It's almost 2018... The future of GPUs for Graphics is:
- Realtime raytracing51%
- REYES/per-pixel geometry13.3%
- CS->Graphics22.1%
- Texture-space shading13.6%
420 votesFinal results
18
12
33
Conversation
Replying to
The future? I feel like raytracing is very much the "now" already! I can't wait to see more crazy SDF stuff from and
2
3
Afaik Dreams doesn't raymarch the SDF, but point-splats it.
1
2
it's a mad hybrid these days! we rasterize rough cubes, then raymarch the sdf from there; then we selectively splat strokes on top. so it's all the things :) it wasn't always, but pure splatting left too many tiny holes :)
5
10
67
Aha! You decided to RT too. Was overdraw killing your perf with splatting? (RT = zero overdraw)
2
1
alas no! it's small rasterized brick trees per object. but we (hierarchical) cull the bricks as we generate them, so overdraw isnt too bad. hierarchical rt would prob be competitive but we evolved here as an accident of history. may have more to say
2
7
not much to add: GPU occlusion cull all the things, raymarch remaining bricks front to back with conservative Z test. points for everything was fun compute, but yeah hole filling was tricky, also needed whole other pipeline to deal with with sparse points when up close
you know how many days and I spent on figuring out this problem assuming you guys have solved it? drove us close to insanity.
2
1
so hole filling is actually doable. I really like that idea of moving surface reconstruction into screen space. more challenging is the retaining of shape features.
1
Show replies
Do you occlusion cull based on last frame’s depth buffer or do you iteratively occlusion cull when rendering the frame?
2
2 pass intraframe; ie back project into last frame, use as oracle to split cubes into early and late; then render early ones, then cull late ones against early ones intraframe.
1
1
6
Show replies