[Pkg-opencl-devel] Bug#994833: OpenCL programs abort when intel-opencl-icd is installed

[Giuseppe Bilotta]

On Wed, Sep 22, 2021 at 8:32 AM Andreas Beckmann <anbe at debian.org> wrote:
> So if it does matter abi-wise for intel-opencl-icd which version of llvm
> libigfoo1 was compiled against, we should model this in the dependency
> chain.

To be able to confirm this, I would need a version of all the libig*
stuff correctly
compiled against LLVM12 (rather than the mix-up we currently have for
libigdfcl1).
I can then test against intel-opencl-icd version 20.x, which is built
with LLVM11,
and (most probably) confirm that the setup is broken.

> It's probably best if all libigfoo1 library packages provide a virtual
> package libigfoo1-llvmXX (don't hardcode it, use substvars) and
> intel-opencl-icd depends on that (in addition to libigfoo1 (>= xx)) to
> specify the specific abi needed.
> (renaming the real package to libigfoo1-llvmXX each time the llvm major
> version changes is probably overkill)

The virtual ABI package sounds like the best choice.
I don't think renaming the real package makes sense,
at least at the moment. It might become useful if there ever is a need
for the different ABIs to be co-installed.

> Are there other users of libigfoo1 besides intel-opencl-icd?

The only one I can see is intel-media-va-driver{,-non-free} depending
on libigdgmm11,
but that particular package doesn't seem to have an LLVM dependency
(directly or not).

> We will proably still run into problems if different ICDs built against
> different LLVM versions are going to be loaded at the same time (e.g.
> pocl/llvm9 and intel/llvm1x) because the different llvm versions seem to
> stomp on each others internal bits. There are bugs open about that ...

FWIW, I haven't had these issues in a while now
(but I'm also building my own pocl, so maybe by chance I'm using the
same LLVM version anyway.)

> I may come up with a patch if time permits.

Much appreciated.

-- 
Giuseppe "Oblomov" Bilotta