[Mesa-dev] nir isinf vs NAN handling
Dave Airlie
airlied at gmail.com
Thu Mar 16 06:04:28 UTC 2017
radv has failed the SPIR-V isinf tests for a good while, I've talked
to Matt and the llvm backend matches on an ordered feq to infinity,
where as we map the nir feq to unordered.
06:00 < arsenm> airlied: unordered eq for nan input would say true. isinf for
NAN says false
So it looks like we should be emitting an unordered feq to the llvm
backend in order to get the right result, I've written a patch to
provide an isinf opcode so the backend can lower it itself, would this
be acceptable, any ideas why nobody else notices?
Dave.
More information about the mesa-dev
mailing list