On 2010-Feb-04 23:56:27 +0000, "Dr. David Kirkby" <david.kir...@onetel.net> wrote: >There is another maths library which can be linked, rather than using >-lm. That at least got around this for the previous case of this.
For that matter, if anyone is aware of a suitably licensed C99 libm, I'd also be interested. Since we're seeing various numeric noise in the double-precision results, what assurance do we have that the 'expected' results are actually correct? Have they been independently verified or are they just what was reported by a particular Sage build? >Ultimately, the FPU uses 64-bit, whereas that in the Intel chip works >to 80-bits internally, but 64-bit when the data is read out. I think >inherently the SPARC chip is less accurate. Well, the x87 FPU natively works in 80-bit mode. On the downside, this means that using C doubles (64-bit) in conjunction with the x87 FPU results in double-rounding - which can also cause problems. AFAIK, only the 32-bit x86 API uses the x87 and the x86_64 API defaults to MMX mode (64-bit). And the SPARC architecture defines a 128-bit 'long double', though I'm not sure what chips implement it. It's (theoretically) possible to deliver correctly rounded 64-bit results using either the SPARC or x87 FPU. Doing so will need diffferent algorithms and the SPARC will probably be slower. -- Peter Jeremy
pgpAzNTi9aES7.pgp
Description: PGP signature
