Routines from Edward Anderson

[jgpallero]

I've seen that from Lapack 3.9.1 onwards some routines of BLAS1 and Lapack were updated with versions from Edward Anderson (https://dl.acm.org/doi/10.1145/3061665). Not only the algorthms were updated, but also the behavior if incX<0, case where in previous versions the subrotine ended as if N<0. This is the case of xNRM2 in BLAS and xLASSQ in Lapack, for example. But I can see that in the functions of Anderson accompanying his paper is also xASUM, which works if incX<0, but in the current version of BLAS in 3.11.0 version of Lapack the classical version from Jack Dongarra (Linpack, 3/11/78) is maintained. Is there any special reason for this?

[langou]

Hi @jgpallero,

No good reason that I know of.

A few "reasons" that I can think of: (1) "Time" to implement, check, write tester (2) usefulness of INCX < 0, (3) reluctance to change the legacy BLAS of 1978, (4) disapproval of allowing INCX < 0, some would rather allow INCX to be an unsigned integer.

Regarding (2), I do not know of many applications that uses INCX < 0. It does seem that INCX < 0 is critical. If someone wants to point to usage of INCX < 0 in existing software, I would appreciate. I think LAPACK uses (used?) INCX < 0 in two places, and these could easily be removed with INCX>=0.

Overall, I also see the interest to have consistency with how INCX behaves. (In particular when it is "easy" to make INCX < 0 happen, so, for example for the BLAS.) Yes, the contributions of Ed Anderson defines behaviors for INCX < 0 that did not exist before.

Julien.