design of functions in miltimin

design of functions in miltimin

[Andrey V. Panov]

I think that the body of take_step() function in 
multimin/directional_minimize.c :

{
  gsl_vector_set_zero (dx);
  gsl_blas_daxpy (-step * lambda, p, dx);
    
  gsl_vector_memcpy (x1, x);
  gsl_blas_daxpy (1.0, dx, x1);
}

can be replaced by more efficient code:

{
int i;
double dx_temp;
for(i = 0; i < dx->size; i++)
{
  dx_temp = -step * lambda * gsl_vector_get(p, i);
  gsl_vector_set(dx, i, dx_temp);
  gsl_vector_set(x1, i, gsl_vector_get(x, i) + dx_temp);
}
}

There are also similar parts of code inside miltimin routines which may be 
replaced.

-- 
Andrey V. Panov
http://canopus.iacp.dvo.ru/~panov/

Re: design of functions in miltimin

[Fabrice Rossi]

Hi.

My understanding of vector calculation optimization is that one should 
rely on BLAS as much as possible. I'm not sure that your code is really 
more efficient than the BLAS code if you use an optimized implementation 
of BLAS such as ATLAS, but of course, I might be wrong. Did you make 
some timing with GSL+ATLAS to compare both solutions?

Fabrice

Andrey V. Panov wrote:

> I think that the body of take_step() function in
> multimin/directional_minimize.c :
>
> {
>   gsl_vector_set_zero (dx);
>   gsl_blas_daxpy (-step * lambda, p, dx);
>
>   gsl_vector_memcpy (x1, x);
>   gsl_blas_daxpy (1.0, dx, x1);
> }
>
> can be replaced by more efficient code:
>
> {
> int i;
> double dx_temp;
> for(i = 0; i < dx->size; i++)
> {
>   dx_temp = -step * lambda * gsl_vector_get(p, i);
>   gsl_vector_set(dx, i, dx_temp);
>   gsl_vector_set(x1, i, gsl_vector_get(x, i) + dx_temp);
> }
> }
>
> There are also similar parts of code inside miltimin routines which 
> may be
> replaced.
>

Re: design of functions in miltimin

[Andrey V. Panov]

Hi.

I disagree. Each call to a cblas function leads to additional for cycle.
There is no much difference when unique call is used (only cost of a function 
call). But in the case of serial 4 calls this leads to 4 for cycles instead 
of one.

On Tuesday 01 July 2003 17:48, Fabrice Rossi wrote:
> Hi.
>
> My understanding of vector calculation optimization is that one should
> rely on BLAS as much as possible. I'm not sure that your code is really
> more efficient than the BLAS code if you use an optimized implementation
> of BLAS such as ATLAS, but of course, I might be wrong. Did you make
> some timing with GSL+ATLAS to compare both solutions?
>
> Fabrice
>
> Andrey V. Panov wrote:
> > I think that the body of take_step() function in
> > multimin/directional_minimize.c :
> >
> > {
> >   gsl_vector_set_zero (dx);
> >   gsl_blas_daxpy (-step * lambda, p, dx);
> >
> >   gsl_vector_memcpy (x1, x);
> >   gsl_blas_daxpy (1.0, dx, x1);
> > }
> >
> > can be replaced by more efficient code:
> >
> > {
> > int i;
> > double dx_temp;
> > for(i = 0; i < dx->size; i++)
> > {
> >   dx_temp = -step * lambda * gsl_vector_get(p, i);
> >   gsl_vector_set(dx, i, dx_temp);
> >   gsl_vector_set(x1, i, gsl_vector_get(x, i) + dx_temp);
> > }
> > }
> >
> > There are also similar parts of code inside miltimin routines which
> > may be
> > replaced.

-- 
Andrey V. Panov
panov /@/ canopus.iacp.dvo.ru