Hi clint, yes, this is ok, no problem. We need to change this though because we
have regular questions on this. It s good. Proves that someone is looking at
the output of the testing sometimes. It s my second email on the issue of the
week. We will fix this. Soon. So bottom line, we are testing some ill
conditionned pencils and checking the forward error and what you see is machine
epsilon times condition number and this is huge indeed. As expected. I
suggested to either remove the test on the forward or divide the result by the
condition number. More later next week. Julien.
________________________________________
From: Clint Whaley [whaley@Domain.Removed]
Sent: Friday, January 13, 2012 4:02 PM
To: lapack@Domain.Removed; Langou, Julien
Cc: whaley@Domain.Removed
Subject: widespread failures in dgd.out
Guys,
If I build lapack 3.3.1 or 3.4.0 from the tarfiles (with f77blas), and then
do a "make double" in TESTING/, then when I scope dgd.out I see:
DXV drivers: 200 out of 5000 tests failed to pass the threshold
When I look at these errors, many are of size 10^3, which often doesn't mean
error at all, but quite a few are more like 10^7, which usually indicates
a real error.
Is this behavior expected and what is the explanation?
Many thanks,
Clint
**************************************************************************
** R. Clint Whaley, PhD ** Assist Prof, UTSA ** www.cs.utsa.edu/~whaley **
**************************************************************************
|