Running this application under a debugger, we can see that LLVM’s OpenMP implementation calls the CUDA API from a static library initializer. According to the CUDA documentation, this is not allowed.
The CUDA interfaces use global state that is initialized during host program initiation and destroyed during host program termination. The CUDA runtime and driver cannot detect if this state is invalid, so using any of these interfaces (implicitly or explicitly) during program initiation (or termination after main) will result in undefined behavior.
Nsight Computes relies on application’s using this interface correctly. We will try to detect this specific scenario and issue a proper error message, but the underlying problem ultimately needs to be addressed by LLVM itself.
#0 0x00007ffff28d9310 in cuInit () from /lib/x86_64-linux-gnu/libcuda.so.1
#1 0x00007ffff42f45f0 in llvm::omp::target::plugin::CUDAPluginTy::initImpl() () from /opt/lib/libomptarget.rtl.cuda.so
#2 0x00007ffff42ff82f in llvm::omp::target::plugin::GenericPluginTy::init() () from /opt/lib/libomptarget.rtl.cuda.so
#3 0x00007ffff42f6529 in llvm::omp::target::plugin::Plugin::Plugin() () from /opt/lib/libomptarget.rtl.cuda.so
#4 0x00007ffff42ffa8e in __tgt_rtl_init_plugin () from /opt/lib/libomptarget.rtl.cuda.so
#5 0x00007ffff7ae374d in RTLsTy::attemptLoadRTL(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, RTLInfoTy&) () from /opt/lib/libomptarget.so.18git
#6 0x00007ffff7ae3435 in RTLsTy::loadRTLs() () from /opt/lib/libomptarget.so.18git
#7 0x00007ffff7ae31a8 in init() () from /opt/lib/libomptarget.so.18git
#8 0x00007ffff7fe0b9a in call_init (l=<optimized out>, argc=argc@entry=1, argv=argv@entry=0x7fffffffe2b8, env=env@entry=0x7fffffffe2c8) at dl-init.c:72
#9 0x00007ffff7fe0ca1 in call_init (env=0x7fffffffe2c8, argv=0x7fffffffe2b8, argc=1, l=<optimized out>) at dl-init.c:30
#10 _dl_init (main_map=0x7ffff7ffe190, argc=1, argv=0x7fffffffe2b8, env=0x7fffffffe2c8) at dl-init.c:119
#11 0x00007ffff7fd013a in _dl_start_user () from /lib64/ld-linux-x86-64.so.2
#12 0x0000000000000001 in ?? ()
#13 0x00007fffffffe574 in ?? ()
#14 0x0000000000000000 in ?? ()