On Mon, Apr 29, 2013 at 11:07 AM, Jordi Gutiérrez Hermoso < jord...@octave.org> wrote:
> On 29 April 2013 14:00, Jordi Gutiérrez Hermoso <jord...@octave.org> > wrote: > > And yes, sparse matrices can be indexed by a single index instead of > > two, like any other matrix. Internally in Octave's source, the > > assumption that a single index of octave_idx_type is available is used > > throughout. > > The specific case where this fails in this instance is octave_idx_type > dim_vector::numel (), which is obtained simply by multiplying each of > the dimensions of the matrix, even for sparse matrices (this is unlike > nnz, so a workaround just for trace.m would be to use nnz instead of > numel, but I think this would still leave some pretty broken sparse > matrices lying around with other problems). > > We would have to change numel () to use some other type that can hold > the result of a larger size, but this is a pretty fundamental function > in Octave. The overall assumption is that you can linearly index up to > numel (). Every place that calls this function would need to be > checked to see what happens if we change its return type to be some > special bigint. > > - Jordi G. H. > I'm not proposing using anything but octave_idx_type for indexing or changing the return type of numel() - I just question why numel() is used for sparse matrices. It should be irrelevant for anything but ccs2full(). Rather than restrict the size of sparse matrices I think it would make more sense to fix problems like this as they come up so that sparse storage is used as it was intended - to reduce storage & op count. -- Ed Meyer
