Python Performance

All benchmarks are platform-independent (run on any computing device with appropriate hardware). CuPy tests require an NVIDIA GPU with CUDA toolkit installed.

Install

This command prepares Python prereqs:

pip install -r requirements.txt

C and Fortran benchmarks requires building first using CMake.

cmake --workflow --preset default

Usage

Iterative benchmarks, here using the pisum algorithm:

python Pisum.py

---

Matrix Multiplication benchmarks:

python Matmul.py

Hypotenuse

For Python, numpy.hypot() is faster up to about a hundred elements, then numpy.sqrt(x2 + y2) becomes slightly faster. The benefit of hypot() is to not overflow for arguments near REALMAX.

For example, in Python:

from math import sqrt, hypot

a=1e154; hypot(a,a); sqrt(a**2+a**2);

1.414213562373095e+154
inf

For Fortran, observe that with Gfortran compiler that sqrt(x**2 + y**2) is slightly faster than hypot(x,y) in general across the tested array sizes.

Execute the Hypot speed test by:

python Hypot.py

Notes

Julia

Julia binaries are often downloaded to a particular directory. Python doesn't pickup .bash_aliases, which is commonly used to point to Julia.

MKL selection

https://software.intel.com/en-us/articles/intel-mkl-link-line-advisor

We give a hint to CMake where your MKL libraries on. For example:

MKLROOT=/opt/intel/mkl cmake ..

Of course this option can be combined with FC.

You can set this environment variable permanently for your convenience (normally you always want to use MKL) by adding to your ~/.bashrc the line:

export MKLROOT=/opt/intel/mkl