BLAS: Matrix C is allowed to be uninitialized for beta=0

[juerghutter]

According to BLAS

!> On entry, BETA specifies the scalar beta. When BETA is
!> supplied as zero then C need not be set on input.

This seems not to be supported by COSMA.

[simonpintarelli]

I'm sorry for the late reply. Unfortunately, I'm unable to tell by trying to understand the COSMA implementation. For Tiled-MM it should work.

It is the same for pbgemm in scalapack, if beta=0, C is allowed to be uninitialized:
https://www.netlib.org/scalapack/explore-html/d6/da2/pdgemm___8c_a72ccfb5c9cafc30a11d67a9a964349e9.html

So it would be expected from a user perspective that this is true for COSMA too.

Is it a requirement for CP2K that C doesn't need to be initialized when beta=0?

[simonpintarelli]

Ok, as discussed yesterday, I'll fill matrix C with zeros for beta=0.

[simonpintarelli]

As it turned out #159 did only fix the C interface, while CP2K was of course using the Fortran pbgemm call.
COSTA@2.3.1 should have fixed this now: https://github.com/eth-cscs/COSTA/releases/tag/v2.3.1
also for older releases of COSMA, excluding the C interface.

[simonpintarelli]

I'm sorry for the delay.

I'm repeating the explanation given in cp2k/cp2k#5044:

The unit tests in COSMA had a flaw, falsely reporting NaN results as success, which led me to believe COSTA was fixed.

I've published a new COSTA release (this contains the actual fix for the problem):
https://github.com/eth-cscs/COSTA/releases/tag/v2.3.2

There is also a new COSMA release, there is no user-facing change. But it corrects the unit tests, now correctly checking for NaN (uninitialized matrix C).
https://github.com/eth-cscs/COSMA/releases/tag/v2.8.4