bit-counter 0.3.1

Package for efficiently counting the number of one bits in a numpy array.

bit-counter

Package for efficiently counting the number of one bits in a numpy array. Implemented as a Python module using Rust, providing high performance counting.

Building

To build this package an installation of Rust and Python with the maturin package is required. The Maturin documentation on maturin local development is a useful reference for more details.

This project is configured by default to target CPUs with the POPCNT instruction. The builds available on PyPI have been built with this configuration. If you require a version for an older CPU without popcnt support, build with the RUSTFLAGS environment variable to exclude the popcnt target-feature. i.e. RUSTFLAGS='-C target-feature=-popcnt' maturin build -r.

By default the bit counting is done with a parallel map across all available CPUs through the use of rayon. Number of threads can be configured with the environment variable RAYON_NUM_THREADS.

Example usage

import numpy as np
from bit_counter import count_ones

arr = np.packbits(np.random.choice([True, False], 1000000))

count_of_true_values = count_ones(arr)

Performance

When built to target a CPU with popcnt support the count_ones method provided is substantially faster than a naive np.unpackbits(arr).sum(). Note that a faster bitwise_count function was added in numpy 1.22.0. The count_ones method also doesn't require unpacking the bit packed numpy array, so doesn't require any addtional memory to do the calculation.

For example with the following test code for 100 million True/False values that are packed:

import numpy as np
from bit_counter import count_ones

np.random.seed(42)
arr = np.packbits(np.random.choice([True, False], 100000000))

Using this package on an AMD Ryzen 9 3900X 12-Core Processor yields:

%timeit count_ones(arr)

471 µs ± 46.4 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

Better CPU performance can be had by treating the packed array as uint64:

%timeit count_ones(arr.view(np.uint64))

198 µs ± 47.2 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

Compared to the simple unpack and sum case with numpy 2.2.6:

%timeit np.unpackbits(arr).sum()

45.8 ms ± 879 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

The faster available numpy solution using bitwise_count is:

%timeit np.bitwise_count(arr).sum()

7.12 ms ± 44.8 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

Which is faster again when treating the packed bits as uint64:

%timeit np.bitwise_count(arr.view(np.uint64)).sum()

1.14 ms ± 4.94 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)