topic QR Factorization - support for tall, skinny matrices in Intel® oneAPI Math Kernel Library

QR Factorization - support for tall, skinny matrices

smh — Thu, 08 Dec 2011 14:31:42 GMT

Would it be possible to get an optimized version of the routine cgeqrf, in order to speed up the QR factorization of tall, skinny matrices?

QR Factorization - support for tall, skinny matrices

Gennady_F_Intel — Fri, 09 Dec 2011 04:03:01 GMT

QR factorization routines are threaded but the efficiency of theirs implementation is depended on the input problem sizes.

QR Factorization - support for tall, skinny matrices

Chao_Y_Intel — Fri, 09 Dec 2011 09:19:17 GMT

Hello,

Could you also support the matrix size for the tall, skinny matrix? so we can track it for the future optimization consideration.

Thanks,
Chao

QR Factorization - support for tall, skinny matrices

styc — Sat, 10 Dec 2011 06:35:36 GMT

Following http://www.eecs.berkeley.edu/Pubs/TechRpts/2010/EECS-2010-131.pdf supposedly it is not too hard to produce a DIY version. The algorithm is not quite easy to understand.

QR Factorization - support for tall, skinny matrices

TimP — Sat, 10 Dec 2011 12:46:02 GMT

Investigators at Ohio State/Ohio Supercomputer center made versions of caqr using cilk+ and OpenMP. They offered to speak on it at SC11 conference. It links MKL but doesn't make significant use of it. The cilk+ organized for stride 1 inner loops runs well on Xeon 5680.

QR Factorization - support for tall, skinny matrices

smh — Mon, 19 Dec 2011 14:27:13 GMT

The matrix size of interest has a lead dimension of up to 250,000 with the other dimension on the order of one or two hundred.