@@ -345,33 +345,32 @@ contains
345345 #:endfor
346346 call nvtxEndRange ! Packbuf
347347
348+
348349 #:for rdma_mpi in [False, True]
349350 if (rdma_mpi .eqv. ${' .true. ' if rdma_mpi else ' .false. ' }$) then
350351 #:if rdma_mpi
351352 #:call GPU_HOST_DATA(use_device=' [ib_buff_send, ib_buff_recv]' )
352- call nvtxStartRange("IB-MARKER-SENDRECV-RDMA")
353-
354- call MPI_SENDRECV( &
355- ib_buff_send, buffer_count, MPI_INTEGER, dst_proc, send_tag, &
356- ib_buff_recv, buffer_count, MPI_INTEGER, src_proc, recv_tag, &
357- MPI_COMM_WORLD, MPI_STATUS_IGNORE, ierr)
358-
359- call nvtxEndRange ! RHS-MPI-SENDRECV-(NO)-RDMA
353+
354+ call nvtxStartRange("IB-MARKER-SENDRECV-RDMA")
355+ call MPI_SENDRECV( &
356+ ib_buff_send, buffer_count, MPI_INTEGER, dst_proc, send_tag, &
357+ ib_buff_recv, buffer_count, MPI_INTEGER, src_proc, recv_tag, &
358+ MPI_COMM_WORLD, MPI_STATUS_IGNORE, ierr)
359+ call nvtxEndRange
360360
361361 #:endcall GPU_HOST_DATA
362362 $:GPU_WAIT()
363363 #:else
364364 call nvtxStartRange("IB-MARKER-DEV2HOST")
365365 $:GPU_UPDATE(host=' [ib_buff_send]' )
366366 call nvtxEndRange
367- call nvtxStartRange("IB-MARKER-SENDRECV-NO-RMDA")
368367
368+ call nvtxStartRange("IB-MARKER-SENDRECV-NO-RMDA")
369369 call MPI_SENDRECV( &
370- ib_buff_send, buffer_count, MPI_INTEGER, dst_proc, send_tag, &
371- ib_buff_recv, buffer_count, MPI_INTEGER, src_proc, recv_tag, &
372- MPI_COMM_WORLD, MPI_STATUS_IGNORE, ierr)
373-
374- call nvtxEndRange ! RHS-MPI-SENDRECV-(NO)-RDMA
370+ ib_buff_send, buffer_count, MPI_INTEGER, dst_proc, send_tag, &
371+ ib_buff_recv, buffer_count, MPI_INTEGER, src_proc, recv_tag, &
372+ MPI_COMM_WORLD, MPI_STATUS_IGNORE, ierr)
373+ call nvtxEndRange
375374
376375 call nvtxStartRange("IB-MARKER-HOST2DEV")
377376 $:GPU_UPDATE(device=' [ib_buff_recv]' )
0 commit comments