Benchmarks

The compiled package is more than a hundred times faster than the pure-Python one, on CPython and on MicroPython.

Milliseconds per call, drawing into a 160 × 128 RGB565 framebuffer:

Call

CPython, compiled

CPython, pure

MicroPython, compiled

MicroPython, pure

fill_rect, the whole buffer

0.0008

3.15

0.0059

1.79

filled circle, radius 40

0.0075

1.67

0.0063

1.21

filled round_rect, 120 × 80, radius 10

0.0029

1.77

0.0047

1.08

text16, 20 characters

0.0037

0.68

0.0042

1.41

blit, a 64 × 64 sprite

0.024

3.57

0.021

4.40

Measured on one laptop (Intel Core Ultra 7 155H, Linux): CPython 3.12.3 with omnipython-pygraphics 0.1.1, and MicroPython 1.30 preview unix builds, one with the compiled module built in and one without. Neither pure run had numpy or ulab to call on. A board is far slower than a laptop; the ratio between the columns is what carries over.

The pure package fills RGB565 rectangles with numpy when it can import it (ulab’s numpy on MicroPython), which makes its fill_rect many times faster.

Run it yourself

tools/benchmark.py is the script behind the table. It prints which package it imported, then one line per call.

python tools/benchmark.py

Run it from a directory other than the repository root to get the installed package. To time the pure package on CPython, put lib/ first on the path from inside Python, because the repository root comes before PYTHONPATH:

cd /tmp && python -c "import sys; sys.path.insert(0, '/path/to/pygraphics/lib'); exec(open('/path/to/pygraphics/tools/benchmark.py').read())"

On MicroPython, a compiled module built into the firmware wins over any folder on the path, so time the pure package on a build without it:

MICROPYPATH=lib micropython tools/benchmark.py