Hi,
Am 14.02.2019 um 22:03 schrieb DarkZeros mailszeros@gmail.com:
Hi,
It looks to me is a simple clamping issue, it fails just on the borders. But it would be interesting to see if forcing the wrapping fixes it, or that GPU just is unable to wrap around fetch4 sampling.
Yes, after looking at it a bit more I agree. If I force clamping I get the same results on evergreen and Wine (minus results accidentally accepted by broken()). I have attached test results of Polaris and evergreen with addressmode U and V = clamp. My r500 box is just compiling Gentoo updates, so I'll get comparison results there later.
I see two choices: Either accept the Polaris results as broken, or force clamping everywhere to get the same results everywhere and adjust the expected values accordingly. In theory we could also just not test the borders, but I don't see how that would be better than clamping. My preference is option 2 because with all the ways fetch4 is broken the test accepts too large an amount of results already.