Clang CUDA Support Failing

[Axel Huebl]

Hi,

clang 3.9+ supports compiling CUDA C++ files (.cu):
  http://llvm.org/docs/CompileCudaWithLLVM.html#prerequisites

I installed the packages
  - clang-3.9
  - nvidia-cuda-toolkit (8.0)

on Ubuntu 16.04 (via the latest docker image: ubuntu:16.04).

My example file is just any .cu file, e.g. this one:
  https://github.com/ax3l/cpp_snippets/blob/master/cuda_mandelbrot/mandelbrot.cu

If I now run
  clang++ main.cu

or
  clang++ --cuda-path=/usr main.cu

I get the error

  clang: error: cannot find CUDA installation. Provide its path via --cuda-path, or pass -nocudainc to build without CUDA includes.

This is the same bug as reported upstream in
  https://bugs.llvm.org/show_bug.cgi?id=26966

due to the "scattered install" of CUDA in Debian/Ubuntu.

Is there a way to fix this from the packaging side?

[Launchpad Janitor (janitor)]

This bug was fixed in the package nvidia-cuda-toolkit - 8.0.61-3

---------------
nvidia-cuda-toolkit (8.0.61-3) unstable; urgency=medium

  * Use dpkg makefile snippets instead of manual changelog parsing.
  * Add monolithic cuda tree shim in /usr/lib/cuda/.
    Use for 'clang++ --cuda-path=/usr/lib/cuda'. (LP: #1706326)
  * Update Lintian overrides.

 -- Andreas Beckmann <email address hidden> Thu, 23 Nov 2017 05:23:45 +0100

[Axel Huebl (axel-huebl)]

Thank you, that's a great solution! :)

[Launchpad Janitor (janitor)]

Status changed to 'Confirmed' because the bug affects multiple users.

[Axel Huebl (axel-huebl)]

I'm not sure if invalid is the correct resolution status but the issue was solved now in the nvidia-cuda-toolkit package. It now installs CUDA in a monolithic path.