search

hughperkins / DeepCL

OpenCL library to train deep convolutional neural networks
Mozilla Public License 2.0
865 stars 199 forks source link

clBlas error #127

Open skn123 opened 6 years ago

skn123 commented 6 years ago

Looks like you will have to change the clBlas files in your recursive pull. /home/naths/srcs/DeepCL/clMathLibraries/clBLAS/src/library/blas/trtri/diag_dtrtri_lower_128_16.cpp:65:2: error: embedding a #pragma directive within macro arguments is not supported

pragma unroll\n

^ /home/naths/srcs/DeepCL/clMathLibraries/clBLAS/src/library/blas/trtri/diag_dtrtri_lower_128_16.cpp:141:2: error: embedding a #pragma directive within macro arguments is not supported

pragma unroll\n

^ /home/naths/srcs/DeepCL/clMathLibraries/clBLAS/src/library/blas/trtri/diag_dtrtri_lower_128_16.cpp:166:2: error: embedding a #pragma directive within macro arguments is not supported

pragma unroll\n

The error has been fixed here: http://lists.alioth.debian.org/pipermail/debian-science-commits/2016-January/056094.html

hughperkins commented 6 years ago

Upgrading clBLAS is a ton of work. It might be easier just to backport the fix, into the current clBLAS fork, ie to https://github.com/hughperkins/clBLAS/tree/2.11-bugfixes (I think).

Thoughts?

skn123 commented 6 years ago

Your call :) I have no problem either way. Its just that I have clBlas as en external Library also and would ideally like to have clBlas as an external build rather than part of a superbuild. In that way, both conditions are satisfied. Thoughts?

hughperkins commented 6 years ago

Its just that I have clBlas as en external Library also and would ideally like to have clBlas as an external build rather than part of a superbuild

Ah. Personally, I found there were a few bugs to work around each time I upgraded... Since I was using an NVIDIA GPU at the time, those were both blocker bugs for me, and not really supported by AMD :-) . I suppose I was probably the only NVIDIA user of clBLAS :-P

If you want to dabble in upgrading to the latest version of clBLAS please go ahead. It would be most convenient if it was tested on all of ubuntu, Mac and windows, but no reason why you couldnt create a branch on your favored OS, and someone else could come along and fix the branch for one or more of the other OSes.

skn123 commented 6 years ago

Looks like this was fixed here https://github.com/clMathLibraries/clBLAS/pull/216/commits/9d4c312a012f6da830c57cd7db8f3d335fd81a9c But I think the pull does not reflect it in DeepCL

hughperkins commented 6 years ago

Thoughts on backporting this into the version of clBLAS used by DeepCL?

Alternatively, thoughts on upgrading DeepCL to latest version of clBLAS? (we could do this in a branch perhaps)