[Mesa-dev] Testing OpenCL with gegl

[Jan Vesely]

I readded the mesa list again, as the info might be interesting to
others.

On Sat, 2014-06-21 at 11:50 +0200, Bruno Jimenez wrote:
> On Fri, 2014-06-20 at 17:01 -0400, Jan Vesely wrote:
> > On Fri, 2014-06-20 at 21:58 +0200, Bruno Jimenez wrote:
> > > [snip]
> > I also found a reference to BABL_PATH env var in the babl code, it can
> > provide a list of directories to search for extensions.
> > you can try setting it to proper location. by default it looks in
> > $PREFIX/$LIBDIR/babl-0.1 (set during configure phase), so a symlnk might
> > also be an option
> 
> This was it! Setting BABL_PATH to the path of the extensions lib has
> worked. Thanks.
> 
> Now the tests have started passing, although edge-laplace have hanged
> completely my gpu and I have had to hard-reboot.
> 
> Which tests are failing for you? maybe I can just launch them and see if
> it happens too here.

Right now, the failure count is at 4:
 - clones.xml: r600_blit.c:592:r600_resource_copy_region: Assertion
`u_max_sample(dst) == u_max_sample(src)' failed.
 - gegl.xml: r600_blit.c:592:r600_resource_copy_region: Assertion
`u_max_sample(dst) == u_max_sample(src)' failed.
 - noise-cell.xml: %lsr.iv861 = phi [3 x float]* [ %closest, %
land.rhs.i69.preheader ], [ %60, %for.cond24.i64 ]
Do not know how to replace this instruction with vector op
 - stretch-contrast.xml:

It used to be only noise-cell.xml and stretch-contrast.xml.

noise-cell needs support for large program scope arrays (same reason the
second patch I sent you disables color kernels). it's a known problem in
both llvm and clover. although the original failure was different from
the currently reported one. (might be a new bug)

stretch-contrast needs memory fences (libclc and llvm) I started some
work on libclc side [0], but it'll need more changes on llvm side.

These are known and afaik they affect radeonsi too.
[0]http://www.pcc.me.uk/pipermail/libclc-dev/2014-April/000284.html


assertion failures in clones.xml and gegl.xml are new and introduced by
the opencl mem pool changes. I haven't bisected but
'96a95f mesa: Copy Geom.UsesEndPrimitive when cloning a geometry
program.'
works OK. you might want to have a look as you know the code better.

I also used the two fixes I sent so that the colors.xml won't hang

> > > 
> > > Sorry for any inconvenience I may be causing.
> > 
> > no worries, it's good that someone else is trying to get the thing
> > running on mesa/clover
> 
> Just for the r600g driver I'm afraid, if it's a radeonsi issue I won't
> be able to help, neither if it's in LLVM.

I'm also on r600, and I believe cypress and turks should use much of the
same code paths. so it should be possible for you to get at least the
same pass rate as my gpu.

I'm sure the amd guys would appreciate bug reports :)


regards,
Jan

> 
> Thanks for everything!
> Bruno
> 
> > Jan
> > 
> > > 
> > > Thanks in advance!
> > > Bruno
> > > 
> > > > regards,
> > > > Jan
> > > [snip]
> > > 
> > 
> 
> 

-- 
Jan Vesely <jan.vesely at rutgers.edu>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 836 bytes
Desc: This is a digitally signed message part
URL: <http://lists.freedesktop.org/archives/mesa-dev/attachments/20140621/8cc99773/attachment.sig>