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)
+ , old_buf_(buf_)
+ , 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 Request::init_buffer(int count){
- 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 (not smpi_process()->replaying() &&
+ ((((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_;
+ 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);
}
}
}
- old_buf_ = old_buf;
}
bool Request::match_recv(void* a, void* b, simgrid::kernel::activity::CommImpl*)
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 */
+ MPI_Request request;
if(op==MPI_OP_NULL){
request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src),
dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
MPI_Request Request::rma_recv_init(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 */
aid_t source = MPI_PROC_NULL;
if (src == MPI_ANY_SOURCE)
source = MPI_ANY_SOURCE;
else if (src != MPI_PROC_NULL)
source = comm->group()->actor(src);
+ MPI_Request request;
if(op==MPI_OP_NULL){
request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, source,
dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
MPI_Request Request::ibsend(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(),
- dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
- MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_BSEND);
+ auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
+ dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
+ MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_BSEND);
if(dst != MPI_PROC_NULL)
request->start();
return request;
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(),
- dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
- MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND);
+ auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
+ dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
+ MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND);
if(dst != MPI_PROC_NULL)
request->start();
return request;
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(),
- dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
- MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SSEND | MPI_REQ_SEND);
+ auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
+ dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
+ MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SSEND | MPI_REQ_SEND);
if(dst != MPI_PROC_NULL)
request->start();
return request;
MPI_Request Request::irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
{
- MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
aid_t source = MPI_PROC_NULL;
if (src == MPI_ANY_SOURCE)
source = MPI_ANY_SOURCE;
else if (src != MPI_PROC_NULL)
source = comm->group()->actor(src);
- request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
- source, simgrid::s4u::this_actor::get_pid(), tag, comm,
- MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV);
+ auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, source,
+ simgrid::s4u::this_actor::get_pid(), tag, comm, MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV);
if(src != MPI_PROC_NULL)
request->start();
return request;
int Request::recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
{
- MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
- request = irecv(buf, count, datatype, src, tag, comm);
+ MPI_Request request = irecv(buf, count, datatype, src, tag, comm);
int retval = wait(&request,status);
request = nullptr;
return retval;
void Request::bsend(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(),
- dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
- MPI_REQ_NON_PERSISTENT | MPI_REQ_SEND | MPI_REQ_BSEND);
+ auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
+ dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
+ MPI_REQ_NON_PERSISTENT | MPI_REQ_SEND | MPI_REQ_BSEND);
if(dst != MPI_PROC_NULL)
request->start();
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(),
- dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
- MPI_REQ_NON_PERSISTENT | MPI_REQ_SEND);
+ auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
+ dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
+ MPI_REQ_NON_PERSISTENT | MPI_REQ_SEND);
if(dst != MPI_PROC_NULL)
request->start();
wait(&request, MPI_STATUS_IGNORE);
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(),
- dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
- MPI_REQ_NON_PERSISTENT | MPI_REQ_SSEND | MPI_REQ_SEND);
+ auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
+ dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
+ MPI_REQ_NON_PERSISTENT | MPI_REQ_SSEND | MPI_REQ_SEND);
if(dst != MPI_PROC_NULL)
request->start();
//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_);
std::vector<simgrid::kernel::activity::CommImpl*> comms;
comms.reserve(count);
- int i;
*flag = 0;
int ret = MPI_SUCCESS;
*index = MPI_UNDEFINED;
std::vector<int> map; /** Maps all matching comms back to their location in requests **/
- for(i = 0; i < count; i++) {
+ for (int i = 0; i < count; i++) {
if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action_ && not(requests[i]->flags_ & MPI_REQ_PREPARED)) {
comms.push_back(static_cast<simgrid::kernel::activity::CommImpl*>(requests[i]->action_.get()));
map.push_back(i);
static int nsleeps = 1;
if(smpi_test_sleep > 0)
simgrid::s4u::this_actor::sleep_for(nsleeps * smpi_test_sleep);
+ ssize_t i;
try{
i = simcall_comm_testany(comms.data(), comms.size()); // The i-th element in comms matches!
} catch (const Exception&) {
XBT_DEBUG("Exception in testany");
return 0;
}
-
+
if (i != -1) { // -1 is not MPI_UNDEFINED but a SIMIX return code. (nothing matches)
*index = map[i];
if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_GENERALIZED) &&
}else{
*outflag=0;
}
- if (ret != MPI_SUCCESS)
+ if (ret != MPI_SUCCESS)
error = 1;
}else{
Status::empty(pstat);
status[i] = *pstat;
}
}
- if(error==1)
+ if (error == 1)
return MPI_ERR_IN_STATUS;
- else
+ else
return MPI_SUCCESS;
}
/*(seconds * flop/s -> total flops)*/ smpi_iprobe_sleep * speed * maxrate)
->set_name("iprobe")
/* Not the entire CPU can be used when iprobing: This is important for
- * the energy consumption caused by polling with iprobes.
+ * the energy consumption caused by polling with iprobes.
* Note also that the number of flops that was
* computed above contains a maxrate factor and is hence reduced (maxrate < 1)
*/
- ->set_bound(maxrate*speed)
+ ->set_bound(maxrate * speed)
->start()
->wait();
}
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() &&
+ (((req->flags_ & MPI_REQ_ACCUMULATE) != 0) || (datatype->flags() & DT_FLAG_DERIVED))) {
+ if (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 */
// This is a finished detached request, let's return this one
comms.clear(); // don't do the waitany call afterwards
index = i;
- if (requests[index] != MPI_REQUEST_NULL && (requests[index])->flags_ & MPI_REQ_NBC)
+ if (requests[index]->flags_ & MPI_REQ_NBC)
finish_nbc_requests(&requests[index], 0);
finish_wait(&requests[i], status); // cleanup if refcount = 0
if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
}
if (not comms.empty()) {
XBT_DEBUG("Enter waitany for %zu comms", comms.size());
- int i;
+ ssize_t i;
try{
i = simcall_comm_waitany(comms.data(), comms.size(), -1);
} catch (const CancelException&) {
}
}
- 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;
}
indices[count] = index;
count++;
for (int i = 0; i < incount; i++) {
- if (i!=index && requests[i] != MPI_REQUEST_NULL
- && not(requests[i]->flags_ & MPI_REQ_FINISHED)) {
+ if (i != index && requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & MPI_REQ_FINISHED)) {
test(&requests[i], pstat,&flag);
if (flag==1){
indices[count] = i;