Bug in dlasd4 in 3.4.2

[andreasnoackjensen]

This small program gives either sigma=NaN or the wrong answer 0.5 in LAPACK 3.4.2 but not in previous versions. Also, the answer changes if I comment out the write statements.

program test

implicit none

double precision :: v(5), z(5), delta(5), work(5), sigma, rho, dnrm2
integer :: info

v = (/0.1d0, 0.2d0, 0.3d0, 0.4d0, 0.5d0/)
z = (/0.1d0, 0.2d0, 0.3d0, 0.4d0, 0.5d0/)
rho = dnrm2(5_8, z, 1_8)
write(*,*) "z", z
z = z / rho
rho = rho**2
write(*,*) "v", v
write(*,*) "z", z
write(*,*) "rho", rho
call dlasd4(5_8, 5_8, v, z, delta, rho, sigma, work, info)
write(*,*) "sigma", sigma
write(*,*) "info", info
end program

[lawrence mulholland]

I think this is the problem

dlasd4.f
line 268
TAU = RHO
should be
TAU2 = RHO

Lawrence

PS I think NAG should give somebody who performs the Lapack builds and tests
a NAG compiler licence to find things like this. It shows up right away as an unitialized
variable.

[akobotov]

Hello,

Could you please check the latest version of dlasd4 on NETLIB LAPACK SVN https://icl.cs.utk.edu/svn/lapack-dev/l ... C/dlasd4.f. Very similar issue was corrected there at rev 1337.

Thank you,
Alexander

[lawrence mulholland]

OK looks like latest repo version is quite a bit different and doesn't have the same bug.
Presumably this is a revision for the next release.
Some might still might want the one line fix for 3.4.2 in the meantime.