3 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_mpi, smpi,
4 "Logging specific to SMPI (mpi)");
6 int SMPI_MPI_Init(int *argc, char ***argv)
8 smpi_process_init(argc,argv);
13 int SMPI_MPI_Finalize()
16 smpi_process_finalize();
20 // right now this just exits the current node, should send abort signal to all
21 // hosts in the communicator (TODO)
22 int SMPI_MPI_Abort(MPI_Comm comm, int errorcode)
28 int SMPI_MPI_Comm_size(MPI_Comm comm, int *size)
30 int retval = MPI_SUCCESS;
35 retval = MPI_ERR_COMM;
36 } else if (NULL == size) {
47 int SMPI_MPI_Comm_rank(MPI_Comm comm, int *rank)
49 int retval = MPI_SUCCESS;
54 retval = MPI_ERR_COMM;
55 } else if (NULL == rank) {
58 *rank = smpi_mpi_comm_rank(comm);
66 int SMPI_MPI_Type_size(MPI_Datatype datatype, size_t * size)
68 int retval = MPI_SUCCESS;
72 if (NULL == datatype) {
73 retval = MPI_ERR_TYPE;
74 } else if (NULL == size) {
77 *size = datatype->size;
85 int SMPI_MPI_Barrier(MPI_Comm comm)
87 int retval = MPI_SUCCESS;
92 retval = MPI_ERR_COMM;
94 retval = smpi_mpi_barrier(comm);
102 int SMPI_MPI_Irecv(void *buf, int count, MPI_Datatype datatype, int src,
103 int tag, MPI_Comm comm, MPI_Request * request)
105 int retval = MPI_SUCCESS;
109 retval = smpi_create_request(buf, count, datatype, src, 0, tag, comm,
111 if (NULL != *request && MPI_SUCCESS == retval) {
112 retval = smpi_mpi_irecv(*request);
120 int SMPI_MPI_Recv(void *buf, int count, MPI_Datatype datatype, int src,
121 int tag, MPI_Comm comm, MPI_Status * status)
123 int retval = MPI_SUCCESS;
124 smpi_mpi_request_t request;
128 retval = smpi_create_request(buf, count, datatype, src, 0, tag, comm,
130 if (NULL != request && MPI_SUCCESS == retval) {
131 retval = smpi_mpi_irecv(request);
132 if (MPI_SUCCESS == retval) {
133 retval = smpi_mpi_wait(request, status);
135 xbt_mallocator_release(smpi_global->request_mallocator, request);
143 int SMPI_MPI_Isend(void *buf, int count, MPI_Datatype datatype, int dst,
144 int tag, MPI_Comm comm, MPI_Request * request)
146 int retval = MPI_SUCCESS;
150 retval = smpi_create_request(buf, count, datatype, 0, dst, tag, comm,
152 if (NULL != *request && MPI_SUCCESS == retval) {
153 retval = smpi_mpi_isend(*request);
161 int SMPI_MPI_Send(void *buf, int count, MPI_Datatype datatype, int dst,
162 int tag, MPI_Comm comm)
164 int retval = MPI_SUCCESS;
165 smpi_mpi_request_t request;
169 retval = smpi_create_request(buf, count, datatype, 0, dst, tag, comm,
171 if (NULL != request && MPI_SUCCESS == retval) {
172 retval = smpi_mpi_isend(request);
173 if (MPI_SUCCESS == retval) {
174 smpi_mpi_wait(request, MPI_STATUS_IGNORE);
176 xbt_mallocator_release(smpi_global->request_mallocator, request);
185 * MPI_Wait and friends
187 int SMPI_MPI_Wait(MPI_Request * request, MPI_Status * status)
189 return smpi_mpi_wait(*request, status);
192 int SMPI_MPI_Waitall(int count, MPI_Request requests[], MPI_Status status[]) {
193 return smpi_mpi_waitall(count, requests,status);
195 int SMPI_MPI_Waitany(int count, MPI_Request requests[], int *index, MPI_Status status[]) {
196 return smpi_mpi_waitany(count, requests, index,status);
201 int SMPI_MPI_Bcast(void *buf, int count, MPI_Datatype datatype, int root,
205 int retval = MPI_SUCCESS;
207 smpi_mpi_request_t request;
211 rank = smpi_mpi_comm_rank(comm);
214 retval = smpi_create_request(buf, count, datatype, root,
215 (root + 1) % comm->size, 0, comm, &request);
216 request->forward = comm->size - 1;
217 smpi_mpi_isend(request);
219 retval = smpi_create_request(buf, count, datatype, MPI_ANY_SOURCE, rank,
221 smpi_mpi_irecv(request);
224 smpi_mpi_wait(request, MPI_STATUS_IGNORE);
225 xbt_mallocator_release(smpi_global->request_mallocator, request);
236 * debugging helper function
238 static void print_buffer_int( void *buf, int len, const char *msg, int rank) {
240 printf("**[%d] %s: ",rank,msg);
241 for (tmp=0;tmp<len;tmp++) {
243 printf("[%d]", v[tmp] );
253 int SMPI_MPI_Reduce( void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype,
254 MPI_Op op, int root, MPI_Comm comm )
256 int retval = MPI_SUCCESS;
261 smpi_mpi_request_t *tabrequest;
262 smpi_mpi_request_t request;
266 rank = smpi_mpi_comm_rank(comm);
269 if (rank != root) { // if i am not ROOT, simply send my buffer to root
272 print_buffer_int( sendbuf, count, xbt_strdup("sndbuf"),rank);
274 retval = smpi_create_request(sendbuf, count, datatype, rank, root, tag , comm, &request);
275 smpi_mpi_isend(request);
276 smpi_mpi_wait(request, MPI_STATUS_IGNORE);
277 xbt_mallocator_release(smpi_global->request_mallocator, request);
280 // i am the ROOT: wait for all buffers by creating one request by sender
282 tabrequest = xbt_malloc((size-1)*sizeof(smpi_mpi_request_t));
284 void **tmpbufs = xbt_malloc((size-1)*sizeof(void *));
285 for (i=0; i<size-1; i++) {
286 // we need 1 buffer per request to store intermediate receptions
287 tmpbufs[i] = xbt_malloc(count*datatype->size);
289 memcpy(recvbuf,sendbuf,count*datatype->size*sizeof(char)); // initiliaze recv buf with my own snd buf
291 // i can not use: 'request->forward = size-1;' (which would progagate size-1 receive reqs)
292 // since we should op values as soon as one receiving request matches.
293 for (i=0; i<size-1; i++) {
294 // reminder: for smpi_create_request() the src is always the process sending.
295 src = i < root ? i : i+1;
296 retval = smpi_create_request(tmpbufs[i], count, datatype,
298 tag, comm, &(tabrequest[i]));
299 if (NULL != tabrequest[i] && MPI_SUCCESS == retval) {
300 if (MPI_SUCCESS == retval) {
301 smpi_mpi_irecv(tabrequest[i]);
305 // now, wait for completion of all irecv's.
306 for (i=0; i<size-1; i++) {
307 int index = MPI_UNDEFINED;
308 smpi_mpi_waitany(size-1, tabrequest, &index, MPI_STATUS_IGNORE);
311 printf("MPI_Waitany() unblocked: root received (completes req[index=%d])\n",index);
312 print_buffer_int( tmpbufs[index], count, bprintf("tmpbufs[index=%d] (value received)",index),rank);
316 op->func (tmpbufs[index],recvbuf,&count,&datatype);
318 print_buffer_int( recvbuf, count, xbt_strdup("rcvbuf"),rank);
321 //xbt_mallocator_release(smpi_global->request_mallocator, tabrequest[i]);
322 xbt_free( tmpbufs[index]);
324 xbt_free(tabrequest);
333 // used by comm_split to sort ranks based on key values
334 int smpi_compare_rankkeys(const void *a, const void *b);
335 int smpi_compare_rankkeys(const void *a, const void *b)
354 int SMPI_MPI_Comm_split(MPI_Comm comm, int color, int key,
357 int retval = MPI_SUCCESS;
360 smpi_mpi_request_t request;
362 smpi_mpi_status_t status;
366 // FIXME: need to test parameters
368 index = smpi_process_index();
369 rank = comm->index_to_rank_map[index];
374 // root node does most of the real work
376 int colormap[comm->size];
377 int keymap[comm->size];
378 int rankkeymap[comm->size * 2];
380 smpi_mpi_communicator_t tempcomm = NULL;
387 // FIXME: use scatter/gather or similar instead of individual comms
388 for (i = 1; i < comm->size; i++) {
389 retval = smpi_create_request(colorkey, 2, MPI_INT, MPI_ANY_SOURCE,
390 rank, MPI_ANY_TAG, comm, &request);
391 smpi_mpi_irecv(request);
392 smpi_mpi_wait(request, &status);
393 colormap[status.MPI_SOURCE] = colorkey[0];
394 keymap[status.MPI_SOURCE] = colorkey[1];
395 xbt_mallocator_release(smpi_global->request_mallocator, request);
398 for (i = 0; i < comm->size; i++) {
399 if (MPI_UNDEFINED == colormap[i]) {
402 // make a list of nodes with current color and sort by keys
404 for (j = i; j < comm->size; j++) {
405 if (colormap[i] == colormap[j]) {
406 colormap[j] = MPI_UNDEFINED;
407 rankkeymap[count * 2] = j;
408 rankkeymap[count * 2 + 1] = keymap[j];
412 qsort(rankkeymap, count, sizeof(int) * 2, &smpi_compare_rankkeys);
415 tempcomm = xbt_new(s_smpi_mpi_communicator_t, 1);
416 tempcomm->barrier_count = 0;
417 tempcomm->size = count;
418 tempcomm->barrier_mutex = SIMIX_mutex_init();
419 tempcomm->barrier_cond = SIMIX_cond_init();
420 tempcomm->rank_to_index_map = xbt_new(int, count);
421 tempcomm->index_to_rank_map = xbt_new(int, smpi_global->process_count);
422 for (j = 0; j < smpi_global->process_count; j++) {
423 tempcomm->index_to_rank_map[j] = -1;
425 for (j = 0; j < count; j++) {
426 indextmp = comm->rank_to_index_map[rankkeymap[j * 2]];
427 tempcomm->rank_to_index_map[j] = indextmp;
428 tempcomm->index_to_rank_map[indextmp] = j;
430 for (j = 0; j < count; j++) {
431 if (rankkeymap[j * 2]) {
432 retval = smpi_create_request(&j, 1, MPI_INT, 0,
433 rankkeymap[j * 2], 0, comm, &request);
434 request->data = tempcomm;
435 smpi_mpi_isend(request);
436 smpi_mpi_wait(request, &status);
437 xbt_mallocator_release(smpi_global->request_mallocator, request);
439 *comm_out = tempcomm;
446 retval = smpi_create_request(colorkey, 2, MPI_INT, rank, 0, 0, comm,
448 smpi_mpi_isend(request);
449 smpi_mpi_wait(request, &status);
450 xbt_mallocator_release(smpi_global->request_mallocator, request);
451 if (MPI_UNDEFINED != color) {
452 retval = smpi_create_request(colorkey, 1, MPI_INT, 0, rank, 0, comm,
454 smpi_mpi_irecv(request);
455 smpi_mpi_wait(request, &status);
456 *comm_out = request->data;
465 double SMPI_MPI_Wtime( void )
467 return ( SIMIX_get_clock() );