|
|
|
|
Re: 7811 - Arithmetic Processor - AP-1 [message #380340 is a reply to message #380334] |
Tue, 29 January 2019 20:36 |
|
Originally posted by: R.Kiefer.SPAEM
Michael J. Mahon wrote:
> From a //e usage point of view, the AP-1 and other coprocessor cards like
> it (even the FastMath with its 68881 chip) suffer from an 8-bit data path
> to 64-bit registers, with all the overhead in loading operands and storing
> the result that it implies.
Yes, they do. But the 6888x suffer more than the 9511. Advantage of the
6888x: more MHz and more registers.
With the 9511 you can multiply two 16bit integers by writing 5 bytes to
the 9511, waiting about 45 cycles and reading back 2 bytes after polling
for success. The same with division of two 16bit integers. Means
estimated 115 cycles for these operations.
The 65(C)02 @1MHz is not able to do this job. My accelerator board
(65C02 @12,5MHz from 1987) runs 1437 cycles in the same time. This one
can beat the 9511 in 16bit integer operations but not in floating point.
A multiplication or division of two floating point values needs writing
of 9 bytes to the APU, about 80 cycles for multiplication and reading of
4 bytes. Altogether: 210 cycles. No chance to do that with a 65(C)02
@1MHz.
Cycle count means AppleBus cycles (1MHz).
> As a practical matter, using the card for Applesoft's FP calculations is
> less of a speedup (even with math-heavy code) than a Zip Chip (with which
> most of these cards are incompatible because they use the /RDY line
> (IIRC?).
Yes, the /RDY signal .... I think neither my Transwarp nor my
Schaetzle&Bsteh (12,5MHz) would run beside the 9511 board.
> The alternative design is to use polling to detect when the FP op is
> complete, but that's even slower.
If you run special software for the 9511 you can run the 6502 and the
9511 strict parallel like the 68881 beside the 68020, IMHO.
Regards
Ralf
|
|
|