- Added operator+(Matrix3x4), operator*(float), RetrieveScaleSq and GetReciprocalScaled to Matrix3x4. Used by Cloth CPU Linear Skinning. These operation will be performant as they use SIMD.
- Modified so there are no virtual functions calls at vertex level.
- Caching indices to simplify the loop when applying skinning.
- Caching static variables Matrix3x4 zero and DualQuaternion zero to avoid creating it for every vertex.
- Removing branching to skip joints when the weight is zero, these cases are rarely and this improves performance by removing branching from loops at vertex level.
- Changing skinning influences so it's a continuous block of memory.
- Caching the vector size() if a variable instead of directly using it in a for loop.
The code that adds a $type field for pointers where needed was still assuming that custom serializer were always for primitives, which isn't the case anymore. This changes updates the behavior to allow $type to be added to those as well as long as they use an object. This does now however rely more heavily on earlier checks that the data needs a $type because it otherwise can't tell the difference between a primitive getting a default value (an empty object). In the original code this situation would have resulted in failed serialization though, so it's unlikely to be a problem.
Aaron Ruiz Mora
AMZN-koppersr
Chris Burel
amzn-mike
Terry Michaels
amzn-mike
greerdv
hultonha
alexpete