quantlib-dev

[Quantlib-dev] fd operators update

[Peter C. <pca...@gm...>]

Hi,

when updating to the current trunk I notice that my operators do not 
compile any more due to an extended interface of FdmLinearOpComposite. 
Seems I have to implement

virtual Disposable<std::vector<SparseMatrix> > toMatrixDecomp() const=0;

and looking at other operators I think I should return a vector of 
SparseMatrix'es corresponding to

apply_direction(0,...)
apply_direction(1,...)
...
apply_direction(n,...)
apply_mixed(...)

, yes ? What is the improvement when using UBLAS ? Sorry in case I 
overlooked any documentation on this.

thank you
Peter

Re: [Quantlib-dev] fd operators update

[Ralph S. <ori...@we...>]

Hi Peter,

yes, either you define QL_NO_UBLAS_SUPPORT or you implement toMatrixDecomp() as you described it. Providing toMatrixDecomp() has the advantage that it delivers the components for toMatrix() of FdmLinearOp and in this way you get a matrix representation of the P(I)DE as a SparseMatrix, such that you can e. g. directly invert it in order to solve it.

This was e.g. used by Klaus Spanderen in his GPU example

	http://hpcquantlib.wordpress.com/2012/12/27/multi-dimensional-finite-difference-methods-on-a-gpu/

where the QuantLib fdm framework was brought on the GPU via the toMatrix() interface and the CUDA cuSPARSE library.

Furthermore, you can use the matrix representation to precondition the finite difference solution.


Best regards
Ralph

Am 20.04.2013 um 20:25 schrieb Peter Caspers:

> Hi,
> 
> when updating to the current trunk I notice that my operators do not 
> compile any more due to an extended interface of FdmLinearOpComposite. 
> Seems I have to implement
> 
> virtual Disposable<std::vector<SparseMatrix> > toMatrixDecomp() const=0;
> 
> and looking at other operators I think I should return a vector of 
> SparseMatrix'es corresponding to
> 
> apply_direction(0,...)
> apply_direction(1,...)
> ...
> apply_direction(n,...)
> apply_mixed(...)
> 
> , yes ? What is the improvement when using UBLAS ? Sorry in case I 
> overlooked any documentation on this.
> 
> thank you
> Peter
> 
> 
> 
> 
> 
> ------------------------------------------------------------------------------
> Precog is a next-generation analytics platform capable of advanced
> analytics on semi-structured data. The platform includes APIs for building
> apps and a phenomenal toolset for data science. Developers can use
> our toolset for easy data analysis & visualization. Get a free account!
> http://www2.precog.com/precogplatform/slashdotnewsletter
> _______________________________________________
> QuantLib-dev mailing list
> Qua...@li...
> https://lists.sourceforge.net/lists/listinfo/quantlib-dev

Re: [Quantlib-dev] fd operators update

[Klaus S. <kl...@sp...>]

Hi Peter

the rational behind this method was to get an easy way to create sparse 
matrices out of a FdmLinearOpComposite or more precise to get an 
boost::numeric::ublas::compressed_matrix<Real> object. This class can then be 
used as an interface to highly efficient sparse matrix library like MKL or 
cusip. I have seen significant performance improvements while moving to a 
"tailor-made" sparse matrix library.

But looking at your posting I now realize that I've broken backwards 
compatibility and we should define 

virtual Disposable<std::vector<SparseMatrix> > toMatrixDecomp() const {
	QL_FAIL("method not implemented");
}

Then you don't have implement a method to provide functionality you are not 
interest in. What do you think? 

regards
 Klaus


On Saturday, April 20, 2013 08:25:24 PM Peter Caspers wrote:
> Hi,
> 
> when updating to the current trunk I notice that my operators do not
> compile any more due to an extended interface of FdmLinearOpComposite.
> Seems I have to implement
> 
> virtual Disposable<std::vector<SparseMatrix> > toMatrixDecomp() const=0;
> 
> and looking at other operators I think I should return a vector of
> SparseMatrix'es corresponding to
> 
> apply_direction(0,...)
> apply_direction(1,...)
> ...
> apply_direction(n,...)
> apply_mixed(...)
> 
> , yes ? What is the improvement when using UBLAS ? Sorry in case I
> overlooked any documentation on this.
> 
> thank you
> Peter
> 
> 
> 
> 
> 
> ----------------------------------------------------------------------------
> -- Precog is a next-generation analytics platform capable of advanced
> analytics on semi-structured data. The platform includes APIs for building
> apps and a phenomenal toolset for data science. Developers can use our
> toolset for easy data analysis & visualization. Get a free account!
> http://www2.precog.com/precogplatform/slashdotnewsletter
> _______________________________________________
> QuantLib-dev mailing list
> Qua...@li...
> https://lists.sourceforge.net/lists/listinfo/quantlib-dev

Re: [Quantlib-dev] fd operators update

[Peter C. <pca...@gm...>]

Thank you Klaus, Ralph, this is very interesting. Very nice blog too. 
The default implementation surely looks like a good idea (though for me 
backward compatibility is not highest priority and in this case it is 
really easy to stay compliant).
kind regards
Peter

Am 21.04.2013 13:39, schrieb Klaus Spanderen:
>
> Hi Peter
>
> the rational behind this method was to get an easy way to create 
> sparse matrices out of a FdmLinearOpComposite or more precise to get an
>
> boost::numeric::ublas::compressed_matrix<Real> object. This class can 
> then be used as an interface to highly efficient sparse matrix library 
> like MKL or cusip. I have seen significant performance improvements 
> while moving to a "tailor-made" sparse matrix library.
>
> But looking at your posting I now realize that I've broken backwards 
> compatibility and we should define
>
> virtual Disposable<std::vector<SparseMatrix> > toMatrixDecomp() const {
>
> QL_FAIL("method not implemented");
>
> }
>
> Then you don't have implement a method to provide functionality you 
> are not interest in. What do you think?
>
> regards
>
> Klaus
>
> On Saturday, April 20, 2013 08:25:24 PM Peter Caspers wrote:
>
> > Hi,
>
> >
>
> > when updating to the current trunk I notice that my operators do not
>
> > compile any more due to an extended interface of FdmLinearOpComposite.
>
> > Seems I have to implement
>
> >
>
> > virtual Disposable<std::vector<SparseMatrix> > toMatrixDecomp() const=0;
>
> >
>
> > and looking at other operators I think I should return a vector of
>
> > SparseMatrix'es corresponding to
>
> >
>
> > apply_direction(0,...)
>
> > apply_direction(1,...)
>
> > ...
>
> > apply_direction(n,...)
>
> > apply_mixed(...)
>
> >
>
> > , yes ? What is the improvement when using UBLAS ? Sorry in case I
>
> > overlooked any documentation on this.
>
> >
>
> > thank you
>
> > Peter
>
> >
>
> >
>
> >
>
> >
>
> >
>
> > 
> ----------------------------------------------------------------------------
>
> > -- Precog is a next-generation analytics platform capable of advanced
>
> > analytics on semi-structured data. The platform includes APIs for 
> building
>
> > apps and a phenomenal toolset for data science. Developers can use our
>
> > toolset for easy data analysis & visualization. Get a free account!
>
> > http://www2.precog.com/precogplatform/slashdotnewsletter
>
> > _______________________________________________
>
> > QuantLib-dev mailing list
>
> > Qua...@li...
>
> > https://lists.sourceforge.net/lists/listinfo/quantlib-dev
>