float8

FP8 E4M3FN arithmetic library for Go, commonly used in quantized ML inference.

Part of the Zerfoo ML ecosystem.

Features

• IEEE 754 FP8 E4M3FN format — 1 sign, 4 exponent, 3 mantissa bits
• Fast lookup tables — optional pre-computed tables for arithmetic and conversion
• Full arithmetic — add, subtract, multiply, divide, sqrt, abs, neg
• No infinities — the E4M3FN variant uses the infinity encoding for additional finite values
• Zero dependencies — pure Go, no CGo

Installation

go get github.com/zerfoo/float8

Requires Go 1.26+.

Quick Start

package main

import (
    "fmt"
    "github.com/zerfoo/float8"
)

func main() {
    a := float8.FromFloat32(3.14)
    b := float8.FromFloat32(2.71)

    sum := a.Add(b)
    product := a.Mul(b)

    fmt.Printf("a = %f\n", a.ToFloat32())
    fmt.Printf("a + b = %f\n", sum.ToFloat32())
    fmt.Printf("a * b = %f\n", product.ToFloat32())
}

Format

Field	Bits	Description
Sign	1	0 = positive, 1 = negative
Exponent	4	Biased by 7, range [-6, 7]
Mantissa	3	3 explicit + 1 implicit leading bit

Special values: ±0 (exp=0, mant=0), NaN (exp=1111, mant=111). No infinities.

Performance Modes

// Enable lookup tables for faster arithmetic (trades memory for speed)
float8.EnableFastArithmetic()
float8.EnableFastConversion()

Used By

• ztensor — GPU-accelerated tensor library

License

Apache 2.0