PATCH: R600 OpenCV fixes

[Tom Stellard]

On Wed, Jan 15, 2014 at 04:29:17PM -0500, Jan Vesely wrote:
> On Tue, 2014-01-14 at 16:24 -0500, Jan Vesely wrote:
> > Hi,
> > 
> > On Mon, 2014-01-13 at 16:07 -0500, Jan Vesely wrote:
> > > Hi,
> > > 
> > > I tested these patches and they fix the piglit test I submitted, however
> > > they seem to introduce stack corruption. See
> > > https://bugs.freedesktop.org/show_bug.cgi?id=73571 for details.
> > 
> > I dug a bit deeper into this.
> > 
> > The third patch introduces (or uncovers) the stack corruption crash
> > mentioned in the above bug report (description).
> > 
> > The other patches are OK (run all gegl tests without crashing). The
> > second patch even fixes what I thought was a rounding issue in one of
> > the gegl tests. If it helps, all but the third one have my
> > Acked-by: Jan Vesely <jan.vesely at rutgers.edu>
> > 
> > I also ran piglit and there is no difference between master, master +
> > 1,2,4,5. Adding patch 3 fixes the newly submitted 'program-scope-arrays'
> > test.
> 
> Finally, I figured out what's going on. After enabling asserts (dunno
> why I ever built without it), the failure changes to:
> unsigned int llvm::MCRegisterClass::getRegister(unsigned int) const:
> Assertion `i < getNumRegs() && "Register number out of range!"' failed.
> 
> The tested kernel uses array of 256 floats. It explains why small test
> (piglit) passes and gegl fails. Sorry for the noise. All patches have my
> acked by.
> 
> Just a question. is it possible to support large arrays or is the proper
> solution to fail kernel compilation?
> 

Thanks for tracking this down.  It is possible to support larger arrays.  We
would have to store them in a scratch buffer, which the compiler doesn't support
yet.

-Tom

> regards,
> Jan
> 
> > 
> > 
> > regards,
> > Jan
> > 
> > 
> > > 
> > > regards,
> > > Jan
> > > 
> > > On Thu, 2013-12-19 at 21:01 -0800, Tom Stellard wrote:
> > > > Hi,
> > > > 
> > > > There patches fix a few bugs uncovered by the OpenCV test suite.
> > > > 
> > > > Please Review.
> > > > -Tom
> > > > _______________________________________________
> > > > llvm-commits mailing list
> > > > llvm-commits at cs.uiuc.edu
> > > > http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
> > > 
> > 
> 
> -- 
> Jan Vesely <jan.vesely at rutgers.edu>