Skip to content
This repository has been archived by the owner on Nov 17, 2023. It is now read-only.

Dual stream cudnn Convolution backward() with MXNET_GPU_WORKER_NSTREAMS=2. #14006

Merged
merged 11 commits into from
Feb 24, 2019
Merged

Dual stream cudnn Convolution backward() with MXNET_GPU_WORKER_NSTREAMS=2. #14006

Changes from 1 commit
Commits
Show all changes
11 commits
Select commit Hold shift + click to select a range
ff759f2
Dual stream conv backward(). Enable with MXNET_GPU_WORKER_NSTREAMS=2.
DickJC123 Sep 26, 2018
d52774d
Fix for MSVC compiler.
DickJC123 Jan 29, 2019
1cf5c67
Fix cpplint.
DickJC123 Jan 29, 2019
e5b57a8
Add MXNET_GPU_WORKER_NSTREAMS env var documentation.
DickJC123 Jan 29, 2019
d16e85f
Improve test function and commenting.
DickJC123 Jan 30, 2019
b44d096
Merge remote-tracking branch 'mxnet/master' into parallel_conv_dgrad_…
DickJC123 Jan 31, 2019
838bda0
Add description of proper aux stream use using events.
DickJC123 Jan 31, 2019
c3f3320
RAII rework to simplify usage within operators.
DickJC123 Feb 9, 2019
db5075f
Fix cpplint.
DickJC123 Feb 10, 2019
ea56c91
Expand testing to cover all engines.
DickJC123 Feb 11, 2019
790a998
Fix NaiveEngine shutdown segfault on CentOS7.
DickJC123 Feb 14, 2019
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Prev Previous commit
Next Next commit
Fix cpplint.
  • Loading branch information
@DickJC123
DickJC123 committed Jan 29, 2019
commit 1cf5c67370b32d8e41b9669f03071dd8560383f6
3 changes: 2 additions & 1 deletion src/operator/nn/cudnn/cudnn_convolution-inl.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1038,7 +1038,8 @@ class CuDNNConvolutionOp {
// Always allocates at least one word.
mshadow::Tensor<gpu, 1, DType> AllocateTempWorkspace(const OpContext &ctx, size_t size_bytes) {
mshadow::Stream<gpu> *s = ctx.get_stream<gpu>();
size_t size_words = std::max<size_t>(1, RoundToMultiple(size_bytes, sizeof(DType)) / sizeof(DType));
size_t size_words =
std::max<size_t>(1, RoundToMultiple(size_bytes, sizeof(DType)) / sizeof(DType));
return ctx.requested[conv::kTempSpace].get_space_typed<gpu, 1, DType>(
mshadow::Shape1(size_words), s);
}
Expand Down