Problems with BLAS Documentation

[bogdan-barbu]

First of all, since BLAST doesn't document the legacy BLAS but instead refers to the original Fortran 77 definition, I've noticed a few inconsistencies which I am going to be a bit pedantic about:

1. The Quick Reference Guide is missing two level 1 BLAS routines that appear in the original publication, Basic Linear Algebra Subprograms for FORTRAN Usage, namely the extended-precision dot products DQDOTI and DQDOTA. They also appear in an older quick reference that can be found on the Web site (see the second page, which is labeled as page 3, since they don't appear directly in the table).
2. The are at least three versions of the CBLAS C header. The first is the one linked in annex B of the BLAST and can be found here. I'm not sure why it contains typesetting instructions at the beginning and end, I guess it was meant to be printed somewhere, but it contains two typos: cblas_chpr and cblas_zhpr take a const float *alpha and const double *alpha parameter, respectively, when these shouldn't be pointers. Given that this header is normative and linked from the official specification, I think this should be fixed. The second header, which I believe was taken directly from an old version of LAPACK, can be found here and actually fixes the previous problem but also has a cblas_xerbla which I believe is a vendor-specific error handler. So long as this is removed, this could probably just replace the header linked from the spec; maybe the extern "C" should also be removed given the spec doesn't care about C++ and that's up to the implementor to add if they want to. Apart from that, the only differences between the two seem to be in terms of whitespace. The last header is the one from the current LAPACK, and has all sorts of LAPACK-specific extensions. At any rate, none of these header seem to have the routines mentioned above, at point 1.

I would also suggest, if there aren't any copyright issues, to maybe post the level 1/2/3 documents on the Web site. The BLAST standard only references them but anyone who wants to implement legacy BLAS with the guidance from Annex B has to go search for this stuff on the Internet on their own. It would be nice to have everything in one place.

[mgates3]

dqdoti and dqdota were indeed in the original paper by Lawson, et al., but are not widely implemented. (Honestly, I'd never heard of them before now.) Part of the issue probably is there is no standard extended precision format, so there's no way to provide a portable interface. Whether Fortran's real(kind=16) or C/C++ long double is a 128-bit IEEE quad, double-double, 80-bit extended float, just 64-bit double, or supported at all, is compiler implementation defined. CBLAS in particular never implemented these routines, and it would be difficult to do in a portable manner.