4 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_base, smpi,
5 "Logging specific to SMPI (base)");
6 XBT_LOG_EXTERNAL_CATEGORY(smpi_base);
7 XBT_LOG_EXTERNAL_CATEGORY(smpi_bench);
8 XBT_LOG_EXTERNAL_CATEGORY(smpi_kernel);
9 XBT_LOG_EXTERNAL_CATEGORY(smpi_mpi);
10 XBT_LOG_EXTERNAL_CATEGORY(smpi_mpi_dt);
11 XBT_LOG_EXTERNAL_CATEGORY(smpi_coll);
12 XBT_LOG_EXTERNAL_CATEGORY(smpi_receiver);
13 XBT_LOG_EXTERNAL_CATEGORY(smpi_sender);
14 XBT_LOG_EXTERNAL_CATEGORY(smpi_util);
16 void smpi_process_init(int* argc, char*** argv) {
18 smpi_process_data_t data;
21 proc = SIMIX_process_self();
22 index = atoi((*argv)[1]);
23 data = smpi_process_remote_data(index);
24 SIMIX_process_set_data(proc, data);
27 memmove(&(*argv)[1], &(*argv)[2], sizeof(char *) * (*argc - 2));
28 (*argv)[(*argc) - 1] = NULL;
31 DEBUG2("<%d> New process in the game: %p", index, proc);
34 void smpi_process_destroy(void) {
35 int index = smpi_process_index();
37 DEBUG1("<%d> Process left the game", index);
40 /* MPI Low level calls */
41 MPI_Request smpi_mpi_isend(void* buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) {
44 request = xbt_new(s_smpi_mpi_request_t, 1);
46 request->src = smpi_comm_rank(comm);
49 request->size = smpi_datatype_size(datatype) * count;
50 request->complete = 0;
51 smpi_process_post_send(comm, request);
52 request->pair = SIMIX_network_isend(request->rdv, request->size, -1.0, buf, request->size, request);
56 MPI_Request smpi_mpi_irecv(void* buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm) {
59 request = xbt_new(s_smpi_mpi_request_t, 1);
62 request->dst = smpi_comm_rank(comm);
64 request->size = smpi_datatype_size(datatype) * count;
65 request->complete = 0;
66 smpi_process_post_recv(request);
67 request->pair = SIMIX_network_irecv(request->rdv, buf, &request->size);
71 void smpi_mpi_recv(void* buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status* status) {
74 request = smpi_mpi_irecv(buf, count, datatype, src, tag, comm);
75 smpi_mpi_wait(&request, status);
78 void smpi_mpi_send(void* buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm) {
81 request = smpi_mpi_isend(buf, count, datatype, src, tag, comm);
82 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
85 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) {
86 MPI_Request requests[2];
89 requests[0] = smpi_mpi_isend(sendbuf, sendcount, sendtype, dst, sendtag, comm);
90 requests[1] = smpi_mpi_irecv(recvbuf, recvcount, recvtype, src, recvtag, comm);
91 smpi_mpi_waitall(2, requests, stats);
92 if(status != MPI_STATUS_IGNORE) {
93 // Copy receive status
94 memcpy(status, &stats[1], sizeof(MPI_Status));
98 static void finish_wait(MPI_Request* request, MPI_Status* status) {
99 MPI_Request data = (MPI_Request)SIMIX_communication_get_data((*request)->pair);
101 xbt_assert0(data != MPI_REQUEST_NULL, "Erroneous situation");
102 if(status != MPI_STATUS_IGNORE) {
103 status->MPI_SOURCE = (*request)->src;
104 status->MPI_TAG = (*request)->tag;
105 status->MPI_ERROR = MPI_SUCCESS;
107 DEBUG3("finishing wait for %p [data = %p, complete = %d]", *request, data, data->complete);
108 // data == *request if sender is first to finish its wait
109 // data != *request if receiver is first to finish its wait
110 if(data->complete == 0) {
111 // first arrives here
113 if(data != *request) {
114 // receveiver cleans its part
118 // second arrives here
119 if(data != *request) {
120 // receiver cleans everything
125 *request = MPI_REQUEST_NULL;
128 int smpi_mpi_test(MPI_Request* request, MPI_Status* status) {
129 MPI_Request data = (MPI_Request)SIMIX_communication_get_data((*request)->pair);
130 int flag = data && data->complete == 1;
133 finish_wait(request, status);
138 int smpi_mpi_testany(int count, MPI_Request requests[], int* index, MPI_Status* status) {
142 *index = MPI_UNDEFINED;
144 for(i = 0; i < count; i++) {
145 if(requests[i] != MPI_REQUEST_NULL) {
146 data = (MPI_Request)SIMIX_communication_get_data(requests[i]->pair);
147 if(data != MPI_REQUEST_NULL && data->complete == 1) {
148 finish_wait(&requests[i], status);
158 void smpi_mpi_wait(MPI_Request* request, MPI_Status* status) {
159 MPI_Request data = (MPI_Request)SIMIX_communication_get_data((*request)->pair);
161 DEBUG6("wait for request %p (%p: %p) [src = %d, dst = %d, tag = %d]",
162 *request, (*request)->pair, data, (*request)->src, (*request)->dst, (*request)->tag);
163 // data is null if receiver waits before sender enters the rdv
164 if(data == MPI_REQUEST_NULL || data->complete == 0) {
165 SIMIX_network_wait((*request)->pair, -1.0);
167 SIMIX_communication_destroy((*request)->pair);
169 finish_wait(request, status);
172 int smpi_mpi_waitany(int count, MPI_Request requests[], MPI_Status* status) {
178 index = MPI_UNDEFINED;
180 // First check for already completed requests
181 for(i = 0; i < count; i++) {
182 if(requests[i] != MPI_REQUEST_NULL) {
183 data = (MPI_Request)SIMIX_communication_get_data(requests[i]->pair);
184 if(data != MPI_REQUEST_NULL && data->complete == 1) {
186 SIMIX_communication_destroy(requests[index]->pair); // always succeeds (but cleans the simix layer)
191 if(index == MPI_UNDEFINED) {
192 // Otherwise, wait for a request to complete
193 comms = xbt_dynar_new(sizeof(smx_comm_t), NULL);
194 map = xbt_new(int, count);
196 DEBUG0("Wait for one of");
197 for(i = 0; i < count; i++) {
198 if(requests[i] != MPI_REQUEST_NULL && requests[i]->complete == 0) {
199 DEBUG4(" request %p [src = %d, dst = %d, tag = %d]",
200 requests[i], requests[i]->src, requests[i]->dst, requests[i]->tag);
201 xbt_dynar_push(comms, &requests[i]->pair);
207 index = SIMIX_network_waitany(comms);
211 xbt_dynar_free_container(&comms);
213 if(index != MPI_UNDEFINED) {
214 finish_wait(&requests[index], status);
220 void smpi_mpi_waitall(int count, MPI_Request requests[], MPI_Status status[]) {
225 index = smpi_mpi_waitany(count, requests, &stat);
226 if(index == MPI_UNDEFINED) {
229 if(status != MPI_STATUS_IGNORE) {
230 memcpy(&status[index], &stat, sizeof(stat));
232 // Move the last request to the found position
233 requests[index] = requests[count - 1];
234 requests[count - 1] = MPI_REQUEST_NULL;
239 int smpi_mpi_waitsome(int incount, MPI_Request requests[], int* indices, MPI_Status status[]) {
244 for(i = 0; i < count; i++) {
245 if(requests[i] != MPI_REQUEST_NULL) {
246 data = (MPI_Request)SIMIX_communication_get_data(requests[i]->pair);
247 if(data != MPI_REQUEST_NULL && data->complete == 1) {
248 finish_wait(&requests[i], status != MPI_STATUS_IGNORE ? &status[i] : MPI_STATUS_IGNORE);
257 void smpi_mpi_bcast(void* buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm) {
258 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
259 nary_tree_bcast(buf, count, datatype, root, comm, 4);
262 void smpi_mpi_barrier(MPI_Comm comm) {
263 // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
264 nary_tree_barrier(comm, 4);
267 void smpi_mpi_gather(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm) {
268 int system_tag = 666;
269 int rank, size, src, index, sendsize, recvsize;
270 MPI_Request* requests;
272 rank = smpi_comm_rank(comm);
273 size = smpi_comm_size(comm);
275 // Send buffer to root
276 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
278 sendsize = smpi_datatype_size(sendtype);
279 recvsize = smpi_datatype_size(recvtype);
280 // Local copy from root
281 memcpy(&((char*)recvbuf)[root * recvcount * recvsize], sendbuf, sendcount * sendsize * sizeof(char));
282 // Receive buffers from senders
283 requests = xbt_new(MPI_Request, size - 1);
285 for(src = 0; src < size; src++) {
287 requests[index] = smpi_mpi_irecv(&((char*)recvbuf)[src * recvcount * recvsize], recvcount, recvtype, src, system_tag, comm);
291 // Wait for completion of irecv's.
292 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
297 void smpi_mpi_gatherv(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int* recvcounts, int* displs, MPI_Datatype recvtype, int root, MPI_Comm comm) {
298 int system_tag = 666;
299 int rank, size, src, index, sendsize;
300 MPI_Request* requests;
302 rank = smpi_comm_rank(comm);
303 size = smpi_comm_size(comm);
305 // Send buffer to root
306 smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
308 sendsize = smpi_datatype_size(sendtype);
309 // Local copy from root
310 memcpy(&((char*)recvbuf)[displs[root]], sendbuf, sendcount * sendsize * sizeof(char));
311 // Receive buffers from senders
312 requests = xbt_new(MPI_Request, size - 1);
314 for(src = 0; src < size; src++) {
316 requests[index] = smpi_mpi_irecv(&((char*)recvbuf)[displs[src]], recvcounts[src], recvtype, src, system_tag, comm);
320 // Wait for completion of irecv's.
321 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
326 void smpi_mpi_allgather(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, MPI_Comm comm) {
327 int system_tag = 666;
328 int rank, size, other, index, sendsize, recvsize;
329 MPI_Request* requests;
331 rank = smpi_comm_rank(comm);
332 size = smpi_comm_size(comm);
333 sendsize = smpi_datatype_size(sendtype);
334 recvsize = smpi_datatype_size(recvtype);
335 // Local copy from self
336 memcpy(&((char*)recvbuf)[rank * recvcount * recvsize], sendbuf, sendcount * sendsize * sizeof(char));
337 // Send/Recv buffers to/from others;
338 requests = xbt_new(MPI_Request, 2 * (size - 1));
340 for(other = 0; other < size; other++) {
342 requests[index] = smpi_mpi_isend(sendbuf, sendcount, sendtype, other, system_tag, comm);
344 requests[index] = smpi_mpi_irecv(&((char*)recvbuf)[other * recvcount * recvsize], recvcount, recvtype, other, system_tag, comm);
348 // Wait for completion of all comms.
349 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
353 void smpi_mpi_allgatherv(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int* recvcounts, int* displs, MPI_Datatype recvtype, MPI_Comm comm) {
354 int system_tag = 666;
355 int rank, size, other, index, sendsize, recvsize;
356 MPI_Request* requests;
358 rank = smpi_comm_rank(comm);
359 size = smpi_comm_size(comm);
360 sendsize = smpi_datatype_size(sendtype);
361 recvsize = smpi_datatype_size(recvtype);
362 // Local copy from self
363 memcpy(&((char*)recvbuf)[displs[rank]], sendbuf, sendcount * sendsize * sizeof(char));
364 // Send buffers to others;
365 requests = xbt_new(MPI_Request, 2 * (size - 1));
367 for(other = 0; other < size; other++) {
369 requests[index] = smpi_mpi_isend(sendbuf, sendcount, sendtype, other, system_tag, comm);
371 requests[index] = smpi_mpi_irecv(&((char*)recvbuf)[displs[other]], recvcounts[other], recvtype, other, system_tag, comm);
375 // Wait for completion of all comms.
376 smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
380 void smpi_mpi_scatter(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm) {
381 int system_tag = 666;
382 int rank, size, dst, index, sendsize, recvsize;
383 MPI_Request* requests;
385 rank = smpi_comm_rank(comm);
386 size = smpi_comm_size(comm);
388 // Recv buffer from root
389 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm, MPI_STATUS_IGNORE);
391 sendsize = smpi_datatype_size(sendtype);
392 recvsize = smpi_datatype_size(recvtype);
393 // Local copy from root
394 memcpy(recvbuf, &((char*)sendbuf)[root * sendcount * sendsize], recvcount * recvsize * sizeof(char));
395 // Send buffers to receivers
396 requests = xbt_new(MPI_Request, size - 1);
398 for(dst = 0; dst < size; dst++) {
400 requests[index] = smpi_mpi_isend(&((char*)sendbuf)[dst * sendcount * sendsize], sendcount, sendtype, dst, system_tag, comm);
404 // Wait for completion of isend's.
405 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
410 void smpi_mpi_scatterv(void* sendbuf, int* sendcounts, int* displs, MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm) {
411 int system_tag = 666;
412 int rank, size, dst, index, sendsize, recvsize;
413 MPI_Request* requests;
415 rank = smpi_comm_rank(comm);
416 size = smpi_comm_size(comm);
418 // Recv buffer from root
419 smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm, MPI_STATUS_IGNORE);
421 sendsize = smpi_datatype_size(sendtype);
422 recvsize = smpi_datatype_size(recvtype);
423 // Local copy from root
424 memcpy(recvbuf, &((char*)sendbuf)[displs[root]], recvcount * recvsize * sizeof(char));
425 // Send buffers to receivers
426 requests = xbt_new(MPI_Request, size - 1);
428 for(dst = 0; dst < size; dst++) {
430 requests[index] = smpi_mpi_isend(&((char*)sendbuf)[displs[dst]], sendcounts[dst], sendtype, dst, system_tag, comm);
434 // Wait for completion of isend's.
435 smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
440 void smpi_mpi_reduce(void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm) {
441 int system_tag = 666;
442 int rank, size, src, index, datasize;
443 MPI_Request* requests;
446 rank = smpi_comm_rank(comm);
447 size = smpi_comm_size(comm);
449 // Send buffer to root
450 smpi_mpi_send(sendbuf, count, datatype, root, system_tag, comm);
452 datasize = smpi_datatype_size(datatype);
453 // Local copy from root
454 memcpy(recvbuf, sendbuf, count * datasize * sizeof(char));
455 // Receive buffers from senders
456 //TODO: make a MPI_barrier here ?
457 requests = xbt_new(MPI_Request, size - 1);
458 tmpbufs = xbt_new(void*, size - 1);
460 for(src = 0; src < size; src++) {
462 tmpbufs[index] = xbt_malloc(count * datasize);
463 requests[index] = smpi_mpi_irecv(tmpbufs[index], count, datatype, src, system_tag, comm);
467 // Wait for completion of irecv's.
468 for(src = 0; src < size - 1; src++) {
469 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
470 if(index == MPI_UNDEFINED) {
473 smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
475 for(index = 0; index < size - 1; index++) {
476 xbt_free(tmpbufs[index]);
483 void smpi_mpi_allreduce(void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) {
484 smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
485 smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
488 FIXME: buggy implementation
490 int system_tag = 666;
491 int rank, size, other, index, datasize;
492 MPI_Request* requests;
495 rank = smpi_comm_rank(comm);
496 size = smpi_comm_size(comm);
497 datasize = smpi_datatype_size(datatype);
498 // Local copy from self
499 memcpy(recvbuf, sendbuf, count * datasize * sizeof(char));
500 // Send/Recv buffers to/from others;
501 //TODO: make a MPI_barrier here ?
502 requests = xbt_new(MPI_Request, 2 * (size - 1));
503 tmpbufs = xbt_new(void*, size - 1);
505 for(other = 0; other < size; other++) {
507 tmpbufs[index / 2] = xbt_malloc(count * datasize);
508 requests[index] = smpi_mpi_isend(sendbuf, count, datatype, other, system_tag, comm);
509 requests[index + 1] = smpi_mpi_irecv(tmpbufs[index / 2], count, datatype, other, system_tag, comm);
513 // Wait for completion of all comms.
514 for(other = 0; other < 2 * (size - 1); other++) {
515 index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
516 if(index == MPI_UNDEFINED) {
519 if((index & 1) == 1) {
520 // Request is odd: it's a irecv
521 smpi_op_apply(op, tmpbufs[index / 2], recvbuf, &count, &datatype);
524 for(index = 0; index < size - 1; index++) {
525 xbt_free(tmpbufs[index]);