Re: [Lapackpp-devel] bugs

[Christian S. <sti...@tu...>]

Am Samstag, 5. August 2006 20:14 schrieb Do bi:
> Please can you hint me on how to reduce the memory needs of a 10^5 x 10^5
> Matrix in Lapackpp? 1) How can I use floats instead for the matrix?

Use LaGenMatFloat instead of LaGenMatDouble.

>   2) What are the options in Lapack in terms of taking advantage of the
> symmetry, and using just half the matrix, and still solve A . x = b ?

Use LaSymmMatDouble from include/symd.h, but the header file hasn't been 
proof-read by myself (obvious because of almost missing documentation) so you 
need to watch out for potential errors in the code. 

Christian

> Christian Stimming <sti...@tu...> wrote:
>
>   Am Freitag, 4. August 2006 18:57 schrieb Do bi:
> > Hi Christian sorry I mistakenly sent to your personal address the first
> > time. This is my mportant enquiry:
> >
> > I noticed that LaGenMatDouble cannot be a size more than (10000,10000).
> > Please tell me what to do. I need larger matrices.
> > Thanks.
>
> Have you spent some thought about the required memory for the actual
> storage of a 10^4 x 10^4 general matrix? Obviously it has 10^8 entries. The
> "double" data type needs 8 bytes each ("float" needs 4 bytes), so this
> matrix would need 8*10^8 bytes in memory or in other words roughly 800
> MBytes (in one continuous block). You don't have that amount of memory
> available. That's the whole point of all the other matrix data types which
> can be used if you don't need one "general dense" matrix (but instead any
> kind of symmetry or triangularity or whatever.)
>
> So, no, it's not a bug of lapack/lapackpp that you cannot create
> LaGenMatDouble of that size. You cannot do this unless you're on some kind
> of supercomputer. You have to formulate your problem in a different way so
> that you don't need a LaGenMatDouble.