suessspeise / matmul_benchmark Goto Github PK

This is a study on how fast embedded Python in Fortran can be. It compares different matrix multiplication implementations: two implementations using nested loops (in both Fortran and embedded Python) and two implementations that use BLAS libraries (Fortran's intrinsic matmul() and Numpy's np.matmul()).

During compilation via make different options can be specified.

• Compiler (FC): gfortran or ifort
• Timer (TIMER): CPU_TIME, OMP_WTIME or ICON_WALLCLOCK_TIMER
• Optimisation (OPT_LEVEL): any, e.g. 2
• RUN_SET will produce a binary that runs a set of tests, instead of receiving a configuration as arguments

A simple make will compile with default settings as configured in the makefile.

The makefile is made to work on DKRZ's Mistral cluster, but this is only visible in the $PYTHON_PREFIX that is hard coded and the use of module names for the the distinction of Python versions. They makefile can easily adapted by exchanging the $PYTHON_PREFIX with your systems path.

Compile in the desired configuration (see above), e.g.

make FC=gfortran OPT_LEVEL=2 TIMER=CPU_TIME

Then call from the terminal with the desired configuration, e.g.

./test.exe m 100 10

Here the three input arguments are

• implementation, choosing between
  • m: Fortran instrinsic matmul()
  • e: Fortran, nested loops
  • n: Numpy's np.matmul()
  • p: Python, nested loops
• matrix edge length, resulting in edge ** 2 elements
• repetitions of the same operation, a mean will be given

Alternatively you can compile using the RUN_SET keyword which will produce a binary that runs a hard coded set of implementations and matrix sizes. This is used in the run_set.sh.

work in progress