CMake Error: CUDA_CUDA_LIBRARY (ADVANCED) set to NOTFOUND

[ElD4n1]

Hi, I'm trying to cross-compile my own C++ application using the ZED SDK and CUDA and get some very strange error in cmake:

| -- Found CUDA: /home/datastorage/Projects/B-tastic.Device.Firmware/build/tmp/work/armv8a_tegra210-poky-linux/zed-svo-recording/1.0+gitAUTOINC+8d3267ceeb-r0/recipe-sysroot-native/usr/local/cuda-10.2 (found suitable version "10.2", minimum required is "10.2")
| -- Configuring done
| CMake Error: The following variables are used in this project, but they are set to NOTFOUND.
| Please set them or make sure they are set and tested correctly in the CMake files:
| CUDA_CUDA_LIBRARY (ADVANCED)
|     linked by target "ZED_SVO_Recording" in directory /home/datastorage/Projects/B-tastic.Device.Firmware/build/tmp/work/armv8a_tegra210-poky-linux/zed-svo-recording/1.0+gitAUTOINC+8d3267ceeb-r0/git

First it says Found CUDA and then it complains that CUDA_CUDA_LIBRARY (ADVANCED) is set to NOTFOUND? Has this anything to do with the (ADVANCED) and there is something else missing?

Here is my recipe, which I derived from #440:

LICENSE = "CLOSED"
LIC_FILES_CHKSUM = ""

SRC_URI = "<link to our private git repo>"

PV = "1.0+git${SRCPV}"
SRCREV = "8d3267ceebfe6e3bf82a56f7983573ec94d714d7"

S = "${WORKDIR}/git"

DEPENDS = "zed-sdk openblas"
RDEPENDS:${PN} = "zed-sdk openblas"

inherit cmake cuda

EXTRA_OECMAKE = "-DBLAS_LIBRARIES=${libdir}/libopenblas.so -DCMAKE_LIBRARY_PATH=/usr/local/cuda-10.2/lib/stubs/"

do_install() {
    install -D ZED_SVO_Recording -t ${D}/usr/local/bin/ -m 777
}

FILES:${PN} += "/usr/local/bin/ZED_SVO_Recording"

I also tried this (see the last EXTRA_OECMAKE arg above), but with no success.

Here the CMakeLists.txt of my program:

CMAKE_MINIMUM_REQUIRED(VERSION 2.4)
PROJECT(ZED_SVO_Recording)

option(LINK_SHARED_ZED "Link with the ZED SDK shared executable" ON)

if (NOT LINK_SHARED_ZED AND MSVC)
    message(FATAL_ERROR "LINK_SHARED_ZED OFF : ZED SDK static libraries not available on Windows")
endif()

if(COMMAND cmake_policy)
	cmake_policy(SET CMP0003 OLD)
	cmake_policy(SET CMP0015 OLD)
endif(COMMAND cmake_policy)

if (NOT CMAKE_BUILD_TYPE OR CMAKE_BUILD_TYPE STREQUAL "")
SET(CMAKE_BUILD_TYPE "RelWithDebInfo")
endif()

SET(EXECUTABLE_OUTPUT_PATH ".")

find_package(ZED 2 REQUIRED)
find_package(CUDA ${ZED_CUDA_VERSION} EXACT REQUIRED)

IF(NOT WIN32) 
    SET(SPECIAL_OS_LIBS "pthread" "X11")
ENDIF()
  
include_directories(${CUDA_INCLUDE_DIRS})
include_directories(${ZED_INCLUDE_DIRS})
include_directories(${CMAKE_CURRENT_SOURCE_DIR}/include)

link_directories(${ZED_LIBRARY_DIR})
link_directories(${CUDA_LIBRARY_DIRS})

ADD_EXECUTABLE(${PROJECT_NAME} include/utils.hpp src/main.cpp)
add_definitions(-std=c++14 -O3)

if (LINK_SHARED_ZED)
    SET(ZED_LIBS ${ZED_LIBRARIES} ${CUDA_CUDA_LIBRARY} ${CUDA_CUDART_LIBRARY} ${CUDA_NPP_LIBRARIES_ZED})
else()
    SET(ZED_LIBS ${ZED_STATIC_LIBRARIES} ${CUDA_CUDA_LIBRARY} ${CUDA_LIBRARY})
endif()

TARGET_LINK_LIBRARIES(${PROJECT_NAME} ${ZED_LIBS})

if(INSTALL_SAMPLES)
    LIST(APPEND SAMPLE_LIST ${PROJECT_NAME})
    SET(SAMPLE_LIST "${SAMPLE_LIST}" PARENT_SCOPE)
endif()

[ichergui]

Hey @ElD4n1

Could you share the result of the following command ?

$ bitbake -e YOUR_RECIPE | grep -i "^DEPENDS"

[madisongh]

First it says Found CUDA and then it complains that CUDA_CUDA_LIBRARY (ADVANCED) is set to NOTFOUND? Has this anything to do with the (ADVANCED) and there is something else missing?

The old and long-deprecated FindCUDA CMake module has trouble locating the CUDA driver library (libcuda.so) in some cases, and evidently this is one of them. The more modern approach is to use CUDA language support in CMake, which should be better about using the lib/stub directory when linking.

Otherwise, you may have to find a way to append the library (either the stub or, if you add tegra-libraries to the DEPENDS setting in your recipe, the real copy, which would be at /usr/lib/libcuda.so in your recipe sysroot) to the CUDA_LIBRARIES variable when you need to link against the driver library directly.

[ElD4n1]

Yes, by using set_property(TARGET ${PROJECT_NAME} PROPERTY CUDA_SEPARABLE_COMPILATION ON) instead of the find_package approach, configuring the build now succeeds. Many thanks!