Undefined symbols when linking CUDA 10.1 on macOS 10.13

[mateuszwojt]

I'm trying to compile this library on a Mac, after some modifications. I'm running into an issue when linking CUDA. It's a series of errors, but there's a common pattern, where I get:

Undefined symbols for architecture x86_64:
  "_cuArray3DCreate_v2", referenced from:
      nvdb::Allocator::AllocateTextureGPU(nvdb::DataPtr&, unsigned char, nvdb::Vector3D<int>, bool, unsigned long long) in gvdb_allocator.cpp.o
     (maybe you meant: __ZN6cudart24__fun_cuArray3DCreate_v2E)

Every undefined symbol ends with this v2E suffix, which appears to exist only on macOS platform. I tried the same on Windows and the linking process there is working fine. I'm not sure where this E comes from, couldn't find any info in the documentation. Is that a macOS-specific thing?

My environment:

• macOS 10.13.6
• CUDA 10.1
• CUDA driver 418.163

[NBickford-NV]

Hi GreenEminence!

This is interesting - do the samples that come with CUDA build for you (they should be in a folder such as /Developer/NVIDIA/CUDA-10.1/samples)? If so, then I'm guessing this is probably a bug in GVDB's new CMakeLists (atsource/gvdb_library/CMakeLists.txt) - it's been tested on Windows and Ubuntu, but not on macOS yet.

Looking at the CMakeLists.txt file, I'm not seeing anything that stands out so far as being a bug, though - we enable CUDA support using

project(gvdb LANGUAGES CUDA CXX C)

configure the CUDA C++ standard using

if(NOT DEFINED CMAKE_CUDA_STANDARD)
    set(CMAKE_CUDA_STANDARD 11)
    set(CMAKE_CUDA_STANDARD_REQUIRED ON)
endif()

and then link CUDA using

if(WIN32)
    target_link_libraries(gvdb PUBLIC cuda.lib)
elseif(UNIX)
    find_package(X11 REQUIRED)
    target_link_libraries(gvdb
        PUBLIC  cuda
                ${X11_LIBRARIES})
endif()

This should hopefully go down the second branch - UNIX seems to be defined for macOS platforms in CMake, right? (Maybe there's something else that needs to be done in the CMake file for macOS platforms?)

Thank you!

[djspecialagent]

Would be so awesome if gvdb could be run on Mac!

On Jul 31, 2020, at 2:55 PM, Neil Bickford notifications@github.com wrote:

 Hi GreenEminence!

This is interesting - do the samples that come with CUDA build for you (they should be in a folder such as /Developer/NVIDIA/CUDA-10.1/samples)? If so, then I'm guessing this is probably a bug in GVDB's new CMakeLists (atsource/gvdb_library/CMakeLists.txt) - it's been tested on Windows and Ubuntu, but not on macOS yet.

Looking at the CMakeLists.txt file, I'm not seeing anything that stands out so far as being a bug, though - we enable CUDA support using

project(gvdb LANGUAGES CUDA CXX C) configure the CUDA C++ standard using

if(NOT DEFINED CMAKE_CUDA_STANDARD) set(CMAKE_CUDA_STANDARD 11) set(CMAKE_CUDA_STANDARD_REQUIRED ON) endif() and then link CUDA using

if(WIN32) target_link_libraries(gvdb PUBLIC cuda.lib) elseif(UNIX) find_package(X11 REQUIRED) target_link_libraries(gvdb PUBLIC cuda ${X11_LIBRARIES}) endif() This should hopefully go down the second branch - UNIX seems to be defined for macOS platforms in CMake, right? (Maybe there's something else that needs to be done in the CMake file for macOS platforms?)

Thank you!

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe.

[NBickford-NV]

Hi all - just checking in on the questions in the previous post to see if there have been any updates!

Also, just quickly checking - is it possible this could be a compiler mismatch? On Mac OS, CUDA 10.1 update 2 requires a specific version of Xcode and the Clang compiler and toolchain - see https://docs.nvidia.com/cuda/archive/10.1/cuda-installation-guide-mac-os-x/index.html#system-requirements. You can verify your installation using https://docs.nvidia.com/cuda/archive/10.1/cuda-installation-guide-mac-os-x/index.html#compiler-verification.

[mateuszwojt]

Hi @NeilBickford-NV sorry for such a late response. I couldn't make it link correctly on Mac, even after confirming all specific versions of the compiler and toolchains are correct. I think I'm gonna pass on this one, as developing anything CUDA-related on a Mac is becoming quite difficult due to lack of modern GPUs and driver support. Thanks a lot! Cheers!

[NBickford-NV]

Ah, OK, sorry about that!