Would like to see DOPLA, at least, maintained.

[whaley]

Guys,

I saw that the timing package has been eliminated from LAPACK. That's unfortunate, but I understand the economy of the thing. I would like to suggest that DOPLA be re-introduced, and kept in the package regardless of the other timer routines.

DOPLA provides operation counts for most LAPACK routines. As far as I know, it is the only repository that lists them all (though I see that the single precision, which can be used to compute the complex, is available in LAWN41, appA). I fear that as LAPACK changes, paper documents will not be maintained, but perhaps DOPLA could.

I just published a paper where I used FLOP rates. Because I was surveying several LAPACK routines with complex counts, I reference DOPLA for the full FLOP count, and had to find it at a mirror site, since it had been deleted (in the last week or so) from netlib!

I know that we have non-deterministic algorithms in LAPACK for which we cannot provide flop counts. What I'd like to see is even this documented in DLOPA: eg. a comment that the FLOP count is unknown, and maybe a zero return value.

It is easy to think people know FLOP counts, but frankly I didn't until I looked them up in DOPLA for this paper, and I'm probably more informed on these issues than most users. I didn't even know about LAWN41 until I searched this list . . .

Anyway, perhaps it is too odd to have a FLOP counter with no timing routines; if so, I think there needs to some official documentation that is kept up to date that lists this vital information; without it, we get people using opaque flop counts, and HPC timings become even less useful than present.

Thanks,
Clint