Re: [Gcl-devel] Floating-point performance of GCL? Bug?

[Nicolas Neuss]

Camm Maguire <address@hidden> writes:

> Atlas does not require non-gcc.  None such is available on Debian, and
> numerous packages in this distro make use of atlas.  The L1 blas
> operations you report above cannot be compared with L3 routines -- the
> former will basically be bound by the memory bandwidth for reasonable
> cases at best.  Your 'short' numbers are likely highly misleading, as
> they presuppose finding the entire vector in some level of cache,
> which is very likely not the case on a general call.


OK.  But the Atlas L3 routines get their performance by doing daxpy on
short vectors fitting in the cache.  So I think that DAXPY-short should
yield an upper bound also to L3 performance (given that function call
overhead does not count for daxpying two vectors of length 256).

Thanks, Nicolas.