Re: Fast copy of buffer data into a matrix

[John Swensen]

David Bateman wrote:
> John Swensen wrote:
> > FYI, here is how I ended up copying a unsigned char buffer into a Matrix 
> > of type uint8.  It ended up taking about 50% of the time of doing it 
> > element by element.
> >
> > static Matrix m1(settings.width, settings.height);
> > static uint8NDArray m = octave_value(m1).uint8_array_value();
> > static octave_uint8* tmp = m.fortran_vec();
> > memcpy( tmp, capturebuffer, settings.height*settings.width );
> > return octave_value(m.transpose());   
> >
> >   
> >     
> Shouldn't that be
>
> static unit8NDArray m (dim_vector(settings.width, settings.height));
> static octave_uint8* tmp = m.fortran_vec();
> memcpy( tmp, capturebuffer, settings.height*settings.width );
> return octave_value(m.transpose());
>
> You might be able to do away with the transpose as well, as the
> transpose means that there is a second copy of m in memory at the same
> time. However, for it to be efficient, it would have to be some sort of
> blocked algorithm so that cache limitations are taken into account..
>
> D.
>
>   
Is there a difference between:
1) static unit8NDArray m (dim_vector(settings.width, settings.height));
and
2) static uint8NDArray m = octave_value(m1).uint8_array_value();
?
Both of them seem to work.  Also, I have to do the transpose because the 
data from the camera is in row major form and octave is column major (or 
the other way around, I always get confused).

John