eugene@ames.UUCP (Eugene Miya) (09/13/85)
Laura interpreted correctly what I said about methods in computer
science tending to be sloppy. We have a Cyber 205 here, and three
of us are trying to program a test to determine whether triadic
arithmetic units are cost effective. [see the beginnings of the
Null Hypothesis?] The issues involved include: memory contention,
pipeline startup, symmetry, different operations, system overhead
[does the OS decide to page half way thru your 65K long array?],
the quality of the system clock, fooling potentially smart compilers,
etc.
We've have three days without writing much code. We thought the
test would start as:
Test
Time:
Loop:
T = A * B
D = T + C
Versus
Time:
Loop:
D = A * B + C
where A,B,C,D,T are all contiguous arrays. What factors are extraneous?
What factors are significant? What things can be subtracted out as
overhead? The above test turn out to be too naive. A smart compiler
should recognize the above expression and perform a strength reduction
operation and the times should be equal. What about register allocation?
And so forth. This has become an experiment design of about 5 factors.
Architectures I know using triads include the Cyber 205 and the FPS-series.
It does not yet appear cost-effective in micros or non-"vector" CPUs.
Many of the methods in computer science would leave us with simple
but (naive) tests. This is the iterative (self-correcting) beauty of
the sciences. Oh, for a simpler field :-).
From the Rock of Ages Home for Retired Hackers:
--eugene miya
NASA Ames Research Center
{hplabs,ihnp4,dual,hao,decwrl,allegra}!ames!aurora!eugene
emiya@ames-vmsb