Add realistic CUDA API function names to test stack traces

- Add fake cuLaunchKernel, cuGraphLaunch, cudaLaunchKernel, cudaGraphLaunch
  functions that appear in the call stack when the profiler captures samples
- Use __attribute__((noinline)) to prevent compiler from inlining these
- Use random kernel names for graph kernels instead of hardcoded "graph_kernel"
- Add cuda_runtime.h include for runtime API types

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Author: gnurizen

test/CMakeLists.txt

@@ -0,0 +1,59 @@
+cmake_minimum_required(VERSION 3.18)
+project(parcagpu_test C)
+
+# CUDA include paths
+if(NOT DEFINED CUDA_ROOT)
+    set(CUDA_ROOT "/usr/local/cuda" CACHE PATH "CUDA installation directory")
+endif()
+
+set(CUDA_INCLUDE_DIRS
+    "${CUDA_ROOT}/include"
+    "${CUDA_ROOT}/extras/CUPTI/include"
+)
+
+# Mock CUPTI library
+add_library(cupti SHARED mock_cupti.c)
+target_include_directories(cupti PRIVATE ${CUDA_INCLUDE_DIRS})
+set_target_properties(cupti PROPERTIES
+    C_STANDARD 11
+    C_STANDARD_REQUIRED ON
+    POSITION_INDEPENDENT_CODE ON
+)
+
+# Detect CUDA major version from CUPTI library
+execute_process(
+    COMMAND bash -c "ls ${CUDA_ROOT}/lib64/libcupti.so.* 2>/dev/null | grep -oE 'libcupti\\.so\\.[0-9]+$' | head -1 | grep -oE '[0-9]+$'"
+    OUTPUT_VARIABLE CUDA_MAJOR_VERSION
+    OUTPUT_STRIP_TRAILING_WHITESPACE
+)
+if(NOT CUDA_MAJOR_VERSION)
+    set(CUDA_MAJOR_VERSION "12")
+endif()
+message(STATUS "Detected CUDA major version: ${CUDA_MAJOR_VERSION}")
+
+# Create versioned symlink matching installed CUDA
+add_custom_command(TARGET cupti POST_BUILD
+    COMMAND ${CMAKE_COMMAND} -E create_symlink
+    $<TARGET_FILE_NAME:cupti>
+    ${CMAKE_CURRENT_BINARY_DIR}/libcupti.so.${CUDA_MAJOR_VERSION}
+)
+
+# Test executable
+add_executable(test_cupti_prof test_cupti_prof.c)
+target_include_directories(test_cupti_prof PRIVATE ${CUDA_INCLUDE_DIRS})
+target_compile_options(test_cupti_prof PRIVATE
+    -D_POSIX_C_SOURCE=199309L
+    -Wall
+    -Wextra
+)
+target_link_libraries(test_cupti_prof PRIVATE dl pthread)
+set_target_properties(test_cupti_prof PROPERTIES
+    C_STANDARD 11
+    C_STANDARD_REQUIRED ON
+)
+
+# Installation
+install(TARGETS cupti test_cupti_prof
+    RUNTIME DESTINATION bin
+    LIBRARY DESTINATION lib
+)