Some libraries to steal ideas from

baxissimo · Joined: 23 Oct 2006 Posts: 241 Location: Tokyo, Japan

These are all linear algebra-oriented matrix libraries:

An alpha of MTL 4 was just released recently.
http://www.osl.iu.edu/research/mtl/mtl4/

Boost ublas:
http://www.boost.org/libs/numeric/ublas/doc/index.htm

Boost multi_array:
http://www.boost.org/libs/multi_array/doc/user.html

Schani recommended FLENS:
http://flens.sourceforge.net/

I haven't run across a C++ library geared towards general N-dimensional arrays. Oh wait -- Don points out below that Blitz++ is such a beast.

Blitz++
http://www.oonumerics.org/blitz/

And on the NumPy mailing list someone suggested the computer vision / image library "Vigra":
http://kogs-www.informatik.uni-hamburg.de/~koethe/vigra/

Which reminds me of the other generic image package that you hear about these days -- Adobe's GIL (Generic Image Library), which also support arbitrary N-dimensional storage:
http://opensource.adobe.com/gil/

The GNU Scientific Library (GSL) apparently has an N-dim array thing. (It's in C so it's not a 'class' per se):
http://www.gnu.org/software/gsl/manual/html_node/index.html

For Python of course there's NumPy.

Lush:
http://lush.sourceforge.net/lush-manual/index.html

Don Clugston · Posted: Fri Nov 02, 2007 5:17 am Post subject:

sandford · Joined: 16 Jan 2007 Posts: 3

baxissimo · Joined: 23 Oct 2006 Posts: 241 Location: Tokyo, Japan

fawzi · Joined: 19 Feb 2008 Posts: 9 Location: Berlin, Germany

I gave a look at all the libraries.

Indeed the cisstVector seems nice for a user that needs all of that, but it is not what I want to do.
I working on a class/template for uniformly spaced (normally dense) multidimensional arrays with a dimension that although constant for each object once created it is not known at compile time.
The rank of the object on the other hand must be known at compile time.

Having multidimensional arrays with constant sizes (and very often you want only one or two dimensions constant) immediately increases the the complexity very much, and makes a clean design and uniform handling much less difficult (it is easy to have holes in the implementation of given corner cases).

This won't be the right class for 3D vectors, but for array of 3D vectors it might well be.

time will tell Smile

Don Clugston · Posted: Thu Jul 24, 2008 4:58 am Post subject:

It's worth knowing that the next DMD2 release (probably in a couple of days -- I have a pre-release already) contains array operations.
So you can write stuff like:

double [] a;
double [] b;

a += b[3..$-8]*3.5;

and have it converted into a DAXPY call.
Which will mean we'll pretty much have BLAS1 built into the language.
The implications for a matrix library are quite interesting...

baxissimo · Joined: 23 Oct 2006 Posts: 241 Location: Tokyo, Japan

Don Clugston · Posted: Mon Jul 28, 2008 1:46 am Post subject:

fawzi · Joined: 19 Feb 2008 Posts: 9 Location: Berlin, Germany

Well my multidimensional array library is now basically ready for primetime.

You can give it a look at

http://github.com/fawzi/narray/tree/master/INSTALL.txt

I think that the design is quite sound (3. iteration) and the performance reasonable.

There is still to do, but it has already a meaningful functionality, and performance.

Unfortunately the direct indexing of 3D arrays (that have larger overhead than 1D/2D) has a rather large overhead over the optimal strategy (10-15x) because the compiler is not smart enough to optimize it fully, but it is easy to avoid direct indexing, and as reference implementation direct indexing is still acceptable.

You can use it like this:
import frm.narray.NArray;

auto a=zeros!(float)([5,5]);
auto b=zeros!(float)(5);
auto c=dot(a,b);
diag(a)=arange(5);
a[0,0]=12.5f;
auto ev=eigh(a);