Arrayfunc

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 Centos has been replaced by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.

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.

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.

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.

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.

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.