#68881 support
3 messages in this thread
Leon:
I just did some bencmark testing using LATTIC C and a loop of 70000
multiplies and divides. The test involves inlined 881 code, standard Amiga
IEEE libraries, IEEE without coprocessor, and FFP. Part of the speed of
the 881 code is due to the small loop used. Part is due to the the pure
speed of the 881/882 processor.
IEEE no 881 4.20 sec
FFP 2.50 sec
IEEE Amiga 1.250 sec
881 inline 0.366 sec
At nearly 10 times the speed, it would be very desireable to have a
compiler switch which would generate 881 requiring code. At twice the
speed, it would be desirable to have IEEE double presision library support.
This does not cover the need for double presision math for several
applications such as
Simulations (particularly continuous or mixed type problems.
Non-linear optimizations.
Dynamic programing.
Linear Algebra.
Need I go on?
>> Gary <<
Gary:
My understanding was that the FFP was on a fast Amiga (020 or 030) was as
fast as IEEE libraries with the 68881. However, if this is not the case
then I stand corrected. Since, Benchmark currently does not have IEEE
support my suggestion is to use the routines from the AmigaTech/Lib 12
which give you access to the IEEE library from Benchmark. Not an ideal
solution, but certainly workable.
>> Leon Frenkel <<
Gary, according to your message the speedup you acheived for inline
IEEE/881 code vs. FFP was about 85%. Note this about the same speedup
(83%) I found for M2Sprint using the LONGREAL type vs. REAL when an 882 is
present. — Art