10

I am writing some code right now and I have a placeholder with matmul that seems to be working pretty well, but I'd like to use a LAPACK dgemm implementation. I am only using gfortran right now and getting very good speeds with matmul, but I wonder if I can get better.

The current call is:

C = transpose(matmul( transpose(A), B))

where A, B, and C are non-square, double precision matrices. I can easily write a wrapper for dgemm with the current gfortran implementation of LAPACK, but I like that I can do this all as a function (rather than worrying about call for a surbroutine and having to deal with the transpose).

I am wondering if I compile with ifort and include the MKL, will this matmul magically change to a MKL dgemm function for me with no wrapper?

Ross
  • 2,130
  • 14
  • 25
drjrm3
  • 4,474
  • 10
  • 53
  • 91

1 Answers1

14

You don't want all MATMULs to be dgemm, it is not profitable for very small matrices.

Gfortran does what you want

-fexternal-blas This option will make gfortran generate calls to BLAS functions for some matrix operations like MATMUL, instead of using our own algorithms, if the size of the matrices involved is larger than a given limit (see -fblas-matmul-limit). This may be profitable if an optimized vendor BLAS library is available. The BLAS library will have to be specified at link time.

and you can even change the size limit for switching to BLAS by -fblas-matmul-limit=n

You can easily use MKL this way in gfortran.

Intel Fortran has something similar

[no - ] opt-matmul This option enables [disables] a compiler - generated Matrix Multiply (matmul) library call by identifying matrix multiplicat ion loop nests , if any , and replacing them with a matmul library call for improved performance. This option is enabled by default if options / O3 ( - O3) and /Qparallel ( - parallel) are specified. This option has no effect unless option / O2 ( - O2) or higher is set.

Vladimir F Героям слава
  • 57,977
  • 4
  • 76
  • 119