OSSForks/pytorch
1
0
Fork 0
mirror of https://github.com/zebrajr/pytorch.git synced 2025-12-06 12:20:52 +01:00
Code Issues Actions 38 Packages Projects Releases Wiki Activity

Skip compiledWithCuDNN() call for mobile to avoid segfault (#71775)

Browse Source 
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/71775

Mobile is running into segfaults at the `compiledWithCuDNN()` call as described in T110194934. This fix works around this with an #ifdef following the approach done [here](d32b7d9585/aten/src/ATen/native/Convolution.cpp (L1076-L1088)). TBD how to fix the underlying cause.

Test Plan: Imported from OSS

Reviewed By: anjali411

Differential Revision: D33778888

Pulled By: jbschlosser

fbshipit-source-id: 2a22b2eaa858ee6adf5b3c25a1c470c6aebc3f87
(cherry picked from commit e90a6bb402)
This commit is contained in:
Joel Schlosser 2022-01-26 08:55:54 -08:00 committed by PyTorch MergeBot
parent 0891c908bb
commit e04ade92ae
1 changed files with 9 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
aten/src/ATen/native/Convolution.cpp
View File

@ -176,6 +176,11 @@ auto ConvParams::needs_64bit_indexing_no_split(const at::Tensor& input, const at
}
auto ConvParams::use_cudnn(const at::Tensor& input, const at::Tensor& weight) const -> bool {
// Note [Mobile check segfaults]
// cudnn and miopen are guaranteed not to be on mobile, and T102591915 / T110194934 suggest
// that maybe the compiledWithCuDNN() check sometimes segfaults (though I can't imagine how)
#if !defined(C10_MOBILE)
if (needs_64bit_indexing_no_split(input, weight)) {
return false;
}
@ -199,6 +204,9 @@ auto ConvParams::use_cudnn(const at::Tensor& input, const at::Tensor& weight) co
}
}
return !is_output_padding_big();
#else
return false;
#endif
}
auto ConvParams::use_miopen(const at::Tensor& input, const at::Tensor& weight, bool bias_defined) const -> bool {
@ -1074,9 +1082,7 @@ static inline at::MemoryFormat determine_backend_memory_format(
at::MemoryFormat backend_memory_format = at::MemoryFormat::Contiguous;
auto k = weight.ndimension();
#if !defined(C10_MOBILE)
// cudnn and miopen are guaranteed not to be on mobile, and T102591915
// suggests that maybe the cudnn condition sometimes segfaults (though
// I can't imagine how)
// See Note [Mobile check segfaults]
if (detail::getCUDAHooks().compiledWithCuDNN()) {
backend_memory_format = cudnn_conv_suggest_memory_format(input, weight);
}