On Wed, Dec 19, 2018 at 12:22 PM Eric Anholt <e...@anholt.net> wrote:
> Jonathan Marek <jonat...@marek.ca> writes: > > > Note: the backend must take care that uniform index is now a float > > This makes me think that lowering ints to float should be done near the > end of the compile (followed by maybe an algebraic and a dce). As is, I > think nir_lower_io() is going to do bad things to dereferences of i/o > arrays. > > That said, it looks like this will be fixing way more than it regresses, > so I would go along with it. > I used to agree but I'm actually inclined to go the other way. Because NIR is typeles, you have no idea in a lower pass whether a particular value is an int or a float. In particular, you have no idea for immediates. Of course, doing it this early would mean that we'd need a version of nir_lower_io that produces floats but that doesn't seem entirely unreasonable.
_______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev
