# CMake CUDA opinions
**URL:** https://discourse.cmake.org/t/cmake-cuda-opinions/13160
**Category:** Code
**Tags:** os:linux, lang:cuda
**Created:** [December 9, 2024, 2:31pm UTC](https://discourse.cmake.org/t/cmake-cuda-opinions/13160 "2024-12-09T14:31:00Z")
**Posts on this page:** 2
**Page:** 1
### Author:  [@eduardz1](https://discourse.cmake.org/u/eduardz1)
#### Post date: [December 9, 2024, 2:31pm UTC](https://discourse.cmake.org/t/cmake-cuda-opinions/13160/1 "2024-12-09T14:31:00Z")
This is a [cross post from the CUDA discourse](https://discourse.llvm.org/t/cmake-cuda-opinions/83561/1), it was suggested that I ask here instead.
I have a repository that is structured in this way, following the example from [ModernCMake - Extended example](https://gitlab.com/CLIUtils/modern-cmake/-/tree/master/examples/extended-project?ref_type=heads). I’m replacing with `...` extra information.
```auto
├── apps
│ ├── CMakeLists.txt
│ └── main.cpp
├── CMakeLists.txt
├── include
│ └── cuda_path_tracer
│ ├── error.cuh
│ ├── image.hpp
│ ...
├── Makefile
├── src
│ ├── CMakeLists.txt
│ ├── error.cu
│ ├── image.cpp
│ ...
└── tests
├── CMakeLists.txt
├── test_error.cpp
...
```
Can somebody help me figure out if I’m doing things correctly in the CMake configuration? I seem to keep getting errors and never seem to get to a configuration that just works. At the moment it compiles and runs fine with nvcc, but with clang++ it throws some
```auto
clang++-20: error: no such file or directory: 'CUDA'
clang++-20: error: no such file or directory: 'devices'
clang++-20: error: no such file or directory: 'found.'
clang++-20: error: no such file or directory: 'CUDA'
clang++-20: error: no such file or directory: 'devices'
clang++-20: error: no such file or directory: 'found.'
clang++-20: warning: CUDA version 12.6 is only partially supported [-Wunknown-cuda-version]
clang++-20: error: unsupported CUDA gpu architecture: sm_No
```
I’m compiling the project with the following commands:
```makefile
CC := cmake
override FLAGS := -DCMAKE_EXPORT_COMPILE_COMMANDS=ON \
-DCMAKE_CUDA_COMPILER=clang++-20 \
-DCMAKE_CXX_COMPILER=clang++-20 \
-DCMAKE_CUDA_ARCHITECTURES=native
# override FLAGS := -DCMAKE_EXPORT_COMPILE_COMMANDS=ON \
# -DCMAKE_CUDA_COMPILER=nvcc \
# -DCMAKE_CXX_COMPILER=g++
TARGET := cuda_path_tracer
.PHONY: build run clean test
build:
$(CC) -S . -B build $(FLAGS)
$(CC) --build build
test: build
./build/tests/tests
run: build
./build/apps/$(TARGET)
clean:
rm -rf *build*
```
my root CMakeList.txt is
```cmake
cmake_minimum_required(VERSION 3.27...3.31)
if(NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
set(CMAKE_CUDA_ARCHITECTURES 90)
endif()
project(
cuda_path_tracer
VERSION 1.0
LANGUAGES CXX CUDA
)
# Only set these properties if this is the main project, not included as a library
if(CMAKE_PROJECT_NAME STREQUAL PROJECT_NAME)
set(CMAKE_CXX_EXTENSIONS OFF)
set_property(GLOBAL PROPERTY USE_FOLDERS ON)
include(CTest)
endif()
include(FetchContent)
add_subdirectory(src)
add_subdirectory(apps)
if((CMAKE_PROJECT_NAME STREQUAL PROJECT_NAME OR CUDA_PATH_TRACER_CMAKE_BUILD_TESTING)
AND BUILD_TESTING)
add_subdirectory(tests)
endif()
```
the one in the `src` folder is
```cmake
set(HEADER_LIST
"${cuda_path_tracer_SOURCE_DIR}/include/cuda_path_tracer/error.cuh"
"${cuda_path_tracer_SOURCE_DIR}/include/cuda_path_tracer/image.hpp"
...
)
add_library(cuda_path_tracer_lib
error.cu
image.cpp
...
${HEADER_LIST}
)
target_include_directories(cuda_path_tracer_lib
PUBLIC
${CMAKE_SOURCE_DIR}/include
${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES}
)
target_compile_features(cuda_path_tracer_lib PUBLIC cxx_std_20)
set_target_properties(cuda_path_tracer_lib
PROPERTIES
CUDA_SEPARABLE_COMPILATION ON
CUDA_RESOLVE_DEVICE_SYMBOLS ON
)
source_group(
TREE "${PROJECT_SOURCE_DIR}/include"
PREFIX "Header Files"
FILES ${HEADER_LIST}
)
```
The one in the `apps` is
```cmake
add_executable(cuda_path_tracer main.cpp)
set_target_properties(cuda_path_tracer
PROPERTIES
CUDA_SEPARABLE_COMPILATION ON
CUDA_RESOLVE_DEVICE_SYMBOLS ON
)
target_compile_features(cuda_path_tracer PRIVATE cxx_std_20)
target_link_libraries(cuda_path_tracer PRIVATE cuda_path_tracer_lib)
```
while the one in `tests` is:
```cmake
FetchContent_Declare(
Catch2
GIT_REPOSITORY https://github.com/catchorg/Catch2.git
GIT_TAG v3.7.1
)
FetchContent_MakeAvailable(Catch2)
add_executable(tests
test_error.cpp
...
)
include(Catch)
catch_discover_tests(tests)
set_target_properties(tests
PROPERTIES
CUDA_SEPARABLE_COMPILATION ON
)
target_compile_features(tests PRIVATE cxx_std_20)
target_link_libraries(tests PRIVATE cuda_path_tracer_lib Catch2::Catch2WithMain)
add_test(NAME test_all COMMAND tests)
```
Is there something that I’m doing wrong/that can be improved? Initially I had CUDA\_ARCHITECTURES defined in the `set_target_properties` of the inner CMakeList.txt, but this lead to correct compilation both with nvcc and clang++ but errors at runtime due to library linkage issues.
I’m sorry if it’s such an open-ended question, but I’m new to CMake and I want to learn it right
---
### Author:  [@jwm](https://discourse.cmake.org/u/jwm)
#### Post date: [October 8, 2025, 8:01pm UTC](https://discourse.cmake.org/t/cmake-cuda-opinions/13160/2 "2025-10-08T20:01:46Z")
This is a bit old, but a couple of things jump out at me.
First, why are you showing us a Makefile? Are you saying that is what CMake generated? Or did you write a Makefile to call CMake?
Second, why are you overriding the CUDA compiler with `clang++-20`? I don’t expect your CUDA code to compile without using `nvcc`.
In general, the less you specify the better. Try a vanilla project that just defines the project and declares a target. Grow from there.