[Pkg-opencl-devel] Moving to LLVM 3.9 (was: pocl)

[Andreas Boll]

Hi,

we are currently in the same situation again. Mesa has moved to LLVM 3.9
and thus beignet will automatically switch with the next upload/binNMU
to LLVM 3.9 too [1]. Unfortunately pocl 0.13 isn't compatible with LLVM
3.9. There's a git commit with fixes for 3.9 but I don't know if it's
enough [2]. Fedora Rawhide has moved to a git snapshot of pocl 0.14 to
get compatibility with LLVM 3.9 [3].

What should we do with pocl?

There're already reports of random crashes with:
- Blender + mesa-opencl-icd + pocl [4]
- Blender + beignet-opencl-icd [5]
- Related: kdevelop (was LLVM 3.8) + mesa (LLVM 3.9) crashed [6]

And there's a LLVM bug which might cause these issues [7].

Thanks,
Andreas

[1] https://sources.debian.net/src/beignet/1.2.1-1/debian/rules/#L16
[2] https://github.com/pocl/pocl/commit/08742e77004ec4ffdb081d2c493ef994b075cfcb
[3] https://apps.fedoraproject.org/packages/pocl/changelog/
[4] https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=846162#5
[5] https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=848792
[6] https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=847447
[7] https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=848368

2016-04-23 11:37 GMT+02:00 Giuseppe Bilotta <giuseppe.bilotta at gmail.com>:
> My limited experience with building pocl manually in debian unstable
> has taught me a couple of things:
>
> * all ICDs based on LLVM have to be built using the same version of
> LLVM, or one will get random crashes and segfaults during execution;
> this means that Pocl must be built with the same LLVM as
> mesa-opencl-icd and beignet;
> * Pocl needs at least LLVM 3.7, but Debian defaults to LLVM 3.6, so
> the LLVM version must be specified manually (by the way, why is 3.6
> still the default in Debian?)
> * automatic build dependencies are affected by
> https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=792842
> * there is a known issue with LLVM 3.7 (at least the version that
> ships in Debian unstable) producing the wrong shift/rotation
> instructions for short/ushort vector types on Haswell CPUs, which
> causes some tests to fail.