Re: slow char({str})

[Daniel J Sebald]

Ben Abbott wrote:
> On Jun 21, 2009, at 8:32 PM, Daniel J Sebald wrote:
>
> > Ben Abbott wrote:
> >
> > > I noticed a substantial amount of time is needed to convert a cell   
> > > string into a string
> > > octave:50> tic; char ({repmat("a", [1, 100000])}); toc
> > > Elapsed time is 11.9 seconds.
> > > but ...
> > > octave:51> tic; char ({repmat("a", [1, 100000])}{1}); toc
> > > Elapsed time is 0.001117 seconds.
> > > Can someone with skilled in c++ take a look?
> >
> >
> > The hard part is locating a built-in function named "char".
> >
> > Dan
>
>
> It is the first function defined in .../src/strfns.cc

Ah.  I was searching the help text I saw running Octave, but the text has been 
updated.  Anyway...


> Fixing it is not a task I trust myself with (as my experience with c\c+ 
> + is mimicing the work of others).

Yes, there is a lot there and really those knowing it best should make changes, 
just to be consistent.  There is a lot to follow because the 
convert_to_str_internal() routine goes back to the base class.  Perhaps the 
other matrix types have special, more efficient member functions overriding the 
base class.  My first guess is that a slow point might be:

octave_value::octave_value (const Cell& c, bool is_csl)
 : rep (is_csl
         ? dynamic_cast<octave_base_value *> (new octave_cs_list (c))
         : dynamic_cast<octave_base_value *> (new octave_cell (c)))
{
}

in ov.cc.  Maybe there is a lot of overhead with cells in creating stuff on the heap, 
i.e., "new".

There is also this strange comment in ov.h:

// This will eventually go away, but for now it can be used to
// simplify the transition to the new octave_value class hierarchy,
// which uses octave_base_value instead of octave_value for the type
// of octave_value::rep.
#define OV_REP_TYPE octave_base_value

Maybe I've help someone narrow down things.  (Or thrown a red herring.)

Dan