On Thu Aug 31 17:51:27 2023 +0000, Matteo Bruni wrote:
256K ULPs, which is crazy to me. See https://gitlab.winehq.org/wine/wine/-/merge_requests/3172#note_43940 Assuming there isn't some other explanation for the huge discrepancy, I think we should simply consider QXL not a suitable test platform. In this particular case, though, testing filtered sampling is not the point, while testing shader translation is, and the updated test does a better job of that as well IMO.
Oh, ha ha, you know what it is? QXL is giving us 2.47058839e-001, and:
```
0x3f / 0xff
0.24705882352941178 ```
It's rendering into an R8G8B8A8_UNORM render target and then converting that *back* to float. That can't possibly be in spec.