LAPACK/ScaLAPACK Development

E.g.

Code: Select all

program test_dlapy2
    implicit none

    double precision :: myzero, dlapy2, x, y

    write(*,*) 'write 0.0 and hit enter'
    read(*,*) myzero

    x = 0.0d0/myzero

    write(*,*) 'x is ', x

    y = dlapy2(1.0d0, x)

    write(*,*) 'y is ', y

end program test_dlapy2


returns

Code: Select all

 write 0.0 and hit enter
0.0
 x is                        NaN
 y is    1.4142135623730951


This has severe implications because it can cause xlarfg to hang during scaling if alpha=0 and x has a nan. I.e. all code that uses elementary reflectors is vulnerable to this as e.g. reported for the SVD in https://github.com/JuliaLang/julia/issues/21757.

It appears that Intel has fixed this in MKL without upstreaming their fix.

Hi Andreas, thanks for the feedback. This is fixed. See link below. Cheers, Julien.
https://github.com/Reference-LAPACK/lapack/issues/147
https://github.com/Reference-LAPACK/lapack/commit/a1e853bd005a2107a0a43f92782a158a9fd6e56c