Floating point emulation

[Gunther]

I've to write a bit software for an embedded system (80386 CPU, unfortunately without 80387). So, I need a good 32 bit floating point emulation library. Can anyone help?

Gunther

[japheth]

I've to write a bit software for an embedded system (80386 CPU, unfortunately without 80387). So, I need a good 32 bit floating point emulation library. Can anyone help?

I've created such a "library" ( it is a dll in PE format, based on the Open Watcom 387 emulation code ).

However, I'm selling it ( source + binary ).

[Gunther]

I've created such a "library" ( it is a dll in PE format, based on the Open Watcom 387 emulation code ).

That doesn't help much, because the OS is Open DOS.

Gunther

[clive]

Does it have to emulate, or could you just use a floating point library that doesn't use the x87 at all. Watcom, my favorite 386 era 32-bit compiler had MATH3R.LIB and MATH3S.LIB to do exactly this.

[japheth]

I've created such a "library" ( it is a dll in PE format, based on the Open Watcom 387 emulation code ).

That doesn't help much, because the OS is Open DOS.

If "OpenDOS" is MS-DOS compatible, it's no problem. However, a requirement is that your software has to run in 32-bit protected-mode

[Gunther]

Does it have to emulate, or could you just use a floating point library that doesn't use the x87 at all.

Clive, the trick is that there isn't much memory and only a small hard disk. I can't use HLL and therefore, the applications must be written in assembly language. The best solution would be the usage of emulation procedures at the source code level.

However, a requirement is that your software has to run in 32-bit protected-mode

I've to check that.

Gunther

[dedndave]

using a good library can work as well as emulated FP
in some ways, a library is better because it can perform entire functions, rather than emulating single instructions
emulated FP code isn't all that fast   :P
shouldn't be much difference between the two, in terms of size
the main advantage of emulated code is, you can easily run the same code on machines that have an FPU and those that don't

the problem you are going to have is finding it   :bg
everyone is used to having an FPU present

i have a very old 16 bit lib that i wrote back in the 80's
it wouldn't be much use to you, i'm afraid
i may update it someday, just as a math refresher   :P

[clive]

Clive, the trick is that there isn't much memory and only a small hard disk. I can't use HLL and therefore, the applications must be written in assembly language. The best solution would be the usage of emulation procedures at the source code level.

Exactly how small? I'm used to working with 128KB or less in ROM/FLASH, and no hard-drive. It is not as if emulating the 387 instruction set is free of substantial costs in speed/space, and it has to emulate all possible instructions, rather than a well written float library which will actually dump unused functionality. Now you could implement a subset, but that has to be clearly documented, so that someone down the road doesn't assume that because it compiled/linked that it is going to work.

The 32-bit 387 emulation package in Microsoft C 7 had a 16-20KB footprint. Watcom's 387 EMU is pushing close to 20KB. Both are almost certainly hand coded assembler.

[FORTRANS]

Hi,

   Until you find a 32-bit solution, you can try out EM87 on
this site (a TSR).  Not perfect, but it works okay on most FPU
instructions.

EM87 (13 KB)
Version 1.2 by Ron Kimball

http://hp200lx.net/super2.html

Regards,

Steve N.

[Gunther]

I could solve my problem. I need only addition, subtraction, multiplication, and division of floating point numbers; so I've written that small library and it works well.

Gunther

[dedndave]

oh - that's easy stuff - lol
it is things like exponentiation and trig functions that are hard to perfect

[Gunther]

oh - that's easy stuff - lol

Yes, easy to write.

things like exponentiation and trig functions that are hard to perfect

Yep, Taylor series, CORDIC algorithms etc. A hard job.

Gunther