* All rights reserved. */
/* 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. */
+ * under the terms of the license (GNU LGPL) which comes with this package. */
#include "private.h"
#include "xbt/time.h"
+#include "mc/mc.h"
+#include "xbt/replay.h"
+#include <errno.h>
+#include "surf/surf.h"
-XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_base, smpi,
- "Logging specific to SMPI (base)");
-XBT_LOG_EXTERNAL_CATEGORY(smpi_base);
-XBT_LOG_EXTERNAL_CATEGORY(smpi_bench);
-XBT_LOG_EXTERNAL_CATEGORY(smpi_kernel);
-XBT_LOG_EXTERNAL_CATEGORY(smpi_mpi);
-XBT_LOG_EXTERNAL_CATEGORY(smpi_mpi_dt);
-XBT_LOG_EXTERNAL_CATEGORY(smpi_coll);
-XBT_LOG_EXTERNAL_CATEGORY(smpi_receiver);
-XBT_LOG_EXTERNAL_CATEGORY(smpi_sender);
-XBT_LOG_EXTERNAL_CATEGORY(smpi_util);
+
+XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_base, smpi, "Logging specific to SMPI (base)");
+
+
+static int match_recv(void* a, void* b, smx_action_t ignored) {
+ MPI_Request ref = (MPI_Request)a;
+ MPI_Request req = (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_assert(ref, "Cannot match recv against null reference");
+ xbt_assert(req, "Cannot match recv against null request");
+ return (ref->src == MPI_ANY_SOURCE || req->src == ref->src)
+ && (ref->tag == MPI_ANY_TAG || req->tag == ref->tag);
+}
+
+static int match_send(void* a, void* b,smx_action_t ignored) {
+ MPI_Request ref = (MPI_Request)a;
+ MPI_Request req = (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_assert(ref, "Cannot match send against null reference");
+ xbt_assert(req, "Cannot match send against null request");
+ return (req->src == MPI_ANY_SOURCE || req->src == ref->src)
+ && (req->tag == MPI_ANY_TAG || req->tag == ref->tag);
+}
static MPI_Request build_request(void *buf, int count,
MPI_Datatype datatype, int src, int dst,
{
MPI_Request request;
+ void *old_buf;
+
request = xbt_new(s_smpi_mpi_request_t, 1);
+
+ s_smpi_subtype_t *subtype = datatype->substruct;
+
+ if(datatype->has_subtype == 1){
+ // This part handles the problem of non-contignous memory
+ old_buf = buf;
+ buf = malloc(count*smpi_datatype_size(datatype));
+ if (flags & SEND) {
+ subtype->serialize(old_buf, buf, count, datatype->substruct);
+ }
+ }
+
request->buf = buf;
+ // This part handles the problem of non-contignous memory (for the
+ // unserialisation at the reception)
+ request->old_buf = old_buf;
+ request->old_type = datatype;
+
request->size = smpi_datatype_size(datatype) * count;
request->src = src;
request->dst = dst;
request->tag = tag;
request->comm = comm;
- request->rdv = NULL;
- request->pair = NULL;
- request->complete = 0;
- request->match = MPI_REQUEST_NULL;
+ request->action = NULL;
request->flags = flags;
#ifdef HAVE_TRACING
request->send = 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->count=0;
+ }
+}
+
+void smpi_action_trace_run(char *path)
+{
+ char *name;
+ xbt_dynar_t todo;
+ xbt_dict_cursor_t cursor;
+
+ action_fp=NULL;
+ if (path) {
+ action_fp = fopen(path, "r");
+ xbt_assert(action_fp != NULL, "Cannot open %s: %s", path,
+ strerror(errno));
+ }
+
+ if (!xbt_dict_is_empty(action_queues)) {
+ XBT_WARN
+ ("Not all actions got consumed. If the simulation ended successfully (without deadlock), you may want to add new processes to your deployment file.");
+
+
+ xbt_dict_foreach(action_queues, cursor, name, todo) {
+ XBT_WARN("Still %lu actions for %s", xbt_dynar_length(todo), name);
+ }
+ }
+
+ if (path)
+ fclose(action_fp);
+ xbt_dict_free(&action_queues);
+ action_queues = xbt_dict_new_homogeneous(NULL);
+}
+
+static void smpi_mpi_request_free_voidp(void* request)
+{
+ MPI_Request req = 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 =
- build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag,
- comm, PERSISTENT | SEND);
+ build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag,
+ comm, PERSISTENT | SEND);
return request;
}
int src, int tag, MPI_Comm comm)
{
MPI_Request request =
- build_request(buf, count, datatype, src, smpi_comm_rank(comm), tag,
- comm, PERSISTENT | RECV);
+ build_request(buf, count, datatype, src, smpi_comm_rank(comm), tag,
+ comm, PERSISTENT | RECV);
return request;
}
void smpi_mpi_start(MPI_Request request)
{
- xbt_assert0(request->complete == 0,
- "Cannot start a non-finished communication");
- if ((request->flags & RECV) == RECV) {
- smpi_process_post_recv(request);
+ smx_rdv_t mailbox;
+ int detached = 0;
+
+ xbt_assert(!request->action,
+ "Cannot (re)start a non-finished communication");
+ if(request->flags & RECV) {
print_request("New recv", request);
- request->pair =
- SIMIX_network_irecv(request->rdv, request->buf, &request->size);
+ if (request->size < xbt_cfg_get_int(_surf_cfg_set, "smpi/async_small_thres"))
+ mailbox = smpi_process_mailbox_small();
+ else
+ mailbox = smpi_process_mailbox();
+
+ // FIXME: SIMIX does not yet support non-contiguous datatypes
+ request->action = simcall_comm_irecv(mailbox, request->buf, &request->size, &match_recv, request);
} else {
- smpi_process_post_send(request->comm, request); // FIXME
print_request("New send", request);
- request->pair =
- SIMIX_network_isend(request->rdv, request->size, -1.0,
- request->buf, request->size, NULL);
+ if (request->size < xbt_cfg_get_int(_surf_cfg_set, "smpi/async_small_thres")) { // eager mode
+ mailbox = smpi_process_remote_mailbox_small(
+ smpi_group_index(smpi_comm_group(request->comm), request->dst));
+ }else{
+ XBT_DEBUG("Send request %p is not in the permanent receive mailbox (buf: %p)",request,request->buf);
+ mailbox = smpi_process_remote_mailbox(
+ smpi_group_index(smpi_comm_group(request->comm), request->dst));
+ }
+ if (request->size < 64*1024 ) { //(FIXME: this limit should be configurable)
+ void *oldbuf = request->buf;
+ detached = 1;
+ request->buf = malloc(request->size);
+ if (oldbuf)
+ memcpy(request->buf,oldbuf,request->size);
+ XBT_DEBUG("Send request %p is detached; buf %p copied into %p",request,oldbuf,request->buf);
+ }
+
+ request->action =
+ simcall_comm_isend(mailbox, request->size, -1.0,
+ request->buf, request->size,
+ &match_send,
+ &smpi_mpi_request_free_voidp, // how to free the userdata if a detached send fails
+ request,
+ // detach if msg size < eager/rdv switch limit
+ detached);
+
+#ifdef HAVE_TRACING
+ /* FIXME: detached sends are not traceable (request->action == NULL) */
+ if (request->action)
+ simcall_set_category(request->action, TRACE_internal_smpi_get_category());
+#endif
+
}
+
}
void smpi_mpi_startall(int count, MPI_Request * requests)
{
int i;
- for (i = 0; i < count; i++) {
+ for(i = 0; i < count; i++) {
smpi_mpi_start(requests[i]);
}
}
int dst, int tag, MPI_Comm comm)
{
MPI_Request request =
- build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag,
- comm, NON_PERSISTENT | SEND);
+ build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag,
+ comm, NON_PERSISTENT | SEND);
return request;
}
int dst, int tag, MPI_Comm comm)
{
MPI_Request request =
- smpi_isend_init(buf, count, datatype, dst, tag, comm);
+ smpi_isend_init(buf, count, datatype, dst, tag, comm);
smpi_mpi_start(request);
return request;
int src, int tag, MPI_Comm comm)
{
MPI_Request request =
- build_request(buf, count, datatype, src, smpi_comm_rank(comm), tag,
- comm, NON_PERSISTENT | RECV);
-
+ build_request(buf, count, datatype, src, smpi_comm_rank(comm), tag,
+ comm, NON_PERSISTENT | RECV);
return request;
}
int src, int tag, MPI_Comm comm)
{
MPI_Request request =
- smpi_irecv_init(buf, count, datatype, src, tag, comm);
+ smpi_irecv_init(buf, count, datatype, src, tag, comm);
smpi_mpi_start(request);
return request;
int tag, MPI_Comm comm, MPI_Status * status)
{
MPI_Request request;
-
request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
smpi_mpi_wait(&request, status);
}
+
+
void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst,
int tag, MPI_Comm comm)
{
MPI_Status stats[2];
requests[0] =
- smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
+ smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
requests[1] =
- smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
+ smpi_irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
smpi_mpi_startall(2, requests);
smpi_mpi_waitall(2, requests, stats);
- if (status != MPI_STATUS_IGNORE) {
+ if(status != MPI_STATUS_IGNORE) {
// Copy receive status
memcpy(status, &stats[1], sizeof(MPI_Status));
}
static void finish_wait(MPI_Request * request, MPI_Status * status)
{
- if (status != MPI_STATUS_IGNORE) {
- status->MPI_SOURCE = (*request)->src;
- status->MPI_TAG = (*request)->tag;
+ MPI_Request req = *request;
+ // if we have a sender, we should use its data, and not the data from the receive
+ if((req->action)&&
+ (req->src==MPI_ANY_SOURCE || req->tag== MPI_ANY_TAG))
+ req = (MPI_Request)SIMIX_comm_get_src_data((*request)->action);
+
+ if(status != MPI_STATUS_IGNORE) {
+ status->MPI_SOURCE = req->src;
+ status->MPI_TAG = req->tag;
status->MPI_ERROR = MPI_SUCCESS;
- status->count = SIMIX_communication_get_dst_buf_size((*request)->pair);
+ // FIXME: really this should just contain the count of receive-type blocks,
+ // right?
+ status->count = req->size;
}
- SIMIX_communication_destroy((*request)->pair);
- print_request("finishing wait", *request);
- if ((*request)->complete == 1) {
- SIMIX_rdv_destroy((*request)->rdv);
- } else {
- (*request)->match->complete = 1;
- (*request)->match->match = MPI_REQUEST_NULL;
+ req = *request;
+
+ print_request("Finishing", req);
+ MPI_Datatype datatype = req->old_type;
+ if(datatype->has_subtype == 1){
+ // This part handles the problem of non-contignous memory
+ // the unserialization at the reception
+ s_smpi_subtype_t *subtype = datatype->substruct;
+ if(req->flags & RECV) {
+ subtype->unserialize(req->buf, req->old_buf, req->size/smpi_datatype_size(datatype) , datatype->substruct);
+ }
+ //FIXME: I am not sure that if the send is detached we have to free
+ //the sender buffer thus I do it only for the reciever
+ if(req->flags & RECV) free(req->buf);
}
- if (((*request)->flags & NON_PERSISTENT) == NON_PERSISTENT) {
+
+ if(req->flags & NON_PERSISTENT) {
smpi_mpi_request_free(request);
} else {
- (*request)->rdv = NULL;
- (*request)->pair = NULL;
+ req->action = NULL;
}
}
-int smpi_mpi_test(MPI_Request * request, MPI_Status * status)
-{
- int flag = (*request)->complete;
+int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
+ int flag;
- if (flag) {
- smpi_mpi_wait(request, status);
+ //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or smpi_mpi_testall before)
+ if ((*request)->action == NULL)
+ flag = 1;
+ else
+ flag = simcall_comm_test((*request)->action);
+ if(flag) {
+ finish_wait(request, status);
+ }else{
+ smpi_empty_status(status);
}
return flag;
}
int smpi_mpi_testany(int count, MPI_Request requests[], int *index,
MPI_Status * status)
{
- int i, flag;
+ xbt_dynar_t comms;
+ int i, flag, size;
+ int* map;
*index = MPI_UNDEFINED;
flag = 0;
- for (i = 0; i < count; i++) {
- if (requests[i] != MPI_REQUEST_NULL && requests[i]->complete) {
- smpi_mpi_wait(&requests[i], status);
- *index = i;
- flag = 1;
- break;
+ if(count > 0) {
+ comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
+ map = xbt_new(int, count);
+ size = 0;
+ for(i = 0; i < count; i++) {
+ if((requests[i]!=MPI_REQUEST_NULL) && requests[i]->action) {
+ xbt_dynar_push(comms, &requests[i]->action);
+ map[size] = i;
+ size++;
+ }
+ }
+ if(size > 0) {
+ 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);
+ flag = 1;
+ }
+ }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){
+ if (smpi_mpi_test(&requests[i], pstat)!=1){
+ flag=0;
+ }
+ }else{
+ smpi_empty_status(pstat);
+ }
+ if(status != MPI_STATUSES_IGNORE) {
+ memcpy(&status[i], pstat, sizeof(*pstat));
}
}
return flag;
}
+void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
+ int flag=0;
+ //FIXME find another wait 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);
+ if(!flag) {
+ simcall_process_sleep(0.0001);
+ }
+ }
+}
+
+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, smpi_comm_rank(comm), tag,
+ comm, NON_PERSISTENT | RECV);
+
+ // behave like a receive, but don't do it
+ smx_rdv_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(_surf_cfg_set, "smpi/async_small_thres")>0){
+ mailbox = smpi_process_mailbox_small();
+ XBT_DEBUG("trying to probe the perm recv mailbox");
+ request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
+ }
+ if (request->action==NULL){
+ mailbox = smpi_process_mailbox();
+ XBT_DEBUG("trying to probe the other mailbox");
+ request->action = simcall_comm_iprobe(mailbox, request->src, request->tag, &match_recv, (void*)request);
+ }
+
+ if(request->action){
+ MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action);
+ *flag=true;
+ if(status != MPI_STATUS_IGNORE) {
+ status->MPI_SOURCE = req->src;
+ status->MPI_TAG = req->tag;
+ status->MPI_ERROR = MPI_SUCCESS;
+ status->count = req->size;
+ }
+ }
+ else *flag=false;
+ smpi_mpi_request_free(&request);
+
+ return;
+}
+
void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
{
- print_request("wait", *request);
- SIMIX_network_wait((*request)->pair, -1.0);
- finish_wait(request, status);
+ print_request("Waiting", *request);
+ if ((*request)->action != NULL) { // this is not a detached send
+ simcall_comm_wait((*request)->action, -1.0);
+ finish_wait(request, status);
+ }
+ // FIXME for a detached send, finish_wait is not called:
}
int smpi_mpi_waitany(int count, MPI_Request requests[],
int *map;
index = MPI_UNDEFINED;
- if (count > 0) {
- // First check for already completed requests
- for (i = 0; i < count; i++) {
- if (requests[i] != MPI_REQUEST_NULL && requests[i]->complete) {
- index = i;
- smpi_mpi_wait(&requests[index], status);
- break;
+ if(count > 0) {
+ // Wait for a request to complete
+ comms = xbt_dynar_new(sizeof(smx_action_t), NULL);
+ map = xbt_new(int, count);
+ size = 0;
+ XBT_DEBUG("Wait for one of");
+ for(i = 0; i < count; i++) {
+ if((requests[i] != MPI_REQUEST_NULL) && (requests[i]->action != NULL)) {
+ print_request("Waiting any ", requests[i]);
+ xbt_dynar_push(comms, &requests[i]->action);
+ map[size] = i;
+ size++;
}
}
- if (index == MPI_UNDEFINED) {
- // Otherwise, wait for a request to complete
- comms = xbt_dynar_new(sizeof(smx_comm_t), NULL);
- map = xbt_new(int, count);
- size = 0;
- DEBUG0("Wait for one of");
- for (i = 0; i < count; i++) {
- if (requests[i] != MPI_REQUEST_NULL && requests[i]->complete == 0) {
- print_request(" ", requests[i]);
- xbt_dynar_push(comms, &requests[i]->pair);
- map[size] = i;
- size++;
- }
- }
- if (size > 0) {
- index = SIMIX_network_waitany(comms);
- index = map[index];
+ 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);
}
- xbt_free(map);
- xbt_dynar_free(&comms);
}
+ xbt_free(map);
+ xbt_dynar_free(&comms);
}
+
+ if (index==MPI_UNDEFINED)
+ smpi_empty_status(status);
+
return index;
}
void smpi_mpi_waitall(int count, MPI_Request requests[],
MPI_Status status[])
{
- int index;
+ int index, c;
MPI_Status stat;
-
- while (count > 0) {
- index = smpi_mpi_waitany(count, requests, &stat);
- if (index == MPI_UNDEFINED) {
- break;
+ MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
+ //tag invalid requests in the set
+ for(c = 0; c < count; c++) {
+ if(requests[c]==MPI_REQUEST_NULL || requests[c]->dst == MPI_PROC_NULL ){
+ if(status != MPI_STATUSES_IGNORE)
+ smpi_empty_status(&status[c]);
+ }else if(requests[c]->src == MPI_PROC_NULL ){
+ if(status != MPI_STATUSES_IGNORE) {
+ smpi_empty_status(&status[c]);
+ status[c].MPI_SOURCE=MPI_PROC_NULL;
+ }
}
- if (status != MPI_STATUS_IGNORE) {
- memcpy(&status[index], &stat, sizeof(stat));
+ }
+
+ for(c = 0; c < count; c++) {
+ if(MC_IS_ENABLED) {
+ smpi_mpi_wait(&requests[c], pstat);
+ index = c;
+ } else {
+ index = smpi_mpi_waitany(count, requests, pstat);
+ if(index == MPI_UNDEFINED) {
+ break;
+ }
+ if(status != MPI_STATUSES_IGNORE) {
+ memcpy(&status[index], pstat, sizeof(*pstat));
+
+ }
}
- // FIXME: check this -v
- // Move the last request to the found position
- requests[index] = requests[count - 1];
- requests[count - 1] = MPI_REQUEST_NULL;
- count--;
}
}
int smpi_mpi_waitsome(int incount, MPI_Request requests[], int *indices,
MPI_Status status[])
{
- int i, count;
+ int i, count, index;
+ MPI_Status stat;
+ MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
count = 0;
- for (i = 0; i < incount; i++) {
- if (requests[i] != MPI_REQUEST_NULL && requests[i]->complete) {
- smpi_mpi_wait(&requests[i],
- status !=
- MPI_STATUS_IGNORE ? &status[i] : MPI_STATUS_IGNORE);
- indices[count] = i;
+ 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) {
+ memcpy(&status[index], pstat, sizeof(*pstat));
+ }
+ }else{
+ return MPI_UNDEFINED;
}
}
return count;
}
+int smpi_mpi_testsome(int incount, MPI_Request requests[], int *indices,
+ MPI_Status status[])
+{
+ int i, count, count_dead;
+ MPI_Status stat;
+ MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
+
+ count = 0;
+ for(i = 0; i < incount; i++) {
+ if((requests[i] != MPI_REQUEST_NULL)) {
+ if(smpi_mpi_test(&requests[i], pstat)) {
+ indices[count] = i;
+ count++;
+ if(status != MPI_STATUSES_IGNORE) {
+ memcpy(&status[i], pstat, sizeof(*pstat));
+ }
+ }
+ }else{
+ count_dead++;
+ }
+ }
+ if(count_dead==incount)return MPI_UNDEFINED;
+ else return count;
+}
+
void smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root,
MPI_Comm comm)
{
int root, MPI_Comm comm)
{
int system_tag = 666;
- int rank, size, src, index, sendsize, recvsize;
+ int rank, size, src, index;
+ MPI_Aint lb = 0, recvext = 0;
MPI_Request *requests;
rank = smpi_comm_rank(comm);
size = smpi_comm_size(comm);
- if (rank != root) {
+ if(rank != root) {
// Send buffer to root
smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
} else {
- sendsize = smpi_datatype_size(sendtype);
- recvsize = smpi_datatype_size(recvtype);
+ // FIXME: check for errors
+ smpi_datatype_extent(recvtype, &lb, &recvext);
// Local copy from root
- memcpy(&((char *) recvbuf)[root * recvcount * recvsize], sendbuf,
- sendcount * sendsize * sizeof(char));
+ smpi_datatype_copy(sendbuf, sendcount, sendtype,
+ (char *)recvbuf + root * recvcount * recvext, recvcount, recvtype);
// Receive buffers from senders
requests = xbt_new(MPI_Request, size - 1);
index = 0;
- for (src = 0; src < size; src++) {
- if (src != root) {
- requests[index] = smpi_irecv_init(&((char *) recvbuf)
- [src * recvcount * recvsize],
- recvcount, recvtype, src,
- system_tag, comm);
+ for(src = 0; src < size; src++) {
+ if(src != root) {
+ requests[index] = smpi_irecv_init((char *)recvbuf + src * recvcount * recvext,
+ recvcount, recvtype,
+ src, system_tag, comm);
index++;
}
}
MPI_Datatype recvtype, int root, MPI_Comm comm)
{
int system_tag = 666;
- int rank, size, src, index, sendsize;
+ int rank, size, src, index;
+ MPI_Aint lb = 0, recvext = 0;
MPI_Request *requests;
rank = smpi_comm_rank(comm);
size = smpi_comm_size(comm);
- if (rank != root) {
+ if(rank != root) {
// Send buffer to root
smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
} else {
- sendsize = smpi_datatype_size(sendtype);
+ // FIXME: check for errors
+ smpi_datatype_extent(recvtype, &lb, &recvext);
// Local copy from root
- memcpy(&((char *) recvbuf)[displs[root]], sendbuf,
- sendcount * sendsize * sizeof(char));
+ smpi_datatype_copy(sendbuf, sendcount, sendtype,
+ (char *)recvbuf + displs[root] * recvext,
+ recvcounts[root], recvtype);
// Receive buffers from senders
requests = xbt_new(MPI_Request, size - 1);
index = 0;
- for (src = 0; src < size; src++) {
- if (src != root) {
+ for(src = 0; src < size; src++) {
+ if(src != root) {
requests[index] =
- smpi_irecv_init(&((char *) recvbuf)[displs[src]],
- recvcounts[src], recvtype, src, system_tag,
- comm);
+ smpi_irecv_init((char *)recvbuf + displs[src] * recvext,
+ recvcounts[src], recvtype, src, system_tag, comm);
index++;
}
}
MPI_Comm comm)
{
int system_tag = 666;
- int rank, size, other, index, sendsize, recvsize;
+ int rank, size, other, index;
+ MPI_Aint lb = 0, recvext = 0;
MPI_Request *requests;
rank = smpi_comm_rank(comm);
size = smpi_comm_size(comm);
- sendsize = smpi_datatype_size(sendtype);
- recvsize = smpi_datatype_size(recvtype);
+ // FIXME: check for errors
+ smpi_datatype_extent(recvtype, &lb, &recvext);
// Local copy from self
- memcpy(&((char *) recvbuf)[rank * recvcount * recvsize], sendbuf,
- sendcount * sendsize * sizeof(char));
+ smpi_datatype_copy(sendbuf, sendcount, sendtype,
+ (char *)recvbuf + rank * recvcount * recvext, recvcount,
+ recvtype);
// Send/Recv buffers to/from others;
requests = xbt_new(MPI_Request, 2 * (size - 1));
index = 0;
- for (other = 0; other < size; other++) {
- if (other != rank) {
+ for(other = 0; other < size; other++) {
+ if(other != rank) {
requests[index] =
- smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
- comm);
+ smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
+ comm);
index++;
- requests[index] = smpi_irecv_init(&((char *) recvbuf)
- [other * recvcount * recvsize],
+ requests[index] = smpi_irecv_init((char *)recvbuf + other * recvcount * recvext,
recvcount, recvtype, other,
system_tag, comm);
index++;
MPI_Datatype recvtype, MPI_Comm comm)
{
int system_tag = 666;
- int rank, size, other, index, sendsize, recvsize;
+ int rank, size, other, index;
+ MPI_Aint lb = 0, recvext = 0;
MPI_Request *requests;
rank = smpi_comm_rank(comm);
size = smpi_comm_size(comm);
- sendsize = smpi_datatype_size(sendtype);
- recvsize = smpi_datatype_size(recvtype);
+ // FIXME: check for errors
+ smpi_datatype_extent(recvtype, &lb, &recvext);
// Local copy from self
- memcpy(&((char *) recvbuf)[displs[rank]], sendbuf,
- sendcount * sendsize * sizeof(char));
+ smpi_datatype_copy(sendbuf, sendcount, sendtype,
+ (char *)recvbuf + displs[rank] * recvext,
+ recvcounts[rank], recvtype);
// Send buffers to others;
requests = xbt_new(MPI_Request, 2 * (size - 1));
index = 0;
- for (other = 0; other < size; other++) {
- if (other != rank) {
+ for(other = 0; other < size; other++) {
+ if(other != rank) {
requests[index] =
- smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
- comm);
+ smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
+ comm);
index++;
requests[index] =
- smpi_irecv_init(&((char *) recvbuf)[displs[other]],
- recvcounts[other], recvtype, other, system_tag,
- comm);
+ smpi_irecv_init((char *)recvbuf + displs[other] * recvext, recvcounts[other],
+ recvtype, other, system_tag, comm);
index++;
}
}
int root, MPI_Comm comm)
{
int system_tag = 666;
- int rank, size, dst, index, sendsize, recvsize;
+ int rank, size, dst, index;
+ MPI_Aint lb = 0, sendext = 0;
MPI_Request *requests;
rank = smpi_comm_rank(comm);
size = smpi_comm_size(comm);
- if (rank != root) {
+ if(rank != root) {
// Recv buffer from root
smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
MPI_STATUS_IGNORE);
} else {
- sendsize = smpi_datatype_size(sendtype);
- recvsize = smpi_datatype_size(recvtype);
+ // FIXME: check for errors
+ smpi_datatype_extent(sendtype, &lb, &sendext);
// Local copy from root
- memcpy(recvbuf, &((char *) sendbuf)[root * sendcount * sendsize],
- recvcount * recvsize * sizeof(char));
+ smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext,
+ sendcount, sendtype, recvbuf, recvcount, recvtype);
// Send buffers to receivers
requests = xbt_new(MPI_Request, size - 1);
index = 0;
- for (dst = 0; dst < size; dst++) {
- if (dst != root) {
- requests[index] = smpi_isend_init(&((char *) sendbuf)
- [dst * sendcount * sendsize],
+ for(dst = 0; dst < size; dst++) {
+ if(dst != root) {
+ requests[index] = smpi_isend_init((char *)sendbuf + dst * sendcount * sendext,
sendcount, sendtype, dst,
system_tag, comm);
index++;
MPI_Datatype recvtype, int root, MPI_Comm comm)
{
int system_tag = 666;
- int rank, size, dst, index, sendsize, recvsize;
+ int rank, size, dst, index;
+ MPI_Aint lb = 0, sendext = 0;
MPI_Request *requests;
rank = smpi_comm_rank(comm);
size = smpi_comm_size(comm);
- if (rank != root) {
+ if(rank != root) {
// Recv buffer from root
smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm,
MPI_STATUS_IGNORE);
} else {
- sendsize = smpi_datatype_size(sendtype);
- recvsize = smpi_datatype_size(recvtype);
+ // FIXME: check for errors
+ smpi_datatype_extent(sendtype, &lb, &sendext);
// Local copy from root
- memcpy(recvbuf, &((char *) sendbuf)[displs[root]],
- recvcount * recvsize * sizeof(char));
+ smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root],
+ sendtype, recvbuf, recvcount, recvtype);
// Send buffers to receivers
requests = xbt_new(MPI_Request, size - 1);
index = 0;
- for (dst = 0; dst < size; dst++) {
- if (dst != root) {
+ for(dst = 0; dst < size; dst++) {
+ if(dst != root) {
requests[index] =
- smpi_isend_init(&((char *) sendbuf)[displs[dst]],
- sendcounts[dst], sendtype, dst, system_tag,
- comm);
+ smpi_isend_init((char *)sendbuf + displs[dst] * sendext, sendcounts[dst],
+ sendtype, dst, system_tag, comm);
index++;
}
}
MPI_Comm comm)
{
int system_tag = 666;
- int rank, size, src, index, datasize;
+ int rank, size, src, index;
+ MPI_Aint lb = 0, dataext = 0;
MPI_Request *requests;
void **tmpbufs;
rank = smpi_comm_rank(comm);
size = smpi_comm_size(comm);
- if (rank != root) {
+ if(rank != root) {
// Send buffer to root
smpi_mpi_send(sendbuf, count, datatype, root, system_tag, comm);
} else {
- datasize = smpi_datatype_size(datatype);
+ // FIXME: check for errors
+ smpi_datatype_extent(datatype, &lb, &dataext);
// Local copy from root
- memcpy(recvbuf, sendbuf, count * datasize * sizeof(char));
+ if (sendbuf && recvbuf)
+ smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
// Receive buffers from senders
//TODO: make a MPI_barrier here ?
requests = xbt_new(MPI_Request, size - 1);
tmpbufs = xbt_new(void *, size - 1);
index = 0;
- for (src = 0; src < size; src++) {
- if (src != root) {
- tmpbufs[index] = xbt_malloc(count * datasize);
+ for(src = 0; src < size; src++) {
+ if(src != root) {
+ // FIXME: possibly overkill we we have contiguous/noncontiguous data
+ // mapping...
+ tmpbufs[index] = xbt_malloc(count * dataext);
requests[index] =
- smpi_irecv_init(tmpbufs[index], count, datatype, src,
- system_tag, comm);
+ smpi_irecv_init(tmpbufs[index], count, datatype, src,
+ system_tag, comm);
index++;
}
}
// Wait for completion of irecv's.
smpi_mpi_startall(size - 1, requests);
- for (src = 0; src < size - 1; src++) {
+ for(src = 0; src < size - 1; src++) {
index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
- if (index == MPI_UNDEFINED) {
+ XBT_DEBUG("finished waiting any request with index %d", index);
+ if(index == MPI_UNDEFINED) {
break;
}
- smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
+ if(op) /* op can be MPI_OP_NULL that does nothing */
+ smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
}
- for (index = 0; index < size - 1; index++) {
+ for(index = 0; index < size - 1; index++) {
xbt_free(tmpbufs[index]);
}
xbt_free(tmpbufs);
{
smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
-
-/*
-FIXME: buggy implementation
-
- int system_tag = 666;
- int rank, size, other, index, datasize;
- MPI_Request* requests;
- void** tmpbufs;
-
- rank = smpi_comm_rank(comm);
- size = smpi_comm_size(comm);
- datasize = smpi_datatype_size(datatype);
- // Local copy from self
- memcpy(recvbuf, sendbuf, count * datasize * sizeof(char));
- // Send/Recv buffers to/from others;
- //TODO: make a MPI_barrier here ?
- requests = xbt_new(MPI_Request, 2 * (size - 1));
- tmpbufs = xbt_new(void*, size - 1);
- index = 0;
- for(other = 0; other < size; other++) {
- if(other != rank) {
- tmpbufs[index / 2] = xbt_malloc(count * datasize);
- requests[index] = smpi_mpi_isend(sendbuf, count, datatype, other, system_tag, comm);
- requests[index + 1] = smpi_mpi_irecv(tmpbufs[index / 2], count, datatype, other, system_tag, comm);
- index += 2;
- }
- }
- // Wait for completion of all comms.
- for(other = 0; other < 2 * (size - 1); other++) {
- index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
- if(index == MPI_UNDEFINED) {
- break;
- }
- if((index & 1) == 1) {
- // Request is odd: it's a irecv
- smpi_op_apply(op, tmpbufs[index / 2], recvbuf, &count, &datatype);
- }
- }
- for(index = 0; index < size - 1; index++) {
- xbt_free(tmpbufs[index]);
- }
- xbt_free(tmpbufs);
- xbt_free(requests);
-*/
}
void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
{
int system_tag = 666;
- int rank, size, other, index, datasize;
- int total;
+ int rank, size, other, index;
+ MPI_Aint lb = 0, dataext = 0;
MPI_Request *requests;
void **tmpbufs;
rank = smpi_comm_rank(comm);
size = smpi_comm_size(comm);
- datasize = smpi_datatype_size(datatype);
+
+ // FIXME: check for errors
+ smpi_datatype_extent(datatype, &lb, &dataext);
+
// Local copy from self
- memcpy(recvbuf, sendbuf, count * datasize * sizeof(char));
+ smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype);
+
// Send/Recv buffers to/from others;
- total = rank + (size - (rank + 1));
- requests = xbt_new(MPI_Request, total);
+ requests = xbt_new(MPI_Request, size - 1);
tmpbufs = xbt_new(void *, rank);
index = 0;
- for (other = 0; other < rank; other++) {
- tmpbufs[index] = xbt_malloc(count * datasize);
+ for(other = 0; other < rank; other++) {
+ // FIXME: possibly overkill we we have contiguous/noncontiguous data
+ // mapping...
+ tmpbufs[index] = xbt_malloc(count * dataext);
requests[index] =
- smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
- comm);
+ smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
+ comm);
index++;
}
- for (other = rank + 1; other < size; other++) {
+ for(other = rank + 1; other < size; other++) {
requests[index] =
- smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
+ smpi_isend_init(sendbuf, count, datatype, other, system_tag, comm);
index++;
}
// Wait for completion of all comms.
smpi_mpi_startall(size - 1, requests);
- for (other = 0; other < total; other++) {
+ for(other = 0; other < size - 1; other++) {
index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
- if (index == MPI_UNDEFINED) {
+ if(index == MPI_UNDEFINED) {
break;
}
- if (index < rank) {
+ if(index < rank) {
// #Request is below rank: it's a irecv
smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
}
}
- for (index = 0; index < size - 1; index++) {
+ for(index = 0; index < rank; index++) {
xbt_free(tmpbufs[index]);
}
xbt_free(tmpbufs);
