If I run the serial cuda tests on a Tesla 2070 card with the SPFP model,
I get 34 test diffs. To me they all look like roundoff, but I'm not happy
with asking users to decide for themselves, without more guidance from
the developers.
Most of the SPFP tests use a relative tolerance of 1.e-4. If we changed
that to 5.e-4, we would get rid of all but two of the test failures. Is this
a reasonable thing to do? Is there another way to help users new to GPUs to
interpret the test output?
...thx...dac
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Wed Mar 19 2014 - 09:00:03 PDT
