
@ -1177,6 +1177,19 @@ namespace dlib
const tensor& biases
)
{
// Function cudnnConvolutionBiasActivationForward should only be called with CUDNN_ACTIVATION_IDENTITY when
// the chosen forward algorithm is CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM, as cuDNN documentation explicitly says.
// In case the algorithm is different, perform the forward pass and bias addition separately.
if (forward_algo != CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM)
{
(*this)(add_to_output, output, data, filters);
tt::add(1, output, 1, biases);
return;
}
DLIB_CASSERT(is_same_object(output,data) == false);
DLIB_CASSERT(is_same_object(output,filters) == false);
DLIB_CASSERT(filters.k() == data.k());

22
dlib/dnn/layers.h
View File

@ -233,17 +233,19 @@ namespace dlib
padding_y_,
padding_x_);
conv(false, output,
sub.get_output(),
filters(params,0));
// For some reason, doing this is sometimes slower than two separate calls
// conv(false, output,
// sub.get_output(),
// filters(params,0),
// biases(params, filters.size()));
if (use_bias)
tt::add(1,output,1,biases(params,filters.size()));
{
conv(false, output,
sub.get_output(),
filters(params,0),
biases(params, filters.size()));
}
else
{
conv(false, output,
sub.get_output(),
filters(params,0));
}
}
template <typename SUBNET>