Small complex matrix performance issue

[AndrewC]

There seems to be some kind of significant setup overhead with MKL routines. When doing operations on "small" (6x6) double precision complex matrices (zgemm), MKL performs somewhat worse than "regular" LAPACK/BLAS. MKL pulls ahead with larger matrices.
Unfortunately, though it is possible to "hand-code" an approach for these 6x6 matrices, a "naive" matrix multiply loop is much slower than using zgemm, MKL or not.

Andrew

[TimP]

MKL ZGEMM performance for complex data types would not be outstanding, unless you have a recent release, with an SSE3 specific version. Even then, for such a small matrix, it may be possible to write an in-line version which is faster. If you are able to use the MKL 8.0, with the version optimized for your processor, and you still find this slowdown, please submit a problem report on premier.intel.com.

[AndrewC]

Sorry,
I should have said I am using MKL 8.0 on a Pentium D