X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/cac63707a2f233d0a979a67403336292eb2a3038..25d4b13f5b74c01ccf1da9cc2a0ffe934ea7da1a:/src/smpi/smpi_base.c diff --git a/src/smpi/smpi_base.c b/src/smpi/smpi_base.c index d8fd995a56..0fc8226a36 100644 --- a/src/smpi/smpi_base.c +++ b/src/smpi/smpi_base.c @@ -1,5 +1,12 @@ +/* Copyright (c) 2007, 2008, 2009, 2010. The SimGrid Team. + * All rights reserved. */ + +/* This program is free software; you can redistribute it and/or modify it + * under the terms of the license (GNU LGPL) which comes with this package. */ + #include "private.h" #include "xbt/time.h" +#include "mc/mc.h" XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_base, smpi, "Logging specific to SMPI (base)"); @@ -7,195 +14,720 @@ XBT_LOG_EXTERNAL_CATEGORY(smpi_base); XBT_LOG_EXTERNAL_CATEGORY(smpi_bench); XBT_LOG_EXTERNAL_CATEGORY(smpi_kernel); XBT_LOG_EXTERNAL_CATEGORY(smpi_mpi); +XBT_LOG_EXTERNAL_CATEGORY(smpi_mpi_dt); +XBT_LOG_EXTERNAL_CATEGORY(smpi_coll); XBT_LOG_EXTERNAL_CATEGORY(smpi_receiver); XBT_LOG_EXTERNAL_CATEGORY(smpi_sender); XBT_LOG_EXTERNAL_CATEGORY(smpi_util); -smpi_mpi_global_t smpi_mpi_global = NULL; +static int match_recv(void* a, void* b) { + MPI_Request ref = (MPI_Request)a; + MPI_Request req = (MPI_Request)b; + + xbt_assert(ref, "Cannot match recv against null reference"); + xbt_assert(req, "Cannot match recv against null request"); + return (ref->src == MPI_ANY_SOURCE || req->src == ref->src) + && (ref->tag == MPI_ANY_TAG || req->tag == ref->tag); +} + +static int match_send(void* a, void* b) { + MPI_Request ref = (MPI_Request)a; + MPI_Request req = (MPI_Request)b; -void smpi_mpi_land_func(void *a, void *b, int *length, - MPI_Datatype * datatype); + xbt_assert(ref, "Cannot match send against null reference"); + xbt_assert(req, "Cannot match send against null request"); + return (req->src == MPI_ANY_SOURCE || req->src == ref->src) + && (req->tag == MPI_ANY_TAG || req->tag == ref->tag); +} -void smpi_mpi_land_func(void *a, void *b, int *length, - MPI_Datatype * datatype) +static MPI_Request build_request(void *buf, int count, + MPI_Datatype datatype, int src, int dst, + int tag, MPI_Comm comm, unsigned flags) { - int i; - if (*datatype == smpi_mpi_global->mpi_int) { - int *x = a, *y = b; - for (i = 0; i < *length; i++) { - y[i] = x[i] && y[i]; - } - } + MPI_Request request; + + request = xbt_new(s_smpi_mpi_request_t, 1); + request->buf = buf; + // FIXME: this will have to be changed to support non-contiguous datatypes + request->size = smpi_datatype_size(datatype) * count; + request->src = src; + request->dst = dst; + request->tag = tag; + request->comm = comm; + request->action = NULL; + request->flags = flags; +#ifdef HAVE_TRACING + request->send = 0; + request->recv = 0; +#endif + return request; } -void smpi_mpi_sum_func(void *a, void *b, int *length, - MPI_Datatype * datatype); +/* MPI Low level calls */ +MPI_Request smpi_mpi_send_init(void *buf, int count, MPI_Datatype datatype, + int dst, int tag, MPI_Comm comm) +{ + MPI_Request request = + build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag, + comm, PERSISTENT | SEND); + + return request; +} -void smpi_mpi_sum_func(void *a, void *b, int *length, MPI_Datatype * datatype) +MPI_Request smpi_mpi_recv_init(void *buf, int count, MPI_Datatype datatype, + int src, int tag, MPI_Comm comm) { - int i; - if (*datatype == smpi_mpi_global->mpi_int) { - int *x = a, *y = b; - for (i = 0; i < *length; i++) { - y[i] = x[i] + y[i]; + MPI_Request request = + build_request(buf, count, datatype, src, smpi_comm_rank(comm), tag, + comm, PERSISTENT | RECV); + + return request; +} + +void smpi_mpi_start(MPI_Request request) +{ + smx_rdv_t mailbox; + int detached = 0; + + xbt_assert(!request->action, + "Cannot (re)start a non-finished communication"); + if(request->flags & RECV) { + print_request("New recv", request); + mailbox = smpi_process_mailbox(); + // FIXME: SIMIX does not yet support non-contiguous datatypes + request->action = SIMIX_req_comm_irecv(mailbox, request->buf, &request->size, &match_recv, request); + } else { + print_request("New send", request); + mailbox = smpi_process_remote_mailbox( + smpi_group_index(smpi_comm_group(request->comm), request->dst)); + // FIXME: SIMIX does not yet support non-contiguous datatypes + + if (request->size < 64*1024 ) { // eager mode => detached send (FIXME: this limit should be configurable) + void *oldbuf = request->buf; + detached = 1; + request->buf = malloc(request->size); + memcpy(request->buf,oldbuf,request->size); + XBT_DEBUG("Send request %p is detached; buf %p copied into %p",request,oldbuf,request->buf); + } else { + XBT_DEBUG("Send request %p is not detached (buf: %p)",request,request->buf); } + request->action = + SIMIX_req_comm_isend(mailbox, request->size, -1.0, + request->buf, request->size, &match_send, request, + // detach if msg size < eager/rdv switch limit + detached); + +#ifdef HAVE_TRACING + SIMIX_req_set_category (request->action, TRACE_internal_smpi_get_category()); +#endif } } -int smpi_mpi_comm_rank(smpi_mpi_communicator_t comm) +void smpi_mpi_startall(int count, MPI_Request * requests) { - return comm->index_to_rank_map[smpi_host_index()]; + int i; + + for(i = 0; i < count; i++) { + smpi_mpi_start(requests[i]); + } } -void smpi_process_init() +void smpi_mpi_request_free(MPI_Request * request) { - smx_host_t host; - int i; - smpi_host_data_t hdata; + xbt_free(*request); + *request = MPI_REQUEST_NULL; +} - smpi_global->running_hosts_count++; +MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype, + int dst, int tag, MPI_Comm comm) +{ + MPI_Request request = + build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag, + comm, NON_PERSISTENT | SEND); - // initialize some local variables - host = SIMIX_host_self(); + return request; +} - hdata = xbt_new(s_smpi_host_data_t, 1); - SIMIX_host_set_data(host, hdata); - SIMIX_process_set_data(SIMIX_process_self(),hdata); +MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype, + int dst, int tag, MPI_Comm comm) +{ + MPI_Request request = + smpi_isend_init(buf, count, datatype, dst, tag, comm); - for (i = 0; i < smpi_global->host_count && host != smpi_global->hosts[i]; i++); + smpi_mpi_start(request); + return request; +} - hdata->index = i; - hdata->mutex = SIMIX_mutex_init(); - hdata->cond = SIMIX_cond_init(); +MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype, + int src, int tag, MPI_Comm comm) +{ + MPI_Request request = + build_request(buf, count, datatype, src, smpi_comm_rank(comm), tag, + comm, NON_PERSISTENT | RECV); - hdata->pending_recv_request_queue = xbt_fifo_new(); + return request; +} - hdata->main = SIMIX_process_self(); - hdata->sender = SIMIX_process_create("smpi_sender", - smpi_sender, hdata, - SIMIX_host_get_name(SIMIX_host_self()), 0, NULL, - /*props */ NULL); - hdata->receiver = SIMIX_process_create("smpi_receiver", - smpi_receiver, hdata, - SIMIX_host_get_name(SIMIX_host_self()), 0, NULL, - /*props */ NULL); +MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype, + int src, int tag, MPI_Comm comm) +{ + MPI_Request request = + smpi_irecv_init(buf, count, datatype, src, tag, comm); - smpi_global->main_processes[hdata->index] = SIMIX_process_self(); - return; + smpi_mpi_start(request); + return request; } -void smpi_process_finalize() +void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src, + int tag, MPI_Comm comm, MPI_Status * status) { - int i; - smpi_host_data_t hdata = SIMIX_host_get_data(SIMIX_host_self()); + MPI_Request request; - i = --smpi_global->running_hosts_count; + request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm); + smpi_mpi_wait(&request, status); +} - SIMIX_mutex_destroy(smpi_host_mutex()); - SIMIX_cond_destroy(smpi_host_cond()); - xbt_fifo_free(hdata->pending_recv_request_queue); +void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst, + int tag, MPI_Comm comm) +{ + MPI_Request request; - if (0 >= i) { + request = smpi_mpi_isend(buf, count, datatype, dst, tag, comm); + smpi_mpi_wait(&request, MPI_STATUS_IGNORE); +} - // wake up senders/receivers - /* MQ: (FIXME) Don't do so: it breaks since some hosts are already gone - for (i = 0; i < smpi_global->host_count; i++) { - smpi_host_data_t remote_hdata = SIMIX_process_get_data(smpi_global->main_processes[i]); +void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype, + int dst, int sendtag, void *recvbuf, int recvcount, + MPI_Datatype recvtype, int src, int recvtag, + MPI_Comm comm, MPI_Status * status) +{ + MPI_Request requests[2]; + MPI_Status stats[2]; + + requests[0] = + smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm); + requests[1] = + smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm); + smpi_mpi_startall(2, requests); + smpi_mpi_waitall(2, requests, stats); + if(status != MPI_STATUS_IGNORE) { + // Copy receive status + memcpy(status, &stats[1], sizeof(MPI_Status)); + } +} - if (SIMIX_process_is_suspended(remote_hdata->sender)) - SIMIX_process_resume(remote_hdata->sender); +int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype) +{ + return status->count / smpi_datatype_size(datatype); +} - if (SIMIX_process_is_suspended(remote_hdata->receiver)) - SIMIX_process_resume(remote_hdata->receiver); - }*/ +static void finish_wait(MPI_Request * request, MPI_Status * status) +{ + MPI_Request req = *request; + + if(status != MPI_STATUS_IGNORE) { + status->MPI_SOURCE = req->src; + status->MPI_TAG = req->tag; + status->MPI_ERROR = MPI_SUCCESS; + // FIXME: really this should just contain the count of receive-type blocks, + // right? + status->count = req->size; + } + print_request("Finishing", req); + if(req->flags & NON_PERSISTENT) { + smpi_mpi_request_free(request); + } else { + req->action = NULL; + } +} - SIMIX_mutex_destroy(smpi_mpi_global->mpi_comm_world->barrier_mutex); - SIMIX_cond_destroy(smpi_mpi_global->mpi_comm_world->barrier_cond); - xbt_free(smpi_mpi_global->mpi_comm_world); +int smpi_mpi_test(MPI_Request * request, MPI_Status * status) { +int flag; + + if ((*request)->action == NULL) + flag = 1; + else + flag = SIMIX_req_comm_test((*request)->action); + if(flag) { + smpi_mpi_wait(request, status); + } + return flag; +} - xbt_free(smpi_mpi_global->mpi_byte); - xbt_free(smpi_mpi_global->mpi_int); - xbt_free(smpi_mpi_global->mpi_double); +int smpi_mpi_testany(int count, MPI_Request requests[], int *index, + MPI_Status * status) +{ + xbt_dynar_t comms; + int i, flag, size; + int* map; + + *index = MPI_UNDEFINED; + flag = 0; + if(count > 0) { + comms = xbt_dynar_new(sizeof(smx_action_t), NULL); + map = xbt_new(int, count); + size = 0; + for(i = 0; i < count; i++) { + if(requests[i]->action) { + xbt_dynar_push(comms, &requests[i]->action); + map[size] = i; + size++; + } + } + if(size > 0) { + i = SIMIX_req_comm_testany(comms); + // FIXME: MPI_UNDEFINED or does SIMIX have a return code? + if(i != MPI_UNDEFINED) { + *index = map[i]; + smpi_mpi_wait(&requests[*index], status); + flag = 1; + } + } + xbt_free(map); + xbt_dynar_free(&comms); + } + return flag; +} - xbt_free(smpi_mpi_global->mpi_land); - xbt_free(smpi_mpi_global->mpi_sum); +void smpi_mpi_wait(MPI_Request * request, MPI_Status * status) +{ + print_request("Waiting", *request); + if ((*request)->action != NULL ) + SIMIX_req_comm_wait((*request)->action, -1.0); + finish_wait(request, status); +} - xbt_free(smpi_mpi_global); +int smpi_mpi_waitany(int count, MPI_Request requests[], + MPI_Status * status) +{ + xbt_dynar_t comms; + int i, size, index; + int *map; + + index = MPI_UNDEFINED; + if(count > 0) { + // Wait for a request to complete + comms = xbt_dynar_new(sizeof(smx_action_t), NULL); + map = xbt_new(int, count); + size = 0; + XBT_DEBUG("Wait for one of"); + for(i = 0; i < count; i++) { + if((requests[i] != MPI_REQUEST_NULL) && (requests[i]->action != NULL)) { + print_request(" ", requests[i]); + xbt_dynar_push(comms, &requests[i]->action); + map[size] = i; + size++; + } + } + if(size > 0) { + i = SIMIX_req_comm_waitany(comms); + // FIXME: MPI_UNDEFINED or does SIMIX have a return code? + if (i != MPI_UNDEFINED) { + index = map[i]; + finish_wait(&requests[index], status); + } + } + xbt_free(map); + xbt_dynar_free(&comms); } + return index; } -int smpi_mpi_barrier(smpi_mpi_communicator_t comm) +void smpi_mpi_waitall(int count, MPI_Request requests[], + MPI_Status status[]) { - - SIMIX_mutex_lock(comm->barrier_mutex); - ++comm->barrier_count; - if (comm->barrier_count > comm->size) { // only happens on second barrier... - comm->barrier_count = 0; - } else if (comm->barrier_count == comm->size) { - SIMIX_cond_broadcast(comm->barrier_cond); + int index, c; + MPI_Status stat; + MPI_Status *pstat = status == MPI_STATUS_IGNORE ? MPI_STATUS_IGNORE : &stat; + + for(c = 0; c < count; c++) { + if(MC_IS_ENABLED) { + smpi_mpi_wait(&requests[c], pstat); + index = c; + } else { + index = smpi_mpi_waitany(count, requests, pstat); + if(index == MPI_UNDEFINED) { + break; + } + } + if(status != MPI_STATUS_IGNORE) { + memcpy(&status[index], pstat, sizeof(*pstat)); + } } - while (comm->barrier_count < comm->size) { - SIMIX_cond_wait(comm->barrier_cond, comm->barrier_mutex); +} + +int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices, + MPI_Status status[]) +{ + int i, count, index; + + count = 0; + for(i = 0; i < incount; i++) { + if(smpi_mpi_testany(incount, requests, &index, status)) { + indices[count] = index; + count++; + } } - SIMIX_mutex_unlock(comm->barrier_mutex); + return count; +} - return MPI_SUCCESS; +void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root, + MPI_Comm comm) +{ + // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI)) + nary_tree_bcast(buf, count, datatype, root, comm, 4); } -int smpi_mpi_isend(smpi_mpi_request_t request) +void smpi_mpi_barrier(MPI_Comm comm) { - smpi_host_data_t hdata = SIMIX_host_get_data(SIMIX_host_self()); - int retval = MPI_SUCCESS; + // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI)) + nary_tree_barrier(comm, 4); +} - if (NULL == request) { - retval = MPI_ERR_INTERN; +void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype, + void *recvbuf, int recvcount, MPI_Datatype recvtype, + int root, MPI_Comm comm) +{ + int system_tag = 666; + int rank, size, src, index; + MPI_Aint lb = 0, recvext = 0; + MPI_Request *requests; + + rank = smpi_comm_rank(comm); + size = smpi_comm_size(comm); + if(rank != root) { + // Send buffer to root + smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm); } else { - xbt_fifo_push(smpi_global->pending_send_request_queues[hdata->index], request); + // FIXME: check for errors + smpi_datatype_extent(recvtype, &lb, &recvext); + // Local copy from root + smpi_datatype_copy(sendbuf, sendcount, sendtype, + (char *)recvbuf + root * recvcount * recvext, recvcount, recvtype); + // Receive buffers from senders + requests = xbt_new(MPI_Request, size - 1); + index = 0; + for(src = 0; src < size; src++) { + if(src != root) { + requests[index] = smpi_irecv_init((char *)recvbuf + src * recvcount * recvext, + recvcount, recvtype, + src, system_tag, comm); + index++; + } + } + // Wait for completion of irecv's. + smpi_mpi_startall(size - 1, requests); + smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE); + xbt_free(requests); + } +} - if (SIMIX_process_is_suspended(hdata->sender)) { - SIMIX_process_resume(hdata->sender); +void 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 system_tag = 666; + int rank, size, src, index; + MPI_Aint lb = 0, recvext = 0; + MPI_Request *requests; + + rank = smpi_comm_rank(comm); + size = smpi_comm_size(comm); + if(rank != root) { + // Send buffer to root + smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm); + } else { + // FIXME: check for errors + smpi_datatype_extent(recvtype, &lb, &recvext); + // Local copy from root + smpi_datatype_copy(sendbuf, sendcount, sendtype, + (char *)recvbuf + displs[root] * recvext, + recvcounts[root], recvtype); + // Receive buffers from senders + requests = xbt_new(MPI_Request, size - 1); + index = 0; + for(src = 0; src < size; src++) { + if(src != root) { + requests[index] = + smpi_irecv_init((char *)recvbuf + displs[src] * recvext, + recvcounts[src], recvtype, src, system_tag, comm); + index++; + } } + // Wait for completion of irecv's. + smpi_mpi_startall(size - 1, requests); + smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE); + xbt_free(requests); } +} - return retval; +void smpi_mpi_allgather(void *sendbuf, int sendcount, + MPI_Datatype sendtype, void *recvbuf, + int recvcount, MPI_Datatype recvtype, + MPI_Comm comm) +{ + int system_tag = 666; + int rank, size, other, index; + MPI_Aint lb = 0, recvext = 0; + MPI_Request *requests; + + rank = smpi_comm_rank(comm); + size = smpi_comm_size(comm); + // FIXME: check for errors + smpi_datatype_extent(recvtype, &lb, &recvext); + // Local copy from self + smpi_datatype_copy(sendbuf, sendcount, sendtype, + (char *)recvbuf + rank * recvcount * recvext, recvcount, + recvtype); + // Send/Recv buffers to/from others; + requests = xbt_new(MPI_Request, 2 * (size - 1)); + index = 0; + for(other = 0; other < size; other++) { + if(other != rank) { + requests[index] = + smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag, + comm); + index++; + requests[index] = smpi_irecv_init((char *)recvbuf + other * recvcount * recvext, + recvcount, recvtype, other, + system_tag, comm); + index++; + } + } + // Wait for completion of all comms. + smpi_mpi_startall(2 * (size - 1), requests); + smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE); + xbt_free(requests); } -int smpi_mpi_irecv(smpi_mpi_request_t request) +void smpi_mpi_allgatherv(void *sendbuf, int sendcount, + MPI_Datatype sendtype, void *recvbuf, + int *recvcounts, int *displs, + MPI_Datatype recvtype, MPI_Comm comm) { - int retval = MPI_SUCCESS; - smpi_host_data_t hdata = SIMIX_host_get_data(SIMIX_host_self()); + int system_tag = 666; + int rank, size, other, index; + MPI_Aint lb = 0, recvext = 0; + MPI_Request *requests; + + rank = smpi_comm_rank(comm); + size = smpi_comm_size(comm); + // FIXME: check for errors + smpi_datatype_extent(recvtype, &lb, &recvext); + // Local copy from self + smpi_datatype_copy(sendbuf, sendcount, sendtype, + (char *)recvbuf + displs[rank] * recvext, + recvcounts[rank], recvtype); + // Send buffers to others; + requests = xbt_new(MPI_Request, 2 * (size - 1)); + index = 0; + for(other = 0; other < size; other++) { + if(other != rank) { + requests[index] = + smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag, + comm); + index++; + requests[index] = + smpi_irecv_init((char *)recvbuf + displs[other] * recvext, recvcounts[other], + recvtype, other, system_tag, comm); + index++; + } + } + // Wait for completion of all comms. + smpi_mpi_startall(2 * (size - 1), requests); + smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE); + xbt_free(requests); +} - if (NULL == request) { - retval = MPI_ERR_INTERN; +void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype, + void *recvbuf, int recvcount, MPI_Datatype recvtype, + int root, MPI_Comm comm) +{ + int system_tag = 666; + int rank, size, dst, index; + MPI_Aint lb = 0, sendext = 0; + MPI_Request *requests; + + rank = smpi_comm_rank(comm); + size = smpi_comm_size(comm); + if(rank != root) { + // Recv buffer from root + smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm, + MPI_STATUS_IGNORE); } else { - xbt_fifo_push(hdata->pending_recv_request_queue, request); - - if (SIMIX_process_is_suspended(hdata->receiver)) { - SIMIX_process_resume(hdata->receiver); + // FIXME: check for errors + smpi_datatype_extent(sendtype, &lb, &sendext); + // Local copy from root + smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext, + sendcount, sendtype, recvbuf, recvcount, recvtype); + // Send buffers to receivers + requests = xbt_new(MPI_Request, size - 1); + index = 0; + for(dst = 0; dst < size; dst++) { + if(dst != root) { + requests[index] = smpi_isend_init((char *)sendbuf + dst * sendcount * sendext, + sendcount, sendtype, dst, + system_tag, comm); + index++; + } } + // Wait for completion of isend's. + smpi_mpi_startall(size - 1, requests); + smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE); + xbt_free(requests); } - - return retval; } -int smpi_mpi_wait(smpi_mpi_request_t request, smpi_mpi_status_t * status) +void 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, dst, index; + MPI_Aint lb = 0, sendext = 0; + MPI_Request *requests; + + rank = smpi_comm_rank(comm); + size = smpi_comm_size(comm); + if(rank != root) { + // Recv buffer from root + smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm, + MPI_STATUS_IGNORE); + } else { + // FIXME: check for errors + smpi_datatype_extent(sendtype, &lb, &sendext); + // Local copy from root + smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root], + sendtype, recvbuf, recvcount, recvtype); + // Send buffers to receivers + requests = xbt_new(MPI_Request, size - 1); + index = 0; + for(dst = 0; dst < size; dst++) { + if(dst != root) { + requests[index] = + smpi_isend_init((char *)sendbuf + displs[dst] * sendext, sendcounts[dst], + sendtype, dst, system_tag, comm); + index++; + } + } + // Wait for completion of isend's. + smpi_mpi_startall(size - 1, requests); + smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE); + xbt_free(requests); + } +} - if (NULL == request) { - retval = MPI_ERR_INTERN; +void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count, + MPI_Datatype datatype, MPI_Op op, int root, + MPI_Comm comm) +{ + int system_tag = 666; + int rank, size, src, index; + MPI_Aint lb = 0, dataext = 0; + MPI_Request *requests; + void **tmpbufs; + + rank = smpi_comm_rank(comm); + size = smpi_comm_size(comm); + if(rank != root) { + // Send buffer to root + smpi_mpi_send(sendbuf, count, datatype, root, system_tag, comm); } else { - SIMIX_mutex_lock(request->mutex); - while (!request->completed) { - SIMIX_cond_wait(request->cond, request->mutex); + // FIXME: check for errors + smpi_datatype_extent(datatype, &lb, &dataext); + // Local copy from root + smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype); + // Receive buffers from senders + //TODO: make a MPI_barrier here ? + requests = xbt_new(MPI_Request, size - 1); + tmpbufs = xbt_new(void *, size - 1); + index = 0; + for(src = 0; src < size; src++) { + if(src != root) { + // FIXME: possibly overkill we we have contiguous/noncontiguous data + // mapping... + tmpbufs[index] = xbt_malloc(count * dataext); + requests[index] = + smpi_irecv_init(tmpbufs[index], count, datatype, src, + system_tag, comm); + index++; + } + } + // Wait for completion of irecv's. + smpi_mpi_startall(size - 1, requests); + for(src = 0; src < size - 1; src++) { + index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE); + if(index == MPI_UNDEFINED) { + break; + } + smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype); } - if (NULL != status) { - status->MPI_SOURCE = request->src; - status->MPI_TAG = request->tag; - status->MPI_ERROR = MPI_SUCCESS; + for(index = 0; index < size - 1; index++) { + xbt_free(tmpbufs[index]); } - SIMIX_mutex_unlock(request->mutex); + xbt_free(tmpbufs); + xbt_free(requests); } +} + +void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count, + MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) +{ + smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm); + smpi_mpi_bcast(recvbuf, count, datatype, 0, comm); +} - return retval; +void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count, + MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) +{ + int system_tag = 666; + int rank, size, other, index; + MPI_Aint lb = 0, dataext = 0; + MPI_Request *requests; + void **tmpbufs; + + rank = smpi_comm_rank(comm); + size = smpi_comm_size(comm); + + // FIXME: check for errors + smpi_datatype_extent(datatype, &lb, &dataext); + + // Local copy from self + smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype); + + // Send/Recv buffers to/from others; + requests = xbt_new(MPI_Request, size - 1); + tmpbufs = xbt_new(void *, rank); + index = 0; + for(other = 0; other < rank; other++) { + // FIXME: possibly overkill we we have contiguous/noncontiguous data + // mapping... + tmpbufs[index] = xbt_malloc(count * dataext); + requests[index] = + smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag, + comm); + index++; + } + for(other = rank + 1; other < size; other++) { + requests[index] = + smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm); + index++; + } + // Wait for completion of all comms. + smpi_mpi_startall(size - 1, requests); + for(other = 0; other < size - 1; other++) { + index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE); + if(index == MPI_UNDEFINED) { + break; + } + if(index < rank) { + // #Request is below rank: it's a irecv + smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype); + } + } + for(index = 0; index < rank; index++) { + xbt_free(tmpbufs[index]); + } + xbt_free(tmpbufs); + xbt_free(requests); }