v10

Small float support for DL/ML (no breaking changes):

• fixed and significantly accelerated "to/from" scalar conversions
• implemented accelerated path for 4-components vectors (3x-4x faster than scalar, SSE4 required)
• implemented accelerated path for 8-components vectors (4x-6x faster than scalar, AVX2 required)
• added emulation for some missing AVX2 intrinsics
• added new types float8_e4m3_t[2, 4, 8], float8_e5m2_t[2, 4, 8] and float16_t8
• added conversion operators to convert to/from a small float

v9

Avoided potential NaN in Geometry::ReconstructViewPosition

v8

Fixed potential FXC warning "loop only executes for N iteration(s), forcing loop to unroll"

v7

Various fixes for GCC/Clang/Arm64

v1.4

Renamed internal swizzling macros to avoid potential conflicts with user code

v1.3

Official release