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.
size_ = datatype->size() * count;
datatype->ref();
comm_->ref();
+ if(op != MPI_REPLACE && op != MPI_OP_NULL)
+ op_->ref();
action_ = nullptr;
- detached_ = 0;
+ detached_ = false;
detached_sender_ = nullptr;
real_src_ = 0;
- truncated_ = 0;
+ truncated_ = false;
real_size_ = 0;
real_tag_ = 0;
if (flags & MPI_REQ_PERSISTENT)
refcount_ = 1;
else
refcount_ = 0;
- op_ = MPI_REPLACE;
cancelled_ = 0;
+ generalized_funcs=nullptr;
+ nbc_requests_=nullptr;
+ nbc_requests_size_=0;
}
void Request::ref(){
Comm::unref((*request)->comm_);
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;
if(ref->tag_ == MPI_ANY_TAG)
ref->real_tag_ = req->tag_;
if(ref->real_size_ < req->real_size_)
- ref->truncated_ = 1;
- if(req->detached_==1)
+ ref->truncated_ = true;
+ if (req->detached_)
ref->detached_sender_=req; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
if(req->cancelled_==0)
req->cancelled_=-1;//mark as uncancellable
if(req->tag_ == MPI_ANY_TAG)
req->real_tag_ = ref->tag_;
if(req->real_size_ < ref->real_size_)
- req->truncated_ = 1;
- if(ref->detached_==1)
+ req->truncated_ = true;
+ if (ref->detached_)
req->detached_sender_=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
if(req->cancelled_==0)
req->cancelled_=-1;//mark as uncancellable
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;
}
}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;
}
((flags_ & MPI_REQ_RMA) != 0 ||
static_cast<int>(size_) < simgrid::config::get_value<int>("smpi/send-is-detached-thresh"))) {
void *oldbuf = nullptr;
- detached_ = 1;
+ detached_ = true;
XBT_DEBUG("Send request %p is detached", this);
this->ref();
if (not(old_type_->flags() & DT_FLAG_DERIVED)) {
//if we are giving back the control to the user without waiting for completion, we have to inject timings
double sleeptime = 0.0;
- if (detached_ != 0 || ((flags_ & (MPI_REQ_ISEND | MPI_REQ_SSEND)) != 0)) { // issend should be treated as isend
+ if (detached_ || ((flags_ & (MPI_REQ_ISEND | MPI_REQ_SSEND)) != 0)) { // issend should be treated as isend
// isend and send timings may be different
sleeptime = ((flags_ & MPI_REQ_ISEND) != 0)
? simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->oisend(size_)
/* FIXME: detached sends are not traceable (action_ == nullptr) */
if (action_ != nullptr) {
- std::string category = smpi_process()->get_tracing_category();
- simgrid::simix::simcall([this, category] { this->action_->set_category(category); });
+ boost::static_pointer_cast<kernel::activity::CommImpl>(action_)->set_tracing_category(
+ smpi_process()->get_tracing_category());
}
if (async_small_thresh != 0 || ((flags_ & MPI_REQ_RMA) != 0))
// 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);
*flag = 1;
if (((*request)->flags_ & MPI_REQ_PREPARED) == 0) {
if (*flag) {
finish_wait(request,status);
if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_GENERALIZED)){
+ MPI_Status* mystatus;
if(status==MPI_STATUS_IGNORE){
mystatus=new MPI_Status();
Status::empty(mystatus);
int i;
*flag = 0;
int ret = MPI_SUCCESS;
- MPI_Status* mystatus;
*index = MPI_UNDEFINED;
std::vector<int> map; /** Maps all matching comms back to their location in requests **/
} else {
finish_wait(&requests[*index],status);
if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_GENERALIZED)){
+ MPI_Status* mystatus;
if(status==MPI_STATUS_IGNORE){
mystatus=new MPI_Status();
Status::empty(mystatus);
{
MPI_Status stat;
MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
- int flag, error=0;
+ int flag;
+ int error = 0;
int ret=MPI_SUCCESS;
*outflag = 1;
for(int i=0; i<count; i++){
nsleeps++;
}
unref(&request);
+ xbt_assert(request == MPI_REQUEST_NULL);
}
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)
- && ((req->flags_ & MPI_REQ_GENERALIZED) == 0)) {
+ if (not(req->detached_ && ((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);
status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
- status->MPI_ERROR = req->truncated_ != 0 ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
+ status->MPI_ERROR = req->truncated_ ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
// this handles the case were size in receive differs from size in send
status->count = req->real_size_;
}
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)->nbc_requests_[i]->flags_ & MPI_REQ_RECV ){
+ 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);
+ (*request)=MPI_REQUEST_NULL;
+ return ret;
+ }
+
(*request)->print_request("Waiting");
if ((*request)->flags_ & MPI_REQ_PREPARED) {
Status::empty(status);
indices[count] = index;
count++;
for (int i = 0; i < incount; i++) {
- if((requests[i] != MPI_REQUEST_NULL)) {
+ if (requests[i] != MPI_REQUEST_NULL) {
test(&requests[i], pstat,&flag);
if (flag==1){
indices[count] = i;
}
+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();
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_;
+}
+
}
}
