My feeling is that these tests are more obscure than they need to be. I'd rather have a test for each of the derivative variants (fine, coarse and "unspecified"), and test all four the possible positions in a 2x2 tile in each of the tests. If the results vary wildly, use quantization. The point of these tests, I think, is not checking how precisely the video card is doing derivatives, but rather check that (in both the HLSL compiler and SMx -> SPIR-V compiler) we're correctly passing derivation operations over.