It looks like grpdelay in octave accepts different arguments than the matlab version

I guess the problem with this filter is that at ω = 0 we have undefined or indeterminate angle.

H(0) = (1 - e^(-j 0)) / (1 - 2e^(-j 0)) = 0 / (-1)

arg(0) = ?

In another approach, the numerator might be expressed by:
B(ω) = e^(-jω/2)(e^(jω/2) - e^(-jω/2)) = e^(-jω/2)(2j sin(ω/2))
arg(B(ω)) = -ω/2 + angle(2j sin(ω/2))

and the denominator by:
A(ω) = e^(-jω/2)(3jsin(ω/2) - cos(ω/2))
ang(A(ω)) = -ω/2 + arctan((3sin(ω/2)) / (-cos(ω/2)))

angle(H(ω)) = angle(2j sin(ω/2)) - arctan((3sin(ω/2)) / (-cos(ω/2)))

for ω > 0
angle(H(ω)) = π/2 - arctan((3sin(ω/2)) / (-cos(ω/2)))

for ω = 0 it is undefined.

That is causing the warning.

Here is the compared result of grpdelay in GNU Octave and Matlab for that particular filter:

GNU Octave
octave:1> pkg load signal
octave:2> b = [1 -1]; a = [1 -2];
octave:3> [gd, w] = grpdelay (b, a);
octave:4> [gd(1:10)'; w(1:10)']
ans =

  -1.0000  -1.4999  -1.4995  -1.4990  -1.4982  -1.4972  -1.4959  -1.4945  -1.4928  -1.4909
        0   0.0061   0.0123   0.0184   0.0245   0.0307   0.0368   0.0430   0.0491   0.0552


MATLAB

>> b = [1 -1]; a = [1 -2];
>> [gd, w] = grpdelay (b, a);
>> [gd(1:10)'; w(1:10)']

ans =

   -1.5000   -1.4999   -1.4995   -1.4990   -1.4982   -1.4972   -1.4959   -1.4945   -1.4928   -1.4909
         0    0.0061    0.0123    0.0184    0.0245    0.0307    0.0368    0.0430    0.0491    0.0552


Maybe we should point this to Paul Kienzle and Julius O. Smith III who have created the grpdelay.m function in GNU Octave.

To avoid this warning in the tests, we could just remove this one.

I had no warning here.

'grpdelay' is a function from the file /usr/share/octave/packages/signal-1.4.3/grpdelay.m

The function grpdelay.m used follows attached.

(file #54891)

Implementation of the function islinphase.