31 Aug
2023
31 Aug
'23
9:17 a.m.
For this one it would be raising the last argument to the compare_vec4() call on line 1180 to higher powers of 2 until the test passes.
It turns out this only starts passing with 256 * 1024 ULPs of tolerance, which seems insane to me. I'll try to change the test to avoid interpolation from the magnification filter instead.