On Sun, Apr 28, 2013 at 11:58:00PM -0400, Erik Schnetter wrote:
This is not really news, given that all test cases are run automatically after every commit: https://trac.einsteintoolkit.org/; look for "Status".
You are right. This has been steady for some months now, in no small part due to these automated tests - thanks to Barry and Ian. I was referring to a more distant past, when it was a bit like this:
http://geek-and-poke.com/2011/08/hudson-status-monitor.html
Speaking about tests: I did a quick test on spine (the machine that currently doesn't show failures). This uses gcc, and I only changed the optimization from -O2 to -Ofast. The result is some testsuite failures, most of them the 'usual set': Exact, GRHydro: weno, Symmetry thorns.
-Ofast (apart other things) uses some non-conforming math, something that is not default for gcc. The intel compiler, on the other hand, does some non-conforming math by default (but very likely not the same subset as gcc with -Ofast). Part of the testsuite failures could just be due to this: different compiler defaults and/or options. Also here: this is not news. We are likely not talking about a lot of bugs in the code, but more likely about how much inaccuracy we are willing to trade for speed.
Frank