> AFAICS it's the opposite, translational gradient (I admit I don't understand
> how it is calculated in pov) but no rotational gradient. A problem in adding
> another gradient is that it will cost a lot of memory (the existing gradient
> already uses 3*3*sizeof(float) for each gather location !). Plus, for low
> error_bound the gather locations get closer and the gradients become less
> usefull (maybe a keywork to deactivate the gradient could be interesting for
> people who want to save some RAM in those cases with many gather locations,
> or for people who use only gather locations from a saved radiosity file)
> Yet, I'll be happy to test your patch :)

IMHO gradients can be removed totally from this implementation.
The reason is we can use better interpolation techniques and
give more oversampling of irradiance gathers in those areas where
variance is bigger. Another speeding up technique is to not use so
much samples during gather in areas partially occluded by nearby
objects... we can obtain this info from average distance to
objects calculated during hemisphere gather.

And one more which would conserve memory.... resign from
floats in normals stored in irradiance cache.. it can be stored in polar
coordinates (only 2 bytes) instead of 3 floats... when 
irradiance map count reaches a few milions... it would be nice
idea to free some memory.

Post a reply to this message