arrayfunc 8.5.2

Fast array processing functions

Authors:

Michael Griffin

Version:

8.5.2 for 2023-10-05

Copyright:

2014 - 2023

License:

This document may be distributed under the Apache 2.0 License.

Language:

Python 3.6 or later

---

Introduction

The arrayfunc module provides high speed array processing functions for use with the standard Python array module. These functions are patterned after the functions in the standard Python Itertools and math module together with some additional ones from other sources.

The purpose of these functions is to perform mathematical calculations on arrays faster than using native Python.

See full documentation at: http://arrayfunc.readthedocs.io/en/latest/

If you are installing on an ARM platform such as the Raspberry Pi, see the installation notes at the end before attempting to install from PyPI using PIP.

---

Function Summary

The functions fall into several categories.

Filling Arrays

Function	Description
count	Fill an array with evenly spaced values using a start and step values.
cycle	Fill an array with evenly spaced values using a start, stop, and step values, and repeat until the array is filled.
repeat	Fill an array with a specified value.

Filtering Arrays

Function	Description
afilter	Select values from an array based on a boolean criteria.
compress	Select values from an array based on another array of boolean values.
dropwhile	Select values from an array starting from where a selected criteria fails and proceding to the end.
takewhile	Like dropwhile, but starts from the beginning and stops when the criteria fails.

Examining and Searching Arrays

Function	Description
findindex	Returns the index of the first value in an array to meet the specified criteria.
findindices	Searches an array for the array indices which meet the specified criteria and writes the results to a second array. Also returns the number of matches found.

Summarising Arrays

Function	Description
aany	Returns True if any element in an array meets the selected criteria.
aall	Returns True if all element in an array meet the selected criteria.
amax	Returns the maximum value in the array.
amin	Returns the minimum value in the array.
asum	Calculate the arithmetic sum of an array.

Data Conversion

Function	Description
convert	Convert arrays between data types. The data will be converted into the form required by the output array.

Mathematical operator functions

Function	Equivalent to
add	x + y
truediv	x / y
floordiv	x // y
mod	x % y
mul	x * y
neg	-x
pow	x**y or math.pow(x, y)
pow2	x * x or math.pow(x, 2)
pow3	x * x * x or math.pow(x, 3)
sub	x - y
abs_	abs(x)

Comparison operator functions

Function	Equivalent to
eq	x == y
gt	x > y
ge	x >= y
lt	x < y
le	x <= y
ne	x != y

Bitwise operator functions

Function	Equivalent to
and_	x & y
or_	x | y
xor	x ^ y
invert	~x
lshift	x << y
rshift	x >> y

Power and logarithmic functions

Function	Equivalent to
exp	math.exp(x)
expm1	math.expm1(x)
log	math.log(x)
log10	math.log10(x)
log1p	math.log1p(x)
log2	math.log2(x)
sqrt	math.sqrt(x)

Hyperbolic functions

Function	Equivalent to
acosh	math.acosh(x)
asinh	math.asinh(x)
atanh	math.atanh(x)
cosh	math.cosh(x)
sinh	math.sinh(x)
tanh	math.tanh(x)

Trigonometric functions

Function	Equivalent to
acos	math.acos(x)
asin	math.asin(x)
atan	math.atan(x)
atan2	math.atan2(x, y)
cos	math.cos(x)
hypot	math.hypot(x, y)
sin	math.sin(x)
tan	math.tan(x)

Angular conversion

Function	Equivalent to
degrees	math.degrees(x)
radians	math.radians(x)

Number-theoretic and representation functions

Function	Equivalent to
ceil	math.ceil(x)
copysign	math.copysign(x, y)
fabs	math.fabs(x)
factorial	math.factorial(x)
floor	math.floor(x)
fmod	math.fmod(x, y)
isfinite	math.isfinite(x)
isinf	math.isinf(x)
isnan	math.isnan(x)
ldexp	math.ldexp(x, y)
trunc	math.trunc(x)

Special functions

Function	Equivalent to
erf	math.erf(x)
erfc	math.erfc(x)
gamma	math.gamma(x)
lgamma	math.lgamma(x)

Additional functions

Function	Equivalent to
fma	fma(x, y, z) or x * y + z

Attributes

In addition to functions, a set of attributes are provided representing the platform specific maximum and minimum numerical values for each array type. These attributes are part of the “arraylimits” module.

---

Supported Array Types

Arrayfunc supports all standard Python 3.x array types.

---

Performance

Average performance increase on x86_64 Ubuntu with GCC is 100 times faster than native Python. Performance will vary depending on the function, operation, array data type used, and whether overflow checking is enabled, with the performance increase ranging from 50% to 3000 times.

Other platforms show similar improvements.

Detailed performance figures are listed in the full documentation.

---

Platform support

Arrayfunc is written in ‘C’ and uses the standard C libraries to implement the underlying math functions. Arrayfunc has been tested on the following platforms.

OS	Hardware	Bits	Compiler	Python Version
Debian 12	i686	32	GCC	3.11.2
Debian 12	x86_64	64	GCC	3.11.2
Ubuntu 22.04	x86_64	64	GCC	3.10.12
Ubuntu 23.04	x86_64	64	GCC	3.11.4
opensuse-leap 15.4	x86_64	64	GCC	3.6.15
almalinux 9.2	x86_64	64	GCC	3.9.16
alpine 3.18.4	i686	32	GCC	3.11.6
FreeBSD 13.2	amd64	64	Clang	3.9.18
OpenBSD 7.3	amd64	64	Clang	3.10.13
MS Windows 10	AMD64	64	MSC	3.12.0
MS Windows 11	AMD64	64	MSC	3.12.0
Raspbian 11	armv7l	32	GCC	3.9.2
Ubuntu 22.04	aarch64	64	GCC	3.10.12

amd64 is another name for x86_64 and does not indicate the CPU brand. armv7l is 32 bit ARM. The test hardware is a Raspberry Pi 3. aarch64 is 64 bit ARM. The test hardware is a Raspberry Pi 4.

• The Rasberry Pi 3 tests were conducted on a Raspberry Pi 3 ARM CPU running in 32 bit mode.

• The Ubuntu ARM tests were conducted on a Raspberry Pi 4 ARM CPU running in 64 bit mode.

• All others were conducted using VMs running on x86 hardware.

---

Installation

Please note that this is a Python 3 package. To install using Pip, you will need (with Debian package in brackets):

• The appropriate C compiler and header files (gcc and build-essential).

• The Python3 development headers (python3-dev).

• Pip3 together with the corresponding Setuptools (python3-pip).

example:

# Install from PyPI.
pip3 install arrayfunc
# Force install from PyPI source instead of using a binary wheel.
pip3 install --user --force-reinstall --no-binary=:all: arrayfunc
# Install from a local copy of the source package (Linux).
pip3 install --no-index --find-links=. arrayfunc
# Install a local package as a user package.
pip3 install --user --no-index --find-links=. arrayfunc
# Windows, FreeBSD, and OpenBSD seems to use "pip" instead
# of "pip3" for some reason.
pip install arrayfunc

Newer versions of OpenBSD and FreeBSD will not install this package correctly when running setup.py directly. Use pip to install, even for local package installs. Testing of this package has been changed to use only pip (or pip3) in order to provide a common testing method for all platforms. Testing using setup.py directly is no longer done.

Recent versions of PyPI seem to be building their own binary wheels for some platforms using their own infrastruction. This may result in an invalid ARM binary on Raspberry Pi.

If you have difficulties, then either download the tar.gz version and install it locally (see the above instructions for a local install). Alternatively, see the above example for how to force a binary install instead of using a wheel. There is also a bash script called “setupuser.sh” which will call setup. py directly with the appropriate parameters.

The setup.py file has platform detection code which it uses to pass the correct flags to the C compiler. For ARM, this includes the CPU type. If you are using an ARM CPU type which is not recognized then setup.py may not compile in SIMD features. You can experiment with modifying setup.py to add new ARM models, but be sure that anything you try is compatible with the existing ones.

Installing on Linux with PIP and PEP-668

PEP-668 (PEPs describe changes to Python) introduced a new feature which can affect how packages are installed with PIP. If PIP is configured to be EXTERNALLY-MANAGED it will refuse to install a package outside of a virtual environment.

The intention of this is to prevent conflicts between packages which are installed using the system package manager, and ones which are installed using PIP.

Linux distros which are affeced by this include the latest versions of Debian and Ubuntu.

As this package is a library which is intended to be used by other applications, there is no one right way to install it, whether inside or outside of a virtual environment. Review the options available with PIP to see what is suitable for your application.

For testing purposes this package was installed by setting the environment variable PIP_BREAK_SYSTEM_PACKAGES to “1”, which effectively disables this feature in PIP.

example:

export PIP_BREAK_SYSTEM_PACKAGES=1

---

Release History

• 8.5.2 - Update to testing and support. There were no code changes.

  Python version on Windows 10 and 11 was updated to version 12.

• 8.5.1 - Update to testing and support. There were no code. changes.

  Ubuntu version updated to 23.04. AlmaLinux updated to 9.1. Alpine Linux updated to 3.17.3. FreeBSD updated to 13.2. OpenBSD updated to 7.3. On Ubuntu 23.04, the installation method has changed due to how PEP-668 was implemented by Debian and how this affects “pip”. Some other distros may experience the same problems if they made the same changes. See the README.TxT for details.

• 8.5.0 - Added pyproject.toml file to satisfy Python 3.11 requirements.

  Updated build scripts to use python3 -m build instead of calling setup.py directly. Test targets were updated, Ubuntu 20.04 was dropped, Ubuntu 22.10 was added, FreeBSD python version upgraded to 3.9, OpenBSD upgraded to 7.2, Windows 10 Python upgraded to 3.11, Windows 11 Python upgraded to 3.11. Removed duplicate assignment in parameter parsing return data in arrayparams_asum.c. Added __version__ attribute to allow checking package version number at run time. Added version unit test. Updated setup.py and other files to allow the version number to be automatically updated from a single source at build time.

• 8.4.1 - Minor bug fix for asum for unsigned integer SIMD on ARM. This

  corrects the function return type for SIMD operations on ARM. No incorrect behaviour was found in the original, but this change was made to ensure correctness.

• 8.4.0 - Major performance improvements for asum through the use of SIMD and

  other optimizations.

• 8.3.0 - Fixed the effects of an apparent compiler bug affecting 32 bit

  x86 only for function asum. Tested and verified on 32 bit Debian and 32 bit Alpine. This would in a few very specific circumstances result in the sum of a float array (array code ‘f’) exceeding the valid range for a float instead of returning infinity. The fix forces the result to infinity in these cases. Also tested with new releases of Alma 9 and Alpine 3.16.

• 8.2.0 - Update to testing and support. Tested with new releases of Ubuntu

  22.04 and OpenBSD 7.1. Changed “simdsupport” to also report the architecture the binary was compiled for. “Simdsupport” is only used for testing and benchmarking and is not a stable part of the release.

• 8.1.2 - Bump to correct minor documentation error in README.rst.

• 8.1.1 - Update to testing and support. Raspberry Pi 32 bit OS updated to

  version 2022-04-04. Update to setup.py to improve ARM version detection.

• 8.1.0 - Update to testing and support. Centos has been replaced by

  AlmaLinux due to Red Hat ending long term support for Centos. No actual code changes.

• 8.0.1 - Technical bump to version number to include update information.

• 8.0.0 - Performance improvements in add, sub, mul, neg, abs, ceil, floor,

  trunc, sqrt, degrees, radians. Asum will now use error checking with floating point SIMD by default where available. Benchmarks and unit tests have been updated accordingly.

• 7.2.0 - Performance improvements in asum and pow. Asum will now use error

  checking with floating point SIMD on x86_64 by default. Pow has special cases for powers of 2 and 3 on integer arrays which allow for much greater performance. Pow will now raise a value error exception if an attempt to raise to a negative number. This makes it it more compatible with Python. New functions pow2 and pow3 added which raise array values to powers of 2 and 3 respectively. These have additional optimisations beyond pow, particularly with floating point arrays. Benchmarks for add, floordiv, mod, mul, pow, sub, and truediv have been changed to make them run the expanded range of tests much faster.

• 7.1.0 - This is a bugfix release to correct mod, mul, and pow. This affects

  integer overflow checking at extremes, particularly with the greatest magnitude negative number on signed arrays. Certain combinations of numbers may have produced an overflow error when the result was at the negative margin of the numeric range (e.g. -128 for array type ‘b’ when -2 is raised to the power of 7). The errors have been fixed, including adding special cases. Also, when 1 or -1 was raised to a very large power this would cause the algorithm to work for a very long time to produce an answer (e.g. 1 to the power of 4 billion). This is now detected and a special case added to short circuit the calculation to produce the answer. The unit tests for these and related functions have been updated to include a much wider range of test data.

• 7.0.0 - Major speed improvements to add, sub, mul, abs, neg using SIMD with

  overflow checking on integer array types. SIMD is now active as the default on integer arrays with smaller word sizes for these functions. Major speed improvements on x86 for lshift and rshift by adding SIMD support to addition integer array types. This was already present on ARM. Added benchmark for “convert” (this was missing). Debian test platforms were updated to latest versions (11).

• 6.2.0 - Updated benchmarks to make each one a separate file. Centos and

  OpenSuse test platforms updated to latest versions.

• 6.1.1 - Documentation updated and version number bumped to reflect testing

  with Ubuntu 21.04, FreeBSD 13.0, and OpenBSD 6.9. No code changes.

• 6.1.0 - Changed convguardbands to narrow -ve guard bands by 1 to handle

  LLVM warning. Changed setup.py to detect Raspberry Pi 4 and set the compiler args accordingly. Added support for Pi 4. Dropped testing of 64 bit mode on Pi 3.

• 6.0.1 - Documentation updated to reflect testing with the release version

  of Ubuntu 20.04 ARM (Rasberry Pi), Ubuntu 2010 (x86-64), OpenBSD 6.8, and Python 3.9 on Windows. No code changes and no change in version number.

• 6.0.0 - Documentation updated to reflect testing with the release version

  of Ubuntu 20.04. No code changes and no change in version number.

• 6.0.0 - Added SIMD support for ARMv8 AARCH64. This is 64 bit ARM on a

  Raspberry Pi3 when running 64 bit Ubuntu. Raspbian is 32 bit only and has 64 bit SIMD vectors. 64 bit ARM has 128 bit SIMD vectors and so offers improved performance.

• 5.1.1 - Updated and improved help documentation. Also updated test

  platforms and retested.

• 5.1.0 - This is a bug fix release only, centred around SIMD issues on

  x86-64 with GCC. In a previous release some of the x86-64 SIMD code had been changed to take advantage of a sort of assisted auto-vectorisation present in GCC. However, certain operations on certain integer sizes with certain array types will cause GCC to generate incorrect x86 SIMD operations, producting integer overflow. The functions known to be affected are aall, aany, findindex (B, H, I arrays), eq, ge, gt, le, lt, ne (B, H, I arrays), and rshift (h, i arrays). ARM was not affected. All auto-vectorisation, where used, has been changed back to manually generated SIMD operations for both x86 and ARM. Rshift no longer uses SIMD operations for b, B, h, or i arrays on x86. Lshift no longer supports SIMD operations on b or B arrays on x86. Add and sub no longer use SIMD for B, H, and I arrays on x86. Mul no longer uses SIMD on x86 for any array types. Where SIMD functionality has been removed on x86, it of course is still supported through normal portable CPU instructions. ARM SIMD support was not affected by these changes. Lost SIMD acceleration will be returned to x86 in a later release where possible after the necessary research has been conducted. Unit tests have been updated to cover a greater range of integer values to test for this problem. Platforms using compilers other than GCC were not affected by this, as they did not use SIMD anyway. The main effect of this present change is that some calculations may be slower for some array types. The problem with GCC generating incorrect SIMD instructions in some circumstances is apparently a known (but obscure) issue. This will be avoided in future releases by sticking with manual SIMD built-ins. Some source code files have updated date stamps in this release but no substantive code changes due to the template system used to auto-generate code.

• 5.0.0 - The main focus of this release has been adding SIMD

  acceleration support to the ARMv7 platform (e.g. Raspberry Pi 3). Also added SIMD support to ‘lshift’ and ‘rshift’ on x86-64 and ARM. Changed arrayparamsbase to fix compiler warning on newer versions of GCC, but no change in actual operation. Updated supported OS versions tested, and added OpenBSD to supported platform list.

• 4.3.1 - Numerous performance inprovements through the use of SIMD

  acceleration in many functions. See the documentation to see which functions are affected. Restrictions on the use of non-finite data in parameters has been relaxed where possible. Repeat now allows non-finite data as fill values. For findindices, if no matches are found the result code is now 0 (zero) instead of -1.

• 4.2.0 - Added fma function. This has no equivalent in the Python

  standard library but is equivalent to x * y + z. Also changed list of supported platforms to update FreeBSD to version 12 and added Centos 7.

• 4.1.0 - Added isfinite function.

• 4.0.1 - Repeat upload to synchronise source and Windows binary “wheel”

  version. PyPI was not happy with the previous attempt.

• 4.0.0 - Major revision with many changes. Amap, starmap, and acalc were

  replaced with new individual functions. This change was made to provides a simpler and more consistent interface which is tailored to the individual function rather than attempting to make one parameter format fit all. The “disovfl” parameter has been named to “matherrors” in order to better reflect that it encompasses more than just integer overflow. Support for the “bytes” type has been removed. The Raspberry Pi has been added as a supported platform.

• 3.1.0 - Added log2 to amap, amapi, and acalc.

• 3.0.0 - Changed package format to “Wheel” files. No functional changes.

• 2.1.1 - Fixed missing header files in PyPI package. No functional changes.

• 2.0.0 - Many changes. Updated MS Windows support to 3.6 and latest compiler.

  This in turn brought the Windows version up to feature parity with the other versions. Changed supported MS Windows version from 32 bit to 64 bit. Added SIMD support for some functions which provided a significant performance for those affected. Updated supported versions of Debian and FreeBSD to current releases.

• 1.1.0 - Added support for math constants math.pi and math.e.

• 1.0.0 - First release.