single float to double float api

[jpbruyere]

SIMD instructions target minimum 128 bits registers for floats, so for 2 dimensions vectors, the minimal alignment is 128bits, this makes use of SIMD only possible for 2*64bit float component vectors. So I'm planning to test and maybe move the vkvg api to double float instead of single float to be able to use SIMD.

[jpbruyere]

First tests with double vec2d with _m128d show no improvements with _m_add, mul, div. perfs are even worse (20% slower than vec2 (2*float) and normal arithmetic, load/store overhead for single ops on 2d vectors of doubles cant bring speedup.
SIMD usage must be considered for full algorithm with few load/store ops.

[Volatus]

@jpbruyere Doesn't the compiler do some amount of vectorization at -O3 level? (At least I've seen this happen in C++, not sure about C).

[jpbruyere]

The idea here was to try to optimize atomic arithmetic operations by hands, but I was missing the load/store overheads. I guess a lot of vectorization is already done by the compilers. This subject is a hard one, this tests was my first attempt with simd ops.

[jpbruyere]

Most of the floating point operations will get float types promoted to double, it would maybe be wise to use double in computations where possible...