Re: questions about recent changes

[Jaroslav Hajek]

On Thu, Mar 26, 2009 at 5:29 PM, John W. Eaton <address@hidden> wrote:
> You made the following change to Array.h:
>
>  -  dim_vector dims (void) const { return dimensions; }
>  +  // Return a const-reference so that dims ()(i) works efficiently.
>  +  const dim_vector& dims (void) const { return dimensions; }
>
> I'm not opposed to making a change like this, but we could make
> similar changes to many other functions in Octave, so maybe we should
> defer this until we can look at doing it more globally.  It seems odd
> to me to just make the change in this one place.  Also, I'm not sure I
> see the need to do this at all since dim_vectors are reference counted
> and copying them should be fairly fast (copy a pointer and increment a
> reference count).  Also, how often is "dims()(i)" actually used?  Does
> making this change have a noticeable effect on performance?

That's what I thought until recently, too; unfortunately, the reality
is much worse.
The reason is that dims () creates (until now) a non-const dim_vector
expression, and subsequently the non-const version of operator () was
called, which in turn forces a copy (!), because the object is aliased
(as you noted).

I think dims()(i) is not rare because there is no other method for
querying a single dimension; if there was, the change would not be
important.

I agree this should be changed in more places, at least in
octave_value and octave_base_value. I just didn't have enough time
today.

> You also changed dim_vecctor::numel so that it is now
>
>  octave_idx_type numel (int n = 0) const
>  {
>    int n_dims = length ();
>
>    octave_idx_type retval = 1;
>
>    for (int i = n; i < n_dims; i++)
>      retval *= elem (i);
>
>    return retval;
>  }
>
> With this definition and an object with dimensions [1, 0, 2, 3],
> dims.numel(2) will return 6.  Is that what you want?

Yes.

> I suppose this
> could help with doing things like
>
>  x = zeros (1, 0, 2, 3)
>  x(1,:,:)
>    => [](1,0,6)
>
> Is it used anywhere yet?

I just used it in Ffind, but I certainly needed it many times in the
past. See also dim_vector::redim.

> Should the Array::numel and octave_value::numel functions also accept
> an optional argument to specify the starting dimension?  Maybe it
> would be helpful to add a comment explaining the intent of N here.
>
> jwe
>

Possibly. But since it can be done with octave_value::dims ().numel
(i), I don't see that as very important. I just wanted the operation
to be easily doable without writing a loop. I guess there will be some
equivalent loops scattered over the sources.

cheers

-- 
RNDr. Jaroslav Hajek
computing expert & GNU Octave developer
Aeronautical Research and Test Institute (VZLU)
Prague, Czech Republic
url: www.highegg.matfyz.cz