[Mesa-stable] [Mesa-dev] [PATCH 3/3] amd: Apply elf relocations and allow code with relocations

[Jan Vesely]

On Sat, 2019-06-15 at 07:38 +0200, Dieter Nützel wrote:
> Am 14.06.2019 08:13, schrieb Jan Vesely:
> > On Thu, 2019-06-13 at 21:20 +0200, Dieter Nützel wrote:
> > > Am 13.06.2019 07:10, schrieb Marek Olšák:
> > > > FYI, I just pushed the new linker.
> > > > 
> > > > Marek
> > > 
> > > Thank you very much Marek and _Nicolai_ for this GREAT stuff.
> > > It brings back some speed after 1/8 drop with glmark2, lately.
> > > Maybe my amd-staging-drm-next tree (5.2-rc1) didn't honor the kernel
> > > mitigation parameter right.
> > > 
> > > @Jan
> > > Go ahead with your nice relocation and image work.
> > > Send me what you have in the works.
> > 
> > The relocation work is no longer needed as the new linker handles
> > things.
> > The corruption is caused either by (still faulty) conversion builtins,
> > or incorrect buffer coherence handling. Both need fixing, but I'm not
> > sure which one is to blame in this case.
> > 
> > > Latest Mesa git (with Nicolai's new linker) let all 3 luxmark versions
> > > run.
> > > Only 'Hotel lobby' (with v3.0 and v3.1) show some corruption but do 
> > > NOT
> > > crash any longer. Numbers for 'Neumann TLM-102 SE' (medium) show 
> > > ~43000K
> > > (!!!).
> > > 
> > > https://nam02.safelinks.protection.outlook.com/?url=https%3A%2F%2Fwww.phoronix.com%2Fforums%2Fforum%2Fphoronix%2Flatest-phoronix-articles%2F1106085-linux-kernel-set-to-expose-hidden-nvidia-hda-controllers-helping-laptop-users%3Fp%3D1106199%23post1106199&data=02%7C01%7Cjan.vesely%40cs.rutgers.edu%7Ca6eda55e70a546c57cfa08d6f153b93a%7Cb92d2b234d35447093ff69aca6632ffe%7C1%7C0%7C636961739242915247&sdata=HK9Shj%2B8monTnuXeWu%2BgYT77EjqcXT7NOtnpoiyeQNY%3D&reserved=0
> > > 
> > > Blender crash as expected ;-)
> > > 
> > > /home/dieter> trying to save userpref at
> > > /home/dieter/.config/blender/2.79/config/userpref.blend ok
> > > Read blend: /data/Blender/barbershop_interior_gpu.blend
> > > scripts disabled for "/data/Blender/barbershop_interior_gpu.blend",
> > > skipping 'generate_customprops.py'
> > > skipping driver 'var', automatic scripts are disabled
> > > skipping driver 'var', automatic scripts are disabled
> > > skipping driver 'var', automatic scripts are disabled
> > > skipping driver 'var', automatic scripts are disabled
> > > skipping driver 'var', automatic scripts are disabled
> > > skipping driver 'var', automatic scripts are disabled
> > > skipping driver 'var', automatic scripts are disabled
> > > skipping driver 'var', automatic scripts are disabled
> > > skipping driver 'var', automatic scripts are disabled
> > > Device init success
> > > Compiling OpenCL program split
> > > Kernel compilation of split finished in 8.41s.
> > > 
> > > Compiling OpenCL program base
> > > Kernel compilation of base finished in 4.55s.
> > > 
> > > Compiling OpenCL program denoising
> > > Kernel compilation of denoising finished in 2.08s.
> > > 
> > > blender: ../src/gallium/drivers/radeonsi/si_compute.c:319:
> > > si_set_global_binding: Assertion `first + n <= MAX_GLOBAL_BUFFERS'
> > > failed.
> > > 
> > > [1]    Abbruch                       blender (core dumped)
> > 
> > The number of max global buffers was bumped in 06bf56725d to fix
> > similar crash in luxmark. I guess it needs another bump.
> 
> Hello Jan,
> 
> I'm so blind...
> ...bumping it 48 and 64 (first try) works. 33 not ;-)
> We shouldn't waste to much memory.

Feel free to post a patch. I'm not sure at which point Marek wants to
switch to dynamic allocation (or if at all), but there's no limit in
OCL so we might end up bumping this every time a new app pushes
against the limit.

> Now, let's start with the libclc work.
> Luxmark 'Hotel' is very blocky and Blender 'barbershop_interior_gpu' 
> mostly black. I have some images.
> 
> Shouldn't we better open a new ticket. Any hints for a good name?
> Or do we have one already? I can put my pictures, there.
> Simpler scenes work, but mostly gray (without colors/texture).

Feel free to create a llvm bug for libclc. The best reproducer is
probably OCL CTS convert test failures.

There are several buffer synchronization bugs reported for clover, so
I don't think we need a new one.

sorry for the delay, my day job projects require more time and
attention than usual.

Jan

> 
> Dieter

-- 
Jan Vesely <jan.vesely at rutgers.edu>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 833 bytes
Desc: This is a digitally signed message part
URL: <https://lists.freedesktop.org/archives/mesa-stable/attachments/20190620/5cba0b49/attachment.sig>