On Wed Apr 26 14:21:51 2023 +0000, Francisco Casas wrote:
It is meaningful, notice that, on the tests where there are more than one sampler, I am sampling `float(0.22, 0.22)` which only gives a rounded value for the `filter point point point` sampler, because it does a nearest neighbor sampling.
My problem is that if I change `point point point` to `linear linear linear` the test still passes, at lest with the WARP driver in my Win 11 virtual machine (for both instances of `point point point`). If I change 0.22 to anything strictly larger then 0.25, then it doesn't pass anymore. I'm not sure exactly how sample coordinates are mapped to the texture, but it would seem that for too small coordinates the sampling is considered "on the border" instead of between the two samples, so even if it's linear it doesn't really do a linear interpolation. If you change 0.22 to 0.3 that seems better.