Yes, it is definitely my recommendation to use marrays everywhere. I am starting to get in the habit of doing so myself. One of my ideas is to extend CL array operations to marrays so that you need not convert back and forth. It is also my intent that other math software libraries would also use marrays (or an extension of them), so that they become a standard way to interchange data between libraries.
There is more overhead for marrays than plain arrays, but on SBCL where the native arrays are used in GSL directly, there is very little, and there is no copying. I recommend that you build it with marray and then see if you have performance problems. However, if you are talking matrices of 1000000x1000000, I think you will have problems in any representation on almost any kind of hardware. For single floats, one such matrix is 4TB. In that case, I recommend you rethink the problem structure. If these are sparse matrices, you should use a sparse matrix representation. GSL doesn't have such a representation, but others have been interested in having it so it might be worth posting something to the GSL mailing list -- perhaps someone has written a GSL-compatible sparse array representation.
Liam
On Sat, Apr 11, 2009 at 8:12 AM, Malcolm Reynolds malcolm.reynolds@gmail.com wrote:
Hi all, hope this is the right place for this kind of request.
I'm writing a program which will be doing Machine Learning over graph data structures. In it I'm going to be needing to represent a bunch of matrices, some sparse, some definitely not sparse, and do fairly basic operations on these matrices (add/ subtract/multiply, pseudoinverse, probably some others). I've started a basic prototype using the standard CL arrays but I always knew I would need GSLL to calculate a pseudo inverse.
My question is, would it be a reasonably sensible decision to just use marrays for the whole program rather than worrying about converting to/ from CL-type arrays all the time? Are there any places where CL-arrays beat marrays either in terms of memory usage, access speed, or anything else? I might be needing to scale this system up be dealing with N*N square matrices where N is on the order of hundreds of thousands or even millions, if that makes any difference to the recommendation.
Thanks in advance for any help!
Malcolm Reynolds
Gsll-devel mailing list Gsll-devel@common-lisp.net http://common-lisp.net/cgi-bin/mailman/listinfo/gsll-devel