Le 09/09/13 02:32, Adam Estrada a écrit :
Yep. JAMA looks good to me as well. I am not all that familiar with it and
think that JBLAS would be good too but that seems like it would add quite a
few unneeded dependencies.
I think that JBLAS could play a role a little bit later. For example
when drawing a contour map from a set of stations at random locations
(not on a grid), one method - among others - involve inverting a matrix
having as many rows and columns than stations. For such large matrix,
the industrially-proven LAPACK algorithms against numerical
instabilities would be very appreciated.
But in the particular case of the sis-referencing module, the matrix
size is the number of spatio-temporal dimensions + 1. For this reason I
expect matrix bigger than 5x5 to be quite unusual (while possible - for
example meteorologists use 2 time axes). For such small matrix, I think
that JAMA would be perfect. From my reading of javadoc, JBLAS uses pure
Java for O(1) operations like addition and multiplication by a vector,
but still uses native code for multiplication by an other matrix and
inversion (which we need).
Thanks all for your feedbacks!
Martin
