870646c@aucs.UUCP (11/25/87)
I am going out on a limb here, there is a fellow here at school from Germany that says that in Germany there are upgrades to allow you to install a 68881 math chip in the present ST's. Well I thought I would get ahold of Atari and ask them, well it seems that Michtron in Germany is indeed offering this upgrade??? Could someone out there please verify this to be true. And if it is true what are the chances of some the compilers being upgraded to use the upgrade(it sure would make CAD 2.0 fly). later Barry
ws1i+@andrew.cmu.edu.UUCP (11/30/87)
Heck, if you can get this upgrade who cares about the compilers! I'll write my own C library so that I can use it! Does anybody know anything about it?
t68@nikhefh.UUCP (Jos Vermaseren) (12/02/87)
In article <sVgPeNy00W0syMM0Z7@andrew.cmu.edu>, ws1i+@andrew.cmu.edu (William Manchester Shubert) writes: > > Heck, if you can get this upgrade who cares about the compilers! > I'll write my own C library so that I can use it! Does anybody know anything > about it? In Germany there are two companies making a plug in 68881 board. You plug the (little) board on top of a chip and put some connectors somewhere else and presto. I only know the name of one of these companies: Weide elektronik Regerstr. 34 D-4010 Hilden W.Germany They charge something around DM1000 for it. The 68881 is memory mapped for its communication so one cannot get the same speed as with the link between the 68020 and the 68881. The 68881 runs at 16Mhz. Most of the time is spent in sending numbers to and from the coprocessor. A good gain can be obtained when special dedicated assembler routines can perform subtasks in the registers of the coprocessor or when many built in functions ( sin and cos etc ) are used. For a computational program the typical gain is a factor three to four over a software library, but this can run up to a factor 50 for a benchmark like the Savage test. It will indeed be necessary to do some hacking inside the compilers library as the only software support that is provided with the package is in the form of a set of external routines so that a multiplication is done via c = mult(a,b) rather than c = a*b which is a royal pain in the neck and renders the thing almost useless. A decent hacker has an easy time to disassemble the c/fortran/pascal library of his favourite compiler and replace the appropriate routines by calls to the coprocessor. There are complications though: some compilers don't use IEEE format: Most notably the Mark Williams compiler. Both the Fortran compilers ( ProFortran and AC/Fortran (=Absoft) ) use IEEE and so does Lattice C. I assume that the ProPascal package also uses IEEE. Good Luck. Jos Vermaseren T68@nikhefh.uucp