All Level 1 optimizations are carried on in the
BLDdir/tune/blas/level1 directory and its subdirectories. Under this
directory are subdirs with names corresponding to the generic name of the
routine in question (eg. AXPY, IAMAX, DOT, etc). It is in these subdirs that
the user should place the routines to test and time.
A great deal of the performance win to be had on the Level 1 BLAS,
particularly for long vectors, comes from using data prefetch.
ATLAS now includes a prefetch header file (described in
Section 6.1), which makes prefetch instructions for
various systems available for C programmers.