This commit contains all the work related on the AMD megakernel split work
which was mainly done by Varun Sundar, George Kyriazis and Lenny Wang, plus
some help from Sergey Sharybin, Martijn Berger, Thomas Dinges and likely
someone else which we're forgetting to mention.
Currently only AMD cards are enabled for the new split kernel, but it is
possible to force split opencl kernel to be used by setting the following
environment variable: CYCLES_OPENCL_SPLIT_KERNEL_TEST=1.
Not all the features are supported yet, and that being said no motion blur,
camera blur, SSS and volumetrics for now. Also transparent shadows are
disabled on AMD device because of some compiler bug.
This kernel is also only implements regular path tracing and supporting
branched one will take a bit. Branched path tracing is exposed to the
interface still, which is a bit misleading and will be hidden there soon.
More feature will be enabled once they're ported to the split kernel and
tested.
Neither regular CPU nor CUDA has any difference, they're generating the
same exact code, which means no regressions/improvements there.
Based on the research paper:
https://research.nvidia.com/sites/default/files/publications/laine2013hpg_paper.pdf
Here's the documentation:
https://docs.google.com/document/d/1LuXW-CV-sVJkQaEGZlMJ86jZ8FmoPfecaMdR-oiWbUY/edit
Design discussion of the patch:
https://developer.blender.org/T44197
Differential Revision: https://developer.blender.org/D1200
This commit implements heuristic which allows to skip nodes pushed to the stack
from intersection if distance to them is larger than the distance to the current
intersection.
This should solve speed regression which i didn't notice in the original QBVH
commit (which could have because i had WIP version of this patch applied in my
local branch).
From quick tests speed seems to be much closer to what is was with regular BVH.
There's still some possible code cleanup, but they'll need a bit of assembly
code check and now i want to make it so artists can happily use Cycles over the
holidays.
Old algorithm:
Raytrace from one transparent surface to the next step by step. To minimize
overhead in cases where we don't need transparent shadows, we first trace a
regular shadow ray. We check if the hit primitive was potentially transparent,
and only in that case start marching. this gives extra ray cast for the cases
were we do want transparency.
New algorithm:
We trace a single ray. If it hits any opaque surface, or more than a given
number of transparent surfaces is hit, then we consider the geometry to be
entirely blocked. If not, all transparent surfaces will be recorded and we
will shade them one by one to determine how much light is blocked. This all
happens in one scene intersection function.
Recording all hits works well in some cases but may be slower in others. If
we have many semi-transparent hairs, one intersection may be faster because
you'd be reinteresecting the same hairs a lot with each step otherwise. If
however there is mostly binary transparency then we may be recording many
unnecessary intersections when one of the first surfaces blocks all light.
We found that this helps quite nicely in some scenes, on koro.blend this can
give a 50% reduction in render time, on the pabellon barcelona scene and a
forest scene with transparent leaves it was 30%. Some other files rendered
maybe 1% or 2% slower, but this seems a reasonable tradeoff.
Differential Revision: https://developer.blender.org/D473
This now supports multiple steps and subframe sampling of motion.
There is one difference for object and camera transform motion blur. It still
only supports two steps there, but the transforms are now sampled at subframe
times instead of the previous and next frame and then interpolated/extrapolated.
This will give different render results in some cases but it's more accurate.
Part of the code is from the summer of code project by Gavin Howard, but it has
been significantly rewritten and extended.
