NextCloud/dlib
1
0
Fork 0
mirror of https://github.com/davisking/dlib.git synced 2024-11-01 10:14:53 +08:00
Code Issues Packages Projects Releases Wiki Activity

add support for cudnn 8.0

Browse Source
This commit is contained in:
Davis King 2020-06-20 09:43:17 -04:00
parent f8887d8cbb
commit fe803b566f
1 changed files with 85 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

85
dlib/cuda/cudnn_dlibapi.cpp
View File

@ -751,6 +751,31 @@ namespace dlib
backward_filters_workspace.reset(); backward_filters_workspace.reset();
} }
// Given an array of cudnn algorithm performance results, like
// cudnnConvolutionFwdAlgoPerf_t, pick the best one to use.
template <typename T>
decltype(std::declval<T>().algo) pick_best_algorithm(const std::vector<T> &perf_results)
{
DLIB_CASSERT(!perf_results.empty());
CHECK_CUDNN(perf_results[0].status);
if (dnn_prefer_fastest_algorithms())
return perf_results[0].algo;
// Otherwise we find the algorithm that has a good status and uses the least amount
// of memory.
size_t best_memory = std::numeric_limits<size_t>::max();
decltype(std::declval<T>().algo) best_alg;
for (auto&& perf : perf_results)
{
if (perf.status == CUDNN_STATUS_SUCCESS && perf.memory < best_memory)
{
best_memory = perf.memory;
best_alg = perf.algo;
}
}
return best_alg;
}
void tensor_conv:: void tensor_conv::
setup( setup(
const tensor& data, const tensor& data,
@ -841,6 +866,25 @@ namespace dlib
// Pick which forward algorithm we will use and allocate the necessary // Pick which forward algorithm we will use and allocate the necessary
// workspace buffer. // workspace buffer.
cudnnConvolutionFwdAlgo_t forward_best_algo; cudnnConvolutionFwdAlgo_t forward_best_algo;
#if CUDNN_MAJOR >= 8
{
int num_possilbe_algorithms = 0;
CHECK_CUDNN(cudnnGetConvolutionForwardAlgorithmMaxCount(context(), &num_possilbe_algorithms));
std::vector<cudnnConvolutionFwdAlgoPerf_t> perf_results(num_possilbe_algorithms);
int num_algorithms = 0;
CHECK_CUDNN(cudnnFindConvolutionForwardAlgorithm(
context(),
descriptor(data),
(const cudnnFilterDescriptor_t)filter_handle,
(const cudnnConvolutionDescriptor_t)conv_handle,
descriptor(dest_desc),
num_possilbe_algorithms,
&num_algorithms,
perf_results.data()));
perf_results.resize(num_algorithms);
forward_best_algo = pick_best_algorithm(perf_results);
}
#else
CHECK_CUDNN(cudnnGetConvolutionForwardAlgorithm( CHECK_CUDNN(cudnnGetConvolutionForwardAlgorithm(
context(), context(),
descriptor(data), descriptor(data),
@ -850,6 +894,7 @@ namespace dlib
dnn_prefer_fastest_algorithms()?CUDNN_CONVOLUTION_FWD_PREFER_FASTEST:CUDNN_CONVOLUTION_FWD_NO_WORKSPACE, dnn_prefer_fastest_algorithms()?CUDNN_CONVOLUTION_FWD_PREFER_FASTEST:CUDNN_CONVOLUTION_FWD_NO_WORKSPACE,
std::numeric_limits<size_t>::max(), std::numeric_limits<size_t>::max(),
&forward_best_algo)); &forward_best_algo));
#endif
forward_algo = forward_best_algo; forward_algo = forward_best_algo;
CHECK_CUDNN(cudnnGetConvolutionForwardWorkspaceSize( CHECK_CUDNN(cudnnGetConvolutionForwardWorkspaceSize(
context(), context(),
@ -863,6 +908,25 @@ namespace dlib
// Pick which backward data algorithm we will use and allocate the // Pick which backward data algorithm we will use and allocate the
// necessary workspace buffer. // necessary workspace buffer.
cudnnConvolutionBwdDataAlgo_t backward_data_best_algo; cudnnConvolutionBwdDataAlgo_t backward_data_best_algo;
#if CUDNN_MAJOR >= 8
{
int num_possilbe_algorithms = 0;
CHECK_CUDNN(cudnnGetConvolutionBackwardFilterAlgorithmMaxCount(context(), &num_possilbe_algorithms));
std::vector<cudnnConvolutionBwdDataAlgoPerf_t> perf_results(num_possilbe_algorithms);
int num_algorithms = 0;
CHECK_CUDNN(cudnnFindConvolutionBackwardDataAlgorithm(
context(),
(const cudnnFilterDescriptor_t)filter_handle,
descriptor(dest_desc),
(const cudnnConvolutionDescriptor_t)conv_handle,
descriptor(data),
num_possilbe_algorithms,
&num_algorithms,
perf_results.data()));
perf_results.resize(num_algorithms);
backward_data_best_algo = pick_best_algorithm(perf_results);
}
#else
CHECK_CUDNN(cudnnGetConvolutionBackwardDataAlgorithm( CHECK_CUDNN(cudnnGetConvolutionBackwardDataAlgorithm(
context(), context(),
(const cudnnFilterDescriptor_t)filter_handle, (const cudnnFilterDescriptor_t)filter_handle,
@ -872,6 +936,7 @@ namespace dlib
dnn_prefer_fastest_algorithms()?CUDNN_CONVOLUTION_BWD_DATA_PREFER_FASTEST:CUDNN_CONVOLUTION_BWD_DATA_NO_WORKSPACE, dnn_prefer_fastest_algorithms()?CUDNN_CONVOLUTION_BWD_DATA_PREFER_FASTEST:CUDNN_CONVOLUTION_BWD_DATA_NO_WORKSPACE,
std::numeric_limits<size_t>::max(), std::numeric_limits<size_t>::max(),
&backward_data_best_algo)); &backward_data_best_algo));
#endif
backward_data_algo = backward_data_best_algo; backward_data_algo = backward_data_best_algo;
CHECK_CUDNN(cudnnGetConvolutionBackwardDataWorkspaceSize( CHECK_CUDNN(cudnnGetConvolutionBackwardDataWorkspaceSize(
@ -886,6 +951,25 @@ namespace dlib
// Pick which backward filters algorithm we will use and allocate the // Pick which backward filters algorithm we will use and allocate the
// necessary workspace buffer. // necessary workspace buffer.
cudnnConvolutionBwdFilterAlgo_t backward_filters_best_algo; cudnnConvolutionBwdFilterAlgo_t backward_filters_best_algo;
#if CUDNN_MAJOR >= 8
{
int num_possilbe_algorithms = 0;
CHECK_CUDNN(cudnnGetConvolutionBackwardFilterAlgorithmMaxCount(context(), &num_possilbe_algorithms));
std::vector<cudnnConvolutionBwdFilterAlgoPerf_t> perf_results(num_possilbe_algorithms);
int num_algorithms = 0;
CHECK_CUDNN(cudnnFindConvolutionBackwardFilterAlgorithm(
context(),
descriptor(data),
descriptor(dest_desc),
(const cudnnConvolutionDescriptor_t)conv_handle,
(const cudnnFilterDescriptor_t)filter_handle,
num_possilbe_algorithms,
&num_algorithms,
perf_results.data()));
perf_results.resize(num_algorithms);
backward_filters_best_algo = pick_best_algorithm(perf_results);
}
#else
CHECK_CUDNN(cudnnGetConvolutionBackwardFilterAlgorithm( CHECK_CUDNN(cudnnGetConvolutionBackwardFilterAlgorithm(
context(), context(),
descriptor(data), descriptor(data),
@ -895,6 +979,7 @@ namespace dlib
dnn_prefer_fastest_algorithms()?CUDNN_CONVOLUTION_BWD_FILTER_PREFER_FASTEST:CUDNN_CONVOLUTION_BWD_FILTER_NO_WORKSPACE, dnn_prefer_fastest_algorithms()?CUDNN_CONVOLUTION_BWD_FILTER_PREFER_FASTEST:CUDNN_CONVOLUTION_BWD_FILTER_NO_WORKSPACE,
std::numeric_limits<size_t>::max(), std::numeric_limits<size_t>::max(),
&backward_filters_best_algo)); &backward_filters_best_algo));
#endif
// cuDNN 5.1 has a bug that causes // cuDNN 5.1 has a bug that causes
// cudnnGetConvolutionBackwardFilterAlgorithm() to pick the winograd // cudnnGetConvolutionBackwardFilterAlgorithm() to pick the winograd
// algorithm even for cases where cuDNN doesn't support it, leading to // algorithm even for cases where cuDNN doesn't support it, leading to