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(const void* buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm, unsigned flags, MPI_Op op)
+ : buf_(const_cast<void*>(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.
if ((((flags & MPI_REQ_RECV) != 0) && ((flags & MPI_REQ_ACCUMULATE) != 0)) || (datatype->flags() & DT_FLAG_DERIVED)) {
// This part handles the problem of non-contiguous memory
- old_buf = buf;
+ old_buf = const_cast<void*>(buf);
if (count==0){
buf_ = nullptr;
}else {
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;
if((*request)->refcount_==0){
if ((*request)->flags_ & MPI_REQ_GENERALIZED){
((*request)->generalized_funcs)->free_fn(((*request)->generalized_funcs)->extra_state);
+ delete (*request)->generalized_funcs;
}else{
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;
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_)
+ if(((ref->src_ == MPI_ANY_SOURCE && (ref->comm_->group()->rank(req->src_) != MPI_UNDEFINED)) || req->src_ == ref->src_)
&& ((ref->tag_ == MPI_ANY_TAG && req->tag_ >=0) || req->tag_ == ref->tag_)){
//we match, we can transfer some values
if(ref->src_ == MPI_ANY_SOURCE)
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
xbt_assert(ref, "Cannot match send against null reference");
xbt_assert(req, "Cannot match send against null request");
- if((req->src_ == MPI_ANY_SOURCE || req->src_ == ref->src_)
+ if(((req->src_ == MPI_ANY_SOURCE && (req->comm_->group()->rank(ref->src_) != MPI_UNDEFINED)) || req->src_ == ref->src_)
&& ((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_;
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
/* factories, to hide the internal flags from the caller */
-MPI_Request Request::send_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
+MPI_Request Request::send_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
{
return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
MPI_REQ_PERSISTENT | MPI_REQ_SEND | MPI_REQ_PREPARED);
}
-MPI_Request Request::ssend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
+MPI_Request Request::ssend_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
{
return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
comm->group()->actor(dst)->get_pid(), tag, comm,
MPI_REQ_PERSISTENT | MPI_REQ_SSEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
}
-MPI_Request Request::isend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
+MPI_Request Request::isend_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
{
return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
comm->group()->actor(dst)->get_pid(), tag, comm,
}
-MPI_Request Request::rma_send_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
+MPI_Request Request::rma_send_init(const void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
MPI_Op op)
{
MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
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;
}
MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
}
-MPI_Request Request::isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
+MPI_Request Request::isend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
{
MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
return request;
}
-MPI_Request Request::issend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
+MPI_Request Request::issend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
{
MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
request = nullptr;
}
-void Request::send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
+void Request::send(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
{
MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
request = nullptr;
}
-void Request::ssend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
+void Request::ssend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
{
MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
request = nullptr;
}
-void Request::sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
+void Request::sendrecv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
void *recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag,
MPI_Comm comm, MPI_Status * status)
{
((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))
static int nsleeps = 1;
int ret = MPI_SUCCESS;
- // are we testing a request meant for non blocking comms ?
+ // 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(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 ((*request)->action_ != nullptr){
try{
*flag = simcall_comm_test((*request)->action_);
- }catch (xbt_ex& e) {
+ } catch (const xbt_ex&) {
*flag = 0;
return ret;
}
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 **/
simcall_process_sleep(nsleeps*smpi_test_sleep);
try{
i = simcall_comm_testany(comms.data(), comms.size()); // The i-th element in comms matches!
- }catch (xbt_ex& e) {
+ } catch (const xbt_ex&) {
+ XBT_DEBUG("Exception in testany");
return 0;
}
} 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);
if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_NON_PERSISTENT))
requests[*index] = MPI_REQUEST_NULL;
+ XBT_DEBUG("Testany - returning with index %d", *index);
*flag=1;
}
nsleeps = 1;
nsleeps++;
}
} else {
+ XBT_DEBUG("Testany on inactive handles, returning flag=1 but empty status");
//all requests are null or inactive, return true
*flag = 1;
+ *index = MPI_UNDEFINED;
Status::empty(status);
}
{
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 flag;
-// Request::get_status(req,&flag,status);
}
req->print_request("Finishing");
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(static_cast<unsigned char*>(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);
try{
// this is not a detached send
simcall_comm_wait((*request)->action_, -1.0);
- }catch (xbt_ex& e) {
+ } catch (const xbt_ex&) {
XBT_VERB("Request cancelled");
}
}
try{
// this is not a detached send
i = simcall_comm_waitany(comms.data(), comms.size(), -1);
- }catch (xbt_ex& e) {
- XBT_INFO("request %d cancelled ",i);
+ } catch (const xbt_ex&) {
+ XBT_INFO("request %d cancelled ", i);
return i;
}
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;
if(req != MPI_REQUEST_NULL && req->action_ != nullptr) {
req->iprobe(req->src_, req->tag_, req->comm_, flag, status);
- if(flag)
+ if(*flag)
return MPI_SUCCESS;
}
if (req != MPI_REQUEST_NULL &&
}
*flag=1;
- if(status != MPI_STATUS_IGNORE) {
+ 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_;
(*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) = new s_smpi_mpi_generalized_request_funcs_t;
((*request)->generalized_funcs)->query_fn=query_fn;
((*request)->generalized_funcs)->free_fn=free_fn;
((*request)->generalized_funcs)->cancel_fn=cancel_fn;
}
void Request::set_nbc_requests(MPI_Request* reqs, int size){
- nbc_requests_=reqs;
- nbc_requests_size_=size;
+ nbc_requests_size_ = size;
+ if (size > 0) {
+ nbc_requests_ = reqs;
+ } else {
+ delete[] reqs;
+ nbc_requests_ = nullptr;
+ }
+}
+
+int Request::get_nbc_requests_size(){
+ return nbc_requests_size_;
+}
+
+MPI_Request* Request::get_nbc_requests(){
+ return nbc_requests_;
}
}
