No general kernels here

The Level 1 BLAS are in general too basic to be written in terms of simpler kernels. Therefore, each Level 1 routine must be pretty much optimized individually. The only real reuse of kernels comes from either complex-to-real reuse, or one BLAS routine simplifying to another.

For an example of complex-to-real reuse, consider ZNRM2 which, when called with incX = 1, can be simply implemented as a call to DNRM2 with 2*N. An example of a routine simplifying to another would be calling DSCAL with alpha = 0.0, which devolves to a call to the primitive ATL_dset.

Therefore, if you are planning to optimize a particular case, be sure to read the appopriate section below to make sure that the case you want to optimize is not implemented by a call to another routine.

As with other ATLAS optimizations, each routine has its own kernel index file, one for each precision (eg., AXPY/dcases.dsc indexes the various DAXPY implementations that should be tested and timed during the install process). All of these index files follow the format below, though they leave out unneeded parameters (eg., SCAL, which operates on only one vector, will not have an entry for incY or BETA).