Changes from 2.8.0 to 2.8.1
---------------------------
* Fixed dependency list.
* Added ``pyproject.toml`` and modernize the ``setup.py`` script. Thanks to
Antonio Valentino for the PR.
Changes from 2.7.3 to 2.8.0
---------------------------
* Wheels for Python 3.10 are now provided.
* Support for Python 2.7 and 3.5 has been discontinued.
* All residual support for Python 2.X syntax has been removed, and therefore
the setup build no longer makes calls to the `2to3` script. The `setup.py`
has been refactored to be more modern.
* The examples on how to link into Intel VML/MKL/oneAPI now use the dynamic
library.
Changes from 2.7.2 to 2.7.3
---------------------------
- Pinned Numpy versions to minimum supported version in an effort to alleviate
issues seen in Windows machines not having the same MSVC runtime installed as
was used to build the wheels.
- ARMv8 wheels are now available, thanks to `odidev` for the pull request.
Changes from 2.7.1 to 2.7.2
---------------------------
- Support for Python 2.7 and 3.5 is deprecated and will be discontinued when
`cibuildwheels` and/or GitHub Actions no longer support these versions.
- Wheels are now provided for Python 3.7, 3.5, 3.6, 3.7, 3.8, and 3.9 via
GitHub Actions.
- The block size is now exported into the namespace as `numexpr.__BLOCK_SIZE1__`
as a read-only value.
- If using MKL, the number of threads for VML is no longer forced to 1 on loading
the module. Testing has shown that VML never runs in multi-threaded mode for
the default BLOCKSIZE1 of 1024 elements, and forcing to 1 can have deleterious
effects on NumPy functions when built with MKL.
- Use of `ndarray.tostring()` in tests has been switch to `ndarray.tobytes()`
for future-proofing deprecation of `.tostring()`, if the version of NumPy is
greater than 1.9.
- Added a utility method `get_num_threads` that returns the (maximum) number of
threads currently in use by the virtual machine. The functionality of
`set_num_threads` whereby it returns the previous value has been deprecated
and will be removed in 2.8.X.
Install the new interchangeable BLAS system created by Thomas Orgis,
currently supporting Netlib BLAS/LAPACK, OpenBLAS, cblas, lapacke, and
Apple's Accelerate.framework. This system allows the user to select any
BLAS implementation without modifying packages or using package options, by
setting PKGSRC_BLAS_TYPES in mk.conf. See mk/blas.buildlink3.mk for details.
This commit should not alter behavior of existing packages as the system
defaults to Netlib BLAS/LAPACK, which until now has been the only supported
implementation.
Details:
Add new mk/blas.buildlink3.mk for inclusion in dependent packages
Install compatible Netlib math/blas and math/lapack packages
Update math/blas and math/lapack MAINTAINER approved by adam@
OpenBLAS, cblas, and lapacke will follow in separate commits
Update direct dependents to use mk/blas.buildlink3.mk
Perform recursive revbump
Changes from 2.7.0 to 2.7.1
- Python 3.8 support has been added.
- Python 3.4 support is discontinued.
- The tests are now compatible with NumPy 1.18.
- `site.cfg.example` was updated to use the `libraries` tag instead of `mkl_libs`,
which is recommended for newer version of NumPy.
Changes from 2.6.9 to 2.7.0
- The default number of 'safe' threads has been restored to the historical limit
of 8, if the environment variable "NUMEXPR_MAX_THREADS" has not been set.
- Thanks to @eltoder who fixed a small memory leak.
- Support for Python 2.6 has been dropped, as it is no longer available via
TravisCI.
- A typo in the test suite that had a less than rather than greater than symbol
in the NumPy version check has been corrected thanks to dhomeier.
- The file `site.cfg` was being accidently included in the sdists on PyPi.
It has now been excluded.
Changes from 2.6.8 to 2.6.9
- Thanks to Mike Toews for more robust handling of the thread-setting
environment variables.
- With Appveyor updating to Python 3.7.1, wheels for Python 3.7 are now
available in addition to those for other OSes.
Changes from 2.6.7 to 2.6.8
---------------------------
- Add check to make sure that f_locals is not actually f_globals when we
do the f_locals clear to avoid the 310 memory leak issue.
- Compare NumPy versions using distutils.version.LooseVersion to avoid issue
312 when working with NumPy development versions.
- As part of multibuild, wheels for Python 3.7 for Linux and MacOSX are now
available on PyPI
Changes from 2.6.6 to 2.6.7
* Thanks to Lehman Garrison for finding and fixing a bug that exhibited memory leak-like behavior. The use in numexpr.evaluate of sys._getframe combined with .f_locals from that frame object results an extra refcount on objects in the frame that calls numexpr.evaluate, and not evaluate’s frame. So if the calling frame remains in scope for a long time (such as a procedural script where numexpr is called from the base frame) garbage collection would never occur.
* Imports for the numexpr.test submodule were made lazy in the numexpr module.
Changes from 2.6.4 to 2.6.5
- The maximum thread count can now be set at import-time by setting the
environment variable 'NUMEXPR_MAX_THREADS'. The default number of
max threads was lowered from 4096 (which was deemed excessive) to 64.
- A number of imports were removed (pkg_resources) or made lazy (cpuinfo) in
order to speed load-times for downstream packages (such as pandas, sympy,
and tables). Import time has dropped from about 330 ms to 90 ms. Thanks to
Jason Sachs for pointing out the source of the slow-down.
- Thanks to Alvaro Lopez Ortega for updates to benchmarks to be compatible with
Python 3.
- Travis and AppVeyor now fail if the test module fails or errors.
- Thanks to Mahdi Ben Jelloul for a patch that removed a bug where constants
in where calls would raise a ValueError.
- Fixed a bug whereby all-constant power operations would lead to infinite
recursion.
Changes from 2.6.3 to 2.6.4
- Christoph Gohkle noticed a lack of coverage for the 2.6.3
`floor` and `ceil` functions for MKL that caused seg-faults in
test, so thanks to him for that.
Changes from 2.6.2 to 2.6.3
- Documentation now available at readthedocs.io_.
- Support for floor() and ceil() functions added by Caleb P. Burns.
- NumPy requirement increased from 1.6 to 1.7 due to changes in iterator
flags.
- Sphinx autodocs support added for documentation on readthedocs.org.
- Fixed a bug where complex constants would return an error, fixing
problems with `sympy` when using NumExpr as a backend.
- Fix for 277 whereby arrays of shape (1,...) would be reduced as
if they were full reduction. Behavoir now matches that of NumPy.
- String literals are automatically encoded into 'ascii' bytes for
convience.
Updates to keep with API changes in newer NumPy versions
Removed several warnings
Fix bugs in function stringcontains()
Detection of the POWER processor
Fix pow result casting
Fix integers to negative integer powers
Detect numpy exceptions in expression evaluation
Better handling of RC versions
Prompted by beta.repology.org.
Changes from 2.6.0 to 2.6.1
* Fixed a performance regression in some situations as consequence of
increasing too much the BLOCK_SIZE1 constant. After more careful
benchmarks (both in VML and non-VML modes), the value has been set
again to 1024 (down from 8192). The benchmarks have been made with
a relatively new processor (Intel Xeon E3-1245 v5 @ 3.50GHz), so
they should work well for a good range of processors again.
* Added NetBSD support to CPU detection. Thanks to Thomas Klausner.
Fix CPU detection patch while here.
Changes from 2.5.2 to 2.6.0
===========================
- Introduced a new re_evaluate() function for re-evaluating the
previous executed array expression without any check. This is meant
for accelerating loops that are re-evaluating the same expression
repeatedly without changing anything else than the operands. If
unsure, use evaluate() which is safer.
- The BLOCK_SIZE1 and BLOCK_SIZE2 constants have been re-checked in
order to find a value maximizing most of the benchmarks in bench/
directory. The new values (8192 and 16 respectively) give somewhat
better results (~5%) overall. The CPU used for fine tuning is a
relatively new Haswell processor (E3-1240 v3).
- The '--name' flag for `setup.py` returning the name of the package
is honored now (issue #215).
Changes from 2.5.1 to 2.5.2
===========================
- conj() and abs() actually added as VML-powered functions, preventing
the same problems than log10() before (PR #212). Thanks to Tom Kooij
for the fix!
Changes from 2.5 to 2.5.1
=========================
- Fix for log10() and conj() functions. These produced wrong results
when numexpr was compiled with Intel's MKL (which is a popular build
since Anaconda ships it by default) and non-contiguous data (issue
#210). Thanks to Arne de Laat and Tom Kooij for reporting and
providing a nice test unit.
- Fix that allows numexpr-powered apps to be profiled with pympler.
Thanks to @nbecker.
Changes from 2.4.6 to 2.5
=========================
- Added locking for allowing the use of numexpr in multi-threaded
callers (this does not prevent numexpr to use multiple cores
simultaneously). (PR #199, Antoine Pitrou, PR #200, Jenn Olsen).
- Added new min() and max() functions (PR #195, CJ Carey).
Changes from 2.4.5 to 2.4.6
===========================
- Fixed some UserWarnings in Solaris (PR #189, Graham Jones).
- Better handling of MSVC defines. (#168, Francesc Alted).
Changes from 2.4.4 to 2.4.5
===========================
- Undone a 'fix' for a harmless data race. (#185 Benedikt Reinartz,
Francesc Alted).
- Ignore NumPy warnings (overflow/underflow, divide by zero and
others) that only show up in Python3. Masking these warnings in
tests is fine because all the results are checked to be
valid. (#183, Francesc Alted).
Changes from 2.4.3 to 2.4.4
===========================
- Fix bad #ifdef for including stdint on Windows (PR #186, Mike Sarahan).
Changes from 2.4.3 to 2.4.4
===========================
* Honor OMP_NUM_THREADS as a fallback in case NUMEXPR_NUM_THREADS is not
set. Fixes#161. (PR #175, Stefan Erb).
* Added support for AppVeyor (PR #178 Andrea Bedini)
* Fix to allow numexpr to be imported after eventlet.monkey_patch(),
as suggested in #118 (PR #180 Ben Moran).
* Fix harmless data race that triggers false positives in ThreadSanitizer.
(PR #179, Clement Courbet).
* Fixed some string tests on Python 3 (PR #182, Antonio Valentino).
Changes from 2.4.2 to 2.4.3
===========================
* Comparisons with empty strings work correctly now. Fixes#121 and
PyTables #184.
Changes from 2.4.1 to 2.4.2
===========================
* Improved setup.py so that pip can query the name and version without
actually doing the installation. Thanks to Joris Borgdorff.
Changes from 2.4 to 2.4.1
=========================
* Added more configuration examples for compiling with MKL/VML
support. Thanks to Davide Del Vento.
* Symbol MKL_VML changed into MKL_DOMAIN_VML because the former is
deprecated in newer MKL. Thanks to Nick Papior Andersen.
* Better determination of methods in `cpuinfo` module. Thanks to Marc
Jofre.
* Improved NumPy version determination (handy for 1.10.0). Thanks
to Åsmund Hjulstad.
* Benchmarks run now with both Python 2 and Python 3. Thanks to Zoran
Plesivčak.
Changes from 2.3.1 to 2.4
=========================
* A new `contains()` function has been added for detecting substrings
in strings. Only plain strings (bytes) are supported for now. See
PR #135 and ticket #142. Thanks to Marcin Krol.
* New version of setup.py that allows better management of NumPy
dependency. See PR #133. Thanks to Aleks Bunin.
Changes from 2.3 to 2.3.1
=========================
* Added support for shift-left (<<) and shift-right (>>) binary operators.
See PR #131. Thanks to fish2000!
* Removed the rpath flag for the GCC linker, because it is probably
not necessary and it chokes to clang.
Changes from 2.2.1 to 2.2.2
===========================
* The `copy_args` argument of `NumExpr` function has been brought
back. This has been mainly necessary for compatibility with
PyTables < 3.0, which I decided to continue to support. Fixed#115.
* The `__nonzero__` method in `ExpressionNode` class has been
commented out. This is also for compatibility with PyTables < 3.0.
See #24 for details.
* Fixed the type of some parameters in the C extension so that s390
architecture compiles. Fixes#116. Thank to Antonio Valentino for
reporting and the patch.
Changes from 2.2 to 2.2.1
=========================
* Fixes a secondary effect of "from numpy.testing import `*`", where
division is imported now too, so only then necessary functions from
there are imported now. Thanks to Christoph Gohlke for the patch.
Changes from 2.1 to 2.2
=======================
* [LICENSE] Fixed a problem with the license of the
numexpr/win32/pthread.{c,h} files emulating pthreads on Windows
platforms. After persmission from the original authors is granted,
these files adopt the MIT license and can be redistributed without
problems. See issue #109 for details
(https://code.google.com/p/numexpr/issues/detail?id=110).
* [ENH] Improved the algorithm to decide the initial number of threads
to be used. This was necessary because by default, numexpr was
using a number of threads equal to the detected number of cores, and
this can be just too much for moder systems where this number can be
too high (and counterporductive for performance in many cases).
Now, the 'NUMEXPR_NUM_THREADS' environment variable is honored, and
in case this is not present, a maximum number of *8* threads are
setup initially. The new algorithm is fully described in the Users
Guide now in the note of 'General routines' section:
https://code.google.com/p/numexpr/wiki/UsersGuide#General_routines.
Closes#110.
* [ENH] numexpr.test() returns `TestResult` instead of None now.
Closes#111.
* [FIX] Modulus with zero with integers no longer crashes the
interpreter. It nows puts a zero in the result. Fixes#107.
* [API CLEAN] Removed `copy_args` argument of `evaluate`. This should
only be used by old versions of PyTables (< 3.0).
* [DOC] Documented the `optimization` and `truediv` flags of
`evaluate` in Users Guide
(https://code.google.com/p/numexpr/wiki/UsersGuide).
Changes from 2.0.1 to 2.1
===========================
* Dropped compatibility with Python < 2.6.
* Improve compatibiity with Python 3:
- switch from PyString to PyBytes API (requires Python >= 2.6).
- fixed incompatibilities regarding the int/long API
- use the Py_TYPE macro
- use the PyVarObject_HEAD_INIT macro instead of PyObject_HEAD_INIT
* Fixed several issues with different platforms not supporting
multithreading or subprocess properly (see tickets #75 and #77).
* Now, when trying to use pure Python boolean operators, 'and',
'or' and 'not', an error is issued suggesting that '&', '|' and
'~' should be used instead (fixes#24).
Changes from 2.0 to 2.0.1
=========================
* Added compatibility with Python 2.5 (2.4 is definitely not supported
anymore).
* `numexpr.evaluate` is fully documented now, in particular the new
`out`, `order` and `casting` parameters.
* Reduction operations are fully documented now.
* Negative axis in reductions are not supported (they have never been
actually), and a `ValueError` will be raised if they are used.
Changes from 1.x series to 2.0
==============================
- Added support for the new iterator object in NumPy 1.6 and later.
This allows for better performance with operations that implies
broadcast operations, fortran-ordered or non-native byte orderings.
Performance for other scenarios is preserved (except for very small
arrays).
- Division in numexpr is consistent now with Python/NumPy. Fixes#22
and #58.
- Constants like "2." or "2.0" must be evaluated as float, not
integer. Fixes#59.
- `evaluate()` function has received a new parameter `out` for storing
the result in already allocated arrays. This is very useful when
dealing with large arrays, and a allocating new space for keeping
the result is not acceptable. Closes#56.
- Maximum number of threads raised from 256 to 4096. Machines with a
higher number of cores will still be able to import numexpr, but
limited to 4096 (which is an absurdly high number already).
changes:
-Added support for multi-threading in pure C
-refactorization of the opcode machinery, Added a couple of opcodes
-fixes
-release GIL during computations for better resource usage for
multithreaded apps
