ramenhut / half

IEEE 754 16-bit Float

half

IEEE 754 16-bit Floating Point Format

This is a simple 16 bit floating point storage interface. It is intended to serve as a learning aid for students, and is not in an optimized form.

This was designed for the following scenarios:

1. Reduced file storage costs for terrain maps consisting of floating point 3D position, normal, texture coordinates, and color data.

2. In memory support for the half format as supported by NVIDIA and AMD GPUs.

Convention:

 Float32:                     Float16:

 1 bit sign                   1 bit sign
 8 bit exponent               5 bit exponent
 23 bit mantissa              10 bit mantissa
 Bias of 127                  Bias of 15  

Special Values:

 +-Zero:          s,  0e, 0m
 +-Denormalized:  s,  0e, (1 -> max)m
 +-Normalized:    s, (1 -> [max-1])e, m
 +-Infinity:      s, (all 1)e, (all 0s)m
 +-SNaN:          s, (all 1)e, (1 -> [max-high_bit])m
 +-QNaN:          s, (all 1)e, (high_bit -> all 1s)m