https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120216
--- Comment #2 from Benjamin Schulz <schulz.benjamin at googlemail dot com> --- Hi, the reason why I am mentioning this is the following: With an nvidia card that supports HMM, like mine, the nvidia nvcc compiler can, if one starts it with the managed parameter, be set such that it can load the c++ standard template library classes into the gpu... This is frequently pointed to as a trick in the nvidia developer forums for people who want to use the standard template library on gpu, since similarly like gcc, nvcc has no mapping support yet for stl classes, (unless one uses hmm)... That gcc needs more than hmm but also pageable memory for the shared memory requires directive is a bit unfortunate...
