Crate minifloat

A const generic library for minifloats

This crate provides emulation of minifloats up to 16 bits. This is done with two generic structs, F8 and F16, which take up to 8 and 16 bits of storage respectively. Many parameters are configurable, including

• Exponent width
• Significand (mantissa) precision
• (F8-only) Exponent bias
• (F8-only) NaN encodings: IEEE, FN, or FNUZ

Note that there is always a sign bit, so F8<4, 3> already uses up all 8 bits: 1 sign bit, 4 exponent bits, and 3 significand bits.

Structs

F8

Minifloat taking up to 8 bits with configurable bias and NaN encoding

F16

Minifloat taking up to 16 bits

Enums

NanStyle

NaN encoding style

Traits

Minifloat

Generic trait for minifloat types

Transmute

Mutual transmutation

Type Aliases

bf16

F16<8, 7>, bfloat16 format

f16

F16<5, 10>, IEEE binary16, half precision