The fp64 resuts match J in all 3 arrayfire backends for me. The fp32 differ on all 3 backends. I think AMD gpus also match with fp64.
On Sunday, January 30, 2022, 05:33:08 p.m. EST, 'Mike Day' via Programming <programm...@jsoftware.com> wrote: So ...bench/matmul.ijs now has mpx, which now calls mptime, and returns benchmarks for f64 and f32. It’s fairly obvious what’s meant, but I’m getting two comments comparing J and AF results, but they don’t state which of f36 or f64 applies. I’m copy-typing, again, so typos rule: J and AF results match exactly J and AF results have max abs difference of: 31263.2 (I don’t yet know whether the 2nd is something of concern!) Cheers, Mike Sent from my iPad > On 30 Jan 2022, at 18:14, Eric Iverson <eric.b.iver...@gmail.com> wrote: > > The arrayfire addon updated with minor fixes and changes. > > Note that the mp_bench.ijs script has been renamed to be matmul.ijs. > > It might be stable enough now to start looking at what it can do for > performance. This is especially true if you have a higher end nvidia card. > ---------------------------------------------------------------------- > 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
