Yes, my point was that I don't think you can reliably determine in advance whether a matrix product scan will invoke floating point imprecision.
Marshall On Mon, Jan 09, 2023 at 08:25:43PM -0500, Henry Rich wrote: > The implementation is floating point. > > Henry Rich > > On Mon, Jan 9, 2023, 8:24 PM Marshall Lochbaum <mwlochb...@gmail.com> wrote: > > > The matrices in my example are integral. > > > > Marshall > > > > On Mon, Jan 09, 2023 at 05:20:49PM -0800, Elijah Stone wrote: > > > I've thought in the past that it would be nice to have a mode for > > > reproducible fp, but there is none yet (try +/ vs ([+])/). But it is > > moot > > > in this case, as the matrices in question were integral. > > > > > > (I don't think strict interpretation per se is necessary; just something > > > reliable and consistent. EG I imagine picking and committing to a way of > > > associating +/ which requires log space but admits efficient > > implementation > > > on any sort of parallel architecture.) > > > > > > On Mon, 9 Jan 2023, Marshall Lochbaum wrote: > > > > > > > I don't think you'd want to reorder matrix multiplications as that can > > > > change the results by an arbitrary amount. Here's an example where > > > > multiplication on one side cancels but not on the other side. > > > > > > > > (+/ .*&.>/ , +/ .*&.>~/@:|.) 1e6 1e7;(1e13+=i.2);1 _1 > > > > ┌────────┬──────────┐ > > > > │_9000000│_8.99482e6│ > > > > └────────┴──────────┘ > > > > > > > > Marshall > > > > > > > > On Mon, Jan 09, 2023 at 01:45:32PM -0500, Henry Rich wrote: > > > > > If f is known to be associative, f/\ proceeds from start to end in > > linear > > > > > time. Elijah has thoughts about how you could inform the > > interpreter that a > > > > > function is associative. Till that is implemented, you can use > > f/\.&.|. > > > > > which has linear time, or you can use Fold. > > > > > > > > > > Henry Rich > > > > > > > > > > On 1/9/2023 1:23 PM, Omar Antolín Camarena wrote: > > > > > > Am I right in thinking that the J interpreter does use a linear > > time algorithm for u/\ for some u including + and *? If so, maybe the case > > of +/ . * is worth adding. > > > > > > > > > > > > > > > > > > ---------------------------------------------------------------------- > > > > > For information about J forums see > > http://www.jsoftware.com/forums.htm > > > > ---------------------------------------------------------------------- > > > > For information about J forums see http://www.jsoftware.com/forums.htm > > > ---------------------------------------------------------------------- > > > For information about J forums see http://www.jsoftware.com/forums.htm > > ---------------------------------------------------------------------- > > For information about J forums see http://www.jsoftware.com/forums.htm > > > ---------------------------------------------------------------------- > For information about J forums see http://www.jsoftware.com/forums.htm ---------------------------------------------------------------------- For information about J forums see http://www.jsoftware.com/forums.htm
