johguenther
commented
3 years ago What I found so far:
- default sampling rate looks ok (essential it corresponds to the number of voxels along the ray, i.e. viewing the scaled low-resolution side of the volume renders much faster); if at all it is computed too low
- we currently do not analyse the content of the volume, if is is mostly homogeneous the sampling rate is still high (depending on the potential variation, not the actual one),in the SciVis renderer – the pathtracer will partly exploit homogeneous regions
- however, scaling one dimension of the volume not via the
xfmparameter of the instance, but using the volume'sgridSpacingparameter results in 2x faster rendering (in my reproducer), despite that the same number of samples is taken - writing this down it occurs to me: which volume type are you using? if it is actually not a structured regular, but an unstructured one, then I need to start investigation again... Also, in that case a workaround could be to "bake" that non-uniform scaling into the vertex positions to see whether this helps
StasJ
clyne
For certain volumes the automatically determined sampling rate is unusable. For example, one volume spans the following volume:
and then this volume is scaled in the Z axis by about 100x using the
xfmparameter. The default sampling rate results in a 60x20 pixel rendering taking about 3 minutes. SettingvolumeSamplingRateto anything over 1/500 does not give improved image quality and a 1080p image renders in 2-3 seconds.I suspect the issue is that the automatically determined sampling rate does not take into account the volume's scaling factor.