This is very interesting. Do you guys have a standard set of scripts for performance testing?
> On Mar 12, 2017, at 8:28 PM, Henry Rich <[email protected]> wrote: > > Complex +/ . * does not use AVX yet. > > Henry Rich > > On 3/12/2017 7:47 PM, Don Guinn wrote: >> Tried a 1000 by 1000 complex matrix and compared inverting it and then the >> matrix product of it and its inverse on J805 and J806 avx. Saw less than a >> 10% improvement with avx. Then tried a real 5000 by 5000 matrix and saw a >> speedup of 1.6 on the inverse and 2.2 on the multiply. Ran it several times >> back and forth to minimize outside interference. Probably the complex >> numbers aren't well vectorized yet. >> >> My laptop is kind of old. 8G memory and Intel(R) Core(TM) i7-3619QM CPU @ >> 2.30GHz running Windows 10. Had upgraded from the original Windows 7 and it >> doesn't run really well, but maybe wait a while before getting a new one to >> let the latest avx mature some. >> ---------------------------------------------------------------------- >> 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
