#73 Quad precision support

Dreaming
open
5
2011-02-14
2011-01-20
Anonymous
No

With quad precision support implemented in the upcoming GCC 4.6 (http://gcc.gnu.org/gcc-4.6/changes.html), it'd be nice to have it in ATLAS. I don't know if there's a standard interface for quad precision in BLAS, and I don't know about LAPACK interaction either.

Discussion

  • Nope, there's no standard quad precision interface to either the BLAS or lapack. Gcc has long supported a version of quad, that gave you 80-bit (x87 unit) precision rather than 64-bit. Quad precision is one of those straightforward but nontrivial tasks that could be done, but won't until/unless there is a large demand . . .

    Cheers,
    Clint

     
    • assigned_to: nobody --> rwhaley
     
  • Note that quad precision arithmetic is so slow (~100x slower than hardware arithmetic if I recall from my last benchmark), that BLAS operations are likely to be arithmetic-bound rather than memory-bound.

    That means that all of the fancy blocking algorithms etcetera that ATLAS does are probably overkill. You might as well just use the reference BLAS with OpenMP directives to parallelize it.