Comments on: Testing Graphics Code, 4 years later That's an inspiring write up, Aras. That’s an inspiring write up, Aras.

]]>
By: Shaun Kime/2011/06/17/testing-graphics-code-4-years-later/#comment-5751 Shaun Kime Fri, 17 Jun 2011 13:41:26 +0000 On PCs most often it's just like "up to 5 pixels being wrong" on a 400x300 or so image. Usually at object intersections or edges of very thin / small triangles, for some reason we haven't figured out once in a while the GPU would just pick a pixel from another object. When a programmer screws something up, for example breaks some half-texel offset calculation on D3D9, a lot of tests start failing because all 2D/UI stuff is suddenly shifted by almost a pixel (or blurred), and that is very much more than 5 pixels in the whole image. So the test is either "passes" or "fails", and if it fails it needs to be investigated. Of course it might be that the new change is "ok" and expected. For example if we decide to improve shadowmap filtering, then all tests who have that will have different output. A programmer then reviews the new results and commits them into the "template images" as the new "golden truth". On PCs most often it’s just like “up to 5 pixels being wrong” on a 400×300 or so image. Usually at object intersections or edges of very thin / small triangles, for some reason we haven’t figured out once in a while the GPU would just pick a pixel from another object.

When a programmer screws something up, for example breaks some half-texel offset calculation on D3D9, a lot of tests start failing because all 2D/UI stuff is suddenly shifted by almost a pixel (or blurred), and that is very much more than 5 pixels in the whole image.

So the test is either “passes” or “fails”, and if it fails it needs to be investigated. Of course it might be that the new change is “ok” and expected. For example if we decide to improve shadowmap filtering, then all tests who have that will have different output. A programmer then reviews the new results and commits them into the “template images” as the new “golden truth”.

]]>
By: Kevin Gadd/2011/06/17/testing-graphics-code-4-years-later/#comment-5721 Kevin Gadd Fri, 17 Jun 2011 05:51:50 +0000