SIMD backends for math functions

.github/workflows/simd-backends.yml

@@ -0,0 +1,226 @@
+# SPDX-FileCopyrightText: 2026 CExA-project
+# SPDX-License-Identifier: MIT or Apache-2.0 with LLVM-exception
+
+name: SIMD Backends Tests
+
+on:
+  pull_request:
+    paths:
+      - '.github/workflows/simd-backends.yml'
+      - 'simd-backends/**'
+      - '!simd-backends/README.md'
+  push:
+    branches:
+      - main
+
+concurrency:
+  group: ${{ github.event_name }}-${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
+
+jobs:
+  unit-tests:
+    continue-on-error: true
+
+    strategy:
+      matrix:
+        kokkos-version: [4.6.00, 5.1.0]
+
+    runs-on: ubuntu-latest
+
+    env:
+      CMAKE_GENERATOR: Ninja
+      Kokkos_ROOT: ${{ github.workspace }}/opt/kokkos
+      sleef_ROOT: ${{ github.workspace }}/opt/sleef
+      CexaSimdBackends_ROOT: ${{ github.workspace }}/simd-backends/install
+    steps:
+      - name: Checkout Code
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+      - uses: actions/cache@cdf6c1fa76f9f475f3d7449005a359c84ca0f306 # v5.0.3
+        id: cache-sleef
+        with:
+          path: opt/sleef
+          key: simd-backends-sleef-ubuntu-latest
+      - name: Checkout Sleef
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+        if: steps.cache-sleef.outputs.cache-hit != 'true'
+        with:
+          repository: shibatch/sleef
+          ref: 3.9.0
+          path: sleef
+      - name: Install Sleef
+        if: steps.cache-sleef.outputs.cache-hit != 'true'
+        run: |
+          cmake \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DCMAKE_C_COMPILER=gcc \
+            -DCMAKE_CXX_COMPILER=g++ \
+            -DCMAKE_INSTALL_PREFIX="$sleef_ROOT" \
+            -DSLEEF_BUILD_GNUABI_LIBS=OFF \
+            -DSLEEF_BUILD_TESTS=OFF \
+            -DSLEEF_ENABLE_TLFLOAT=OFF \
+            -DSLEEF_ENABLE_FFTW=OFF \
+            -DSLEEF_ENABLE_MPFR=OFF \
+            -B sleef/build -S sleef
+          cmake --build sleef/build -j $(nproc) --target install
+      - name: Checkout Kokkos
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+        with:
+          repository: kokkos/kokkos
+          ref: ${{ matrix.kokkos-version }}
+          path: kokkos
+      - name: Install kokkos
+        run: |
+          cmake \
+            -DCMAKE_CXX_COMPILER=g++ \
+            -DCMAKE_BUILD_TYPE=RelWithDebInfo \
+            -DCMAKE_INSTALL_PREFIX="$Kokkos_ROOT" \
+            -DKokkos_ARCH_NATIVE=ON \
+            -B kokkos/build -S kokkos
+          cmake --build kokkos/build -j $(nproc) --target install
+      - name: Configure
+        working-directory: ${{ github.workspace }}/simd-backends
+        run:
+          cmake
+            -DCMAKE_CXX_COMPILER=g++
+            -DCMAKE_BUILD_TYPE=RelWithDebInfo
+            -DCMAKE_INSTALL_PREFIX="$CexaSimdBackends_ROOT"
+            -DCEXA_SIMD_ENABLE_SLEEF=ON
+            -DCEXA_SIMD_ENABLE_TESTS=ON
+            -B build -S .
+      - name: Build
+        working-directory: ${{ github.workspace }}/simd-backends
+        run: cmake --build build -j $(nproc)
+      - name: Install
+        working-directory: ${{ github.workspace }}/simd-backends
+        run: cmake --install build
+      - name: Test install
+        working-directory: ${{ github.workspace }}/simd-backends/test/install_test
+        run: |
+          cmake \
+            -DCMAKE_CXX_COMPILER=g++ \
+            -DCMAKE_BUILD_TYPE=RelWithDebInfo \
+            -B build -S .
+          cmake --build build -j $(nproc)
+          ./build/main
+      - name: Patch the Kokkos SIMD unit tests to use the Sleef backend
+        working-directory: ${{ github.workspace }}/kokkos/simd/unit_tests
+        run: |
+          cp "$CexaSimdBackends_ROOT/include/CEXA_SIMD_SLEEF.hpp" include
+          sed -i '/kokkos_add_executable_and_test/a find_package(sleef REQUIRED)\ntarget_link_libraries(Kokkos_UnitTest_SIMD PRIVATE sleef::sleef)' CMakeLists.txt
+          sed -i '/include <Kokkos_SIMD.hpp>/a #include <CEXA_SIMD_SLEEF.hpp>' include/TestSIMD_MathOps.hpp
+      - name: Run the Kokkos SIMD unit tests using the Sleef backend
+        working-directory: ${{ github.workspace }}/kokkos
+        run: |
+          cmake \
+            -DCMAKE_CXX_COMPILER=g++ \
+            -DCMAKE_BUILD_TYPE=RelWithDebInfo \
+            -DKokkos_ARCH_NATIVE=ON \
+            -DKokkos_ENABLE_TESTS=ON \
+            -B build -S .
+          cmake --build build -j $(nproc) --target Kokkos_UnitTest_SIMD
+          ctest --test-dir build --output-on-failure -R SIMD
+
+  performance-tests:
+    continue-on-error: true
+
+    runs-on: ubuntu-latest
+
+    env:
+      CMAKE_GENERATOR: Ninja
+      sleef_ROOT: ${{ github.workspace }}/opt/sleef
+      compare_ROOT: ${{ github.workspace }}/opt/compare
+      benchmark_VERSION: v1.9.5
+    steps:
+      - name: Checkout Code
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+      - uses: actions/cache@cdf6c1fa76f9f475f3d7449005a359c84ca0f306 # v5.0.3
+        id: cache-sleef
+        with:
+          path: opt/sleef
+          key: simd-backends-sleef-ubuntu-latest
+      - name: Checkout Sleef
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+        if: steps.cache-sleef.outputs.cache-hit != 'true'
+        with:
+          repository: shibatch/sleef
+          ref: 3.9.0
+          path: sleef
+      - name: Install Sleef
+        if: steps.cache-sleef.outputs.cache-hit != 'true'
+        run: |
+          cmake \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DCMAKE_C_COMPILER=gcc \
+            -DCMAKE_CXX_COMPILER=g++ \
+            -DCMAKE_INSTALL_PREFIX="$sleef_ROOT" \
+            -DSLEEF_BUILD_GNUABI_LIBS=OFF \
+            -DSLEEF_BUILD_TESTS=OFF \
+            -DSLEEF_ENABLE_TLFLOAT=OFF \
+            -DSLEEF_ENABLE_FFTW=OFF \
+            -DSLEEF_ENABLE_MPFR=OFF \
+            -B sleef/build -S sleef
+          cmake --build sleef/build -j $(nproc) --target install
+      - run: echo "PYTHON_VERSION=$(python3 --version | sed 's/ /_/')" >> $GITHUB_ENV
+      - uses: actions/cache@cdf6c1fa76f9f475f3d7449005a359c84ca0f306 # v5.0.3
+        id: cache-gbench-compare
+        with:
+          path: ${{ env.compare_ROOT }}
+          key: simd-backends-gbench-ubuntu-latest-${{ env.benchmark_VERSION }}-${{ env.PYTHON_VERSION }}
+      - name: Checkout google benchmark
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+        if: steps.cache-gbench-compare.outputs.cache-hit != 'true'
+        with:
+          repository: google/benchmark
+          ref: ${{ env.benchmark_VERSION }}
+          path: benchmark
+      - name: Get the compare script from google benchmark
+        if: steps.cache-gbench-compare.outputs.cache-hit != 'true'
+        run: |
+          mkdir -p ${{ env.compare_ROOT }}
+          cp -r benchmark/tools/* ${{ env.compare_ROOT }}
+      - name: Install python dependencies
+        working-directory: ${{ env.compare_ROOT }}
+        if: steps.cache-gbench-compare.outputs.cache-hit != 'true'
+        run: |
+          python3 -m venv ./venv
+          source ./venv/bin/activate
+          pip3 install -r ./requirements.txt
+      - name: Checkout Kokkos
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+        with:
+          repository: kokkos/kokkos
+          ref: 5.1.0
+          path: kokkos
+      - name: Run the Kokkos SIMD perf tests
+        # We use "math .[^c]" as benchmark_filter to only run benchmarks for
+        # math functions and to exclude the scalar ones, ".[^c]" will exclude
+        # "scalar" but not "avx", "neon" or "sve".
+        # We also increase the number of iterations to get more stable results,
+        # we can afford it since we don't run the scalar benchmarks.
+        working-directory: ${{ github.workspace }}/kokkos
+        run: |
+          sed -i 's/Iterations(3)/Iterations(300)/' simd/perf_tests/include/PerfTest_Host.hpp
+          cmake \
+            -DCMAKE_CXX_COMPILER=g++ \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DKokkos_ARCH_NATIVE=ON \
+            -DKokkos_ENABLE_BENCHMARKS=ON \
+            -B build -S .
+          cmake --build build -j $(nproc) --target Kokkos_PerformanceTest_SIMD_Host
+          ./build/simd/perf_tests/Kokkos_PerformanceTest_SIMD_Host --benchmark_out=baseline.json --benchmark_filter="math .[^c]"
+      - name: Patch the Kokkos SIMD perf tests to use the Sleef backend
+        working-directory: ${{ github.workspace }}/kokkos/simd/perf_tests
+        run: |
+          cp "${{ github.workspace }}/simd-backends/src/CEXA_SIMD_SLEEF.hpp" include
+          sed -i '/KOKKOS_IMPL_SIMD_DEVICE_PERFTEST/a find_package(sleef REQUIRED)\ntarget_link_libraries(Kokkos_PerformanceTest_SIMD_Host PRIVATE sleef::sleef)' CMakeLists.txt
+          sed -i '/include <Kokkos_SIMD.hpp>/a #include <CEXA_SIMD_SLEEF.hpp>' include/PerfTest_Host.hpp
+      - name: Run the Kokkos SIMD perf tests using the Sleef backend
+        working-directory: ${{ github.workspace }}/kokkos
+        run: |
+          cmake --build build -j $(nproc) --target Kokkos_PerformanceTest_SIMD_Host
+          ./build/simd/perf_tests/Kokkos_PerformanceTest_SIMD_Host --benchmark_out=sleef.json --benchmark_filter="math .[^c]"
+      - name: Compare the perf test results
+        working-directory: ${{ github.workspace }}/kokkos
+        run: |
+          source ${{ env.compare_ROOT }}/venv/bin/activate
+          ${{ env.compare_ROOT }}/compare.py benchmarks baseline.json sleef.json

simd-backends/CMakeLists.txt

@@ -0,0 +1,90 @@
+# SPDX-FileCopyrightText: 2026 CExA-project
+# SPDX-License-Identifier: MIT or Apache-2.0 with LLVM-exception
+
+cmake_minimum_required(VERSION 3.23)
+
+project(simd-backends VERSION 0.1.0)
+
+find_package(Kokkos 4.6 REQUIRED)
+
+option(
+  CEXA_SIMD_ENABLE_SLEEF
+  "Use the sleef library to implement the simd math functions"
+  OFF
+)
+option(CEXA_SIMD_ENABLE_SVML "Use the svml to implement the simd math functions" OFF)
+option(CEXA_SIMD_ENABLE_TESTS "Enable accuracy tests" OFF)
+
+if(CEXA_SIMD_ENABLE_SLEEF AND CEXA_SIMD_ENABLE_SVML)
+  message(FATAL_ERROR "Only one backend can be enabled at a time")
+endif()
+
+if(CEXA_SIMD_ENABLE_SLEEF)
+  find_package(sleef 3.6 QUIET)
+  if(NOT sleef_FOUND)
+    include(FetchContent)
+
+    set(SLEEF_BUILD_GNUABI_LIBS OFF)
+    set(SLEEF_BUILD_TESTS OFF)
+    set(SLEEF_ENABLE_TLFLOAT OFF)
+    set(SLEEF_ENABLE_FFTW OFF)
+    set(SLEEF_ENABLE_MPFR OFF)
+    list(APPEND CMAKE_MESSAGE_INDENT "[sleef] ")
+    FetchContent_Declare(
+      sleef
+      URL https://github.com/shibatch/sleef/archive/refs/tags/3.9.0.tar.gz
+      DOWNLOAD_EXTRACT_TIMESTAMP TRUE
+      DOWNLOAD_NO_PROGRESS TRUE
+    )
+    FetchContent_MakeAvailable(sleef)
+    list(POP_BACK CMAKE_MESSAGE_INDENT)
+  endif()
+endif()
+
+if(CEXA_SIMD_ENABLE_SVML AND NOT ${CMAKE_CXX_COMPILER_ID} STREQUAL IntelLLVM)
+  message(
+    FATAL_ERROR
+    "The SVML simd backend is only available with Intel compilers"
+  )
+endif()
+
+add_subdirectory(src)
+
+if(CEXA_SIMD_ENABLE_TESTS)
+  add_subdirectory(test)
+endif()
+
+# Install
+include(GNUInstallDirs)
+
+include(CMakePackageConfigHelpers)
+write_basic_package_version_file(
+  CexaSimdBackendsConfigVersion.cmake
+  COMPATIBILITY AnyNewerVersion
+)
+
+install(
+  TARGETS simd-backends
+  EXPORT CexaSimdBackends
+  FILE_SET HEADERS
+)
+
+install(
+  EXPORT CexaSimdBackends
+  FILE CexaSimdBackendsTargets.cmake
+  NAMESPACE cexa::
+  DESTINATION "${CMAKE_INSTALL_LIBDIR}/cmake/CexaSimdBackends"
+)
+
+configure_package_config_file(
+  cmake/CexaSimdBackendsConfig.cmake.in
+  "${CMAKE_CURRENT_BINARY_DIR}/CexaSimdBackendsConfig.cmake"
+  INSTALL_DESTINATION "${CMAKE_INSTALL_LIBDIR}/cmake/CexaSimdBackends"
+)
+
+install(
+  FILES
+    "${CMAKE_CURRENT_BINARY_DIR}/CexaSimdBackendsConfig.cmake"
+    "${CMAKE_CURRENT_BINARY_DIR}/CexaSimdBackendsConfigVersion.cmake"
+  DESTINATION "${CMAKE_INSTALL_LIBDIR}/cmake/CexaSimdBackends"
+)

simd-backends/README.md

@@ -0,0 +1,56 @@
+<!--
+SPDX-FileCopyrightText: 2026 CExA-project
+
+SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
+-->
+
+# SIMD backends
+
+This header only library provides an interface to replace the Kokkos math
+functions for SIMD types with calls to different external libraries.
+
+## Build
+
+The available cmake configuration options are:
+- `CEXA_SIMD_ENABLE_SLEEF`: use the sleef library
+- `CEXA_SIMD_ENABLE_SVML`: use the Intel SVML library (only available with intel compilers)
+- `CEXA_SIMD_ENABLE_TESTS`: build the tests
+
+The sleef backend requires sleef v3.6.0 or later.
+
+## Usage
+
+With CMake, you can use the `find_package` command as shown in the
+example below. If not installed to a system location, you should provide the
+path to the library using the `-DCexaSimdBackends_ROOT=<path to the install location>`
+option.
+
+```cmake
+cmake_minimum_required(VERSION 3.23)
+
+project(test)
+
+find_package(Kokkos REQUIRED)
+find_package(CexaSimdBackends REQUIRED)
+
+add_executable(main main.cpp)
+target_link_libraries(main PRIVATE Kokkos::kokkos cexa::simd-backends)
+```
+
+You can then include `CEXA_SIMD_Backends.hpp` in your code and use the Kokkos simd math
+functions as usual.
+
+```cpp
+#include <Kokkos_Core.hpp>
+#include <Kokkos_SIMD.hpp>
+#include <CEXA_SIMD_Backends.hpp>
+
+int main(int argc, char* argv[]) {
+  Kokkos::ScopeGuard kokkos_scope(argc, argv);
+
+  Kokkos::Experimental::simd<float> vec(1.f);
+  vec = Kokkos::exp(vec);
+
+  return 0;
+}
+```

simd-backends/cmake/CexaSimdBackendsConfig.cmake.in

@@ -0,0 +1,17 @@
+# SPDX-FileCopyrightText: 2026 CExA-project
+# SPDX-License-Identifier: MIT or Apache-2.0 with LLVM-exception
+
+@PACKAGE_INIT@
+
+include(CMakeFindDependencyMacro)
+
+find_dependency(Kokkos)
+
+set(CEXA_SIMD_ENABLE_SLEEF @CEXA_SIMD_ENABLE_SLEEF@)
+if(CEXA_SIMD_ENABLE_SLEEF)
+  find_dependency(sleef)
+endif()
+
+include("${CMAKE_CURRENT_LIST_DIR}/CexaSimdBackendsTargets.cmake")
+
+check_required_components(CexaSimdBackends)