-static MPI_Request build_request(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
- unsigned flags)
-{
- MPI_Request request = NULL;
-
- void *old_buf = NULL;
-
- request = xbt_new(s_smpi_mpi_request_t, 1);
-
- s_smpi_subtype_t *subtype = static_cast<s_smpi_subtype_t*>(datatype->substruct);
-
- if((((flags & RECV) != 0) && ((flags & ACCUMULATE) !=0)) || (datatype->sizeof_substruct != 0)){
- // This part handles the problem of non-contiguous memory
- old_buf = buf;
- buf = count==0 ? NULL : xbt_malloc(count*smpi_datatype_size(datatype));
- if ((datatype->sizeof_substruct != 0) && ((flags & SEND) != 0)) {
- subtype->serialize(old_buf, buf, count, datatype->substruct);
- }
- }
-
- request->buf = buf;
- // This part handles the problem of non-contiguous memory (for the unserialisation at the reception)
- request->old_buf = old_buf;
- request->old_type = datatype;
-
- request->size = smpi_datatype_size(datatype) * count;
- smpi_datatype_use(datatype);
- request->src = src;
- request->dst = dst;
- request->tag = tag;
- request->comm = comm;
- smpi_comm_use(request->comm);
- request->action = nullptr;
- request->flags = flags;
- request->detached = 0;
- request->detached_sender = nullptr;
- request->real_src = 0;
-
- request->truncated = 0;
- request->real_size = 0;
- request->real_tag = 0;
- if (flags & PERSISTENT)
- request->refcount = 1;
- else
- request->refcount = 0;
- request->op = MPI_REPLACE;
- request->send = 0;
- request->recv = 0;
-
- return request;
-}
-
-void smpi_empty_status(MPI_Status * status)
-{
- if(status != MPI_STATUS_IGNORE) {
- status->MPI_SOURCE = MPI_ANY_SOURCE;
- status->MPI_TAG = MPI_ANY_TAG;
- status->MPI_ERROR = MPI_SUCCESS;
- status->count=0;
- }
-}
-
-static void smpi_mpi_request_free_voidp(void* request)
-{
- MPI_Request req = static_cast<MPI_Request>(request);
- smpi_mpi_request_free(&req);
-}
-
-/* MPI Low level calls */
-MPI_Request smpi_mpi_send_init(void *buf, int count, MPI_Datatype datatype,
- int dst, int tag, MPI_Comm comm)
-{
- MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
- request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, smpi_process_index(),
- smpi_group_index(smpi_comm_group(comm), dst), tag, comm, PERSISTENT | SEND | PREPARED);
- return request;
-}
-
-MPI_Request smpi_mpi_ssend_init(void *buf, int count, MPI_Datatype datatype,
- int dst, int tag, MPI_Comm comm)
-{
- MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
- request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, smpi_process_index(),
- smpi_group_index(smpi_comm_group(comm), dst), tag, comm, PERSISTENT | SSEND | SEND | PREPARED);
- return request;
-}
-
-MPI_Request smpi_mpi_recv_init(void *buf, int count, MPI_Datatype datatype,
- int src, int tag, MPI_Comm comm)
-{
- MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
- request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype,
- src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src),
- smpi_process_index(), tag, comm, PERSISTENT | RECV | PREPARED);
- return request;
-}
-
-void smpi_mpi_start(MPI_Request request)
-{
- smx_mailbox_t mailbox;
-
- xbt_assert(request->action == nullptr, "Cannot (re-)start unfinished communication");
- request->flags &= ~PREPARED;
- request->flags &= ~FINISHED;
- request->refcount++;
-
- if ((request->flags & RECV) != 0) {
- print_request("New recv", request);
-
- int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
-
- xbt_mutex_t mut = smpi_process_mailboxes_mutex();
- if (async_small_thresh != 0 || (request->flags & RMA) != 0)
- xbt_mutex_acquire(mut);
-
- if (async_small_thresh == 0 && (request->flags & RMA) == 0 ) {
- mailbox = smpi_process_mailbox();
- } else if (((request->flags & RMA) != 0) || static_cast<int>(request->size) < async_small_thresh){
- //We have to check both mailboxes (because SSEND messages are sent to the large mbox).
- //begin with the more appropriate one : the small one.
- mailbox = smpi_process_mailbox_small();
- XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %p (in case of SSEND)?", mailbox);
- smx_synchro_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, static_cast<void*>(request));
-
- if(action == nullptr){
- mailbox = smpi_process_mailbox();
- XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox);
- action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, static_cast<void*>(request));
- if(action == nullptr){
- XBT_DEBUG("Still nothing, switch back to the small mailbox : %p", mailbox);
- mailbox = smpi_process_mailbox_small();
- }
- }else{
- XBT_DEBUG("yes there was something for us in the large mailbox");
- }
- }else{
- mailbox = smpi_process_mailbox_small();
- XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
- smx_synchro_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
-
- if(action == nullptr){
- XBT_DEBUG("No, nothing in the permanent receive mailbox");
- mailbox = smpi_process_mailbox();
- }else{
- XBT_DEBUG("yes there was something for us in the small mailbox");
- }
- }
-
- //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
- double sleeptime = (request->detached != 0) ? smpi_or(request->size) : 0.0;
- if(sleeptime > 0.0){
- simcall_process_sleep(sleeptime);
- XBT_DEBUG("receiving size of %zu : sleep %f ", request->size, smpi_or(request->size));
- }
-
- // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
- request->real_size=request->size;
- request->action = simcall_comm_irecv(SIMIX_process_self(), mailbox, request->buf, &request->real_size, &match_recv,
- ! smpi_process_get_replaying()? &smpi_comm_copy_buffer_callback
- : &smpi_comm_null_copy_buffer_callback, request, -1.0);
- XBT_DEBUG("recv simcall posted");
-
- if (async_small_thresh != 0 || (request->flags & RMA) != 0 )
- xbt_mutex_release(mut);
- } else {
- int receiver = request->dst;
-
- int rank = request->src;
- if (TRACE_smpi_view_internals()) {
- TRACE_smpi_send(rank, rank, receiver,request->size);
- }
- print_request("New send", request);
-
- //if we are giving back the control to the user without waiting for completion, we have to inject timings
- double sleeptime = 0.0;
- if(request->detached != 0 || ((request->flags & (ISEND|SSEND)) != 0)){// issend should be treated as isend
- //isend and send timings may be different
- sleeptime = ((request->flags & ISEND) != 0)? smpi_ois(request->size) : smpi_os(request->size);
- }
-
- if(sleeptime > 0.0){
- simcall_process_sleep(sleeptime);
- XBT_DEBUG("sending size of %zu : sleep %f ", request->size, smpi_os(request->size));
- }
-
- int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
-
- xbt_mutex_t mut=smpi_process_remote_mailboxes_mutex(receiver);
-
- if (async_small_thresh != 0 || (request->flags & RMA) != 0)
- xbt_mutex_acquire(mut);
-
- if (!(async_small_thresh != 0 || (request->flags & RMA) !=0)) {
- mailbox = smpi_process_remote_mailbox(receiver);
- }
- else if (((request->flags & RMA) != 0) || static_cast<int>(request->size) < async_small_thresh) { // eager mode
- mailbox = smpi_process_remote_mailbox(receiver);
- XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox);
- smx_synchro_t action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, static_cast<void*>(request));
- if (action == nullptr) {
- if ((request->flags & SSEND) == 0){
- mailbox = smpi_process_remote_mailbox_small(receiver);
- XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %p", mailbox);
- }
- else {
- mailbox = smpi_process_remote_mailbox_small(receiver);
- XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %p?", mailbox);
- action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, static_cast<void*>(request));
- if (action == nullptr) {
- XBT_DEBUG("No, we are first, send to large mailbox");
- mailbox = smpi_process_remote_mailbox(receiver);
- }
- }
- }
- else {
- XBT_DEBUG("Yes there was something for us in the large mailbox");
- }
- }
- else {
- mailbox = smpi_process_remote_mailbox(receiver);
- XBT_DEBUG("Send request %p is in the large mailbox %p (buf: %p)",mailbox, request,request->buf);
- }
-
- void* buf = request->buf;
- if ( ((request->flags & SSEND) == 0) && (((request->flags & RMA) != 0) ||
- (static_cast<int>(request->size) < xbt_cfg_get_int("smpi/send-is-detached-thresh")))) {
- void *oldbuf = NULL;
- request->detached = 1;
- XBT_DEBUG("Send request %p is detached", request);
- request->refcount++;
- if(request->old_type->sizeof_substruct == 0){
- oldbuf = request->buf;
- if (!smpi_process_get_replaying() && oldbuf != NULL && request->size!=0){
- if((smpi_privatize_global_variables != 0)
- && (static_cast<char*>(request->buf) >= smpi_start_data_exe)
- && (static_cast<char*>(request->buf) < smpi_start_data_exe + smpi_size_data_exe )){
- XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
- smpi_switch_data_segment(request->src);
- }
- buf = xbt_malloc(request->size);
- memcpy(buf,oldbuf,request->size);
- XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
- }
- }
- }
-
- // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
- request->real_size=request->size;
- request->action = simcall_comm_isend(SIMIX_process_from_PID(request->src+1), mailbox, request->size, -1.0,
- buf, request->real_size, &match_send,
- &xbt_free_f, // how to free the userdata if a detached send fails
- !smpi_process_get_replaying() ? &smpi_comm_copy_buffer_callback
- : &smpi_comm_null_copy_buffer_callback, request,
- // detach if msg size < eager/rdv switch limit
- request->detached);
- XBT_DEBUG("send simcall posted");
-
- /* FIXME: detached sends are not traceable (request->action == nullptr) */
- if (request->action != nullptr)
- simcall_set_category(request->action, TRACE_internal_smpi_get_category());
-
- if (async_small_thresh != 0 || ((request->flags & RMA)!=0))
- xbt_mutex_release(mut);
- }
-}
-
-void smpi_mpi_startall(int count, MPI_Request * requests)
-{
- if(requests== nullptr)
- return;
-
- for(int i = 0; i < count; i++) {
- smpi_mpi_start(requests[i]);
- }
-}
-
-void smpi_mpi_request_free(MPI_Request * request)
-{
- if((*request) != MPI_REQUEST_NULL){
- (*request)->refcount--;
- if((*request)->refcount<0) xbt_die("wrong refcount");
-
- if((*request)->refcount==0){
- smpi_datatype_unuse((*request)->old_type);
- smpi_comm_unuse((*request)->comm);
- print_request("Destroying", (*request));
- xbt_free(*request);
- *request = MPI_REQUEST_NULL;
- }else{
- print_request("Decrementing", (*request));
- }
- }else{
- xbt_die("freeing an already free request");
- }
-}
-
-MPI_Request smpi_rma_send_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
- MPI_Op op)
-{
- MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
- if(op==MPI_OP_NULL){
- request = build_request(buf==MPI_BOTTOM ? NULL : buf , count, datatype, src, dst, tag,
- comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED);
- }else{
- request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, src, dst, tag,
- comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED | ACCUMULATE);
- request->op = op;
- }
- return request;
-}
-
-MPI_Request smpi_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 = NULL; /* MC needs the comm to be set to NULL during the call */
- if(op==MPI_OP_NULL){
- request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, src, dst, tag,
- comm, RMA | NON_PERSISTENT | RECV | PREPARED);
- }else{
- request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, src, dst, tag,
- comm, RMA | NON_PERSISTENT | RECV | PREPARED | ACCUMULATE);
- request->op = op;
- }
- return request;
-}
-
-MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
-{
- MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
- request = build_request(buf==MPI_BOTTOM ? NULL : buf , count, datatype, smpi_process_index(),
- smpi_group_index(smpi_comm_group(comm), dst), tag,comm, PERSISTENT | ISEND | SEND | PREPARED);
- return request;
-}
-
-MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
-{
- MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
- request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, smpi_process_index(),
- smpi_group_index(smpi_comm_group(comm), dst), tag, comm, NON_PERSISTENT | ISEND | SEND);
- smpi_mpi_start(request);
- return request;
-}
-
-MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
-{
- MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
- request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, smpi_process_index(),
- smpi_group_index(smpi_comm_group(comm), dst), tag,comm, NON_PERSISTENT | ISEND | SSEND | SEND);
- smpi_mpi_start(request);
- return request;
-}
-
-MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
-{
- MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
- request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
- smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag,
- comm, PERSISTENT | RECV | PREPARED);
- return request;
-}
-
-MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
-{
- MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
- request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
- smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag, comm,
- NON_PERSISTENT | RECV);
- smpi_mpi_start(request);
- return request;
-}
-
-void smpi_mpi_recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
-{
- MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
- request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
- smpi_mpi_wait(&request, status);
- request = NULL;
-}
-
-void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
-{
- MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
- request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, smpi_process_index(),
- smpi_group_index(smpi_comm_group(comm), dst), tag, comm, NON_PERSISTENT | SEND);
-
- smpi_mpi_start(request);
- smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
- request = NULL;
-}
-
-void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
-{
- MPI_Request request = NULL; /* MC needs the comm to be set to NULL during the call */
- request = build_request(buf==MPI_BOTTOM ? NULL : buf, count, datatype, smpi_process_index(),
- smpi_group_index(smpi_comm_group(comm), dst), tag, comm, NON_PERSISTENT | SSEND | SEND);
-
- smpi_mpi_start(request);
- smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
- request = NULL;
-}
-
-void smpi_mpi_sendrecv(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)
-{
- MPI_Request requests[2];
- MPI_Status stats[2];
- int myid=smpi_process_index();
- if ((smpi_group_index(smpi_comm_group(comm), dst) == myid) && (smpi_group_index(smpi_comm_group(comm), src) == myid)){
- smpi_datatype_copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype);
- return;
- }
- requests[0] = smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
- requests[1] = smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
- smpi_mpi_startall(2, requests);
- smpi_mpi_waitall(2, requests, stats);
- smpi_mpi_request_free(&requests[0]);
- smpi_mpi_request_free(&requests[1]);
- if(status != MPI_STATUS_IGNORE) {
- // Copy receive status
- *status = stats[1];
- }
-}
-
-int smpi_mpi_get_count(MPI_Status * status, MPI_Datatype datatype)
-{
- return status->count / smpi_datatype_size(datatype);
-}
-
-static void finish_wait(MPI_Request * request, MPI_Status * status)
-{
- MPI_Request req = *request;
- smpi_empty_status(status);
-
- if(!((req->detached != 0) && ((req->flags & SEND) != 0)) && ((req->flags & PREPARED) == 0)){
- if(status != MPI_STATUS_IGNORE) {
- int src = req->src == MPI_ANY_SOURCE ? req->real_src : req->src;
- status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(req->comm), 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;
- // this handles the case were size in receive differs from size in send
- status->count = req->real_size;
- }
-
- print_request("Finishing", req);
- MPI_Datatype datatype = req->old_type;
-
- if(((req->flags & ACCUMULATE) != 0) || (datatype->sizeof_substruct != 0)){
- if (!smpi_process_get_replaying()){
- if( smpi_privatize_global_variables != 0 && (static_cast<char*>(req->old_buf) >= smpi_start_data_exe)
- && ((char*)req->old_buf < smpi_start_data_exe + smpi_size_data_exe )){
- XBT_VERB("Privatization : We are unserializing to a zone in global memory - Switch data segment ");
- smpi_switch_data_segment(smpi_process_index());
- }
- }
-
- if(datatype->sizeof_substruct != 0){
- // This part handles the problem of non-contignous memory the unserialization at the reception
- s_smpi_subtype_t *subtype = static_cast<s_smpi_subtype_t*>(datatype->substruct);
- if(req->flags & RECV)
- subtype->unserialize(req->buf, req->old_buf, req->real_size/smpi_datatype_size(datatype) ,
- datatype->substruct, req->op);
- xbt_free(req->buf);
- }else if(req->flags & RECV){//apply op on contiguous buffer for accumulate
- int n =req->real_size/smpi_datatype_size(datatype);
- smpi_op_apply(req->op, req->buf, req->old_buf, &n, &datatype);
- xbt_free(req->buf);
- }
- }
- }
-
- if (TRACE_smpi_view_internals() && ((req->flags & RECV) != 0)){
- int rank = smpi_process_index();
- int src_traced = (req->src == MPI_ANY_SOURCE ? req->real_src : req->src);
- TRACE_smpi_recv(rank, src_traced, rank);
- }
-
- if(req->detached_sender!=NULL){
- smpi_mpi_request_free(&(req->detached_sender));
- }
- if(req->flags & PERSISTENT)
- req->action = NULL;
- req->flags |= FINISHED;
-
- smpi_mpi_request_free(request);
-}
-
-int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
- //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
-
- // to avoid deadlocks if used as a break condition, such as
- // while (MPI_Test(request, flag, status) && flag) {
- // }
- // because the time will not normally advance when only calls to MPI_Test are made -> deadlock
- // multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
- static int nsleeps = 1;
- if(smpi_test_sleep > 0)
- simcall_process_sleep(nsleeps*smpi_test_sleep);
-
- smpi_empty_status(status);
- int flag = 1;
- if (((*request)->flags & PREPARED) == 0) {
- if ((*request)->action != NULL)
- flag = simcall_comm_test((*request)->action);
- if (flag) {
- finish_wait(request, status);
- nsleeps=1;//reset the number of sleeps we will do next time
- if (*request != MPI_REQUEST_NULL && ((*request)->flags & PERSISTENT)==0)
- *request = MPI_REQUEST_NULL;
- }else{
- nsleeps++;
- }
- }
- return flag;
-}
-
-int smpi_mpi_testany(int count, MPI_Request requests[], int *index, MPI_Status * status)
-{
- xbt_dynar_t comms;
- int i;
- int* map;
- int flag = 0;
- int size = 0;
-
- *index = MPI_UNDEFINED;
- comms = xbt_dynar_new(sizeof(smx_synchro_t), NULL);
- map = xbt_new(int, count);
- for(i = 0; i < count; i++) {
- if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action && !(requests[i]->flags & PREPARED)) {
- xbt_dynar_push(comms, &requests[i]->action);
- map[size] = i;
- size++;
- }
- }
- if(size > 0) {
- //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);
-
- i = simcall_comm_testany(comms);
- // not MPI_UNDEFINED, as this is a simix return code
- if(i != -1) {
- *index = map[i];
- finish_wait(&requests[*index], status);
- if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags & NON_PERSISTENT))
- requests[*index] = MPI_REQUEST_NULL;
- flag = 1;
- nsleeps=1;
- }else{
- nsleeps++;
- }
- }else{
- //all requests are null or inactive, return true
- flag=1;
- smpi_empty_status(status);
- }
- xbt_free(map);
- xbt_dynar_free(&comms);
-
- return flag;
-}
-
-int smpi_mpi_testall(int count, MPI_Request requests[], MPI_Status status[])
-{
- MPI_Status stat;
- MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
- int flag=1;
- int i;
- for(i=0; i<count; i++){
- if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED)) {
- if (smpi_mpi_test(&requests[i], pstat)!=1){
- flag=0;
- }else{
- requests[i]=MPI_REQUEST_NULL;
- }
- }else{
- smpi_empty_status(pstat);
- }
- if(status != MPI_STATUSES_IGNORE) {
- status[i] = *pstat;
- }
- }
- return flag;
-}
-
-void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
- int flag=0;
- //FIXME find another way to avoid busy waiting ?
- // the issue here is that we have to wait on a nonexistent comm
- while(flag==0){
- smpi_mpi_iprobe(source, tag, comm, &flag, status);
- XBT_DEBUG("Busy Waiting on probing : %d", flag);
- }
-}
-
-void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
-
- MPI_Request request = build_request(NULL, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
- smpi_group_index(smpi_comm_group(comm), source), smpi_comm_rank(comm), tag, comm, PERSISTENT | RECV);
-
- // to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
- // (especially when used as a break condition, such as while(MPI_Iprobe(...)) ... )
- // multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
- static int nsleeps = 1;
- if(smpi_iprobe_sleep > 0)
- simcall_process_sleep(nsleeps*smpi_iprobe_sleep);
- // behave like a receive, but don't do it
- smx_mailbox_t mailbox;
-
- print_request("New iprobe", request);
- // We have to test both mailboxes as we don't know if we will receive one one or another
- if (xbt_cfg_get_int("smpi/async-small-thresh") > 0){
- mailbox = smpi_process_mailbox_small();
- XBT_DEBUG("Trying to probe the perm recv mailbox");
- request->action = simcall_comm_iprobe(mailbox, 0, request->src, request->tag, &match_recv, static_cast<void*>(request));
- }
-
- if (request->action == nullptr){
- mailbox = smpi_process_mailbox();
- XBT_DEBUG("trying to probe the other mailbox");
- request->action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, static_cast<void*>(request));
- }
-
- if (request->action != nullptr){
- simgrid::simix::Comm *sync_comm = static_cast<simgrid::simix::Comm*>(request->action);
- MPI_Request req = static_cast<MPI_Request>(sync_comm->src_data);
- *flag = 1;
- if(status != MPI_STATUS_IGNORE && (req->flags & PREPARED) == 0) {
- status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src);
- status->MPI_TAG = req->tag;
- status->MPI_ERROR = MPI_SUCCESS;
- status->count = req->real_size;
- }
- nsleeps = 1;//reset the number of sleeps we will do next time
- }
- else {
- *flag = 0;
- nsleeps++;
- }
- smpi_mpi_request_free(&request);
-
- return;
-}
-
-void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
-{
- print_request("Waiting", *request);
- if ((*request)->flags & PREPARED) {
- smpi_empty_status(status);
- return;
- }
-
- if ((*request)->action != NULL)
- // this is not a detached send
- simcall_comm_wait((*request)->action, -1.0);
-
- finish_wait(request, status);
- if (*request != MPI_REQUEST_NULL && (((*request)->flags & NON_PERSISTENT)!=0))
- *request = MPI_REQUEST_NULL;
-}
-
-int smpi_mpi_waitany(int count, MPI_Request requests[], MPI_Status * status)
-{
- xbt_dynar_t comms;
- int i;
- int size = 0;
- int index = MPI_UNDEFINED;
- int *map;
-
- if(count > 0) {
- // Wait for a request to complete
- comms = xbt_dynar_new(sizeof(smx_synchro_t), NULL);
- map = xbt_new(int, count);
- XBT_DEBUG("Wait for one of %d", count);
- for(i = 0; i < count; i++) {
- if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags & PREPARED) && !(requests[i]->flags & FINISHED)) {
- if (requests[i]->action != NULL) {
- XBT_DEBUG("Waiting any %p ", requests[i]);
- xbt_dynar_push(comms, &requests[i]->action);
- map[size] = i;
- size++;
- }else{
- //This is a finished detached request, let's return this one
- size=0;//so we free the dynar but don't do the waitany call
- index=i;
- finish_wait(&requests[i], status);//cleanup if refcount = 0
- if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
- requests[i]=MPI_REQUEST_NULL;//set to null
- break;
- }
- }
- }
- if(size > 0) {
- i = simcall_comm_waitany(comms);
-
- // not MPI_UNDEFINED, as this is a simix return code
- if (i != -1) {
- index = map[i];
- finish_wait(&requests[index], status);
- if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags & NON_PERSISTENT))
- requests[index] = MPI_REQUEST_NULL;
- }
- }
- xbt_free(map);
- xbt_dynar_free(&comms);
- }
-
- if (index==MPI_UNDEFINED)
- smpi_empty_status(status);
-
- return index;
-}
-
-int smpi_mpi_waitall(int count, MPI_Request requests[], MPI_Status status[])
-{
- int index, c;
- MPI_Status stat;
- MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
- int retvalue = MPI_SUCCESS;
- //tag invalid requests in the set
- if (status != MPI_STATUSES_IGNORE) {
- for (c = 0; c < count; c++) {
- if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL || (requests[c]->flags & PREPARED)) {
- smpi_empty_status(&status[c]);
- } else if (requests[c]->src == MPI_PROC_NULL) {
- smpi_empty_status(&status[c]);
- status[c].MPI_SOURCE = MPI_PROC_NULL;
- }
- }
- }
- for(c = 0; c < count; c++) {
-
- if (MC_is_active() || MC_record_replay_is_active()) {
- smpi_mpi_wait(&requests[c], pstat);
- index = c;
- } else {
- index = smpi_mpi_waitany(count, requests, pstat);
- if (index == MPI_UNDEFINED)
- break;
- if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
- requests[index]=MPI_REQUEST_NULL;
- }
- if (status != MPI_STATUSES_IGNORE) {
- status[index] = *pstat;
- if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
- retvalue = MPI_ERR_IN_STATUS;
- }
- }
-
- return retvalue;
-}
-
-int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
-{
- int i;
- int count = 0;
- int index;
- MPI_Status stat;
- MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
-
- for(i = 0; i < incount; i++)
- {
- index=smpi_mpi_waitany(incount, requests, pstat);
- if(index!=MPI_UNDEFINED){
- indices[count] = index;
- count++;
- if(status != MPI_STATUSES_IGNORE) {
- status[index] = *pstat;
- }
- if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags & NON_PERSISTENT))
- requests[index]=MPI_REQUEST_NULL;
- }else{
- return MPI_UNDEFINED;
- }
- }
- return count;
-}
-
-int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
-{
- int i;
- int count = 0;
- int count_dead = 0;
- MPI_Status stat;
- MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
-
- for(i = 0; i < incount; i++) {
- if((requests[i] != MPI_REQUEST_NULL)) {
- if(smpi_mpi_test(&requests[i], pstat)) {
- indices[i] = 1;
- count++;
- if(status != MPI_STATUSES_IGNORE) {
- status[i] = *pstat;
- }
- if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->flags & NON_PERSISTENT)
- requests[i]=MPI_REQUEST_NULL;
- }
- }else{
- count_dead++;
- }
- }
- if(count_dead==incount)
- return MPI_UNDEFINED;
- else return count;
-}