X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/47158a9055045f3bce00e77b9010e6a8b0e23802..845590d7b4f0790c34191697d1b6c28d50d1ccbd:/src/smpi/mpi/smpi_request.cpp diff --git a/src/smpi/mpi/smpi_request.cpp b/src/smpi/mpi/smpi_request.cpp index 5bbd84ea23..5cf3a4990b 100644 --- a/src/smpi/mpi/smpi_request.cpp +++ b/src/smpi/mpi/smpi_request.cpp @@ -2,7 +2,8 @@ /* This program is free software; you can redistribute it and/or modify it * under the terms of the license (GNU LGPL) which comes with this package. */ - +#include "simgrid/s4u/Mutex.hpp" +#include "simgrid/s4u/ConditionVariable.hpp" #include "smpi_request.hpp" #include "mc/mc.h" @@ -15,10 +16,10 @@ #include "smpi_op.hpp" #include "src/kernel/activity/CommImpl.hpp" #include "src/mc/mc_replay.hpp" -#include "src/simix/ActorImpl.hpp" #include "src/smpi/include/smpi_actor.hpp" #include "xbt/config.hpp" + #include XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_request, smpi, "Logging specific to SMPI (request)"); @@ -35,8 +36,8 @@ extern void (*smpi_comm_copy_data_callback)(simgrid::kernel::activity::CommImpl* namespace simgrid{ namespace smpi{ -Request::Request(void* buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm, unsigned flags) - : buf_(buf), old_type_(datatype), src_(src), dst_(dst), tag_(tag), comm_(comm), flags_(flags) +Request::Request(void* buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm, unsigned flags, MPI_Op op) + : buf_(buf), old_type_(datatype), src_(src), dst_(dst), tag_(tag), comm_(comm), flags_(flags), op_(op) { void *old_buf = nullptr; // FIXME Handle the case of a partial shared malloc. @@ -57,6 +58,8 @@ Request::Request(void* buf, int count, MPI_Datatype datatype, int src, int dst, size_ = datatype->size() * count; datatype->ref(); comm_->ref(); + if(op != MPI_REPLACE && op != MPI_OP_NULL) + op_->ref(); action_ = nullptr; detached_ = 0; detached_sender_ = nullptr; @@ -68,44 +71,10 @@ Request::Request(void* buf, int count, MPI_Datatype datatype, int src, int dst, refcount_ = 1; else refcount_ = 0; - op_ = MPI_REPLACE; cancelled_ = 0; -} - -MPI_Comm Request::comm(){ - return comm_; -} - -int Request::src(){ - return src_; -} - -int Request::dst(){ - return dst_; -} - -int Request::tag(){ - return tag_; -} - -int Request::flags(){ - return flags_; -} - -int Request::detached(){ - return detached_; -} - -MPI_Datatype Request::type(){ - return old_type_; -} - -size_t Request::size(){ - return size_; -} - -size_t Request::real_size(){ - return real_size_; + generalized_funcs=nullptr; + nbc_requests_=nullptr; + nbc_requests_size_=0; } void Request::ref(){ @@ -121,11 +90,18 @@ void Request::unref(MPI_Request* request) xbt_die("Whoops, wrong refcount"); } if((*request)->refcount_==0){ - Datatype::unref((*request)->old_type_); + if ((*request)->flags_ & MPI_REQ_GENERALIZED){ + ((*request)->generalized_funcs)->free_fn(((*request)->generalized_funcs)->extra_state); + }else{ Comm::unref((*request)->comm_); - (*request)->print_request("Destroying"); - delete *request; - *request = MPI_REQUEST_NULL; + Datatype::unref((*request)->old_type_); + } + if ((*request)->op_!=MPI_REPLACE && (*request)->op_!=MPI_OP_NULL) + Op::unref(&(*request)->op_); + + (*request)->print_request("Destroying"); + delete *request; + *request = MPI_REQUEST_NULL; }else{ (*request)->print_request("Decrementing"); } @@ -229,8 +205,7 @@ MPI_Request Request::rma_send_init(void *buf, int count, MPI_Datatype datatype, request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src)->get_pid(), comm->group()->actor(dst)->get_pid(), tag, comm, MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED | - MPI_REQ_ACCUMULATE); - request->op_ = op; + MPI_REQ_ACCUMULATE, op); } return request; } @@ -254,8 +229,7 @@ MPI_Request Request::rma_recv_init(void *buf, int count, MPI_Datatype datatype, }else{ request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src)->get_pid(), comm->group()->actor(dst)->get_pid(), tag, comm, - MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED | MPI_REQ_ACCUMULATE); - request->op_ = op; + MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED | MPI_REQ_ACCUMULATE, op); } return request; } @@ -361,7 +335,7 @@ void Request::sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,int d void Request::start() { - s4u::MailboxPtr mailbox; + s4u::Mailbox* mailbox; xbt_assert(action_ == nullptr, "Cannot (re-)start unfinished communication"); flags_ &= ~MPI_REQ_PREPARED; @@ -543,28 +517,59 @@ void Request::cancel() (boost::static_pointer_cast(this->action_))->cancel(); } -int Request::test(MPI_Request * request, MPI_Status * status) { +int Request::test(MPI_Request * request, MPI_Status * status, int* flag) { //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or testall before) // to avoid deadlocks if used as a break condition, such as // while (MPI_Test(request, flag, status) && flag) dostuff... // because the time will not normally advance when only calls to MPI_Test are made -> deadlock // multiplier to the sleeptime, to increase speed of execution, each failed test will increase it static int nsleeps = 1; + int ret = MPI_SUCCESS; + + // Are we testing a request meant for non blocking collectives ? + // If so, test all the subrequests. + if ((*request)->nbc_requests_size_>0){ + ret = testall((*request)->nbc_requests_size_, (*request)->nbc_requests_, flag, MPI_STATUSES_IGNORE); + if(*flag){ + delete[] (*request)->nbc_requests_; + (*request)->nbc_requests_size_=0; + unref(request); + } + return ret; + } + if(smpi_test_sleep > 0) simcall_process_sleep(nsleeps*smpi_test_sleep); + MPI_Status* mystatus; Status::empty(status); - int flag = 1; + *flag = 1; if (((*request)->flags_ & MPI_REQ_PREPARED) == 0) { if ((*request)->action_ != nullptr){ try{ - flag = simcall_comm_test((*request)->action_); + *flag = simcall_comm_test((*request)->action_); }catch (xbt_ex& e) { - return 0; + *flag = 0; + return ret; } } - if (flag) { + if (*request != MPI_REQUEST_NULL && + ((*request)->flags_ & MPI_REQ_GENERALIZED) + && !((*request)->flags_ & MPI_REQ_COMPLETE)) + *flag=0; + if (*flag) { finish_wait(request,status); + if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_GENERALIZED)){ + if(status==MPI_STATUS_IGNORE){ + mystatus=new MPI_Status(); + Status::empty(mystatus); + }else{ + mystatus=status; + } + ret = ((*request)->generalized_funcs)->query_fn(((*request)->generalized_funcs)->extra_state, mystatus); + if(status==MPI_STATUS_IGNORE) + delete mystatus; + } nsleeps=1;//reset the number of sleeps we will do next time if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_PERSISTENT) == 0) *request = MPI_REQUEST_NULL; @@ -572,21 +577,27 @@ int Request::test(MPI_Request * request, MPI_Status * status) { nsleeps++; } } - return flag; + return ret; } -int Request::testsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[]) +int Request::testsome(int incount, MPI_Request requests[], int *count, int *indices, MPI_Status status[]) { - int count = 0; + int ret = MPI_SUCCESS; + int error=0; int count_dead = 0; + int flag = 0; MPI_Status stat; MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat; + *count = 0; for (int i = 0; i < incount; i++) { if (requests[i] != MPI_REQUEST_NULL) { - if (test(&requests[i], pstat)) { + ret = test(&requests[i], pstat, &flag); + if(ret!=MPI_SUCCESS) + error = 1; + if(flag) { indices[i] = 1; - count++; + (*count)++; if (status != MPI_STATUSES_IGNORE) status[i] = *pstat; if ((requests[i] != MPI_REQUEST_NULL) && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT)) @@ -596,19 +607,22 @@ int Request::testsome(int incount, MPI_Request requests[], int *indices, MPI_Sta count_dead++; } } - if(count_dead==incount) - return MPI_UNDEFINED; - else return count; + if(count_dead==incount)*count=MPI_UNDEFINED; + if(error!=0) + return MPI_ERR_IN_STATUS; + else + return MPI_SUCCESS; } -int Request::testany(int count, MPI_Request requests[], int *index, MPI_Status * status) +int Request::testany(int count, MPI_Request requests[], int *index, int* flag, MPI_Status * status) { std::vector comms; comms.reserve(count); int i; - int flag = 0; - + *flag = 0; + int ret = MPI_SUCCESS; + MPI_Status* mystatus; *index = MPI_UNDEFINED; std::vector map; /** Maps all matching comms back to their location in requests **/ @@ -631,36 +645,59 @@ int Request::testany(int count, MPI_Request requests[], int *index, MPI_Status * if (i != -1) { // -1 is not MPI_UNDEFINED but a SIMIX return code. (nothing matches) *index = map[i]; - finish_wait(&requests[*index],status); - flag = 1; - nsleeps = 1; - if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_NON_PERSISTENT)) { - requests[*index] = MPI_REQUEST_NULL; + if (requests[*index] != MPI_REQUEST_NULL && + (requests[*index]->flags_ & MPI_REQ_GENERALIZED) + && !(requests[*index]->flags_ & MPI_REQ_COMPLETE)) { + *flag=0; + } else { + finish_wait(&requests[*index],status); + if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_GENERALIZED)){ + if(status==MPI_STATUS_IGNORE){ + mystatus=new MPI_Status(); + Status::empty(mystatus); + }else{ + mystatus=status; + } + ret=(requests[*index]->generalized_funcs)->query_fn((requests[*index]->generalized_funcs)->extra_state, mystatus); + if(status==MPI_STATUS_IGNORE) + delete mystatus; } + + if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_NON_PERSISTENT)) + requests[*index] = MPI_REQUEST_NULL; + *flag=1; + } + nsleeps = 1; } else { nsleeps++; } } else { //all requests are null or inactive, return true - flag = 1; + *flag = 1; Status::empty(status); } - return flag; + return ret; } -int Request::testall(int count, MPI_Request requests[], MPI_Status status[]) +int Request::testall(int count, MPI_Request requests[], int* outflag, MPI_Status status[]) { MPI_Status stat; MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat; - int flag=1; + int flag, error=0; + int ret=MPI_SUCCESS; + *outflag = 1; for(int i=0; iflags_ & MPI_REQ_PREPARED)) { - if (test(&requests[i], pstat)!=1){ + ret = test(&requests[i], pstat, &flag); + if (flag){ flag=0; + requests[i]=MPI_REQUEST_NULL; }else{ - requests[i]=MPI_REQUEST_NULL; + *outflag=0; } + if (ret != MPI_SUCCESS) + error = 1; }else{ Status::empty(pstat); } @@ -668,7 +705,10 @@ int Request::testall(int count, MPI_Request requests[], MPI_Status status[]) status[i] = *pstat; } } - return flag; + if(error==1) + return MPI_ERR_IN_STATUS; + else + return MPI_SUCCESS; } void Request::probe(int source, int tag, MPI_Comm comm, MPI_Status* status){ @@ -707,7 +747,7 @@ void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* ->wait(); } // behave like a receive, but don't do it - s4u::MailboxPtr mailbox; + s4u::Mailbox* mailbox; request->print_request("New iprobe"); // We have to test both mailboxes as we don't know if we will receive one one or another @@ -757,7 +797,9 @@ void Request::finish_wait(MPI_Request* request, MPI_Status * status) return; } - if (not((req->detached_ != 0) && ((req->flags_ & MPI_REQ_SEND) != 0)) && ((req->flags_ & MPI_REQ_PREPARED) == 0)) { + if (not((req->detached_ != 0) && ((req->flags_ & MPI_REQ_SEND) != 0)) + && ((req->flags_ & MPI_REQ_PREPARED) == 0) + && ((req->flags_ & MPI_REQ_GENERALIZED) == 0)) { if(status != MPI_STATUS_IGNORE) { int src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_; status->MPI_SOURCE = req->comm_->group()->rank(src); @@ -817,12 +859,37 @@ void Request::finish_wait(MPI_Request* request, MPI_Status * status) unref(request); } -void Request::wait(MPI_Request * request, MPI_Status * status) +int Request::wait(MPI_Request * request, MPI_Status * status) { + int ret=MPI_SUCCESS; + // Are we waiting on a request meant for non blocking collectives ? + // If so, wait for all the subrequests. + if ((*request)->nbc_requests_size_>0){ + ret = waitall((*request)->nbc_requests_size_, (*request)->nbc_requests_, MPI_STATUSES_IGNORE); + for (int i = 0; i < (*request)->nbc_requests_size_; i++) { + if((*request)->buf_!=nullptr && (*request)->nbc_requests_[i]!=MPI_REQUEST_NULL){//reduce case + void * buf=(*request)->nbc_requests_[i]->buf_; + if((*request)->old_type_->flags() & DT_FLAG_DERIVED) + buf=(*request)->nbc_requests_[i]->old_buf_; + if((*request)->op_!=MPI_OP_NULL){ + int count=(*request)->size_/ (*request)->old_type_->size(); + (*request)->op_->apply(buf, (*request)->buf_, &count, (*request)->old_type_); + } + smpi_free_tmp_buffer(buf); + } + if((*request)->nbc_requests_[i]!=MPI_REQUEST_NULL) + Request::unref(&((*request)->nbc_requests_[i])); + } + delete[] (*request)->nbc_requests_; + (*request)->nbc_requests_size_=0; + unref(request); + return ret; + } + (*request)->print_request("Waiting"); if ((*request)->flags_ & MPI_REQ_PREPARED) { Status::empty(status); - return; + return ret; } if ((*request)->action_ != nullptr){ @@ -834,10 +901,28 @@ void Request::wait(MPI_Request * request, MPI_Status * status) } } + if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_GENERALIZED)){ + MPI_Status* mystatus; + if(!((*request)->flags_ & MPI_REQ_COMPLETE)){ + ((*request)->generalized_funcs)->mutex->lock(); + ((*request)->generalized_funcs)->cond->wait(((*request)->generalized_funcs)->mutex); + ((*request)->generalized_funcs)->mutex->unlock(); + } + if(status==MPI_STATUS_IGNORE){ + mystatus=new MPI_Status(); + Status::empty(mystatus); + }else{ + mystatus=status; + } + ret = ((*request)->generalized_funcs)->query_fn(((*request)->generalized_funcs)->extra_state, mystatus); + if(status==MPI_STATUS_IGNORE) + delete mystatus; + } finish_wait(request,status); if (*request != MPI_REQUEST_NULL && (((*request)->flags_ & MPI_REQ_NON_PERSISTENT) != 0)) *request = MPI_REQUEST_NULL; + return ret; } int Request::waitany(int count, MPI_Request requests[], MPI_Status * status) @@ -959,21 +1044,30 @@ int Request::waitall(int count, MPI_Request requests[], MPI_Status status[]) int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[]) { int count = 0; + int flag = 0; + int index = 0; MPI_Status stat; MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat; + index = waitany(incount, (MPI_Request*)requests, pstat); + if(index==MPI_UNDEFINED) return MPI_UNDEFINED; + if(status != MPI_STATUSES_IGNORE) { + status[count] = *pstat; + } + indices[count] = index; + count++; for (int i = 0; i < incount; i++) { - int index = waitany(incount, requests, pstat); - if(index!=MPI_UNDEFINED){ - indices[count] = index; - count++; - if(status != MPI_STATUSES_IGNORE) { - status[index] = *pstat; + if((requests[i] != MPI_REQUEST_NULL)) { + test(&requests[i], pstat,&flag); + if (flag==1){ + indices[count] = i; + if(status != MPI_STATUSES_IGNORE) { + status[count] = *pstat; + } + if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT)) + requests[i]=MPI_REQUEST_NULL; + count++; } - if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_NON_PERSISTENT)) - requests[index] = MPI_REQUEST_NULL; - }else{ - return MPI_UNDEFINED; } } return count; @@ -1005,5 +1099,72 @@ void Request::free_f(int id) } } + +int Request::get_status(MPI_Request req, int* flag, MPI_Status * status){ + *flag=0; + + if(req != MPI_REQUEST_NULL && req->action_ != nullptr) { + req->iprobe(req->src_, req->tag_, req->comm_, flag, status); + if(*flag) + return MPI_SUCCESS; + } + if (req != MPI_REQUEST_NULL && + (req->flags_ & MPI_REQ_GENERALIZED) + && !(req->flags_ & MPI_REQ_COMPLETE)) { + *flag=0; + return MPI_SUCCESS; + } + + *flag=1; + if(req != MPI_REQUEST_NULL && + status != MPI_STATUS_IGNORE) { + int src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_; + status->MPI_SOURCE = req->comm_->group()->rank(src); + status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_; + status->MPI_ERROR = req->truncated_ ? MPI_ERR_TRUNCATE : MPI_SUCCESS; + status->count = req->real_size_; + } + return MPI_SUCCESS; +} + +int Request::grequest_start( MPI_Grequest_query_function *query_fn, MPI_Grequest_free_function *free_fn, MPI_Grequest_cancel_function *cancel_fn, void *extra_state, MPI_Request *request){ + + *request = new Request(); + (*request)->flags_ |= MPI_REQ_GENERALIZED; + (*request)->flags_ |= MPI_REQ_PERSISTENT; + (*request)->refcount_ = 1; + ((*request)->generalized_funcs)=xbt_new0(s_smpi_mpi_generalized_request_funcs_t ,1); + ((*request)->generalized_funcs)->query_fn=query_fn; + ((*request)->generalized_funcs)->free_fn=free_fn; + ((*request)->generalized_funcs)->cancel_fn=cancel_fn; + ((*request)->generalized_funcs)->extra_state=extra_state; + ((*request)->generalized_funcs)->cond = simgrid::s4u::ConditionVariable::create(); + ((*request)->generalized_funcs)->mutex = simgrid::s4u::Mutex::create(); + return MPI_SUCCESS; +} + +int Request::grequest_complete( MPI_Request request){ + if ((!(request->flags_ & MPI_REQ_GENERALIZED)) || request->generalized_funcs->mutex==NULL) + return MPI_ERR_REQUEST; + request->generalized_funcs->mutex->lock(); + request->flags_ |= MPI_REQ_COMPLETE; // in case wait would be called after complete + request->generalized_funcs->cond->notify_one(); + request->generalized_funcs->mutex->unlock(); + return MPI_SUCCESS; +} + +void Request::set_nbc_requests(MPI_Request* reqs, int size){ + nbc_requests_=reqs; + nbc_requests_size_=size; +} + +int Request::get_nbc_requests_size(){ + return nbc_requests_size_; +} + +MPI_Request* Request::get_nbc_requests(){ + return nbc_requests_; +} + } }