nixpkgs/pkgs/development/python-modules/numba/default.nix
2024-03-31 04:48:26 +00:00

131 lines
3.9 KiB
Nix

{ lib
, stdenv
, pythonAtLeast
, pythonOlder
, fetchFromGitHub
, python
, buildPythonPackage
, setuptools
, numpy
, llvmlite
, libcxx
, importlib-metadata
, substituteAll
, runCommand
, config
# CUDA-only dependencies:
, addOpenGLRunpath ? null
, cudaPackages ? {}
# CUDA flags:
, cudaSupport ? config.cudaSupport
}:
let
inherit (cudaPackages) cudatoolkit;
in buildPythonPackage rec {
# Using an untagged version, with numpy 1.25 support, when it's released
# also drop the versioneer patch in postPatch
version = "0.59.1";
pname = "numba";
pyproject = true;
disabled = pythonOlder "3.8" || pythonAtLeast "3.13";
src = fetchFromGitHub {
owner = "numba";
repo = "numba";
rev = "refs/tags/${version}";
# Upstream uses .gitattributes to inject information about the revision
# hash and the refname into `numba/_version.py`, see:
#
# - https://git-scm.com/docs/gitattributes#_export_subst and
# - https://github.com/numba/numba/blame/5ef7c86f76a6e8cc90e9486487294e0c34024797/numba/_version.py#L25-L31
#
# Hence this hash may change if GitHub / Git will change it's behavior.
# Hopefully this will not happen until the next release. We are fairly sure
# that upstream relies on those strings to be valid, that's why we don't
# use `forceFetchGit = true;`.` If in the future we'll observe the hash
# changes too often, we can always use forceFetchGit, and inject the
# relevant strings ourselves, using `sed` commands, in extraPostFetch.
hash = "sha256-4udpgLLHbHNtxPiYVkj+gxAjTWV3ClZOv98Y313/qbc=";
};
env.NIX_CFLAGS_COMPILE = lib.optionalString stdenv.isDarwin "-I${lib.getDev libcxx}/include/c++/v1";
nativeBuildInputs = [
numpy
] ++ lib.optionals cudaSupport [
addOpenGLRunpath
];
propagatedBuildInputs = [
numpy
llvmlite
setuptools
] ++ lib.optionals (pythonOlder "3.9") [
importlib-metadata
] ++ lib.optionals cudaSupport [
cudatoolkit
cudatoolkit.lib
];
patches = lib.optionals cudaSupport [
(substituteAll {
src = ./cuda_path.patch;
cuda_toolkit_path = cudatoolkit;
cuda_toolkit_lib_path = cudatoolkit.lib;
})
];
postFixup = lib.optionalString cudaSupport ''
find $out -type f \( -name '*.so' -or -name '*.so.*' \) | while read lib; do
addOpenGLRunpath "$lib"
patchelf --set-rpath "${cudatoolkit}/lib:${cudatoolkit.lib}/lib:$(patchelf --print-rpath "$lib")" "$lib"
done
'';
# run a smoke test in a temporary directory so that
# a) Python picks up the installed library in $out instead of the build files
# b) we have somewhere to put $HOME so some caching tests work
# c) it doesn't take 6 CPU hours for the full suite
checkPhase = ''
runHook preCheck
pushd $(mktemp -d)
HOME=. ${python.interpreter} -m numba.runtests -m $NIX_BUILD_CORES numba.tests.test_usecases
popd
runHook postCheck
'';
pythonImportsCheck = [
"numba"
];
passthru.tests = {
# CONTRIBUTOR NOTE: numba also contains CUDA tests, though these cannot be run in
# this sandbox environment. Consider running similar commands to those below outside the
# sandbox manually if you have the appropriate hardware; support will be detected
# and the corresponding tests enabled automatically.
# Also, the full suite currently does not complete on anything but x86_64-linux.
fullSuite = runCommand "${pname}-test" {} ''
pushd $(mktemp -d)
# pip and python in $PATH is needed for the test suite to pass fully
PATH=${python.withPackages (p: [ p.numba p.pip ])}/bin:$PATH
HOME=$PWD python -m numba.runtests -m $NIX_BUILD_CORES
popd
touch $out # stop Nix from complaining no output was generated and failing the build
'';
};
meta = with lib; {
description = "Compiling Python code using LLVM";
homepage = "https://numba.pydata.org/";
license = licenses.bsd2;
mainProgram = "numba";
maintainers = with maintainers; [ fridh ];
};
}