X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/4f377cfbcae89e70585c675eb978c7c8a33dd773..db17323a11a3089f492c2cae1be7913673689d5c:/src/smpi/smpi_pmpi.c diff --git a/src/smpi/smpi_pmpi.c b/src/smpi/smpi_pmpi.c index d6938f5f7f..7859738e24 100644 --- a/src/smpi/smpi_pmpi.c +++ b/src/smpi/smpi_pmpi.c @@ -50,6 +50,28 @@ int PMPI_Finalize(void) return MPI_SUCCESS; } +int PMPI_Finalized(int* flag) +{ + *flag=smpi_process_finalized(); + return MPI_SUCCESS; +} + +int PMPI_Get_version (int *version,int *subversion){ + *version = MPI_VERSION; + *subversion= MPI_SUBVERSION; + return MPI_SUCCESS; +} + +int PMPI_Get_library_version (char *version,int *len){ + int retval = MPI_SUCCESS; + smpi_bench_end(); + snprintf(version,MPI_MAX_LIBRARY_VERSION_STRING,"SMPI Version %d.%d. Copyright The Simgrid Team 2007-2013",SIMGRID_VERSION_MAJOR, + SIMGRID_VERSION_MINOR); + *len = strlen(version) > MPI_MAX_LIBRARY_VERSION_STRING ? MPI_MAX_LIBRARY_VERSION_STRING : strlen(version); + smpi_bench_begin(); + return retval; +} + int PMPI_Init_thread(int *argc, char ***argv, int required, int *provided) { if (provided != NULL) { @@ -131,6 +153,11 @@ int PMPI_Address(void *location, MPI_Aint * address) return retval; } +int PMPI_Get_address(void *location, MPI_Aint * address) +{ + return PMPI_Address(location, address); +} + int PMPI_Type_free(MPI_Datatype * datatype) { int retval; @@ -179,6 +206,11 @@ int PMPI_Type_get_extent(MPI_Datatype datatype, MPI_Aint * lb, MPI_Aint * extent return retval; } +int PMPI_Type_get_true_extent(MPI_Datatype datatype, MPI_Aint * lb, MPI_Aint * extent) +{ + return PMPI_Type_get_extent(datatype, lb, extent); +} + int PMPI_Type_extent(MPI_Datatype datatype, MPI_Aint * extent) { int retval; @@ -271,7 +303,6 @@ int PMPI_Group_free(MPI_Group * group) if (group == NULL) { retval = MPI_ERR_ARG; } else { - if(*group!= smpi_comm_group(MPI_COMM_WORLD))// do not free the group of the comm_world smpi_group_destroy(*group); *group = MPI_GROUP_NULL; retval = MPI_SUCCESS; @@ -318,14 +349,17 @@ int PMPI_Group_translate_ranks(MPI_Group group1, int n, int *ranks1, MPI_Group group2, int *ranks2) { int retval, i, index; - smpi_bench_end(); if (group1 == MPI_GROUP_NULL || group2 == MPI_GROUP_NULL) { retval = MPI_ERR_GROUP; } else { for (i = 0; i < n; i++) { - index = smpi_group_index(group1, ranks1[i]); - ranks2[i] = smpi_group_rank(group2, index); + if(ranks1[i]==MPI_PROC_NULL){ + ranks2[i]=MPI_PROC_NULL; + }else{ + index = smpi_group_index(group1, ranks1[i]); + ranks2[i] = smpi_group_rank(group2, index); + } } retval = MPI_SUCCESS; } @@ -384,7 +418,6 @@ int PMPI_Group_union(MPI_Group group1, MPI_Group group2, smpi_group_set_mapping(*newgroup, proc2, i); } } - smpi_group_use(*newgroup); retval = MPI_SUCCESS; } smpi_bench_begin(); @@ -424,7 +457,6 @@ int PMPI_Group_intersection(MPI_Group group1, MPI_Group group2, } } } - smpi_group_use(*newgroup); retval = MPI_SUCCESS; } smpi_bench_begin(); @@ -461,7 +493,6 @@ int PMPI_Group_difference(MPI_Group group1, MPI_Group group2, MPI_Group * newgro } } } - smpi_group_use(*newgroup); retval = MPI_SUCCESS; } smpi_bench_begin(); @@ -482,6 +513,11 @@ int PMPI_Group_incl(MPI_Group group, int n, int *ranks, MPI_Group * newgroup) *newgroup = MPI_GROUP_EMPTY; } else if (n == smpi_group_size(group)) { *newgroup = group; + if(group!= smpi_comm_group(MPI_COMM_WORLD) + && group != MPI_GROUP_NULL + && group != smpi_comm_group(MPI_COMM_SELF) + && group != MPI_GROUP_EMPTY) + smpi_group_use(group); } else { *newgroup = smpi_group_new(n); for (i = 0; i < n; i++) { @@ -489,7 +525,6 @@ int PMPI_Group_incl(MPI_Group group, int n, int *ranks, MPI_Group * newgroup) smpi_group_set_mapping(*newgroup, index, i); } } - smpi_group_use(*newgroup); retval = MPI_SUCCESS; } smpi_bench_begin(); @@ -508,6 +543,11 @@ int PMPI_Group_excl(MPI_Group group, int n, int *ranks, MPI_Group * newgroup) } else { if (n == 0) { *newgroup = group; + if(group!= smpi_comm_group(MPI_COMM_WORLD) + && group != MPI_GROUP_NULL + && group != smpi_comm_group(MPI_COMM_SELF) + && group != MPI_GROUP_EMPTY) + smpi_group_use(group); } else if (n == smpi_group_size(group)) { *newgroup = MPI_GROUP_EMPTY; } else { @@ -532,7 +572,6 @@ int PMPI_Group_excl(MPI_Group group, int n, int *ranks, MPI_Group * newgroup) xbt_free(to_exclude); } - smpi_group_use(*newgroup); retval = MPI_SUCCESS; } smpi_bench_begin(); @@ -556,28 +595,41 @@ int PMPI_Group_range_incl(MPI_Group group, int n, int ranges[][3], size = 0; for (i = 0; i < n; i++) { for (rank = ranges[i][0]; /* First */ - rank >= 0 && rank <= ranges[i][1]; /* Last */ - rank += ranges[i][2] /* Stride */ ) { + rank >= 0; /* Last */ + ) { size++; + + rank += ranges[i][2]; /* Stride */ + if (ranges[i][0] ranges[i][1]) + break; + }else{ + if(rank < ranges[i][1]) + break; + } } } - if (size == smpi_group_size(group)) { - *newgroup = group; - } else { - *newgroup = smpi_group_new(size); - j = 0; - for (i = 0; i < n; i++) { - for (rank = ranges[i][0]; /* First */ - rank >= 0 && rank <= ranges[i][1]; /* Last */ - rank += ranges[i][2] /* Stride */ ) { - index = smpi_group_index(group, rank); - smpi_group_set_mapping(*newgroup, index, j); - j++; - } + + *newgroup = smpi_group_new(size); + j = 0; + for (i = 0; i < n; i++) { + for (rank = ranges[i][0]; /* First */ + rank >= 0; /* Last */ + ) { + index = smpi_group_index(group, rank); + smpi_group_set_mapping(*newgroup, index, j); + j++; + rank += ranges[i][2]; /* Stride */ + if (ranges[i][0] ranges[i][1]) + break; + }else{ + if(rank < ranges[i][1]) + break; + } } } } - smpi_group_use(*newgroup); retval = MPI_SUCCESS; } smpi_bench_begin(); @@ -587,7 +639,7 @@ int PMPI_Group_range_incl(MPI_Group group, int n, int ranges[][3], int PMPI_Group_range_excl(MPI_Group group, int n, int ranges[][3], MPI_Group * newgroup) { - int retval, i, newrank, rank, size, index, add; + int retval, i, rank, newrank,oldrank, size, index, add; smpi_bench_end(); if (group == MPI_GROUP_NULL) { @@ -597,41 +649,65 @@ int PMPI_Group_range_excl(MPI_Group group, int n, int ranges[][3], } else { if (n == 0) { *newgroup = group; + if(group!= smpi_comm_group(MPI_COMM_WORLD) + && group != MPI_GROUP_NULL + && group != smpi_comm_group(MPI_COMM_SELF) + && group != MPI_GROUP_EMPTY) + smpi_group_use(group); } else { size = smpi_group_size(group); for (i = 0; i < n; i++) { for (rank = ranges[i][0]; /* First */ - rank >= 0 && rank <= ranges[i][1]; /* Last */ - rank += ranges[i][2] /* Stride */ ) { + rank >= 0; /* Last */ + ) { size--; + + rank += ranges[i][2]; /* Stride */ + if (ranges[i][0] ranges[i][1]) + break; + }else{ + if(rank < ranges[i][1]) + break; + } } } if (size == 0) { *newgroup = MPI_GROUP_EMPTY; } else { *newgroup = smpi_group_new(size); - newrank = 0; + newrank=0; + oldrank=0; while (newrank < size) { + add=1; for (i = 0; i < n; i++) { - add = 1; - for (rank = ranges[i][0]; /* First */ - rank >= 0 && rank <= ranges[i][1]; /* Last */ - rank += ranges[i][2] /* Stride */ ) { - if (rank == newrank) { - add = 0; - break; + for (rank = ranges[i][0];rank >= 0;){ + if(rank==oldrank){ + add=0; + break; + } + + rank += ranges[i][2]; /* Stride */ + + if (ranges[i][0] ranges[i][1]) + break; + }else{ + if(rank < ranges[i][1]) + break; } - } - if (add == 1) { - index = smpi_group_index(group, newrank); - smpi_group_set_mapping(*newgroup, index, newrank); } } - newrank++; //added to avoid looping, need to be checked .. + if(add==1){ + index = smpi_group_index(group, oldrank); + smpi_group_set_mapping(*newgroup, index, newrank); + newrank++; + } + oldrank++; } } } - smpi_group_use(*newgroup); + retval = MPI_SUCCESS; } smpi_bench_begin(); @@ -700,6 +776,11 @@ int PMPI_Comm_group(MPI_Comm comm, MPI_Group * group) retval = MPI_ERR_ARG; } else { *group = smpi_comm_group(comm); + if(*group!= smpi_comm_group(MPI_COMM_WORLD) + && *group != MPI_GROUP_NULL + && *group != smpi_comm_group(MPI_COMM_SELF) + && *group != MPI_GROUP_EMPTY) + smpi_group_use(*group); retval = MPI_SUCCESS; } smpi_bench_begin(); @@ -760,7 +841,11 @@ int PMPI_Comm_create(MPI_Comm comm, MPI_Group group, MPI_Comm * newcomm) retval = MPI_ERR_GROUP; } else if (newcomm == NULL) { retval = MPI_ERR_ARG; - } else { + } else if(smpi_group_rank(group,smpi_process_index())==MPI_UNDEFINED){ + *newcomm= MPI_COMM_NULL; + retval = MPI_SUCCESS; + }else{ + *newcomm = smpi_comm_new(group); retval = MPI_SUCCESS; } @@ -986,7 +1071,7 @@ int PMPI_Isend(void *buf, int count, MPI_Datatype datatype, int dst, *request = MPI_REQUEST_NULL; retval = MPI_SUCCESS; } else if (dst >= smpi_group_size(smpi_comm_group(comm)) || dst <0){ - retval = MPI_ERR_COMM; + retval = MPI_ERR_RANK; } else if (count < 0) { retval = MPI_ERR_COUNT; } else if (buf==NULL && count > 0) { @@ -1031,7 +1116,7 @@ int PMPI_Issend(void* buf, int count, MPI_Datatype datatype, int dst, int tag, M *request = MPI_REQUEST_NULL; retval = MPI_SUCCESS; } else if (dst >= smpi_group_size(smpi_comm_group(comm)) || dst <0){ - retval = MPI_ERR_COMM; + retval = MPI_ERR_RANK; } else if (count < 0) { retval = MPI_ERR_COUNT; } else if (buf==NULL && count > 0) { @@ -1077,7 +1162,7 @@ int PMPI_Recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, status->MPI_SOURCE = MPI_PROC_NULL; retval = MPI_SUCCESS; } else if (src!=MPI_ANY_SOURCE && (src >= smpi_group_size(smpi_comm_group(comm)) || src <0)){ - retval = MPI_ERR_COMM; + retval = MPI_ERR_RANK; } else if (count < 0) { retval = MPI_ERR_COUNT; } else if (buf==NULL && count > 0) { @@ -1123,7 +1208,7 @@ int PMPI_Send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, } else if (dst == MPI_PROC_NULL) { retval = MPI_SUCCESS; } else if (dst >= smpi_group_size(smpi_comm_group(comm)) || dst <0){ - retval = MPI_ERR_COMM; + retval = MPI_ERR_RANK; } else if (count < 0) { retval = MPI_ERR_COUNT; } else if (buf==NULL && count > 0) { @@ -1167,7 +1252,7 @@ int PMPI_Ssend(void* buf, int count, MPI_Datatype datatype, int dst, int tag, MP } else if (dst == MPI_PROC_NULL) { retval = MPI_SUCCESS; } else if (dst >= smpi_group_size(smpi_comm_group(comm)) || dst <0){ - retval = MPI_ERR_COMM; + retval = MPI_ERR_RANK; } else if (count < 0) { retval = MPI_ERR_COUNT; } else if (buf==NULL && count > 0) { @@ -1219,7 +1304,7 @@ int PMPI_Sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype, retval = MPI_SUCCESS; }else if (dst >= smpi_group_size(smpi_comm_group(comm)) || dst <0 || (src!=MPI_ANY_SOURCE && (src >= smpi_group_size(smpi_comm_group(comm)) || src <0))){ - retval = MPI_ERR_COMM; + retval = MPI_ERR_RANK; } else if (sendcount < 0 || recvcount<0) { retval = MPI_ERR_COUNT; } else if ((sendbuf==NULL && sendcount > 0)||(recvbuf==NULL && recvcount>0)) { @@ -1235,7 +1320,6 @@ int PMPI_Sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype, int src_traced = smpi_group_index(smpi_comm_group(comm), src); TRACE_smpi_ptp_in(rank, src_traced, dst_traced, __FUNCTION__); TRACE_smpi_send(rank, rank, dst_traced); - TRACE_smpi_send(rank, src_traced, rank); #endif @@ -1245,7 +1329,6 @@ int PMPI_Sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype, #ifdef HAVE_TRACING TRACE_smpi_ptp_out(rank, src_traced, dst_traced, __FUNCTION__); - TRACE_smpi_recv(rank, rank, dst_traced); TRACE_smpi_recv(rank, src_traced, rank); TRACE_smpi_computing_in(rank); #endif @@ -1263,18 +1346,18 @@ int PMPI_Sendrecv_replace(void *buf, int count, MPI_Datatype datatype, //TODO: suboptimal implementation void *recvbuf; int retval; - if ((datatype == MPI_DATATYPE_NULL)||(datatype->has_subtype==1)) { + if (datatype == MPI_DATATYPE_NULL) { retval = MPI_ERR_TYPE; } else if (count < 0) { retval = MPI_ERR_COUNT; } else { - int size = smpi_datatype_size(datatype) * count; + int size = smpi_datatype_get_extent(datatype) * count; recvbuf = xbt_new(char, size); retval = MPI_Sendrecv(buf, count, datatype, dst, sendtag, recvbuf, count, datatype, src, recvtag, comm, status); if(retval==MPI_SUCCESS){ - memcpy(buf, recvbuf, size * sizeof(char)); + smpi_datatype_copy(recvbuf, count, datatype, buf, count, datatype); } xbt_free(recvbuf); @@ -1363,6 +1446,7 @@ int PMPI_Iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* statu } else if (comm == MPI_COMM_NULL) { retval = MPI_ERR_COMM; } else if (source == MPI_PROC_NULL) { + *flag=TRUE; smpi_empty_status(status); status->MPI_SOURCE = MPI_PROC_NULL; retval = MPI_SUCCESS; @@ -1392,9 +1476,8 @@ int PMPI_Wait(MPI_Request * request, MPI_Status * status) : -1; TRACE_smpi_computing_out(rank); - MPI_Group group = smpi_comm_group((*request)->comm); - int src_traced = smpi_group_index(group, (*request)->src); - int dst_traced = smpi_group_index(group, (*request)->dst); + int src_traced = (*request)->src; + int dst_traced = (*request)->dst; int is_wait_for_receive = (*request)->recv; TRACE_smpi_ptp_in(rank, src_traced, dst_traced, __FUNCTION__); #endif @@ -1479,7 +1562,7 @@ int PMPI_Waitall(int count, MPI_Request requests[], MPI_Status status[]) int valid_count = 0; for (i = 0; i < count; i++) { MPI_Request req = requests[i]; - if(req){ + if(req!=MPI_REQUEST_NULL){ srcs[valid_count] = req->src; dsts[valid_count] = req->dst; recvs[valid_count] = req->recv; @@ -1517,7 +1600,7 @@ int PMPI_Waitsome(int incount, MPI_Request requests[], int *outcount, int retval; smpi_bench_end(); - if (outcount == NULL || indices == NULL) { + if (outcount == NULL) { retval = MPI_ERR_ARG; } else { *outcount = smpi_mpi_waitsome(incount, requests, indices, status); @@ -1533,7 +1616,7 @@ int PMPI_Testsome(int incount, MPI_Request requests[], int* outcount, int retval; smpi_bench_end(); - if (outcount == NULL || indices == NULL) { + if (outcount == NULL) { retval = MPI_ERR_ARG; } else { *outcount = smpi_mpi_testsome(incount, requests, indices, status); @@ -1558,7 +1641,7 @@ int PMPI_Bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm c if (comm == MPI_COMM_NULL) { retval = MPI_ERR_COMM; } else { - smpi_mpi_bcast(buf, count, datatype, root, comm); + mpi_coll_bcast_fun(buf, count, datatype, root, comm); retval = MPI_SUCCESS; } #ifdef HAVE_TRACING @@ -1582,7 +1665,7 @@ int PMPI_Barrier(MPI_Comm comm) if (comm == MPI_COMM_NULL) { retval = MPI_ERR_COMM; } else { - smpi_mpi_barrier(comm); + mpi_coll_barrier_fun(comm); retval = MPI_SUCCESS; } #ifdef HAVE_TRACING @@ -1608,12 +1691,26 @@ int PMPI_Gather(void *sendbuf, int sendcount, MPI_Datatype sendtype, #endif if (comm == MPI_COMM_NULL) { retval = MPI_ERR_COMM; - } else if (sendtype == MPI_DATATYPE_NULL - || recvtype == MPI_DATATYPE_NULL) { + } else if ((( sendbuf != MPI_IN_PLACE) && (sendtype == MPI_DATATYPE_NULL)) || + ((smpi_comm_rank(comm) == root) && (recvtype == MPI_DATATYPE_NULL))){ retval = MPI_ERR_TYPE; + } else if ((( sendbuf != MPI_IN_PLACE) && (sendcount <0)) || + ((smpi_comm_rank(comm) == root) && (recvcount <0))){ + retval = MPI_ERR_COUNT; } else { - smpi_mpi_gather(sendbuf, sendcount, sendtype, recvbuf, recvcount, + + char* sendtmpbuf = (char*) sendbuf; + int sendtmpcount = sendcount; + MPI_Datatype sendtmptype = sendtype; + if( (smpi_comm_rank(comm) == root) && (sendbuf == MPI_IN_PLACE )) { + sendtmpcount=0; + sendtmptype=recvtype; + } + + mpi_coll_gather_fun(sendtmpbuf, sendtmpcount, sendtmptype, recvbuf, recvcount, recvtype, root, comm); + + retval = MPI_SUCCESS; } #ifdef HAVE_TRACING @@ -1639,13 +1736,24 @@ int PMPI_Gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype, #endif if (comm == MPI_COMM_NULL) { retval = MPI_ERR_COMM; - } else if (sendtype == MPI_DATATYPE_NULL - || recvtype == MPI_DATATYPE_NULL) { + } else if ((( sendbuf != MPI_IN_PLACE) && (sendtype == MPI_DATATYPE_NULL)) || + ((smpi_comm_rank(comm) == root) && (recvtype == MPI_DATATYPE_NULL))){ retval = MPI_ERR_TYPE; + } else if (( sendbuf != MPI_IN_PLACE) && (sendcount <0)){ + retval = MPI_ERR_COUNT; } else if (recvcounts == NULL || displs == NULL) { retval = MPI_ERR_ARG; } else { - smpi_mpi_gatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts, + + char* sendtmpbuf = (char*) sendbuf; + int sendtmpcount = sendcount; + MPI_Datatype sendtmptype = sendtype; + if( (smpi_comm_rank(comm) == root) && (sendbuf == MPI_IN_PLACE )) { + sendtmpcount=0; + sendtmptype=recvtype; + } + + smpi_mpi_gatherv(sendtmpbuf, sendtmpcount, sendtmptype, recvbuf, recvcounts, displs, recvtype, root, comm); retval = MPI_SUCCESS; } @@ -1671,12 +1779,22 @@ int PMPI_Allgather(void *sendbuf, int sendcount, MPI_Datatype sendtype, #endif if (comm == MPI_COMM_NULL) { retval = MPI_ERR_COMM; - } else if (sendtype == MPI_DATATYPE_NULL - || recvtype == MPI_DATATYPE_NULL) { + } else if ((( sendbuf != MPI_IN_PLACE) && (sendtype == MPI_DATATYPE_NULL)) || + (recvtype == MPI_DATATYPE_NULL)){ retval = MPI_ERR_TYPE; + } else if ((( sendbuf != MPI_IN_PLACE) && (sendcount <0)) || + (recvcount <0)){ + retval = MPI_ERR_COUNT; } else { - smpi_mpi_allgather(sendbuf, sendcount, sendtype, recvbuf, recvcount, - recvtype, comm); + + if(sendbuf == MPI_IN_PLACE) { + sendbuf=((char*)recvbuf)+smpi_datatype_get_extent(recvtype)*recvcount*smpi_comm_rank(comm); + sendcount=recvcount; + sendtype=recvtype; + } + + mpi_coll_allgather_fun(sendbuf, sendcount, sendtype, recvbuf, recvcount, + recvtype, comm); retval = MPI_SUCCESS; } #ifdef HAVE_TRACING @@ -1700,13 +1818,22 @@ int PMPI_Allgatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype, #endif if (comm == MPI_COMM_NULL) { retval = MPI_ERR_COMM; - } else if (sendtype == MPI_DATATYPE_NULL - || recvtype == MPI_DATATYPE_NULL) { + } else if ((( sendbuf != MPI_IN_PLACE) && (sendtype == MPI_DATATYPE_NULL)) || + (recvtype == MPI_DATATYPE_NULL)){ retval = MPI_ERR_TYPE; + } else if (( sendbuf != MPI_IN_PLACE) && (sendcount <0)){ + retval = MPI_ERR_COUNT; } else if (recvcounts == NULL || displs == NULL) { retval = MPI_ERR_ARG; } else { - smpi_mpi_allgatherv(sendbuf, sendcount, sendtype, recvbuf, recvcounts, + + if(sendbuf == MPI_IN_PLACE) { + sendbuf=((char*)recvbuf)+smpi_datatype_get_extent(recvtype)*displs[smpi_comm_rank(comm)]; + sendcount=recvcounts[smpi_comm_rank(comm)]; + sendtype=recvtype; + } + + mpi_coll_allgatherv_fun(sendbuf, sendcount, sendtype, recvbuf, recvcounts, displs, recvtype, comm); retval = MPI_SUCCESS; } @@ -1734,11 +1861,15 @@ int PMPI_Scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype, #endif if (comm == MPI_COMM_NULL) { retval = MPI_ERR_COMM; - } else if (sendtype == MPI_DATATYPE_NULL - || recvtype == MPI_DATATYPE_NULL) { + } else if (((smpi_comm_rank(comm)==root) && (sendtype == MPI_DATATYPE_NULL)) + || ((recvbuf !=MPI_IN_PLACE) && (recvtype == MPI_DATATYPE_NULL))) { retval = MPI_ERR_TYPE; } else { - smpi_mpi_scatter(sendbuf, sendcount, sendtype, recvbuf, recvcount, + if (recvbuf == MPI_IN_PLACE) { + recvtype=sendtype; + recvcount=sendcount; + } + mpi_coll_scatter_fun(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, root, comm); retval = MPI_SUCCESS; } @@ -1765,12 +1896,16 @@ int PMPI_Scatterv(void *sendbuf, int *sendcounts, int *displs, #endif if (comm == MPI_COMM_NULL) { retval = MPI_ERR_COMM; - } else if (sendtype == MPI_DATATYPE_NULL - || recvtype == MPI_DATATYPE_NULL) { - retval = MPI_ERR_TYPE; } else if (sendcounts == NULL || displs == NULL) { retval = MPI_ERR_ARG; + } else if (((smpi_comm_rank(comm)==root) && (sendtype == MPI_DATATYPE_NULL)) + || ((recvbuf !=MPI_IN_PLACE) && (recvtype == MPI_DATATYPE_NULL))) { + retval = MPI_ERR_TYPE; } else { + if (recvbuf == MPI_IN_PLACE) { + recvtype=sendtype; + recvcount=sendcounts[smpi_comm_rank(comm)]; + } smpi_mpi_scatterv(sendbuf, sendcounts, displs, sendtype, recvbuf, recvcount, recvtype, root, comm); retval = MPI_SUCCESS; @@ -1800,7 +1935,9 @@ int PMPI_Reduce(void *sendbuf, void *recvbuf, int count, } else if (datatype == MPI_DATATYPE_NULL || op == MPI_OP_NULL) { retval = MPI_ERR_ARG; } else { - smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, root, comm); + + mpi_coll_reduce_fun(sendbuf, recvbuf, count, datatype, op, root, comm); + retval = MPI_SUCCESS; } #ifdef HAVE_TRACING @@ -1811,6 +1948,21 @@ int PMPI_Reduce(void *sendbuf, void *recvbuf, int count, return retval; } +int PMPI_Reduce_local(void *inbuf, void *inoutbuf, int count, + MPI_Datatype datatype, MPI_Op op){ + int retval; + + smpi_bench_end(); + if (datatype == MPI_DATATYPE_NULL || op == MPI_OP_NULL) { + retval = MPI_ERR_ARG; + } else { + smpi_op_apply(op, inbuf, inoutbuf, &count, &datatype); + retval=MPI_SUCCESS; + } + smpi_bench_begin(); + return retval; +} + int PMPI_Allreduce(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) { @@ -1829,8 +1981,21 @@ int PMPI_Allreduce(void *sendbuf, void *recvbuf, int count, } else if (op == MPI_OP_NULL) { retval = MPI_ERR_OP; } else { - smpi_mpi_allreduce(sendbuf, recvbuf, count, datatype, op, comm); + + char* sendtmpbuf = (char*) sendbuf; + if( sendbuf == MPI_IN_PLACE ) { + sendtmpbuf = (char *)xbt_malloc(count*smpi_datatype_get_extent(datatype)); + smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype); + } + + mpi_coll_allreduce_fun(sendtmpbuf, recvbuf, count, datatype, op, comm); + + if( sendbuf == MPI_IN_PLACE ) { + xbt_free(sendtmpbuf); + } + retval = MPI_SUCCESS; + } #ifdef HAVE_TRACING TRACE_smpi_collective_out(rank, -1, __FUNCTION__); @@ -1869,15 +2034,41 @@ int PMPI_Scan(void *sendbuf, void *recvbuf, int count, 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) { - int retval, i, size, count; - int *displs; - int rank = comm != MPI_COMM_NULL ? smpi_process_index() : -1; - + 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 @@ -1890,19 +2081,48 @@ int PMPI_Reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts, } else if (recvcounts == NULL) { retval = MPI_ERR_ARG; } else { - /* arbitrarily choose root as rank 0 */ - /* TODO: faster direct implementation ? */ - size = smpi_comm_size(comm); - count = 0; - displs = xbt_new(int, size); - for (i = 0; i < size; i++) { - count += recvcounts[i]; - displs[i] = 0; + void* sendtmpbuf=sendbuf; + if(sendbuf==MPI_IN_PLACE){ + sendtmpbuf=recvbuf; } - smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm); - smpi_mpi_scatterv(recvbuf, recvcounts, displs, datatype, recvbuf, - recvcounts[rank], datatype, 0, comm); - xbt_free(displs); + + mpi_coll_reduce_scatter_fun(sendtmpbuf, recvbuf, recvcounts, + 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_block(void *sendbuf, void *recvbuf, int recvcount, + MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) +{ + int retval,i; + 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 if (recvcount < 0) { + retval = MPI_ERR_ARG; + } else { + int count=smpi_comm_size(comm); + int* recvcounts=(int*)xbt_malloc(count); + for (i=0; i