|
|
On 5/4/20 6:13 AM, William F Pokorny wrote:
...
>
> So in my mind parametric now use-ably slow, instead of never-use slow. :-)
>
...
Yesterday got some testing infrastructure going for the parametric; I'd
never bothered given I didn't previously use it. Ran a larger matrix of
options including turning precompute completely off and have better
information on lowering the gradient.
Basically you can lower gradients as you increase the precompute depth.
The stuff I have tends to use depths of 8 to 16 say so taking the
max_gradient directly to 0.001 almost always works.
With a sphere, MG 0.001 still works at a precompute depth of 4, but at 3
you get only a partial surface(1) and the MG needs to jump up to 0.5 or
so. With no precompute you need MG 1.0.
(1) - My sphere scene. Results will vary depending on bounding, camera etc.
While I expect most everyone runs with precompute, this makes wrong my
comment the max gradient being defaulted to 1.0 a bad default. Perhaps
not having precompute on by default with a much lower gradient is... ;-)
On the other end, the benefit from lowering max_gradient below 0.001
diminishes with the sphere pretty rapidly once at a depth of say 8. You
can lower MG more at 8, but the gains really slight. Might be a reason
to do < 0.001 for accurate results on complicated shapes, but doesn't
look to be much of one for performance.
So! My updated recommendation would be to start with:
accuracy 0.0005
max_gradient 0.001
precompute 8, x,y,z
over current defaulting. Of course, what you need will depend on what
your are doing equation wise, bounding wise etc.
I've not dug into why the behavior is what it is. The apparent
exponential MG jump a low/no precompute a little surprising to me, but
yep...
Bill P.
Post a reply to this message
|
|