a lot of the hierarchical z is tree-style with varying resolution, in ways that might be NDA'd but are relatively clear if you work with the PS4 sdk or see AMD's docs
-
-
Replying to @antumbral
I don’t see how that meaningfully differs from traditional cascaded shadow maps, unless we’re talking about different things? My point is that hierarchical Z is inexact, while irregular Z is exact.
1 reply 0 retweets 0 likes -
Replying to @pcwalton
hi-z is exact if you take all the layers, and inexact if you don't. the inexactness is used for early culling and tests. if you have direct access to the layers you can do clever things with it
1 reply 0 retweets 0 likes -
Replying to @antumbral @pcwalton
it's definitely not the ideal implementation they describe and it's only screen-space, so it doesn't solve the same problem - but in practice it does give you per-pixel level accuracy without needing to store every pixel.
1 reply 0 retweets 0 likes -
Replying to @antumbral @pcwalton
the hierarchical z implementations i've interacted with are basically space partitioning where if the high level (coarse) data is 100% correct the high-frequency z is just dropped.
2 replies 0 retweets 0 likes -
Replying to @antumbral
Yeah, we lean on the hierarchical Z a lot in WebRender actually because we aggressively Z-cull opaque objects and hi-Z is <3 for quick z-reject of large tiles. I want to start doing more aggressive things too; e.g. marking RGBA PNGs that don’t actually have any alpha as opaque
1 reply 0 retweets 1 like -
Replying to @pcwalton
analyzing rgba bitmaps for that got me huge wins on low end hardware in the old days. If you're doing that you can probably identify large transparent margins and kill those cheaply too.
1 reply 0 retweets 0 likes -
Replying to @antumbral @pcwalton
from looking around it seems like all the production hi-z algorithms are NDA'd, which makes sense if they're a competitive advantage. The hierarchical rasterization described here can be mapped to something approaching irregular Z though http://www.drdobbs.com/parallel/rasterization-on-larrabee/217200602?pgno=3 …
1 reply 0 retweets 0 likes -
Replying to @antumbral @pcwalton
oh fyi a few other bitmap analysis things you can fit into the same scan pass that were big wins: bitmaps that are all one color (with or without alpha) -> solid fills or 1x1 bitmaps bitmaps that only have data in one channel or are grayscale -> R8 instead of RGBA32
1 reply 0 retweets 0 likes -
Replying to @antumbral @pcwalton
speak of the devil, the variable rate shading + texture space shading here could probably be used for full blown irregular buffers https://www.nvidia.com/content/dam/en-zz/Solutions/design-visualization/technologies/turing-architecture/NVIDIA-Turing-Architecture-Whitepaper.pdf …
1 reply 0 retweets 0 likes
Semi-related: Learning about the theorem that shows that raytracing and rasterization are formally equivalent was kind of a mind-opening moment for me. It made me a believer in rasterization over raytracing…right as the industry was starting to accept real time raytracing :)
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.