The problem was detected during LLVM pull down testing: http://ci.llvm.intel.com:8010/#/builders/37/builds/1152 It looks like a bug in diagnostics (AFAIK CUDA supports FP16) introduced by cf6cc662e and the tests passed before that patch.