Skip to content

Latest commit

 

History

History
98 lines (89 loc) · 3.34 KB

readme.md

File metadata and controls

98 lines (89 loc) · 3.34 KB
Raw

f24

These routines are for a non-standard 24-bit float format. While there is a standard for binary16 and binary32, there is not one for 24-bit floats. This does give me some artistic liberties, but I tried to be sensible :)

Format

The format I went with has 1 bit for the sign and 7 bits for the exponent, and these are packed into a byte. The significand is 17 bits, but the top bit is treated as an implicit 1, and so it is actually encoded as a 16-bit word.

Special values (0, inf, NaN)

The exponent has a bias of +63, or 0b0111111 in binary. 0 is encoded as an exponent of -63, regardless of the significand. inf is encoded as 2^64, and NaN is encoded as 2^64*1.x where x is any non-zero value.

Example values

Example values:

              s eeeeeee mmmmmmmmmmmmmmmm
2.16841e-19   0 0000001 0000000000000000  smallest representable positive number
1.84466e19    0 1111110 1111111111111111  largest representable positive number
1.0           0 0111111 0000000000000000
2.0           0 1000000 0000000000000000
-1.0          1 0111111 0000000000000000
-0.5          1 0111110 0000000000000000
3.1416        0 1000000 1001001000100000
+0            0 0000000 ----------------
-0            1 0000000 ----------------
+inf          0 1111111 0000000000000000
-inf          1 1111111 0000000000000000
NaN           - 1111111 xxxxxxxxxxxxxxxx some non-zero value so as not to be interpreted as inf

Input/Output

I aimed to keep input/output consistent. Floats are stored in registers AHL and CDE, where A and C hold the sign (top bit) and exponent (bottom bits), and HL and DE hold the bottom 16 bits of the significand. AHL always holds the first input (often referred to in these documents as x), and CDE always holds the second floating-point input (referred to as y in these docs). In all cases where output is a float, the result is in AHL.

There are three "strange" routines:

  • f24rand has no input
  • f24cmp takes float inputs, but returns output in the zero flag and carry flag
  • f24div_pow2 takes as second input an unsigned integer in register B.

To summarize:

  • AHL is x, the first operand (if required)
  • CDE is y, the second operand (if required)
  • AHL is the output (if output is a float)

Routines

These float routines are for the most part prefixed with f24 to distinguish them. Most of them are in their own file, but some files contain several routines. These cases are noted in the chart below.

f24abs       |x|
f24acosh     acosh(x)
f24acos      acos(x)
f24add       x+y
f24amean     (x+y)/2
f24asinh     asinh(x)
f24asin      asin(x)
f24atanh     atanh(x)
f24atan      atan(x)
f24bg        1/BG(x,y)  (this is a special-purpose routine)
f24cmp       compare x to y
f24cosh      cosh(x)
f24cos       cos(x)
f24neg       -x
f24div       x/y
f24div2      x/2        (included in f24amean.z80)
f24div_pow2  x/2^k
f24exp       e^x
f24geomean   sqrt(x*y)
f24log10     log10(x)
f24log2      log2(x)
f24logy      log_y(x)
f24log       log(x)     (a.k.a. ln(x))
f24mod1      x mod 1
f24mul2      x*2
f24mul3      x*3
f24mul       x*y
f24pow10     10^x
f24pow2      2^x
f24pow       x^y
f24rand      random number on [0,1)
f24rsub      -x+y       (included with f24sub.z80)
f24sinh      sinh(x)
f24sin       sin(x)
f24sqrt      sqrt(x)
f24sqr       x*x
f24sub       x-y
f24tanh      tanh(x)
f24tan       tan(x)