X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/1e356ccf3f40973c033d35177172838dd5b7854f..d30fcaa41c0aabda5afef028a793b9e295a4aadb:/src/smpi/smpi_mpi.c diff --git a/src/smpi/smpi_mpi.c b/src/smpi/smpi_mpi.c index ca52f3fc11..926728888a 100644 --- a/src/smpi/smpi_mpi.c +++ b/src/smpi/smpi_mpi.c @@ -1,11 +1,14 @@ + + #include "private.h" +#include "smpi_coll_private.h" XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_mpi, smpi, "Logging specific to SMPI (mpi)"); int SMPI_MPI_Init(int *argc, char ***argv) { - smpi_process_init(argc,argv); + smpi_process_init(argc, argv); smpi_bench_begin(); return MPI_SUCCESS; } @@ -181,42 +184,213 @@ int SMPI_MPI_Send(void *buf, int count, MPI_Datatype datatype, int dst, return retval; } +/** + * MPI_Wait and friends + **/ int SMPI_MPI_Wait(MPI_Request * request, MPI_Status * status) { return smpi_mpi_wait(*request, status); } +int SMPI_MPI_Waitall(int count, MPI_Request requests[], MPI_Status status[]) +{ + return smpi_mpi_waitall(count, requests, status); +} + +int SMPI_MPI_Waitany(int count, MPI_Request requests[], int *index, + MPI_Status status[]) +{ + return smpi_mpi_waitany(count, requests, index, status); +} + +/** + * MPI_Bcast + **/ + +/** + * flat bcast + **/ +int flat_tree_bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm); +int flat_tree_bcast(void *buf, int count, MPI_Datatype datatype, int root, + MPI_Comm comm) +{ + int rank; + int retval = MPI_SUCCESS; + smpi_mpi_request_t request; + + rank = smpi_mpi_comm_rank(comm); + if (rank == root) { + retval = smpi_create_request(buf, count, datatype, root, + (root + 1) % comm->size, 0, comm, &request); + request->forward = comm->size - 1; + smpi_mpi_isend(request); + } else { + retval = smpi_create_request(buf, count, datatype, MPI_ANY_SOURCE, rank, + 0, comm, &request); + smpi_mpi_irecv(request); + } + + smpi_mpi_wait(request, MPI_STATUS_IGNORE); + xbt_mallocator_release(smpi_global->request_mallocator, request); + + return(retval); + +} + +/** + * Bcast user entry point + **/ int SMPI_MPI_Bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm) { int retval = MPI_SUCCESS; + + smpi_bench_end(); + + //retval = flat_tree_bcast(buf, count, datatype, root, comm); + retval = binomial_tree_bcast(buf, count, datatype, root, comm); + + smpi_bench_begin(); + + return retval; +} + + + +#ifdef DEBUG_REDUCE +/** + * debugging helper function + **/ +static void print_buffer_int(void *buf, int len, const char *msg, int rank) +{ + int tmp, *v; + printf("**[%d] %s: ", rank, msg); + for (tmp = 0; tmp < len; tmp++) { + v = buf; + printf("[%d]", v[tmp]); + } + printf("\n"); + free(msg); +} +#endif + +/** + * MPI_Reduce + **/ +int SMPI_MPI_Reduce(void *sendbuf, void *recvbuf, int count, + MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm) +{ + int retval = MPI_SUCCESS; int rank; + int size; + int i; + int tag = 0; + smpi_mpi_request_t *tabrequest; smpi_mpi_request_t request; smpi_bench_end(); rank = smpi_mpi_comm_rank(comm); + size = comm->size; - if (rank == root) { - retval = smpi_create_request(buf, count, datatype, root, - (root + 1) % comm->size, 0, comm, &request); - request->forward = comm->size - 1; + if (rank != root) { // if i am not ROOT, simply send my buffer to root + +#ifdef DEBUG_REDUCE + print_buffer_int(sendbuf, count, xbt_strdup("sndbuf"), rank); +#endif + retval = + smpi_create_request(sendbuf, count, datatype, rank, root, tag, comm, + &request); smpi_mpi_isend(request); + smpi_mpi_wait(request, MPI_STATUS_IGNORE); + xbt_mallocator_release(smpi_global->request_mallocator, request); + } else { - retval = smpi_create_request(buf, count, datatype, MPI_ANY_SOURCE, rank, - 0, comm, &request); - smpi_mpi_irecv(request); + // i am the ROOT: wait for all buffers by creating one request by sender + int src; + tabrequest = xbt_malloc((size - 1) * sizeof(smpi_mpi_request_t)); + + void **tmpbufs = xbt_malloc((size - 1) * sizeof(void *)); + for (i = 0; i < size - 1; i++) { + // we need 1 buffer per request to store intermediate receptions + tmpbufs[i] = xbt_malloc(count * datatype->size); + } + memcpy(recvbuf, sendbuf, count * datatype->size * sizeof(char)); // initiliaze recv buf with my own snd buf + + // i can not use: 'request->forward = size-1;' (which would progagate size-1 receive reqs) + // since we should op values as soon as one receiving request matches. + for (i = 0; i < size - 1; i++) { + // reminder: for smpi_create_request() the src is always the process sending. + src = i < root ? i : i + 1; + retval = smpi_create_request(tmpbufs[i], count, datatype, + src, root, tag, comm, &(tabrequest[i])); + if (NULL != tabrequest[i] && MPI_SUCCESS == retval) { + if (MPI_SUCCESS == retval) { + smpi_mpi_irecv(tabrequest[i]); + } + } + } + // now, wait for completion of all irecv's. + for (i = 0; i < size - 1; i++) { + int index = MPI_UNDEFINED; + smpi_mpi_waitany(size - 1, tabrequest, &index, MPI_STATUS_IGNORE); + +#ifdef DEBUG_REDUCE + printf + ("MPI_Waitany() unblocked: root received (completes req[index=%d])\n", + index); + print_buffer_int(tmpbufs[index], count, + bprintf("tmpbufs[index=%d] (value received)", index), + rank); +#endif + + // arg 2 is modified + op->func(tmpbufs[index], recvbuf, &count, &datatype); +#ifdef DEBUG_REDUCE + print_buffer_int(recvbuf, count, xbt_strdup("rcvbuf"), rank); + +#endif + //xbt_mallocator_release(smpi_global->request_mallocator, tabrequest[i]); + xbt_free(tmpbufs[index]); + } + xbt_free(tabrequest); + xbt_free(tmpbufs); } - smpi_mpi_wait(request, MPI_STATUS_IGNORE); - xbt_mallocator_release(smpi_global->request_mallocator, request); - smpi_bench_begin(); return retval; } +/** + * MPI_Allreduce + * + * Same as MPI_REDUCE except that the result appears in the receive buffer of all the group members. + **/ +int SMPI_MPI_Allreduce( void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, + MPI_Op op, MPI_Comm comm ); +int SMPI_MPI_Allreduce( void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, + MPI_Op op, MPI_Comm comm ) +{ + int retval = MPI_SUCCESS; + int root=0; // arbitrary choice + + smpi_bench_end(); + + retval = SMPI_MPI_Reduce( sendbuf, recvbuf, count, datatype, op, root, comm); + if (MPI_SUCCESS != retval) + return(retval); + retval = SMPI_MPI_Bcast( recvbuf, count, datatype, root, comm); + + smpi_bench_begin(); + return( retval ); +} + + + + + // used by comm_split to sort ranks based on key values int smpi_compare_rankkeys(const void *a, const void *b); int smpi_compare_rankkeys(const void *a, const void *b) @@ -348,3 +522,8 @@ int SMPI_MPI_Comm_split(MPI_Comm comm, int color, int key, return retval; } + +double SMPI_MPI_Wtime(void) +{ + return (SIMIX_get_clock()); +}