Embree

- Improved performance of occlusion ray packets by up to 50%.
- Fixed detection of Clang for CMake 3 under MacOSX
- Fixed AVX code compilation issue with GCC 7 compiler caused by
explicit use of vzeroupper intrinsics.
- Fixed an issue where Clang address sanitizer reported an error in
the internal tasking system.
- Added fix to compile on 32 bit Linux distribution.
- Fixed some wrong relative include paths in Embree.
- Improved performance of robust single ray mode by 5%.
- Added EMBREE_INSTALL_DEPENDENCIES option (default OFF) to enable
installing of Embree dependencies.
- Fixed performance regression for occlusion ray streams.
- Reduced temporary memory requirements of BVH builder for curves and
line segments.
- Fixed performance regression for user geometries and packet ray tracing.
- Fixed bug where wrong closest hit was reported for very curvy hair segment.

- Improved packet ray tracing performance for coherent rays by 10-60%
(requires RTC_INTERSECT_COHERENT flag).
- Improved ray tracing performance for incoherent rays on
AVX-512 architectures by 5%.
- Improved ray tracing performance for streams of incoherent rays
by 5-15%.
- Fixed tbb_debug.lib linking error under Windows.
- Fast coherent ray stream and packet code paths now also work in robust mode.
- Using less aggressive prefetching for large BVH nodes which
results in 1-2% higher ray tracing performance.
- Precompiled binaries have stack-protector enabled, except for
traversal kernels. BVH builders can be slightly slower due to this
change. If you want stack-protectors disabled please turn off
EMBREE_STACK_PROTECTOR in cmake and build the binaries yourself.
- When enabling ISAs individually, the 8-wide BVH was previously only
available when the AVX ISA was also selected. This issue is now
fixed, and one can enable only AVX2 and still get best
performance by using an 8-wide BVH.
- Fixed rtcOccluded1 and rtcOccluded1Ex API functions which were
broken in Intel® ISPC.
- Providing MSI installer for Windows.

- Bugfix in the robust triangle intersector that rarely caused NaNs.
- Fixed bug in hybrid traversal kernel when BVH leaf was entered with no
active rays. This rarely caused crashes when used with instancing.
- Fixed bug introduced in Embree 2.16.2 which caused instancing not to
work properly when a smaller than the native SIMD width was
used in ray packet mode.
- Fixed bug in the curve geometry intersector that caused rendering
artefacts for Bézier curves with p0=p1 and/or p2=p3.
- Fixed bug in the curve geometry intersector that caused hit results
with NaNs to be reported.
- Fixed masking bug that caused rare cracks in curve geometry.
- Enabled support for SSE2 in precompiled binaries again.

- Bugfix in the ribbon intersector for hair primitives. Non-normalized
rays caused wrong intersection distance to be reported.

- Increased accuracy for handling subdivision surfaces. This fixes
cracks when using displacement mapping but reduces performance
at irregular vertices.
- Fixed a bug where subdivision geometry was not properly updated
when modifying only the tessellation rate and vertex array.

- Fixed bug that caused NULL ray query context in intersection
filter when instancing was used.
- Fixed an issue where uv's where outside the triangle (or quad) for
very small triangles (or quads). In robust mode we improved the uv
calculation to avoid that issue, in fast mode we accept that
inconsistency for better performance.
- Changed UV encoding for non-quad subdivision patches to
allow a subpatch UV range of `[-0.5,1.5[`. Using this new encoding
one can use finite differences to calculate derivatives if required.
Please adjust your code in case you rely on the old encoding.