John Travers <[email protected]> writes:

> Thanks for this. It wasn't obvious to me. I recompiled and get much
> better performance.

Great.

> OK, this solved my problem. Although I couldn't add zeros on the
> diagonal (the software I use to write the sparse matrices optimises
> this out), so I put a diagonal of 1e-300. Do you think this will cause
> inaccuracies in the calculations (so far it seems OK).

Anything at least 1e-16 smaller than the other entries is essentially
0.0 as far as floating point arithmetic is concerned.  One possible
concern with 1e-300 is creating floating point exceptions due to
denormals.  For example, if you get a NaN and then turn on trapping FP
exceptions (-fp_trap in PETSc) and try to debug, you might get a false
positive when the 1e-300 creates a denormal.

Attachment: pgpTgMIzLAAZt.pgp
Description: PGP signature

Reply via email to