[Octave-bug-tracker] [bug #60273] Wrong result in GSVD

[Rik]

Follow-up Comment #26, bug #60273 (project octave):

@Dmitri: Could you run the test on Ryzen and capture the log?  My guess is the
failures are simply due to the tolerance of the comparison in the assert()
statements.

My own testing showed that LAPACK seems to vary quite a bit with how accurate
the return values are and that the accuracy depends on the characteristics of
the input.  For example, if one of the inputs is rank-deficient then it
matters whether that input is the first one, matrix A, or the second one,
matrix B.  The orientation of the matrix also affected results.  If A was 3x5
accuracy was one value but if it was 5x3 it was another.

When I wrote the BIST tests I updated the tolerances so the tests would pass
on my machine.  I don't think there's any problem with increasing some of the
tolerances by as much as 50 eps.

    _______________________________________________________

Reply to this item at:

  <https://savannah.gnu.org/bugs/?60273>

_______________________________________________
  Message sent via Savannah
  https://savannah.gnu.org/