Improving Performance for Small Size Problems

The overhead of calling an Intel® oneAPI Math Kernel Library function for small problem sizes can be significant when the functionhas a large number of parameters or internally checks parameter errors. To reduce the performance overhead for these small size problems, the Intel® oneAPI Math Kernel Librarydirect callfeature works in conjunction with the compiler to preprocess the calling parameters to supported Intel® oneAPI Math Kernel Library functions and directly call or inline special optimized small-matrix kernels that bypass error checking.For a list of functions supporting direct call, see Limitations of the Direct Call.

To activate the feature, do the following:

• Compile your C or Fortran code with the preprocessor macro depending on whether a threaded or sequential mode of Intel® oneAPI Math Kernel Library is required by supplying the compiler option as explained below:

  Intel® oneAPI Math Kernel Library Mode	Macro	Compiler Option
  Threaded	MKL_DIRECT_CALL	-DMKL_DIRECT_CALL
  Sequential	MKL_DIRECT_CALL_SEQ	-DMKL_DIRECT_CALL_SEQ

• For Fortran applications:

  • Enable preprocessor by using the -fpp option for Intel® Fortran Compiler.
  • Include the Intel® oneAPI Math Kernel Library Fortran include filemkl_direct_call.fi.

Intel® oneAPI Math Kernel Library skips error checking and intermediate function calls if the problem size is small enough (for example: a call to a function that supports direct call, such asdgemm, with matrix ranks smaller than 50).

• Using MKL_DIRECT_CALL in C Applications
  
• Using MKL_DIRECT_CALL in Fortran Applications
  
• Limitations of the Direct Call