Depth pyramid conservativeness

[eqxba]

Hi! I'm implementing 2-pass HZB occlusion culling and while inspecting your implementation I've noticed that your depth pyramid might not always be conservative, so I'd like to clarify if it's indeed so or I understand it wrong.

Let's examine simplified 1D example that I've been thinking of:

1. Let's say you have original depth target of an odd dimension - let it be 7
2. Your depth pyramid mip0 dimension will be the previous pow2(7) = 4
3. In the image attached, mip0 texels are numbered in black at the top, mip1 texels are numbered below the UV axis in red and their centers are marked as red dashes on the UV axis
4. Reduction sampler for texel 1 of mip1 (let's call it mip1[1]) will sample position (1 + 0.5) / 4 = 0.375 and so effectively will return min(mip0[2], mip0[3]), but we can see that not only mip0[2] and mip0[3] texels contribute to mip1[1], but also mip0[1] does.
5. Same is the case for mip1[2]

Could you please comment on that?

[zeux]

I believe you are correct - there could in theory be small single pixel errors due to how the power of two version is constructed initially. The mips themselves should be conservative (in the sense that they shouldn’t amplify the error). The motivation for downsampling into a power of two target was to make it easy to do a conservative reduction as if the same issue happened during mip construction, the results would likely be visible in practice.