Switching Github CI workflow to use nanobind. This should fix the err…

.github/workflows/ci.yml

@@ -45,18 +45,18 @@ jobs:
             cmake \
             ninja-build \
             libopenblas-dev \
-            pybind11-dev \
             python3-dev \
             ${{ matrix.compiler == 'clang' && 'clang libomp-dev' || '' }}
 
       - name: Install Python build dependencies
-        # pybind11's CMake config package is needed even though libpybind11-dev
-        # is installed above, in case find_package(pybind11) resolves via pip's
-        # copy instead - installing both covers either path.
-        run: pip install pybind11
+        run: python -m pip install nanobind rich
 
-      - name: Build and test (Release)
-        run: python build.py Release
+      - name: Build (Release)
+        # --no-tests: CMakeLists.txt no longer registers any CTest tests
+        # (the old test/benchmark executables were removed), so running
+        # ctest here would just fail on "no tests found" for no useful
+        # reason.
+        run: python build.py Release --no-tests
 
       - name: Upload build log
         if: always()
@@ -105,7 +105,7 @@ jobs:
             !C:/vcpkg/buildtrees
             !C:/vcpkg/packages
             !C:/vcpkg/downloads
-          key: vcpkg-windows-openblas-pybind11-v1
+          key: vcpkg-windows-openblas-v1
 
       - name: Bootstrap vcpkg
         if: steps.vcpkg-cache.outputs.cache-hit != 'true'
@@ -114,14 +114,10 @@ jobs:
           C:/vcpkg/bootstrap-vcpkg.bat
 
       - name: Install vcpkg dependencies
-        # Mirrors CMakeLists.txt: only OpenBLAS and pybind11 come from
-        # vcpkg. SLEEF and Google Benchmark are fetched from source by CMake
-        # itself (see FetchContent_Declare in CMakeLists.txt) and need no
-        # package manager step at all.
-        run: C:/vcpkg/vcpkg.exe install openblas pybind11 --triplet x64-windows
+        run: C:/vcpkg/vcpkg.exe install openblas --triplet x64-windows
 
       - name: Install Python build dependencies
-        run: pip install pybind11
+        run: python -m pip install nanobind rich
 
       - name: Configure (CMake, using the windows-clang preset)
         # Uses CMakePresets.json's "windows-clang" preset directly, rather
@@ -133,14 +129,6 @@ jobs:
       - name: Build
         run: cmake --build build/Release --config Release --parallel 2>&1 | Tee-Object -FilePath windows-build.log
 
-      - name: Run tests
-        run: |
-          $exe = Get-ChildItem -Path build/Release -Recurse -Filter "DeepityTests.exe" | Select-Object -First 1
-          if (-not $exe) {
-            throw "DeepityTests.exe not found under build/Release - check windows-build.log"
-          }
-          & $exe.FullName
-
       - name: Upload logs
         if: always()
         uses: actions/upload-artifact@v4

CMakeLists.txt

@@ -1,14 +1,12 @@
 cmake_minimum_required(VERSION 3.21)
 
 project(deepity C CXX)
-enable_testing()
 
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
 set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
 set(CMAKE_ARCHIVE_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
 
 option(DEEPITY_ENABLE_CUDA "Build with CUDA support" ON)
-option(DEEPITY_BUILD_TESTS "Build the DeepityTests target" OFF)
 option(DEEPITY_BUILD_PYTHON_BINDINGS "Build the Python bindings target" ON)
 # Recommended if on Intel CPUs
 option(DEEPITY_USE_MKL "Use Intel MKL instead of OpenBLAS (Intel CPUs only -- MKL has a documented history of deliberately worse performance on non-Intel CPUs via runtime dispatch)" OFF)
@@ -259,29 +257,6 @@ set(CMAKE_REQUIRED_QUIET ON)
 FetchContent_MakeAvailable(sleef)
 set(CMAKE_REQUIRED_QUIET OFF)
 
-# --- Google Benchmark -----------------------------------------------
-
-if(DEEPITY_BUILD_TESTS)
-    if(CMAKE_CXX_COMPILER_ID MATCHES "Clang")
-        add_compile_options(-Wno-c2y-extensions)
-    endif()
-
-    set(BENCHMARK_ENABLE_WERROR OFF CACHE BOOL "" FORCE)
-    set(BENCHMARK_ENABLE_PEDANTIC OFF CACHE BOOL "" FORCE)
-    set(BENCHMARK_ENABLE_TESTING OFF CACHE BOOL "" FORCE)
-    set(BENCHMARK_ENABLE_GTEST_TESTS OFF CACHE BOOL "" FORCE)
-
-    FetchContent_Declare(
-        benchmark
-        GIT_REPOSITORY https://github.com/google/benchmark.git
-        GIT_TAG v1.9.1
-    )
-
-    set(CMAKE_REQUIRED_QUIET ON)
-    FetchContent_MakeAvailable(benchmark)
-    set(CMAKE_REQUIRED_QUIET OFF)
-endif()
-
 # --- Deepity library ------------------------------------------------
 
 add_library(Deepity
@@ -351,28 +326,6 @@ set_target_properties(Deepity PROPERTIES
     INTERPROCEDURAL_OPTIMIZATION_RELEASE ${DEEPITY_IPO_SUPPORTED}
 )
 
-# --- Runtime DLL copying (Windows) -----------------------------------
-# $<TARGET_RUNTIME_DLLS:...> only knows about DLLs CMake can see a full
-# IMPORTED_LOCATION for (e.g. openblas_dll above). It can't discover
-# libomp.dll on its own, so DEEPITY_OMP_DLL (resolved earlier) is copied
-# alongside it explicitly. Centralized here instead of repeating both
-# copy steps for every executable below.
-function(deepity_copy_runtime_dlls target)
-    add_custom_command(TARGET ${target} POST_BUILD
-        COMMAND ${CMAKE_COMMAND} -E $<IF:$<BOOL:$<TARGET_RUNTIME_DLLS:${target}>>,copy_if_different,true>
-        $<TARGET_RUNTIME_DLLS:${target}>
-        $<TARGET_FILE_DIR:${target}>
-        COMMAND_EXPAND_LISTS
-    )
-    if(DEEPITY_OMP_DLL)
-        add_custom_command(TARGET ${target} POST_BUILD
-            COMMAND ${CMAKE_COMMAND} -E copy_if_different
-            "${DEEPITY_OMP_DLL}"
-            $<TARGET_FILE_DIR:${target}>
-        )
-    endif()
-endfunction()
-
 # --- pydeepity ------------------------------------------------------
 
 if(DEEPITY_BUILD_PYTHON_BINDINGS AND NOT CMAKE_BUILD_TYPE STREQUAL "Debug")
@@ -424,113 +377,3 @@ if(WIN32)
         install(FILES ${DEEPITY_OMP_DLL} DESTINATION pydeepity)
     endif()
 endif()
-
-# --- Tests ----------------------------------------------------------
-
-if(DEEPITY_BUILD_TESTS)
-    add_executable(DirectKPVerify tests/tDirectKPVerify.cpp)
-    target_link_libraries(DirectKPVerify PRIVATE Deepity)
-    set_target_properties(DirectKPVerify PROPERTIES RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
-    add_test(NAME DirectKPVerify COMMAND DirectKPVerify)
-
-    if(WIN32)
-        deepity_copy_runtime_dlls(DirectKPVerify)
-    endif()
-endif()
-
-# --- Profiling ------------------------------------------------------
-
-if(DEEPITY_BUILD_TESTS)
-    add_executable(GaussSeidelMiddleLayerVerify tests/tGaussSeidelMiddleLayerVerify.cpp)
-    target_link_libraries(GaussSeidelMiddleLayerVerify PRIVATE Deepity)
-    set_target_properties(GaussSeidelMiddleLayerVerify PROPERTIES RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
-    add_test(NAME GaussSeidelMiddleLayerVerify COMMAND GaussSeidelMiddleLayerVerify)
-
-    if(WIN32)
-        deepity_copy_runtime_dlls(GaussSeidelMiddleLayerVerify)
-    endif()
-endif()
-
-if(DEEPITY_BUILD_TESTS)
-    add_library(DeepityProfiled STATIC
-        src/DiscriminativePCLayer.cpp
-        src/RBLayer.cpp
-        src/ConvPCLayer.cpp
-        src/SimpleConvPCLayer.cpp
-        src/SimplePCLayer.cpp
-        src/DiscriminativePCNetwork.cpp
-        src/ConvPCNetwork.cpp
-        src/SimplePCNetwork.cpp
-        src/ModelIO.cpp
-        src/StreamAlignedBatcher.cpp
-    )
-
-    if(DEEPITY_USE_MKL AND DEEPITY_BLAS_RESOLVED)
-        target_compile_definitions(DeepityProfiled PUBLIC DEEPITY_USE_MKL)
-    endif()
-
-    target_include_directories(DeepityProfiled PUBLIC
-        ${CMAKE_CURRENT_SOURCE_DIR}/include
-        ${BLAS_INCLUDE_DIR}
-    )
-    target_include_directories(DeepityProfiled SYSTEM PUBLIC
-        ${sleef_SOURCE_DIR}/include
-        ${sleef_BINARY_DIR}/include
-    )
-    target_link_libraries(DeepityProfiled PUBLIC
-        ${BLAS_LIBRARIES}
-        OpenMP::OpenMP_CXX
-        sleef
-    )
-
-    target_compile_definitions(DeepityProfiled PUBLIC SLEEF_STATIC_LIBS PCN_PROFILE)
-
-    if(WIN32)
-        target_compile_definitions(DeepityProfiled PUBLIC NOMINMAX)
-    endif()
-
-    if(MSVC)
-        target_compile_options(DeepityProfiled PRIVATE /O2 /fp:fast /openmp:llvm)
-    elseif(CMAKE_CXX_COMPILER_ID MATCHES "Clang")
-        target_compile_options(DeepityProfiled PRIVATE -O3 -ffast-math -fvectorize -fslp-vectorize)
-    else()
-        target_compile_options(DeepityProfiled PRIVATE -O3 -ffast-math -ftree-vectorize -ftree-slp-vectorize)
-    endif()
-
-    add_executable(DeepityProfile tests/tProfile.cpp)
-    target_link_libraries(DeepityProfile PRIVATE DeepityProfiled)
-    target_compile_definitions(DeepityProfile PRIVATE PCN_PROFILE)
-
-    set_target_properties(DeepityProfile PROPERTIES
-        RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin
-    )
-    add_test(NAME DeepityProfile COMMAND DeepityProfile)
-
-    if(WIN32)
-        deepity_copy_runtime_dlls(DeepityProfile)
-    endif()
-
-    # --- README benchmark: std-library vs Deep:: activations ---------
-    add_executable(ActivationBenchmark tests/tActivations.cpp)
-    target_link_libraries(ActivationBenchmark PRIVATE DeepityProfiled benchmark)
-
-    set_target_properties(ActivationBenchmark PROPERTIES
-        RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin
-    )
-
-    if(WIN32)
-        deepity_copy_runtime_dlls(ActivationBenchmark)
-    endif()
-
-    # --- README benchmark: sustained GFLOPS during a real train step -
-    add_executable(GflopsBenchmark tests/tGflopsBenchmark.cpp)
-    target_link_libraries(GflopsBenchmark PRIVATE DeepityProfiled benchmark)
-
-    set_target_properties(GflopsBenchmark PROPERTIES
-        RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin
-    )
-
-    if(WIN32)
-        deepity_copy_runtime_dlls(GflopsBenchmark)
-    endif()
-endif()

deepity_build/cmake_runner.py

@@ -15,12 +15,7 @@ def find_generator() -> tuple[str | None, str]:
     ninja = shutil.which("ninja")
     return ninja, "Ninja" if ninja else "CMake"
 
-
 def configure_command(config: BuildConfig, ninja: str | None, pgo_phase: str | None = None) -> list[str]:
-    """pgo_phase: None (no PGO), "GENERATE", or "USE" -- distinguishes
-    which pass of the two-pass PGO workflow this configure call is for.
-    A single config.pgo boolean can't express this on its own, since
-    both passes share the same BuildConfig."""
     profile = config.profile
 
     cmd = [
@@ -35,47 +30,45 @@ def configure_command(config: BuildConfig, ninja: str | None, pgo_phase: str | N
         f"-DDEEPITY_ARCH_FLAGS={profile.unix_flags}",
     ]
 
+    if ninja:
+        cmd.extend(["-G", "Ninja"])
+
     if pgo_phase in ("GENERATE", "USE"):
         cmd.append(f"-DDEEPITY_PGO_MODE={pgo_phase}")
         cmd.append(f"-DDEEPITY_PGO_DATA_DIR={config.pgo_data_dir}")
 
     if profile.msvc_flags:
         cmd.append(f"-DDEEPITY_MSVC_ARCH_FLAGS={profile.msvc_flags}")
 
-    if sys.platform == "win32":
-        if ninja:
-            cmd.extend(["-G", "Ninja"])
-
-            # CC being unset doesn't mean clang isn't in play -- CMake can
-            # auto-detect and pick it up on its own (as this project's own
-            # builds have shown), so check what's actually on PATH rather
-            # than an environment variable that may never have been set.
-            clang_path = shutil.which("clang")
-            cc_env = os.environ.get("CC", "").lower()
-            using_clang = clang_path is not None or "clang" in cc_env
-
-            if using_clang:
-                # Locate libomp.lib next to the clang.exe on PATH
-                omp_lib = None
-                if clang_path:
-                    llvm_lib_dir = Path(clang_path).parent.parent / "lib"
-                    candidate = llvm_lib_dir / "libomp.lib"
-                    if candidate.is_file():
-                        omp_lib = candidate.as_posix()
-
-                cmd.extend([
-                    f"-DCMAKE_C_COMPILER=clang",
-                    f"-DCMAKE_CXX_COMPILER=clang++",
-                    "-DOpenMP_C_FLAGS=-fopenmp",
-                    "-DOpenMP_CXX_FLAGS=-fopenmp",
-                    "-DOpenMP_C_LIB_NAMES=omp",
-                    "-DOpenMP_CXX_LIB_NAMES=omp",
-                ])
-                if omp_lib:
-                    cmd.append(f"-DOpenMP_omp_LIBRARY={omp_lib}")
+    if sys.platform == "win32" and ninja:
+        # CC being unset doesn't mean clang isn't in play -- CMake can
+        # auto-detect and pick it up on its own (as this project's own
+        # builds have shown), so check what's actually on PATH rather
+        # than an environment variable that may never have been set.
+        clang_path = shutil.which("clang")
+        cc_env = os.environ.get("CC", "").lower()
+        using_clang = clang_path is not None or "clang" in cc_env
+
+        if using_clang:
+            omp_lib = None
+            if clang_path:
+                llvm_lib_dir = Path(clang_path).parent.parent / "lib"
+                candidate = llvm_lib_dir / "libomp.lib"
+                if candidate.is_file():
+                    omp_lib = candidate.as_posix()
+
+            cmd.extend([
+                f"-DCMAKE_C_COMPILER=clang",
+                f"-DCMAKE_CXX_COMPILER=clang++",
+                "-DOpenMP_C_FLAGS=-fopenmp",
+                "-DOpenMP_CXX_FLAGS=-fopenmp",
+                "-DOpenMP_C_LIB_NAMES=omp",
+                "-DOpenMP_CXX_LIB_NAMES=omp",
+            ])
+            if omp_lib:
+                cmd.append(f"-DOpenMP_omp_LIBRARY={omp_lib}")
     return cmd
 
-
 def build_command(config: BuildConfig) -> list[str]:
     return [
         "cmake",
@@ -98,4 +91,4 @@ def test_command(config: BuildConfig) -> list[str]:
         "--output-on-failure",
         "-j",
         str(config.jobs),
-    ]
+    ]