pydivsufsort: bindings to libdivsufsort

pydivsufsort prebuilds libdivsufsort as a shared library and includes it in a Python package with bindings. Wheels are built for Linux, macOS and Windows (32 and 64 bits) using cibuildwheel and GitHub Actions. Basically, you should be able to install it with pip install pydivsufsort on any system and it should work out of the box. If it doesn't, please create an issue.

Features:

• bindings to divsufsort that return numpy arrays
• handle string, bytes and almost any integer data type (e.g. int64) and not only char
• algorithms work even for non char inputs
• additional string algorithms coded in Cython

Installation

On Linux, macOS and Windows:

python -m pip install pydivsufsort

We provide precompiled wheels for common systems using cibuildwheel, and a source distribution for Unix systems. Manual compilation on Windows might require some tweaking, please create an issue.

Features

All methods support string, bytes and numpy array inputs, including datatypes greater than uint8_t (e.g. uint64_t). Below are the signatures of all methods exposed by pydivsufsort. To import a method, just do from pydivsufsort import method_name. All methods are documented in the docstrings. You can display them with help(method_name).

A nicer interface to reuse computations lazily is provided in WonderString but currently undocumented. Please create an issue if you are interested.

Methods exposed from libdivsufsort

• divsufsort(string): suffix array
• bw_transform(string): Burrows-Wheeler transform
• inverse_bw_transform(idx, string): inverse Burrows-Wheeler transform
• sa_search(string, suffix_array, pattern): search for a pattern in a suffix array

Additional string algorithms

• kasai(string, suffix_array=None): LCP array computation (lazily computes the suffix array if not provided)
• lcp_segtree(string, suffix_array=None, lcp=None): build a segment tree for LCP queries (lazily computes the suffix array and LCP array if not provided)
• lcp_query(segtree, queries): query a segment tree for LCP queries. Queries are pairs of indices.
• levenshtein(string1, string2): Levenshtein distance
• most_frequent_substrings(lcp, length, limit=0, minimum_count=1): most frequent substrings. See the docstring for details.
• common_substrings(string1, string2, limit=25): common substrings between two strings.
• min_rotation(string): minimum rotation of a string
• longest_previous_factor(string, suffix_array=None, lcp=None): longest previous factor array (used in the Lempel-Ziv factorization)
• lempel_ziv_factorization(lpf, complexity: bool = False): Lempel-Ziv factorization
• lempel_ziv_complexity(string, suffix_array=None, lcp=None): Lempel-Ziv complexity
• kmp_censor_stream(censor, string): Censor a stream (like a generator of string) using the KMP algorithm

Example usage

from pydivsufsort import divsufsort, kasai

string_inp = "banana$"
string_suffix_array = divsufsort(string_inp)
string_lcp_array = kasai(string_inp, string_suffix_array)
print(string_suffix_array, string_lcp_array)
# [6 5 3 1 0 4 2] [0 1 3 0 0 2 0]

# You can also convert the string input to integers first

import numpy as np

int_inp = np.unique(np.array(list(string_inp)), return_inverse=True)[1]
int_suffix_array = divsufsort(int_inp)
int_lcp_array = kasai(int_inp, int_suffix_array)
print(int_suffix_array, int_lcp_array)
# [6 5 3 1 0 4 2] [0 1 3 0 0 2 0]

Development

You can install locally with

pip install -e .

A useful command to iterate quickly when changing Cython code is

python setup.py build_ext --inplace && pytest -s

Profiling

Profiling can be activated with the environment variable PROFILE:

PROFILE=1 python setup.py build_ext --inplace && pytest -s

Here is an example with line_profiler (requires pip install "line_profiler<4"):

import line_profiler
from pydivsufsort import common_substrings
from pydivsufsort.stringalg import (
    _common_substrings,
    repeated_substrings,
)

s1 = "banana" * 10000
s2 = "ananas" * 10000

func = common_substrings
profile = line_profiler.LineProfiler(func)
profile.add_function(_common_substrings)
profile.add_function(repeated_substrings)
profile.runcall(func, s1, s2, limit=15)
profile.print_stats()

Testing

pytest

Technical details (for performance tweaks)

libdivsufsort is compiled in both 32 and 64 bits, as the 32 bits version is faster. pydivsufsort automatically chooses to use the 32 bits version when possible (aka when the input size is less than 2**31-1).

For best performance, use contiguous arrays. If you have a sliced array, pydivsufsort converts it automatically with numpy.ascontiguousarray.

The precompiled libraries use OpenMP. You can disable it by setting the env variable OMP_NUM_THREADS=1, and it will yield the same performance as the version compiled without OpenMP

The original libdivsufsort only supports char as the base type. pydivsufsort can handle arrays of any integer type (even signed), by encoding each element as multiple chars, which makes the computation slower. If your values use an integer type that is bigger than required, but they span over a small contiguous range, pydivsufsort will automatically change their type (see #6).

Acknowledgements

• Yuta Mori for writing libdivsufsort
• Sean Law for initiating this project and contributing

Citing

If you have used this software in a scientific publication, please cite it using the following BibLaTeX code:

@software{pydivsufsort,
  author       = {Louis Abraham},
  title        = {pydivsufsort},
  year         = 2023,
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.7932458},
  howpublished = {\url{https://github.com/louisabraham/pydivsufsort}},
}