"reduce", "0 1 0",
"reducescatter", "0.52 1 0.52",
"scan", "1 0.58 0.23",
+ "exscan", "1 0.54 0.25",
"scatterv", "0.52 0 0.52",
"scatter", "1 0.74 0.54",
"computing", "0 1 1",
MPI_Datatype datatype, MPI_Op op, MPI_Comm comm);
void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
MPI_Datatype datatype, MPI_Op op, MPI_Comm comm);
+void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count,
+ MPI_Datatype datatype, MPI_Op op, MPI_Comm comm);
void nary_tree_bcast(void *buf, int count, MPI_Datatype datatype, int root,
MPI_Comm comm, int arity);
xbt_free(tmpbufs);
xbt_free(requests);
}
+
+void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count,
+ MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
+{
+ int system_tag = 888;
+ int rank, size, other, index;
+ MPI_Aint lb = 0, dataext = 0;
+ MPI_Request *requests;
+ void **tmpbufs;
+ int recvbuf_is_empty=1;
+ rank = smpi_comm_rank(comm);
+ size = smpi_comm_size(comm);
+
+ // FIXME: check for errors
+ smpi_datatype_extent(datatype, &lb, &dataext);
+
+ // Send/Recv buffers to/from others;
+ requests = xbt_new(MPI_Request, size - 1);
+ tmpbufs = xbt_new(void *, rank);
+ index = 0;
+ for(other = 0; other < rank; other++) {
+ // FIXME: possibly overkill we we have contiguous/noncontiguous data
+ // mapping...
+ tmpbufs[index] = xbt_malloc(count * dataext);
+ requests[index] =
+ smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
+ comm);
+ index++;
+ }
+ for(other = rank + 1; other < size; other++) {
+ requests[index] =
+ smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
+ index++;
+ }
+ // Wait for completion of all comms.
+ smpi_mpi_startall(size - 1, requests);
+ for(other = 0; other < size - 1; other++) {
+ index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
+ if(index == MPI_UNDEFINED) {
+ break;
+ }
+ if(index < rank) {
+ if(recvbuf_is_empty) smpi_datatype_copy(tmpbufs[index], count, datatype, recvbuf, count, datatype);
+ // #Request is below rank: it's a irecv
+ else smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
+ }
+ }
+ for(index = 0; index < rank; index++) {
+ xbt_free(tmpbufs[index]);
+ }
+ xbt_free(tmpbufs);
+ xbt_free(requests);
+}
return retval;
}
+int PMPI_Exscan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype,
+ MPI_Op op, MPI_Comm comm){
+ int retval;
+
+ smpi_bench_end();
+#ifdef HAVE_TRACING
+ int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1;
+ TRACE_smpi_computing_out(rank);
+ TRACE_smpi_collective_in(rank, -1, __FUNCTION__);
+#endif
+ if (comm == MPI_COMM_NULL) {
+ retval = MPI_ERR_COMM;
+ } else if (datatype == MPI_DATATYPE_NULL) {
+ retval = MPI_ERR_TYPE;
+ } else if (op == MPI_OP_NULL) {
+ retval = MPI_ERR_OP;
+ } else {
+ smpi_mpi_exscan(sendbuf, recvbuf, count, datatype, op, comm);
+ retval = MPI_SUCCESS;
+ }
+#ifdef HAVE_TRACING
+ TRACE_smpi_collective_out(rank, -1, __FUNCTION__);
+ TRACE_smpi_computing_in(rank);
+#endif
+ smpi_bench_begin();
+ return retval;
+}
+
int PMPI_Reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts,
MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
{
NOT_YET_IMPLEMENTED
}
-int PMPI_Exscan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype,
- MPI_Op op, MPI_Comm comm){
- NOT_YET_IMPLEMENTED
-}
-
int PMPI_Comm_set_name(MPI_Comm comm, char* name){
NOT_YET_IMPLEMENTED
}