CUDA 10.0 compiler can't figure out templates of nlohmann::json

[nadrino]

While trying to compile GUNDAM on L4T Ubuntu 18.04 (https://wiki.switchroot.org/wiki/linux/l4t-ubuntu-bionic-installation-guide) with the GPU fonctionalities, nvcc was not able to parse some template functions of the nlohmann::json requested by some ROOT classes:

to be copy-pasted (waiting for the log)

Here is the version of nvcc:

ablanche@ablanche-laptop:~/Documents/work/build/gundam$ /usr/local/cuda/bin/nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2019 NVIDIA Corporation
Built on Mon_Mar_11_22:13:24_CDT_2019
Cuda compilation tools, release 10.0, V10.0.326

And the CMake command line:

cmake \
  -D CMAKE_INSTALL_PREFIX:PATH=$INSTALL_DIR/gundam  \
  -D CMAKE_BUILD_TYPE=Release \
  -D WITH_CUDA_LIB=ON \
  -D CMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc \
  -D CMAKE_CUDA_ARCHITECTURES=native \
  $REPO_DIR/gundam/.

Apparently devs from the JSON lib have gave up supporting this version of CUDA as this is related to a compiler bug: nlohmann/json-ci#8

Unless we find a workaround for this, I suggest to require a minimum version for CUDA (apparently v11 should be fine?).

[nadrino]

There might be a solution for this: in princple the CacheManager lib don't need ROOT at all. The compiler still request it from the headers that are coming for some of the #include.

Some of the dials might require ROOT, like TSpline3 for instance, but it could be useful to see if we could externalize the conversion from ROOT types to GPU-friendly types

[nadrino]

In order to get rid of ROOT headers in sublying dependencies, we would need to clean the headers from ROOT includes and only use them in the .cpp files (which are not seen by the CUDA compiler)