tape: Add functions to encode and decode float16

This commit is contained in:
Sasha Koshka 2025-07-21 15:58:32 -04:00
parent f009a970cd
commit 5d84636b55
2 changed files with 129 additions and 0 deletions

View File

@ -112,6 +112,13 @@ func (this *Decoder) ReadUintN(bytes int) (value uint64, n int, err error) {
return value, n, nil
}
// ReadFloat16 decodes a 16-bit floating point value from the input reader.
func (this *Decoder) ReadFloat16() (value float32, n int, err error) {
bits, nn, err := this.ReadUint16()
n += nn; if err != nil { return 0, n, err }
return math.Float32frombits(f16bitsToF32bits(bits)), n, nil
}
// ReadFloat32 decldes a 32-bit floating point value from the input reader.
func (this *Decoder) ReadFloat32() (value float32, n int, err error) {
bits, nn, err := this.ReadUint32()
@ -132,3 +139,54 @@ func (this *Decoder) ReadTag() (value Tag, n int, err error) {
n += nn; if err != nil { return 0, n, err }
return Tag(uncasted), n, nil
}
// f16bitsToF32bits returns uint32 (float32 bits) converted from specified uint16.
// Taken from https://github.com/x448/float16/blob/v0.8.4/float16
//
// MIT License
//
// Copyright (c) 2019 Montgomery Edwards⁴⁴⁸ and Faye Amacker
//
// Permission is hereby granted, free of charge, to any person obtaining a copy
// of this software and associated documentation files (the "Software"), to deal
// in the Software without restriction, including without limitation the rights
// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
// copies of the Software, and to permit persons to whom the Software is
// furnished to do so, subject to the following conditions:
//
// The above copyright notice and this permission notice shall be included in all
// copies or substantial portions of the Software.
func f16bitsToF32bits(in uint16) uint32 {
// All 65536 conversions with this were confirmed to be correct
// by Montgomery Edwards⁴⁴⁸ (github.com/x448).
sign := uint32(in&0x8000) << 16 // sign for 32-bit
exp := uint32(in&0x7c00) >> 10 // exponenent for 16-bit
coef := uint32(in&0x03ff) << 13 // significand for 32-bit
if exp == 0x1f {
if coef == 0 {
// infinity
return sign | 0x7f800000 | coef
}
// NaN
return sign | 0x7fc00000 | coef
}
if exp == 0 {
if coef == 0 {
// zero
return sign
}
// normalize subnormal numbers
exp++
for coef&0x7f800000 == 0 {
coef <<= 1
exp--
}
coef &= 0x007fffff
}
return sign | ((exp + (0x7f - 0xf)) << 23) | coef
}

View File

@ -102,6 +102,11 @@ func (this *Encoder) WriteUintN(value uint64, bytes int) (n int, err error) {
return n, nil
}
// WriteFloat16 encodes a 16-bit floating point value to the output writer.
func (this *Encoder) WriteFloat16(value float32) (n int, err error) {
return this.WriteUint16(f32bitsToF16bits(math.Float32bits(value)))
}
// WriteFloat32 encodes a 32-bit floating point value to the output writer.
func (this *Encoder) WriteFloat32(value float32) (n int, err error) {
return this.WriteUint32(math.Float32bits(value))
@ -116,3 +121,69 @@ func (this *Encoder) WriteFloat64(value float64) (n int, err error) {
func (this *Encoder) WriteTag(value Tag) (n int, err error) {
return this.WriteUint8(uint8(value))
}
// f32bitsToF16bits returns uint16 (Float16 bits) converted from the specified float32.
// Conversion rounds to nearest integer with ties to even.
// Taken from https://github.com/x448/float16/blob/v0.8.4/float16
//
// MIT License
//
// Copyright (c) 2019 Montgomery Edwards⁴⁴⁸ and Faye Amacker
//
// Permission is hereby granted, free of charge, to any person obtaining a copy
// of this software and associated documentation files (the "Software"), to deal
// in the Software without restriction, including without limitation the rights
// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
// copies of the Software, and to permit persons to whom the Software is
// furnished to do so, subject to the following conditions:
//
// The above copyright notice and this permission notice shall be included in all
// copies or substantial portions of the Software.
func f32bitsToF16bits(u32 uint32) uint16 {
// Translated from Rust to Go by Montgomery Edwards⁴⁴⁸ (github.com/x448).
// All 4294967296 conversions with this were confirmed to be correct by x448.
// Original Rust implementation is by Kathryn Long (github.com/starkat99) with MIT license.
sign := u32 & 0x80000000
exp := u32 & 0x7f800000
coef := u32 & 0x007fffff
if exp == 0x7f800000 {
// NaN or Infinity
nanBit := uint32(0)
if coef != 0 {
nanBit = uint32(0x0200)
}
return uint16((sign >> 16) | uint32(0x7c00) | nanBit | (coef >> 13))
}
halfSign := sign >> 16
unbiasedExp := int32(exp>>23) - 127
halfExp := unbiasedExp + 15
if halfExp >= 0x1f {
return uint16(halfSign | uint32(0x7c00))
}
if halfExp <= 0 {
if 14-halfExp > 24 {
return uint16(halfSign)
}
coef := coef | uint32(0x00800000)
halfCoef := coef >> uint32(14-halfExp)
roundBit := uint32(1) << uint32(13-halfExp)
if (coef&roundBit) != 0 && (coef&(3*roundBit-1)) != 0 {
halfCoef++
}
return uint16(halfSign | halfCoef)
}
uHalfExp := uint32(halfExp) << 10
halfCoef := coef >> 13
roundBit := uint32(0x00001000)
if (coef&roundBit) != 0 && (coef&(3*roundBit-1)) != 0 {
return uint16((halfSign | uHalfExp | halfCoef) + 1)
}
return uint16(halfSign | uHalfExp | halfCoef)
}