X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/d30fcaa41c0aabda5afef028a793b9e295a4aadb..5a185bb60f4c7b77ebccc4948d17f0dc43a9d6a5:/src/smpi/smpi_base.c diff --git a/src/smpi/smpi_base.c b/src/smpi/smpi_base.c index e5af68d34c..2bc6b976de 100644 --- a/src/smpi/smpi_base.c +++ b/src/smpi/smpi_base.c @@ -1,379 +1,787 @@ +/* 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" +#include "xbt/replay.h" +#include XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_base, smpi, "Logging specific to SMPI (base)"); -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_receiver); -XBT_LOG_EXTERNAL_CATEGORY(smpi_sender); -XBT_LOG_EXTERNAL_CATEGORY(smpi_util); - -smpi_mpi_global_t smpi_mpi_global = NULL; - -/** - * Operations of MPI_OP : implemented=land,sum,min,max - **/ -void smpi_mpi_land_func(void *a, void *b, int *length, - MPI_Datatype * datatype); - -void smpi_mpi_land_func(void *a, void *b, int *length, - MPI_Datatype * datatype) -{ - 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]; - } - } + +static int match_recv(void* a, void* b, smx_action_t ignored) { + 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); } -/** - * sum two vectors element-wise - * - * @param a the first vectors - * @param b the second vectors - * @return the second vector is modified and contains the element-wise sums - **/ -void smpi_mpi_sum_func(void *a, void *b, int *length, - MPI_Datatype * datatype); +static int match_send(void* a, void* b,smx_action_t ignored) { + MPI_Request ref = (MPI_Request)a; + MPI_Request req = (MPI_Request)b; + + 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_sum_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_byte) { - char *x = a, *y = b; - for (i = 0; i < *length; i++) { - y[i] = x[i] + y[i]; - } - } else if (*datatype == smpi_mpi_global->mpi_int) { - int *x = a, *y = b; - for (i = 0; i < *length; i++) { - y[i] = x[i] + y[i]; - } - } else if (*datatype == smpi_mpi_global->mpi_float) { - float *x = a, *y = b; - for (i = 0; i < *length; i++) { - y[i] = x[i] + y[i]; - } - } else if (*datatype == smpi_mpi_global->mpi_double) { - double *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_action_trace_run(char *path) +{ + char *name; + xbt_dynar_t todo; + xbt_dict_cursor_t cursor; + + action_fp=NULL; + if (path) { + action_fp = fopen(path, "r"); + xbt_assert(action_fp != NULL, "Cannot open %s: %s", path, + strerror(errno)); + } + + if (!xbt_dict_is_empty(action_queues)) { + XBT_WARN + ("Not all actions got consumed. If the simulation ended successfully (without deadlock), you may want to add new processes to your deployment file."); + + + xbt_dict_foreach(action_queues, cursor, name, todo) { + XBT_WARN("Still %lu actions for %s", xbt_dynar_length(todo), name); } } + + if (path) + fclose(action_fp); + xbt_dict_free(&action_queues); + action_queues = xbt_dict_new_homogeneous(NULL); } -/** - * compute the min of two vectors element-wise - **/ -void smpi_mpi_min_func(void *a, void *b, int *length, - MPI_Datatype * datatype); +static void smpi_mpi_request_free_voidp(void* request) +{ + MPI_Request req = request; + smpi_mpi_request_free(&req); +} -void smpi_mpi_min_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) { - int i; - if (*datatype == smpi_mpi_global->mpi_byte) { - char *x = a, *y = b; - for (i = 0; i < *length; i++) { - y[i] = x[i] < y[i] ? x[i] : y[i]; - } - } else { - if (*datatype == smpi_mpi_global->mpi_int) { - int *x = a, *y = b; - for (i = 0; i < *length; i++) { - y[i] = x[i] < y[i] ? x[i] : y[i]; - } - } else { - if (*datatype == smpi_mpi_global->mpi_float) { - float *x = a, *y = b; - for (i = 0; i < *length; i++) { - y[i] = x[i] < y[i] ? x[i] : y[i]; - } - } else { - if (*datatype == smpi_mpi_global->mpi_double) { - double *x = a, *y = b; - for (i = 0; i < *length; i++) { - y[i] = x[i] < y[i] ? x[i] : y[i]; - } - - } - } - } - } + MPI_Request request = + build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag, + comm, PERSISTENT | SEND); + + return request; } -/** - * compute the max of two vectors element-wise - **/ -void smpi_mpi_max_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) +{ + MPI_Request request = + build_request(buf, count, datatype, src, smpi_comm_rank(comm), tag, + comm, PERSISTENT | RECV); + + return request; +} -void smpi_mpi_max_func(void *a, void *b, int *length, MPI_Datatype * datatype) +void smpi_mpi_start(MPI_Request request) { - int i; - if (*datatype == smpi_mpi_global->mpi_byte) { - char *x = a, *y = b; - for (i = 0; i < *length; i++) { - y[i] = x[i] > y[i] ? x[i] : y[i]; - } - } else if (*datatype == smpi_mpi_global->mpi_int) { - int *x = a, *y = b; - for (i = 0; i < *length; i++) { - y[i] = x[i] > y[i] ? x[i] : y[i]; + 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); + if (request->size < xbt_cfg_get_int(_surf_cfg_set, "smpi/async_small_thres")) + mailbox = smpi_process_mailbox_small(); + else + mailbox = smpi_process_mailbox(); + + // FIXME: SIMIX does not yet support non-contiguous datatypes + request->action = simcall_comm_irecv(mailbox, request->buf, &request->size, &match_recv, request); + } else { + print_request("New send", request); + + if (request->size < xbt_cfg_get_int(_surf_cfg_set, "smpi/async_small_thres")) { // eager mode => detached send (FIXME: this limit should be configurable) + mailbox = smpi_process_remote_mailbox_small( + smpi_group_index(smpi_comm_group(request->comm), request->dst)); + }else{ + XBT_DEBUG("Send request %p is not in the permanent receive mailbox (buf: %p)",request,request->buf); + mailbox = smpi_process_remote_mailbox( + smpi_group_index(smpi_comm_group(request->comm), request->dst)); } - } else if (*datatype == smpi_mpi_global->mpi_float) { - float *x = a, *y = b; - for (i = 0; i < *length; i++) { - y[i] = x[i] > y[i] ? x[i] : y[i]; + if (request->size < 64*1024 ) { //(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); + mailbox = smpi_process_remote_mailbox( + smpi_group_index(smpi_comm_group(request->comm), request->dst)); } - } else if (*datatype == smpi_mpi_global->mpi_double) { - double *x = a, *y = b; - for (i = 0; i < *length; i++) { - y[i] = x[i] > y[i] ? x[i] : y[i]; + + request->action = + simcall_comm_isend(mailbox, request->size, -1.0, + request->buf, request->size, + &match_send, + &smpi_mpi_request_free_voidp, // how to free the userdata if a detached send fails + request, + // detach if msg size < eager/rdv switch limit + detached); + + #ifdef HAVE_TRACING + /* FIXME: detached sends are not traceable (request->action == NULL) */ + if (request->action) + simcall_set_category(request->action, TRACE_internal_smpi_get_category()); + #endif + } - } } +void smpi_mpi_startall(int count, MPI_Request * requests) +{ + int i; + for(i = 0; i < count; i++) { + smpi_mpi_start(requests[i]); + } +} +void smpi_mpi_request_free(MPI_Request * request) +{ + xbt_free(*request); + *request = MPI_REQUEST_NULL; +} -/** - * tell the MPI rank of the calling process (from its SIMIX process id) - **/ -int smpi_mpi_comm_rank(smpi_mpi_communicator_t comm) +MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype, + int dst, int tag, MPI_Comm comm) { - return comm->index_to_rank_map[smpi_process_index()]; + MPI_Request request = + build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag, + comm, NON_PERSISTENT | SEND); + + return request; } -void smpi_process_init(int *argc, char ***argv) +MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype, + int dst, int tag, MPI_Comm comm) { - smpi_process_data_t pdata; + MPI_Request request = + smpi_isend_init(buf, count, datatype, dst, tag, comm); - // initialize some local variables + smpi_mpi_start(request); + return request; +} - pdata = xbt_new(s_smpi_process_data_t, 1); - SIMIX_process_set_data(SIMIX_process_self(), pdata); +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); - /* get rank from command line, and remove it from argv */ - pdata->index = atoi((*argv)[1]); - DEBUG1("I'm rank %d", pdata->index); - if (*argc > 2) { - memmove((*argv)[1], (*argv)[2], sizeof(char *) * (*argc - 2)); - (*argv)[(*argc) - 1] = NULL; - } - (*argc)--; + return request; +} - pdata->mutex = SIMIX_mutex_init(); - pdata->cond = SIMIX_cond_init(); - pdata->finalize = 0; +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); - pdata->pending_recv_request_queue = xbt_fifo_new(); - pdata->pending_send_request_queue = xbt_fifo_new(); - pdata->received_message_queue = xbt_fifo_new(); + smpi_mpi_start(request); + return request; +} - pdata->main = SIMIX_process_self(); - pdata->sender = SIMIX_process_create("smpi_sender", - smpi_sender, pdata, - SIMIX_host_get_name(SIMIX_host_self()), - 0, NULL, - /*props */ NULL); - pdata->receiver = SIMIX_process_create("smpi_receiver", - smpi_receiver, pdata, - SIMIX_host_get_name(SIMIX_host_self - ()), 0, NULL, - /*props */ NULL); +void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src, + int tag, MPI_Comm comm, MPI_Status * status) +{ + MPI_Request request; - smpi_global->main_processes[pdata->index] = SIMIX_process_self(); - return; + request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm); + smpi_mpi_wait(&request, status); } -void smpi_process_finalize() + + +void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst, + int tag, MPI_Comm comm) { - smpi_process_data_t pdata = SIMIX_process_get_data(SIMIX_process_self()); + MPI_Request request; + + request = smpi_mpi_isend(buf, count, datatype, dst, tag, comm); + smpi_mpi_wait(&request, MPI_STATUS_IGNORE); +} - pdata->finalize = 2; /* Tell sender and receiver to quit */ - SIMIX_process_resume(pdata->sender); - SIMIX_process_resume(pdata->receiver); - while (pdata->finalize > 0) { /* wait until it's done */ - SIMIX_cond_wait(pdata->cond, pdata->mutex); +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)); } +} - SIMIX_mutex_destroy(pdata->mutex); - SIMIX_cond_destroy(pdata->cond); - xbt_fifo_free(pdata->pending_recv_request_queue); - xbt_fifo_free(pdata->pending_send_request_queue); - xbt_fifo_free(pdata->received_message_queue); - xbt_free(pdata); +int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype) +{ + return status->count / smpi_datatype_size(datatype); } -int smpi_mpi_barrier(smpi_mpi_communicator_t comm) +static void finish_wait(MPI_Request * request, MPI_Status * status) { + MPI_Request req = *request; - 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); + 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; } - while (comm->barrier_count < comm->size) { - SIMIX_cond_wait(comm->barrier_cond, comm->barrier_mutex); + print_request("Finishing", req); + if(req->flags & NON_PERSISTENT) { + smpi_mpi_request_free(request); + } else { + req->action = NULL; } - SIMIX_mutex_unlock(comm->barrier_mutex); +} - return MPI_SUCCESS; +int smpi_mpi_test(MPI_Request * request, MPI_Status * status) { +int flag; + + if ((*request)->action == NULL) + flag = 1; + else + flag = simcall_comm_test((*request)->action); + if(flag) { + smpi_mpi_wait(request, status); + } + return flag; } -int smpi_mpi_isend(smpi_mpi_request_t request) +int smpi_mpi_testany(int count, MPI_Request requests[], int *index, + MPI_Status * status) { - smpi_process_data_t pdata = SIMIX_process_get_data(SIMIX_process_self()); - int retval = MPI_SUCCESS; + xbt_dynar_t comms; + int i, flag, size; + int* map; - if (NULL == request) { - retval = MPI_ERR_INTERN; - } else { - xbt_fifo_push(pdata->pending_send_request_queue, request); - SIMIX_process_resume(pdata->sender); + *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 = simcall_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; +} - return retval; +void smpi_mpi_wait(MPI_Request * request, MPI_Status * status) +{ + print_request("Waiting", *request); + if ((*request)->action != NULL) { // this is not a detached send + simcall_comm_wait((*request)->action, -1.0); + finish_wait(request, status); + } + // FIXME for a detached send, finish_wait is not called: } -int smpi_mpi_irecv(smpi_mpi_request_t request) +int smpi_mpi_waitany(int count, MPI_Request requests[], + MPI_Status * status) { - int retval = MPI_SUCCESS; - smpi_process_data_t pdata = SIMIX_process_get_data(SIMIX_process_self()); + 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("Waiting any ", requests[i]); + xbt_dynar_push(comms, &requests[i]->action); + map[size] = i; + size++; + } + } + if(size > 0) { + i = simcall_comm_waitany(comms); - if (NULL == request) { - retval = MPI_ERR_INTERN; - } else { - xbt_fifo_push(pdata->pending_recv_request_queue, request); + // FIXME: MPI_UNDEFINED or does SIMIX have a return code? + if (i != MPI_UNDEFINED) { + index = map[i]; + finish_wait(&requests[index], status); - if (SIMIX_process_is_suspended(pdata->receiver)) { - SIMIX_process_resume(pdata->receiver); + } } + xbt_free(map); + xbt_dynar_free(&comms); } + return index; +} - return retval; +void smpi_mpi_waitall(int count, MPI_Request requests[], + MPI_Status status[]) +{ + 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)); + } + } } -void print_req( smpi_mpi_request_t r ); -void print_req( smpi_mpi_request_t r ) { - printf("***req %p-> src=%d dst=%d tag=%d completed=0x%x consumed=0x%x\n",r,r->src,r->dst,r->tag,r->completed,r->consumed); +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++; + } + } + return count; } +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); +} -/** - * wait and friends ... - **/ -int smpi_mpi_wait(smpi_mpi_request_t request, smpi_mpi_status_t * status) +void smpi_mpi_barrier(MPI_Comm comm) { - 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 { - SIMIX_mutex_lock(request->mutex); - -#ifdef DEBUG_STEPH - print_req( request ); //@@ -#endif - while (!request->completed) { - SIMIX_cond_wait(request->cond, request->mutex); - } - if (NULL != status) { - status->MPI_SOURCE = request->src; - status->MPI_TAG = request->tag; - status->MPI_ERROR = MPI_SUCCESS; + // 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++; + } } - SIMIX_mutex_unlock(request->mutex); + // 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; } -/** - * waitall - **/ -int smpi_mpi_waitall(int count, smpi_mpi_request_t requests[], - smpi_mpi_status_t status[]) +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 cpt; - int index; - int retval; - smpi_mpi_status_t stat; - - for (cpt = 0; cpt < count; cpt++) { - retval = smpi_mpi_waitany(count, requests, &index, &stat); - if (retval != MPI_SUCCESS) - return retval; - if (MPI_STATUS_IGNORE != status) - memcpy(&(status[index]), &stat, sizeof(stat)); - } - return MPI_SUCCESS; + 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); + } } -/** - * waitany - **/ -int smpi_mpi_waitany(int count, smpi_mpi_request_t * requests, int *index, - smpi_mpi_status_t * status) +void smpi_mpi_allgather(void *sendbuf, int sendcount, + MPI_Datatype sendtype, void *recvbuf, + int recvcount, MPI_Datatype recvtype, + MPI_Comm comm) { - int cpt; - - *index = MPI_UNDEFINED; - if (NULL == requests) { - return MPI_ERR_INTERN; + 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++; + } } - /* First check if one of them is already done */ - for (cpt = 0; cpt < count; cpt++) { - printf("...exam req[%d] of msg from [%d]\n",cpt,requests[cpt]->src); - if (requests[cpt]->completed && !requests[cpt]->consumed) { /* got ya */ - printf("...found match req[%d] of msg from [%d]\n",cpt,requests[cpt]->src); - *index = cpt; - goto found_request; + // 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); +} + +void smpi_mpi_allgatherv(void *sendbuf, int sendcount, + MPI_Datatype sendtype, void *recvbuf, + int *recvcounts, int *displs, + 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 + 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++; } } - /* If none found, block */ - /* FIXME: should use a SIMIX_cond_waitany, when implemented. For now, block on the first one */ - while (1) { - for (cpt = 0; cpt < count; cpt++) { + // 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); +} -#ifdef DEBUG_STEPH - print_req( requests[cpt] ); -#endif - if (!requests[cpt]->completed) { /* this one is not done, wait on it */ - printf("... blocked waiting a msg %d->%d, tag=%d\n",requests[cpt]->src,requests[cpt]->dst,requests[cpt]->tag); - while (!requests[cpt]->completed) - SIMIX_cond_wait(requests[cpt]->cond, requests[cpt]->mutex); +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 { + // 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); + } +} - *index = cpt; - goto found_request; +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 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++; } } - if (cpt == count) /* they are all done. Damn user */ - return MPI_ERR_REQUEST; + // Wait for completion of isend's. + smpi_mpi_startall(size - 1, requests); + smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE); + xbt_free(requests); } +} -found_request: -#ifdef DEBUG_STEPH - print_req( requests[cpt] ); -#endif - requests[*index]->consumed = 1; -#ifdef DEBUG_STEPH - print_req( requests[cpt] ); -#endif - printf("...accessing *req[%d]->consumed\n",cpt); - if (NULL != status) { - status->MPI_SOURCE = requests[*index]->src; - status->MPI_TAG = requests[*index]->tag; - status->MPI_ERROR = MPI_SUCCESS; +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 { + // 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); + XBT_VERB("finished waiting any request with index %d", index); + if(index == MPI_UNDEFINED) { + break; + } + smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype); + } + for(index = 0; index < size - 1; index++) { + xbt_free(tmpbufs[index]); + } + xbt_free(tmpbufs); + xbt_free(requests); } - return MPI_SUCCESS; +} +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); +} + +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); }