search

Maratyszcza / NNPACK

Acceleration package for neural networks on multi-core CPUs
BSD 2-Clause "Simplified" License
1.67k stars 317 forks source link

Fix x86 32 bit detection #161

Closed t-vi closed 5 years ago

t-vi commented 5 years ago

When building on 32bit linux ("i386" architecture) on a 64 bit machine, cmake will report CMAKE_SYSTEM_PROCESSOR as x86_64 - which is not intended. This currently causes NNPACK to detect x86-64 target when the psimd would be correct - the result is a mix of x86_64 and x86 object files and unhappy linkers.

t-vi commented 5 years ago

Hello Marat,

thanks for your work on NNPACK!

Here is some background info for the PR:

While investigating https://github.com/pytorch/pytorch/issues/17901, I found that NNPACK won't build on i386 systems due to CMake reporting x86_64 as the CMAKE_SYSTEM_PROCESSOR.

See here for a similar workaround: https://github.com/opencv/opencv/blob/master/cmake/OpenCVDetectCXXCompiler.cmake#L104

Maratyszcza commented 5 years ago

The proposed fix would misdetect x86-64 systems with x32 ABI. As a better work-around, you can explicitly specify -DCMAKE_SYSTEM_PROCESSOR=i386 or -DNNPACK_BACKEND=psimd to avoid x86-64-specific parts of NNPACK.

t-vi commented 5 years ago

Hi Marat,

right. I thought that x32 was much a more unlikely target, but as the developer you'd know better. Would you take a patch using a compile check whether __386__ or _M_I386 is defined and switch on that?

Best regards

Thomas

Maratyszcza commented 5 years ago

Yes, doing preprocessing-time check sounds good with me