Laplacian filters benchmark

This project aims to compare the performance of different implementations of the Laplacian filter (edge detection).

DataSet

The test pictures are from the BSDS500 dataset.

We used the grayscale versions and converted them into portable graymap format (.pgm) to keep the implementation simple, including with languages such as Mojo 🔥 where there is no utility libraries yet.

Some scripts use the inline image protocol offered by iTerm2 to make the user experience more interactive. This also helped us during the development phase to quickly detect implementation errors.

A subset of the dataset can be viewed using the viewpic.py or viewpic.mojo scripts.

Installation

Python

Python implementations have been tested with version 3.11.1. You can install the dependancies with pip

cd python; pip install -r requirements.txt

Mojo 🔥

⚠️ The Mojo implementations have been tested with the pre-release of mojo version 24.3. Mojo is moving fast : you may encounter malfunctions if you try to use it with a newer version. I plan to follow up quickly the syntax updates.

You can follow the instructions to install mojo compiler on Ubuntu linux or MacOs. There is no external dependancies.

How to run the benchmark

Interactive mode

It is possible to interactively test each implementation, for example:

cd python; python naive.py

Likewise, it is possible to test interactively mojo implementations :

cd mojo; mojo naive.mojo

Benchmark mode

Interactive mode works on a subset of the dataset and the display of images may influence the results. The script run.sh allows you to run the battery of tests in batch mode on all the dataset (except the first 3 slowest python implementations) :

./run.sh

Results

Here are the results on my laptop (apple macbook M1 first generation) :

PLATFORM : arm/macOS-14.2.1-arm64-arm-64bit with 8 cpus
Python 3.11.1
mojo 24.1.0 (55ec12d6)
---   compute mean duration on the BSDS500 dataset  ---
PYTHON : naive implementation          : 494.972 ms / picture
PYTHON : numpy element wise product    : 273.633 ms / picture
PYTHON : numpy + numba jit             : 44.914 ms / picture
PYTHON : scipy.signal fftconvolve      : 2.407 ms / picture
PYTHON : scipy.ndimage convolve        : 0.851 ms / picture
PYTHON : scipy.ndimage laplace         : 0.684 ms / picture
PYTHON : opencv filter2D               : 0.363 ms / picture
PYTHON : opencv laplacian              : 0.358 ms / picture
MOJO🔥 : naive implementation          : 0.80554000000000003 ms / picture
MOJO🔥 : vectorized implementation     : 0.14718999999999999 ms / picture
MOJO🔥 : parallelized (2 workers)      : 0.12929000000000002 ms / picture

No living animals were harmed during these experiments.