Embree

- Added device concept to Embree to allow different components of an
application to use Embree without interfering with each other.
- Fixed memory leak in twolevel builder used for dynamic scenes.
- Fixed bug in tessellation cache that caused crashes for subdivision
surfaces.
- Fixed bug in internal task scheduler that caused deadlocks when
using `rtcCommitThread`.
- Improved hit-distance accuracy for thin triangles in robust mode.
- Added support to disable ray packet support in cmake.

- Fixed bug triggered by instantiating motion blur geometry.
- Fixed bug in hit UV coordinates of static subdivision geometries.
- Performance improvements when only changing tessellation levels for
subdivision geometry per frame.
- Added ray packet intersectors for subdivision geometry, resulting in
improved performance for coherent rays.
- Reduced virtual address space usage for static geometries.
- Fixed some AVX2 code paths when compiling with GCC or Clang.
- Bugfix for subdiv patches with non-matching winding order.
- Bugfix in ISA detection of AVX-512.

- Major performance improvements for ray tracing subdivision surfaces,
e.g. up to 2× faster for scenes where only the tessellation levels
are changing per frame, and up to 3× faster for scenes with lots of
crease features
- Initial support for architectures supporting the new 16-wide AVX-512
ISA
- Implemented intersection filter callback support for subdivision
surfaces
- Added `RTC_IGNORE_INVALID_RAYS` CMake option which makes the ray
intersectors more robust against full tree traversal caused by
invalid ray inputs (e.g. INF, NaN, etc)

- Added `rtcInterpolate` function to interpolate per vertex
attributes
- Added `rtcSetBoundaryMode` function that can be used to select the
boundary handling for subdivision surfaces
- Fixed a traversal bug that caused rays with very small ray
direction components to miss geometry
- Performance improvements for the robust traversal mode
- Fixed deadlock when calling `rtcCommit` from multiple
threads on same scene

- On dual socket workstations, the initial BVH build performance
almost doubled through a better memory allocation scheme
- Reduced memory usage for subdivision surface objects with crease
features
- `rtcCommit` performance is robust against unset "flush to zero" and
"denormals are zero" flags. However, enabling these flags in your
application is still recommended
- Reduced memory usage for subdivision surfaces with borders and
infinitely sharp creases
- Lots of internal cleanups and bug fixes for both Intel® Xeon® and
Intel® Xeon Phi™

- Improved hierarchy build performance on both Intel Xeon and Intel
Xeon Phi
- Vastly improved tessellation cache for ray tracing subdivision
surfaces
- Added `rtcGetUserData` API call to query per geometry user pointer
set through `rtcSetUserData`
- Added support for memory monitor callback functions to track and
limit memory consumption
- Added support for progress monitor callback functions to track build
progress and cancel long build operations
- BVH builders can be used to build user defined hierarchies inside
the application (see tutorial [BVH Builder])
- Switched to Intel® TBB as default tasking system on Xeon to get even faster
hierarchy build times and better integration for applications that
also use Intel® TBB
- `rtcCommit` can get called from multiple Intel® TBB threads to join the
hierarchy build operations