topic Thanks Dmitry. I am in doubt in Intel® oneAPI Math Kernel Library

Help using mkl_?imatcopy

Allan_L_ — Thu, 29 Aug 2013 12:11:21 GMT

The mkl_?imatcopy functions require an extra leading dimension parameter, called ldb.
From the documentation, I am not sure what value it should be used for this parameter. As an answer, I would like that the following case be considered:

A = (lda = 7 for row-major)
[x x x x x x x]
[x x a b c x x]
[x x d e f x x]
[x x x x x x x]

Transposed =
[x x x x x x x]
[x x a d b x x]
[x x e c f x x]
[x x x x x x x]

What value for ldb should be used in mkl_?imatcopy?

You cannot do this with mkl_

Dmitry_B_Intel — Thu, 29 Aug 2013 12:34:50 GMT

You cannot do this with mkl_?imatcopy, because elements b and e are not adjacent in the result - this permutation is not a transposition.

parameter ldb is the output stride. When you transpose M-by-N matrix you mean your input(m,n) is located at input[m*LDA+n] and your output(n,m) is located at output[n*LDB+m]

Thanks
Dima

Thanks Dmitry. I am in doubt

Allan_L_ — Thu, 29 Aug 2013 12:54:29 GMT

Thanks Dmitry. I am in doubt now. Can I use mkl_?imatcopy on a matrix block? If not, I can only think of lda being always cols (row major order). Also, for a MxN matrix with lda = N, what would be ldb? M?

Allan,

Evarist_F_Intel — Thu, 29 Aug 2013 18:31:09 GMT

Allan, one of the most popular imatcopy usage is imatcopy('R','T',rows,cols,1.,A,cols,rows). That is exactly what you ask about transpose entire MxN matrix (i.e. ldb=cols). But if you want to transpose a sub-block in a matrix you have to specify ldb slightly different. For example imatcopy('R', 'T', 2, 2, 1., A+6+2, 6, 6) converts A = [ x x x x x x ] [ x x a b x x ] [ x x c d x x ] [ ... ] to A` = [ x x x x x x ] [ x x a c x x ] [ x x b d x x ] [ ... ]

Thanks Evarist. Is there a

Allan_L_ — Thu, 29 Aug 2013 19:01:32 GMT

Thanks Evarist. Is there a scenario where ldb is different from lda?