Added interface to MPI Ibarrier collective (#512)

* Added interface to MPI Ibarrier collective

Added ability to call MPI Ibarrier. Useful for allowing ranks to continue listening for messages or otherwise processing while waiting at a barrier.

* Creates a test for Ibarrier

Tests to ensure the Ibarrier capability works. Each rank >0 hits the Ibarrier and then cycles, waiting for messages from  until they've all processed the messages. Then 0 can hit the barrier. Represents the usage where different ranks can finish their work at the barrier, but still need to potentially listen for requests from other ranks.

* Update collective.md

Added reference to Ibarrier

Author: mikegros

docs/src/collective.md

@@ -4,6 +4,7 @@
 
 ```@docs
 MPI.Barrier
+MPI.Ibarrier
 ```
 
 ## Broadcast

src/collective.jl

@@ -16,6 +16,26 @@ function Barrier(comm::Comm)
     return nothing
 end
 
+"""
+    Ibarrier(comm::Comm)
+
+Blocks until `comm` is synchronized.
+
+If `comm` is an intracommunicator, then it blocks until all members of the group have called it.
+
+If `comm` is an intercommunicator, then it blocks until all members of the other group have called it.
+
+# External links
+$(_doc_external("MPI_Ibarrier"))
+"""
+function Ibarrier(comm::Comm)
+    req = Request()
+    # int MPI_Ibarrier(MPI_Comm comm, MPI_Req req)
+    @mpichk ccall((:MPI_Ibarrier, libmpi), Cint, (MPI_Comm, Ptr{MPI_Request}), comm, req)
+    return req
+end
+
+
 """
     Bcast!(buf, root::Integer, comm::Comm)
 

test/test_Ibarrier.jl

@@ -0,0 +1,74 @@
+using Test
+using MPI
+
+function check_for_query(comm)
+    is_message, status  = MPI.Iprobe(MPI.MPI_ANY_SOURCE, MPI.MPI_ANY_TAG, comm)
+    if is_message
+        recv_id = status.source
+        tag_ind = status.tag
+        return true, recv_id, tag_ind
+    else
+        return false, -1, -1
+    end
+end
+
+MPI.Init()
+comm = MPI.COMM_WORLD
+myrank = MPI.Comm_rank(comm)
+mysize = MPI.Comm_size(comm)
+
+# First rank will send, one-at-a-time, the rank + 4 to each other rank. 
+#     They will then sum these and test if the sum is correct. Each rank > 0
+#     will be waiting at the Ibarrier for messages from 0 in order to test
+#     that the Ibarrier is working properly. 
+# They then communicate back to 0 when they have received all messages to
+#     to allow it to reach the barrier.
+#
+# This is a contrived example but does test to ensure Ibarrier works
+
+if myrank == 0
+  for reps in 1:10
+    for ii in 1:(mysize-1)
+      smsg = MPI.Send(ii + 4, ii, ii, comm)
+    end
+  end
+  for ii in 1:(mysize-1)
+    dummy  = [0]
+    rmsg   = MPI.Recv!(dummy, ii, ii, comm)
+  end
+end
+    
+
+
+all_done = false
+localsum = 0 
+msg_num = 0
+
+barrier_req = MPI.Ibarrier(comm)
+
+
+all_done, barrier_status = MPI.Test!(barrier_req)
+
+while !all_done
+  global all_done
+  global msg_num
+  global myrank
+  is_request, recv_id, tag_ind = check_for_query(comm)
+  if is_request 
+    dummy     = [0]
+    rmsg      = MPI.Recv!(dummy, recv_id, tag_ind, comm)
+    msg_num  += 1
+    global localsum += dummy[1]
+  end # is_request
+  if msg_num == 10
+    smsg = MPI.Send(tag_ind, 0, myrank, comm)
+  end
+  all_done, barrier_status = MPI.Test!(barrier_req)
+end # !all_done 
+
+if myrank > 0
+  @test localsum == 10 * (myrank + 4)
+end
+
+MPI.Finalize()
+@test MPI.Finalized()