pool/python-numba
SHA256
2
0
Fork 1
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Accepting request 946610 from devel:languages:python:numeric

Browse Source 
OBS-URL: https://build.opensuse.org/request/show/946610
OBS-URL: https://build.opensuse.org/package/show/openSUSE:Factory/python-numba?expand=0&rev=31
This commit is contained in:
Dominique Leuenberger 2022-01-15 19:05:10 +00:00 committed by Git OBS Bridge
commit e47af9260d
6 changed files with 148 additions and 267 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
numba-0.54.1.tar.gz
View File

@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:f9dfc803c864edcc2381219b800abf366793400aea55e26d4d5b7d953e14f43f
size 2232957

3
numba-0.55.0.tar.gz Normal file
View File

@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:b221ebd997662a1dcb77e4f09140e022fbfe7577ad0781fc2e0214135dba6cbd
size 2300580

242
numba-pr7483-numpy1_21.patch
View File

@ -1,242 +0,0 @@
From 3a2c5042fa85ac9b8fe398d605d6d373d27f42bb Mon Sep 17 00:00:00 2001
From: Stuart Archibald <stuartarchibald@users.noreply.github.com>
Date: Thu, 14 Oct 2021 09:35:24 +0100
Subject: [PATCH 1/7] Update build matrix for NumPy 1.21 as per NEP-029.
As title.
---
azure-pipelines.yml | 68 ++++++++++++++--------------
buildscripts/azure/azure-windows.yml | 8 ++--
2 files changed, 38 insertions(+), 38 deletions(-)
Index: numba-0.54.1/setup.py
===================================================================
--- numba-0.54.1.orig/setup.py
+++ numba-0.54.1/setup.py
@@ -22,8 +22,7 @@ except ImportError:
min_python_version = "3.7"
max_python_version = "3.10" # exclusive
min_numpy_build_version = "1.11"
-min_numpy_run_version = "1.17"
-max_numpy_run_version = "1.21"
+min_numpy_run_version = "1.18"
min_llvmlite_version = "0.37.0rc1"
max_llvmlite_version = "0.38"
@@ -360,7 +359,7 @@ packages = find_packages(include=["numba
build_requires = ['numpy >={}'.format(min_numpy_build_version)]
install_requires = [
'llvmlite >={},<{}'.format(min_llvmlite_version, max_llvmlite_version),
- 'numpy >={},<{}'.format(min_numpy_run_version, max_numpy_run_version),
+ 'numpy >={}'.format(min_numpy_run_version),
'setuptools',
]
Index: numba-0.54.1/numba/__init__.py
===================================================================
--- numba-0.54.1.orig/numba/__init__.py
+++ numba-0.54.1/numba/__init__.py
@@ -137,10 +137,8 @@ def _ensure_critical_deps():
if PYVERSION < (3, 7):
raise ImportError("Numba needs Python 3.7 or greater")
- if numpy_version < (1, 17):
- raise ImportError("Numba needs NumPy 1.17 or greater")
- elif numpy_version > (1, 20):
- raise ImportError("Numba needs NumPy 1.20 or less")
+ if numpy_version < (1, 18):
+ raise ImportError("Numba needs NumPy 1.18 or greater")
try:
import scipy
Index: numba-0.54.1/README.rst
===================================================================
--- numba-0.54.1.orig/README.rst
+++ numba-0.54.1/README.rst
@@ -50,7 +50,7 @@ Dependencies
* Python versions: 3.7-3.9
* llvmlite 0.37.*
-* NumPy >=1.17,<1.21 (can build with 1.11 for ABI compatibility).
+* NumPy >=1.18 (can build with 1.11 for ABI compatibility).
Optionally:
Index: numba-0.54.1/docs/source/user/5minguide.rst
===================================================================
--- numba-0.54.1.orig/docs/source/user/5minguide.rst
+++ numba-0.54.1/docs/source/user/5minguide.rst
@@ -18,7 +18,7 @@ Out of the box Numba works with the foll
support on M1/Arm64.
* GPUs: Nvidia CUDA.
* CPython
-* NumPy 1.17 - latest
+* NumPy 1.18 - latest
How do I get it?
----------------
Index: numba-0.54.1/numba/np/ufunc/_internal.c
===================================================================
--- numba-0.54.1.orig/numba/np/ufunc/_internal.c
+++ numba-0.54.1/numba/np/ufunc/_internal.c
@@ -276,6 +276,7 @@ static PyMemberDef dufunc_members[] = {
*/
static struct _ufunc_dispatch {
+ /* Note that the following may also hold `_PyCFunctionFastWithKeywords` */
PyCFunctionWithKeywords ufunc_reduce;
PyCFunctionWithKeywords ufunc_accumulate;
PyCFunctionWithKeywords ufunc_reduceat;
@@ -286,7 +287,7 @@ static struct _ufunc_dispatch {
} ufunc_dispatch;
static int
-init_ufunc_dispatch(void)
+init_ufunc_dispatch(int *numpy_uses_fastcall)
{
int result = 0;
PyMethodDef * crnt = PyUFunc_Type.tp_methods;
@@ -329,6 +330,16 @@ init_ufunc_dispatch(void)
result = -1; /* Unknown method */
}
if (result < 0) break;
+
+ /* Check whether NumPy uses fastcall (ufunc.at never uses it) */
+ if (strncmp(crnt_name, "at", 3) != 0) {
+ if (*numpy_uses_fastcall == -1) {
+ *numpy_uses_fastcall = crnt->ml_flags & METH_FASTCALL;
+ }
+ else if (*numpy_uses_fastcall != (crnt->ml_flags & METH_FASTCALL)) {
+ return -1;
+ }
+ }
}
if (result == 0) {
/* Sanity check. */
@@ -344,6 +355,7 @@ init_ufunc_dispatch(void)
return result;
}
+
static PyObject *
dufunc_reduce(PyDUFuncObject * self, PyObject * args, PyObject *kws)
{
@@ -368,6 +380,47 @@ dufunc_outer(PyDUFuncObject * self, PyOb
return ufunc_dispatch.ufunc_outer((PyObject*)self->ufunc, args, kws);
}
+
+/*
+ * The following are the vectorcall versions of the above, since NumPy
+ * uses the FASTCALL/Vectorcall protocol starting with version 1.21.
+ * The only NumPy versions supporting vectorcall use Python 3.7 or higher.
+ */
+static PyObject *
+dufunc_reduce_fast(PyDUFuncObject * self,
+ PyObject *const *args, Py_ssize_t len_args, PyObject *kwnames)
+{
+ return ((_PyCFunctionFastWithKeywords)ufunc_dispatch.ufunc_reduce)(
+ (PyObject*)self->ufunc, args, len_args, kwnames);
+}
+
+static PyObject *
+dufunc_reduceat_fast(PyDUFuncObject * self,
+ PyObject *const *args, Py_ssize_t len_args, PyObject *kwnames)
+{
+ return ((_PyCFunctionFastWithKeywords)ufunc_dispatch.ufunc_reduceat)(
+ (PyObject*)self->ufunc, args, len_args, kwnames);
+}
+
+
+static PyObject *
+dufunc_accumulate_fast(PyDUFuncObject * self,
+ PyObject *const *args, Py_ssize_t len_args, PyObject *kwnames)
+{
+ return ((_PyCFunctionFastWithKeywords)ufunc_dispatch.ufunc_accumulate)(
+ (PyObject*)self->ufunc, args, len_args, kwnames);
+}
+
+
+static PyObject *
+dufunc_outer_fast(PyDUFuncObject * self,
+ PyObject *const *args, Py_ssize_t len_args, PyObject *kwnames)
+{
+ return ((_PyCFunctionFastWithKeywords)ufunc_dispatch.ufunc_outer)(
+ (PyObject*)self->ufunc, args, len_args, kwnames);
+}
+
+
#if NPY_API_VERSION >= 0x00000008
static PyObject *
dufunc_at(PyDUFuncObject * self, PyObject * args)
@@ -568,6 +621,41 @@ static struct PyMethodDef dufunc_methods
{NULL, NULL, 0, NULL} /* sentinel */
};
+
+/*
+ * If Python is new enough, NumPy may use fastcall. In that case we have to
+ * also use fastcall for simplicity and speed.
+ */
+static struct PyMethodDef dufunc_methods_fast[] = {
+ {"reduce",
+ (PyCFunction)dufunc_reduce_fast,
+ METH_FASTCALL | METH_KEYWORDS, NULL },
+ {"accumulate",
+ (PyCFunction)dufunc_accumulate_fast,
+ METH_FASTCALL | METH_KEYWORDS, NULL },
+ {"reduceat",
+ (PyCFunction)dufunc_reduceat_fast,
+ METH_FASTCALL | METH_KEYWORDS, NULL },
+ {"outer",
+ (PyCFunction)dufunc_outer_fast,
+ METH_FASTCALL | METH_KEYWORDS, NULL},
+#if NPY_API_VERSION >= 0x00000008
+ {"at",
+ (PyCFunction)dufunc_at,
+ METH_VARARGS, NULL},
+#endif
+ {"_compile_for_args",
+ (PyCFunction)dufunc__compile_for_args,
+ METH_VARARGS | METH_KEYWORDS,
+ "Abstract method: subclasses should overload _compile_for_args() to compile the ufunc at the given arguments' types."},
+ {"_add_loop",
+ (PyCFunction)dufunc__add_loop,
+ METH_VARARGS,
+ NULL},
+ {NULL, NULL, 0, NULL} /* sentinel */
+};
+
+
static PyObject *
dufunc_getfrozen(PyDUFuncObject * self, void * closure)
{
@@ -681,8 +769,15 @@ MOD_INIT(_internal)
return MOD_ERROR_VAL;
PyDUFunc_Type.tp_new = PyType_GenericNew;
- if (init_ufunc_dispatch() <= 0)
+
+ int numpy_uses_fastcall = -1;
+ if (init_ufunc_dispatch(&numpy_uses_fastcall) <= 0)
return MOD_ERROR_VAL;
+
+ if (numpy_uses_fastcall) {
+ PyDUFunc_Type.tp_methods = dufunc_methods_fast;
+ }
+
if (PyType_Ready(&PyDUFunc_Type) < 0)
return MOD_ERROR_VAL;
Py_INCREF(&PyDUFunc_Type);
Index: numba-0.54.1/numba/tests/test_array_reductions.py
===================================================================
--- numba-0.54.1.orig/numba/tests/test_array_reductions.py
+++ numba-0.54.1/numba/tests/test_array_reductions.py
@@ -725,6 +725,7 @@ class TestArrayReductions(MemoryLeakMixi
self.check_npdatetime(array_argmax)
self.check_nptimedelta(array_argmax)
+ @unittest.skip("https://github.com/numpy/numpy/issues/20376")
def test_median_npdatetime(self):
self.check_nptimedelta(array_median_global)

58
numba-pr7748-random32bitwidth.patch Normal file
View File

@ -0,0 +1,58 @@
From e6df66d52152156ba8bcda64b37f4995bda72d2f Mon Sep 17 00:00:00 2001
From: Graham Markall <gmarkall@nvidia.com>
Date: Fri, 14 Jan 2022 21:24:40 +0000
Subject: [PATCH] Fix #7713: Ensure _prng_random_hash return has correct
bitwidth
When the hash width is 32 bits, get_next_int32 needs to be used because
that returns a 32-bit integer, as opposed to get_next_int, which returns
a 64-bit integer regardless of the supplied bitwidth.
This commit also alters the signature of _prng_random_hash to use
_Py_hash_t as the return type - it will be equal to types.intp, but the
intent should be clearer.
---
numba/cpython/hashing.py | 21 +++++++++++++++++----
1 file changed, 17 insertions(+), 4 deletions(-)
diff --git a/numba/cpython/hashing.py b/numba/cpython/hashing.py
index 5a8673b04b..227fc4e1c5 100644
--- a/numba/cpython/hashing.py
+++ b/numba/cpython/hashing.py
@@ -18,7 +18,8 @@
from numba.core import errors
from numba.core import types, utils
from numba.core.unsafe.bytes import grab_byte, grab_uint64_t
-from numba.cpython.randomimpl import get_state_ptr, get_next_int, const_int
+from numba.cpython.randomimpl import (const_int, get_next_int, get_next_int32,
+ get_state_ptr)
_py38_or_later = utils.PYVERSION >= (3, 8)
_py310_or_later = utils.PYVERSION >= (3, 10)
@@ -137,11 +138,23 @@ def _prng_random_hash(tyctx):
def impl(cgctx, builder, signature, args):
state_ptr = get_state_ptr(cgctx, builder, "internal")
- bits = const_int(types.intp.bitwidth)
- value = get_next_int(cgctx, builder, state_ptr, bits, False)
+ bits = const_int(_hash_width)
+
+ # Why not just use get_next_int() with the correct bitwidth?
+ # get_next_int() always returns an i64, because the bitwidth it is
+ # passed may not be a compile-time constant, so it needs to allocate
+ # the largest unit of storage that may be required. Therefore, if the
+ # hash width is 32, then we need to use get_next_int32() to ensure we
+ # don't return a wider-than-expected hash, even if everything above
+ # the low 32 bits would have been zero.
+ if _hash_width == 32:
+ value = get_next_int32(cgctx, builder, state_ptr)
+ else:
+ value = get_next_int(cgctx, builder, state_ptr, bits, False)
+
return value
- sig = types.intp()
+ sig = _Py_hash_t()
return sig, impl

70
python-numba.changes
View File

@ -1,3 +1,73 @@
-------------------------------------------------------------------
Fri Jan 14 16:55:37 UTC 2022 - Ben Greiner <code@bnavigator.de>
- Update to 0.55.0
* This release includes a significant number important dependency
upgrades along with a number of new features and bug fixes.
* NOTE: Due to NumPy CVE-2021-33430 this release has bypassed the
usual release process so as to promptly provide a Numba release
that supports NumPy 1.21. A single release candidate (RC1) was
made and a few issues were reported, these are summarised as
follows and will be fixed in a subsequent 0.55.1 release.
* Known issues with this release:
- Incorrect result copying array-typed field of structured
array (#7693)
- Two issues in DebugInfo generation (#7726, #7730)
- Compilation failure for hash of floating point values on 32
bit Windows when using Python 3.10 (#7713).
* Support for Python 3.10
* Support for NumPy 1.21
* The minimum supported NumPy version is raised to 1.18 for
runtime (compilation however remains compatible with NumPy
1.11).
* Experimental support for isinstance.
* The following functions are now supported:
- np.broadcast_to
- np.float_power
- np.cbrt
- np.logspace
- np.take_along_axis
- np.average
- np.argmin gains support for the axis kwarg.
- np.ndarray.astype gains support for types expressed as
literal strings.
* For users of the Numba extension API, Numba now has a new error
handling mode whereby it will treat all exceptions that do not
inherit from numba.errors.NumbaException as a “hard error” and
immediately unwind the stack. This makes it much easier to
debug when writing @overloads etc from the extension API as
there’s now no confusion between Python errors and Numba
errors. This feature can be enabled by setting the environment
variable: NUMBA_CAPTURED_ERRORS='new_style'.
* The threading layer selection priority can now be changed via
the environment variable NUMBA_THREADING_LAYER_PRIORITY.
* Support for NVIDIA’s CUDA Python bindings.
* Support for 16-bit floating point numbers and their basic
operations via intrinsics.
* Streams are provided in the Stream.async_done result, making it
easier to implement asynchronous work queues.
* Support for structured types in device arrays, character
sequences in NumPy arrays, and some array operations on nested
arrays.
* Much underlying refactoring to align the CUDA target more
closely with the CPU target, which lays the groudwork for
supporting the high level extension API in CUDA in future
releases.
* Intel also kindly sponsored research and development into
native debug (DWARF) support and handling per-function
compilation flags:
* Line number/location tracking is much improved.
* Numba’s internal representation of containers (e.g. tuples,
arrays) are now encoded as structures.
* Numba’s per-function compilation flags are encoded into the ABI
field of the mangled name of the function such that it’s
possible to compile and differentiate between versions of the
same function with different flags set.
* There are no new general deprecations.
* There are no new CUDA target deprecations.
- Drop numba-pr7483-numpy1_21.patch
- Add numba-pr7748-random32bitwidth.patch -- gh#numba/numba#7748
-------------------------------------------------------------------
Sat Jan 8 22:19:07 UTC 2022 - Ben Greiner <code@bnavigator.de>

39
python-numba.spec
View File

@ -16,9 +16,8 @@
#
%{?!python_module:%define python_module() python-%{**} python3-%{**}}
%{?!python_module:%define python_module() python3-%{**}}
%define skip_python2 1
%define skip_python310 1
%define plainpython python
%global flavor @BUILD_FLAVOR@%{nil}
%if "%{flavor}" == "test"
@ -29,7 +28,7 @@
%bcond_with test
%endif
Name: python-numba%{psuffix}
Version: 0.54.1
Version: 0.55.0
Release: 0
Summary: NumPy-aware optimizing compiler for Python using LLVM
License: BSD-2-Clause
@ -37,26 +36,25 @@ URL: https://numba.pydata.org/
Source: https://files.pythonhosted.org/packages/source/n/numba/numba-%{version}.tar.gz
# PATCH-FIX-UPSTREAM fix-max-name-size.patch -- fix for gh#numba/numba#3876 -- from gh#numba/numba#4373
Patch0: fix-max-name-size.patch
# PATCH-FIX-UPSTREAM support numpy 1.21 -- gh#numba/numba#7176, gh#numba/numba#7483, gh#numpy/numpy#20376
Patch1: numba-pr7483-numpy1_21.patch
# PATCH-FIX-UPSTREAM numba-pr7748-random32bitwidth.patch -- gh#numba/numba#7748
Patch1: numba-pr7748-random32bitwidth.patch
# PATCH-FIX-OPENSUSE skip tests failing due to OBS specifics
Patch3: skip-failing-tests.patch
BuildRequires: %{python_module devel >= 3.7 with %python-devel < 3.10}
BuildRequires: %{python_module devel >= 3.7}
BuildRequires: %{python_module numpy-devel >= 1.18 with %python-numpy-devel < 1.22}
BuildRequires: %{python_module setuptools}
BuildRequires: fdupes
BuildRequires: gcc-c++
BuildRequires: python-rpm-macros
BuildRequires: tbb-devel >= 2021
Requires: python-llvmlite < 0.38
Requires: python-llvmlite >= 0.37
Requires: python-scipy >= 0.16
Requires: (python-llvmlite >= 0.38 with python-llvmlite < 0.39)
Requires: (python-numpy >= 1.18 with python-numpy < 1.22)
Requires(post): update-alternatives
Requires(preun):update-alternatives
Recommends: python-Jinja2
Recommends: python-Pygments
Recommends: python-cffi
Recommends: python-scipy > 1.0
Recommends: python-tbb
%if %{with test}
BuildRequires: %{python_module Jinja2}
@ -69,7 +67,7 @@ BuildRequires: %{python_module numba-devel >= %{version}}
BuildRequires: %{python_module pip}
BuildRequires: %{python_module psutil}
BuildRequires: %{python_module pytest}
BuildRequires: %{python_module scipy >= 0.16}
BuildRequires: %{python_module scipy >= 1.0}
BuildRequires: %{python_module tbb}
%endif
%python_subpackages
@ -102,17 +100,13 @@ Requires: %{plainpython}(abi) = %{python_version}
This package contains files for developing applications using numba.
%prep
%setup -q -n numba-%{version}
%autopatch -p1
%autosetup -p1 -n numba-%{version}
sed -i -e '1{/env python/ d}' numba/misc/appdirs.py
# timeouts randomly in OBS
# random timeouts in OBS
rm numba/tests/test_typedlist.py
# if we reduced the amount of tests too much:
# sed -i -e '/check_testsuite_size/ s/5000/3000/' numba/tests/test_runtests.py
# our setup imports distutils. Not sure why, but should not be a problem.
sed -i -e "/def test_laziness/,/def/ {/'distutils',/ d}" numba/tests/test_import.py
sed -i -e '1{/env python/ d}' numba/misc/appdirs.py
%build
%if !%{with test}
@ -136,13 +130,14 @@ sed 's|^%{buildroot}|%%exclude |' devel-files0-%{$python_bin_suffix}.files > dev
%check
%if %{with test}
mv numba numba_temp
export NUMBA_PARALLEL_DIAGNOSTICS=1
%{python_expand # test the installed package
# test the installed package, not the source without compiled modules
mkdir emtpytestdir
pushd emtpytestdir
%{python_expand # numbatests: check specific tests with `osc build -M test --define="numbatests <testnames>"`
%{_bindir}/numba-%{$python_bin_suffix} -s
$python -m numba.runtests -v -b --exclude-tags='long_running' -m %{_smp_build_ncpus} -- numba.tests
$python -m numba.runtests -v -b --exclude-tags='long_running' -m %{_smp_build_ncpus} -- %{?!numbatests:numba.tests}%{?numbatests}
}
mv numba_temp numba
popd
%endif
%if !%{with test}