Support LLVM 22 with Numba 0.66

.github/workflows/build-upload-wheels.yml

@@ -52,7 +52,7 @@ jobs:
         # Set LLVM_VERSION for the host to forward to the cibuildwheel
         # environment.
         env:
-          LLVM_VERSION: "20.1.8"
+          LLVM_VERSION: "22.1.8"
         run: python -m cibuildwheel --output-dir wheelhouse
 
       - uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02
@@ -69,7 +69,7 @@ jobs:
 
       - name: Build sdist
         env:
-          LLVM_VERSION: "20.1.8"
+          LLVM_VERSION: "22.1.8"
         run: pipx run build --sdist
 
       - uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02
@@ -85,12 +85,7 @@ jobs:
       matrix:
         os: [ubuntu-latest, macos-latest, ubuntu-24.04-arm]
         python-version: ['3.10', '3.11', '3.12', '3.13', '3.14']
-        numba-version: ['0.62.0', '0.62.1', '0.63.0', '0.63.1']
-        exclude:
-          - python-version: '3.14'
-            numba-version: '0.62.0'
-          - python-version: '3.14'
-            numba-version: '0.62.1'
+        numba-version: ['0.66.0']
 
     steps:
       - name: Download built wheels

README.md

@@ -45,6 +45,7 @@ conda install -c python-for-hpc -c conda-forge pyomp
 
 | PyOMP | Numba           |
 | ----- | --------------- |
+| 0.6.x | 0.66.x          |
 | 0.5.x | 0.62.x - 0.63.x |
 | 0.4.x | 0.61.x          |
 | 0.3.x | 0.57.x - 0.60.x |

buildscripts/cibuildwheel/setup-miniforge3.sh

@@ -60,4 +60,4 @@ source "_stage/miniforge3/bin/activate" base
 
 # Create conda environment with tools and libraries for the LLVM_VERSION.
 echo "Installing llvmdev ${LLVM_VERSION}..."
-conda create -n llvmdev-${LLVM_VERSION} --override-channels -c conda-forge -q -y clang=${LLVM_VERSION} clangxx=${LLVM_VERSION} clang-tools=${LLVM_VERSION} llvmdev=${LLVM_VERSION} zstd
+conda create -n llvmdev-${LLVM_VERSION} --override-channels -c conda-forge -q -y clang=${LLVM_VERSION} clangxx=${LLVM_VERSION} clang-tools=${LLVM_VERSION} llvmdev=${LLVM_VERSION} lld=${LLVM_VERSION} zstd

buildscripts/conda-recipes/pyomp/meta.yaml

@@ -1,5 +1,5 @@
 {% set version = environ.get('GIT_DESCRIBE_TAG', '0.0.0').lstrip('v') %}
-{% set LLVM_VERSION = environ.get('LLVM_VERSION', '20.1.8') %}
+{% set LLVM_VERSION = environ.get('LLVM_VERSION', '22.1.8') %}
 {% set LLVM_VERSION_MAJOR = LLVM_VERSION.split('.')[0] %}
 {% set LLVM_VERSION_MINOR = LLVM_VERSION.split('.')[1] %}
 
@@ -34,6 +34,7 @@ requirements:
     - sysroot_linux-aarch64  # [aarch64]
     - cmake
     - ninja
+    - patch # [linux]
     - setuptools_scm
     - elfutils # [linux]
     - libffi # [linux]
@@ -45,24 +46,23 @@ requirements:
     - sysroot_linux-aarch64  # [aarch64]
     - setuptools
     - setuptools_scm
-    - numba >=0.62, <0.64
+    - numba >=0.66, <0.67
     - clang {{ LLVM_VERSION }}
     - clangxx {{ LLVM_VERSION }}
     - clang-tools {{ LLVM_VERSION }}
     - llvmdev {{ LLVM_VERSION }}
+    - lld {{ LLVM_VERSION }} # [linux]
     - zlib
     # require llvm-openmp for the openmp cpu runtime.
-    - llvm-openmp 21.1.8.* # [osx]
-    - llvm-openmp {{ LLVM_VERSION }} # [not osx]
+    - llvm-openmp {{ LLVM_VERSION }}
     - elfutils # [linux]
     - libffi # [linux]
   run:
     - python
     - setuptools
-    - numba >=0.62, <0.64
+    - numba >=0.66, <0.67
     # require llvm-openmp for the openmp cpu runtime.
-    - llvm-openmp 21.1.8.* # [osx]
-    - llvm-openmp {{ LLVM_VERSION }} # [not osx]
+    - llvm-openmp {{ LLVM_VERSION }}
     - lark
     - cffi
 

buildscripts/conda-recipes/pyomp/run_test.sh

@@ -12,7 +12,7 @@ export PYTHONFAULTHANDLER=1
 # of low accuracy SVML libm replacements in ufunc loops.
 _NPY_CMD='from numba.misc import numba_sysinfo;\
           sysinfo=numba_sysinfo.get_sysinfo();\
-          print(sysinfo["NumPy AVX512_SKX detected"] and
+          print("AVX512_SKX" in sysinfo["NumPy Supported SIMD features"] and
                 sysinfo["NumPy Version"]>="1.22")'
 NUMPY_DETECTS_AVX512_SKX_NP_GT_122=$(python -c "$_NPY_CMD")
 echo "NumPy >= 1.22 with AVX512_SKX detected: $NUMPY_DETECTS_AVX512_SKX_NP_GT_122"

buildscripts/modal/test_gpu.py

@@ -8,7 +8,7 @@
 
 
 PYTHON_VERSIONS = ("3.10", "3.11", "3.12", "3.13", "3.14")
-NUMBA_VERSION = "0.63.1"
+NUMBA_VERSION = "0.66.0"
 MINIFORGE_VERSION = "26.3.2-3"
 MINIFORGE_SHA256 = "848194851a98903134187fbb4ab50efe87b003e0c0f808f97644b7524a62bf2c"
 WHEEL_DIRECTORY = (

docs/source/installation.rst

@@ -24,6 +24,8 @@ supported.
 +--------+---------------------+
 | PyOMP  | Numba               |
 +========+=====================+
+| 0.6.x  | 0.66.x              |
++--------+---------------------+
 | 0.5.x  | 0.62.x - 0.63.x     |
 +--------+---------------------+
 | 0.4.x  | 0.61.x              |

docs/source/openmp.rst

@@ -356,6 +356,7 @@ The following table shows tested combinations of PyOMP, Numba, Python, LLVM, and
    ===================== ==================== ==================== ============ ================================
    PyOMP                 Numba                Python               LLVM         Supported Platforms
    ===================== ==================== ==================== ============ ================================
+   0.6.x                 0.66.x               3.10 - 3.14          22.x         linux-64, osx-arm64, linux-arm64
    0.5.x                 0.62.x - 0.63.x      3.10 - 3.14          20.x         linux-64, osx-arm64, linux-arm64
    0.4.x                 0.61.x               3.10 - 3.13          15.x         linux-64, osx-arm64, linux-arm64
    0.3.x                 0.57.x - 0.60.x      3.9 - 3.12           14.x         linux-64, osx-arm64, linux-arm64

pyproject.toml

@@ -17,7 +17,7 @@ classifiers = [
     "Intended Audience :: Developers",
     "Topic :: Software Development :: Compilers",
 ]
-dependencies = ["numba>=0.62, <0.64", "lark", "cffi", "setuptools"]
+dependencies = ["numba>=0.66, <0.67", "lark", "cffi", "setuptools"]
 maintainers = [
     { name = "Giorgis Georgakoudis", email = "georgakoudis1@llnl.gov" },
 ]
@@ -57,15 +57,18 @@ PIP_NO_INPUT = "1"
 
 [tool.cibuildwheel.linux]
 before-all = [
-    "yum install -y elfutils-libelf-devel libffi-devel clang-devel-20.1.8 llvm-devel-20.1.8",
+    "yum install -y elfutils-libelf-devel libffi-devel patch",
+    "bash buildscripts/cibuildwheel/setup-miniforge3.sh",
 ]
 
 [tool.cibuildwheel.linux.environment]
+# Conda's C++ runtime is newer than the manylinux system runtime.
+LDFLAGS = "-static-libstdc++ -static-libgcc"
 ENABLE_BUNDLED_LIBOMP = "1"
 ENABLE_BUNDLED_LIBOMPTARGET = "1"
-LLVM_DIR = "/usr/lib64/cmake/llvm"
-CC = "/usr/bin/clang"
-CXX = "/usr/bin/clang++"
+LLVM_DIR = "${PWD}/_stage/miniforge3/envs/llvmdev-${LLVM_VERSION}/lib/cmake/llvm"
+CC = "${PWD}/_stage/miniforge3/envs/llvmdev-${LLVM_VERSION}/bin/clang"
+CXX = "${PWD}/_stage/miniforge3/envs/llvmdev-${LLVM_VERSION}/bin/clang++"
 
 [tool.cibuildwheel.macos]
 before-all = ["bash buildscripts/cibuildwheel/setup-miniforge3.sh"]

setup.py

@@ -84,6 +84,10 @@ def _build_cmake(self, ext: CMakeExtension):
             lib_dir = Path(self.build_lib) / "numba/openmp/libs"
 
         extra_cmake_args = self._env_toolchain_args(ext)
+        if ext.name.startswith("libompdevice-"):
+            extra_cmake_args.append(
+                f"-DLLVM_LIBRARY_OUTPUT_INTDIR={build_dir.absolute() / 'lib'}"
+            )
         # Set RPATH.
         if sys.platform.startswith("linux"):
             extra_cmake_args.append(r"-DCMAKE_INSTALL_RPATH=$ORIGIN")
@@ -359,13 +363,32 @@ def _check_true(env_var):
             cmake_args=[
                 "-DOPENMP_STANDALONE_BUILD=ON",
                 "-DLLVM_ENABLE_RUNTIMES=offload",
+                "-DLLVM_INCLUDE_TESTS=OFF",
                 # Avoid conflicts in manylinux builds with packaged clang/llvm
                 # under /usr/include and its gcc-toolset provided header files.
                 "-DCMAKE_NO_SYSTEM_FROM_IMPORTED=ON",
             ],
         )
     )
 
+    # LLVM 22 builds GPU device bitcode separately from libomptarget.
+    for target in ("nvptx64-nvidia-cuda", "amdgcn-amd-amdhsa"):
+        ext_modules.append(
+            CMakeExtension(
+                f"libompdevice-{target}",
+                setup=PrepareOpenMP,
+                source_dir=PrepareOpenMP.get_source_dir().parent / "openmp",
+                install_dir="openmp",
+                cmake_args=[
+                    f"-DLLVM_DEFAULT_TARGET_TRIPLE={target}",
+                    "-DLLVM_INCLUDE_TESTS=OFF",
+                    # Host CPU flags from Conda do not apply to GPU bitcode.
+                    "-DCMAKE_C_FLAGS=",
+                    "-DCMAKE_CXX_FLAGS=",
+                ],
+            )
+        )
+
 
 setup(
     ext_modules=ext_modules,

src/numba/openmp/libs/openmp/patches/22.1.8/0002-Link-statically-LLVM.patch

@@ -0,0 +1,20 @@
+--- a/offload/plugins-nextgen/CMakeLists.txt
++++ b/offload/plugins-nextgen/CMakeLists.txt
+@@ -4,6 +4,7 @@
+ add_subdirectory(common)
+ function(add_target_library target_name lib_name)
+   add_llvm_library(${target_name} STATIC
++    DISABLE_LLVM_LINK_LLVM_DYLIB
+     LINK_COMPONENTS
+       AggressiveInstCombine
+       Analysis
+--- a/offload/libomptarget/CMakeLists.txt
++++ b/offload/libomptarget/CMakeLists.txt
+@@ -8,6 +8,7 @@
+
+ add_llvm_library(omptarget
+   SHARED
++  DISABLE_LLVM_LINK_LLVM_DYLIB
+
+   device.cpp
+   interface.cpp

src/numba/openmp/libs/openmp/patches/22.1.8/0003-Do-not-build-liboffload.patch

@@ -0,0 +1,10 @@
+--- a/offload/CMakeLists.txt
++++ b/offload/CMakeLists.txt
+@@ -391,7 +391,6 @@
+ # Build target agnostic offloading library.
+ add_subdirectory(libomptarget)
+
+-add_subdirectory(liboffload)
+
+ # Add tests.
+ if(OFFLOAD_INCLUDE_TESTS)