Does not happen with Branched Path, there it's fine. I am wondering why the Step Size affects this, it should not as this is a homogeneous volume.

Have just done a quick retest on the uploaded file, not sure if the extra info will help - on the intel Xeon/GTX 770 machine from before. I'm looking in the viewport rather than final render, but I'm not aware of differences. There does appear to be a difference between CPU+Branched Path and all other mixes of CPU/GPU & Branched/Normal:

GPU:
I do still see the same artifacts in Branched Path here. Using 4 samples (with x1 in all specific categories) and a step size of 1 I can see a dozen bright spots around the circle, more do seem to appear with more samples. Using a step of 0.01 (rather extreme) and watching the render one sample at a time the artifacts do show up in the first few samples (in the same pixels as before), but are much dimmer and get cleaned up by the averaging across samples very quickly.

This looks to be consistent with what I'm seeing in normal path tracing with the same sample/step count.

CPU:
If I switch to CPU rendering, path tracing behaves exactly the same and shows up these artifacts. Branched Path as you say looks flawless, at any step size I've tried (0.01, 0.1, 1 are all fine). Decidedly weird.

Idle musing: I did wonder (as a programmer but blender newb) if it's running into a numerical inaccuracy, missing the geometry at the intersection and assuming some "maximum" brightness over the volume step when this issue occurs, so a longer step contributes more overall light to the pixel than a short step.

It's most likely a ray intersection precision issue, where we miss an intersection and as a result assume the volume extends to infinity. This means it will hit the Max Steps, and with a larger step size that distance will be larger, leading to more emission.

The precision issue would be fixed ideally, but perhaps it's also a good idea to discard volumes segments with no finite exit point.

This actually happens with both branched and regular path tracers here. We talked with @Thomas Dinges (dingto) about roots of the issue in the IRC, but worth dumping some thoughts here.

Basically, what's most likely is happening is wrong volume stack when ray hits close enough to the intersection. This could happen due to two reasons:

1. Ray leaves first volume and ray_offset() pushes ray inside second volume, which leads to a missing volume enter event.
2. Ray enters first volume, keeps traversing and hits exact (in machine precision terms) intersection of two volumes. It's undefined which volume will be picked up here, could be any of them. This will either lead to a missing volume enter event (similar to 1) or will lead to missing volume exit event (causing bright spots).

Surely discarding volumes which are extended to infinity will solve bright spots from appearing, but it's just a partial problem caused by precision issues. We still wouldn't be able to deal with missing volume enter event.

Not sure we'll technically consider this a bug tho, precision issues happens here and there..

This issue could also be fixed by recording all transparent intersections along a ray instead of just the first. We've discussed that as an optimization before, but it would also help with this precision issue.

Seems some help here is needed. So here's a quick patch which implements the idea of ignoring volumes with no exit point (one of the possible ways to go with that at least):

But that opens other types of questions, like whether volume was never exit since we missed some event or because we've run out of camera clipping distance and actual exit will happen later and currently it's all fine.

As for recording all transparent intersections: we talked it about it in context of camera rays. While it's possible to get all transparent intersections for camera ray in regular integrator, there'll be some open issues still:

• Implementing intersect_all for GPU will mean bumping memory usage quite a bit again, alternatively GPU will have the artifact unsolved
• It wouldn't work for secondary rays

So perhaps it should be both: recording all transparent intersections for camera rays and then still have volume stack "validation", so secondary rays and GPU are still free from really visible artifacts.

@Thomas Dinges (dingto), are you actually looking into the report still?

Haven't looked into this since last years EGSR. Still think it's worth to implement "intersect all transparent" for camera rays.

That wouldn't solve the root of the problem -- you'll still have fireflies when secondary ray hits those points. It is still required some trickery to avoid infinite volumes.

If you want i can steal this report and continue with those tricks i've mentioned above.

After looking a lot in the behavior of volume with different camera setups (especially, when clipping distance is inside the volume) came to conclusion current render results aren't really correct anyway (at least from own tests). So perhaps introducing some behavior difference is not that bad.

Finished the proposed change in D2212, let's move discussion there.