Cycles: Switch to reformulated Pluecker ray/triangle intersection
authorSergey Sharybin <sergey.vfx@gmail.com>
Mon, 27 Mar 2017 15:06:37 +0000 (17:06 +0200)
committerSergey Sharybin <sergey.vfx@gmail.com>
Tue, 28 Mar 2017 15:26:47 +0000 (17:26 +0200)
commit6ea54fe9ffe2b2514990fdf3489ca53d05ce449a
tree3026ce049c43bf8e0d56fbad57e3cbb844444fc4
parent69aa6577b3dfea5d8a6d915fad7fb7650d8b6278
Cycles: Switch to reformulated Pluecker ray/triangle intersection

The intention of this commit it to address issues mentioned in the
reports T43865,T50164 and T50452.

The code is based on Embree code with some extra vectorization
to speed up single ray to single triangle intersection.

Unfortunately, such a fix is not coming for free. There is some
slowdown for AVX2 processors, mainly due to different vectorization
code, which caused different number of instructions to be executed
and different instructions-per-cycle counters. But on another hand
this commit makes pre-AVX2 platforms such as AVX and SSE4.1 a bit
faster. The prerformance goes as following:

              2.78c AVX2   2.78c AVX   Patch AVX2         Patch AVX
BMW            05:21.09     06:05.34    05:32.97 (+3.5%)   05:34.97 (-8.5%)
Classroom      16:55.36     18:24.51    17:10.41 (+1.4%)   17:15.87 (-6.3%)
Fishy Cat      08:08.49     08:36.26    08:09.19 (+0.2%)   08:12.25 (-4.7%
Koro           11:22.54     11:45.24    11:13.25 (-1.5%)   11:43.81 (-0.3%)
Barcelone      14:18.32     16:09.46    14:15.20 (-0.4%)   14:25.15 (-10.8%)

On GPU the performance is about 1.5-2% slower in my tests on GTX1080
but afraid we can't do much as a part of this chaneg here and
consider it a price to pay for more proper intersection check.

Made in collaboration with Maxym Dmytrychenko, big thanks to him!

Reviewers: brecht, juicyfruit, lukasstockner97, dingto

Differential Revision: https://developer.blender.org/D1574
13 files changed:
intern/cycles/kernel/bvh/bvh_shadow_all.h
intern/cycles/kernel/bvh/bvh_subsurface.h
intern/cycles/kernel/bvh/bvh_traversal.h
intern/cycles/kernel/bvh/bvh_volume.h
intern/cycles/kernel/bvh/bvh_volume_all.h
intern/cycles/kernel/bvh/qbvh_shadow_all.h
intern/cycles/kernel/bvh/qbvh_subsurface.h
intern/cycles/kernel/bvh/qbvh_traversal.h
intern/cycles/kernel/bvh/qbvh_volume.h
intern/cycles/kernel/bvh/qbvh_volume_all.h
intern/cycles/kernel/geom/geom_motion_triangle_intersect.h
intern/cycles/kernel/geom/geom_triangle_intersect.h
intern/cycles/util/util_math_intersect.h