|
|
Ublas : |
From: Janek Kozicki (janek_listy_at_[hidden])
Date: 2006-12-29 10:31:17
Matthias Troyer said: (by the date of Fri, 29 Dec 2006 12:03:31 +0100)
> I see serious problems with this approach:
>
> * a sparse matrix/vector is serialized as dense
> * a copy is made
> * no use is made of the array optimizations in the CVS HEAD
>
> The performance, especially on large matrices will be very bad, and
> in many cases - especially when a dense copy is attempted for a large
> sparse matrix there will just not be enough memory.
Hello
Great that someone bothered to read that source :) Of course you are
absolutely right. It is a bit troubling that so many years passed and
there is still no ublas serialization. I have written this serialization
just to have it working for my project.
It would be very good if someone with time and ublas knowledge has
written this serialization stuff. Unfortunately due to lack of time
I can't do that, and simply use just this inefficient implentation.
Considering that there is a demand for this serialization I opt for
placing this implementation in CVS HEAD, so that people who need it -
will have it. Eventually someone with bigger demends (like huge
sparse matrices) will try to use it, and then will write his own
version that later he will submit here as a patch. And voila! We will
have an efficient implementation by then.
If nobody will submit that efficient version, it will mean that it's
not needed, and this version works just fine.
Better have something that works than nothing ;>
> Unfortunately ublas serialization is much more involved.
>
> > Those files should be placed in boost/serialization directory.
>
> No, they should be placed with ublas.
OK, I used them in that directory, but it's not much different for me.
-- Janek Kozicki |