gs_device_shmem.F90

Go to the documentation of this file.

! Copyright (c) 2020-2024, 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_device_shmem
  use num_types, only : rp, c_rp
  use gs_comm, only : gs_comm_t, gs_vec_nc
  use stack, only : stack_i4_t
  use htable, only : htable_i4_t
  use device
  use comm, only : pe_size, pe_rank, neko_comm
  use mpi_f08, only : mpi_allreduce, mpi_alltoall, mpi_integer, mpi_max
  use utils, only : neko_error
  use, intrinsic :: iso_c_binding, only : c_sizeof, c_int32_t, c_int64_t, &
       c_ptr, c_null_ptr, c_size_t, c_associated
  implicit none
  private
 
  type, private :: gs_device_shmem_buf_t
     integer, allocatable :: ndofs(:)
     integer, allocatable :: offset(:)
     integer, allocatable :: remote_offset(:)
     integer :: total
     type(c_ptr) :: buf_d = c_null_ptr 
     type(c_ptr) :: buf_v_d = c_null_ptr 
     type(c_ptr) :: dof_d = c_null_ptr 
   contains
     procedure, pass(this) :: init => gs_device_shmem_buf_init
     procedure, pass(this) :: free => gs_device_shmem_buf_free
  end type gs_device_shmem_buf_t
 
  type, public, extends(gs_comm_t) :: gs_device_shmem_t
     type(gs_device_shmem_buf_t) :: send_buf
     type(gs_device_shmem_buf_t) :: recv_buf
     type(c_ptr), allocatable :: stream(:)
     type(c_ptr), allocatable :: event(:)
     integer :: iter = 0
     type(c_ptr) :: done_sig_d = c_null_ptr
     type(c_ptr) :: ready_sig_d = c_null_ptr
   contains
     procedure, pass(this) :: init => gs_device_shmem_init
     procedure, pass(this) :: free => gs_device_shmem_free
     procedure, pass(this) :: nbsend => gs_device_shmem_nbsend
     procedure, pass(this) :: nbrecv => gs_device_shmem_nbrecv
     procedure, pass(this) :: nbwait => gs_device_shmem_nbwait
     procedure, pass(this) :: nbsend_vec => gs_device_shmem_nbsend_vec
     procedure, pass(this) :: nbrecv_vec => gs_device_shmem_nbrecv_vec
     procedure, pass(this) :: nbwait_vec => gs_device_shmem_nbwait_vec
  end type gs_device_shmem_t
 
 
#if defined (HAVE_CUDA) && defined(HAVE_NVSHMEM)
 
  interface
     subroutine cudamalloc_nvshmem(ptr, size) &
          bind(c, name = 'cudamalloc_nvshmem')
       use, intrinsic :: iso_c_binding
       implicit none
       type(c_ptr) :: ptr
       integer(c_size_t), value :: size
     end subroutine cudamalloc_nvshmem
  end interface
 
  interface
     subroutine cudafree_nvshmem(ptr) &
          bind(c, name = 'cudafree_nvshmem')
       use, intrinsic :: iso_c_binding
       implicit none
       type(c_ptr) :: ptr
     end subroutine cudafree_nvshmem
  end interface
 
  interface
     subroutine cuda_gs_pack_and_push(u_d, buf_d, dof_d, offset, n, stream, &
          dest_rank, rbuf_d, roffset, iter, done_d, ready_d, mype) &
          bind(c, name = 'cuda_gs_pack_and_push')
       use, intrinsic :: iso_c_binding
       implicit none
       integer(c_int), value :: n, offset, dest_rank, roffset, iter, mype
       type(c_ptr), value :: u_d, buf_d, dof_d, stream, rbuf_d, done_d, &
            ready_d
     end subroutine cuda_gs_pack_and_push
  end interface
 
  interface
     subroutine cuda_gs_pack_and_push_wait(stream, iter, done_d, src_rank) &
          bind(c, name = 'cuda_gs_pack_and_push_wait')
       use, intrinsic :: iso_c_binding
       implicit none
       integer(c_int), value :: iter, src_rank
       type(c_ptr), value :: stream, done_d
     end subroutine cuda_gs_pack_and_push_wait
  end interface
 
  interface
     subroutine cuda_gs_post_ready(stream, iter, ready_d, mype, src_rank) &
          bind(c, name = 'cuda_gs_post_ready')
       use, intrinsic :: iso_c_binding
       implicit none
       integer(c_int), value :: iter, mype, src_rank
       type(c_ptr), value :: stream, ready_d
     end subroutine cuda_gs_post_ready
  end interface
 
  interface
     subroutine cuda_gs_unpack(u_d, op, buf_d, dof_d, offset, n, stream) &
          bind(c, name = 'cuda_gs_unpack')
       use, intrinsic :: iso_c_binding
       implicit none
       integer(c_int), value :: op, offset, n
       type(c_ptr), value :: u_d, buf_d, dof_d, stream
     end subroutine cuda_gs_unpack
  end interface
 
  interface
     subroutine cuda_gs_pack_and_push_vec(u_d, buf_d, dof_d, offset, n, nc, &
          ns, stream, dest_rank, rbuf_d, roffset, iter, done_d, ready_d, &
          mype) bind(c, name = 'cuda_gs_pack_and_push_vec')
       use, intrinsic :: iso_c_binding
       implicit none
       integer(c_int), value :: n, nc, ns, offset, dest_rank, roffset, iter
       integer(c_int), value :: mype
       type(c_ptr), value :: u_d, buf_d, dof_d, stream, rbuf_d, done_d, &
            ready_d
     end subroutine cuda_gs_pack_and_push_vec
  end interface
 
  interface
     subroutine cuda_gs_unpack_vec(u_d, op, buf_d, dof_d, offset, n, nc, ns, &
          stream) bind(c, name = 'cuda_gs_unpack_vec')
       use, intrinsic :: iso_c_binding
       implicit none
       integer(c_int), value :: op, offset, n, nc, ns
       type(c_ptr), value :: u_d, buf_d, dof_d, stream
     end subroutine cuda_gs_unpack_vec
  end interface
#endif
 
contains
 
  subroutine gs_device_shmem_buf_init(this, pe_order, dof_stack, mark_dupes)
    class(gs_device_shmem_buf_t), intent(inout) :: this
    integer, allocatable, intent(inout) :: pe_order(:)
    type(stack_i4_t), allocatable, intent(inout) :: dof_stack(:)
    logical, intent(in) :: mark_dupes
    integer, allocatable :: dofs(:)
    integer :: i, j, total, max_total
    integer(c_size_t) :: sz
    type(htable_i4_t) :: doftable
    integer :: dupe, marked, k
    real(c_rp) :: rp_dummy
    integer(c_int32_t) :: i4_dummy
 
    allocate(this%ndofs(size(pe_order)))
    allocate(this%offset(size(pe_order)))
    allocate(this%remote_offset(size(pe_order)))
 
    do i = 1, size(pe_order)
       this%remote_offset(i) = -1
    end do
 
    total = 0
    do i = 1, size(pe_order)
       this%ndofs(i) = dof_stack(pe_order(i))%size()
       this%offset(i) = total
       total = total + this%ndofs(i)
    end do
 
    call mpi_allreduce(total, max_total, 1, mpi_integer, mpi_max, neko_comm)
 
    this%total = total
 
    sz = c_sizeof(rp_dummy) * max_total
#ifdef HAVE_NVSHMEM
    call cudamalloc_nvshmem(this%buf_d, sz)
    ! Fused vector symmetric buffer, sized for up to GS_VEC_NC components.
    sz = c_sizeof(rp_dummy) * gs_vec_nc * max_total
    call cudamalloc_nvshmem(this%buf_v_d, sz)
#endif
 
    sz = c_sizeof(i4_dummy) * total
    call device_alloc(this%dof_d, sz)
 
    if (mark_dupes) call doftable%init(2*total)
    allocate(dofs(total))
 
    ! Copy from dof_stack into dofs, optionally marking duplicates with doftable
    marked = 0
    do i = 1, size(pe_order)
       ! %array() breaks on cray
       select type (arr => dof_stack(pe_order(i))%data)
       type is (integer)
          do j = 1, this%ndofs(i)
             k = this%offset(i) + j
             if (mark_dupes) then
                if (doftable%get(arr(j), dupe) .eq. 0) then
                   if (dofs(dupe) .gt. 0) then
                      dofs(dupe) = -dofs(dupe)
                      marked = marked + 1
                   end if
                   dofs(k) = -arr(j)
                   marked = marked + 1
                else
                   call doftable%set(arr(j), k)
                   dofs(k) = arr(j)
                end if
             else
                dofs(k) = arr(j)
             end if
          end do
       end select
    end do
 
    call device_memcpy(dofs, this%dof_d, total, host_to_device, sync = .true.)
 
    deallocate(dofs)
    call doftable%free()
 
  end subroutine gs_device_shmem_buf_init
 
  subroutine gs_device_shmem_buf_free(this)
    class(gs_device_shmem_buf_t), intent(inout) :: this
 
 
    if (allocated(this%ndofs)) deallocate(this%ndofs)
    if (allocated(this%offset)) deallocate(this%offset)
 
#ifdef HAVE_NVSHMEM
    if (c_associated(this%buf_d)) call cudafree_nvshmem(this%buf_d)
    if (c_associated(this%buf_v_d)) call cudafree_nvshmem(this%buf_v_d)
#endif
    if (c_associated(this%dof_d)) call device_free(this%dof_d)
 
  end subroutine gs_device_shmem_buf_free
 
  subroutine gs_device_shmem_init(this, send_pe, recv_pe)
    class(gs_device_shmem_t), intent(inout) :: this
    type(stack_i4_t), intent(inout) :: send_pe
    type(stack_i4_t), intent(inout) :: recv_pe
    integer :: i, nstrm, ierr
    integer, allocatable :: local_offsets(:), remote_offsets(:)
    integer(c_size_t) :: sz
    integer(c_int64_t) :: i64_dummy
 
    call this%init_order(send_pe, recv_pe)
 
    call this%send_buf%init(this%send_pe, this%send_dof, .false.)
    call this%recv_buf%init(this%recv_pe, this%recv_dof, .true.)
 
    ! Exchange, for every send peer, the offset in the receiver's recv buffer
    ! where our slab must land. A single Alltoall keeps the exchange
    ! deadlock-free for arbitrary, non-uniform peer sets (same rationale as
    ! the host OpenSHMEM backend); the lazy pairwise exchange this replaces
    ! indexed send_pe/recv_pe with the same loop index and required uniform,
    ! index-aligned peer lists on every rank.
    allocate(local_offsets(0:pe_size - 1))
    allocate(remote_offsets(0:pe_size - 1))
    local_offsets = -1
    do i = 1, size(this%recv_pe)
       local_offsets(this%recv_pe(i)) = this%recv_buf%offset(i)
    end do
    call mpi_alltoall(local_offsets, 1, mpi_integer, &
         remote_offsets, 1, mpi_integer, neko_comm, ierr)
    do i = 1, size(this%send_pe)
       this%send_buf%remote_offset(i) = remote_offsets(this%send_pe(i))
    end do
    deallocate(local_offsets)
    deallocate(remote_offsets)
 
#if defined(HAVE_HIP) || defined(HAVE_CUDA)
    ! Create a set of non-blocking streams. The per-peer streams, events and
    ! notify signals are indexed over both send_pe (pack-and-push) and
    ! recv_pe (unpack, sync), so size them for the larger of the two peer
    ! lists.
    nstrm = max(size(this%send_pe), size(this%recv_pe))
    allocate(this%stream(nstrm))
    do i = 1, nstrm
       call device_stream_create_with_priority(this%stream(i), 1, &
            strm_high_prio)
    end do
 
    allocate(this%event(nstrm))
    do i = 1, nstrm
       call device_event_create(this%event(i), 2)
    end do
 
#ifdef HAVE_NVSHMEM
    ! Rank-indexed symmetric signal arrays; zero-initialised by
    ! cudamalloc_nvshmem, so the first round's CMP_GE waits on iter-1 = 0
    ! pass immediately.
    sz = c_sizeof(i64_dummy) * pe_size
    call cudamalloc_nvshmem(this%done_sig_d, sz)
    call cudamalloc_nvshmem(this%ready_sig_d, sz)
#endif
#endif
 
    this%iter = 0
    this%vec_supported = .true.
 
  end subroutine gs_device_shmem_init
 
  subroutine gs_device_shmem_free(this)
    class(gs_device_shmem_t), intent(inout) :: this
    integer :: i
 
    call this%send_buf%free()
    call this%recv_buf%free()
 
#ifdef HAVE_NVSHMEM
    ! Collective frees; every rank allocated both arrays exactly once.
    if (c_associated(this%done_sig_d)) then
       call cudafree_nvshmem(this%done_sig_d)
       this%done_sig_d = c_null_ptr
    end if
    if (c_associated(this%ready_sig_d)) then
       call cudafree_nvshmem(this%ready_sig_d)
       this%ready_sig_d = c_null_ptr
    end if
#endif
 
    call this%free_order()
    call this%free_dofs()
 
#if defined(HAVE_HIP) || defined(HAVE_CUDA)
    if (allocated(this%stream)) then
       do i = 1, size(this%stream)
          call device_stream_destroy(this%stream(i))
       end do
       deallocate(this%stream)
    end if
#endif
 
  end subroutine gs_device_shmem_free
 
  subroutine gs_device_shmem_nbsend(this, u, n, tag, deps, strm)
    class(gs_device_shmem_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
    integer :: i
    type(c_ptr) :: u_d
 
    u_d = device_get_ptr(u)
 
    ! Order each per-peer stream after the gather (deps); the pack-and-push
    ! kernels in nbwait are stream-ordered behind this wait. The historic
    ! host-side device_sync here papered over a send-buffer reuse race: the
    ! pack-and-push kernel used to overwrite the send buffer BEFORE its
    ! ready handshake, while the previous round's non-blocking put could
    ! still be draining it. The kernel now performs the handshake before
    ! packing (see gs_nvshmem_kernels.h), which closes that race properly,
    ! so no host synchronisation is needed.
    do i = 1, size(this%send_pe)
       call device_stream_wait_event(this%stream(i), deps, 0)
    end do
 
    ! We do the rest in the "wait" routine below
 
  end subroutine gs_device_shmem_nbsend
 
  subroutine gs_device_shmem_nbrecv(this, tag)
    class(gs_device_shmem_t), intent(inout) :: this
    integer, intent(in) :: tag
    integer :: i
 
    ! We do everything in the "wait" routine below
 
  end subroutine gs_device_shmem_nbrecv
 
  subroutine gs_device_shmem_nbwait(this, u, n, op, strm)
    class(gs_device_shmem_t), intent(inout) :: this
    integer, intent(in) :: n
    real(kind=rp), dimension(n), intent(inout) :: u
    type(c_ptr), intent(inout) :: strm
    integer :: op, done_req, i
    type(c_ptr) :: u_d
 
    u_d = device_get_ptr(u)
#ifdef HAVE_NVSHMEM
    ! One round counter per gs op; the rank-indexed signal slots make the
    ! exchange independent of peer counts and peer-list ordering.
    this%iter = this%iter + 1
 
    ! Push our slab to every send peer. The kernel waits for the peer's
    ! ready signal (previous round consumed) before overwriting the send
    ! slab, then puts with a done signal.
    do i = 1, size(this%send_pe)
       call cuda_gs_pack_and_push(u_d, &
            this%send_buf%buf_d, &
            this%send_buf%dof_d, &
            this%send_buf%offset(i), &
            this%send_buf%ndofs(i), &
            this%stream(i), &
            this%send_pe(i), &
            this%recv_buf%buf_d, &
            this%send_buf%remote_offset(i), &
            this%iter, this%done_sig_d, this%ready_sig_d, pe_rank)
    end do
 
    ! For every recv peer: wait until its slab has landed, reduce it into u,
    ! then post our ready signal so the peer may start its next round.
    do done_req = 1, size(this%recv_pe)
       call cuda_gs_pack_and_push_wait(this%stream(done_req), this%iter, &
            this%done_sig_d, this%recv_pe(done_req))
       call cuda_gs_unpack(u_d, op, &
            this%recv_buf%buf_d, &
            this%recv_buf%dof_d, &
            this%recv_buf%offset(done_req), &
            this%recv_buf%ndofs(done_req), &
            this%stream(done_req))
       call cuda_gs_post_ready(this%stream(done_req), this%iter, &
            this%ready_sig_d, pe_rank, this%recv_pe(done_req))
       call device_event_record(this%event(done_req), this%stream(done_req))
    end do
 
    ! Sync non-blocking streams
    do done_req = 1, size(this%recv_pe)
       call device_stream_wait_event(strm, &
            this%event(done_req), 0)
    end do
#endif
  end subroutine gs_device_shmem_nbwait
 
  subroutine gs_device_shmem_nbsend_vec(this, u, n, nc, tag, deps, strm)
    class(gs_device_shmem_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
    integer :: i
 
    ! Order each per-peer stream after the gather/copies (deps); no host
    ! synchronisation needed, see gs_device_shmem_nbsend.
    do i = 1, size(this%send_pe)
       call device_stream_wait_event(this%stream(i), deps, 0)
    end do
 
  end subroutine gs_device_shmem_nbsend_vec
 
  subroutine gs_device_shmem_nbrecv_vec(this, tag, nc)
    class(gs_device_shmem_t), intent(inout) :: this
    integer, intent(in) :: tag, nc
  end subroutine gs_device_shmem_nbrecv_vec
 
  subroutine gs_device_shmem_nbwait_vec(this, u, n, nc, op, strm)
    class(gs_device_shmem_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, done_req, i
    type(c_ptr) :: u_d
 
    u_d = device_get_ptr(u)
#ifdef HAVE_NVSHMEM
    ! One round counter per gs op (shared with the scalar path; all ranks
    ! execute the same op sequence, so counters stay in lockstep).
    this%iter = this%iter + 1
 
    ! Push our nc-component slab to every send peer.
    do i = 1, size(this%send_pe)
       call cuda_gs_pack_and_push_vec(u_d, &
            this%send_buf%buf_v_d, &
            this%send_buf%dof_d, &
            this%send_buf%offset(i), &
            this%send_buf%ndofs(i), &
            nc, n, &
            this%stream(i), &
            this%send_pe(i), &
            this%recv_buf%buf_v_d, &
            this%send_buf%remote_offset(i), &
            this%iter, this%done_sig_d, this%ready_sig_d, pe_rank)
    end do
 
    ! For every recv peer: wait until its slab has landed, reduce it into u,
    ! then post our ready signal so the peer may start its next round.
    do done_req = 1, size(this%recv_pe)
       call cuda_gs_pack_and_push_wait(this%stream(done_req), this%iter, &
            this%done_sig_d, this%recv_pe(done_req))
       call cuda_gs_unpack_vec(u_d, op, &
            this%recv_buf%buf_v_d, &
            this%recv_buf%dof_d, &
            this%recv_buf%offset(done_req), &
            this%recv_buf%ndofs(done_req), &
            nc, n, &
            this%stream(done_req))
       call cuda_gs_post_ready(this%stream(done_req), this%iter, &
            this%ready_sig_d, pe_rank, this%recv_pe(done_req))
       call device_event_record(this%event(done_req), this%stream(done_req))
    end do
 
    ! Sync non-blocking streams
    do done_req = 1, size(this%recv_pe)
       call device_stream_wait_event(strm, &
            this%event(done_req), 0)
    end do
#endif
  end subroutine gs_device_shmem_nbwait_vec
 
end module gs_device_shmem