# Include directories and libraries
include_directories(../../../Common)

# Source file
set(SRC_FILES
    matrixMulDrv.cpp
)

# Add target for matrixMulDrv
add_executable(matrixMulDrv ${SRC_FILES})
set_target_properties(matrixMulDrv PROPERTIES CUDA_SEPARABLE_COMPILATION ON)

target_include_directories(matrixMulDrv PRIVATE
    ${CUDAToolkit_INCLUDE_DIRS}
)

target_link_libraries(matrixMulDrv PUBLIC
    CUDA::cuda_driver
)

set(CUDA_FATBIN_FILE "${CMAKE_CURRENT_BINARY_DIR}/matrixMul_kernel64.fatbin")
set(CUDA_KERNEL_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/matrixMul_kernel.cu")

add_custom_command(
    OUTPUT ${CUDA_FATBIN_FILE}
    COMMAND ${CMAKE_CUDA_COMPILER} ${INCLUDES} ${ALL_CCFLAGS} ${GENCODE_FLAGS} -o ${CUDA_FATBIN_FILE} -fatbin ${CUDA_KERNEL_SOURCE}
    DEPENDS ${CUDA_KERNEL_SOURCE}
    COMMENT "Building CUDA fatbin: ${CUDA_FATBIN_FILE}"
)

# Create a dummy target for fatbin generation
add_custom_target(generate_fatbin ALL DEPENDS ${CUDA_FATBIN_FILE})

# Ensure matrixMulDrv depends on the fatbin
add_dependencies(matrixMulDrv generate_fatbin)