[Pkg-fglrx-devel] Bug#793488: amd-opencl-icd: Update to 1:15.7-1 broke opencl

[andrey.gursky at e-mail.ua]

23.09.2015, 20:10, "Andreas Beckmann" <anbe at debian.org>:
> On 2015-09-23 19:49, andrey.gursky at e-mail.ua wrote:
>>  23.09.2015, 19:44, "Andreas Beckmann" <anbe at debian.org>:
>>>  On 2015-09-23 18:22, andrey.gursky at e-mail.ua wrote:
>>>>   dear maintainers, the package amd-opencl-icd 15.7 contains a different libamdocl64.so than in the 15.7 fglrx release (and even the one residing on debian server). But how could this happen?
>>>  Can someone please check the package in jessie-backports?
>>  Andreas, if you mean [1], then I downloaded [2] and it contains a "broken" one:
>
> Thanks.
>
> The build process runs
>
>   strip --remove-section=.comment --remove-section=.note --strip-unneeded debian/amd-opencl-icd/usr/lib/x86_64-linux-gnu/libamdocl64.so
>
> which is regular procedure for all shared libraries and should be a safe
> operation, but the library provided by AMD seems to be fragile this time.

Indeed. I applied strip as you speciefied to the original .so and received exactly the same as in the deb package. Then it is either a kind of "protection" from AMD or bug in strip/binutils?

> Is the amd-ocl-icd:i386 package affected as well? (I.e. does it not work,
> having "just" a different file size due to stripping is not enough.)
>
> Running clinfo (the i386 binary!) should be sufficient to list the
> available devices.

I downloaded the 32bit packages and run clinfo locally:
LD_LIBRARY_PATH=. OPENCL_VENDOR_PATH=. ./clinfo
but it crashed. Then I installed them properly with aptitude and it crashs either. The same with original 15.7 and also 15.9. But even wine32 can run opencl code, thus it should actually work on amd64, though I haven't tested it since a while.

Andrey