On 6/4/2012 5:37 PM, Lukas Reichlin wrote: > On 03.06.2012, at 21:45, Sergei Steshenko wrote: > > > > I would not call your ATLAS defective. For various reasons, results from > ATLAS and Reference BLAS/LAPACK can differ. Remember that ATLAS stands for > _Automatically_ _Tuned_ Linear Algebra Software :-) > > Your results from the failing tests are probably not wrong: Just compare the > observed and the expected state-space models. There are several ways to do > this: > - poles and zeros should be the same > - Hankel singular values should be the same > - plots from time and frequency responses look the same > - solve the equations below for the diagonal transformation matrix T, its > entries should be -1 and 1. (Because in your example, the absolute values are > the same) > > Aexp = T \ Aobs * T > Bexp = T \ Bobs > Cexp = Cobs * T > > To sum up: if there is a state transformation between observed and expected > model, your results are correct. > > Remember that the SLICOT authors recommend reference BLAS/LAPACK and use it > for their test cases (which I use for test_control). If confidence in correct > results is more important to you than a few milliseconds speed advantage from > automatic tuning, then you should follow their advice. > > Hope this helps, > Lukas >
Lukas, I had same of Sergei's faults (2 of 3) and I am using the reference blas/lapack 3.4.1 I expect a package test function to behave reasonable correctly for most of the system. Borderline case or alternative valid solution should be avoided. Regards Marco ------------------------------------------------------------------------------ Live Security Virtual Conference Exclusive live event will cover all the ways today's security and threat landscape has changed and how IT managers can respond. Discussions will include endpoint security, mobile security and the latest in malware threats. http://www.accelacomm.com/jaw/sfrnl04242012/114/50122263/ _______________________________________________ Octave-dev mailing list Octave-dev@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/octave-dev
