Half-precision IEEE-754 binary16 float compressed into 2 bytes (1 sign + 5 exponent + 10 mantissa bits, no infinity/NaN encoding), used to halve the memory footprint of large float buffers (e.g. sample or FFT data) at the cost of precision: range +-131008.0, minimum positive normal ~6.1035156E-5, ~3.3 significant decimal digits. Unlike float32, this is a genuinely lossy compression – converting a value in and back out will generally not return the exact original float.
More...
Half-precision IEEE-754 binary16 float compressed into 2 bytes (1 sign + 5 exponent + 10 mantissa bits, no infinity/NaN encoding), used to halve the memory footprint of large float buffers (e.g. sample or FFT data) at the cost of precision: range +-131008.0, minimum positive normal ~6.1035156E-5, ~3.3 significant decimal digits. Unlike float32, this is a genuinely lossy compression – converting a value in and back out will generally not return the exact original float.
There is no assumption of hardware fp16 support: conversion to/from a native 32-bit float is done with portable integer bit manipulation (see half_to_float()/float_to_half()), and all arithmetic (+, -, *, /) is performed by converting both operands to float, computing natively, and returning float – there is no direct half+half instruction path. This makes float16 suitable as a compact storage type, not as a type to run tight numeric loops in.
- Author
- Phil Schatzmann
- Copyright
- GPLv3
