ulab: Crunch Numbers fast in CircuitPython

by Jeff Epler

published March 06, 2020, last edited January 22, 2025

posted in Sensors CircuitPython Programming/ MicroPython / CircuitPython

Intermediate

Project guide

A Simple Benchmark

In this simple benchmark, we compare two ways of finding the amplitude of a signal. normalized_rms computes it in a traditional way, handling each number one by one in Python code. normalized_rms_ulab computes more quickly by working on groups of numbers in a ulab array at the same time. ulab.numpy.std computes most quickly by moving all operations from Python to ulab.

Download Project Bundle

Copy Code

# SPDX-FileCopyrightText: 2020 Jeff Epler for Adafruit Industries
# SPDX-FileCopyrightText: 2020 Zoltán Vörös for Adafruit Industries
#
# SPDX-License-Identifier: MIT

import time
import math
from ulab import numpy as np

def mean(values):
    return sum(values) / len(values)

def normalized_rms(values):
    minbuf = int(mean(values))
    samples_sum = sum(
        float(sample - minbuf) * (sample - minbuf)
        for sample in values
    )

    return math.sqrt(samples_sum / len(values))

def normalized_rms_ulab(values):
    # this function works with ndarrays only
    minbuf = np.mean(values)
    values = values - minbuf
    samples_sum = np.sum(values * values)
    return math.sqrt(samples_sum / len(values))

# Instead of using sensor data, we generate some data
# The amplitude is 5000 so the rms should be around 5000/1.414 = 3536
nums_list = [int(8000 + math.sin(i) * 5000) for i in range(100)]
nums_array = np.array(nums_list)

def timeit(s, f, n=100):
    t0 = time.monotonic_ns()
    for _ in range(n):
        x = f()
    t1 = time.monotonic_ns()
    r = (t1 - t0) * 1e-6 / n
    print("%-30s : %8.3fms [result=%f]" % (s, r, x))

print("Computing the RMS value of 100 numbers")
timeit("traditional", lambda: normalized_rms(nums_list))
timeit("ulab, with ndarray, some implementation in python", lambda: normalized_rms_ulab(nums_array))
timeit("ulab only, with list", lambda: np.std(nums_list))
timeit("ulab only, with ndarray", lambda: np.std(nums_array))

# SPDX-FileCopyrightText: 2020 Jeff Epler for Adafruit Industries
# SPDX-FileCopyrightText: 2020 Zoltán Vörös for Adafruit Industries
#
# SPDX-License-Identifier: MIT

import time
import math
from ulab import numpy as np

def mean(values):
    return sum(values) / len(values)

def normalized_rms(values):
    minbuf = int(mean(values))
    samples_sum = sum(
        float(sample - minbuf) * (sample - minbuf)
        for sample in values
    )

    return math.sqrt(samples_sum / len(values))

def normalized_rms_ulab(values):
    # this function works with ndarrays only
    minbuf = np.mean(values)
    values = values - minbuf
    samples_sum = np.sum(values * values)
    return math.sqrt(samples_sum / len(values))

# Instead of using sensor data, we generate some data
# The amplitude is 5000 so the rms should be around 5000/1.414 = 3536
nums_list = [int(8000 + math.sin(i) * 5000) for i in range(100)]
nums_array = np.array(nums_list)

def timeit(s, f, n=100):
    t0 = time.monotonic_ns()
    for _ in range(n):
        x = f()
    t1 = time.monotonic_ns()
    r = (t1 - t0) * 1e-6 / n
    print("%-30s : %8.3fms [result=%f]" % (s, r, x))

print("Computing the RMS value of 100 numbers")
timeit("traditional", lambda: normalized_rms(nums_list))
timeit("ulab, with ndarray, some implementation in python", lambda: normalized_rms_ulab(nums_array))
timeit("ulab only, with list", lambda: np.std(nums_list))
timeit("ulab only, with ndarray", lambda: np.std(nums_array))

On my Metro M4, the ulab code computes almost exactly the same value, but over 40 times faster. Take care if running this on a board with an LCD display such as the CLUE, printing the results on the display takes a lot longer than the computation itself, and completely distorts the results.

Download File

Copy Code

traditional               :    2.951ms [result=3535.843611]
ulab, algorithm in python :    0.251ms [result=3535.853624]
ulab only, with list      :    0.336ms [result=3535.854340]
ulab only, with ndarray   :    0.068ms [result=3535.854340]

traditional               :    2.951ms [result=3535.843611]
ulab, algorithm in python :    0.251ms [result=3535.853624]
ulab only, with list      :    0.336ms [result=3535.854340]
ulab only, with ndarray   :    0.068ms [result=3535.854340]

Page last edited January 21, 2025