How to correctly build a C++ CUDA project using Cmake on Linux?

Question

linux

It is not possible to write Cmake correctly so that it is possible to build the project. An undefined reference to `main' error is thrown.
Project structure:
main.cpp
include
-- CUDA_wrappers.hpp
src
-- CUDA.cu

cmake_minimum_required(VERSION 3.8)
project(Mandelbrot_set)

set(CMAKE_CXX_STANDARD 11)

set(SOURCE_FILES main.cpp include/CUDA_wrappers.hpp src/CUDA.cu)

# LIBGD Package
find_package(GD REQUIRED)
include_directories(${GD_INCLUDE_DIR})

# CUDA Package
find_package(CUDA REQUIRED)
include_directories(${CUDA_INCLUDE_DIRS})

# Compile CU files
list( APPEND CUDA_NVCC_FLAGS "-gencode arch=compute_20,code=sm_20; -std=c++11")
cuda_compile(CudaKernel src/CUDA.cu)
cuda_add_library(CudaKernels src/CUDA.cu)

# COMPILE AND LINK
cuda_add_executable(Mandelbrot_set $(SOURCE_FILES) CudaKernels)
target_link_libraries(Mandelbrot_set ${GD_LIBRARIES} ${CUDA_LIBRARIES})

I do not understand Cmake, I tried various options from the network.
CUDA seems to work correctly, it sees my adapter and correctly displays its characteristics.
CUDA_wrappers.hpp is included in main.cpp and CUDA.cu. A function is declared in CUDA_wrappers.hpp which is defined in CUDA.cu. This function is called in main.cpp
If more details are needed, I'll let you know.

Answer 1

cuda_compile(CudaKernel src/CUDA.cu)
cuda_add_library(CudaKernels src/CUDA.cu)

Suspiciously different identifiers, is this exactly how it should be? Why this cuda_compile at all? If you use it, then you need to substitute what it puts in CudaKernel into cuda_add_library, for examplecuda_add_library(CudaKernels ${CudaKernel})
Strange construction. If CudaKernels is a target, what does it do in this list, where should the sources be?
And should src/CUDA.cu be compiled twice?