Le 29/08/2014 12:58, Dang, Christophe a écrit :
Just to close the subject:
I tried to implement the algorithm with sparse matrices, and it is less
efficient than scanning over one dimension: 7 times faster than the naive
algorithm.
Yes, it was somewhat expected. Also for memory consumption, the sparse
encoding becomes interesting vs the dense encoding only when the
sparsity becomes smaller than .. 50%.
There was a quite recent thread on bugzilla, about the relevance of
keeping a sparse encoding for the result of cos(SP) where SP is a sparse
(so with a good fraction of null values).
It was finally -- wisely -- decided (after some tests) to switch to a
dense encoding for that case (as it should be for any function turning 0
into a non-null value).
Samuel
_______________________________________________
users mailing list
users@lists.scilab.org
http://lists.scilab.org/mailman/listinfo/users
