This looks good.  I adjusted the use of newlines and spacing of end-of-line comments to follow Octave norms in this cset https://hg.savannah.gnu.org/hgweb/octave/rev/ffc7bb75ea3e.

Marking bug as Fixed and closing report.

ok. figured out how to avoid the sort for k. I realized k is just a mapping of sorted x, or x(i) back to x, such that x(i)(k) = x.  that can be gotten from k and i with:

k = i #cheap way to initialize k to the right size
k(i) = 1:n;

 I'm pretty sure there's something similar that can be done to remove the need to sort for s, but don't have that yet.

Excellent! That works and it is now just five lines. Thanks but put your name on the patch.

(file #55729)

Thanks for cleaning it up. Much better without for-loops. I'd suggest removing the commented for-loop for l but adding a comment to the code for j that refers to the example. See attachment.

This needs to reverse three mappings:
  input->sorted->unique->stable
It was always going to be a bit obscure.

Immediately after I hit "Submit" I realised that I hadn't fixed the white space completely. Oops.

Thanks for the offer, but I don't think this is sufficient to be called a "contributor" so I'd prefer to remain a lurker.

(file #55728)

no problem with the timing. you've seen how long some bugs sit :)

i did a look through the patch. it seems to work fairly well. A few issues - octave tends to perform poorly with for loops.  It's not that noticeable with small inputs, but for larger arrays/loops it can get significant.

I attached a revised patch where I was able to replace your loops with some vectorized equivalents.  Still passes all the tests. 
see a quite timing comparison on the two:

## small input borrowed from a BIST, not too much different
A = [4,5,6; 1,2,3; 4,5,6];

#loops
tic; for idx = 1:10000, [y,i,j] = unique(A,"rows","stable");endfor,toc
Elapsed time is 1.78322 seconds.

#vectorized
tic; for idx = 1:10000, [y,i,j] = unique(A,"rows","stable");endfor,toc
Elapsed time is 1.59985 seconds.

## bigger array
A = randi([1 10], 1000, 3);

#loops
> tic; for idx = 1:100, [y,i,j] = unique(a,"rows","stable");endfor,toc
Elapsed time is 1.523 seconds.

#vectorized
tic; for idx = 1:100, [y,i,j] = unique(a,"rows","stable");endfor,toc
Elapsed time is 0.121731 seconds.

I also cleaned up some of the formatting (mostly whitespace around functions and operators - see the style guide link below.), separated two added tests and added a <bug#> tag to them so the test suite can trace them back if a regression ever occurs.


I'm still wondering if that could get streamlined a bit more.  There are a few calls to sort that I think might be redundant with the multi-level mapping, and my find(u(s)==j') to project into dim2 followed by cleanup math is a bit of a hack that might not be necessary with the right mapping.   I left the old code in as comments for now as a reminder, but they should get trimmed before being finished. let me know what you think.

also, assuming we push this for octave 10 in the near future, would you like to be in the contributor list (assuming your name as shown here and with email from your savannah profile?)

(file #55727)

Apologies for the slow response. Your examples with arrays are easy to fix with numel() instead of length(). I am looking at "rows".

ok, thanks for the explanation. updating the bug details. will give it a look.

It was my understanding that the current version of unique did have an option for up to three outputs.  Can you explain more about what yours does differently?

See https://docs.octave.org/latest/Sets.html