Thema "Hey Yongbei," in Intel® oneAPI Math Kernel Library

is there some mkl function for transposing rectangle matrix in mpi?

yongbei_ma_ — Sun, 08 Mar 2015 02:35:41 GMT

thanks.

Hi yongbei m.

Ying_H_Intel — Mon, 09 Mar 2015 03:55:18 GMT

Hi yongbei m.

It seems two questions here

R1. transposing function

R2. MPI support

So are you working on large matrix , which are distributed on nodes in a cluster?

You may check the MKL manual

for R1.

mkl_?imatcopy
Performs scaling and in-place transposition/copying of
matrices

mkl_?omatcopy
Performs scaling and out-place transposition/copying
of matrices.
Syntax
Fortran:
call mkl_somatcopy(ordering, trans, rows, cols, alpha, a, lda, b, ldb)
call mkl_domatcopy(ordering, trans, rows, cols, alpha, a, lda, b, ldb)
call mkl_comatcopy(ordering, trans, rows, cols, alpha, a, lda, b, ldb

For R2,

Please see ScaLAPACK Auxiliary and Utility Routines 7,which support MPI. But seems no direct function of the purpose.

On the other hand, as you see, most of function of Scalapack support Transpose as input. What is the follow operation after you get the transpose of the matrix? it would be better if there are chance to combine the transpose and other functionality together.

Best Regards,

Ying

Hey Yongbei,

Sebastian_L_1 — Thu, 13 Aug 2015 09:17:16 GMT

Hey Yongbei, maybe the functions pstran/pdtran/pctranu/pztranu are what you are looking for? These PBLAS functions should "Transpose a distributed real/complex matrix". I haven't used these functions, yet. I wonder a little bit, why the MPI communicator is not needed in function call. Maybe someone else is experienced with pctran? Unfortunately there are no p?tran examples provided in the Intel installation. Best regards Sebastian

Hi Sebastian,

Ying_H_Intel — Fri, 14 Aug 2015 03:40:56 GMT

Hi Sebastian,

thanks for the reply. right, these functions should be able to transposes a distributed matrix. The function call is like below pstran(m, n, alpha, a, ia, ja, desca, beta, c, ic, jc, descc), but it does need mpi communicator out of functions, also need create desca, descc, which have the position and mpi rank information. So the function don't need MPI communicator itself, but encaps MPI communicator internally. .
We have some discussion about the matrix distrubution and MKL PBLAS function call in forum thread like

https://software.intel.com/en-us/forums/topic/561007

https://software.intel.com/en-us/forums/topic/562259. ;

It also include some examples. You can distribute the matrix manually like the sample or call functions, pstran, psadd (also suppose transpose) You may refer to them.

Best Regards,

Ying