X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/5f758ee0ce7c9908ead65c3b6eb4229e55ee8a27..ae2d6a900ad8864d500f3e6fb63b6074519cb364:/src/smpi/smpi_mpi.c diff --git a/src/smpi/smpi_mpi.c b/src/smpi/smpi_mpi.c index 301a642113..20ff68f150 100644 --- a/src/smpi/smpi_mpi.c +++ b/src/smpi/smpi_mpi.c @@ -1,11 +1,26 @@ +/* $Id$tag */ + +/* smpi_mpi.c -- + * + * Eventually will contain the user level MPI primitives and its corresponding + * internal wrapper. The implementations of these primitives should go to specific + * files. For example, a SMPI_MPI_Bcast() in this file, should call the wrapper + * smpi_mpi_bcast(), which decides which implementation to call. Currently, it + * calls nary_tree_bcast() in smpi_coll.c. (Stéphane Genaud). + * */ + + + #include "private.h" +#include "smpi_coll_private.h" +#include "smpi_mpi_dt_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; } @@ -63,35 +78,28 @@ int SMPI_MPI_Comm_rank(MPI_Comm comm, int *rank) return retval; } -int SMPI_MPI_Type_size(MPI_Datatype datatype, size_t * size) -{ - int retval = MPI_SUCCESS; - - smpi_bench_end(); - if (NULL == datatype) { - retval = MPI_ERR_TYPE; - } else if (NULL == size) { - retval = MPI_ERR_ARG; - } else { - *size = datatype->size; - } - - smpi_bench_begin(); - - return retval; -} +/** + * Barrier + **/ int SMPI_MPI_Barrier(MPI_Comm comm) { int retval = MPI_SUCCESS; + int arity=4; smpi_bench_end(); if (NULL == comm) { retval = MPI_ERR_COMM; } else { - retval = smpi_mpi_barrier(comm); + + /* + * original implemantation: + * retval = smpi_mpi_barrier(comm); + * this one is unrealistic: it just cond_waits, means no time. + */ + retval = nary_tree_barrier( comm, arity ); } smpi_bench_begin(); @@ -99,14 +107,17 @@ int SMPI_MPI_Barrier(MPI_Comm comm) return retval; } + + int SMPI_MPI_Irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Request * request) { int retval = MPI_SUCCESS; + int rank; smpi_bench_end(); - - retval = smpi_create_request(buf, count, datatype, src, 0, tag, comm, + rank = smpi_mpi_comm_rank(comm); + retval = smpi_create_request(buf, count, datatype, src, rank, tag, comm, request); if (NULL != *request && MPI_SUCCESS == retval) { retval = smpi_mpi_irecv(*request); @@ -121,11 +132,13 @@ int SMPI_MPI_Recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status) { int retval = MPI_SUCCESS; + int rank; smpi_mpi_request_t request; smpi_bench_end(); - retval = smpi_create_request(buf, count, datatype, src, 0, tag, comm, + rank = smpi_mpi_comm_rank(comm); + retval = smpi_create_request(buf, count, datatype, src, rank, tag, comm, &request); if (NULL != request && MPI_SUCCESS == retval) { retval = smpi_mpi_irecv(request); @@ -144,10 +157,12 @@ int SMPI_MPI_Isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm, MPI_Request * request) { int retval = MPI_SUCCESS; + int rank; smpi_bench_end(); - retval = smpi_create_request(buf, count, datatype, 0, dst, tag, comm, + rank = smpi_mpi_comm_rank(comm); + retval = smpi_create_request(buf, count, datatype, rank, dst, tag, comm, request); if (NULL != *request && MPI_SUCCESS == retval) { retval = smpi_mpi_isend(*request); @@ -158,15 +173,20 @@ int SMPI_MPI_Isend(void *buf, int count, MPI_Datatype datatype, int dst, return retval; } +/** + * MPI_Send user level + **/ int SMPI_MPI_Send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) { int retval = MPI_SUCCESS; + int rank; smpi_mpi_request_t request; smpi_bench_end(); - retval = smpi_create_request(buf, count, datatype, 0, dst, tag, comm, + rank = smpi_mpi_comm_rank(comm); + retval = smpi_create_request(buf, count, datatype, rank, dst, tag, comm, &request); if (NULL != request && MPI_SUCCESS == retval) { retval = smpi_mpi_isend(request); @@ -181,114 +201,452 @@ int SMPI_MPI_Send(void *buf, int count, MPI_Datatype datatype, int dst, return retval; } + +/** + * MPI_Sendrecv internal level + **/ +int smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype, int dest, int sendtag, + void *recvbuf, int recvcount, MPI_Datatype recvtype, int source, int recvtag, + MPI_Comm comm, MPI_Status *status) +{ +int rank; +int retval = MPI_SUCCESS; +smpi_mpi_request_t srequest; +smpi_mpi_request_t rrequest; + + rank = smpi_mpi_comm_rank(comm); + + /* send */ + retval = smpi_create_request(sendbuf, sendcount, sendtype, + rank,dest,sendtag, + comm, &srequest); + smpi_mpi_isend(srequest); + + /* recv */ + retval = smpi_create_request(recvbuf, recvcount, recvtype, + source, rank,recvtag, + comm, &rrequest); + smpi_mpi_irecv(rrequest); + + smpi_mpi_wait(srequest, MPI_STATUS_IGNORE); + //printf("[%d] isend request src=%d dst=%d tag=%d COMPLETED (retval=%d) \n",rank,rank,dest,sendtag,retval); + smpi_mpi_wait(rrequest, MPI_STATUS_IGNORE); + //printf("[%d] irecv request src=%d -> dst=%d tag=%d COMPLETED (retval=%d)\n",rank,source,rank,recvtag,retval); + + return(retval); +} +/** + * MPI_Sendrecv user entry point + **/ +int SMPI_MPI_Sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype, int dest, int sendtag, + void *recvbuf, int recvcount, MPI_Datatype recvtype, int source, int recvtag, + MPI_Comm comm, MPI_Status *status) +{ +int retval = MPI_SUCCESS; + + smpi_bench_end(); + smpi_mpi_sendrecv( sendbuf, sendcount, sendtype, dest, sendtag, + recvbuf, recvcount, recvtype, source, recvtag, + comm, status); + smpi_bench_begin(); + + return retval; + + +} + +/** + * MPI_Wait and friends + **/ int SMPI_MPI_Wait(MPI_Request * request, MPI_Status * status) { - return smpi_mpi_wait(*request, status); + int retval; + + smpi_bench_end(); + retval = smpi_mpi_wait(*request, status); + smpi_bench_begin(); + return retval; } -int SMPI_MPI_Waitall(int count, MPI_Request requests[], MPI_Status status[]) { - return smpi_mpi_waitall(count, requests,status); +int SMPI_MPI_Waitall(int count, MPI_Request requests[], MPI_Status status[]) +{ + int retval; + + smpi_bench_end(); + retval = smpi_mpi_waitall(count, requests, status); + smpi_bench_begin(); + return retval; } -int SMPI_MPI_Waitany(int count, MPI_Request requests[], int *index, MPI_Status status[]) { - return smpi_mpi_waitany(count, requests, index,status); + +int SMPI_MPI_Waitany(int count, MPI_Request requests[], int *index, + MPI_Status status[]) +{ + int retval; + + smpi_bench_end(); + retval = smpi_mpi_waitany(count, requests, index, status); + smpi_bench_begin(); + return retval; } + /** * 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 internal level + **/ +int smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root, + MPI_Comm comm) +{ + int retval = MPI_SUCCESS; + int rank = smpi_mpi_comm_rank(comm); + + DEBUG1("<%d> entered smpi_mpi_bcast(). Calls nary_tree_bcast()",rank); + //retval = flat_tree_bcast(buf, count, datatype, root, comm); + retval = nary_tree_bcast(buf, count, datatype, root, comm, 2 ); + 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(); + smpi_mpi_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, 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); +} +static void print_buffer_double(void *buf, int len, char *msg, int rank) +{ + int tmp; + double *v; + printf("**[%d] %s: ", rank, msg); + for (tmp = 0; tmp < len; tmp++) { + v = buf; + printf("[%lf]", v[tmp]); + } + printf("\n"); + free(msg); +} + + +#endif +/** + * MPI_Reduce internal level + **/ +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 system_tag = 666; + smpi_mpi_request_t *requests; + smpi_mpi_request_t request; + + smpi_bench_end(); + + rank = smpi_mpi_comm_rank(comm); + size = comm->size; + DEBUG1("<%d> entered smpi_mpi_reduce()",rank); + + 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, system_tag, comm, + &request); + smpi_mpi_isend(request); + smpi_mpi_wait(request, MPI_STATUS_IGNORE); + xbt_mallocator_release(smpi_global->request_mallocator, request); + + } else { + // i am the ROOT: wait for all buffers by creating one request by sender + int src; + requests = 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); + } + // root: initiliaze recv buf with my own snd buf + memcpy(recvbuf, sendbuf, count * datatype->size * sizeof(char)); + + // 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, system_tag, comm, &(requests[i])); + if (NULL != requests[i] && MPI_SUCCESS == retval) { + if (MPI_SUCCESS == retval) { + smpi_mpi_irecv(requests[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, requests, &index, MPI_STATUS_IGNORE); + DEBUG3("<%d> waitany() unblocked by reception (completes request[%d]) (%d reqs remaining)", + rank,index,size-i-2); +#ifdef DEBUG_REDUCE + 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_free(tmpbufs[index]); + /* FIXME: with the following line, it generates an + * [xbt_ex/CRITICAL] Conditional list not empty 162518800. + * Fixed ? + */ + xbt_mallocator_release(smpi_global->request_mallocator, requests[index]); + } + xbt_free(requests); + xbt_free(tmpbufs); + } + return retval; +} + +/** + * MPI_Reduce user entry point + **/ +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; + + smpi_bench_end(); + + retval = smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, root, comm); + + 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 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( sendbuf, count, datatype, root, comm); + smpi_bench_end(); + return( retval ); +} + + +/** + * MPI_Scatter user entry point + **/ +int SMPI_MPI_Scatter(void *sendbuf, int sendcount, MPI_Datatype datatype, + void *recvbuf, int recvcount, MPI_Datatype recvtype, + int root, MPI_Comm comm) +{ int retval = MPI_SUCCESS; + int i; + int cnt=0; int rank; + int tag=0; + char *cptr; // to manipulate the void * buffers + smpi_mpi_request_t *requests; smpi_mpi_request_t request; + smpi_mpi_status_t status; + smpi_bench_end(); rank = smpi_mpi_comm_rank(comm); + requests = xbt_malloc((comm->size-1) * sizeof(smpi_mpi_request_t)); 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); - } + // i am the root: distribute my sendbuf + //print_buffer_int(sendbuf, comm->size, xbt_strdup("rcvbuf"), rank); + cptr = sendbuf; + for (i=0; i < comm->size; i++) { + if ( i!=root ) { // send to processes ... + + retval = smpi_create_request((void *)cptr, sendcount, + datatype, root, i, tag, comm, &(requests[cnt])); + if (NULL != requests[cnt] && MPI_SUCCESS == retval) { + if (MPI_SUCCESS == retval) { + smpi_mpi_isend(requests[cnt]); + } + } + cnt++; + } + else { // ... except if it's me. + memcpy(recvbuf, (void *)cptr, recvcount*recvtype->size*sizeof(char)); + } + cptr += sendcount*datatype->size; + } + for(i=0; irequest_mallocator, requests[i]); - smpi_mpi_wait(request, MPI_STATUS_IGNORE); - xbt_mallocator_release(smpi_global->request_mallocator, request); + } + } + else { // i am a non-root process: wait data from the root + retval = smpi_create_request(recvbuf,recvcount, + recvtype, root, rank, tag, comm, &request); + if (NULL != request && MPI_SUCCESS == retval) { + if (MPI_SUCCESS == retval) { + smpi_mpi_irecv(request); + } + } + smpi_mpi_wait(request, &status); + xbt_mallocator_release(smpi_global->request_mallocator, request); + } + xbt_free(requests); smpi_bench_begin(); return retval; } + /** - * MPI_Reduce + * MPI_Alltoall user entry point + * + * Uses the logic of OpenMPI (upto 1.2.7 or greater) for the optimizations + * ompi/mca/coll/tuned/coll_tuned_module.c **/ - -int SMPI_MPI_Reduce( void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, - MPI_Op op, int root, MPI_Comm comm ) +int SMPI_MPI_Alltoall(void *sendbuf, int sendcount, MPI_Datatype datatype, + void *recvbuf, int recvcount, MPI_Datatype recvtype, + MPI_Comm comm) { int retval = MPI_SUCCESS; + int block_dsize; int rank; - int size; - int i; - smpi_mpi_request_t *tabrequest; smpi_bench_end(); rank = smpi_mpi_comm_rank(comm); - size = comm->size; + block_dsize = datatype->size * sendcount; + DEBUG2("<%d> optimized alltoall() called. Block size sent to each rank: %d bytes.",rank,block_dsize); - tabrequest = xbt_malloc((size)*sizeof(smpi_mpi_request_t)); + if ((block_dsize < 200) && (comm->size > 12)) { + retval = smpi_coll_tuned_alltoall_bruck(sendbuf, sendcount, datatype, + recvbuf, recvcount, recvtype, comm); - if (rank != root) { // if i am not root, simply send my buffer to root - retval = smpi_create_request(sendbuf, count, datatype, - rank, root, 0, comm, &(tabrequest[rank])); - smpi_mpi_isend(tabrequest[rank]); - smpi_mpi_wait(tabrequest[rank], MPI_STATUS_IGNORE); - //printf("DEBUG: rank %d sent my sendbuf to root (rank %d)\n",rank,root); + } else if (block_dsize < 3000) { + retval = smpi_coll_tuned_alltoall_basic_linear(sendbuf, sendcount, datatype, + recvbuf, recvcount, recvtype, comm); } else { - // i am the root: wait for all buffers by creating requests - // 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; isize; i++) { - if ( rank != i ) { // except for me - // reminder: for smpi_create_request() the src is always the process sending. - retval = smpi_create_request(recvbuf, count, datatype, MPI_ANY_SOURCE, root, - 0, 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. - // FIXME: we should implement smpi_wait_all for a more asynchronous behavior - for (i=0; isize; i++) { - if ( rank != i ) { // except for me - smpi_mpi_wait(tabrequest[i], MPI_STATUS_IGNORE); - - // FIXME: the core part is here. To be written ... - - fprintf(stderr,"[smpi] %s:%d : MPI_Reduce *Not yet implemented*.\n",__FILE__,__LINE__); - } - } + retval = smpi_coll_tuned_alltoall_pairwise(sendbuf, sendcount, datatype, + recvbuf, recvcount, recvtype, comm); } - for (i=0; isize; i++) - xbt_mallocator_release(smpi_global->request_mallocator, tabrequest[i]); smpi_bench_begin(); return retval; } +/** + * MPI_Alltoallv user entry point + * + * As in OpenMPI, alltoallv is not optimized + * ompi/mca/coll/basic/coll_basic_alltoallv.c + **/ +int SMPI_MPI_Alltoallv(void *sendbuf, int *scounts, int *sdisps, MPI_Datatype datatype, + void *recvbuf, int *rcounts, int *rdisps, MPI_Datatype recvtype, + MPI_Comm comm) +{ + int retval = MPI_SUCCESS; + int rank; + + smpi_bench_end(); + rank = smpi_mpi_comm_rank(comm); + DEBUG1("<%d> basic alltoallv() called.",rank); + + retval = smpi_coll_basic_alltoallv(sendbuf, scounts, sdisps, datatype, + recvbuf, rcounts, rdisps, recvtype, + 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) @@ -421,7 +779,230 @@ int SMPI_MPI_Comm_split(MPI_Comm comm, int color, int key, return retval; } -double SMPI_MPI_Wtime( void ) +double SMPI_MPI_Wtime(void) +{ + double time; + + smpi_bench_end(); + time = SIMIX_get_clock(); + smpi_bench_begin(); + return time; +} + +int SMPI_MPI_Gather(void* sendbuf, int sendcount, MPI_Datatype sendtype, + void* recvbuf, int recvcount, MPI_Datatype recvtype, + int root, MPI_Comm comm) +{ + int retval = MPI_SUCCESS; + int system_tag = 666; + int rank, size; + + smpi_bench_end(); + rank = smpi_mpi_comm_rank(comm); + size = comm->size; + if(rank != root) { + // Send buffer to root + smpi_mpi_request_t request; + + retval = smpi_create_request(sendbuf, sendcount, sendtype, + rank, root, system_tag, comm, &request); + smpi_mpi_isend(request); + smpi_mpi_wait(request, MPI_STATUS_IGNORE); + xbt_mallocator_release(smpi_global->request_mallocator, request); + } else { + // Receive buffers from senders + int src; + smpi_mpi_request_t* requests; + + requests = xbt_malloc((size-1) * sizeof(smpi_mpi_request_t)); + for(src = 0; src < size; src++) { + if(src == root) { + // Local copy from root + memcpy(&((char*)recvbuf)[src*recvcount*recvtype->size], + sendbuf, sendcount*sendtype->size*sizeof(char)); + } else { + int index = src < root ? src : src - 1; + retval = smpi_create_request(&((char*)recvbuf)[src*recvcount*recvtype->size], + recvcount, recvtype, src, root, system_tag, + comm, &requests[index]); + if(NULL != requests[index] && MPI_SUCCESS == retval) { + smpi_mpi_irecv(requests[index]); + } + } + } + // Wait for completion of irecv's. + for(src = 0; src < size - 1; src++) { + int index = MPI_UNDEFINED; + smpi_mpi_waitany(size - 1, requests, &index, MPI_STATUS_IGNORE); + xbt_mallocator_release(smpi_global->request_mallocator, requests[index]); + } + xbt_free(requests); + } + smpi_bench_begin(); + return retval; +} + +int SMPI_MPI_Gatherv(void* sendbuf, int sendcount, MPI_Datatype sendtype, + void* recvbuf, int* recvcounts, int* displs, MPI_Datatype recvtype, + int root, MPI_Comm comm) +{ + int retval = MPI_SUCCESS; + int system_tag = 666; + int rank, size; + + smpi_bench_end(); + rank = smpi_mpi_comm_rank(comm); + size = comm->size; + if(rank != root) { + // Send buffer to root + smpi_mpi_request_t request; + + retval = smpi_create_request(sendbuf, sendcount, sendtype, + rank, root, system_tag, comm, &request); + smpi_mpi_isend(request); + smpi_mpi_wait(request, MPI_STATUS_IGNORE); + xbt_mallocator_release(smpi_global->request_mallocator, request); + } else { + // Receive buffers from senders + int src; + smpi_mpi_request_t* requests; + + requests = xbt_malloc((size-1) * sizeof(smpi_mpi_request_t)); + for(src = 0; src < size; src++) { + if(src == root) { + // Local copy from root + memcpy(&((char*)recvbuf)[displs[src]], + sendbuf, sendcount*sendtype->size*sizeof(char)); + } else { + int index = src < root ? src : src - 1; + retval = smpi_create_request(&((char*)recvbuf)[displs[src]], + recvcounts[src], recvtype, src, root, system_tag, + comm, &requests[index]); + if(NULL != requests[index] && MPI_SUCCESS == retval) { + smpi_mpi_irecv(requests[index]); + } + } + } + // Wait for completion of irecv's. + for(src = 0; src < size - 1; src++) { + int index = MPI_UNDEFINED; + smpi_mpi_waitany(size - 1, requests, &index, MPI_STATUS_IGNORE); + xbt_mallocator_release(smpi_global->request_mallocator, requests[index]); + } + xbt_free(requests); + } + smpi_bench_begin(); + return retval; +} + +int SMPI_MPI_Scatterv(void* sendbuf, int *sendcounts, int *displs, MPI_Datatype sendtype, + void* recvbuf, int recvcount, MPI_Datatype recvtype, + int root, MPI_Comm comm) +{ + int retval = MPI_SUCCESS; + int system_tag = 666; + int rank, size; + + smpi_bench_end(); + rank = smpi_mpi_comm_rank(comm); + size = comm->size; + if(rank != root) { + // Receive buffer from root + smpi_mpi_request_t request; + + retval = smpi_create_request(recvbuf, recvcount, recvtype, + root, rank, system_tag, comm, &request); + smpi_mpi_isend(request); + smpi_mpi_wait(request, MPI_STATUS_IGNORE); + xbt_mallocator_release(smpi_global->request_mallocator, request); + } else { + // Send buffers to receivers + int dst; + smpi_mpi_request_t* requests; + + requests = xbt_malloc((size-1) * sizeof(smpi_mpi_request_t)); + for(dst = 0; dst < size; dst++) { + if(dst == root) { + // Local copy from root + memcpy(recvbuf, &((char*)sendbuf)[displs[dst]], + sendcounts[dst]*sendtype->size*sizeof(char)); + } else { + int index = dst < root ? dst : dst - 1; + retval = smpi_create_request(&((char*)sendbuf)[displs[dst]], sendcounts[dst], sendtype, + root, dst, system_tag, comm, &requests[index]); + if(NULL != requests[index] && MPI_SUCCESS == retval) { + smpi_mpi_isend(requests[index]); + } + } + } + // Wait for completion of isend's. + for(dst = 0; dst < size - 1; dst++) { + int index = MPI_UNDEFINED; + smpi_mpi_waitany(size - 1, requests, &index, MPI_STATUS_IGNORE); + xbt_mallocator_release(smpi_global->request_mallocator, requests[index]); + } + xbt_free(requests); + } + smpi_bench_begin(); + return retval; +} + +int SMPI_MPI_Reduce_scatter(void* sendbuf, void* recvbuf, int *recvcounts, + MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) +{ + // FIXME: Suboptimal implementation + int retval = MPI_SUCCESS; + int count = 0; + int root = 0; + int i, rank; + int* displs; + + smpi_bench_end(); + rank = smpi_mpi_comm_rank(comm); + displs = xbt_new(int, comm->size); + for(i = 0; i < comm->size; i++) { + count += recvcounts[i]; + displs[i] = 0; + } + retval = smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, root, comm); + retval = SMPI_MPI_Scatterv(recvbuf, recvcounts, displs, datatype, recvbuf, recvcounts[rank], datatype, root, comm); + xbt_free(displs); + smpi_bench_begin(); + return retval; +} + +int SMPI_MPI_Allgather(void* sendbuf, int sendcount, MPI_Datatype sendtype, + void* recvbuf, int recvcount, MPI_Datatype recvtype, MPI_Comm comm) +{ + // FIXME: Suboptimal implementation + int root = 0; + int retval; + + smpi_bench_end(); + retval = SMPI_MPI_Gather(sendbuf, sendcount, sendtype, + recvbuf, recvcount, recvtype, root, comm); + if(retval == MPI_SUCCESS) { + retval = SMPI_MPI_Bcast(recvbuf, recvcount, recvtype, root, comm); + } + smpi_bench_begin(); + return retval; +} + +int SMPI_MPI_Allgatherv(void* sendbuf, int sendcount, MPI_Datatype sendtype, + void* recvbuf, int *recvcounts, int *displs, MPI_Datatype recvtype, + MPI_Comm comm) { - return ( SIMIX_get_clock() ); + // FIXME: Suboptimal implementation + int root = 0; + int last, retval; + + smpi_bench_end(); + retval = SMPI_MPI_Gatherv(sendbuf, sendcount, sendtype, + recvbuf, recvcounts, displs, recvtype, root, comm); + if(retval == MPI_SUCCESS) { + last = comm->size - 1; + retval = SMPI_MPI_Bcast(recvbuf, displs[last] + recvcounts[last], recvtype, root, comm); + } + smpi_bench_begin(); + return retval; }