For Metal it's 5% slower too, but if I render more samples it's about 2%. Initially I thought if self-intersection is handled properly now, then the performance penalty is worth it, but it seems…
@Alaska thanks!
Turns out there are self-intersections in the wdas_cloud scene. Not sure how the bounding boxes are generated and handled for vdb, will investigate later.
@Sergey can you check if the current change already causes a slowdown on GPU in the wdas_cloud scene? On Metal there are lots of fluctuations but this scene always seems slower.
@Sergey ok so the concern is that if things are not removed in order then it's not a stack, and if I insert an object next to itself then it's even less a stack. But I think to call it volume…
@Sergey As long as you remove one from the stack, it shouldn't matter which one, because the rest in the stack are going to be added together. But one problem I can see is when the object…
The motivation of the fix is the bug report #124385, where instances are considered separate objects, but become one after realizing instances. Although this PR fixes the artifacts, the appearance…
@lichtwerk 95d11b0d33 should be simple enough to backport? It's just one line + update one image, do you have any difficulty? For me it's just that I've never backported anything so not familiar…
This is because of self-intersection. With "instances" the spheres are considered different objects, the density accumulates at intersections; with "realized instances" there is only one big…
@Alaska thanks a lot for the list, I've updated your reply with the things that I know are tested, and struck through the bullet point where all sub-bullet points are tested. For the texture…