X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/f6ca129aacb2f3200af4644e0dd236985973e7b6..8a4ded6ee6330d304e1d3f69100eb7ec55c5fd9c:/src/smpi/smpi_base.c diff --git a/src/smpi/smpi_base.c b/src/smpi/smpi_base.c index d7b67c078e..8398a13f0d 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 @@ -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; } @@ -70,13 +71,21 @@ xbt_dynar_t smpi_os_values = NULL; xbt_dynar_t smpi_or_values = NULL; xbt_dynar_t smpi_ois_values = NULL; +double smpi_wtime_sleep = 0.0; +double smpi_iprobe_sleep = 1e-4; +double smpi_test_sleep = 1e-4; + +xbt_dict_t smpi_keyvals = NULL; +int keyval_id=MPI_TAG_UB+1;//avoid collisions + // Methods used to parse and store the values for timing injections in smpi // These are taken from surf/network.c and generalized to have more factors // These methods should be merged with those in surf/network.c (moved somewhere in xbt ?) 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 +94,14 @@ 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 +138,13 @@ 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 +160,13 @@ 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; } @@ -172,21 +182,35 @@ 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; } +double smpi_mpi_wtime(){ + double time; + if (smpi_process_initialized() && !smpi_process_finalized() && !smpi_process_get_sampling()) { + smpi_bench_end(); + time = SIMIX_get_clock(); + //to avoid deadlocks if called too many times + if(smpi_wtime_sleep > 0) simcall_process_sleep(smpi_wtime_sleep); + smpi_bench_begin(); + } else { + time = SIMIX_get_clock(); + } + return time; +} + 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; @@ -194,11 +218,11 @@ static MPI_Request build_request(void *buf, int count, s_smpi_subtype_t *subtype = datatype->substruct; - if(datatype->has_subtype == 1){ + if(((flags & RECV) && (flags & ACCUMULATE)) || (datatype->has_subtype == 1)){ // This part handles the problem of non-contiguous memory old_buf = buf; buf = count==0 ? NULL : xbt_malloc(count*smpi_datatype_size(datatype)); - if (flags & SEND) { + if ((datatype->has_subtype == 1) && (flags & SEND)) { subtype->serialize(old_buf, buf, count, datatype->substruct); } } @@ -218,12 +242,16 @@ 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; - - request->refcount=1; + if(flags & PERSISTENT) + request->refcount = 1; + else + request->refcount = 0; + request->op = MPI_REPLACE; #ifdef HAVE_TRACING request->send = 0; request->recv = 0; @@ -253,8 +281,8 @@ void smpi_action_trace_run(char *path) action_fp=NULL; if (path) { action_fp = fopen(path, "r"); - xbt_assert(action_fp != NULL, "Cannot open %s: %s", path, - strerror(errno)); + if (action_fp == NULL) + xbt_die("Cannot open %s: %s", path, strerror(errno)); } if (!xbt_dict_is_empty(action_queues)) { @@ -283,117 +311,180 @@ 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; } void smpi_mpi_start(MPI_Request request) { smx_rdv_t mailbox; + + xbt_assert(!request->action, "Cannot (re)start a non-finished communication"); + request->flags &= ~PREPARED; + request->flags &= ~FINISHED; + request->refcount++; - xbt_assert(!request->action, - "Cannot (re)start a non-finished communication"); - if(request->flags & RECV) { + if (request->flags & RECV) { print_request("New recv", request); - if (request->size < sg_cfg_get_int("smpi/async_small_thres")) + + if (request->flags & RMA || request->size < sg_cfg_get_int("smpi/async_small_thres")){ + //We have to check both mailboxes (because SSEND messages are sent to the large mbox). begin with the more appropriate one : the small one. mailbox = smpi_process_mailbox_small(); - else - mailbox = smpi_process_mailbox(); - // 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); + XBT_DEBUG("Is there a corresponding send already posted the small mailbox %p (in case of SSEND)?", mailbox); + smx_action_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request); + + if(action ==NULL){ + mailbox = smpi_process_mailbox(); + XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox); + action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request); + if(action ==NULL){ + XBT_DEBUG("Still notching, switch back to the small mailbox : %p", mailbox); + mailbox = smpi_process_mailbox_small(); + } + }else{ + XBT_DEBUG("yes there was something for us in the large mailbox"); + } + }else{ + mailbox = smpi_process_mailbox_small(); + XBT_DEBUG("Is there a corresponding send already posted the small mailbox?"); + smx_action_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request); + + if(action ==NULL){ + XBT_DEBUG("No, nothing in the permanent receive mailbox"); + mailbox = smpi_process_mailbox(); + }else{ + XBT_DEBUG("yes there was something for us in the small mailbox"); + } + } //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)); } + + // 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); + request->action = simcall_comm_irecv(mailbox, request->buf, + &request->real_size, &match_recv, + !smpi_process_get_replaying()? &smpi_comm_copy_buffer_callback + : &smpi_comm_null_copy_buffer_callback, + request, -1.0); + XBT_DEBUG("recv simcall posted"); + } else { - int receiver = request->dst;//smpi_group_index(smpi_comm_group(request->comm), request->dst); + int receiver = request->dst; #ifdef HAVE_TRACING - int rank = smpi_process_index(); + int rank = request->src; if (TRACE_smpi_view_internals()) { - TRACE_smpi_send(rank, rank, receiver); + TRACE_smpi_send(rank, rank, receiver,request->size); } #endif -/* if(receiver == MPI_UNDEFINED) {*/ -/* XBT_WARN("Trying to send a message to a wrong rank");*/ -/* return;*/ -/* }*/ print_request("New send", request); - if (request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode - mailbox = smpi_process_remote_mailbox_small(receiver); + + //if we are giving back the control to the user without waiting for completion, we have to inject timings + 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){ + simcall_process_sleep(sleeptime); + XBT_DEBUG("sending size of %zu : sleep %f ", request->size, smpi_os(request->size)); + } + + if (request->flags & RMA || request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode + mailbox = smpi_process_remote_mailbox(receiver); + XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox); + smx_action_t action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, (void*)request); + if(action ==NULL){ + if (! (request->flags & SSEND)){ + mailbox = smpi_process_remote_mailbox_small(receiver); + XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %p", mailbox); + } else{ + mailbox = smpi_process_remote_mailbox_small(receiver); + XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %p?", mailbox); + action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, (void*)request); + if(action ==NULL){ + XBT_DEBUG("No, we are first, send to large mailbox"); + mailbox = smpi_process_remote_mailbox(receiver); + } + } + }else{ + XBT_DEBUG("Yes there was something for us in the large mailbox"); + } }else{ - XBT_DEBUG("Send request %p is not in the permanent receive mailbox (buf: %p)",request,request->buf); mailbox = smpi_process_remote_mailbox(receiver); + XBT_DEBUG("Send request %p is in the large mailbox %p (buf: %p)",mailbox, request,request->buf); } + + 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; + XBT_DEBUG("Send request %p is detached", request); 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 (!smpi_process_get_replaying() && 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_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment "); + smpi_switch_data_segment(request->src); + } + buf = xbt_malloc(request->size); + memcpy(buf,oldbuf,request->size); + XBT_DEBUG("buf %p copied into %p",oldbuf,buf); } } - XBT_DEBUG("Send request %p is detached; buf %p copied into %p",request,oldbuf,request->buf); } // 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); - - //if we are giving back the control to the user without waiting for completion, we have to inject timings - 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){ - simcall_process_sleep(sleeptime); - XBT_DEBUG("sending size of %zu : sleep %lf ", request->size, smpi_os(request->size)); - } - + smpi_comm_use(request->comm); request->action = - simcall_comm_isend(mailbox, request->size, -1.0, - request->buf, request->real_size, + simcall_comm_isend(SIMIX_process_from_PID(request->src+1), mailbox, request->size, -1.0, + buf, request->real_size, &match_send, - &smpi_mpi_request_free_voidp, // how to free the userdata if a detached send fails + &xbt_free_f, // how to free the userdata if a detached send fails + !smpi_process_get_replaying()? &smpi_comm_copy_buffer_callback + : &smpi_comm_null_copy_buffer_callback, request, // detach if msg size < eager/rdv switch limit request->detached); + XBT_DEBUG("send simcall posted"); + + #ifdef HAVE_TRACING /* FIXME: detached sends are not traceable (request->action == NULL) */ @@ -409,6 +500,7 @@ 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 +509,6 @@ 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"); @@ -428,30 +519,59 @@ 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"); } } + +MPI_Request smpi_rma_send_init(void *buf, int count, MPI_Datatype datatype, + int src, int dst, int tag, MPI_Comm comm, MPI_Op op) +{ + MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */ + if(op==MPI_OP_NULL){ + request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf , count, datatype, src, dst, tag, + comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED); + }else{ + request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag, + comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED | ACCUMULATE); + request->op = op; + } + return request; +} + +MPI_Request smpi_rma_recv_init(void *buf, int count, MPI_Datatype datatype, + int src, int dst, int tag, MPI_Comm comm, MPI_Op op) +{ + MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */ + if(op==MPI_OP_NULL){ + request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag, + comm, RMA | NON_PERSISTENT | RECV | PREPARED); + }else{ + request = build_request(buf==MPI_BOTTOM ? (void*)0 : buf, count, datatype, src, dst, tag, + comm, RMA | NON_PERSISTENT | RECV | PREPARED | ACCUMULATE); + request->op = op; + } + return 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; } @@ -459,31 +579,28 @@ 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; } - - 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 +608,10 @@ 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 +619,25 @@ 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, @@ -539,6 +659,8 @@ void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype, 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]; @@ -553,10 +675,10 @@ 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); @@ -571,53 +693,78 @@ static void finish_wait(MPI_Request * request, MPI_Status * status) print_request("Finishing", req); MPI_Datatype datatype = req->old_type; - if(datatype->has_subtype == 1){ + if((req->flags & ACCUMULATE) || (datatype->has_subtype == 1)){ + if (!smpi_process_get_replaying()){ + 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 "); + smpi_switch_data_segment(smpi_process_index()); + } + } + + if(datatype->has_subtype == 1){ // 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); + 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, req->op); + if(req->detached == 0) free(req->buf); + }else if(req->flags & RECV){//apply op on contiguous buffer for accumulate + int n =req->real_size/smpi_datatype_size(datatype); + smpi_op_apply(req->op, req->buf, req->old_buf, &n, &datatype); } - 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); + + //to avoid deadlocks + //multiplier to the sleeptime, to increase speed of execution, each failed test will increase it + static int nsleeps = 1; + if(smpi_test_sleep > 0) simcall_process_sleep(nsleeps*smpi_test_sleep); + + 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); + nsleeps=1;//reset the number of sleeps we will do next time + if (*request != MPI_REQUEST_NULL && !((*request)->flags & PERSISTENT)) + *request = MPI_REQUEST_NULL; + }else{ + nsleeps++; + } } return flag; } @@ -631,33 +778,41 @@ 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; - } + } + if(size > 0) { + //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it + static int nsleeps = 1; + if(smpi_test_sleep > 0) simcall_process_sleep(nsleeps*smpi_test_sleep); + + 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; + nsleeps=1; }else{ - //all requests are null or inactive, return true - flag=1; - smpi_empty_status(status); + nsleeps++; } - 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; } @@ -671,9 +826,11 @@ 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); @@ -698,15 +855,12 @@ 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"); //multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it static int nsleeps = 1; - - simcall_process_sleep(sleeptime); - + if(smpi_iprobe_sleep > 0) simcall_process_sleep(nsleeps*smpi_iprobe_sleep); // behave like a receive, but don't do it smx_rdv_t mailbox; @@ -715,18 +869,18 @@ void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* if (sg_cfg_get_int("smpi/async_small_thres")>0){ mailbox = smpi_process_mailbox_small(); XBT_DEBUG("trying to probe the perm recv mailbox"); - request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request); + request->action = simcall_comm_iprobe(mailbox, 0, request->src, request->tag, &match_recv, (void*)request); } if (request->action==NULL){ mailbox = smpi_process_mailbox(); XBT_DEBUG("trying to probe the other mailbox"); - request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request); + request->action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request); } 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; @@ -746,11 +900,22 @@ 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: } @@ -769,7 +934,9 @@ 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); @@ -780,6 +947,7 @@ 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; } @@ -792,6 +960,8 @@ 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); @@ -814,7 +984,8 @@ 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]); @@ -823,6 +994,7 @@ 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; @@ -830,6 +1002,7 @@ 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) { @@ -859,6 +1032,7 @@ 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; @@ -879,13 +1053,13 @@ 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++; @@ -942,6 +1116,9 @@ 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); } } @@ -963,12 +1140,13 @@ void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts, displs[i] = count; count += recvcounts[i]; } - tmpbuf=(void*)xbt_malloc(count*smpi_datatype_get_extent(datatype)); + tmpbuf=(void*)smpi_get_tmp_sendbuffer(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); + smpi_free_tmp_buffer(tmpbuf); } void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype, @@ -1006,6 +1184,9 @@ 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); } } @@ -1046,6 +1227,9 @@ 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); } @@ -1085,6 +1269,9 @@ 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); } @@ -1107,8 +1294,10 @@ 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; @@ -1123,6 +1312,9 @@ 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); } } @@ -1146,8 +1338,10 @@ 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; @@ -1162,6 +1356,9 @@ 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); } } @@ -1176,9 +1373,11 @@ void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count, MPI_Request *requests; void **tmpbufs; + char* sendtmpbuf = (char*) sendbuf; if( sendbuf == MPI_IN_PLACE ) { - sendtmpbuf = (char *)recvbuf; + sendtmpbuf = (char *)smpi_get_tmp_sendbuffer(count*smpi_datatype_get_extent(datatype)); + smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype); } rank = smpi_comm_rank(comm); @@ -1208,7 +1407,10 @@ void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count, if(src != root) { // FIXME: possibly overkill we we have contiguous/noncontiguous data // mapping... - tmpbufs[index] = xbt_malloc(count * dataext); + if (!smpi_process_get_replaying()) + tmpbufs[index] = xbt_malloc(count * dataext); + else + tmpbufs[index] = smpi_get_tmp_sendbuffer(count * dataext); requests[index] = smpi_irecv_init(tmpbufs[index], count, datatype, src, system_tag, comm); @@ -1222,15 +1424,21 @@ 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); } - for(index = 0; index < size - 1; index++) { - xbt_free(tmpbufs[index]); - } + for(index = 0; index < size - 1; index++) { + smpi_free_tmp_buffer(tmpbufs[index]); + } xbt_free(tmpbufs); xbt_free(requests); + + if( sendbuf == MPI_IN_PLACE ) { + smpi_free_tmp_buffer(sendtmpbuf); + } } } @@ -1244,7 +1452,7 @@ 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 = 888; + int system_tag = -888; int rank, size, other, index; MPI_Aint lb = 0, dataext = 0; MPI_Request *requests; @@ -1266,7 +1474,73 @@ void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count, for(other = 0; other < rank; other++) { // FIXME: possibly overkill we we have contiguous/noncontiguous data // mapping... - tmpbufs[index] = xbt_malloc(count * dataext); + tmpbufs[index] = smpi_get_tmp_sendbuffer(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) { + // #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++) { + smpi_free_tmp_buffer(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] = smpi_get_tmp_sendbuffer(count * dataext); requests[index] = smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag, comm); @@ -1279,19 +1553,98 @@ 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) { + 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); + } } - 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) { + 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]); + smpi_free_tmp_buffer(tmpbufs[index]); + } + for(index = 0; index < size-1; index++) { + smpi_mpi_request_free(&requests[index]); } xbt_free(tmpbufs); xbt_free(requests); } + + +int smpi_keyval_create(MPI_Copy_function* copy_fn, MPI_Delete_function* delete_fn, int* keyval, void* extra_state){ + + if(!smpi_keyvals) + smpi_keyvals = xbt_dict_new(); + + smpi_key_elem value = (smpi_key_elem) xbt_new0(s_smpi_mpi_key_elem_t,1); + + value->copy_fn=copy_fn; + value->delete_fn=delete_fn; + + *keyval = keyval_id; + + xbt_dict_set(smpi_keyvals,(const char*)keyval,(void*)value, NULL); + keyval_id++; + return MPI_SUCCESS; +} + +int smpi_keyval_free(int* keyval){ + smpi_key_elem elem = xbt_dict_get_or_null(smpi_keyvals, (const char*)keyval); + if(!elem) + return MPI_ERR_ARG; + xbt_dict_remove(smpi_keyvals, (const char*)keyval); + xbt_free(elem); + return MPI_SUCCESS; +} + +int smpi_attr_delete(MPI_Comm comm, int keyval){ + smpi_key_elem elem = xbt_dict_get_or_null(smpi_keyvals, (const char*)&keyval); + if(!elem) + return MPI_ERR_ARG; + if(elem->delete_fn!=MPI_NULL_DELETE_FN){ + void * value; + int flag; + if(smpi_attr_get(comm, keyval, &value, &flag)==MPI_SUCCESS) + elem->delete_fn(comm, keyval, &value, &flag); + } + return smpi_comm_attr_delete(comm, keyval);; +} + +int smpi_attr_get(MPI_Comm comm, int keyval, void* attr_value, int* flag){ + smpi_key_elem elem = xbt_dict_get_or_null(smpi_keyvals, (const char*)&keyval); + if(!elem) + return MPI_ERR_ARG; + return smpi_comm_attr_get(comm, keyval, attr_value, flag);; +} + +int smpi_attr_put(MPI_Comm comm, int keyval, void* attr_value){ + + if(!smpi_keyvals) + smpi_keyvals = xbt_dict_new(); + + smpi_key_elem elem = xbt_dict_get_or_null(smpi_keyvals, (const char*)&keyval); + if(!elem ) + return MPI_ERR_ARG; + return smpi_comm_attr_put(comm, keyval, attr_value);; +} +