X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/726976b66d823569fe7f193e2cc93251435c0d16..d7b725316b632998ed194144be7b4f4b2d433b55:/src/smpi/smpi_base.c diff --git a/src/smpi/smpi_base.c b/src/smpi/smpi_base.c index 837f1e3811..5ad3fa7b8a 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) 2007-2013. 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; } @@ -197,7 +198,7 @@ static MPI_Request build_request(void *buf, int count, if(datatype->has_subtype == 1){ // This part handles the problem of non-contiguous memory old_buf = buf; - buf = xbt_malloc(count*smpi_datatype_size(datatype)); + buf = count==0 ? NULL : xbt_malloc(count*smpi_datatype_size(datatype)); if (flags & SEND) { subtype->serialize(old_buf, buf, count, datatype->substruct); } @@ -218,6 +219,7 @@ 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; @@ -284,8 +286,8 @@ MPI_Request smpi_mpi_send_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) { MPI_Request request = - build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag, - comm, PERSISTENT | SEND); + 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); request->refcount++; return request; } @@ -294,8 +296,8 @@ 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_comm_rank(comm), dst, tag, - comm, PERSISTENT | SSEND | SEND); + 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); request->refcount++; return request; } @@ -304,8 +306,8 @@ 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); + 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); request->refcount++; return request; } @@ -316,8 +318,10 @@ void smpi_mpi_start(MPI_Request request) xbt_assert(!request->action, "Cannot (re)start a non-finished communication"); + if(request->flags & PREPARED)request->flags &= ~PREPARED; 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 @@ -325,6 +329,7 @@ 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); request->action = simcall_comm_irecv(mailbox, request->buf, &request->real_size, &match_recv, request); //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0 @@ -337,12 +342,12 @@ void smpi_mpi_start(MPI_Request request) } else { - int receiver = smpi_group_index(smpi_comm_group(request->comm), request->dst); + int receiver = request->dst;//smpi_group_index(smpi_comm_group(request->comm), request->dst); #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) {*/ @@ -362,16 +367,18 @@ void smpi_mpi_start(MPI_Request request) request->refcount++; if(request->old_type->has_subtype == 0){ oldbuf = request->buf; - if (oldbuf){ + if (oldbuf && request->size!=0){ request->buf = xbt_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); } + // 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; @@ -416,14 +423,17 @@ 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"); if((*request)->refcount==0){ + print_request("Destroying", (*request)); xbt_free(*request); *request = MPI_REQUEST_NULL; + }else{ + print_request("Decrementing", (*request)); + } }else{ xbt_die("freeing an already free request"); @@ -434,8 +444,8 @@ MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) { MPI_Request request = - build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag, - comm, NON_PERSISTENT | SEND); + 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 | PREPARED); return request; } @@ -444,7 +454,7 @@ 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_comm_rank(comm), dst, tag, + 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); @@ -455,7 +465,7 @@ 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_comm_rank(comm), dst, tag, + 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; @@ -467,8 +477,8 @@ 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); + 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 | PREPARED); return request; } @@ -476,7 +486,7 @@ 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, smpi_comm_rank(comm), tag, + 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); @@ -497,9 +507,8 @@ 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_comm_rank(comm), dst, tag, + 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); @@ -508,7 +517,11 @@ void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst, void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) { - MPI_Request request = smpi_mpi_issend(buf, count, datatype, dst, tag, comm); + MPI_Request 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); } @@ -519,7 +532,12 @@ 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] = @@ -543,9 +561,10 @@ static void finish_wait(MPI_Request * request, MPI_Status * status) if(status != MPI_STATUS_IGNORE) smpi_empty_status(status); - if(!(req->detached && req->flags & SEND)){ - if(status != MPI_STATUS_IGNORE) { - status->MPI_SOURCE = req->src == MPI_ANY_SOURCE ? req->real_src : req->src; + 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); status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag; status->MPI_ERROR = req->truncated ? MPI_ERR_TRUNCATE : MPI_SUCCESS; // this handles the case were size in receive differs from size in send @@ -566,6 +585,7 @@ static void finish_wait(MPI_Request * request, MPI_Status * status) } if(req->detached == 0) free(req->buf); } + smpi_comm_unuse(req->comm); smpi_datatype_unuse(datatype); } @@ -574,7 +594,7 @@ static void finish_wait(MPI_Request * request, MPI_Status * status) 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); + int src_traced = (req->src == MPI_ANY_SOURCE ? req->real_src : req->src); TRACE_smpi_recv(rank, src_traced, rank); } } @@ -600,8 +620,8 @@ int smpi_mpi_test(MPI_Request * request, MPI_Status * status) { else flag = simcall_comm_test((*request)->action); if(flag) { - (*request)->refcount++; finish_wait(request, status); + request=MPI_REQUEST_NULL; }else{ smpi_empty_status(status); } @@ -617,33 +637,31 @@ 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) { + 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); + 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; } @@ -660,6 +678,8 @@ int smpi_mpi_testall(int count, MPI_Request requests[], if(requests[i]!= MPI_REQUEST_NULL){ if (smpi_mpi_test(&requests[i], pstat)!=1){ flag=0; + }else{ + requests[i]=MPI_REQUEST_NULL; } }else{ smpi_empty_status(pstat); @@ -678,16 +698,21 @@ void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){ while(flag==0){ smpi_mpi_iprobe(source, tag, comm, &flag, status); XBT_DEBUG("Busy Waiting on probing : %d", flag); - if(!flag) { - simcall_process_sleep(0.0001); - } } } 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, smpi_comm_rank(comm), tag, + + 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); + //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); + // behave like a receive, but don't do it smx_rdv_t mailbox; @@ -707,14 +732,18 @@ 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) { - status->MPI_SOURCE = req->src; + 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; status->count = req->real_size; } + nsleeps=1;//reset the number of sleeps we will do next time + } + else { + *flag = 0; + nsleeps++; } - else *flag = 0; smpi_mpi_request_free(&request); return; @@ -727,7 +756,7 @@ void smpi_mpi_wait(MPI_Request * request, MPI_Status * status) simcall_comm_wait((*request)->action, -1.0); } finish_wait(request, status); - + request=MPI_REQUEST_NULL; // FIXME for a detached send, finish_wait is not called: } @@ -807,6 +836,7 @@ int smpi_mpi_waitall(int count, MPI_Request requests[], index = smpi_mpi_waitany(count, requests, pstat); if (index == MPI_UNDEFINED) break; + requests[index]=MPI_REQUEST_NULL; } if (status != MPI_STATUSES_IGNORE) { status[index] = *pstat; @@ -835,6 +865,7 @@ int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices, if(status != MPI_STATUSES_IGNORE) { status[index] = *pstat; } + requests[index]=MPI_REQUEST_NULL; }else{ return MPI_UNDEFINED; } @@ -854,11 +885,13 @@ 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; } + requests[i]=MPI_REQUEST_NULL; + } }else{ count_dead++; @@ -885,7 +918,7 @@ 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; @@ -919,11 +952,36 @@ void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype, } } + +void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts, + MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) +{ + 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; + displs = xbt_new(int, size); + for (i = 0; i < size; i++) { + displs[i] = count; + count += recvcounts[i]; + } + 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; @@ -963,7 +1021,7 @@ 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; @@ -1002,7 +1060,7 @@ 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; @@ -1040,7 +1098,7 @@ 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; @@ -1055,8 +1113,10 @@ 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; @@ -1079,7 +1139,7 @@ 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; @@ -1094,8 +1154,10 @@ 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; @@ -1118,23 +1180,37 @@ 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); @@ -1167,6 +1243,10 @@ void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count, } xbt_free(tmpbufs); xbt_free(requests); + + if( sendbuf == MPI_IN_PLACE ) { + xbt_free(sendtmpbuf); + } } } @@ -1180,7 +1260,7 @@ 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; @@ -1215,14 +1295,94 @@ 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); + } + } + }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]); + } + 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); + } } - 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++) {