gs_utofu.F90

Go to the documentation of this file.

! Copyright (c) 2026, The Neko Authors
! All rights reserved.
!
! Redistribution and use in source and binary forms, with or without
! modification, are permitted provided that the following conditions
! are met:
!
!   * Redistributions of source code must retain the above copyright
!     notice, this list of conditions and the following disclaimer.
!
!   * Redistributions in binary form must reproduce the above
!     copyright notice, this list of conditions and the following
!     disclaimer in the documentation and/or other materials provided
!     with the distribution.
!
!   * Neither the name of the authors nor the names of its
!     contributors may be used to endorse or promote products derived
!     from this software without specific prior written permission.
!
! THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
! "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
! LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
! FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
! COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
! INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
! BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
! LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
! CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
! LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
! ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
! POSSIBILITY OF SUCH DAMAGE.
!
module gs_utofu
  use num_types, only : rp, i8
  use gs_comm, only : gs_comm_t, gs_comm_utofu, gs_vec_nc
  use gs_ops, only : gs_op_add, gs_op_max, gs_op_min, gs_op_mul
  use stack, only : stack_i4_t
  use comm, only : neko_comm, pe_rank
  use logger, only : neko_log, log_size
#ifdef HAVE_UTOFU
  use mpi_f08, only : mpi_request, mpi_isend, mpi_irecv, mpi_waitall, &
       mpi_integer8, mpi_statuses_ignore
#endif
  use utils, only : neko_error
  !$ use omp_lib, only : omp_get_max_threads, omp_get_thread_num, &
  !$      omp_get_num_threads
  use, intrinsic :: iso_c_binding
  implicit none
  private
 
  integer, parameter :: gs_utofu_xchg_tag = 8123
  integer, parameter :: gs_utofu_xchg_tag_v = 8124
 
  type, public, extends(gs_comm_t) :: gs_utofu_t
     real(kind=rp), pointer :: send_buf(:) => null()
     real(kind=rp), pointer :: recv_buf(:) => null()
     integer, allocatable :: send_len(:), recv_len(:)
     integer, allocatable :: send_offset(:), recv_offset(:)
     integer(c_int64_t), allocatable :: send_off64(:), send_len64(:)
     integer(c_int64_t), allocatable :: rmt_vcq_id(:), rmt_stadd(:)
     integer(c_int64_t), allocatable :: dst_off0(:), dst_off1(:)
     integer(c_int), allocatable :: dst_tag(:)
     integer(c_int), allocatable :: recv_indices(:)
     integer(c_int) :: ncompleted = 0
     integer :: buf_size = 0
     integer :: send_size = 0
     integer :: elem_size = 0
     integer :: parity = 0
     type(c_ptr) :: ctx = c_null_ptr
     real(kind=rp), pointer :: send_buf_v(:) => null()
     real(kind=rp), pointer :: recv_buf_v(:) => null()
     integer(c_int64_t), allocatable :: rmt_vcq_id_v(:), rmt_stadd_v(:)
     integer(c_int64_t), allocatable :: rmt_off_v(:), rmt_bufsz_v(:)
     integer(c_int64_t), allocatable :: send_off64_v(:), send_len64_v(:)
     integer(c_int64_t), allocatable :: dst_off_v(:)
     integer :: parity_v = 0
     type(c_ptr) :: ctx_v = c_null_ptr
   contains
     procedure, pass(this) :: init => gs_utofu_init
     procedure, pass(this) :: free => gs_utofu_free
     procedure, pass(this) :: nbsend => gs_nbsend_utofu
     procedure, pass(this) :: nbrecv => gs_nbrecv_utofu
     procedure, pass(this) :: nbwait => gs_nbwait_utofu
     procedure, pass(this) :: nbsend_vec => gs_nbsend_vec_utofu
     procedure, pass(this) :: nbrecv_vec => gs_nbrecv_vec_utofu
     procedure, pass(this) :: nbwait_vec => gs_nbwait_vec_utofu
  end type gs_utofu_t
 
#ifdef HAVE_UTOFU
 
  integer(c_int64_t) :: gs_utofu_edata_max = 255
  integer(c_int) :: gs_utofu_nrvcq = 1
  logical, save :: gs_utofu_logged = .false.
 
  interface
     subroutine gs_utofu_init_ep(ntni_req, nthreads, ntni_out, nvcq_out, &
          nrvcq_out, edata_max, ierr) bind(c, name = 'gs_utofu_init')
       import :: c_int, c_int64_t
       integer(c_int), value :: ntni_req, nthreads
       integer(c_int) :: ntni_out, nvcq_out, nrvcq_out
       integer(c_int64_t) :: edata_max
       integer(c_int) :: ierr
     end subroutine gs_utofu_init_ep
 
     subroutine gs_utofu_ctx_create(send_buf, send_bytes, recv_buf, &
          recv_bytes, ctx, nrvcq, recv_vcq_id, recv_stadd, ierr) &
          bind(c, name = 'gs_utofu_ctx_create')
       import :: c_ptr, c_size_t, c_int64_t, c_int
       type(c_ptr), value :: send_buf, recv_buf
       integer(c_size_t), value :: send_bytes, recv_bytes
       type(c_ptr) :: ctx
       integer(c_int) :: nrvcq
       integer(c_int64_t) :: recv_vcq_id(*)
       integer(c_int64_t) :: recv_stadd(*)
       integer(c_int) :: ierr
     end subroutine gs_utofu_ctx_create
 
     subroutine gs_utofu_ctx_free(ctx, ierr) bind(c, name = 'gs_utofu_ctx_free')
       import :: c_ptr, c_int
       type(c_ptr), value :: ctx
       integer(c_int) :: ierr
     end subroutine gs_utofu_ctx_free
 
     subroutine gs_utofu_post_puts(ctx, tid, nteam, nsend, send_off, &
          send_len, rmt_vcq_id, rmt_stadd, dst_off0, dst_off1, dst_tag, &
          parity, send_base, elem_size, ierr) &
          bind(c, name = 'gs_utofu_post_puts')
       import :: c_ptr, c_int, c_int64_t
       type(c_ptr), value :: ctx
       integer(c_int), value :: tid, nteam, nsend, parity, elem_size
       integer(c_int64_t) :: send_off(*), send_len(*)
       integer(c_int64_t) :: rmt_vcq_id(*), rmt_stadd(*)
       integer(c_int64_t) :: dst_off0(*), dst_off1(*)
       integer(c_int) :: dst_tag(*)
       integer(c_int64_t), value :: send_base
       integer(c_int) :: ierr
     end subroutine gs_utofu_post_puts
 
     subroutine gs_utofu_poll_recv(ctx, parity, cap, ncompleted, out_idx, &
          ierr) bind(c, name = 'gs_utofu_poll_recv')
       import :: c_ptr, c_int
       type(c_ptr), value :: ctx
       integer(c_int), value :: parity, cap
       integer(c_int) :: ncompleted
       integer(c_int) :: out_idx(*)
       integer(c_int) :: ierr
     end subroutine gs_utofu_poll_recv
 
     subroutine gs_utofu_drain_half(ctx, half, tid, nteam, ierr) &
          bind(c, name = 'gs_utofu_drain_half')
       import :: c_ptr, c_int
       type(c_ptr), value :: ctx
       integer(c_int), value :: half, tid, nteam
       integer(c_int) :: ierr
     end subroutine gs_utofu_drain_half
  end interface
#endif
 
contains
 
  subroutine gs_utofu_init(this, send_pe, recv_pe)
    class(gs_utofu_t), intent(inout) :: this
    type(stack_i4_t), intent(inout) :: send_pe
    type(stack_i4_t), intent(inout) :: recv_pe
#ifdef HAVE_UTOFU
    integer :: i, nsend, nrecv, send_total, recv_total
    integer :: ntni_req, env_len, max_tag, nthreads
    integer(c_int) :: ntni, nvcq, nrv, ierr
    integer(c_int64_t), allocatable :: recv_vcq_id(:), recv_stadd(:)
    integer(c_size_t) :: send_bytes, recv_bytes
    character(len=32) :: env_val
    character(len=LOG_SIZE) :: log_buf
    integer(c_int64_t), allocatable :: msg_out(:,:), msg_in(:,:)
    type(mpi_request), allocatable :: sreq(:), rreq(:)
 
    call this%init_order(send_pe, recv_pe)
 
    nsend = size(this%send_pe)
    nrecv = size(this%recv_pe)
    this%elem_size = int(storage_size(1.0_rp) / 8)
 
    allocate(this%send_len(nsend), this%send_offset(nsend))
    allocate(this%send_off64(nsend), this%send_len64(nsend))
    allocate(this%rmt_vcq_id(nsend), this%rmt_stadd(nsend))
    allocate(this%dst_off0(nsend), this%dst_off1(nsend), this%dst_tag(nsend))
    allocate(this%recv_len(nrecv), this%recv_offset(nrecv))
    allocate(this%recv_indices(nrecv))
 
    ! Local receive layout (one concatenated half; the other half mirrors it).
    recv_total = 0
    do i = 1, nrecv
       this%recv_len(i) = this%recv_dof(this%recv_pe(i))%size()
       this%recv_offset(i) = recv_total
       recv_total = recv_total + this%recv_len(i)
    end do
    this%buf_size = recv_total
 
    ! Local send layout (concatenated per-peer slabs).
    send_total = 0
    do i = 1, nsend
       this%send_len(i) = this%send_dof(this%send_pe(i))%size()
       this%send_offset(i) = send_total
       this%send_off64(i) = int(send_total, c_int64_t)
       this%send_len64(i) = int(this%send_len(i), c_int64_t)
       send_total = send_total + this%send_len(i)
    end do
    this%send_size = send_total
 
    allocate(this%send_buf(max(1, 2 * send_total)))
    allocate(this%recv_buf(max(1, 2 * recv_total)))
 
    ! Create the (process-global) uTofu endpoint and register our buffers.
    ntni_req = 1
    call get_environment_variable("NEKO_GS_UTOFU_NTNI", env_val, env_len)
    if (env_len .gt. 0) read(env_val(1:env_len), *) ntni_req
 
    ! One injection VCQ is requested per OpenMP thread, dealt over the TNIs.
    nthreads = 1
    !$ nthreads = omp_get_max_threads()
 
    call gs_utofu_init_ep(int(ntni_req, c_int), int(nthreads, c_int), &
         ntni, nvcq, gs_utofu_nrvcq, gs_utofu_edata_max, ierr)
    if (ierr .ne. 0) call neko_error("gs_utofu: endpoint init failed")
 
    send_bytes = int(size(this%send_buf), c_size_t) * this%elem_size
    recv_bytes = int(size(this%recv_buf), c_size_t) * this%elem_size
    ! Filled by gs_utofu_ctx_create: one (vcq_id, stadd) pair per receive
    ! VCQ granted to this instance; each receive peer is bound to one pair
    ! below. Initialised to keep the compiler from warning (it cannot see
    ! the C side write).
    allocate(recv_vcq_id(gs_utofu_nrvcq), recv_stadd(gs_utofu_nrvcq))
    recv_vcq_id = 0_c_int64_t
    recv_stadd = 0_c_int64_t
    nrv = 1
    call gs_utofu_ctx_create(c_loc(this%send_buf), send_bytes, &
         c_loc(this%recv_buf), recv_bytes, this%ctx, nrv, recv_vcq_id, &
         recv_stadd, ierr)
    if (ierr .eq. 1) then
       call neko_error("gs_utofu: out of VCQs for the receive queues " // &
            "(lower OMP_NUM_THREADS or NEKO_GS_UTOFU_NRVCQ, or set " // &
            "NEKO_GS_UTOFU_NVCQ)")
    else if (ierr .ne. 0) then
       call neko_error("gs_utofu: buffer registration failed")
    end if
 
    ! Report what the hardware actually granted: requested vs granted can
    ! differ when ranks on a node share its TNIs, or when the VCQ budget
    ! runs dry. Once, on rank 0.
    if (.not. gs_utofu_logged .and. pe_rank .eq. 0) then
       write(log_buf, '(A,I0,A,I0,A,I0,A,I0,A)') 'uTofu inj.   : ', &
            int(nvcq), ' VCQs (', nthreads, ' threads) over ', int(ntni), &
            ' TNIs (', ntni_req, ' requested)'
       call neko_log%message(log_buf)
       write(log_buf, '(A,I0,A,I0,A)') 'uTofu recv   : ', int(nrv), &
            ' of ', int(gs_utofu_nrvcq), ' VCQs (first instance)'
       call neko_log%message(log_buf)
       gs_utofu_logged = .true.
    end if
 
    ! Tell each sender where, and with which tag, to put our slab; learn the
    ! same for each receiver we send to. Message layout per peer:
    !   [recv_stadd, recv_vcq_id, byte_off_parity0, byte_off_parity1, tag].
    allocate(msg_out(5, max(1, nrecv)), msg_in(5, max(1, nsend)))
    allocate(sreq(max(1, nrecv)), rreq(max(1, nsend)))
 
    ! Receive peer i is bound to receive VCQ mod(i-1, nrv) + 1; advertising
    ! that VCQ's STADD/id spreads the incoming slabs over the TNIs.
    do i = 1, nrecv
       msg_out(1, i) = recv_stadd(mod(i - 1, int(nrv)) + 1)
       msg_out(2, i) = recv_vcq_id(mod(i - 1, int(nrv)) + 1)
       msg_out(3, i) = int(this%recv_offset(i), c_int64_t)
       msg_out(4, i) = int(this%buf_size + this%recv_offset(i), c_int64_t)
       msg_out(5, i) = int(i - 1, c_int64_t)
    end do
 
    do i = 1, nsend
       call mpi_irecv(msg_in(:, i), 5, mpi_integer8, this%send_pe(i), &
            gs_utofu_xchg_tag, neko_comm, rreq(i))
    end do
    do i = 1, nrecv
       call mpi_isend(msg_out(:, i), 5, mpi_integer8, this%recv_pe(i), &
            gs_utofu_xchg_tag, neko_comm, sreq(i))
    end do
    if (nsend .gt. 0) call mpi_waitall(nsend, rreq, mpi_statuses_ignore)
    if (nrecv .gt. 0) call mpi_waitall(nrecv, sreq, mpi_statuses_ignore)
 
    max_tag = 0
    do i = 1, nsend
       this%rmt_stadd(i) = msg_in(1, i)
       this%rmt_vcq_id(i) = msg_in(2, i)
       this%dst_off0(i) = msg_in(3, i)
       this%dst_off1(i) = msg_in(4, i)
       this%dst_tag(i) = int(msg_in(5, i), c_int)
       max_tag = max(max_tag, this%dst_tag(i))
    end do
 
    deallocate(msg_out, msg_in, sreq, rreq)
 
    ! The edata carries the slab tag in its high bits and the parity bit in
    ! bit 0, so a neighbour's count must fit (2*tag + 1) in the edata range.
    ! The vector path reuses the same tags, so this check covers it too.
    if (int(2 * max_tag + 1, i8) .gt. int(gs_utofu_edata_max, i8)) &
         call neko_error("gs_utofu: neighbour count exceeds edata capacity")
 
    ! Fused vector path: double-buffered slabs sized for GS_VEC_NC
    ! components, registered under their own context (own receive VCQs).
    allocate(this%send_buf_v(max(1, 2 * gs_vec_nc * send_total)))
    allocate(this%recv_buf_v(max(1, 2 * gs_vec_nc * recv_total)))
    allocate(this%rmt_vcq_id_v(nsend), this%rmt_stadd_v(nsend))
    allocate(this%rmt_off_v(nsend), this%rmt_bufsz_v(nsend))
    allocate(this%send_off64_v(nsend), this%send_len64_v(nsend))
    allocate(this%dst_off_v(nsend))
 
    send_bytes = int(size(this%send_buf_v), c_size_t) * this%elem_size
    recv_bytes = int(size(this%recv_buf_v), c_size_t) * this%elem_size
    recv_vcq_id = 0_c_int64_t
    recv_stadd = 0_c_int64_t
    nrv = 1
    call gs_utofu_ctx_create(c_loc(this%send_buf_v), send_bytes, &
         c_loc(this%recv_buf_v), recv_bytes, this%ctx_v, nrv, recv_vcq_id, &
         recv_stadd, ierr)
    if (ierr .eq. 1) then
       call neko_error("gs_utofu: out of VCQs for the vector receive " // &
            "queues (lower OMP_NUM_THREADS or NEKO_GS_UTOFU_NRVCQ, or " // &
            "set NEKO_GS_UTOFU_NVCQ)")
    else if (ierr .ne. 0) then
       call neko_error("gs_utofu: vector buffer registration failed")
    end if
 
    ! Second metadata exchange for the vector path. Because the component
    ! count nc varies per round, the receiver advertises its UNSCALED layout
    ! -- base STADD, VCQ id, our slab's scalar offset, and its scalar half
    ! size -- and the sender scales by nc and folds the parity in at round
    ! time. Tags are the scalar ones (same peer ordering).
    allocate(msg_out(4, max(1, nrecv)), msg_in(4, max(1, nsend)))
    allocate(sreq(max(1, nrecv)), rreq(max(1, nsend)))
 
    ! Same peer-to-receive-VCQ binding as the scalar path (its own VCQ set).
    do i = 1, nrecv
       msg_out(1, i) = recv_stadd(mod(i - 1, int(nrv)) + 1)
       msg_out(2, i) = recv_vcq_id(mod(i - 1, int(nrv)) + 1)
       msg_out(3, i) = int(this%recv_offset(i), c_int64_t)
       msg_out(4, i) = int(this%buf_size, c_int64_t)
    end do
 
    do i = 1, nsend
       call mpi_irecv(msg_in(:, i), 4, mpi_integer8, this%send_pe(i), &
            gs_utofu_xchg_tag_v, neko_comm, rreq(i))
    end do
    do i = 1, nrecv
       call mpi_isend(msg_out(:, i), 4, mpi_integer8, this%recv_pe(i), &
            gs_utofu_xchg_tag_v, neko_comm, sreq(i))
    end do
    if (nsend .gt. 0) call mpi_waitall(nsend, rreq, mpi_statuses_ignore)
    if (nrecv .gt. 0) call mpi_waitall(nrecv, sreq, mpi_statuses_ignore)
 
    do i = 1, nsend
       this%rmt_stadd_v(i) = msg_in(1, i)
       this%rmt_vcq_id_v(i) = msg_in(2, i)
       this%rmt_off_v(i) = msg_in(3, i)
       this%rmt_bufsz_v(i) = msg_in(4, i)
    end do
 
    deallocate(msg_out, msg_in, sreq, rreq)
 
    ! Runtime kill-switch for the fused vector path (validation aid):
    ! NEKO_GS_UTOFU_VEC=0 makes gs_op_r3 fall back to three scalar rounds.
    ! The vector context stays registered either way; only its use is gated.
    call get_environment_variable("NEKO_GS_UTOFU_VEC", env_val, env_len)
    this%vec_supported = .not. (env_len .gt. 0 .and. env_val(1:1) .eq. '0')
#else
    call neko_error("uTofu support not built; reconfigure with --with-utofu")
#endif
  end subroutine gs_utofu_init
 
  subroutine gs_utofu_free(this)
    class(gs_utofu_t), intent(inout) :: this
#ifdef HAVE_UTOFU
    integer(c_int) :: ierr
 
    if (c_associated(this%ctx)) then
       call gs_utofu_ctx_free(this%ctx, ierr)
       this%ctx = c_null_ptr
    end if
 
    if (c_associated(this%ctx_v)) then
       call gs_utofu_ctx_free(this%ctx_v, ierr)
       this%ctx_v = c_null_ptr
    end if
 
    if (associated(this%send_buf)) then
       deallocate(this%send_buf)
    end if
 
    if (associated(this%recv_buf)) then
       deallocate(this%recv_buf)
    end if
 
    if (associated(this%send_buf_v)) then
       deallocate(this%send_buf_v)
    end if
 
    if (associated(this%recv_buf_v)) then
       deallocate(this%recv_buf_v)
    end if
 
    if (allocated(this%send_len)) then
       deallocate(this%send_len)
    end if
 
    if (allocated(this%recv_len)) then
       deallocate(this%recv_len)
    end if
 
    if (allocated(this%send_offset)) then
       deallocate(this%send_offset)
    end if
 
    if (allocated(this%recv_offset)) then
       deallocate(this%recv_offset)
    end if
 
    if (allocated(this%send_off64)) then
       deallocate(this%send_off64)
    end if
 
    if (allocated(this%send_len64)) then
       deallocate(this%send_len64)
    end if
 
    if (allocated(this%rmt_vcq_id)) then
       deallocate(this%rmt_vcq_id)
    end if
 
    if (allocated(this%rmt_stadd)) then
       deallocate(this%rmt_stadd)
    end if
 
    if (allocated(this%dst_off0)) then
       deallocate(this%dst_off0)
    end if
 
    if (allocated(this%dst_off1)) then
       deallocate(this%dst_off1)
    end if
 
    if (allocated(this%dst_tag)) then
       deallocate(this%dst_tag)
    end if
 
    if (allocated(this%recv_indices)) then
       deallocate(this%recv_indices)
    end if
 
    if (allocated(this%rmt_vcq_id_v)) then
       deallocate(this%rmt_vcq_id_v)
    end if
 
    if (allocated(this%rmt_stadd_v)) then
       deallocate(this%rmt_stadd_v)
    end if
 
    if (allocated(this%rmt_off_v)) then
       deallocate(this%rmt_off_v)
    end if
 
    if (allocated(this%rmt_bufsz_v)) then
       deallocate(this%rmt_bufsz_v)
    end if
 
    if (allocated(this%send_off64_v)) then
       deallocate(this%send_off64_v)
    end if
 
    if (allocated(this%send_len64_v)) then
       deallocate(this%send_len64_v)
    end if
 
    if (allocated(this%dst_off_v)) then
       deallocate(this%dst_off_v)
    end if
 
    call this%free_order()
    call this%free_dofs()
#endif
  end subroutine gs_utofu_free
 
  subroutine gs_nbsend_utofu(this, u, n, tag, deps, strm)
    class(gs_utofu_t), intent(inout) :: this
    integer, intent(in) :: n
    real(kind=rp), dimension(n), intent(inout) :: u
    integer, intent(in) :: tag
    type(c_ptr), intent(inout) :: deps
    type(c_ptr), intent(inout) :: strm
#ifdef HAVE_UTOFU
    integer :: i, j, dst, off, ndst, send_base, tid, nteam
    integer(c_int) :: ierr
    integer, pointer :: sp(:)
 
    ! Entered from inside the gs_op_vector OpenMP parallel region, and every
    ! phase is work-shared: each thread owns one injection VCQ (created
    ! THREAD_SAFE) and fires the puts for its share of the peers itself --
    ! the uTofu substitute for the per-thread MPI_Isend that Fujitsu MPI's
    ! missing MPI_THREAD_MULTIPLE support rules out.
    tid = 0
    nteam = 1
    !$ tid = omp_get_thread_num()
    !$ nteam = omp_get_num_threads()
 
    ! Reclaim the send-buffer half we are about to repack: the team strides
    ! over the injection VCQs so each one is drained by exactly one thread
    ! (no atomics on the completion counters). The half was last used two
    ! rounds ago, so its puts have long completed -- this just reaps their
    ! TCQ notices. The barrier then guarantees the whole half is free before
    ! anyone packs into it.
    call gs_utofu_drain_half(this%ctx, int(this%parity, c_int), &
         int(tid, c_int), int(nteam, c_int), ierr)
    if (ierr .ne. 0) call neko_error("gs_utofu: send drain failed")
    !$omp barrier
 
    ! Pack each peer's slab in one work-shared loop. Subroutine locals are
    ! per-thread private (orphaned region); the implicit barrier at end do
    ! guarantees every slab is packed before any thread fires its puts (a
    ! peer packed by one thread may be sent by another -- the pack and post
    ! partitions differ).
    send_base = this%parity * this%send_size
    !$omp do
    do i = 1, size(this%send_pe)
       dst = this%send_pe(i)
       off = this%send_offset(i)
       ndst = this%send_len(i)
       sp => this%send_dof(dst)%array()
       !OCL NORECURRENCE, NOVREC, NOALIAS
       !DIR$ CONCURRENT
       !DIR$ IVDEP
       !GCC$ ivdep
       !NEC$ IVDEP
       !$omp simd
       do j = 1, ndst
          this%send_buf(send_base + off + j) = u(sp(j))
       end do
    end do
    !$omp end do
 
    ! Fire the one-sided puts: each thread injects its modulo share of the
    ! peers on its own VCQ (or thread 0 injects everything when
    ! NEKO_GS_UTOFU_MASTER_INJECT is set, as an A/B reference). The trailing
    ! barrier holds the team until every put has been issued.
    call gs_utofu_post_puts(this%ctx, int(tid, c_int), int(nteam, c_int), &
         int(size(this%send_pe), c_int), &
         this%send_off64, this%send_len64, this%rmt_vcq_id, this%rmt_stadd, &
         this%dst_off0, this%dst_off1, this%dst_tag, &
         int(this%parity, c_int), int(send_base, c_int64_t), &
         int(this%elem_size, c_int), ierr)
    if (ierr .ne. 0) call neko_error("gs_utofu: put failed")
    !$omp barrier
#else
    call neko_error("uTofu support not built")
#endif
  end subroutine gs_nbsend_utofu
 
  subroutine gs_nbrecv_utofu(this, tag)
    class(gs_utofu_t), intent(inout) :: this
    integer, intent(in) :: tag
  end subroutine gs_nbrecv_utofu
 
  subroutine gs_nbwait_utofu(this, u, n, op, strm)
    class(gs_utofu_t), intent(inout) :: this
    integer, intent(in) :: n
    real(kind=rp), dimension(n), intent(inout) :: u
    type(c_ptr), intent(inout) :: strm
    integer :: op
#ifdef HAVE_UTOFU
    integer :: i, j, k, src, off, nsrc, nreqs, base
    integer(c_int) :: ierr
    integer, pointer :: sp(:)
    ! Bind a local pointer to the receive buffer and index it in the unpack
    ! loops, mirroring how the CAF backend uses its module-level buffer.
    real(kind=rp), pointer :: rbuf(:)
 
    base = this%parity * this%buf_size
    nreqs = size(this%recv_pe)
    rbuf => this%recv_buf
 
    do while (nreqs .gt. 0)
       !$omp master
       ! Spin on the MRQ from the master alone until at least one slab lands.
       ! Polling here (rather than around the whole team) keeps the idle
       ! threads from paying a barrier on every empty poll -- the dominant
       ! cost under interconnect latency.
       this%ncompleted = 0
       do while (this%ncompleted .eq. 0)
          call gs_utofu_poll_recv(this%ctx, int(this%parity, c_int), &
               int(size(this%recv_pe), c_int), this%ncompleted, &
               this%recv_indices, ierr)
          if (ierr .ne. 0) call neko_error("gs_utofu: mrq poll failed")
       end do
       !$omp end master
       !$omp barrier
 
       do k = 1, this%ncompleted
          i = this%recv_indices(k)
          src = this%recv_pe(i)
          off = this%recv_offset(i)
          nsrc = this%recv_len(i)
          sp => this%recv_dof(src)%array()
          select case (op)
          case (gs_op_add)
             !OCL NORECURRENCE, NOVREC, NOALIAS
             !DIR$ CONCURRENT
             !DIR$ IVDEP
             !GCC$ ivdep
             !NEC$ IVDEP
             !$omp do
             do j = 1, nsrc
                u(sp(j)) = u(sp(j)) + rbuf(base + off + j)
             end do
             !$omp end do
          case (gs_op_mul)
             !OCL NORECURRENCE, NOVREC, NOALIAS
             !DIR$ CONCURRENT
             !DIR$ IVDEP
             !GCC$ ivdep
             !NEC$ IVDEP
             !$omp do
             do j = 1, nsrc
                u(sp(j)) = u(sp(j)) * rbuf(base + off + j)
             end do
             !$omp end do
          case (gs_op_min)
             !OCL NORECURRENCE, NOVREC, NOALIAS
             !DIR$ CONCURRENT
             !DIR$ IVDEP
             !GCC$ ivdep
             !NEC$ IVDEP
             !$omp do
             do j = 1, nsrc
                u(sp(j)) = min(u(sp(j)), rbuf(base + off + j))
             end do
             !$omp end do
          case (gs_op_max)
             !OCL NORECURRENCE, NOVREC, NOALIAS
             !DIR$ CONCURRENT
             !DIR$ IVDEP
             !GCC$ ivdep
             !NEC$ IVDEP
             !$omp do
             do j = 1, nsrc
                u(sp(j)) = max(u(sp(j)), rbuf(base + off + j))
             end do
             !$omp end do
          case default
             call neko_error("Unknown operation in gs_nbwait_utofu")
          end select
       end do
 
       nreqs = nreqs - this%ncompleted
       ! Hold the team until every thread has read ncompleted (above) before
       ! the master re-enters the poll and overwrites recv_indices/ncompleted.
       !$omp barrier
    end do
 
    !$omp master
    ! Flip to the other double-buffer half for the next round. Send completions
    ! are NOT waited on here; nbsend drains the half it reuses, keeping the
    ! send-completion wait off the critical path.
    this%parity = 1 - this%parity
    !$omp end master
    !$omp barrier
#else
    call neko_error("uTofu support not built")
#endif
  end subroutine gs_nbwait_utofu
 
  subroutine gs_nbsend_vec_utofu(this, u, n, nc, tag, deps, strm)
    class(gs_utofu_t), intent(inout) :: this
    integer, intent(in) :: n, nc
    real(kind=rp), dimension(nc*n), intent(inout) :: u
    integer, intent(in) :: tag
    type(c_ptr), intent(inout) :: deps
    type(c_ptr), intent(inout) :: strm
#ifdef HAVE_UTOFU
    integer :: i, j, c, dst, off, ndst, send_base, tid, nteam
    integer(c_int) :: ierr
    integer, pointer :: sp(:)
 
    tid = 0
    nteam = 1
    !$ tid = omp_get_thread_num()
    !$ nteam = omp_get_num_threads()
 
    ! Reclaim the vector send-buffer half we are about to repack (see
    ! gs_nbsend_utofu for the ownership rules).
    call gs_utofu_drain_half(this%ctx_v, int(this%parity_v, c_int), &
         int(tid, c_int), int(nteam, c_int), ierr)
    if (ierr .ne. 0) call neko_error("gs_utofu: vector send drain failed")
    !$omp barrier
 
    ! Scale the scalar per-peer layout by this round's nc and fold the
    ! vector parity into the destination offsets. Master-only; the pack
    ! loop's implicit end-do barrier orders these writes before any thread
    ! posts.
    send_base = this%parity_v * (gs_vec_nc * this%send_size)
    !$omp master
    do i = 1, size(this%send_pe)
       this%send_off64_v(i) = int(nc, c_int64_t) * this%send_off64(i)
       this%send_len64_v(i) = int(nc, c_int64_t) * this%send_len64(i)
       this%dst_off_v(i) = int(this%parity_v, c_int64_t) &
            * (gs_vec_nc * this%rmt_bufsz_v(i)) &
            + int(nc, c_int64_t) * this%rmt_off_v(i)
    end do
    !$omp end master
 
    ! Pack each peer's slab: nc consecutive component blocks.
    !$omp do
    do i = 1, size(this%send_pe)
       dst = this%send_pe(i)
       off = this%send_offset(i)
       ndst = this%send_len(i)
       sp => this%send_dof(dst)%array()
       do c = 1, nc
          !OCL NORECURRENCE, NOVREC, NOALIAS
          !DIR$ CONCURRENT
          !DIR$ IVDEP
          !GCC$ ivdep
          !NEC$ IVDEP
          !$omp simd
          do j = 1, ndst
             this%send_buf_v(send_base + nc*off + (c-1)*ndst + j) = &
                  u((c-1)*n + sp(j))
          end do
       end do
    end do
    !$omp end do
 
    ! Fire the puts (see gs_nbsend_utofu). The same dst_off_v serves both
    ! parity slots of the C transport -- the vector parity is already folded
    ! into the offsets; the parity argument still stamps the edata bit.
    call gs_utofu_post_puts(this%ctx_v, int(tid, c_int), &
         int(nteam, c_int), int(size(this%send_pe), c_int), &
         this%send_off64_v, this%send_len64_v, this%rmt_vcq_id_v, &
         this%rmt_stadd_v, this%dst_off_v, this%dst_off_v, this%dst_tag, &
         int(this%parity_v, c_int), int(send_base, c_int64_t), &
         int(this%elem_size, c_int), ierr)
    if (ierr .ne. 0) call neko_error("gs_utofu: vector put failed")
    !$omp barrier
#else
    call neko_error("uTofu support not built")
#endif
  end subroutine gs_nbsend_vec_utofu
 
  subroutine gs_nbrecv_vec_utofu(this, tag, nc)
    class(gs_utofu_t), intent(inout) :: this
    integer, intent(in) :: tag, nc
  end subroutine gs_nbrecv_vec_utofu
 
  subroutine gs_nbwait_vec_utofu(this, u, n, nc, op, strm)
    class(gs_utofu_t), intent(inout) :: this
    integer, intent(in) :: n, nc
    real(kind=rp), dimension(nc*n), intent(inout) :: u
    type(c_ptr), intent(inout) :: strm
    integer :: op
#ifdef HAVE_UTOFU
    integer :: i, j, c, k, src, off, nsrc, nreqs, base
    integer(c_int) :: ierr
    integer, pointer :: sp(:)
    real(kind=rp), pointer :: rbuf(:)
 
    base = this%parity_v * (gs_vec_nc * this%buf_size)
    nreqs = size(this%recv_pe)
    rbuf => this%recv_buf_v
 
    do while (nreqs .gt. 0)
       !$omp master
       ! Spin on the vector MRQ from the master alone (see gs_nbwait_utofu).
       this%ncompleted = 0
       do while (this%ncompleted .eq. 0)
          call gs_utofu_poll_recv(this%ctx_v, int(this%parity_v, c_int), &
               int(size(this%recv_pe), c_int), this%ncompleted, &
               this%recv_indices, ierr)
          if (ierr .ne. 0) &
               call neko_error("gs_utofu: vector mrq poll failed")
       end do
       !$omp end master
       !$omp barrier
 
       do k = 1, this%ncompleted
          i = this%recv_indices(k)
          src = this%recv_pe(i)
          off = this%recv_offset(i)
          nsrc = this%recv_len(i)
          sp => this%recv_dof(src)%array()
          select case (op)
          case (gs_op_add)
             !$omp do
             do j = 1, nsrc
                do c = 1, nc
                   u((c-1)*n + sp(j)) = u((c-1)*n + sp(j)) + &
                        rbuf(base + nc*off + (c-1)*nsrc + j)
                end do
             end do
             !$omp end do
          case (gs_op_mul)
             !$omp do
             do j = 1, nsrc
                do c = 1, nc
                   u((c-1)*n + sp(j)) = u((c-1)*n + sp(j)) * &
                        rbuf(base + nc*off + (c-1)*nsrc + j)
                end do
             end do
             !$omp end do
          case (gs_op_min)
             !$omp do
             do j = 1, nsrc
                do c = 1, nc
                   u((c-1)*n + sp(j)) = min(u((c-1)*n + sp(j)), &
                        rbuf(base + nc*off + (c-1)*nsrc + j))
                end do
             end do
             !$omp end do
          case (gs_op_max)
             !$omp do
             do j = 1, nsrc
                do c = 1, nc
                   u((c-1)*n + sp(j)) = max(u((c-1)*n + sp(j)), &
                        rbuf(base + nc*off + (c-1)*nsrc + j))
                end do
             end do
             !$omp end do
          case default
             call neko_error("Unknown operation in gs_nbwait_vec_utofu")
          end select
       end do
 
       nreqs = nreqs - this%ncompleted
       ! Hold the team until every thread has read ncompleted (above) before
       ! the master re-enters the poll and overwrites recv_indices/ncompleted.
       !$omp barrier
    end do
 
    !$omp master
    ! Flip to the other vector double-buffer half for the next vector round.
    ! Send completions are reaped by the next nbsend_vec's drain.
    this%parity_v = 1 - this%parity_v
    !$omp end master
    !$omp barrier
#else
    call neko_error("uTofu support not built")
#endif
  end subroutine gs_nbwait_vec_utofu
 
end module gs_utofu