X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/b085fe247ae2cf2a6e36863a54f794deb4f342f5..f9e6853d2c5b7b867211700bc12f0ee57f640b30:/src/smpi/smpi_base.c diff --git a/src/smpi/smpi_base.c b/src/smpi/smpi_base.c index 067cd82af0..6295038a3c 100644 --- a/src/smpi/smpi_base.c +++ b/src/smpi/smpi_base.c @@ -1,4 +1,4 @@ -/* Copyright (c) 2007, 2008, 2009, 2010. The SimGrid Team. +/* Copyright (c) 2007-2014. The SimGrid Team. * All rights reserved. */ /* This program is free software; you can redistribute it and/or modify it @@ -12,7 +12,7 @@ #include "simix/smx_private.h" #include "surf/surf.h" #include "simgrid/sg_config.h" - +#include "colls/colls.h" XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_base, smpi, "Logging specific to SMPI (base)"); @@ -25,7 +25,7 @@ static int match_recv(void* a, void* b, smx_action_t ignored) { xbt_assert(ref, "Cannot match recv against null reference"); xbt_assert(req, "Cannot match recv against null request"); if((ref->src == MPI_ANY_SOURCE || req->src == ref->src) - && (ref->tag == MPI_ANY_TAG || req->tag == ref->tag)){ + && ((ref->tag == MPI_ANY_TAG && req->tag >=0) || req->tag == ref->tag)){ //we match, we can transfer some values // FIXME : move this to the copy function ? if(ref->src == MPI_ANY_SOURCE)ref->real_src = req->src; @@ -34,6 +34,7 @@ static int match_recv(void* a, void* b, smx_action_t ignored) { if(req->detached==1){ ref->detached_sender=req; //tie the sender to the receiver, as it is detached and has to be freed in the receiver } + XBT_DEBUG("match succeeded"); return 1; }else return 0; } @@ -46,7 +47,7 @@ static int match_send(void* a, void* b,smx_action_t ignored) { xbt_assert(req, "Cannot match send against null request"); if((req->src == MPI_ANY_SOURCE || req->src == ref->src) - && (req->tag == MPI_ANY_TAG || req->tag == ref->tag)) + && ((req->tag == MPI_ANY_TAG && ref->tag >=0)|| req->tag == ref->tag)) { if(req->src == MPI_ANY_SOURCE)req->real_src = ref->src; if(req->tag == MPI_ANY_TAG)req->real_tag = ref->tag; @@ -54,7 +55,7 @@ static int match_send(void* a, void* b,smx_action_t ignored) { if(ref->detached==1){ req->detached_sender=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver } - + XBT_DEBUG("match succeeded"); return 1; } else return 0; } @@ -76,7 +77,8 @@ xbt_dynar_t smpi_ois_values = NULL; static int factor_cmp(const void *pa, const void *pb) { - return (((s_smpi_factor_t*)pa)->factor > ((s_smpi_factor_t*)pb)->factor); + return (((s_smpi_factor_t*)pa)->factor > ((s_smpi_factor_t*)pb)->factor) ? 1 : + (((s_smpi_factor_t*)pa)->factor < ((s_smpi_factor_t*)pb)->factor) ? -1 : 0; } @@ -85,13 +87,14 @@ static xbt_dynar_t parse_factor(const char *smpi_coef_string) char *value = NULL; unsigned int iter = 0; s_smpi_factor_t fact; + fact.nb_values=0; int i=0; xbt_dynar_t smpi_factor, radical_elements, radical_elements2 = NULL; smpi_factor = xbt_dynar_new(sizeof(s_smpi_factor_t), NULL); radical_elements = xbt_str_split(smpi_coef_string, ";"); xbt_dynar_foreach(radical_elements, iter, value) { - fact.nb_values=0; + memset(&fact, 0, sizeof(s_smpi_factor_t)); radical_elements2 = xbt_str_split(value, ":"); if (xbt_dynar_length(radical_elements2) <2 || xbt_dynar_length(radical_elements2) > 5) xbt_die("Malformed radical for smpi factor!"); @@ -128,13 +131,13 @@ static double smpi_os(double size) double current=0.0; xbt_dynar_foreach(smpi_os_values, iter, fact) { if (size <= fact.factor) { - XBT_DEBUG("os : %lf <= %ld return %f", size, fact.factor, current); + XBT_DEBUG("os : %f <= %ld return %f", size, fact.factor, current); return current; }else{ current=fact.values[0]+fact.values[1]*size; } } - XBT_DEBUG("os : %lf > %ld return %f", size, fact.factor, current); + XBT_DEBUG("os : %f > %ld return %f", size, fact.factor, current); return current; } @@ -150,13 +153,13 @@ static double smpi_ois(double size) double current=0.0; xbt_dynar_foreach(smpi_ois_values, iter, fact) { if (size <= fact.factor) { - XBT_DEBUG("ois : %lf <= %ld return %f", size, fact.factor, current); + XBT_DEBUG("ois : %f <= %ld return %f", size, fact.factor, current); return current; }else{ current=fact.values[0]+fact.values[1]*size; } } - XBT_DEBUG("ois : %lf > %ld return %f", size, fact.factor, current); + XBT_DEBUG("ois : %f > %ld return %f", size, fact.factor, current); return current; } @@ -172,12 +175,12 @@ static double smpi_or(double size) double current=0.0; xbt_dynar_foreach(smpi_or_values, iter, fact) { if (size <= fact.factor) { - XBT_DEBUG("or : %lf <= %ld return %f", size, fact.factor, current); + XBT_DEBUG("or : %f <= %ld return %f", size, fact.factor, current); return current; }else current=fact.values[0]+fact.values[1]*size; } - XBT_DEBUG("or : %lf > %ld return %f", size, fact.factor, current); + XBT_DEBUG("or : %f > %ld return %f", size, fact.factor, current); return current; } @@ -186,7 +189,7 @@ static MPI_Request build_request(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm, unsigned flags) { - MPI_Request request; + MPI_Request request = NULL; void *old_buf = NULL; @@ -218,12 +221,16 @@ static MPI_Request build_request(void *buf, int count, request->flags = flags; request->detached = 0; request->detached_sender = NULL; + request->real_src = 0; request->truncated = 0; request->real_size = 0; request->real_tag = 0; + if(flags & PERSISTENT) + request->refcount = 1; + else + request->refcount = 0; - request->refcount=1; #ifdef HAVE_TRACING request->send = 0; request->recv = 0; @@ -283,30 +290,27 @@ static void smpi_mpi_request_free_voidp(void* request) 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_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag, - comm, PERSISTENT | SEND); - request->refcount++; + MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */ + request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag, + comm, PERSISTENT | SEND | PREPARED); return request; } MPI_Request smpi_mpi_ssend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) { - MPI_Request request = - build_request(buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag, - comm, PERSISTENT | SSEND | SEND); - request->refcount++; + MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */ + request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag, + comm, PERSISTENT | SSEND | SEND | PREPARED); return request; } 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 == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag, - comm, PERSISTENT | RECV); - request->refcount++; + MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */ + request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag, + comm, PERSISTENT | RECV | PREPARED); return request; } @@ -314,10 +318,14 @@ void smpi_mpi_start(MPI_Request request) { smx_rdv_t mailbox; - xbt_assert(!request->action, - "Cannot (re)start a non-finished communication"); - if(request->flags & RECV) { + xbt_assert(!request->action, "Cannot (re)start a non-finished communication"); + request->flags &= ~PREPARED; + request->flags &= ~FINISHED; + request->refcount++; + + if (request->flags & RECV) { print_request("New recv", request); + //FIXME: if receive is posted with a large size, but send is smaller, mailboxes may not match ! if (request->size < sg_cfg_get_int("smpi/async_small_thres")) mailbox = smpi_process_mailbox_small(); else @@ -325,13 +333,16 @@ void smpi_mpi_start(MPI_Request request) // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later request->real_size=request->size; smpi_datatype_use(request->old_type); - request->action = simcall_comm_irecv(mailbox, request->buf, &request->real_size, &match_recv, request); + smpi_comm_use(request->comm); + request->action = simcall_comm_irecv(mailbox, request->buf, + &request->real_size, &match_recv, + request, -1.0); //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0 double sleeptime = request->detached ? smpi_or(request->size) : 0.0; if(sleeptime!=0.0){ simcall_process_sleep(sleeptime); - XBT_DEBUG("receiving size of %zu : sleep %lf ", request->size, smpi_or(request->size)); + XBT_DEBUG("receiving size of %zu : sleep %f ", request->size, smpi_or(request->size)); } } else { @@ -342,7 +353,7 @@ void smpi_mpi_start(MPI_Request request) #ifdef HAVE_TRACING int rank = smpi_process_index(); if (TRACE_smpi_view_internals()) { - TRACE_smpi_send(rank, rank, receiver); + TRACE_smpi_send(rank, rank, receiver,request->size); } #endif /* if(receiver == MPI_UNDEFINED) {*/ @@ -356,15 +367,26 @@ void smpi_mpi_start(MPI_Request request) XBT_DEBUG("Send request %p is not in the permanent receive mailbox (buf: %p)",request,request->buf); mailbox = smpi_process_remote_mailbox(receiver); } + + void* buf = request->buf; if ( (! (request->flags & SSEND)) && (request->size < sg_cfg_get_int("smpi/send_is_detached_thres"))) { void *oldbuf = NULL; request->detached = 1; request->refcount++; if(request->old_type->has_subtype == 0){ oldbuf = request->buf; - if (oldbuf && request->size!=0){ - request->buf = xbt_malloc(request->size); - memcpy(request->buf,oldbuf,request->size); + if (!_xbt_replay_is_active() && oldbuf && request->size!=0){ + if((smpi_privatize_global_variables) + && ((char*)request->buf >= start_data_exe) + && ((char*)request->buf < start_data_exe + size_data_exe ) + ){ + XBT_WARN("Privatization : We are sending from a zone inside global memory. Switch data segment "); + switch_data_segment(smpi_process_index()); + } + + + buf = xbt_malloc(request->size); + memcpy(buf,oldbuf,request->size); } } XBT_DEBUG("Send request %p is detached; buf %p copied into %p",request,oldbuf,request->buf); @@ -373,24 +395,25 @@ void smpi_mpi_start(MPI_Request request) // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later request->real_size=request->size; smpi_datatype_use(request->old_type); + smpi_comm_use(request->comm); //if we are giving back the control to the user without waiting for completion, we have to inject timings - double sleeptime =0.0; + double sleeptime = 0.0; if(request->detached || (request->flags & (ISEND|SSEND))){// issend should be treated as isend //isend and send timings may be different sleeptime = (request->flags & ISEND)? smpi_ois(request->size) : smpi_os(request->size); } - if(sleeptime!=0.0){ + if(sleeptime != 0.0){ simcall_process_sleep(sleeptime); - XBT_DEBUG("sending size of %zu : sleep %lf ", request->size, smpi_os(request->size)); + XBT_DEBUG("sending size of %zu : sleep %f ", request->size, smpi_os(request->size)); } request->action = simcall_comm_isend(mailbox, request->size, -1.0, - request->buf, request->real_size, + buf, request->real_size, &match_send, - &smpi_mpi_request_free_voidp, // how to free the userdata if a detached send fails + &xbt_free, // how to free the userdata if a detached send fails request, // detach if msg size < eager/rdv switch limit request->detached); @@ -409,6 +432,7 @@ void smpi_mpi_start(MPI_Request request) void smpi_mpi_startall(int count, MPI_Request * requests) { int i; + if(requests==NULL) return; for(i = 0; i < count; i++) { smpi_mpi_start(requests[i]); @@ -417,7 +441,6 @@ void smpi_mpi_startall(int count, MPI_Request * requests) void smpi_mpi_request_free(MPI_Request * request) { - if((*request) != MPI_REQUEST_NULL){ (*request)->refcount--; if((*request)->refcount<0) xbt_die("wrong refcount"); @@ -428,7 +451,6 @@ void smpi_mpi_request_free(MPI_Request * request) *request = MPI_REQUEST_NULL; }else{ print_request("Decrementing", (*request)); - } }else{ xbt_die("freeing an already free request"); @@ -438,20 +460,18 @@ void smpi_mpi_request_free(MPI_Request * request) 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_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag, - comm, NON_PERSISTENT | SEND); - + MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */ + request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf , count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag, + comm, PERSISTENT | ISEND | SEND | PREPARED); return request; } MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) { - MPI_Request request = - build_request(buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag, - comm, NON_PERSISTENT | ISEND | SEND); - + MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */ + request = build_request(buf==MPI_BOTTOM?(void*)0:buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag, + comm, NON_PERSISTENT | ISEND | SEND); smpi_mpi_start(request); return request; } @@ -459,9 +479,9 @@ MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype, MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) { - MPI_Request request = - build_request(buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag, - comm, NON_PERSISTENT | ISEND | SSEND | SEND); + MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */ + request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag, + comm, NON_PERSISTENT | ISEND | SSEND | SEND); smpi_mpi_start(request); return request; } @@ -471,19 +491,18 @@ MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype, 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 == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag, - comm, NON_PERSISTENT | RECV); + MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */ + request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag, + comm, PERSISTENT | RECV | PREPARED); return request; } MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm) { - MPI_Request request = - build_request(buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag, - comm, NON_PERSISTENT | RECV); - + MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */ + request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag, + comm, NON_PERSISTENT | RECV); smpi_mpi_start(request); return request; } @@ -491,9 +510,10 @@ MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype, void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status) { - MPI_Request request; + MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */ request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm); smpi_mpi_wait(&request, status); + request = NULL; } @@ -501,23 +521,25 @@ void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src, void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) { - MPI_Request request = - build_request(buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag, - comm, NON_PERSISTENT | SEND); + MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */ + request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag, + comm, NON_PERSISTENT | SEND); + smpi_mpi_start(request); smpi_mpi_wait(&request, MPI_STATUS_IGNORE); - + request = NULL; } void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) { - MPI_Request request = - build_request(buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag, - comm, NON_PERSISTENT | SSEND | SEND); + MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */ + request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag, + comm, NON_PERSISTENT | SSEND | SEND); smpi_mpi_start(request); smpi_mpi_wait(&request, MPI_STATUS_IGNORE); + request = NULL; } void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype, @@ -527,13 +549,20 @@ void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype, { MPI_Request requests[2]; MPI_Status stats[2]; - + int myid=smpi_process_index(); + if ((smpi_group_index(smpi_comm_group(comm), dst) == myid) && (smpi_group_index(smpi_comm_group(comm), src) == myid)) { + smpi_datatype_copy(sendbuf, sendcount, sendtype, + recvbuf, recvcount, recvtype); + return; + } 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); + smpi_mpi_request_free(&requests[0]); + smpi_mpi_request_free(&requests[1]); if(status != MPI_STATUS_IGNORE) { // Copy receive status *status = stats[1]; @@ -548,10 +577,10 @@ int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype) static void finish_wait(MPI_Request * request, MPI_Status * status) { MPI_Request req = *request; - if(status != MPI_STATUS_IGNORE) - smpi_empty_status(status); + smpi_empty_status(status); - if(!(req->detached && req->flags & SEND)){ + if(!(req->detached && req->flags & SEND) + && !(req->flags & PREPARED)){ if(status != MPI_STATUS_IGNORE) { int src = req->src == MPI_ANY_SOURCE ? req->real_src : req->src; status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(req->comm), src); @@ -567,52 +596,63 @@ static void finish_wait(MPI_Request * request, MPI_Status * status) MPI_Datatype datatype = req->old_type; if(datatype->has_subtype == 1){ - // This part handles the problem of non-contignous memory - // the unserialization at the reception + if (!_xbt_replay_is_active()){ + if( smpi_privatize_global_variables + && ((char*)req->old_buf >= start_data_exe) + && ((char*)req->old_buf < start_data_exe + size_data_exe ) + ){ + XBT_VERB("Privatization : We are unserializing to a zone in global memory - Switch data segment "); + switch_data_segment(smpi_process_index()); + } + } + // This part handles the problem of non-contignous memory + // the unserialization at the reception s_smpi_subtype_t *subtype = datatype->substruct; if(req->flags & RECV) { subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) , datatype->substruct); } if(req->detached == 0) free(req->buf); } + smpi_comm_unuse(req->comm); smpi_datatype_unuse(datatype); } #ifdef HAVE_TRACING - if (TRACE_smpi_view_internals()) { - if(req->flags & RECV){ - int rank = smpi_process_index(); - int src_traced = smpi_group_index(smpi_comm_group(req->comm), req->src == MPI_ANY_SOURCE ? req->real_src : req->src); - TRACE_smpi_recv(rank, src_traced, rank); - } + if (TRACE_smpi_view_internals()) { + if(req->flags & RECV){ + int rank = smpi_process_index(); + int src_traced = (req->src == MPI_ANY_SOURCE ? req->real_src : req->src); + TRACE_smpi_recv(rank, src_traced, rank); } + } #endif if(req->detached_sender!=NULL){ smpi_mpi_request_free(&(req->detached_sender)); } - - if(req->flags & NON_PERSISTENT) { - smpi_mpi_request_free(request); - } else { + if(req->flags & PERSISTENT) req->action = NULL; - } + req->flags |= FINISHED; + + smpi_mpi_request_free(request); + } int smpi_mpi_test(MPI_Request * request, MPI_Status * status) { int flag; //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before) - if ((*request)->action == NULL) - flag = 1; - else - flag = simcall_comm_test((*request)->action); - if(flag) { - finish_wait(request, status); - request=MPI_REQUEST_NULL; - }else{ - smpi_empty_status(status); + smpi_empty_status(status); + flag = 1; + if (!((*request)->flags & PREPARED)) { + if ((*request)->action != NULL) + flag = simcall_comm_test((*request)->action); + if (flag) { + finish_wait(request, status); + if (*request != MPI_REQUEST_NULL && !((*request)->flags & PERSISTENT)) + *request = MPI_REQUEST_NULL; + } } return flag; } @@ -626,33 +666,34 @@ int smpi_mpi_testany(int count, MPI_Request requests[], int *index, *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]!=MPI_REQUEST_NULL) && requests[i]->action) { - xbt_dynar_push(comms, &requests[i]->action); - map[size] = i; - size++; - } + 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] != MPI_REQUEST_NULL) && requests[i]->action && + !(requests[i]->flags & PREPARED)) { + xbt_dynar_push(comms, &requests[i]->action); + map[size] = i; + size++; } - if(size > 0) { - i = simcall_comm_testany(comms); - // not MPI_UNDEFINED, as this is a simix return code - if(i != -1) { - *index = map[i]; - finish_wait(&requests[*index], status); - flag = 1; - } - }else{ - //all requests are null or inactive, return true - flag=1; - smpi_empty_status(status); + } + if(size > 0) { + i = simcall_comm_testany(comms); + // not MPI_UNDEFINED, as this is a simix return code + if(i != -1) { + *index = map[i]; + finish_wait(&requests[*index], status); + if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags & NON_PERSISTENT)) + requests[*index] = MPI_REQUEST_NULL; + flag = 1; } - xbt_free(map); - xbt_dynar_free(&comms); + }else{ + //all requests are null or inactive, return true + flag=1; + smpi_empty_status(status); } + xbt_free(map); + xbt_dynar_free(&comms); return flag; } @@ -666,9 +707,11 @@ int smpi_mpi_testall(int count, MPI_Request requests[], int flag=1; int i; for(i=0; iflags & PREPARED)) { if (smpi_mpi_test(&requests[i], pstat)!=1){ flag=0; + }else{ + requests[i]=MPI_REQUEST_NULL; } }else{ smpi_empty_status(pstat); @@ -693,7 +736,7 @@ void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){ void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){ MPI_Request request =build_request(NULL, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), source), smpi_comm_rank(comm), tag, - comm, NON_PERSISTENT | RECV); + comm, PERSISTENT | RECV); //to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls double sleeptime= sg_cfg_get_double("smpi/iprobe"); @@ -721,7 +764,7 @@ void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* if(request->action){ MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action); *flag = 1; - if(status != MPI_STATUS_IGNORE) { + if(status != MPI_STATUS_IGNORE && !(req->flags & PREPARED)) { status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src); status->MPI_TAG = req->tag; status->MPI_ERROR = MPI_SUCCESS; @@ -741,11 +784,22 @@ void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* void smpi_mpi_wait(MPI_Request * request, MPI_Status * status) { print_request("Waiting", *request); + if ((*request)->flags & PREPARED) { + smpi_empty_status(status); + return; + } + if ((*request)->action != NULL) { // this is not a detached send simcall_comm_wait((*request)->action, -1.0); +#ifdef HAVE_MC + if(MC_is_active() && (*request)->action) + (*request)->action->comm.dst_data = NULL; // dangling pointer : dst_data is freed with a wait, need to set it to NULL for system state comparison +#endif } - finish_wait(request, status); + finish_wait(request, status); + if (*request != MPI_REQUEST_NULL && ((*request)->flags & NON_PERSISTENT)) + *request = MPI_REQUEST_NULL; // FIXME for a detached send, finish_wait is not called: } @@ -764,7 +818,9 @@ int smpi_mpi_waitany(int count, MPI_Request requests[], size = 0; XBT_DEBUG("Wait for one of %d", count); for(i = 0; i < count; i++) { - if(requests[i] != MPI_REQUEST_NULL) { + if (requests[i] != MPI_REQUEST_NULL + && !(requests[i]->flags & PREPARED) + && !(requests[i]->flags & FINISHED)) { if (requests[i]->action != NULL) { XBT_DEBUG("Waiting any %p ", requests[i]); xbt_dynar_push(comms, &requests[i]->action); @@ -775,6 +831,7 @@ int smpi_mpi_waitany(int count, MPI_Request requests[], size=0;//so we free the dynar but don't do the waitany call index=i; finish_wait(&requests[i], status);//cleanup if refcount = 0 + if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT)) requests[i]=MPI_REQUEST_NULL;//set to null break; } @@ -787,6 +844,8 @@ int smpi_mpi_waitany(int count, MPI_Request requests[], if (i != -1) { index = map[i]; finish_wait(&requests[index], status); + if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT)) + requests[index] = MPI_REQUEST_NULL; } } xbt_free(map); @@ -809,7 +868,8 @@ int smpi_mpi_waitall(int count, MPI_Request requests[], //tag invalid requests in the set if (status != MPI_STATUSES_IGNORE) { for (c = 0; c < count; c++) { - if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL) { + if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL || + (requests[c]->flags & PREPARED)) { smpi_empty_status(&status[c]); } else if (requests[c]->src == MPI_PROC_NULL) { smpi_empty_status(&status[c]); @@ -818,6 +878,7 @@ int smpi_mpi_waitall(int count, MPI_Request requests[], } } for(c = 0; c < count; c++) { + if (MC_is_active()) { smpi_mpi_wait(&requests[c], pstat); index = c; @@ -825,6 +886,7 @@ int smpi_mpi_waitall(int count, MPI_Request requests[], index = smpi_mpi_waitany(count, requests, pstat); if (index == MPI_UNDEFINED) break; + if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT)) requests[index]=MPI_REQUEST_NULL; } if (status != MPI_STATUSES_IGNORE) { @@ -854,6 +916,7 @@ int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices, if(status != MPI_STATUSES_IGNORE) { status[index] = *pstat; } + if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT)) requests[index]=MPI_REQUEST_NULL; }else{ return MPI_UNDEFINED; @@ -874,13 +937,13 @@ int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices, for(i = 0; i < incount; i++) { if((requests[i] != MPI_REQUEST_NULL)) { if(smpi_mpi_test(&requests[i], pstat)) { - indices[count] = i; + indices[i] = 1; count++; if(status != MPI_STATUSES_IGNORE) { status[i] = *pstat; } + if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->flags & NON_PERSISTENT) requests[i]=MPI_REQUEST_NULL; - } }else{ count_dead++; @@ -907,7 +970,7 @@ 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 system_tag = COLL_TAG_GATHER; int rank, size, src, index; MPI_Aint lb = 0, recvext = 0; MPI_Request *requests; @@ -937,6 +1000,9 @@ void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype, // Wait for completion of irecv's. smpi_mpi_startall(size - 1, requests); smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE); + for(src = 0; src < size-1; src++) { + smpi_mpi_request_free(&requests[src]); + } xbt_free(requests); } } @@ -948,6 +1014,8 @@ void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts, int i, size, count; int *displs; int rank = smpi_process_index(); + void *tmpbuf; + /* arbitrarily choose root as rank 0 */ size = smpi_comm_size(comm); count = 0; @@ -956,17 +1024,19 @@ void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts, displs[i] = count; count += recvcounts[i]; } - mpi_coll_reduce_fun(sendbuf, recvbuf, count, datatype, op, 0, comm); - smpi_mpi_scatterv(recvbuf, recvcounts, displs, datatype, recvbuf, + tmpbuf=(void*)xbt_malloc(count*smpi_datatype_get_extent(datatype)); + mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm); + smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf, recvcounts[rank], datatype, 0, comm); xbt_free(displs); + xbt_free(tmpbuf); } 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 system_tag = COLL_TAG_GATHERV; int rank, size, src, index; MPI_Aint lb = 0, recvext = 0; MPI_Request *requests; @@ -997,6 +1067,9 @@ void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype, // Wait for completion of irecv's. smpi_mpi_startall(size - 1, requests); smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE); + for(src = 0; src < size-1; src++) { + smpi_mpi_request_free(&requests[src]); + } xbt_free(requests); } } @@ -1006,7 +1079,7 @@ void smpi_mpi_allgather(void *sendbuf, int sendcount, int recvcount, MPI_Datatype recvtype, MPI_Comm comm) { - int system_tag = 666; + int system_tag = COLL_TAG_ALLGATHER; int rank, size, other, index; MPI_Aint lb = 0, recvext = 0; MPI_Request *requests; @@ -1037,6 +1110,9 @@ void smpi_mpi_allgather(void *sendbuf, int sendcount, // Wait for completion of all comms. smpi_mpi_startall(2 * (size - 1), requests); smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE); + for(other = 0; other < 2*(size-1); other++) { + smpi_mpi_request_free(&requests[other]); + } xbt_free(requests); } @@ -1045,7 +1121,7 @@ void smpi_mpi_allgatherv(void *sendbuf, int sendcount, int *recvcounts, int *displs, MPI_Datatype recvtype, MPI_Comm comm) { - int system_tag = 666; + int system_tag = COLL_TAG_ALLGATHERV; int rank, size, other, index; MPI_Aint lb = 0, recvext = 0; MPI_Request *requests; @@ -1076,6 +1152,9 @@ void smpi_mpi_allgatherv(void *sendbuf, int sendcount, // Wait for completion of all comms. smpi_mpi_startall(2 * (size - 1), requests); smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE); + for(other = 0; other < 2*(size-1); other++) { + smpi_mpi_request_free(&requests[other]); + } xbt_free(requests); } @@ -1083,7 +1162,7 @@ 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 system_tag = COLL_TAG_SCATTER; int rank, size, dst, index; MPI_Aint lb = 0, sendext = 0; MPI_Request *requests; @@ -1098,8 +1177,10 @@ void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype, // 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); + if(recvbuf!=MPI_IN_PLACE){ + 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; @@ -1114,6 +1195,9 @@ void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype, // Wait for completion of isend's. smpi_mpi_startall(size - 1, requests); smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE); + for(dst = 0; dst < size-1; dst++) { + smpi_mpi_request_free(&requests[dst]); + } xbt_free(requests); } } @@ -1122,7 +1206,7 @@ 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 system_tag = COLL_TAG_SCATTERV; int rank, size, dst, index; MPI_Aint lb = 0, sendext = 0; MPI_Request *requests; @@ -1137,8 +1221,10 @@ void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs, // FIXME: check for errors smpi_datatype_extent(sendtype, &lb, &sendext); // Local copy from root - smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root], + if(recvbuf!=MPI_IN_PLACE){ + 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; @@ -1153,6 +1239,9 @@ void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs, // Wait for completion of isend's. smpi_mpi_startall(size - 1, requests); smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE); + for(dst = 0; dst < size-1; dst++) { + smpi_mpi_request_free(&requests[dst]); + } xbt_free(requests); } } @@ -1161,23 +1250,37 @@ 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 system_tag = COLL_TAG_REDUCE; int rank, size, src, index; MPI_Aint lb = 0, dataext = 0; MPI_Request *requests; void **tmpbufs; + + 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); + } + rank = smpi_comm_rank(comm); size = smpi_comm_size(comm); + //non commutative case, use a working algo from openmpi + if(!smpi_op_is_commute(op)){ + smpi_coll_tuned_reduce_ompi_basic_linear(sendtmpbuf, recvbuf, count, + datatype, op, root, comm); + return; + } + if(rank != root) { // Send buffer to root - smpi_mpi_send(sendbuf, count, datatype, root, system_tag, comm); + smpi_mpi_send(sendtmpbuf, count, datatype, root, system_tag, comm); } else { // FIXME: check for errors smpi_datatype_extent(datatype, &lb, &dataext); // Local copy from root - if (sendbuf && recvbuf) - smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype); + if (sendtmpbuf && recvbuf) + smpi_datatype_copy(sendtmpbuf, count, datatype, recvbuf, count, datatype); // Receive buffers from senders //TODO: make a MPI_barrier here ? requests = xbt_new(MPI_Request, size - 1); @@ -1201,6 +1304,8 @@ void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count, XBT_DEBUG("finished waiting any request with index %d", index); if(index == MPI_UNDEFINED) { break; + }else{ + smpi_mpi_request_free(&requests[index]); } if(op) /* op can be MPI_OP_NULL that does nothing */ smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype); @@ -1210,6 +1315,10 @@ void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count, } xbt_free(tmpbufs); xbt_free(requests); + + if( sendbuf == MPI_IN_PLACE ) { + xbt_free(sendtmpbuf); + } } } @@ -1223,7 +1332,7 @@ void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count, void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) { - int system_tag = 666; + int system_tag = -888; int rank, size, other, index; MPI_Aint lb = 0, dataext = 0; MPI_Request *requests; @@ -1258,19 +1367,105 @@ void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count, } // 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(smpi_op_is_commute(op)){ + 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); + } } - if(index < rank) { - // #Request is below rank: it's a irecv - smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype); + }else{ + //non commutative case, wait in order + for(other = 0; other < size - 1; other++) { + smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE); + if(index < rank) { + smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype); + } } } for(index = 0; index < rank; index++) { xbt_free(tmpbufs[index]); } + for(index = 0; index < size-1; index++) { + smpi_mpi_request_free(&requests[index]); + } + xbt_free(tmpbufs); + xbt_free(requests); +} + +void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count, + MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) +{ + int system_tag = -888; + int rank, size, other, index; + MPI_Aint lb = 0, dataext = 0; + MPI_Request *requests; + void **tmpbufs; + int recvbuf_is_empty=1; + rank = smpi_comm_rank(comm); + size = smpi_comm_size(comm); + + // FIXME: check for errors + smpi_datatype_extent(datatype, &lb, &dataext); + + // 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); + if(smpi_op_is_commute(op)){ + 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) { + if(recvbuf_is_empty){ + smpi_datatype_copy(tmpbufs[index], count, datatype, recvbuf, count, datatype); + recvbuf_is_empty=0; + }else + // #Request is below rank: it's a irecv + smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype); + } + } + }else{ + //non commutative case, wait in order + for(other = 0; other < size - 1; other++) { + smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE); + if(index < rank) { + if(recvbuf_is_empty){ + smpi_datatype_copy(tmpbufs[other], count, datatype, recvbuf, count, datatype); + recvbuf_is_empty=0; + }else smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype); + } + } + } + for(index = 0; index < rank; index++) { + xbt_free(tmpbufs[index]); + } + for(index = 0; index < size-1; index++) { + smpi_mpi_request_free(&requests[index]); + } xbt_free(tmpbufs); xbt_free(requests); }