[test]: Marker-based suites + ctest + ccc-gpu-bench CLI (restructure-tests + extract-benchmarks)

.github/workflows/ci.yml

@@ -100,12 +100,12 @@ jobs:
           PY
 
   # ---------------------------------------------------------------------- #
-  # (c) CPU-only tests: the CPU subset of the suite. GPU, network, and
-  #     undeclared-dep tests are ignored (see BASELINE). The two known
-  #     np.unicode_ test-only failures are deselected until restructure-tests
-  #     (PR D) fixes the NumPy-2 alias.
-  #     TODO(PR D): once pytest markers land, replace the --ignore/--deselect
-  #     list with `-m "not gpu and not slow and not network"`.
+  # (c) CPU-only tests: the CPU subset of the suite, selected by pytest
+  #     markers (`-m "not gpu and not slow and not network"`). GPU tests are
+  #     additionally skipped at collection when cupy/ccc_cuda_ext are absent
+  #     (tests/gpu/conftest.py). The research-only modules import undeclared
+  #     heavy deps (requests/minepy/IPython) and are excluded from the wheel,
+  #     so they are still `--ignore`d here (owned by improve-docs / PR E).
   # ---------------------------------------------------------------------- #
   test-cpu:
     name: CPU tests
@@ -141,11 +141,9 @@ jobs:
           PATH: /opt/conda/bin:/usr/local/cuda/bin:/usr/bin:/bin
         run: |
           python -m pytest tests/ \
-            --ignore=tests/gpu \
             --ignore=tests/test_giant.py \
             --ignore=tests/test_methods.py \
             --ignore=tests/test_plots.py \
             --ignore=tests/test_corr.py \
-            --deselect "tests/test_coef_pval.py::test_cm_numerical_and_categorical_features_perfect_relationship_pvalue" \
-            --deselect "tests/test_coef_pval.py::test_cm_numerical_and_categorical_features_weakly_relationship_pvalue" \
+            -m "not gpu and not slow and not network" \
             -q -ra -p no:cacheprovider -o addopts=""

CMakeLists.txt

@@ -1,6 +1,11 @@
 cmake_minimum_required(VERSION 3.15...3.26)
 project(${SKBUILD_PROJECT_NAME} LANGUAGES CUDA CXX)
 
+# When ON, build the CUDA C++ gtests under tests/cuda_ext/ and register them with
+# ctest. OFF by default so the wheel build never fetches googletest or compiles
+# the native tests. Enable with: cmake -S . -B build -DCCC_BUILD_TESTS=ON
+option(CCC_BUILD_TESTS "Build the CUDA C++ gtests and register them with ctest" OFF)
+
 # Export compile_commands.json so clang-tidy (and editors/LSP) can analyze the
 # CUDA/C++ sources. Run clang-tidy host-only, e.g.:
 #   clang-tidy -p build --extra-arg=--cuda-host-only libs/ccc_cuda_ext/*.cu
@@ -42,14 +47,10 @@ find_package(Python REQUIRED COMPONENTS Interpreter Development.Module)
 find_package(pybind11 CONFIG REQUIRED)
 set(PYBIND11_NEWPYTHON ON)
 
-# Download and configure Google Test
+# Google Test is fetched only when building the native tests (CCC_BUILD_TESTS);
+# the default wheel build must not touch the network for gtest. See the guarded
+# test block near the end of this file.
 include(FetchContent)
-FetchContent_Declare(
-    googletest
-    GIT_REPOSITORY https://github.com/google/googletest.git
-    GIT_TAG v1.15.2  # Adjust version as needed
-)
-FetchContent_MakeAvailable(googletest)
 # Download and configure spdlog
 FetchContent_Declare(
     spdlog
@@ -62,54 +63,6 @@ FetchContent_MakeAvailable(spdlog)
 
 # include_directories("${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES}")
 
-# Setup Gtest
-enable_testing()
-# Function to automatically add tests from a directory
-function(add_tests_from_directory TEST_DIR)
-    # Find all test files in the directory
-    file(GLOB_RECURSE TEST_FILES
-        "${TEST_DIR}/*_test.cu"    # Files ending with _test.cu
-        "${TEST_DIR}/test_*.cu"    # Files starting with test_
-    )
-
-    # Loop through each test file
-    foreach(TEST_FILE ${TEST_FILES})
-        # Get the filename without extension
-        get_filename_component(TEST_NAME ${TEST_FILE} NAME_WE)
-
-        # Create an executable for this test
-        add_executable(${TEST_NAME} ${TEST_FILE} ${headers} ${sources})
-
-        # target_include_directories(${TEST_NAME} PRIVATE
-        #     ${PROJECT_INCLUDE_DIR}  # Add this line
-        #     ${Python_INCLUDE_DIRS}
-        # )
-
-        target_link_libraries(${TEST_NAME} PRIVATE
-            GTest::gtest_main
-            GTest::gtest
-            pybind11::headers
-            pybind11::embed
-            Python::Python
-            # Add your other project libraries here
-            # project_lib
-        )
-
-        # Add the test to CTest
-        add_test(NAME ${TEST_NAME} COMMAND ${TEST_NAME})
-
-        # Set test properties (optional)
-        # Set test properties (optional)
-        set_tests_properties(${TEST_NAME} PROPERTIES
-            TIMEOUT 10  # Timeout in seconds
-            WORKING_DIRECTORY "${CMAKE_RUNTIME_OUTPUT_DIRECTORY}"
-        )
-    endforeach()
-endfunction()
-
-# Specify your test directory and call the function
-# add_tests_from_directory(${CMAKE_CURRENT_SOURCE_DIR}/tests)
-
 # Optional: Set output directories
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
 set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib)
@@ -127,3 +80,68 @@ target_link_libraries(${CUDA_EXT_MODULE_NAME} PRIVATE
 )
 
 install(TARGETS ${CUDA_EXT_MODULE_NAME} LIBRARY DESTINATION .)
+
+# --------------------------------------------------------------------------- #
+# CUDA C++ gtests (opt-in). Guarded by CCC_BUILD_TESTS so the wheel build never
+# fetches googletest or compiles the native tests.
+#   cmake -S . -B build -DCCC_BUILD_TESTS=ON && cmake --build build && ctest --test-dir build
+# --------------------------------------------------------------------------- #
+if(CCC_BUILD_TESTS)
+    message(STATUS "CCC_BUILD_TESTS=ON: configuring CUDA C++ gtests")
+
+    # Embedding a Python interpreter (test_ari_random.cu) needs Development.Embed,
+    # which the default wheel build (Development.Module only) does not require.
+    find_package(Python REQUIRED COMPONENTS Interpreter Development.Embed Development.Module)
+
+    FetchContent_Declare(
+        googletest
+        GIT_REPOSITORY https://github.com/google/googletest.git
+        GIT_TAG v1.15.2
+    )
+    FetchContent_MakeAvailable(googletest)
+    include(GoogleTest)
+    enable_testing()
+
+    # The gtests exercise the ARI kernels in metrics.cu. Compile that translation
+    # unit once into a static library and link it into each test executable so we
+    # do not recompile the (slow) .cu file per test. coef.cu / binder.cu are not
+    # needed by the current tests and are intentionally excluded.
+    add_library(ccc_cuda_ext_testlib STATIC ${CUDA_EXT_DIR}/metrics.cu)
+    target_link_libraries(ccc_cuda_ext_testlib PUBLIC
+        spdlog::spdlog
+        pybind11::headers
+        Python::Python
+    )
+    set_target_properties(ccc_cuda_ext_testlib PROPERTIES
+        POSITION_INDEPENDENT_CODE ON
+    )
+
+    # Discover the gtest source files. test_ari_py.cu is a legacy manual driver
+    # with its own main() (not a gtest); it would clash with gtest_main, so skip it.
+    file(GLOB CCC_TEST_FILES "${CMAKE_CURRENT_SOURCE_DIR}/tests/cuda_ext/test_*.cu")
+    foreach(TEST_FILE ${CCC_TEST_FILES})
+        get_filename_component(TEST_NAME ${TEST_FILE} NAME_WE)
+        if(TEST_NAME STREQUAL "test_ari_py")
+            continue()
+        endif()
+
+        add_executable(${TEST_NAME} ${TEST_FILE})
+        target_link_libraries(${TEST_NAME} PRIVATE
+            ccc_cuda_ext_testlib
+            GTest::gtest_main
+            GTest::gtest
+            pybind11::headers
+            pybind11::embed
+            Python::Python
+        )
+
+        # gtest_discover_tests registers each TEST_P case with ctest. The default
+        # 10s ctest timeout was too tight for the Python-reference tests; raise it.
+        gtest_discover_tests(${TEST_NAME}
+            PROPERTIES
+                TIMEOUT 300
+                WORKING_DIRECTORY "${CMAKE_RUNTIME_OUTPUT_DIRECTORY}"
+            DISCOVERY_TIMEOUT 120
+        )
+    endforeach()
+endif()

libs/ccc/bench/__init__.py

@@ -0,0 +1,8 @@
+"""Standalone benchmark CLI for CCC (``ccc-gpu-bench`` / ``python -m ccc.bench``).
+
+Reproducible, structured GPU-vs-CPU performance measurement decoupled from the
+test suite: end-to-end coefficient benchmarks, ARI-kernel micro-benchmarks and
+CPU parallelism scaling, emitted as JSON Lines / CSV with full environment
+metadata. Runs degrade gracefully on machines without a GPU (CPU-side modes work;
+GPU modes fail fast with a clear message). See ``ccc-gpu-bench --help``.
+"""

libs/ccc/bench/__main__.py

@@ -0,0 +1,8 @@
+"""Enable ``python -m ccc.bench``."""
+
+import sys
+
+from .cli import main
+
+if __name__ == "__main__":
+    sys.exit(main())