"jhu" <nomail@nomail> wrote:
> What are your thoughts on this? The upcoming AVX2 instructions have
> gather/scatter and integer support which should make autovectorization easier.
> Do you think Povray will be able to gain appreciable performance from this?
> Would Povray benefit (if at all) from gather/scatter?

One probably needs to have a closer look. "gather" is not a very fast
instruction on Haswell architecture. "scatter" will only be available with Intel
AVX-512 instructions. In case you are hoping to compute a ray in each vector
element, this won't work for complex scenes: Vectorization and branches don't
work well together and each object intersection might result in different
branches. If at all, I would rather hope for acceleration of individual
routines, e.g. subsurface scattering.

Post a reply to this message