dbcsr_init_lib_pre Subroutine

   SUBROUTINE dbcsr_init_lib_pre(mp_comm, io_unit, accdrv_active_device_id)
      !! Initialize the DBCSR library
      !! Prepares the DBCSR library for use.

#if defined(__LIBXSMM)
      USE libxsmm, ONLY: libxsmm_init
#endif
      TYPE(mp_comm_type), INTENT(IN)  :: mp_comm
      INTEGER, INTENT(IN), OPTIONAL :: io_unit, accdrv_active_device_id

      INTEGER :: numnodes, mynode

#if defined(__DBCSR_ACC)
      INTEGER :: dbcsr_thread_safe, libsmm_acc_thread_safe
#endif

      ! construct defaults which were unknown at compile-time (dbcsr_config_type)
      CALL dbcsr_set_config()

      CALL mp_environ(numnodes, mynode, mp_comm)

      IF (PRESENT(io_unit)) THEN
         ext_io_unit = io_unit
      ELSE
         ext_io_unit = 0
         IF (mynode .EQ. 0) ext_io_unit = default_output_unit
      END IF

      ! if MPI was not initialized in DBCSR, then need to check for the number of communicators
      ! when we finalize DBCSR
      IF (mp_get_comm_count() .EQ. 0) THEN
         check_comm_count = .TRUE.
      END IF
      CALL dbcsr_mp_make_env(mp_env, default_group, mp_comm)

#if defined(__LIBXSMM)
      CALL libxsmm_init()
#endif

      ! Initialize Acc and set active device
      IF (use_acc()) THEN
         IF (PRESENT(accdrv_active_device_id)) THEN
            CALL set_accdrv_active_device_id(accdrv_active_device_id)
         ELSEIF (dbcsr_acc_get_ndevices() > 0) THEN
            ! Use round-robin assignment per rank
            CALL set_accdrv_active_device_id(MOD(mynode, dbcsr_acc_get_ndevices()))
         ELSE
            DBCSR_ABORT("dbcsr_init_lib: No recognized GPU devices")
         END IF
         CALL acc_init()
      END IF

#if defined(__DBCSR_ACC)
      ! Checks related to DBCSR's GPU backend: check consistency in threading level
      libsmm_acc_thread_safe = libsmm_acc_is_thread_safe()  ! 0: not threaded, 1: threaded
      dbcsr_thread_safe = 0  ! not threaded
!$    dbcsr_thread_safe = 1  ! if DBCSR is compiled with openmp, set to threaded
      ! Check whether DBCSR and libsmm_acc (GPU backend) have the same level of threading
      IF (dbcsr_thread_safe /= libsmm_acc_thread_safe) then
         IF (dbcsr_thread_safe /= 0) then
            CALL dbcsr_abort(__LOCATION__, &
                             "DBCSR compiled w/ threading support while libsmm_acc compiled w/o threading support.")
         ELSE
            CALL dbcsr_abort(__LOCATION__, &
                             "DBCSR compiled w/o threading support while libsmm_acc is compiled w/ threading support.")
         END IF
      END IF
#endif
   END SUBROUTINE dbcsr_init_lib_pre