Hi,
Last fall we found that the built-in software floating point library in CodeWarrior was too slow for our robot project, and that we couldn't make controller code changes fast enough with fixed point - too many bugs.
So we wrote our own floating-point package in assembly, kind of like the Fast-Float for the AD Blackfin. It isn't IEEE compliant, it has a 16-bit mantissa (fraction) and 8-bit exponent. Not as much resolution, but much faster. It is written mostly in assembly for the 56F83xx, and uses instructions not recognized by the 56F80x, so probably not compatible without significant rework.
Below is a list of the functions so far, and processor cycle counts. (Cycle counts were done using the CW 7 software simulator, with IEEE single float and int for comparison.) Let me know if there is any interest. I don't know how it would be made available at this point, but if enough people want to use it I'll see what I can do. If you're from Freescale, is there any chance this could be integrated in some way with CodeWarrior? The main problem with it now is that all the operations have to be done as function calls, which gets cumbersome for complicated equations. And, constants have to be pre-converted and written in hex (although macros can help).
Function Name Cycle Count float cycle count int cycle count
FFadd 53 254 6
FFsub 64 264 6
FFmult 41 230 8
FFdiv 87 335 40
FFsin 1061
FFcos 1050
FFatan 808
FFgt 46 157 18
FFgte 46 141 12
FFgtz 22
FFlt 46 164 12
FFlte 46 148 12
FFltz 22
FFeqz 21
S16int2FFloat 22
FFloatRnd2S16int 29
FFloatTrunc2S16int 29
S32int2FFloat 25
U32int2FFloat 27
IEEE2FFloat 50
FFloat2IEEE
FFneg 35
FFabs 35
