On Wed, 07 May 2008 02:01:21 +0200, njh <njh@...1927...> wrote:

On Wed, 7 May 2008, Bill Baxter wrote:

...

I said "seems" :-)

But I was referring more to this list of things it does, of which only two are apparently needed:

I agree, although we will probably use least squares in the near future. Copying the appropriate files across is certainly a reasonable solution, and the one I would advocate at the moment.

(although solving linear systems I would class as something you _don't_ want to be doing yourself, and my past experience timing things showed that it was darned hard to beat the performance of GSL for root finding and linear systems (well that's blas really))

Anyway, I propose we copy the files from gsl exactly as they are, and default to using GSL is available on the grounds that it has code to use SSE/MMX/Transputer when available.

njh

Copying across the needed files could be a sensible solution, even if this means we need finding out all the various inner GSL file dependencies.

GSL is a well established numerical library with really good performance and it supports a lot of platforms. We can't hope to find one library that perfectly fits our needs, and I suppose that we all agree that is not useful to reinvent the wheel. There are problems that own both a geometric and a numerical aspect, and 2geom should offer a specific implementation for these problems, but not for basic numerical algorithms. So, how much should we exploit GSL ? As little as possible but not less. Moreover, I like the idea to make up some wrapper for the needed functionalities, so weakening the source code tie with the library.

Regards, Marco