MPI_Request ref = (MPI_Request)a;
MPI_Request req = (MPI_Request)b;
- xbt_assert0(ref, "Cannot match recv against null reference");
- xbt_assert0(req, "Cannot match recv against null request");
- return req->comm == ref->comm
- && (ref->src == MPI_ANY_SOURCE || req->src == ref->src)
+ 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);
}
MPI_Request ref = (MPI_Request)a;
MPI_Request req = (MPI_Request)b;
- xbt_assert0(ref, "Cannot match send against null reference");
- xbt_assert0(req, "Cannot match send against null request");
- return req->comm == ref->comm
- && (req->src == MPI_ANY_SOURCE || req->src == ref->src)
+ 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);
}
request = xbt_new(s_smpi_mpi_request_t, 1);
request->buf = buf;
+ // FIXME: this will have to be changed to support non-contiguous datatypes
request->size = smpi_datatype_size(datatype) * count;
request->src = src;
request->dst = dst;
return request;
}
+static void smpi_mpi_request_free_voidp(void* request)
+{
+ smpi_mpi_request_free(request);
+}
+
/* MPI Low level calls */
MPI_Request smpi_mpi_send_init(void *buf, int count, MPI_Datatype datatype,
int dst, int tag, MPI_Comm comm)
void smpi_mpi_start(MPI_Request request)
{
smx_rdv_t mailbox;
+ int detached = 0;
- xbt_assert0(!request->action,
+ xbt_assert(!request->action,
"Cannot (re)start a non-finished communication");
if(request->flags & RECV) {
print_request("New recv", request);
mailbox = smpi_process_mailbox();
+ // FIXME: SIMIX does not yet support non-contiguous datatypes
request->action = SIMIX_req_comm_irecv(mailbox, request->buf, &request->size, &match_recv, request);
} else {
print_request("New send", request);
- mailbox = smpi_process_remote_mailbox(request->dst);
- request->action = SIMIX_req_comm_isend(mailbox, request->size, -1.0,
- request->buf, request->size, &match_send, request, 0);
+ mailbox = smpi_process_remote_mailbox(
+ smpi_group_index(smpi_comm_group(request->comm), request->dst));
+ // FIXME: SIMIX does not yet support non-contiguous datatypes
+
+ if (request->size < 64*1024 ) { // eager mode => detached send (FIXME: this limit should be configurable)
+ void *oldbuf = request->buf;
+ detached = 1;
+ request->buf = malloc(request->size);
+ memcpy(request->buf,oldbuf,request->size);
+ XBT_DEBUG("Send request %p is detached; buf %p copied into %p",request,oldbuf,request->buf);
+ } else {
+ XBT_DEBUG("Send request %p is not detached (buf: %p)",request,request->buf);
+ }
+ request->action =
+ SIMIX_req_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
SIMIX_req_set_category (request->action, TRACE_internal_smpi_get_category());
#endif
void smpi_mpi_startall(int count, MPI_Request * requests)
{
- int i;
+ int i;
for(i = 0; i < count; i++) {
smpi_mpi_start(requests[i]);
status->MPI_SOURCE = req->src;
status->MPI_TAG = req->tag;
status->MPI_ERROR = MPI_SUCCESS;
+ // FIXME: really this should just contain the count of receive-type blocks,
+ // right?
status->count = req->size;
}
print_request("Finishing", req);
}
}
-int smpi_mpi_test(MPI_Request * request, MPI_Status * status)
-{
- int flag = SIMIX_req_comm_test((*request)->action);
+int smpi_mpi_test(MPI_Request * request, MPI_Status * status) {
+int flag;
+ if ((*request)->action == NULL)
+ flag = 1;
+ else
+ flag = SIMIX_req_comm_test((*request)->action);
if(flag) {
- smpi_mpi_wait(request, status);
- }
- return flag;
+ smpi_mpi_wait(request, status);
+ }
+ return flag;
}
int smpi_mpi_testany(int count, MPI_Request requests[], int *index,
}
}
if(size > 0) {
- *index = SIMIX_req_comm_testany(comms);
- *index = map[*index];
- if(*index != MPI_UNDEFINED) {
+ i = SIMIX_req_comm_testany(comms);
+ // FIXME: MPI_UNDEFINED or does SIMIX have a return code?
+ if(i != MPI_UNDEFINED) {
+ *index = map[i];
smpi_mpi_wait(&requests[*index], status);
flag = 1;
}
void smpi_mpi_wait(MPI_Request * request, MPI_Status * status)
{
print_request("Waiting", *request);
- SIMIX_req_comm_wait((*request)->action, -1.0);
- finish_wait(request, status);
+ if ((*request)->action != NULL) { // this is not a detached send
+ SIMIX_req_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[],
size = 0;
XBT_DEBUG("Wait for one of");
for(i = 0; i < count; i++) {
- if(requests[i] != MPI_REQUEST_NULL) {
+ if((requests[i] != MPI_REQUEST_NULL) && (requests[i]->action != NULL)) {
print_request(" ", requests[i]);
xbt_dynar_push(comms, &requests[i]->action);
map[size] = i;
}
}
if(size > 0) {
- index = SIMIX_req_comm_waitany(comms);
- index = map[index];
- finish_wait(&requests[index], status);
+ i = SIMIX_req_comm_waitany(comms);
+ // FIXME: MPI_UNDEFINED or does SIMIX have a return code?
+ if (i != MPI_UNDEFINED) {
+ index = map[i];
+ finish_wait(&requests[index], status);
+ }
}
xbt_free(map);
xbt_dynar_free(&comms);
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);
// 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);
+ 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);
// 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) {
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;
smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag,
comm);
index++;
- requests[index] = smpi_irecv_init(&((char *) recvbuf)
- [other * recvcount * recvsize],
- recvcount, recvtype, other,
+ 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;
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);
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],
+ 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);
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) {
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;
// 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));
+ 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);
index = 0;
for(src = 0; src < size; src++) {
if(src != root) {
- tmpbufs[index] = xbt_malloc(count * datasize);
+ // 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_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);
+ // 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);
}
// 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) {
break;
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);