Request::Request(const void* buf, int count, MPI_Datatype datatype, aid_t src, aid_t dst, int tag, MPI_Comm comm,
unsigned flags, MPI_Op op)
: buf_(const_cast<void*>(buf))
- , old_type_(datatype)
+ , type_(datatype)
, size_(datatype->size() * count)
, src_(src)
, dst_(dst)
detached_sender_ = nullptr;
real_src_ = 0;
truncated_ = false;
+ unmatched_types_ = false;
real_size_ = 0;
real_tag_ = 0;
if (flags & MPI_REQ_PERSISTENT)
((*request)->generalized_funcs)->free_fn(((*request)->generalized_funcs)->extra_state);
} else {
Comm::unref((*request)->comm_);
- Datatype::unref((*request)->old_type_);
+ Datatype::unref((*request)->type_);
}
if ((*request)->op_ != MPI_REPLACE && (*request)->op_ != MPI_OP_NULL)
Op::unref(&(*request)->op_);
}
}
+bool Request::match_types(MPI_Datatype stype, MPI_Datatype rtype){
+ bool match = false;
+ if ((stype == rtype) ||
+ //byte and packed always match with anything
+ (stype == MPI_PACKED || rtype == MPI_PACKED || stype == MPI_BYTE || rtype == MPI_BYTE) ||
+ //complex datatypes - we don't properly match these yet, as it would mean checking each subtype recursively.
+ (stype->flags() & DT_FLAG_DERIVED || rtype->flags() & DT_FLAG_DERIVED) ||
+ //duplicated datatypes, check if underlying is ok
+ (stype->duplicated_datatype()!=MPI_DATATYPE_NULL && match_types(stype->duplicated_datatype(), rtype)) ||
+ (rtype->duplicated_datatype()!=MPI_DATATYPE_NULL && match_types(stype, rtype->duplicated_datatype())))
+ match = true;
+ if (!match)
+ XBT_WARN("Mismatched datatypes : sending %s and receiving %s", stype->name().c_str(), rtype->name().c_str());
+ return match;
+}
+
+
bool Request::match_common(MPI_Request req, MPI_Request sender, MPI_Request receiver)
{
xbt_assert(sender, "Cannot match against null sender");
receiver->real_size_=sender->real_size_;
}
}
+ //0-sized datatypes/counts should not interfere and match
+ if ( sender->real_size_ != 0 && receiver->real_size_ != 0 &&
+ !match_types(sender->type_, receiver->type_))
+ receiver->unmatched_types_ = true;
if (sender->detached_)
receiver->detached_sender_ = sender; // tie the sender to the receiver, as it is detached and has to be freed in
// the receiver
void *old_buf = nullptr;
// FIXME Handle the case of a partial shared malloc.
// This part handles the problem of non-contiguous memory (for the unserialization at the reception)
- if ((((flags_ & MPI_REQ_RECV) != 0) && ((flags_ & MPI_REQ_ACCUMULATE) != 0)) || (old_type_->flags() & DT_FLAG_DERIVED)) {
+ if ((((flags_ & MPI_REQ_RECV) != 0) && ((flags_ & MPI_REQ_ACCUMULATE) != 0)) || (type_->flags() & DT_FLAG_DERIVED)) {
// This part handles the problem of non-contiguous memory
old_buf = buf_;
if (count==0){
buf_ = nullptr;
}else {
- buf_ = xbt_malloc(count*old_type_->size());
- if ((old_type_->flags() & DT_FLAG_DERIVED) && ((flags_ & MPI_REQ_SEND) != 0)) {
- old_type_->serialize(old_buf, buf_, count);
+ buf_ = xbt_malloc(count*type_->size());
+ if ((type_->flags() & DT_FLAG_DERIVED) && ((flags_ & MPI_REQ_SEND) != 0)) {
+ type_->serialize(old_buf, buf_, count);
}
}
}
//reinitialize temporary buffer for persistent requests
if(real_size_ > 0 && flags_ & MPI_REQ_FINISHED){
buf_ = old_buf_;
- init_buffer(real_size_/old_type_->size());
+ init_buffer(real_size_/type_->size());
}
flags_ &= ~MPI_REQ_PREPARED;
flags_ &= ~MPI_REQ_FINISHED;
detached_ = true;
XBT_DEBUG("Send request %p is detached", this);
this->ref();
- if (not(old_type_->flags() & DT_FLAG_DERIVED)) {
+ if (not(type_->flags() & DT_FLAG_DERIVED)) {
oldbuf = buf_;
if (not process->replaying() && oldbuf != nullptr && size_ != 0) {
- if ((smpi_cfg_privatization() != SmpiPrivStrategies::NONE) &&
- (static_cast<char*>(buf_) >= smpi_data_exe_start) &&
- (static_cast<char*>(buf_) < smpi_data_exe_start + smpi_data_exe_size)) {
+ if (smpi_switch_data_segment(simgrid::s4u::Actor::by_pid(src_), buf_))
XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
- smpi_switch_data_segment(simgrid::s4u::Actor::by_pid(src_));
- }
+
//we need this temporary buffer even for bsend, as it will be released in the copy callback and we don't have a way to differentiate it
//so actually ... don't use manually attached buffer space.
buf = xbt_malloc(size_);
for(auto& req: (*request)->nbc_requests_){
if((*request)->buf_!=nullptr && req!=MPI_REQUEST_NULL){//reduce case
void * buf=req->buf_;
- if((*request)->old_type_->flags() & DT_FLAG_DERIVED)
+ if((*request)->type_->flags() & DT_FLAG_DERIVED)
buf=req->old_buf_;
if(req->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_);
+ int count=(*request)->size_/ (*request)->type_->size();
+ (*request)->op_->apply(buf, (*request)->buf_, &count, (*request)->type_);
}
smpi_free_tmp_buffer(static_cast<unsigned char*>(buf));
}
//detached send will be finished at the other end
if (not(req->detached_ && ((req->flags_ & MPI_REQ_SEND) != 0))) {
req->print_request("Finishing");
- MPI_Datatype datatype = req->old_type_;
+ MPI_Datatype datatype = req->type_;
// FIXME Handle the case of a partial shared malloc.
if (((req->flags_ & MPI_REQ_ACCUMULATE) != 0) ||
(datatype->flags() & DT_FLAG_DERIVED)) { // && (not smpi_is_shared(req->old_buf_))){
- if (not smpi_process()->replaying() && smpi_cfg_privatization() != SmpiPrivStrategies::NONE &&
- static_cast<char*>(req->old_buf_) >= smpi_data_exe_start &&
- static_cast<char*>(req->old_buf_) < smpi_data_exe_start + smpi_data_exe_size) {
+ if (not smpi_process()->replaying() && smpi_switch_data_segment(simgrid::s4u::Actor::self(), req->old_buf_))
XBT_VERB("Privatization : We are unserializing to a zone in global memory Switch data segment ");
- smpi_switch_data_segment(simgrid::s4u::Actor::self());
- }
if(datatype->flags() & DT_FLAG_DERIVED){
// This part handles the problem of non-contiguous memory the unserialization at the reception
req->action_ = nullptr;
req->flags_ |= MPI_REQ_FINISHED;
- if (req->truncated_) {
+ if (req->truncated_ || req->unmatched_types_) {
char error_string[MPI_MAX_ERROR_STRING];
int error_size;
- PMPI_Error_string(MPI_ERR_TRUNCATE, error_string, &error_size);
+ int errkind;
+ if(req->truncated_ )
+ errkind = MPI_ERR_TRUNCATE;
+ else
+ errkind = MPI_ERR_TYPE;
+ PMPI_Error_string(errkind, error_string, &error_size);
MPI_Errhandler err = (req->comm_) ? (req->comm_)->errhandler() : MPI_ERRHANDLER_NULL;
if (err == MPI_ERRHANDLER_NULL || err == MPI_ERRORS_RETURN)
XBT_WARN("recv - returned %.*s instead of MPI_SUCCESS", error_size, error_string);
else if (err == MPI_ERRORS_ARE_FATAL)
xbt_die("recv - returned %.*s instead of MPI_SUCCESS", error_size, error_string);
else
- err->call((req->comm_), MPI_ERR_TRUNCATE);
+ err->call((req->comm_), errkind);
if (err != MPI_ERRHANDLER_NULL)
simgrid::smpi::Errhandler::unref(err);
MC_assert(not MC_is_active()); /* Only fail in MC mode */
}
}
- if (not accumulates.empty()) {
- std::sort(accumulates.begin(), accumulates.end(), sort_accumulates);
- for (auto& req : accumulates) {
- finish_wait(&req, status);
- }
- }
+ std::sort(accumulates.begin(), accumulates.end(), sort_accumulates);
+ for (auto& req : accumulates)
+ finish_wait(&req, status);
return retvalue;
}
