Possible bug in pcgehdrv

[granat]

The do loop

*
DO 10 J = JL, ILO+JA+NB-IOFF-1, -NB
JB = MIN( JA+IHI-J, NB )

in pcgehdrv makes JB such that the following call to

CALL PCLARFT( 'Forward', 'Columnwise', IHI-K, JB, A, I+1, J,
$ DESCA, TAU, WORK( IPT ), WORK( IPW ) )

adresses TAU at position N, even though TAU only consist of N-1
initiated/computed values from PCGEHRD.

This is a possible bug - anyone who knows who wrote this code?

Best regards
Robert Granat

[granat]

The do loop

*
DO 10 J = JL, ILO+JA+NB-IOFF-1, -NB
JB = MIN( JA+IHI-J, NB )

in pcgehdrv makes JB such that the following call to

CALL PCLARFT( 'Forward', 'Columnwise', IHI-K, JB, A, I+1, J,
$ DESCA, TAU, WORK( IPT ), WORK( IPW ) )

adresses TAU at position N, even though TAU only consist of N-1
initiated/computed values from PCGEHRD.

This is a possible bug - anyone who knows who wrote this code?

Best regards
Robert Granat

By some testing we suggest that the line

JB = MIN( JA+IHI-J, NB )

is replaced by

JB = MIN( JA+IHI-J-1, NB ).

//Robert

[Julien Langou]

Hello Robert,
thanks for the patch. Is there an easy to check this by simply changing the HRD.dat file for example?
Julien.

[granat]

Hello Robert,
thanks for the patch. Is there an easy to check this by simply changing the HRD.dat file for example?
Julien.

What we did was to compile the library using the Pathscale compiler as

pathf90 -O0 -g -trapuv

which means that all variables are initiated with NaN and floating point exceptions are trapped. The NAG compiler should have a similar option.

Observe that the following lines should be modified as well:

@@ -183,7 +183,7 @@
IV = IOFF + 1
I = IA + ILO - 1
J = JA + ILO - 1
- JB = MIN( NB-IOFF, JA+IHI-J )
+ JB = MIN( NB-IOFF, JA+IHI-J-1 )
*
* Compute upper triangular matrix T from TAU.
*

Best regards
Robert Granat and Åke Sandgren