Embree

- Improved performance for streams of coherent (single) rays flagged
with `RTC_INTERSECT_COHERENT`. For such coherent ray streams, e.g.
primary rays, the performance typically improves by 1.3-2×.
- New spatial split BVH builder for triangles, which is 2-6× faster
than the previous version and more memory conservative.
- Improved performance and scalability of all standard BVH builders on
systems with large core counts.
- Fixed `rtcGetBounds` for motion blur scenes.
- Thread affinity is now on by default when running on the latest
Intel® Xeon Phi™ processor.
- Added AVX-512 support for future Intel® Xeon processors.

- Added a new curve geometry which renders the sweep surface of a
circle along a Bézier curve.
- Intersection filters can update the `tfar` ray distance.
- Geometry types can get disabled at compile time.
- Modified and extended the ray stream API.
- Added new callback mechanism for the ray stream API.
- Improved ray stream performance (up to 5-10%).
- Up to 20% faster morton builder on machines with large core counts.
- Lots of optimizations for the second generation Intel® Xeon Phi™
processor codenamed Knights Landing.
- Added experimental support for compressed BVH nodes (reduces node
size to 56-62% of uncompressed size). Compression introduces a
typical performance overhead of ~10%.
- Bugfix in backface culling mode. We do now properly cull the
backfaces and not the frontfaces.
- Feature freeze for the first generation Intel® Xeon Phi™ coprocessor
codenamed Knights Corner. We will still maintain and add bug fixes
to Embree v2.9.0, but Embree 2.10 and future versions will no longer
support it.

- Improved shadow ray performance (10-100% depending on the scene).
- Added initial support for ray streams (10-30% higher performance
depending on ray coherence in the stream).
- Added support to calculate second order derivatives using the
`rtcInterpolate2` function.
- Changed the parametrization for triangular subdivision faces to
the same scheme used for pentagons.
- Added support to query the Embree configuration using the
`rtcDeviceGetParameter` function.

- Added support for setting per geometry tessellation rate (supported
for subdivision and Bézier geometries).
- Added support for motion blurred instances.

- Added support for line segment geometry.
- Added support for quad geometry (replaces triangle-pairs feature).
- Added support for linear motion blur of user geometries.
- Improved performance through AVX-512 optimizations.
- Improved performance of lazy scene build (when using Intel® TBB 4.4 update
2).
- Improved performance through huge page support under linux.

- Internal tasking system supports cancellation of build operations.
- Intel® ISPC mode for robust and compact scenes got significantly faster
(implemented hybrid traversal for bvh4.triangle4v and
bvh4.triangle4i).
- Hair rendering got faster as we fixed some issues with the SAH
heuristic cost factors.
- BVH8 got slight faster for single ray traversal (improved sorting
when hitting more than 4 boxes).
- BVH build performance got up to 30% faster on CPUs with high core
counts (improved parallel partition code).
- High quality build mode again working properly (spatial splits had
been deactivated in v2.7.0 due to some bug).
- Support for merging two adjacent triangles sharing a common edge
into a triangle-pair primitive (can reduce memory consumption and
BVH build times by up to 50% for mostly quad-based input meshes).
- Internal cleanups (reduced number of traversal kernels by more
templating).
- Reduced stack size requirements of BVH builders.
- Fixed crash for dynamic scenes, triggered by deleting all
geometries from the scene.