[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: complex arrays in octfiles

From: c.
Subject: Re: complex arrays in octfiles
Date: Wed, 29 Jun 2011 18:06:19 +0200

On 29 Jun 2011, at 17:37, John W. Eaton wrote:

> On 29-Jun-2011, c. wrote:
> | Can you explain this a bit better? 
> | By saying the slowness is due to copy-on-write semantics, 
> | do you mean that using the indexing operator is making a copy of
> | a each time I access it?
> No, it is checking to see if it needs to make a copy.  If it does,
> then it copies the contents of the array.  The copy will only happen
> once, but the check will happen every time through the loop.

So, if I understand correctly the overhead of

a (i, j, k) = z;

is roughly equivalent to that of

if (true)
 aptr [i + a.dim1 () * (j + a.dim2 () * k)] = z;

in the case at hand?

> In your case, no copy is needed, because you just created the array.
> A copy would be needed if there were multiple "handles" for the same
> data.  For example
>  NDArray x (...);   // Create an array; initial reference count for
>                     // underlying data is 1.
>  NDArray y = x;     // No data is actually copied here, just a
>                     // pointer; the reference count for the
>                     // underlying data is also incremented.
>  x(...) = val;      // Here, the reference count is checked, and
>                     // since it is greater than 1, a copy is made,
>                     // reference counts are adjusted, then the
>                     // assignment is performed.  This way, x changes
>                     // but y does not (copy on write).
> Does that help?

Yes, thanks! 

> | Sorry if the question sounds stupid, I do things this way sometimes in 
> order to have
> | bounds checking, but if I am doing it wrong I have quite a bit of code to 
> fix ;)
> Even if you use the indexing operators, you don't get bounds checking
> unless you enable it with
> before including Array.h.  Because bounds checking slows things down
> considerably, this macro is not defined by default unless Octave is
> configured with --enable-bounds-checking.

Yes, I am aware of this, thanks again.
Actually that allows very to very conveniently 
switch off bounds checking when debugging is finished 
without changing the code.

> jwe


reply via email to

[Prev in Thread] Current Thread [Next in Thread]