1 #include "colls_private.h"
2 #define MPIR_REDUCE_SCATTER_TAG 222
4 static inline int MPIU_Mirror_permutation(unsigned int x, int bits)
6 /* a mask for the high order bits that should be copied as-is */
7 int high_mask = ~((0x1 << bits) - 1);
8 int retval = x & high_mask;
11 for (i = 0; i < bits; ++i) {
12 unsigned int bitval = (x & (0x1 << i)) >> i; /* 0x1 or 0x0 */
13 retval |= bitval << ((bits - i) - 1);
20 int smpi_coll_tuned_reduce_scatter_mpich_pair(void *sendbuf, void *recvbuf, int recvcounts[],
21 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
23 int rank, comm_size, i;
24 MPI_Aint extent, true_extent, true_lb;
27 int mpi_errno = MPI_SUCCESS;
28 int total_count, dst, src;
30 comm_size = smpi_comm_size(comm);
31 rank = smpi_comm_rank(comm);
33 extent =smpi_datatype_get_extent(datatype);
34 smpi_datatype_extent(datatype, &true_lb, &true_extent);
36 if (smpi_op_is_commute(op)) {
40 disps = (int*)xbt_malloc( comm_size * sizeof(int));
43 for (i=0; i<comm_size; i++) {
44 disps[i] = total_count;
45 total_count += recvcounts[i];
48 if (total_count == 0) {
52 if (sendbuf != MPI_IN_PLACE) {
53 /* copy local data into recvbuf */
54 smpi_datatype_copy(((char *)sendbuf+disps[rank]*extent),
55 recvcounts[rank], datatype, recvbuf,
56 recvcounts[rank], datatype);
59 /* allocate temporary buffer to store incoming data */
60 tmp_recvbuf = (void*)xbt_malloc(recvcounts[rank]*(max(true_extent,extent))+1);
61 /* adjust for potential negative lower bound in datatype */
62 tmp_recvbuf = (void *)((char*)tmp_recvbuf - true_lb);
64 for (i=1; i<comm_size; i++) {
65 src = (rank - i + comm_size) % comm_size;
66 dst = (rank + i) % comm_size;
68 /* send the data that dst needs. recv data that this process
69 needs from src into tmp_recvbuf */
70 if (sendbuf != MPI_IN_PLACE)
71 smpi_mpi_sendrecv(((char *)sendbuf+disps[dst]*extent),
72 recvcounts[dst], datatype, dst,
73 MPIR_REDUCE_SCATTER_TAG, tmp_recvbuf,
74 recvcounts[rank], datatype, src,
75 MPIR_REDUCE_SCATTER_TAG, comm,
78 smpi_mpi_sendrecv(((char *)recvbuf+disps[dst]*extent),
79 recvcounts[dst], datatype, dst,
80 MPIR_REDUCE_SCATTER_TAG, tmp_recvbuf,
81 recvcounts[rank], datatype, src,
82 MPIR_REDUCE_SCATTER_TAG, comm,
85 if (is_commutative || (src < rank)) {
86 if (sendbuf != MPI_IN_PLACE) {
88 tmp_recvbuf, recvbuf, &recvcounts[rank],
93 tmp_recvbuf, ((char *)recvbuf+disps[rank]*extent),
94 &recvcounts[rank], &datatype);
95 /* we can't store the result at the beginning of
96 recvbuf right here because there is useful data
97 there that other process/processes need. at the
98 end, we will copy back the result to the
99 beginning of recvbuf. */
103 if (sendbuf != MPI_IN_PLACE) {
105 recvbuf, tmp_recvbuf, &recvcounts[rank], &datatype);
106 /* copy result back into recvbuf */
107 mpi_errno = smpi_datatype_copy(tmp_recvbuf, recvcounts[rank],
109 recvcounts[rank], datatype);
110 if (mpi_errno) return(mpi_errno);
114 ((char *)recvbuf+disps[rank]*extent),
115 tmp_recvbuf, &recvcounts[rank], &datatype);
116 /* copy result back into recvbuf */
117 mpi_errno = smpi_datatype_copy(tmp_recvbuf, recvcounts[rank],
121 recvcounts[rank], datatype);
122 if (mpi_errno) return(mpi_errno);
127 /* if MPI_IN_PLACE, move output data to the beginning of
128 recvbuf. already done for rank 0. */
129 if ((sendbuf == MPI_IN_PLACE) && (rank != 0)) {
130 mpi_errno = smpi_datatype_copy(((char *)recvbuf +
132 recvcounts[rank], datatype,
134 recvcounts[rank], datatype );
135 if (mpi_errno) return(mpi_errno);
142 int smpi_coll_tuned_reduce_scatter_mpich_noncomm(void *sendbuf, void *recvbuf, int recvcounts[],
143 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
145 int mpi_errno = MPI_SUCCESS;
146 int comm_size = smpi_comm_size(comm) ;
147 int rank = smpi_comm_rank(comm);
151 int recv_offset, send_offset;
152 int block_size, total_count, size;
153 MPI_Aint true_extent, true_lb;
159 smpi_datatype_extent(datatype, &true_lb, &true_extent);
163 while (pof2 < comm_size) {
168 /* begin error checking */
169 xbt_assert(pof2 == comm_size); /* FIXME this version only works for power of 2 procs */
171 for (i = 0; i < (comm_size - 1); ++i) {
172 xbt_assert(recvcounts[i] == recvcounts[i+1]);
174 /* end error checking */
176 /* size of a block (count of datatype per block, NOT bytes per block) */
177 block_size = recvcounts[0];
178 total_count = block_size * comm_size;
180 tmp_buf0=( void *)xbt_malloc( true_extent * total_count);
181 tmp_buf1=( void *)xbt_malloc( true_extent * total_count);
182 /* adjust for potential negative lower bound in datatype */
183 tmp_buf0 = (void *)((char*)tmp_buf0 - true_lb);
184 tmp_buf1 = (void *)((char*)tmp_buf1 - true_lb);
186 /* Copy our send data to tmp_buf0. We do this one block at a time and
187 permute the blocks as we go according to the mirror permutation. */
188 for (i = 0; i < comm_size; ++i) {
189 mpi_errno = smpi_datatype_copy((char *)(sendbuf == MPI_IN_PLACE ? recvbuf : sendbuf) + (i * true_extent * block_size), block_size, datatype,
190 (char *)tmp_buf0 + (MPIU_Mirror_permutation(i, log2_comm_size) * true_extent * block_size), block_size, datatype);
191 if (mpi_errno) return(mpi_errno);
198 for (k = 0; k < log2_comm_size; ++k) {
199 /* use a double-buffering scheme to avoid local copies */
200 char *incoming_data = (buf0_was_inout ? tmp_buf1 : tmp_buf0);
201 char *outgoing_data = (buf0_was_inout ? tmp_buf0 : tmp_buf1);
202 int peer = rank ^ (0x1 << k);
206 /* we have the higher rank: send top half, recv bottom half */
210 /* we have the lower rank: recv top half, send bottom half */
214 smpi_mpi_sendrecv(outgoing_data + send_offset*true_extent,
215 size, datatype, peer, MPIR_REDUCE_SCATTER_TAG,
216 incoming_data + recv_offset*true_extent,
217 size, datatype, peer, MPIR_REDUCE_SCATTER_TAG,
218 comm, MPI_STATUS_IGNORE);
219 /* always perform the reduction at recv_offset, the data at send_offset
220 is now our peer's responsibility */
222 /* higher ranked value so need to call op(received_data, my_data) */
224 incoming_data + recv_offset*true_extent,
225 outgoing_data + recv_offset*true_extent,
227 /* buf0_was_inout = buf0_was_inout; */
230 /* lower ranked value so need to call op(my_data, received_data) */
232 outgoing_data + recv_offset*true_extent,
233 incoming_data + recv_offset*true_extent,
235 buf0_was_inout = !buf0_was_inout;
238 /* the next round of send/recv needs to happen within the block (of size
239 "size") that we just received and reduced */
240 send_offset = recv_offset;
243 xbt_assert(size == recvcounts[rank]);
245 /* copy the reduced data to the recvbuf */
246 result_ptr = (char *)(buf0_was_inout ? tmp_buf0 : tmp_buf1) + recv_offset * true_extent;
247 mpi_errno = smpi_datatype_copy(result_ptr, size, datatype,
248 recvbuf, size, datatype);
249 if (mpi_errno) return(mpi_errno);
255 int smpi_coll_tuned_reduce_scatter_mpich_rdb(void *sendbuf, void *recvbuf, int recvcounts[],
256 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
258 int rank, comm_size, i;
259 MPI_Aint extent, true_extent, true_lb;
261 void *tmp_recvbuf, *tmp_results;
262 int mpi_errno = MPI_SUCCESS;
263 int dis[2], blklens[2], total_count, dst;
264 int mask, dst_tree_root, my_tree_root, j, k;
266 MPI_Datatype sendtype, recvtype;
267 int nprocs_completed, tmp_mask, tree_root, is_commutative;
268 comm_size = smpi_comm_size(comm);
269 rank = smpi_comm_rank(comm);
271 extent =smpi_datatype_get_extent(datatype);
272 smpi_datatype_extent(datatype, &true_lb, &true_extent);
274 if (smpi_op_is_commute(op)) {
278 disps = (int*)xbt_malloc( comm_size * sizeof(int));
281 for (i=0; i<comm_size; i++) {
282 disps[i] = total_count;
283 total_count += recvcounts[i];
286 /* noncommutative and (non-pof2 or block irregular), use recursive doubling. */
288 /* need to allocate temporary buffer to receive incoming data*/
289 tmp_recvbuf= (void *) xbt_malloc( total_count*(max(true_extent,extent)));
290 /* adjust for potential negative lower bound in datatype */
291 tmp_recvbuf = (void *)((char*)tmp_recvbuf - true_lb);
293 /* need to allocate another temporary buffer to accumulate
295 tmp_results = (void *)xbt_malloc( total_count*(max(true_extent,extent)));
296 /* adjust for potential negative lower bound in datatype */
297 tmp_results = (void *)((char*)tmp_results - true_lb);
299 /* copy sendbuf into tmp_results */
300 if (sendbuf != MPI_IN_PLACE)
301 mpi_errno = smpi_datatype_copy(sendbuf, total_count, datatype,
302 tmp_results, total_count, datatype);
304 mpi_errno = smpi_datatype_copy(recvbuf, total_count, datatype,
305 tmp_results, total_count, datatype);
307 if (mpi_errno) return(mpi_errno);
311 while (mask < comm_size) {
314 dst_tree_root = dst >> i;
317 my_tree_root = rank >> i;
320 /* At step 1, processes exchange (n-n/p) amount of
321 data; at step 2, (n-2n/p) amount of data; at step 3, (n-4n/p)
322 amount of data, and so forth. We use derived datatypes for this.
324 At each step, a process does not need to send data
325 indexed from my_tree_root to
326 my_tree_root+mask-1. Similarly, a process won't receive
327 data indexed from dst_tree_root to dst_tree_root+mask-1. */
329 /* calculate sendtype */
330 blklens[0] = blklens[1] = 0;
331 for (j=0; j<my_tree_root; j++)
332 blklens[0] += recvcounts[j];
333 for (j=my_tree_root+mask; j<comm_size; j++)
334 blklens[1] += recvcounts[j];
338 for (j=my_tree_root; (j<my_tree_root+mask) && (j<comm_size); j++)
339 dis[1] += recvcounts[j];
341 mpi_errno = smpi_datatype_indexed(2, blklens, dis, datatype, &sendtype);
342 if (mpi_errno) return(mpi_errno);
344 smpi_datatype_commit(&sendtype);
346 /* calculate recvtype */
347 blklens[0] = blklens[1] = 0;
348 for (j=0; j<dst_tree_root && j<comm_size; j++)
349 blklens[0] += recvcounts[j];
350 for (j=dst_tree_root+mask; j<comm_size; j++)
351 blklens[1] += recvcounts[j];
355 for (j=dst_tree_root; (j<dst_tree_root+mask) && (j<comm_size); j++)
356 dis[1] += recvcounts[j];
358 mpi_errno = smpi_datatype_indexed(2, blklens, dis, datatype, &recvtype);
359 if (mpi_errno) return(mpi_errno);
361 smpi_datatype_commit(&recvtype);
364 if (dst < comm_size) {
365 /* tmp_results contains data to be sent in each step. Data is
366 received in tmp_recvbuf and then accumulated into
367 tmp_results. accumulation is done later below. */
369 smpi_mpi_sendrecv(tmp_results, 1, sendtype, dst,
370 MPIR_REDUCE_SCATTER_TAG,
371 tmp_recvbuf, 1, recvtype, dst,
372 MPIR_REDUCE_SCATTER_TAG, comm,
377 /* if some processes in this process's subtree in this step
378 did not have any destination process to communicate with
379 because of non-power-of-two, we need to send them the
380 result. We use a logarithmic recursive-halfing algorithm
383 if (dst_tree_root + mask > comm_size) {
384 nprocs_completed = comm_size - my_tree_root - mask;
385 /* nprocs_completed is the number of processes in this
386 subtree that have all the data. Send data to others
387 in a tree fashion. First find root of current tree
388 that is being divided into two. k is the number of
389 least-significant bits in this process's rank that
390 must be zeroed out to find the rank of the root */
399 tmp_mask = mask >> 1;
401 dst = rank ^ tmp_mask;
403 tree_root = rank >> k;
406 /* send only if this proc has data and destination
407 doesn't have data. at any step, multiple processes
408 can send if they have the data */
410 (rank < tree_root + nprocs_completed)
411 && (dst >= tree_root + nprocs_completed)) {
412 /* send the current result */
413 smpi_mpi_send(tmp_recvbuf, 1, recvtype,
414 dst, MPIR_REDUCE_SCATTER_TAG,
417 /* recv only if this proc. doesn't have data and sender
419 else if ((dst < rank) &&
420 (dst < tree_root + nprocs_completed) &&
421 (rank >= tree_root + nprocs_completed)) {
422 smpi_mpi_recv(tmp_recvbuf, 1, recvtype, dst,
423 MPIR_REDUCE_SCATTER_TAG,
424 comm, MPI_STATUS_IGNORE);
432 /* The following reduction is done here instead of after
433 the MPIC_Sendrecv_ft or MPIC_Recv_ft above. This is
434 because to do it above, in the noncommutative
435 case, we would need an extra temp buffer so as not to
436 overwrite temp_recvbuf, because temp_recvbuf may have
437 to be communicated to other processes in the
438 non-power-of-two case. To avoid that extra allocation,
439 we do the reduce here. */
441 if (is_commutative || (dst_tree_root < my_tree_root)) {
444 tmp_recvbuf, tmp_results, &blklens[0],
447 ((char *)tmp_recvbuf + dis[1]*extent),
448 ((char *)tmp_results + dis[1]*extent),
449 &blklens[1], &datatype);
455 tmp_results, tmp_recvbuf, &blklens[0],
458 ((char *)tmp_results + dis[1]*extent),
459 ((char *)tmp_recvbuf + dis[1]*extent),
460 &blklens[1], &datatype);
462 /* copy result back into tmp_results */
463 mpi_errno = smpi_datatype_copy(tmp_recvbuf, 1, recvtype,
464 tmp_results, 1, recvtype);
465 if (mpi_errno) return(mpi_errno);
469 //smpi_datatype_free(&sendtype);
470 //smpi_datatype_free(&recvtype);
476 /* now copy final results from tmp_results to recvbuf */
477 mpi_errno = smpi_datatype_copy(((char *)tmp_results+disps[rank]*extent),
478 recvcounts[rank], datatype, recvbuf,
479 recvcounts[rank], datatype);
480 if (mpi_errno) return(mpi_errno);