brad@kontron.UUCP (Brad Yearwood) (01/15/87)
I was waiting for somebody else to more-or-less confirm my finding
before posting this particular note. Another posting received this
evening seems to confirm things sufficiently.
Try the following program:
double a[16];
main()
{
int i;
double p, q;
i = 10;
p = 2.0;
q = 3.0;
a[i] = p * q;
}
Examine the generated code. Unless the late hour was impairing
my judgment, the double multiply library routine returns its result
in D0 and D1 (R0 and R1 - whatever you want to call them). The
effective address calculation for a[i] promptly clobbers one or the
other of these registers before doing the stores of each half of
the result. No wonder my Fast Fourier Transforms and digital
filters gave so many zero results that could be cured by breaking
the computation down into tiny bite sized pieces, carefully
avoiding any array element assignment more complex than:
a[i] = <constant or static variable>
I also observed printf hangups in debugging - possibly the partially corrupted
stores create non-numbers which wedge the conversion routine.
When reduced to tiny bite-sized pieces, however, Alcyon lets the inverse FFT
of the FFT of 200 real data points zero-padded to 256, return very much the
original real data, with only 0.0 and -0.0 imaginary components, and 0.0
or -0.0 in the real components of the 56 original zero-fill points. This
is all using the IEEE package - not the fast package.
My numerical luck with Mark Williams C hasn't been quite so good yet, with
substantial fuzz remaining in both the imaginary components and zero-fill.
They _do_ generate correct code, however.
Win a few and lose a few. Maybe us seriously masochistic types who
like the idea of slow floating point image processing on a budget, really
just need a diversified portfolio of compilers to spread our risks.
Disclaimer: This reflects work of strictly personal interest. Though we are
commercially involved in image processing, my work at Kontron lies elsewhere.
Brad Yearwood
Kontron Electronics {voder, pyramid}!kontron!brad
Mountain View, California
(415) 965-7020