Re: test fftfilt fail w/ current tip

[Daniel J Sebald]

On 07/23/2013 01:46 PM, John W. Eaton wrote:
> On 07/23/2013 02:39 PM, Dmitri A. Sergatskov wrote:
> > octave:3> test fftfilt
> > ***** test
> > b = [1 1];
> > x = [1, zeros(1,9)];
> > assert (fftfilt (b, x ), [1 1 0 0 0 0 0 0 0 0] );
> > assert (fftfilt (b, x.'), [1 1 0 0 0 0 0 0 0 0].');
> > assert (fftfilt (b.',x ), [1 1 0 0 0 0 0 0 0 0] );
> > assert (fftfilt (b.',x.'), [1 1 0 0 0 0 0 0 0 0].');
> > assert (fftfilt (b, [x.' x.']), [1 1 0 0 0 0 0 0 0 0].'*[1 1]);
> > assert (fftfilt (b, [x.'+eps x.']) == [1 1 0 0 0 0 0 0 0 0].'*[1 1],
> > [false(10, 1) true(10, 1)]);
> > !!!!! test failed
> > assert (fftfilt (b, [x.' + eps, x.']) == [1, 1, 0, 0, 0, 0, 0, 0, 0,
> > 0].' * [1, 1],[false(10, 1), true(10, 1)]) expected
> > 0 1
> > 0 1
> > 0 1
> > 0 1
> > 0 1
> > 0 1
> > 0 1
> > 0 1
> > 0 1
> > 0 1
> > but got
> > 1 1
> > 0 1
> > 0 1
> > 0 1
> > 0 1
> > 0 1
> > 0 1
> > 0 1
> > 0 1
> > 0 1
> > values do not match
> > shared variables
> > scalar structure containing the fields:
> >
> > b = [](0x0)
> > x = [](0x0)
> > r = [](0x0)
> >
> > hg tip
> > changeset: 17035:08f0e372d006
> > tag: tip
> > user: Torsten <address@hidden <mailto:address@hidden>>
> > date: Mon Jul 22 23:18:50 2013 +0200
> > summary: doc viewer: correct the anchor position of the desired doc
> > entry
> >
> > Test passes with 3.6.4.
> >
> > Dmitri.
> > --
>
> Yes, this has been failing ever since Dan and Jordi did some work on
> fftfilt to fix bug #37297.

I looks like the test that is failing is supposed to test the arithmetic 
of the case when the inputs are not integer.  When the inputs are 
integer (second column), the result is "cast" to integer by dropping the 
floating portion.  In the first column the inputs are not all integer so 
the epsilon should be left intact.  However, perhaps because of machine 
precision the epsilon is being lost and it just so happens the output is 
exactly 1.

Dmitri, on your system, please test a larger discrepancy, say 2*eps or 
3*eps:

assert (fftfilt (b, [x.'+2*eps x.']) == [1 1 0 0 0 0 0 0 0 0].'*[1 1], 
[false(10, 1) true(10, 1)]);

Others have wanted these types of epsilon tests as tight as possible, 
but I'm content with 10*eps.  (I do like the idea though from a few 
weeks ago of making an epsilon function for which the input can be a 
vector containing magnitudes so that accuracy of some functions of 
diminishing value can be checked.)

Dan