dymokomi/luce-js
Source of release 0.6.1
0.6.1 / src / cutils / floats.lucb
#==============================================================================================
#
# floats - Float bit reinterpretation and IEEE half-precision conversion
#
# DESCRIPTION:
# Port of QuickJS cutils.h lines 381-455 (float64_as_uint64, uint64_as_float64,
# fromfp16, tofp16, isfp16nan, isfp16zero).
#
#==============================================================================================
## The bits of `d`.
pub func float64_as_uint64(d: f64) -> u64:
return d.bits()
## The double whose bits are `u64`.
pub func uint64_as_float64(u64_bits: u64) -> f64:
return f64.bits(u64_bits)
## The double value of the IEEE binary16 `v`.
pub func fromfp16(v: u16) -> f64:
var v1: u32 = v & 0x7fff
if v1 >= 0x7c00:
v1 += 0x1f8000 # NaN or infinity
let d = uint64_as_float64(((u64)(v >> 15) << 63) | ((u64)v1 << (52 - 10)))
return d * uint64_as_float64((u64)(1023 + 1008) << 52) # 0x1p1008
## The IEEE binary16 nearest `d` (round to nearest even; NaN becomes 0x7c01).
pub func tofp16(d: f64) -> u16:
var a = float64_as_uint64(d)
let sgn = (u32)(a >> 63)
a = a & 0x7fffffffffffffff
var v: u32
if a > 0x7ff0000000000000:
# nan
v = 0x7c01
elif a < 0x3f10000000000000: # 0x1p-14
# subnormal f16 number or zero
if a <= 0x3e60000000000000: # 0x1p-25
v = 0x0000 # zero
else:
let shift = 1051 - (i32)(a >> 52)
a = ((u64)1 << 52) | (a & (((u64)1 << 52) - 1))
let addend = ((a >> (u64)shift) & 1) + (((u64)1 << (u64)(shift - 1)) - 1)
v = (u32)((a + addend) >> (u64)shift)
else:
# normal number or infinity
a -= 0x3f00000000000000 # adjust the exponent
# round
let addend = ((a >> (52 - 10)) & 1) + (((u64)1 << (52 - 11)) - 1)
v = (u32)((a + addend) >> (52 - 10))
# overflow ?
if v > 0x7c00:
v = 0x7c00
return (u16)(v | (sgn << 15))
## Whether the binary16 `v` is a NaN.
pub func isfp16nan(v: u16) -> bool:
return (v & 0x7FFF) > 0x7C00
## Whether the binary16 `v` is a zero of either sign.
pub func isfp16zero(v: u16) -> bool:
return (v & 0x7FFF) == 0