Alois Steindl said:
>Increasing n to 1024 gives over 200 mflops for the DVF compiler, but
>numerical problems (underflow and overflow) for the Linux compilers.
Hmm, assuming you took the "standard 1000" linpack test program and manually
increased the array sizes (and problem dimension) to 1024, I am not surprised:
this makes it malfunction - overflows or zero-divides - so the Linux compilers
are performing correctly here! (The values in the matrix generated by the
test program are inappropriate the such a problem size with single-precision
arithmetic). Even with non-stop arithmetic you get output like:
norm. resid resid machep x(1) x(n)
NaNQ 0.00000000E+00 1.19209290E-07 -NaNQ NaNQ
which clearly shows there is some sort of problem (this output was not
generated on a PC - I used 2 other machine types to establish that it was
a more general problem).
There are no such problems running the "standard 1000" linpack test program.
Cheers,
--
...........................Malcolm Cohen, NAG Ltd., Oxford, U.K.
([log in to unmask])
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|