futhark-kullback-liebler

The Kullback-Liebler divergence, Hellinger distance, and alpha-divergence in Futhark.

Documentation

Documentation is available here.

Benchmarks

To run the benchmarks:

make
pipenv run python harness.py

and

futhark bench information.fut --backend opencl --runs=100

Comparison

Computation	Array Size	Implementation	Time
Entropy	10000000	Futhark	19.61 ms
Kullback-Liebler Divergence	10000000	Futhark	27.41 ms
Entropy	10000000	Python + Futhark	52.80 ms
Kullback-Liebler Divergence	10000000	Python + Futhark	94.07 ms
Entropy	10000000	Python (SciPy)	233.45 ms
Kullback-Liebler Divergence	10000000	Python (SciPy)	340.83 ms
Entropy	10000000	J	227.37 ms
Entropy	10000000	Haskell Accelerate	46.27 ms
Kullback-Liebler Divergence	10000000	Haskell Accelerate	70.99 ms

(Have a look at accelerate-kullback-liebler for details on its benchmarks).