#include <assert.h>
#include "private.h"
+#include "colls/colls.h"
+
+s_mpi_coll_description_t mpi_coll_alltoall_description[] = {
+ {"ompi",
+ "Ompi alltoall default collective",
+ smpi_coll_tuned_alltoall_ompi},
+
+ {"2dmesh",
+ "Alltoall 2dmesh collective",
+ smpi_coll_tuned_alltoall_2dmesh},
+ {"3dmesh",
+ "Alltoall 3dmesh collective",
+ smpi_coll_tuned_alltoall_3dmesh},
+ /*{"bruck",
+ "Alltoall Bruck collective",
+ smpi_coll_tuned_alltoall_bruck},*/
+ {"pair",
+ "Alltoall pair collective",
+ smpi_coll_tuned_alltoall_pair},
+ {"pair_light_barrier",
+ "Alltoall pair_light_barrier collective",
+ smpi_coll_tuned_alltoall_pair_light_barrier},
+ {"pair_mpi_barrier",
+ "Alltoall pair_mpi_barrier collective",
+ smpi_coll_tuned_alltoall_pair_mpi_barrier},
+ {"rdb",
+ "Alltoall rdb collective",
+ smpi_coll_tuned_alltoall_rdb},
+ {"ring",
+ "Alltoall ring collective",
+ smpi_coll_tuned_alltoall_ring},
+ {"ring_light_barrier",
+ "Alltoall ring_light_barrier collective",
+ smpi_coll_tuned_alltoall_ring_light_barrier},
+ {"ring_light_barrier",
+ "Alltoall ring_light_barrier collective",
+ smpi_coll_tuned_alltoall_ring_light_barrier},
+ {"ring_mpi_barrier",
+ "Alltoall ring_mpi_barrier collective",
+ smpi_coll_tuned_alltoall_ring_mpi_barrier},
+ {"ring_one_barrier",
+ "Alltoall ring_one_barrier collective",
+ smpi_coll_tuned_alltoall_ring_one_barrier},
+ {"simple",
+ "Alltoall simple collective",
+ smpi_coll_tuned_alltoall_simple},
+
+ {"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_allgather_description[] = {
+ {"default",
+ "allgather default collective",
+ smpi_mpi_gather},
+
+ {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_alltoall_fun)(void *, int, MPI_Datatype, void*, int, MPI_Datatype, MPI_Comm);
+int (*mpi_coll_allgather_fun)(void *, int, MPI_Datatype, void*, int, MPI_Datatype, MPI_Comm);
+
struct s_proc_tree {
int PROCTREE_A;
int numChildren;
/**
* Build the tree depending on a process rank (index) and the group size (extent)
- * @param index the rank of the calling process
- * @param extent the total number of processes
+ * @param root the rank of the tree root
+ * @param rank the rank of the calling process
+ * @param size the total number of processes
**/
-static void build_tree(int index, int extent, proc_tree_t * tree)
+static void build_tree(int root, int rank, int size, proc_tree_t * tree)
{
- int places = (*tree)->PROCTREE_A * index;
- int i, ch, pr;
-
- (*tree)->me = index;
- (*tree)->root = 0;
- for (i = 1; i <= (*tree)->PROCTREE_A; i++) {
- ++places;
- ch = (*tree)->PROCTREE_A * index + i + (*tree)->root;
- ch %= extent;
- if (places < extent) {
- (*tree)->child[i - 1] = ch;
- (*tree)->numChildren++;
- }
+ int index = (rank - root + size) % size;
+ int firstChildIdx = index * (*tree)->PROCTREE_A + 1;
+ int i;
+
+ (*tree)->me = rank;
+ (*tree)->root = root;
+
+ for (i = 0; i < (*tree)->PROCTREE_A && firstChildIdx + i < size; i++) {
+ (*tree)->child[i] = (firstChildIdx + i + root) % size;
+ (*tree)->numChildren++;
}
- if (index == (*tree)->root) {
+ if (rank == root) {
(*tree)->isRoot = 1;
} else {
(*tree)->isRoot = 0;
- pr = (index - 1) / (*tree)->PROCTREE_A;
- (*tree)->parent = pr;
+ (*tree)->parent = (((index - 1) / (*tree)->PROCTREE_A) + root) % size;
}
}
* bcast
**/
static void tree_bcast(void *buf, int count, MPI_Datatype datatype,
- int root, MPI_Comm comm, proc_tree_t tree)
+ MPI_Comm comm, proc_tree_t tree)
{
int system_tag = 999; // used negative int but smpi_create_request() declares this illegal (to be checked)
int rank, i;
* anti-bcast
**/
static void tree_antibcast(void *buf, int count, MPI_Datatype datatype,
- int root, MPI_Comm comm, proc_tree_t tree)
+ MPI_Comm comm, proc_tree_t tree)
{
int system_tag = 999; // used negative int but smpi_create_request() declares this illegal (to be checked)
int rank, i;
rank = smpi_comm_rank(comm);
size = smpi_comm_size(comm);
- build_tree(rank, size, &tree);
- tree_bcast(buf, count, datatype, root, comm, tree);
+ build_tree(root, rank, size, &tree);
+ tree_bcast(buf, count, datatype, comm, tree);
free_tree(tree);
}
rank = smpi_comm_rank(comm);
size = smpi_comm_size(comm);
- build_tree(rank, size, &tree);
- tree_antibcast(&dummy, 1, MPI_CHAR, 0, comm, tree);
- tree_bcast(&dummy, 1, MPI_CHAR, 0, comm, tree);
+ build_tree(0, rank, size, &tree);
+ tree_antibcast(&dummy, 1, MPI_CHAR, comm, tree);
+ tree_bcast(&dummy, 1, MPI_CHAR, comm, tree);
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,
{
int system_tag = 888;
int i, rank, size, err, count;
- MPI_Aint lb;
- MPI_Aint sendinc = 0;
- MPI_Aint recvinc = 0;
+ MPI_Aint lb = 0, sendext = 0, recvext = 0;
MPI_Request *requests;
/* Initialize. */
rank = smpi_comm_rank(comm);
size = smpi_comm_size(comm);
XBT_DEBUG("<%d> algorithm alltoall_basic_linear() called.", rank);
- err = smpi_datatype_extent(sendtype, &lb, &sendinc);
- err = smpi_datatype_extent(recvtype, &lb, &recvinc);
- sendinc *= sendcount;
- recvinc *= recvcount;
+ err = smpi_datatype_extent(sendtype, &lb, &sendext);
+ err = smpi_datatype_extent(recvtype, &lb, &recvext);
/* simple optimization */
- err =
- smpi_datatype_copy(&((char *) sendbuf)[rank * sendinc], sendcount,
- sendtype, &((char *) recvbuf)[rank * recvinc],
- 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(&((char *) recvbuf)[i * recvinc], recvcount,
+ smpi_irecv_init((char *)recvbuf + i * recvcount * recvext, recvcount,
recvtype, i, system_tag, comm);
count++;
}
* when messages actually arrive in the order in which they were posted.
* TODO: check the previous assertion
*/
- for (i = (rank + size - 1) % size; i != rank;
- i = (i + size - 1) % size) {
+ for (i = (rank + size - 1) % size; i != rank; i = (i + size - 1) % size) {
requests[count] =
- smpi_isend_init(&((char *) sendbuf)[i * sendinc], sendcount,
+ smpi_isend_init((char *)sendbuf + i * sendcount * sendext, sendcount,
sendtype, i, system_tag, comm);
count++;
}
{
int system_tag = 889;
int i, rank, size, err, count;
- MPI_Aint lb;
- MPI_Aint sendextent = 0;
- MPI_Aint recvextent = 0;
+ MPI_Aint lb = 0, sendext = 0, recvext = 0;
MPI_Request *requests;
/* Initialize. */
rank = smpi_comm_rank(comm);
size = smpi_comm_size(comm);
XBT_DEBUG("<%d> algorithm basic_alltoallv() 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)[senddisps[rank] * sendextent],
+ smpi_datatype_copy((char *)sendbuf + senddisps[rank] * sendext,
sendcounts[rank], sendtype,
- &((char *) recvbuf)[recvdisps[rank] * recvextent],
+ (char *)recvbuf + recvdisps[rank] * recvext,
recvcounts[rank], recvtype);
if (err == MPI_SUCCESS && size > 1) {
/* Initiate all send/recv to/from others. */
continue;
}
requests[count] =
- smpi_irecv_init(&((char *) recvbuf)[recvdisps[i] * recvextent],
+ smpi_irecv_init((char *)recvbuf + recvdisps[i] * recvext,
recvcounts[i], recvtype, i, system_tag, comm);
count++;
}
continue;
}
requests[count] =
- smpi_isend_init(&((char *) sendbuf)[senddisps[i] * sendextent],
+ smpi_isend_init((char *)sendbuf + senddisps[i] * sendext,
sendcounts[i], sendtype, i, system_tag, comm);
count++;
}
