These have stricter pre-conditions than standard mix() and slerp()
- 1) Input quaternions must be unit length.
- 2) The interpolation factor (a) must be in the range [0, 1]
None of these restrictions should be too bad. The reason for these is that it uses fastAcos()
and fastSin(), both of which have a limited allowable range.
In my contrived tests, I observed about a 10x improvement over the standard versions. This is
mostly because of the faster acos/sin operations. The fastSin(__m128) implementation also helps
here because it can do four fastSin() operations simultaneously using SSE (mix() and slerp()
each need three).
A few things here can probably be improved by people a lot smarter then
me, but for the most part things are generally faster.
A few notes:
- A fquatSIMD can be converted to a fmat4x4SIMD using mat4SIMD_cast().
- A tquat<float> can be converted to a fquatSIMD using quatSIMD_cast().
- Some functions are virtually the same as their scalar counterparts
because I've just not been able to get them faster.
- Only the basic functions are implemented. Future plans include fast,
approximate normalize, length and mix/slerp functions.
