I don't remember the specifics, but I recall some of the old conformance
tests being sensitive to float->int conversion in a few areas. Matt, if
you have access to the conform tests and can run the old swrast driver with
your change, that'd be interesting.
-Brian
On Fri, Jul 31, 2015 at 7:52 PM
I have some doubts of this. Given that IROUND was just plain incorrect,
I think we should just use _mesa_lroundevenf() and see what happens.
lroundf() is quite the shocker, my libm's implementation is totalling 40
instructions (!) for it (not counting the last return, vs 1 for lrintf).
Now it's pro
lroundf is the most common replacement. I replaced uses of IROUND()
where there was a comment saying "rounded to nearest integer" with
_mesa_lroundevenf.
IROUND64 is replaced with llroundf.
---
src/mesa/main/drawpix.c | 21 +++--
src/mesa/main/eval.c| 14 ++