Add CI workflow to build and test dpnp with the open source LLVM SYCL compiler

.github/workflows/bump-sycl-deps.yml

@@ -0,0 +1,139 @@
+name: Bump open source SYCL build deps
+
+# Keeps the pinned oclcpuexp driver (intel/llvm YYYY-WWNN release) and oneTBB
+# release used by os-llvm-sycl-build.yml up to date by opening a pull request
+# when a newer upstream release is available.
+
+on:
+  # Check weekly on Saturday, and allow manual runs.
+  schedule:
+    - cron: "0 6 * * 6"
+  workflow_dispatch:
+
+permissions: read-all
+
+# Only one bump run at a time; a newer run supersedes an in-progress one.
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  bump:
+    name: Bump oclcpuexp and oneTBB pins
+
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+
+    permissions:
+      contents: write
+      pull-requests: write
+
+    env:
+      WF: .github/workflows/os-llvm-sycl-build.yml
+      GH_TOKEN: ${{ github.token }}
+
+    steps:
+      - name: Checkout repo
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+
+      - name: Find latest intel/llvm driver release tag
+        id: driver
+        uses: oprypin/find-latest-tag@6957ac556fa6d349727ecabfcaaf9f8e5ee37124 # v1.1.3
+        with:
+          repository: intel/llvm
+          releases-only: true
+          # Driver releases are tagged like "2026-WW28"; ignore nightly-* tags.
+          regex: '^\d{4}-WW\d+$'
+          # Pick the most recently published matching release. Must stay false:
+          # sort-tags: true would semver-sort and mis-parse the YYYY-WWNN tag.
+          sort-tags: false
+
+      - name: Find latest oneTBB release tag
+        id: tbb
+        uses: oprypin/find-latest-tag@6957ac556fa6d349727ecabfcaaf9f8e5ee37124 # v1.1.3
+        with:
+          repository: uxlfoundation/oneTBB
+          releases-only: true
+          regex: '^v\d+\.\d+\.\d+$'
+          # Most recently published matching release (see note above).
+          sort-tags: false
+
+      - name: Resolve asset names and derive values
+        id: resolve
+        env:
+          DRIVER_TAG: ${{ steps.driver.outputs.tag }}
+          TBB_TAG: ${{ steps.tbb.outputs.tag }}
+        run: |
+          set -euo pipefail
+
+          # The oclcpuexp filename embeds its own version and is not derivable
+          # from the tag, so read it back from the driver release assets.
+          OCLCPUEXP_FN=$(gh api "repos/intel/llvm/releases/tags/${DRIVER_TAG}" \
+                          --jq '.assets[].name' \
+                          | grep -E '^oclcpuexp-.*_rel\.tar\.gz$' | head -1)
+          if [[ -z "${OCLCPUEXP_FN}" ]]; then
+              echo "::error::no oclcpuexp asset found in intel/llvm release ${DRIVER_TAG}"
+              exit 1
+          fi
+
+          # oneTBB filenames are derived from the tag (e.g. v2023.1.0).
+          TBB_VER=${TBB_TAG#v}
+
+          {
+            echo "driver_tag=${DRIVER_TAG}"
+            echo "oclcpuexp_fn=${OCLCPUEXP_FN}"
+            echo "tbb_tag=${TBB_TAG}"
+            echo "tbb_fn=oneapi-tbb-${TBB_VER}-lin.tgz"
+            echo "tbb_dir=oneapi-tbb-${TBB_VER}"
+          } >> "${GITHUB_OUTPUT}"
+
+      - name: Apply pins to workflow file
+        env:
+          DRIVER_TAG: ${{ steps.resolve.outputs.driver_tag }}
+          OCLCPUEXP_FN: ${{ steps.resolve.outputs.oclcpuexp_fn }}
+          TBB_TAG: ${{ steps.resolve.outputs.tbb_tag }}
+          TBB_FN: ${{ steps.resolve.outputs.tbb_fn }}
+          TBB_DIR: ${{ steps.resolve.outputs.tbb_dir }}
+        run: |
+          set -euo pipefail
+          # Surgical, comment-preserving edits: rewrite only the value after each
+          # key and keep indentation. "|" delimiter so the "/"-heavy URL is safe.
+          sed -i -E \
+            -e "s|^( *DRIVER_PATH: ).*|\1${DRIVER_TAG}|" \
+            -e "s|^( *OCLCPUEXP_FN: ).*|\1${OCLCPUEXP_FN}|" \
+            -e "s|^( *TBB_URL: ).*|\1https://github.com/uxlfoundation/oneTBB/releases/download/${TBB_TAG}/|" \
+            -e "s|^( *TBB_INSTALL_DIR: ).*|\1${TBB_DIR}|" \
+            -e "s|^( *TBB_FN: ).*|\1${TBB_FN}|" \
+            "${WF}"
+          echo "---- resulting diff ----"
+          git diff -- "${WF}" || true
+
+      - name: Create pull request
+        uses: peter-evans/create-pull-request@5f6978faf089d4d20b00c7766989d076bb2fc7f1 # v8.1.1
+        with:
+          token: ${{ github.token }}
+          add-paths: ${{ env.WF }}
+          branch: bot/bump-sycl-build-deps
+          delete-branch: true
+          labels: autoupdate
+          commit-message: |
+            Bump oclcpuexp/oneTBB pins in os-llvm-sycl-build workflow
+
+            driver: ${{ steps.resolve.outputs.driver_tag }} / ${{ steps.resolve.outputs.oclcpuexp_fn }}
+            oneTBB: ${{ steps.resolve.outputs.tbb_tag }}
+          title: "Bump open source SYCL build dependencies"
+          body: |
+            Automated bump of the pinned open source SYCL build dependencies in
+            `os-llvm-sycl-build.yml`.
+
+            | dependency | new value |
+            | --- | --- |
+            | intel/llvm driver release | `${{ steps.resolve.outputs.driver_tag }}` |
+            | oclcpuexp | `${{ steps.resolve.outputs.oclcpuexp_fn }}` |
+            | oneTBB | `${{ steps.resolve.outputs.tbb_tag }}` |
+
+            The nightly DPC++ compiler is discovered at run time and is not
+            affected by this change.
+
+            Trigger the **Build with Open Source LLVM SYCL compiler** workflow on
+            this branch to validate before merging.

.github/workflows/os-llvm-sycl-build.yml

@@ -0,0 +1,201 @@
+name: Build with Open Source LLVM SYCL compiler
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+  # To be able to be triggered manually
+  workflow_dispatch:
+
+permissions: read-all
+
+# Only one build at a time per ref; a newer run supersedes an in-progress one.
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  install-compiler:
+    name: Build with nightly build of DPC++ toolchain
+
+    runs-on: ubuntu-latest
+    timeout-minutes: 120
+
+    defaults:
+      run:
+        shell: bash -el {0}
+
+    env:
+      DOWNLOAD_URL_PREFIX: https://github.com/intel/llvm/releases/download
+      DRIVER_PATH: 2026-WW28
+      OCLCPUEXP_FN: oclcpuexp-2026.22.6.1.17_160000_rel.tar.gz
+      TBB_URL: https://github.com/uxlfoundation/oneTBB/releases/download/v2023.1.0/
+      TBB_INSTALL_DIR: oneapi-tbb-2023.1.0
+      TBB_FN: oneapi-tbb-2023.1.0-lin.tgz
+      dpctl-repo-path: '${{ github.workspace }}/dpctl'
+
+    steps:
+      - name: Install hwloc
+        run: sudo apt install hwloc
+
+      - name: Install lld
+        run: sudo apt install lld
+
+      - name: Download and install nightly and components
+        env:
+          ARTIFACT_NAME: sycl_linux
+          USE_LATEST_SYCLOS: 1
+        run: |
+          cd /home/runner/work
+          mkdir -p sycl_bundle
+          cd sycl_bundle
+          if [[ "${USE_LATEST_SYCLOS:-0}" -eq "1" ]]; then
+              # get list of shas and tags from remote, filter nightly tags and reverse order
+              LLVM_TAGS=$(git -c 'versionsort.suffix=-' ls-remote --tags --sort='v:refname' https://github.com/intel/llvm.git | \
+                          grep 'refs/tags/nightly-' | awk '{a[i++]=$0} END {for (j=i-1; j>=0;) print a[j--] }')
+              export LLVM_TAGS
+              # initialize
+              unset DEPLOY_NIGHTLY_TAG
+              unset DEPLOY_LLVM_TAG_SHA
+
+              # go through tags and find the most recent one where nighly build binary is available
+              while IFS= read -r NEXT_LLVM_TAG; do
+                  NEXT_LLVM_TAG_SHA=$(echo "${NEXT_LLVM_TAG}" | awk '{print $1}')
+                  export NEXT_LLVM_TAG_SHA
+                  NEXT_NIGHTLY_TAG=$(python3 -c "import sys, urllib.parse as ul; print(ul.quote_plus(sys.argv[1]))" \
+                                      "$(echo "${NEXT_LLVM_TAG}" | awk '{gsub(/^refs\/tags\//, "", $2)} {print $2}')")
+                  export NEXT_NIGHTLY_TAG
+                  if wget -S --spider "${DOWNLOAD_URL_PREFIX}/${NEXT_NIGHTLY_TAG}/${ARTIFACT_NAME}.tar.gz" 2>&1 | grep -q 'HTTP/1.1 200 OK';
+                  then
+                      DEPLOY_NIGHTLY_TAG="${NEXT_NIGHTLY_TAG}"
+                      export DEPLOY_NIGHTLY_TAG
+                      DEPLOY_LLVM_TAG_SHA="${NEXT_LLVM_TAG_SHA}"
+                      export DEPLOY_LLVM_TAG_SHA
+                      break
+                  fi
+              done <<< "${LLVM_TAGS}"
+          else
+              # Use latest known to work tag instead
+              DEPLOY_NIGHTLY_TAG="sycl-nightly%2F20230606"
+              export DEPLOY_NIGHTLY_TAG
+              DEPLOY_LLVM_TAG_SHA=f44d0133d4b0077298f034697a1f3818ff1d6134
+              export DEPLOY_LLVM_TAG_SHA
+          fi
+
+          [[ -n "${DEPLOY_NIGHTLY_TAG}" ]] || exit 1
+          [[ -n "${DEPLOY_LLVM_TAG_SHA}" ]] || exit 1
+          echo "Using ${DEPLOY_NIGHTLY_TAG} corresponding to intel/llvm at ${DEPLOY_LLVM_TAG_SHA}"
+
+          rm -rf "${ARTIFACT_NAME}.tar.gz"
+          wget "${DOWNLOAD_URL_PREFIX}/${DEPLOY_NIGHTLY_TAG}/${ARTIFACT_NAME}.tar.gz"
+          wget "${DOWNLOAD_URL_PREFIX}/${DRIVER_PATH}/${OCLCPUEXP_FN}"
+          wget "${TBB_URL}/${TBB_FN}"
+          mkdir -p dpcpp_compiler
+          tar xf "${ARTIFACT_NAME}.tar.gz" -C dpcpp_compiler
+          mkdir -p oclcpuexp
+          tar xf "${OCLCPUEXP_FN}" -C oclcpuexp
+          tar xf "${TBB_FN}"
+          cp oclcpuexp/x64/libOpenCL.so* dpcpp_compiler/lib/
+
+      - name: Setup Python
+        uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
+        with:
+          python-version: '3.14'
+
+      - name: Checkout repo
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+        with:
+          fetch-depth: 0
+
+      - name: Checkout dpctl repo
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+        with:
+          repository: IntelPython/dpctl
+          fetch-depth: 0
+          path: ${{ env.dpctl-repo-path }}
+
+      - name: Install dpnp dependencies
+        run: |
+          pip install numpy cython setuptools"<80" pytest scikit-build cmake ninja versioneer[toml]==0.29 \
+              mkl-devel-dpcpp onedpl-devel
+
+      - name: Create set_allvars.sh
+        run: |
+          cat << 'EOF' > set_allvars.sh
+          #!/usr/bin/bash
+          export SYCL_BUNDLE_FOLDER=/home/runner/work/sycl_bundle
+          export PATH=${SYCL_BUNDLE_FOLDER}/dpcpp_compiler/bin:${PATH}
+          export LD_LIBRARY_PATH=${SYCL_BUNDLE_FOLDER}/dpcpp_compiler/lib:${LD_LIBRARY_PATH}
+          export LD_LIBRARY_PATH=${SYCL_BUNDLE_FOLDER}/oclcpuexp/x64:${LD_LIBRARY_PATH}
+          export LD_LIBRARY_PATH=${SYCL_BUNDLE_FOLDER}/${TBB_INSTALL_DIR}/lib/intel64/gcc4.8:${LD_LIBRARY_PATH}
+          export OCL_ICD_VENDORS=
+          export OCL_ICD_FILENAMES=libintelocl.so
+          # Help CMake find MKL/oneDPL/TBB shipped in the Python environment.
+          PYTHON_PREFIX=$(python -c "import sys, os; print(os.path.dirname(os.path.dirname(sys.executable)))")
+          export MKLROOT=${PYTHON_PREFIX}
+          export oneDPL_ROOT=${PYTHON_PREFIX}
+          export TBBROOT=${SYCL_BUNDLE_FOLDER}/${TBB_INSTALL_DIR}
+          # The open source DPC++ compiler does not bundle oneDPL/MKL headers.
+          # dpnp extensions include e.g. <oneapi/dpl/numeric> without linking the
+          # oneDPL target, so expose the pip package headers on the default
+          # compiler include path. Use the *_INCLUDE_PATH vars (treated as
+          # -isystem) rather than CPATH (-I) so deprecation warnings from these
+          # headers stay suppressed.
+          export C_INCLUDE_PATH=${PYTHON_PREFIX}/include:${C_INCLUDE_PATH}
+          export CPLUS_INCLUDE_PATH=${PYTHON_PREFIX}/include:${CPLUS_INCLUDE_PATH}
+          EOF
+          chmod +x set_allvars.sh
+          cat set_allvars.sh
+
+      - name: Report compiler version
+        run: |
+          source set_allvars.sh
+          clang++ --version
+
+      - name: Run sycl-ls
+        run: |
+          source set_allvars.sh
+          sycl-ls
+
+      - name: Build and install dpctl
+        run: |
+          source "${GITHUB_WORKSPACE}/set_allvars.sh"
+          # Build dpctl from source with the same nightly so both
+          # share one SYCL runtime.
+          #
+          # Install non-editable so scikit-build runs CMake's install step and
+          # produces a proper package layout (include/, resources/cmake with the
+          # config) that dpnp's find_package(Dpctl) needs. --no-deps avoids
+          # pulling intel-sycl-rt from PyPI, which would reintroduce the
+          # conflicting released libsycl (the nightly one is on LD_LIBRARY_PATH).
+          CC=clang CXX=clang++ pip install --no-build-isolation --no-deps . || exit 1
+        working-directory: ${{ env.dpctl-repo-path }}
+
+      - name: Build dpnp
+        run: |
+          source set_allvars.sh
+          # Build against the oneMath interface, so the math layer is compiled
+          # with the same open source nightly compiler.
+          #
+          # The open source DPC++ nightly ships an IntelSYCLConfig.cmake that
+          # lacks add_sycl_to_target(), so force the vendored module which
+          # provides it.
+          python scripts/build_locally.py --c-compiler=clang --cxx-compiler=clang++ \
+              --compiler-root="${SYCL_BUNDLE_FOLDER}/dpcpp_compiler/bin" \
+              --onemath \
+              --cmake-opts="-DIntelSYCL_DIR=${GITHUB_WORKSPACE}/dpnp/backend/cmake/Modules" || exit 1
+
+      - name: Smoke test
+        run: |
+          source set_allvars.sh
+          python -m dpctl -f
+          python -c "import dpnp; print(dpnp.__version__)"
+
+      - name: Run dpnp/tests
+        env:
+          SYCL_CACHE_PERSISTENT: 1
+          SKIP_TENSOR_TESTS: 0
+        run: |
+          source set_allvars.sh
+          python -m pytest -ra dpnp/tests

dpnp/backend/cmake/Modules/IntelSYCLConfig.cmake

@@ -172,6 +172,18 @@ function(parse_compiler_version compiler_name version_number)
         # Parse Intel Compiler Version
         string(REGEX REPLACE "Intel\\(R\\) (.*) Compiler ([0-9]+\\.[0-9]+\\.[0-9]+) (.*)" "\\2"
                SYCL_VERSION_STRING_MATCH ${INTEL_VERSION_STRING})
+    else()
+        # Open source DPC++/LLVM SYCL compiler reports e.g.
+        # "clang version 23.0.0git (https://github.com/intel/llvm ...)"
+        string(REGEX MATCH "clang version ([0-9]+\\.[0-9]+\\.[0-9]+)"
+               CLANG_VERSION_STRING ${COMPILER_VERSION_STRING})
+        if(CLANG_VERSION_STRING)
+            string(REGEX REPLACE "clang version ([0-9]+\\.[0-9]+\\.[0-9]+)" "\\1"
+                   SYCL_VERSION_STRING_MATCH ${CLANG_VERSION_STRING})
+        endif()
+    endif()
+
+    if(SYCL_VERSION_STRING_MATCH)
         string(REPLACE "." ";" SYCL_VERSION_LIST ${SYCL_VERSION_STRING_MATCH})
         list(GET SYCL_VERSION_LIST 0 VERSION_MAJOR)
         list(GET SYCL_VERSION_LIST 1 VERSION_MINOR)

dpnp/backend/extensions/ufunc/elementwise_functions/erf_funcs.cpp

@@ -87,7 +87,7 @@ struct OutputType
 static int output_typeid_vector[td_ns::num_types];
 
 template <typename fnT, typename T>
-struct TypeMapFactory
+struct ErfTypeMapFactory
 {
     std::enable_if_t<std::is_same<fnT, int>::value, int> get()
     {
@@ -223,7 +223,7 @@ MACRO_DEFINE_IMPL(erfcinv, Erfcinv);
 void init_erf_funcs(py::module_ m)
 {
     using impl::output_typeid_vector;
-    init_dispatch_vector<int, impl::TypeMapFactory>(output_typeid_vector);
+    init_dispatch_vector<int, impl::ErfTypeMapFactory>(output_typeid_vector);
 
     auto erf_result_type_pyapi = [&](const py::dtype &dtype) {
         return py_int::py_unary_ufunc_result_type(dtype, output_typeid_vector);

dpnp/backend/extensions/ufunc/elementwise_functions/float_power.cpp

@@ -85,7 +85,7 @@ struct OutputType
 static int float_power_output_typeid_table[td_ns::num_types][td_ns::num_types];
 
 template <typename fnT, typename T1, typename T2>
-struct TypeMapFactory
+struct FloatPowerTypeMapFactory
 {
     std::enable_if_t<std::is_same<fnT, int>::value, int> get()
     {
@@ -96,7 +96,8 @@ struct TypeMapFactory
 
 static void populate_float_power_dispatch_tables(void)
 {
-    init_dispatch_table<int, TypeMapFactory>(float_power_output_typeid_table);
+    init_dispatch_table<int, FloatPowerTypeMapFactory>(
+        float_power_output_typeid_table);
 }
 } // namespace impl