[Bf-cycles] Volumetric patch update
kartochka22 at yandex.ru
Sat Nov 2 08:21:23 CET 2013
Few bug fixes, one so bad i must note here details. I forgot to
initialize backward volume pdf, and as result MIS was turned off for
cases when path length > 1 segments, in other words for any path more
complex then direct light. It related only to bidirectional mode. Now
few scenes rendering (noise cleaning) speed much more close to unpatched
Cycles or Maxwell "draft" mode. Unfortunately Maxwell render have no
ability to limit bounces to 2-3 (all my tests use less then 4 bounces),
and i have less experience how they look in 8+ bounces, all shadows
become way different, so maybe i got placebo effect. But simple "light
behind corner" scenes definitely got boost.
Of course i was testing that new MCMC feature a lot. Result is mixed.
Pure bidir mode (Mutations=0) is rock solid, always converge to right
image. In opposite, MCMC (Mutations > 0) is sometime fantastically
quick, but result more unpredictable up to unusable. Despite that bad
hack i use to get it work w/o changing tile manager (and as result image
brightness depend on luck how many threads you get in every frame. Yes
it that bad.) , it have additional huge instability in overall picture
brightness for another reason that i failed to find.
And Achilles' heel of all MCMC, as walk can stuck in one pixel burning
it out, if seed hit to in small isolated islands surrounded by black
area that dispersion not enough to jump over, looks very bad
(fireflies). There are many tricks can be used to mitigate it, need more
tests to do it right. For now it will be good idea to keep mutations
low. 250 as Eric Veach suggest in dissertation, or even 25. Looking how
Luxrender have very high "big jump" probability by default, 0.4, i think
it is for same reason. But it degrade efficiency, replacing smart part
of walk by more dumb uniform random.
Looking on current state, maybe it will be good time to get more testing
as experimental feature, more ppl will experiment, earlier we get answer
is it good thing or not, i still not sure. Big fight between PT on GPU
and Bidir on GPU on same input data not happens yet. Showstoppers are
still same asymmetric BSDF (smooth normals/bump unusable), that tile
manager hack (need to be completely rewritten, with proper two new
CUDA/OpenCL kernels for initial seeds and other for main MCMC walk
loop), MIS logic simplification (have idea to handle camera and light
surface uniformly as bounce nodes, it will make code much cleaner and
less lines, but need some work), decrease insane number of data i saving
copying just for nothing. I will try to remove leftovers of debug
things, do not promise more.
-------------- next part --------------
A non-text attachment was scrubbed...
Size: 80839 bytes
Desc: not available
Url : http://lists.blender.org/pipermail/bf-cycles/attachments/20131102/99ceec73/attachment.bin
More information about the Bf-cycles