Dynamic subdivision on buckets

[SonyBoy]

Just finally upgraded to Vray 5 and extremely happy to see the new implementation of dynamic subdivision on buckets when a render is nearing completion. Kudos to Chaos for implementing this, which will go a long way to help alleviating those annoying last bucket hangs that plague certain renders. I am curious about a couple of things:

1) When a bucket that has already been picked up and has started processing then gets subdivided, are the calculations that have already been started retained, or does that bucket then have to start over?

2) I've noticed so far that it seems certain buckets still don't subdivide. Is there some smart analysis going on under the good to determine whether or not it will be faster to subdivide, or better to stick with the existing processes (obviously the answer to this is dependent on the answer to the first question). Or, otherwise, should this be happening and why?

[radoslav.platikanov]

Hi, glad you like it!

1) The computed samples are retained, nothing is thrown away. Otherwise it would be slower than not splitting at all in a lot of cases.
2) A bucket is split only when a CPU thread is left without any work. One of the currently largest buckets is split but otherwise there's no specific order. There isn't much need for anything more complicated due to 1).

[SonyBoy]

Thanks for the clarifications. I am definitely finding that some buckets still simply don't subdivide -- is that expected behavior and if not what could be the cause?

[radoslav.platikanov]

Can you post a video of this happening?

[SonyBoy]

I would, but the site won't let me upload an MP4.

[radoslav.platikanov]

Maybe try zipping it first, or alternatively upload it to GDrive/OneDrive/etc. or send it to our support.

[SonyBoy]

Well, I'm now unable to replicate this, although anecdotally I'll say that the dynamic subdivision of buckets appears to be a bit.... flaky? I've just watched Vray subdivide the render I was attempting to capture for the purposes of this bug report all the way down to one pixel per proc (!). This, of course, is amazing and highly desirable. But this is the same scene as the one that I previously observed max size buckets failing to subdivide on at all. So, go figure.... Anyway, overall I'm still very impressed with the potential of this new behavior, and if I'm able to replicate a scenario where buckets aren't subdividing as expected I'll definitely try to share the evidence.