[Bf-cycles] GPU future

[storm]

> 
> - Is there limitations for implementing complex shaders, like real
> SSS?
Nothing that really hard, I want to do brute force approach first, other
ppls seems not very interested in hacking classic fast but less accurate
variant as dipole or other. More like lack of manpower/interest then GPU
limitation.
> - Is it possible/usefull an hibrid CPU+GPU aproach?
Good question. Personally, i hate hybrid and prefer to offload as much
as pissible to GPU. Internally, modern GPU have huge bandwith between
GPU<->VRAM, and much slower system RAM<->VRAM path. But selected
hardware (AMD :) )can have limits, and force to use hybrid model, as
Luxrender show.
> - It will ever be OSL in GPU?
Not in near future, not sure about Sony vision.
> 
> In general, what is the horizon for GPU in Cycles?
Very bright, but need a lot of work and tests. All modern GPUs have one
or other limited memory mapping support, unfortunately it very different
from usual virtual memory management and require many quirks and
workarounds. Some sort of dynamic BVH is core of that, with swap-like
strategy to keep hot "cells" of BVH in GPU and offload cold "cells",
with more cache friendly QMC sampling maybe.