Support numba 0.66-0.67 with LLVM 22

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

@@ -64,7 +64,7 @@ jobs:
 
           # Select the python variant; 3.14t needs the free-threaded ABI.
           if [[ "${{ matrix.python-version }}" == "3.14t" ]]; then
-            PY_ARGS=(--python "3.14.* *_cp314t" -m buildscripts/conda-recipes/pyomp/freethreading.yaml)
+            PY_ARGS=(--python "3.14.* *_cp314t")
           else
             PY_ARGS=(--python ${{ matrix.python-version }})
           fi

.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
@@ -85,23 +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', '3.14t']
-        numba-version: ['0.62.0', '0.62.1', '0.63.0', '0.63.1', '0.64.0', '0.65.0', '0.65.1']
-        exclude:
-          - python-version: '3.14'
-            numba-version: '0.62.0'
-          - python-version: '3.14'
-            numba-version: '0.62.1'
-          # Numba supports free-threaded Python starting with 0.65.
-          - python-version: '3.14t'
-            numba-version: '0.62.0'
-          - python-version: '3.14t'
-            numba-version: '0.62.1'
-          - python-version: '3.14t'
-            numba-version: '0.63.0'
-          - python-version: '3.14t'
-            numba-version: '0.63.1'
-          - python-version: '3.14t'
-            numba-version: '0.64.0'
+        numba-version: ['0.66.0', '0.67.0']
 
     steps:
       - name: Download built wheels

README.md

@@ -45,11 +45,12 @@ conda install -c python-for-hpc -c conda-forge pyomp
 
 | PyOMP | Numba           |
 | ----- | --------------- |
+| 0.6.x | 0.66.x - 0.67.x |
 | 0.5.x | 0.62.x - 0.65.x |
 | 0.4.x | 0.61.x          |
 | 0.3.x | 0.57.x - 0.60.x |
 
-Python 3.14t (free-threaded) requires Numba 0.65.x.
+Python 3.14t (free-threaded) requires PyOMP 0.5.2 or later and Numba 0.65.x or later.
 
 Besides a standard installation, we also provide the following options to
 quickly try out PyOMP online or through a container.

buildscripts/cibuildwheel/setup-miniforge3.sh

@@ -58,6 +58,7 @@ bash _downloads/miniforge3.sh -b -f -p "_stage/miniforge3"
 echo "Miniforge installed"
 source "_stage/miniforge3/bin/activate" base
 
-# Create conda environment with tools and libraries for the LLVM_VERSION.
+# Create conda environment with tools and libraries for the LLVM_VERSION. lld
+# links the AMDGPU device runtime.
 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/conda_build_config.yaml

@@ -1,8 +1,3 @@
-# --- Free-threading ---
-# Default for the is_freethreading selector in meta.yaml; override for 3.14t.
-is_freethreading:
-  - false
-
 # --- Compilers ---
 c_compiler_version:
   - 9                       # [linux]

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] %}
 
@@ -45,26 +45,24 @@ requirements:
     - sysroot_linux-aarch64  # [aarch64]
     - setuptools
     - setuptools_scm
-    - numba >=0.62, <0.66
-    - numba >=0.65, <0.66  # [is_freethreading]
+    - numba >=0.66, <0.68
     - clang {{ LLVM_VERSION }}
     - clangxx {{ LLVM_VERSION }}
     - clang-tools {{ LLVM_VERSION }}
     - llvmdev {{ LLVM_VERSION }}
+    # lld links the AMDGPU device runtime.
+    - 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, <0.66
-    - numba >=0.65, <0.66  # [is_freethreading]
+    - numba >=0.66, <0.68
     # 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 = "0.65.1"
+NUMBA_VERSION = "0.67.0"
 MINIFORGE_VERSION = "26.3.2-3"
 MINIFORGE_SHA256 = "848194851a98903134187fbb4ab50efe87b003e0c0f808f97644b7524a62bf2c"
 WHEEL_DIRECTORY = (

docs/source/installation.rst

@@ -24,14 +24,16 @@ supported.
 +--------+---------------------+
 | PyOMP  | Numba               |
 +========+=====================+
+| 0.6.x  | 0.66.x - 0.67.x     |
++--------+---------------------+
 | 0.5.x  | 0.62.x - 0.65.x     |
 +--------+---------------------+
 | 0.4.x  | 0.61.x              |
 +--------+---------------------+
 | 0.3.x  | 0.57.x - 0.60.x     |
 +--------+---------------------+
 
-Python 3.14t (free-threaded) requires Numba 0.65.x.
+Python 3.14t (free-threaded) requires PyOMP 0.5.2 or later and Numba 0.65.x or later.
 
 Additional options
 ------------------

docs/source/openmp.rst

@@ -356,7 +356,8 @@ The following table shows tested combinations of PyOMP, Numba, Python, LLVM, and
    ===================== ==================== ==================== ============ ================================
    PyOMP                 Numba                Python               LLVM         Supported Platforms
    ===================== ==================== ==================== ============ ================================
-   0.5.x                 0.62.x - 0.63.x      3.10 - 3.14          20.x         linux-64, osx-arm64, linux-arm64
+   0.6.x                 0.66.x - 0.67.x      3.10 - 3.14          22.x         linux-64, osx-arm64, linux-arm64
+   0.5.x                 0.62.x - 0.65.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
    ===================== ==================== ==================== ============ ================================
@@ -384,7 +385,9 @@ Platform details
 Notes
 ^^^^^
 
-* Python 3.14 free-threaded build (cp314t) is not supported with the current Numba/llvmlite version.
-* LLVM version 20.1.8 is used for the current PyOMP 0.5.x releases.
+* Python 3.14 free-threaded builds (cp314t) are supported starting with PyOMP
+  0.5.2 (Numba 0.65.x).
+* LLVM version 22.1.8 is used for the current PyOMP 0.6.x releases, and LLVM
+  version 20.1.8 for the PyOMP 0.5.x releases.
 * For GPU offloading support, NVIDIA GPU and NVIDIA driver are required on supported Linux platforms.
 * AMD GPU support is in active development.

pyproject.toml

@@ -17,7 +17,7 @@ classifiers = [
     "Intended Audience :: Developers",
     "Topic :: Software Development :: Compilers",
 ]
-dependencies = ["numba>=0.62, <0.66", "lark", "cffi", "setuptools"]
+dependencies = ["numba>=0.66, <0.68", "lark", "cffi", "setuptools"]
 maintainers = [
     { name = "Giorgis Georgakoudis", email = "georgakoudis1@llnl.gov" },
 ]
@@ -55,16 +55,23 @@ USE_CXX11_ABI = "1"
 PIP_NO_INPUT = "1"
 
 [tool.cibuildwheel.linux]
+# The manylinux image has no LLVM_VERSION packages, so use the conda-forge
+# toolchain. libzstd-devel is needed by the conda LLVM cmake exports.
 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 libzstd-devel",
+    "bash buildscripts/cibuildwheel/setup-miniforge3.sh",
 ]
 
 [tool.cibuildwheel.linux.environment]
 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++"
+# Compile against the image's gcc-toolset libstdc++ instead of the conda one
+# that clang selects by default, so the libraries stay manylinux compatible.
+CFLAGS = "--gcc-toolchain=/opt/rh/gcc-toolset-14/root/usr"
+CXXFLAGS = "--gcc-toolchain=/opt/rh/gcc-toolset-14/root/usr"
 
 [tool.cibuildwheel.macos]
 before-all = ["bash buildscripts/cibuildwheel/setup-miniforge3.sh"]

setup.py

@@ -1,5 +1,6 @@
 from pathlib import Path
 import subprocess
+import shlex
 import shutil
 import tarfile
 import urllib
@@ -133,6 +134,9 @@ def _build_cmake(self, ext: CMakeExtension):
             include_dir = install_dir / "lib/cmake"
             if include_dir.exists():
                 shutil.rmtree(include_dir)
+        # Keep only the device runtime bitcode, not its static archive.
+        if ext.name.startswith("devicertl"):
+            (install_dir / "lib/libompdevice.a").unlink(missing_ok=True)
         # Remove symlinks in the install directory to avoid copies.
         for file in install_dir.rglob("*"):
             if file.is_symlink():
@@ -209,9 +213,11 @@ def _build_loader_wrapper(self):
         loader_so = target_lib_dir / f"libpyomp_loader{lib_ext}"
         cc = os.environ.get("CC", "gcc")
 
-        # Compile the wrapper.
+        # Compile the wrapper. Pass CFLAGS so it uses the same toolchain as the
+        # other libraries, e.g., the manylinux --gcc-toolchain.
         cmd = [
             cc,
+            *shlex.split(os.environ.get("CFLAGS", "")),
             "-shared",
             "-fPIC",
             str(loader_c),
@@ -359,13 +365,43 @@ def _check_true(env_var):
             cmake_args=[
                 "-DOPENMP_STANDALONE_BUILD=ON",
                 "-DLLVM_ENABLE_RUNTIMES=offload",
+                # Tests depend on liboffload, which is not built.
+                "-DOFFLOAD_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",
             ],
         )
     )
 
+    # Build the device runtime bitcode (libomptarget-<target>.bc) for each GPU
+    # target, compiled by clang from openmp/device.
+    for triple in ("nvptx64-nvidia-cuda", "amdgcn-amd-amdhsa"):
+        ext_modules.append(
+            CMakeExtension(
+                f"devicertl-{triple}",
+                setup=PrepareOpenMP,
+                source_dir=PrepareOpenMP.get_source_dir(),
+                install_dir="openmp",
+                cmake_args=[
+                    "-DOPENMP_STANDALONE_BUILD=ON",
+                    "-DLLVM_ENABLE_RUNTIMES=openmp",
+                    f"-DLLVM_RUNTIMES_TARGET={triple}",
+                    f"-DLLVM_DEFAULT_TARGET_TRIPLE={triple}",
+                    f"-DCMAKE_C_COMPILER_TARGET={triple}",
+                    f"-DCMAKE_CXX_COMPILER_TARGET={triple}",
+                    # The compiler cannot link host test programs for GPU
+                    # targets.
+                    "-DCMAKE_C_COMPILER_WORKS=ON",
+                    "-DCMAKE_CXX_COMPILER_WORKS=ON",
+                    # Ignore host flags from CFLAGS and CXXFLAGS, such as
+                    # conda's -march, which fail for GPU targets.
+                    "-DCMAKE_C_FLAGS=",
+                    "-DCMAKE_CXX_FLAGS=",
+                ],
+            )
+        )
+
 
 setup(
     ext_modules=ext_modules,