/* 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"
#include "src/kernel/activity/CommImpl.hpp"
#include "src/mc/mc_replay.hpp"
#include "src/smpi/include/smpi_actor.hpp"
-#include "xbt/config.hpp"
-
#include <algorithm>
}
}
}
- // This part handles the problem of non-contiguous memory (for the unserialisation at the reception)
+ // This part handles the problem of non-contiguous memory (for the unserialization at the reception)
old_buf_ = old_buf;
size_ = datatype->size() * count;
datatype->ref();
{
MPI_Request ref = static_cast<MPI_Request>(a);
MPI_Request req = static_cast<MPI_Request>(b);
- XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
+ XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d, id %d against %d",ref->src_,req->src_, ref->tag_, req->tag_,ref->comm_->id(),req->comm_->id());
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->comm_->id()==MPI_UNDEFINED || req->comm_->id() == MPI_UNDEFINED || (ref->comm_->id()==req->comm_->id()))
+ && ((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 (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
+ req->cancelled_ = -1; // mark as uncancelable
XBT_DEBUG("match succeeded");
return 1;
}else return 0;
{
MPI_Request ref = static_cast<MPI_Request>(a);
MPI_Request req = static_cast<MPI_Request>(b);
- XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
+ XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d, id %d against %d",ref->src_,req->src_, ref->tag_, req->tag_,ref->comm_->id(),req->comm_->id());
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((ref->comm_->id()==MPI_UNDEFINED || req->comm_->id() == MPI_UNDEFINED || (ref->comm_->id()==req->comm_->id()))
+ && ((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 (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
+ req->cancelled_ = -1; // mark as uncancelable
XBT_DEBUG("match succeeded");
return 1;
} else
/* factories, to hide the internal flags from the caller */
+MPI_Request Request::bsend_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_SEND | MPI_REQ_PREPARED | MPI_REQ_BSEND);
+}
+
MPI_Request Request::send_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
{
MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
}
+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(),
+ comm->group()->actor(dst)->get_pid(), tag, comm,
+ MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_BSEND);
+ 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 = nullptr;
}
+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(),
+ comm->group()->actor(dst)->get_pid(), tag, comm, MPI_REQ_NON_PERSISTENT | MPI_REQ_SEND | MPI_REQ_BSEND);
+
+ request->start();
+ wait(&request, MPI_STATUS_IGNORE);
+ request = nullptr;
+}
+
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 */
mut->unlock();
} else { /* the RECV flag was not set, so this is a send */
simgrid::smpi::ActorExt* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_));
+ xbt_assert(process, "Actor pid=%d is gone??", dst_);
int rank = src_;
if (TRACE_smpi_view_internals()) {
TRACE_smpi_send(rank, rank, dst_, tag_, size_);
void* buf = buf_;
if ((flags_ & MPI_REQ_SSEND) == 0 &&
- ((flags_ & MPI_REQ_RMA) != 0 ||
+ ((flags_ & MPI_REQ_RMA) != 0 || (flags_ & MPI_REQ_BSEND) != 0 ||
static_cast<int>(size_) < simgrid::config::get_value<int>("smpi/send-is-detached-thresh"))) {
void *oldbuf = nullptr;
detached_ = true;
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_);
memcpy(buf,oldbuf,size_);
XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
}
if(sleeptime > 0.0){
- simcall_process_sleep(sleeptime);
+ simgrid::s4u::this_actor::sleep_for(sleeptime);
XBT_DEBUG("sending size of %zu : sleep %f ", size_, sleeptime);
}
action_ = simcall_comm_isend(
simgrid::s4u::Actor::by_pid(src_)->get_impl(), mailbox->get_impl(), size_, -1.0, buf, real_size_, &match_send,
&xbt_free_f, // how to free the userdata if a detached send fails
- not process->replaying() ? smpi_comm_copy_data_callback : &smpi_comm_null_copy_buffer_callback, this,
+ process->replaying() ? &smpi_comm_null_copy_buffer_callback : smpi_comm_copy_data_callback, this,
// detach if msg size < eager/rdv switch limit
detached_);
XBT_DEBUG("send simcall posted");
}
if(smpi_test_sleep > 0)
- simcall_process_sleep(nsleeps*smpi_test_sleep);
+ simgrid::s4u::this_actor::sleep_for(nsleeps * smpi_test_sleep);
Status::empty(status);
*flag = 1;
if (((*request)->flags_ & MPI_REQ_PREPARED) == 0) {
- if ((*request)->action_ != nullptr){
+ if ((*request)->action_ != nullptr && (*request)->cancelled_ != 1){
try{
*flag = simcall_comm_test((*request)->action_);
- } catch (const xbt_ex&) {
+ } catch (const Exception&) {
*flag = 0;
return ret;
}
int Request::testsome(int incount, MPI_Request requests[], int *count, int *indices, MPI_Status status[])
{
- int ret = MPI_SUCCESS;
int error=0;
int count_dead = 0;
int flag = 0;
*count = 0;
for (int i = 0; i < incount; i++) {
- if (requests[i] != MPI_REQUEST_NULL) {
- ret = test(&requests[i], pstat, &flag);
+ if (requests[i] != MPI_REQUEST_NULL && not (requests[i]->flags_ & MPI_REQ_FINISHED)) {
+ int ret = test(&requests[i], pstat, &flag);
if(ret!=MPI_SUCCESS)
error = 1;
if(flag) {
- indices[i] = 1;
- (*count)++;
+ indices[*count] = i;
if (status != MPI_STATUSES_IGNORE)
- status[i] = *pstat;
+ status[*count] = *pstat;
+ (*count)++;
if ((requests[i] != MPI_REQUEST_NULL) && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
requests[i] = MPI_REQUEST_NULL;
}
//multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
static int nsleeps = 1;
if(smpi_test_sleep > 0)
- simcall_process_sleep(nsleeps*smpi_test_sleep);
+ simgrid::s4u::this_actor::sleep_for(nsleeps * smpi_test_sleep);
try{
i = simcall_comm_testany(comms.data(), comms.size()); // The i-th element in comms matches!
- } catch (const xbt_ex&) {
+ } catch (const Exception&) {
+ XBT_DEBUG("Exception in testany");
return 0;
}
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 *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
int flag;
int error = 0;
- int ret=MPI_SUCCESS;
*outflag = 1;
for(int i=0; i<count; i++){
if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & MPI_REQ_PREPARED)) {
- ret = test(&requests[i], pstat, &flag);
+ int ret = test(&requests[i], pstat, &flag);
if (flag){
flag=0;
requests[i]=MPI_REQUEST_NULL;
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
+ // We have to test both mailboxes as we don't know if we will receive one or another
if (simgrid::config::get_value<int>("smpi/async-small-thresh") > 0) {
mailbox = smpi_process()->mailbox_small();
XBT_DEBUG("Trying to probe the perm recv mailbox");
return;
}
- if (not(req->detached_ && ((req->flags_ & MPI_REQ_SEND) != 0)) && ((req->flags_ & MPI_REQ_PREPARED) == 0) &&
- ((req->flags_ & MPI_REQ_GENERALIZED) == 0)) {
+ if ((req->flags_ & (MPI_REQ_PREPARED | MPI_REQ_GENERALIZED | MPI_REQ_FINISHED)) == 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);
// this handles the case were size in receive differs from size in send
status->count = req->real_size_;
}
+ //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_;
+
+ // 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_privatize_global_variables != 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) {
+ XBT_VERB("Privatization : We are unserializing to a zone in global memory Switch data segment ");
+ smpi_switch_data_segment(simgrid::s4u::Actor::self());
+ }
- req->print_request("Finishing");
- MPI_Datatype datatype = req->old_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_privatize_global_variables != 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) {
- 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-contignous memory the unserialization at the reception
- if ((req->flags_ & MPI_REQ_RECV) && datatype->size() != 0)
- datatype->unserialize(req->buf_, req->old_buf_, req->real_size_/datatype->size() , req->op_);
- xbt_free(req->buf_);
- } else if (req->flags_ & MPI_REQ_RECV) { // apply op on contiguous buffer for accumulate
- if (datatype->size() != 0) {
- int n = req->real_size_ / datatype->size();
- req->op_->apply(req->buf_, req->old_buf_, &n, datatype);
+ if(datatype->flags() & DT_FLAG_DERIVED){
+ // This part handles the problem of non-contiguous memory the unserialization at the reception
+ if ((req->flags_ & MPI_REQ_RECV) && datatype->size() != 0)
+ datatype->unserialize(req->buf_, req->old_buf_, req->real_size_/datatype->size() , req->op_);
+ xbt_free(req->buf_);
+ } else if (req->flags_ & MPI_REQ_RECV) { // apply op on contiguous buffer for accumulate
+ if (datatype->size() != 0) {
+ int n = req->real_size_ / datatype->size();
+ req->op_->apply(req->buf_, req->old_buf_, &n, datatype);
+ }
+ xbt_free(req->buf_);
}
- xbt_free(req->buf_);
}
}
}
//integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
double sleeptime =
simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->orecv(req->real_size());
- if(sleeptime > 0.0){
- simcall_process_sleep(sleeptime);
+ if (sleeptime > 0.0) {
+ simgrid::s4u::this_actor::sleep_for(sleeptime);
XBT_DEBUG("receiving size of %zu : sleep %f ", req->real_size_, sleeptime);
}
unref(&(req->detached_sender_));
try{
// this is not a detached send
simcall_comm_wait((*request)->action_, -1.0);
- } catch (const xbt_ex&) {
+ } catch (const Exception&) {
XBT_VERB("Request cancelled");
}
}
try{
// this is not a detached send
i = simcall_comm_waitany(comms.data(), comms.size(), -1);
- } catch (const xbt_ex&) {
+ } catch (const Exception&) {
XBT_INFO("request %d cancelled ", i);
return i;
}
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) {
indices[count] = index;
count++;
for (int i = 0; i < incount; i++) {
- if (requests[i] != MPI_REQUEST_NULL) {
+ 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;
char key[KEY_SIZE];
if(id==MPI_FORTRAN_REQUEST_NULL)
return static_cast<MPI_Request>(MPI_REQUEST_NULL);
- return static_cast<MPI_Request>(F2C::f2c_lookup()->at(get_key_id(key, id)));
-}
-
-int Request::add_f()
-{
- if (F2C::f2c_lookup() == nullptr) {
- F2C::set_f2c_lookup(new std::unordered_map<std::string, F2C*>);
- }
- char key[KEY_SIZE];
- (*(F2C::f2c_lookup()))[get_key_id(key, F2C::f2c_id())] = this;
- F2C::f2c_id_increment();
- return F2C::f2c_id()-1;
+ return static_cast<MPI_Request>(F2C::f2c_lookup()->at(get_key(key,id)));
}
void Request::free_f(int id)
{
if (id != MPI_FORTRAN_REQUEST_NULL) {
char key[KEY_SIZE];
- F2C::f2c_lookup()->erase(get_key_id(key, id));
+ F2C::f2c_lookup()->erase(get_key(key, id));
}
}
-
int Request::get_status(MPI_Request req, int* flag, MPI_Status * status){
*flag=0;