| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This moves dispatching for `__array_function__` into a C-wrapper. This
helps speed for multiple reasons:
* Avoids one additional dispatching function call to C
* Avoids the use of `*args, **kwargs` which is slower.
* For simple NumPy calls we can stay in the faster "vectorcall" world
This speeds up things generally a little, but can speed things up a lot
when keyword arguments are used on lightweight functions, for example::
np.can_cast(arr, dtype, casting="same_kind")
is more than twice as fast with this.
There is one alternative in principle to get best speed: We could inline
the "relevant argument"/dispatcher extraction. That changes behavior in
an acceptable but larger way (passes default arguments).
Unless the C-entry point seems unwanted, this should be a decent step
in the right direction even if we want to do that eventually, though.
Closes gh-20790
Closes gh-18547 (although not quite sure why)
|
| |
|
|
|
| |
Move the possible_flags dictionary to a global value so it is not
re-constructed each call.
|
| |
|
|
|
| |
Document that other Python sequences are acceptable as arguments, not
just lists.
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
* DEP: remove NPY_ARRAY_UPDATEIFCOPY, deprecated in 1.14
* remove more UPDATEIFCOPY
* typo: add missing comma
* remove a few more UPDATEIFCOPY
* Add release note
* remove UPDATEIFCOPY from comment (from review)
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The array creation functions have the most to gain:
1. np.asarray is 4 times faster and commonly used.
2. Other functions are wrapped using __array_function__ in Python
making it more difficult
This commit (unfortunatly) has to do a few things:
* Modify __array_function__ C-side dispatching to accomodate
the fastcall argument convention.
* Move asarray, etc. to C after removing all "fast paths" from
np.array (simplifying the code)
* Fixup imports, since asarray was imported directly in a few places
* Replace some places where `np.array` was probably used for speed
instead of np.asarray or similar. (or by accident in 1 or 2 places)
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This PR adds the implementation of NEP-35's like= argument, allowing dispatch of array creation functions with __array_function__ based on a reference array.
* ENH: Add like= kwarg via __array_function__ dispatcher to asarray
* ENH: Add new function for __array_function__ dispatching from C
This new function allows dispatching from C directly, while also
implementing the new `like=` argument, requiring only minimal
changes to existing array creation functions that need to add
support for that argument.
* ENH: Add like= support to numpy.array
The implementation uses array_implement_c_array_function, thus
introducing minimal complexity to the original _array_fromobject
code.
* BUG: Fix like= dispatcher for np.full
* ENH: Remove np.asarray like= dispatcher via Python
np.asarray can rely on np.array's C dispatcher instead.
* TST: Add some tests for like= argument
Tests comprise some of the functions that have been implemented already:
* np.array (C dispatcher)
* np.asarray (indirect C dispatcher via np.array)
* np.full (Python dispatcher)
* np.ones (Python dispatcher)
* ENH: Remove like= argument during array_implement_array_function
* ENH: Add like= kwarg to ones and full
* BUG: prevent duplicate removal of `like=` argument
* ENH: Make `like=` a keyword-only argument
* ENH: Use PyUnicode_InternFromString in arrayfunction_override
Replace PyUnicode_FromString by PyUnicode_InternFromString to cache
"like" string.
* ENH: Check for arrayfunction_override import errors
Check and handle errors on importing NumPy's Python functions
* BUG: Fix array_implement_c_array_function error handling
* ENH: Handle exceptions with C implementation of `like=`
* ENH: Add `like=` dispatch for all asarray functions
Using Python dispatcher for all of them. Using the C dispatcher
directly on the `np.array` call can result in incorrect behavior.
Incorrect behavior may happen if the downstream library's
implementation is different or if not all keyword arguments are
supported.
* ENH: Simplify handling of exceptions with `like=`
* TST: Add test for exception handling with `like=`
* ENH: Add support for `like=` to `np.empty` and `np.zeros`
* TST: Add `like=` tests for `np.empty` and `np.zeros`
* ENH: Add `like=` to remaining multiarraymodule.c functions
Functions are:
* np.arange
* np.frombuffer
* np.fromfile
* np.fromiter
* np.fromstring
* TST: Add tests for multiarraymodule.c functions with like=
Functions are:
* np.arange
* np.frombuffer
* np.fromfile
* np.fromiter
* np.fromstring
* ENH: Add `like=` support to more creation functions
Support for the following functions is added:
* np.eye
* np.fromfunction
* np.genfromtxt
* np.identity
* np.loadtxt
* np.tri
* TST: Add `like=` tests for multiple functions
Tests for the following functions are added:
* np.eye
* np.fromfunction
* np.genfromtxt
* np.identity
* np.loadtxt
* np.tri
* TST: Reduce code duplication in `like=` tests
* DOC: Document `like=` in functions that support it
Add documentations for the following functions:
* np.array
* np.arange
* np.asarray
* np.asanyarray
* np.ascontiguousarray
* np.asfortranarray
* np.require
* np.empty
* np.full
* np.ones
* np.zeros
* np.identity
* np.eye
* np.tri
* np.frombuffer
* np.fromfile
* np.fromiter
* np.fromstring
* np.loadtxt
* np.genfromtxt
* ENH: Add `like=` to numpy/__init__.pyi stubs
* BUG: Remove duplicate `like=` dispatching in as*array
Functions `np.asanyarray`, `np.contiguousarray` and `np.fortranarray`
were dispatching both via their definitions and `np.array` calls,
the latter should be avoided.
* BUG: Fix missing check in array_implement_array_function
* BUG: Add missing keyword-only markers in stubs
* BUG: Fix duplicate keyword-only marker in array stub
* BUG: Fix syntax error in numpy/__init__.pyi
* BUG: Fix more syntax errors in numpy/__init__.pyi
* ENH: Intern arrayfunction_override strings in multiarraymodule
* STY: Add missing brackets to arrayfunction_override.c
* MAINT: Remove arrayfunction_override dict check for kwarg
* TST: Assert that 'like' is not in TestArrayLike kwargs
* MAINT: Rename array_implement_c_array_function(_creation)
This is done to be more explicit as to its usage being intended for
array creation functions only.
* MAINT: Use NotImplemented to indicate fallback to default
* TST: Test that results with `like=np.array` are correct
* TST: Avoid duplicating MyArray code in TestArrayLike
* TST: Don't delete temp file, it may cause issues with Windows
* TST: Don't rely on eval in TestArrayLike
* TST: Use lambda with StringIO in TestArrayLike
* ENH: Avoid unnecessary Py_XDECREF in arrayfunction_override
* TST: Make TestArrayLike more readable
* ENH: Cleaner error handling in arrayfunction_override
* ENH: Simplify array_implement_c_array_function_creation
* STY: Add missing spaces to multiarraymodule.c
* STY: C99 declaration style in arrayfunction_override.c
* ENH: Simplify arrayfunction_override.c further
Remove cleanup label from array_implementation_c_array_function,
simplifying the code. Fix unitialized variable warning in
array_implementation_array_function_internal.
* DOC: Use string replacement for `like=` documentation
Avoid repeating the full text for the `like=` argument by storing it as
a variable and using `replace` on each docstring.
* DOC: Update `like=` docstring
* TST: Test like= with array not implementing __array_function__
* TST: Add missing asanyarray to TestArrayLike
* ENH: Use helper function for like= dispatching
Avoid dispatching like= from Python implementation functions to improve
their performance. This is achieved by only calling a dispatcher
function when like is passed by the users.
* ENH: Rename array_function_dispatch kwarg to public_api
* BUG: Add accidentally removed decorator for np.eye back
* DOC: Add set_array_function_like_doc function
The function keeps Python files cleaner and resolve errors when __doc__
is not defined due to PYTHONOPTIMIZE or -OO .
* DOC: Add mention to like= kwarg being experimental
* TST: Test like= with not implemented downstream function
* DOC: Fix like= docstring reference to NEP 35.
* ENH: Prevent silent errors if public_api is not callable
* ENH: Make set_array_function_like_doc a decorator
* ENH: Simplify `_*_with_like` functions
* BUG: Fix multiple `like=` dispatching in `require`
* MAINT: Remove now unused public_api from array_function_dispatch
Co-authored-by: Sebastian Berg <sebastian@sipsolutions.net>
|
| |
|
|
|
|
|
| |
(#16811)
* DOC: update parameter choices for asarray, asarray_contiguous, asarray_chkfinite converters
Co-authored-by: sun <sun@vosdbt.org>
|
| |
|
|
|
| |
As numpy is Python 3 only, these import statements are now unnecessary
and don't alter runtime behavior.
|
| | |
|
|
|
This is a direct move, with some tweaks to imports.
This breaks a cyclic imports between `core.numeric` and `core.fromnumeric`.
This doesn't affect the value of `np.core.numeric.__all__` which keeps code doing `from numpy.core.numeric import *` working.
|
