#include <assert.h>
#include "private.h"
+#include "colls/colls.h"
+
+s_mpi_coll_description_t mpi_coll_allgather_description[] = {
+ {"default",
+ "allgather default collective",
+ smpi_mpi_allgather},
+COLL_ALLGATHERS(COLL_DESCRIPTION, COLL_COMMA),
+ {NULL, NULL, NULL} /* this array must be NULL terminated */
+};
+
+s_mpi_coll_description_t mpi_coll_allreduce_description[] = {
+ {"default",
+ "allreduce default collective",
+ smpi_mpi_allreduce},
+COLL_ALLREDUCES(COLL_DESCRIPTION, COLL_COMMA),
+ {NULL, NULL, NULL} /* this array must be NULL terminated */
+};
+
+s_mpi_coll_description_t mpi_coll_alltoall_description[] = {
+ {"ompi",
+ "Ompi alltoall default collective",
+ smpi_coll_tuned_alltoall_ompi},
+COLL_ALLTOALLS(COLL_DESCRIPTION, COLL_COMMA),
+ {"bruck",
+ "Alltoall Bruck (SG) collective",
+ smpi_coll_tuned_alltoall_bruck},
+ {"basic_linear",
+ "Alltoall basic linear (SG) collective",
+ smpi_coll_tuned_alltoall_basic_linear},
+ {"pairwise",
+ "Alltoall pairwise (SG) collective",
+ smpi_coll_tuned_alltoall_pairwise},
+ {NULL, NULL, NULL} /* this array must be NULL terminated */
+};
+
+s_mpi_coll_description_t mpi_coll_bcast_description[] = {
+ {"default",
+ "allgather default collective",
+ smpi_mpi_bcast},
+COLL_BCASTS(COLL_DESCRIPTION, COLL_COMMA),
+ {NULL, NULL, NULL} /* this array must be NULL terminated */
+};
+
+s_mpi_coll_description_t mpi_coll_reduce_description[] = {
+ {"default",
+ "allgather default collective",
+ smpi_mpi_reduce},
+COLL_REDUCES(COLL_DESCRIPTION, COLL_COMMA),
+ {NULL, NULL, NULL} /* this array must be NULL terminated */
+};
+
+
+
+/** Displays the long description of all registered models, and quit */
+void coll_help(const char *category, s_mpi_coll_description_t * table)
+{
+ int i;
+ printf("Long description of the %s models accepted by this simulator:\n",
+ category);
+ for (i = 0; table[i].name; i++)
+ printf(" %s: %s\n", table[i].name, table[i].description);
+}
+
+int find_coll_description(s_mpi_coll_description_t * table,
+ const char *name)
+{
+ int i;
+ char *name_list = NULL;
+
+ for (i = 0; table[i].name; i++)
+ if (!strcmp(name, table[i].name)) {
+ return i;
+ }
+ name_list = strdup(table[0].name);
+ for (i = 1; table[i].name; i++) {
+ name_list =
+ xbt_realloc(name_list,
+ strlen(name_list) + strlen(table[i].name) + 3);
+ strcat(name_list, ", ");
+ strcat(name_list, table[i].name);
+ }
+ xbt_die("Model '%s' is invalid! Valid models are: %s.", name, name_list);
+ return -1;
+}
+
+
XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_coll, smpi,
"Logging specific to SMPI (coll)");
+int (*mpi_coll_allgather_fun)(void *, int, MPI_Datatype, void*, int, MPI_Datatype, MPI_Comm);
+int (*mpi_coll_allreduce_fun)(void *sbuf, void *rbuf, int rcount, MPI_Datatype dtype, MPI_Op op, MPI_Comm comm);
+int (*mpi_coll_alltoall_fun)(void *, int, MPI_Datatype, void*, int, MPI_Datatype, MPI_Comm);
+int (*mpi_coll_bcast_fun)(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm com);
+int (*mpi_coll_reduce_fun)(void *buf, void *rbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm);
+
struct s_proc_tree {
int PROCTREE_A;
int numChildren;
free_tree(tree);
}
+int smpi_coll_tuned_alltoall_ompi(void *sendbuf, int sendcount,
+ MPI_Datatype sendtype, void *recvbuf,
+ int recvcount, MPI_Datatype recvtype,
+ MPI_Comm comm)
+{
+ int size, sendsize;
+ size = smpi_comm_size(comm);
+ sendsize = smpi_datatype_size(sendtype) * sendcount;
+ if (sendsize < 200 && size > 12) {
+ return
+ smpi_coll_tuned_alltoall_bruck(sendbuf, sendcount, sendtype,
+ recvbuf, recvcount, recvtype,
+ comm);
+ } else if (sendsize < 3000) {
+ return
+ smpi_coll_tuned_alltoall_basic_linear(sendbuf, sendcount,
+ sendtype, recvbuf,
+ recvcount, recvtype, comm);
+ } else {
+ return
+ smpi_coll_tuned_alltoall_pairwise(sendbuf, sendcount, sendtype,
+ recvbuf, recvcount, recvtype,
+ comm);
+ }
+}
+
/**
* Alltoall Bruck
*
* Openmpi calls this routine when the message size sent to each rank < 2000 bytes and size < 12
+ * FIXME: uh, check smpi_pmpi again, but this routine is called for > 12, not
+ * less...
**/
int smpi_coll_tuned_alltoall_bruck(void *sendbuf, int sendcount,
MPI_Datatype sendtype, void *recvbuf,
int system_tag = 777;
int i, rank, size, err, count;
MPI_Aint lb;
- MPI_Aint sendextent = 0;
- MPI_Aint recvextent = 0;
+ MPI_Aint sendext = 0;
+ MPI_Aint recvext = 0;
MPI_Request *requests;
// FIXME: check implementation
rank = smpi_comm_rank(comm);
size = smpi_comm_size(comm);
XBT_DEBUG("<%d> algorithm alltoall_bruck() called.", rank);
- err = smpi_datatype_extent(sendtype, &lb, &sendextent);
- err = smpi_datatype_extent(recvtype, &lb, &recvextent);
+ err = smpi_datatype_extent(sendtype, &lb, &sendext);
+ err = smpi_datatype_extent(recvtype, &lb, &recvext);
/* Local copy from self */
err =
- smpi_datatype_copy(&((char *) sendbuf)[rank * sendextent], sendcount,
- sendtype, &((char *) recvbuf)[rank * recvextent],
+ smpi_datatype_copy((char *)sendbuf + rank * sendcount * sendext,
+ sendcount, sendtype,
+ (char *)recvbuf + rank * recvcount * recvext,
recvcount, recvtype);
if (err == MPI_SUCCESS && size > 1) {
/* Initiate all send/recv to/from others. */
continue;
}
requests[count] =
- smpi_irecv_init(&((char *) recvbuf)[i * recvextent], recvcount,
+ smpi_irecv_init((char *)recvbuf + i * recvcount * recvext, recvcount,
recvtype, i, system_tag, comm);
count++;
}
continue;
}
requests[count] =
- smpi_isend_init(&((char *) sendbuf)[i * sendextent], sendcount,
+ smpi_isend_init((char *)sendbuf + i * sendcount * sendext, sendcount,
sendtype, i, system_tag, comm);
count++;
}
}
/**
- * Alltoall basic_linear
+ * Alltoall basic_linear (STARMPI:alltoall-simple)
**/
int smpi_coll_tuned_alltoall_basic_linear(void *sendbuf, int sendcount,
MPI_Datatype sendtype,
err = smpi_datatype_extent(sendtype, &lb, &sendext);
err = smpi_datatype_extent(recvtype, &lb, &recvext);
/* simple optimization */
- err = smpi_datatype_copy(sendbuf + rank * sendcount * sendext, sendcount,
- sendtype,
- recvbuf + rank * recvcount * recvext, recvcount,
- recvtype);
+ err = smpi_datatype_copy((char *)sendbuf + rank * sendcount * sendext,
+ sendcount, sendtype,
+ (char *)recvbuf + rank * recvcount * recvext,
+ recvcount, recvtype);
if (err == MPI_SUCCESS && size > 1) {
/* Initiate all send/recv to/from others. */
requests = xbt_new(MPI_Request, 2 * (size - 1));
count = 0;
for (i = (rank + 1) % size; i != rank; i = (i + 1) % size) {
requests[count] =
- smpi_irecv_init(recvbuf + i * recvcount * recvext, recvcount,
+ smpi_irecv_init((char *)recvbuf + i * recvcount * recvext, recvcount,
recvtype, i, system_tag, comm);
count++;
}
*/
for (i = (rank + size - 1) % size; i != rank; i = (i + size - 1) % size) {
requests[count] =
- smpi_isend_init(sendbuf + i * sendcount * sendext, sendcount,
+ smpi_isend_init((char *)sendbuf + i * sendcount * sendext, sendcount,
sendtype, i, system_tag, comm);
count++;
}
err = smpi_datatype_extent(recvtype, &lb, &recvext);
/* Local copy from self */
err =
- smpi_datatype_copy(sendbuf + senddisps[rank] * sendext, sendcounts[rank],
- sendtype,
- recvbuf + recvdisps[rank] * recvext, recvcounts[rank],
- recvtype);
+ smpi_datatype_copy((char *)sendbuf + senddisps[rank] * sendext,
+ sendcounts[rank], sendtype,
+ (char *)recvbuf + recvdisps[rank] * recvext,
+ recvcounts[rank], recvtype);
if (err == MPI_SUCCESS && size > 1) {
/* Initiate all send/recv to/from others. */
requests = xbt_new(MPI_Request, 2 * (size - 1));
continue;
}
requests[count] =
- smpi_irecv_init(recvbuf + recvdisps[i] * recvext, recvcounts[i],
- recvtype, i, system_tag, comm);
+ smpi_irecv_init((char *)recvbuf + recvdisps[i] * recvext,
+ recvcounts[i], recvtype, i, system_tag, comm);
count++;
}
/* Now create all sends */
continue;
}
requests[count] =
- smpi_isend_init(sendbuf + senddisps[i] * sendext, sendcounts[i],
- sendtype, i, system_tag, comm);
+ smpi_isend_init((char *)sendbuf + senddisps[i] * sendext,
+ sendcounts[i], sendtype, i, system_tag, comm);
count++;
}
/* Wait for them all. */
