CMAKE_CUDA_COMPILER was set, but COMPILER_ID and COMPILER_VERSION were not. I just updated to 3.10.2 and they were giving values I would expect. I dug around under the diffs on Modules but got really turned around.
I would like to set my minimum required version to whenever this was introduced, as I would like to wield it.
2. Is there any existing work / discussion on the new world order (CUDA as a full-fledged language) and Clang? I noticed that the following happens:
leaving room for a Compiler/Clang-CUDA.cmake to be created. Is there an eventual goal to write this? I think I can maybe help get a prototype working if that was the goal for this setup.
3. Supposing CMake + Clang + CUDA was working, the working assumption would be that a user sets CUDACXX to clang++, right? I’m trying to understand if Clang as a cuda compiler already works and I have bad local configurations, or if they are needing to be written.
I think I have some answers to some of these, but as I make progress will probably reach out to the LLVM mailing lists for help / suggestions (particularly with separable compilation and device linking).
I’m still ramping up to speed on developing CMake. You had mentioned CMake_TEST_CUDA which will be helpful down the line, but intermediately is it “ok” to just modify *.cmake modules in-place underneath the installed Modules and Modules/Compiler directories? I’ve got a custom install / am not concerned about breaking things in it, I’m just wondering if this is a typical practice or if I should be re-compiling and re-installing every time I make changes.