On Tue, Sep 16, 2014 at 1:36 AM, <Graeme.Winter@diamond.ac.uk> wrote:

Quick nag if I may… if people have tests which employ random inputs (e.g. random unit cell, random amplitudes and phases) please could they set a seed or select from some seeds which should give random values which work, or fix their tests so that all random values work (ideally!)

 

That way when tests fail people know something is wrong, rather than everything is fine but the random values drawn are troublesome…


Agreed.  We have a large set of tests in the separate phenix_regression directory that are machine-specific but anything that's part of the CCTBX repository needs to be portable.  (And ideally, this should include most of the CCTBX-only tests.)  The test errors as of a couple of nights ago are here:

http://cci.lbl.gov/cctbx_build/show_error_summary.cgi?build_tag=2014_09_14_2205

At least one of these appears to be systematic but there are several purely numeric differences on specific platforms.

-Nat