CMakeLists.txt

cmake_minimum_required(VERSION 3.18)

project(matrix-ops-lib
    VERSION     0.1.0
    DESCRIPTION "GPU-accelerated matrix operations library with CUDA tensor core support"
    LANGUAGES   CXX CUDA
)

# ============================================================================
# C++ / CUDA standards
# ============================================================================
set(CMAKE_CXX_STANDARD  17)
set(CMAKE_CXX_STANDARD_REQUIRED ON)
set(CMAKE_CUDA_STANDARD 17)
set(CMAKE_CUDA_STANDARD_REQUIRED ON)

# ============================================================================
# Default build type
# ============================================================================
if(NOT CMAKE_BUILD_TYPE)
    set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)
endif()

# ============================================================================
# CUDA architecture selection
# Defaults to Ampere (SM 8.0) + Hopper (SM 9.0).  Override with:
#   cmake -DCMAKE_CUDA_ARCHITECTURES="70;80;89;90" ..
# ============================================================================
if(NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
    set(CMAKE_CUDA_ARCHITECTURES 80 90)
    message(STATUS "CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES} (default)")
endif()

# ============================================================================
# Find CUDA libraries
# ============================================================================
find_package(CUDAToolkit REQUIRED)
message(STATUS "CUDA version: ${CUDAToolkit_VERSION}")

# ============================================================================
# Custom CMake modules
# ============================================================================
list(APPEND CMAKE_MODULE_PATH "${CMAKE_SOURCE_DIR}/cmake")
include(TensorCoreDetection)

# ============================================================================
# Compiler flags
# ============================================================================

# Enable all CUDA warnings and optimise for the target architecture.
set(CUDA_COMMON_FLAGS
    --use_fast_math               # Enable FP32 fast math intrinsics
    --expt-relaxed-constexpr      # Allow constexpr in device code
    --expt-extended-lambda        # Allow lambdas in device code
    -Xcompiler=-Wall              # Host compiler warnings
    -Xcompiler=-Wextra
    -Xcompiler=-Wno-unused-parameter
)

if(CMAKE_BUILD_TYPE STREQUAL "Debug")
    list(APPEND CUDA_COMMON_FLAGS -g -G)   # Device-side debug info
else()
    list(APPEND CUDA_COMMON_FLAGS -O3 --ptxas-options=-v)
endif()

# ============================================================================
# Core library sources
# ============================================================================
set(MATLIB_SOURCES
    src/blas/gemm/gemm_naive.cu
    src/blas/gemm/gemm_tiled.cu
    src/blas/gemm/gemm_registers.cu
    src/blas/gemm/gemm_tensor_core.cu
    src/blas/gemm/gemm_bf16.cu
    src/blas/gemm/gemm_dispatcher.cu
    src/blas/gemm/fp_convert.cu
    src/blas/gemv.cu
    src/blas/trsm.cu
    src/blas/batched_gemm.cu
    src/sparse/spmv_csr.cu
    src/sparse/merge_path_spmv.cu
)

# Static library (used by tests and benchmarks).
add_library(matlib_static STATIC ${MATLIB_SOURCES})

target_include_directories(matlib_static PUBLIC
    "${CMAKE_SOURCE_DIR}/include"
    "${CMAKE_SOURCE_DIR}/kernels"
)

target_link_libraries(matlib_static PUBLIC
    CUDA::cudart
    CUDA::cublas
    CUDA::cusolver
    CUDA::cusparse
)

target_compile_options(matlib_static PRIVATE
    "$<$<COMPILE_LANGUAGE:CUDA>:${CUDA_COMMON_FLAGS}>"
)

# ============================================================================
# Unit tests
# ============================================================================
option(MATLIB_BUILD_TESTS "Build unit and integration tests" ON)

if(MATLIB_BUILD_TESTS)
    find_package(GTest QUIET)
    if(NOT GTest_FOUND)
        # Automatically download GoogleTest via FetchContent.
        include(FetchContent)
        FetchContent_Declare(
            googletest
            GIT_REPOSITORY https://github.com/google/googletest.git
            GIT_TAG        v1.14.0
        )
        FetchContent_MakeAvailable(googletest)
    endif()

    enable_testing()

    # Macro to add a test executable with common setup.
    macro(add_matlib_test name source)
        add_executable(${name} ${source})
        target_include_directories(${name} PRIVATE "${CMAKE_SOURCE_DIR}/include")
        target_link_libraries(${name} PRIVATE
            matlib_static
            GTest::gtest_main
            CUDA::cudart
            CUDA::cublas
            CUDA::cusolver
        )
        add_test(NAME ${name} COMMAND ${name})
    endmacro()

    add_matlib_test(test_gemm           tests/unit/test_gemm.cpp)
    add_matlib_test(test_trsm           tests/unit/test_trsm.cpp)
    add_matlib_test(test_tensor_core    tests/unit/test_tensor_core.cpp)
    add_matlib_test(test_tensor_gemm    tests/unit/test_tensor_core.cpp)
    add_matlib_test(test_bf16_gemm      tests/unit/test_bf16_gemm.cpp)
    add_matlib_test(test_accuracy       tests/numerical/test_accuracy.cpp)
    add_matlib_test(test_conditioning   tests/numerical/test_conditioning.cpp)
    add_matlib_test(test_cublas_compat  tests/integration/test_cublas_compat.cpp)
endif()

# ============================================================================
# Benchmarks
# ============================================================================
option(MATLIB_BUILD_BENCHMARKS "Build benchmark executables" ON)

if(MATLIB_BUILD_BENCHMARKS)
    macro(add_matlib_bench name source)
        add_executable(${name} ${source})
        target_include_directories(${name} PRIVATE "${CMAKE_SOURCE_DIR}/include")
        target_link_libraries(${name} PRIVATE
            matlib_static
            CUDA::cudart
            CUDA::cublas
        )
    endmacro()

    add_matlib_bench(gemm_bench          benchmarks/gemm_bench.cpp)
    add_matlib_bench(gemm_sweep          benchmarks/gemm_sweep.cpp)
    add_matlib_bench(compare_cublas      benchmarks/compare_cublas.cpp)
    add_matlib_bench(kernel_search       benchmarks/autotuning/kernel_search.cpp)
endif()

# ============================================================================
# Python bindings (optional - requires pybind11)
# ============================================================================
option(MATLIB_BUILD_PYTHON "Build Python pybind11 extension" OFF)

if(MATLIB_BUILD_PYTHON)
    include(PythonBindings)
endif()

# ============================================================================
# Install
# ============================================================================
install(TARGETS matlib_static
    ARCHIVE DESTINATION lib
)
install(DIRECTORY include/matlib DESTINATION include)
install(FILES include/matlib.hpp DESTINATION include)

# ============================================================================
# Summary
# ============================================================================
message(STATUS "")
message(STATUS "matrix-ops-lib configuration summary:")
message(STATUS "  Build type        : ${CMAKE_BUILD_TYPE}")
message(STATUS "  CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
message(STATUS "  FP16 tensor cores : ${HAS_TC}")
message(STATUS "  BF16 tensor cores : ${HAS_BF16}")
message(STATUS "  Build tests       : ${MATLIB_BUILD_TESTS}")
message(STATUS "  Build benchmarks  : ${MATLIB_BUILD_BENCHMARKS}")
message(STATUS "  Build Python ext  : ${MATLIB_BUILD_PYTHON}")
message(STATUS "")