4 ms·
Presumably the WGPs are working on rays which start close to each other in space, which causes the potential occupancy problems mentioned. I wonder if you could
by mastax 3y ago
Presumably the WGPs are working on rays which start close to each other in space, which causes the potential occupancy problems mentioned. I wonder if you could shuffle the rays around a bit to help occupancy at the possible expense of cache hit rate.
- d_tr 3y agoWhy would shuffling the rays around help with occupancy? I 'd also guess that worse locality could easily obliterate any other benefits, but you never know unless you test...
- mastax 3y agoI'm working off the hypothesis in the article: > I wonder if RT work is a lot less predictable than rasterization workloads, making workload distribution harder. For example, some rays might hit a matte, opaque surface and terminate early. If one shader engine casts a batch of rays that all terminate early, it could end up with a lot less work even if it’s given the same number of rays to start with.
- d_tr 3y agoYou would end up with shader engines working for longer on average but with more "dead lanes" then, with no real benefit, right? Sounds like more flexible scheduling would ultimately be needed.