packing IEEE754 single
I want to make a function in pure Lua that generates a fraction (23 bits), an exponent (8 bits), and a sign (1 bit) from a number, so that the number is approximately equal to math.ldexp(fraction, exponent - 127) * (sign == 1 and -1 or 1)
, and then packs the generated values into 32 bits.
A certain function in the math library caught my attention:
The frexp function breaks down the floating-point value (v) into a mantissa (m) and an exponent (n), such that the absolute value of m is greater than or equal to 0.5 and less than 1.0, and v = m * 2^n.
Note that math.ldexp is the inverse operation.
However, I can't think of any way to pack non-integer numbers properly. As the the mantissa returned by this function is not an integer, I'm not sure if I can use it.
Is there any efficient way to do something similar to math.frexp()
which returns an integer as the mantissa? Or is there perhaps a better way to pack numbers in the IEEE754 single-precision floating-point format in Lua?
Thank you in advance.
Edit
I hereby present the (hopefully) final version of the functions I made:
function PackIEEE754(number)
if number == 0 then
return string.char(0x00, 0x00, 0x00, 0x00)
elseif number ~= number then
return string.char(0xFF, 0xFF, 0xFF, 0xFF)
else
local sign = 0x00
if number < 0 then
sign = 0x80
number = -number
end
local mantissa, exponent = math.frexp(number)
exponent = exponent + 0x7F
if exponent <= 0 then
mantissa = math.ldexp(mantissa, exponent - 1)
exponent = 0
elseif exponent > 0 then
if exponent >= 0xFF then
return string.char(sign + 0x7F, 0x80, 0x00, 0x00)
elseif exponent == 1 then
exponent = 0
else
mantissa = mantissa * 2 - 1
exponent = exponent - 1
end
end
mantissa = math.floor(math.ldexp(mantissa, 23) + 0.5)
return string.char(
sign + math.floor(exponent / 2),
(exponent % 2) * 0x80 + math.floor(mantissa / 0x10000),
math.floor(mantissa / 0x100) % 0x100,
mantissa % 0x100)
end
end
function UnpackIEEE754(packed)
local b1, b2, b3, b4 = string.byte(packed, 1, 4)
local exponent = (b1 % 0x80) * 0x02 + math.floor(b2 / 0x80)
local mantissa = math.ldexp(((b2 % 0x80) * 0x100 + b3) * 0x100 + b4, -23)
if exponent == 0xFF then
if mantissa > 0 then
return 0 / 0
else
mantissa = math.huge
exponent = 0x7F
end
elseif exponent > 0 then
mantissa = mantissa + 1
else
exponent = exponent + 1
end
if b1 >= 0x80 then
mantissa = -mantissa
end
return math.ldexp(mantissa, exponent - 0x7F)
end
I improved the way to utilise the implicit bit and added proper support for special values such as NaN and infinity. I based the formatting on that of the script catwell linked to.
I thank both of you for your great advice.
Multiply the significand you get from math.frexp()
by 2^24, and subtract 24 from the exponent to compensate. Now the significand is an integer. Note that the significand is 24 bits, not 23 (you need to account for the implicit bit in the IEEE-754 encoding).
Have you seen this?
I think it does what you want in a slighty simpler way.
链接地址: http://www.djcxy.com/p/15002.html上一篇: 优化的多项式极大极小近似反正切[
下一篇: 包装IEEE754单