stable ABI support

.github/workflows/wheels.yml

@@ -15,7 +15,7 @@ jobs:
       fail-fast: false
       matrix:
         os: [ubuntu-latest, macos-13, macos-14, windows-latest]
-        python: [cp38, cp39, cp310, cp311, cp312, cp313]
+        python: [cp38, cp39, cp310, cp311, cp312, cp312_stable, cp313]
         exclude:
           # The first Python version to target Apple arm64 architectures is 3.9
           - os: macos-14
@@ -56,14 +56,14 @@ jobs:
       run: |
         stable_cp=$(echo ${{ matrix.python }} | cut -d_ -f1) &&
         echo "CIBW_BUILD=${stable_cp}-*" >> $GITHUB_ENV &&
-        echo "CIBW_CONFIG_SETTINGS=\"wheel.py-api=cp312\" \"cmake.args=-DDRJIT_STABLE_ABI=ON\"" >> $GITHUB_ENV
+        echo "CIBW_CONFIG_SETTINGS=\"wheel.py-api=cp312\" \"cmake.args=-DFASTSWEEP_STABLE_ABI=ON\"" >> $GITHUB_ENV
 
     - name: Prepare cibuildwheel environment for stable ABI wheel (Windows)
       if: ${{ endsWith(matrix.python, '_stable') && runner.os == 'Windows' }}
       run: |
         $stable_cp = '${{ matrix.python }}' -split '_'
         echo "CIBW_BUILD=$($stable_cp[0])-*" >> $env:GITHUB_ENV
-        echo "CIBW_CONFIG_SETTINGS=wheel.py-api=cp312 cmake.args=-DDRJIT_STABLE_ABI=ON" >> $env:GITHUB_ENV
+        echo "CIBW_CONFIG_SETTINGS=wheel.py-api=cp312 cmake.args=-DFASTSWEEP_STABLE_ABI=ON" >> $env:GITHUB_ENV
 
     #########################
     # Build and store wheels

CMakeLists.txt

@@ -2,8 +2,11 @@ cmake_minimum_required(VERSION 3.15...3.19)
 
 project(fastsweep DESCRIPTION "Fast sweep solver")
 
-option(FASTWEEPING_USE_CUDA "Use high-performance CUDA kernels instead of
-                             Dr.Jit's GPU mode" ON)
+option(FASTSWEEP_USE_CUDA "Use high-performance CUDA kernels instead of
+                            Dr.Jit's GPU mode" ON)
+option(FASTSWEEP_STABLE_ABI "Build Python extension using the CPython stable ABI?
+                             (Only relevant when using scikit-build)" OFF)
+mark_as_advanced(FASTSWEEP_STABLE_ABI)
 
 if (NOT SKBUILD)
   message(WARNING "\
@@ -49,11 +52,27 @@ endif()
 list(APPEND CMAKE_PREFIX_PATH "${FASTSWEEP_DRJIT_CMAKE_DIR}")
 find_package(drjit CONFIG REQUIRED)
 
-if (FASTWEEPING_USE_CUDA)
+if (FASTSWEEP_USE_CUDA)
   add_definitions(-DFASTSWEEPING_USE_CUDA)
 endif()
 
-nanobind_add_module(fastsweep_ext NB_DOMAIN drjit NB_STATIC src/main.cpp src/distance_marcher.cpp)
+if (SKBUILD)
+  # Enable LTO only for release builds targeting PyPI (~5% binary size reduction)
+  set(FASTSWEEP_DIST_FLAGS LTO)
+  if (FASTSWEEP_STABLE_ABI)
+    list(APPEND FASTSWEEP_DIST_FLAGS STABLE_ABI)
+  endif()
+endif()
+
+nanobind_add_module(
+  fastsweep_ext
+  NB_DOMAIN
+  drjit
+  NB_STATIC
+  ${FASTSWEEP_DIST_FLAGS}
+  src/main.cpp
+  src/distance_marcher.cpp
+)
 
 target_link_libraries(fastsweep_ext PUBLIC drjit-core)
 target_compile_features(fastsweep_ext PRIVATE cxx_std_17)