1 /* Copyright (c) 2013-2017. The SimGrid Team.
2 * All rights reserved. */
4 /* This program is free software; you can redistribute it and/or modify it
5 * under the terms of the license (GNU LGPL) which comes with this package. */
7 #include "../colls_private.hpp"
10 static inline int MPIU_Mirror_permutation(unsigned int x, int bits)
12 /* a mask for the high order bits that should be copied as-is */
13 int high_mask = ~((0x1 << bits) - 1);
14 int retval = x & high_mask;
17 for (i = 0; i < bits; ++i) {
18 unsigned int bitval = (x & (0x1 << i)) >> i; /* 0x1 or 0x0 */
19 retval |= bitval << ((bits - i) - 1);
27 int Coll_reduce_scatter_mpich_pair::reduce_scatter(void *sendbuf, void *recvbuf, int recvcounts[],
28 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
30 int rank, comm_size, i;
31 MPI_Aint extent, true_extent, true_lb;
34 int mpi_errno = MPI_SUCCESS;
35 int total_count, dst, src;
37 comm_size = comm->size();
40 extent =datatype->get_extent();
41 datatype->extent(&true_lb, &true_extent);
43 if (op->is_commutative()) {
47 disps = (int*)xbt_malloc( comm_size * sizeof(int));
50 for (i=0; i<comm_size; i++) {
51 disps[i] = total_count;
52 total_count += recvcounts[i];
55 if (total_count == 0) {
60 if (sendbuf != MPI_IN_PLACE) {
61 /* copy local data into recvbuf */
62 Datatype::copy(((char *)sendbuf+disps[rank]*extent),
63 recvcounts[rank], datatype, recvbuf,
64 recvcounts[rank], datatype);
67 /* allocate temporary buffer to store incoming data */
68 tmp_recvbuf = (void*)smpi_get_tmp_recvbuffer(recvcounts[rank] * std::max(true_extent, extent) + 1);
69 /* adjust for potential negative lower bound in datatype */
70 tmp_recvbuf = (void *)((char*)tmp_recvbuf - true_lb);
72 for (i=1; i<comm_size; i++) {
73 src = (rank - i + comm_size) % comm_size;
74 dst = (rank + i) % comm_size;
76 /* send the data that dst needs. recv data that this process
77 needs from src into tmp_recvbuf */
78 if (sendbuf != MPI_IN_PLACE)
79 Request::sendrecv(((char *)sendbuf+disps[dst]*extent),
80 recvcounts[dst], datatype, dst,
81 COLL_TAG_SCATTER, tmp_recvbuf,
82 recvcounts[rank], datatype, src,
83 COLL_TAG_SCATTER, comm,
86 Request::sendrecv(((char *)recvbuf+disps[dst]*extent),
87 recvcounts[dst], datatype, dst,
88 COLL_TAG_SCATTER, tmp_recvbuf,
89 recvcounts[rank], datatype, src,
90 COLL_TAG_SCATTER, comm,
93 if (is_commutative || (src < rank)) {
94 if (sendbuf != MPI_IN_PLACE) {
95 if (op != MPI_OP_NULL)
96 op->apply(tmp_recvbuf, recvbuf, &recvcounts[rank], datatype);
99 if (op != MPI_OP_NULL)
100 op->apply(tmp_recvbuf, ((char*)recvbuf + disps[rank] * extent), &recvcounts[rank], datatype);
101 /* we can't store the result at the beginning of
102 recvbuf right here because there is useful data
103 there that other process/processes need. at the
104 end, we will copy back the result to the
105 beginning of recvbuf. */
109 if (sendbuf != MPI_IN_PLACE) {
110 if (op != MPI_OP_NULL)
111 op->apply(recvbuf, tmp_recvbuf, &recvcounts[rank], datatype);
112 /* copy result back into recvbuf */
114 Datatype::copy(tmp_recvbuf, recvcounts[rank], datatype, recvbuf, recvcounts[rank], datatype);
119 if (op != MPI_OP_NULL)
120 op->apply(((char*)recvbuf + disps[rank] * extent), tmp_recvbuf, &recvcounts[rank], datatype);
121 /* copy result back into recvbuf */
122 mpi_errno = Datatype::copy(tmp_recvbuf, recvcounts[rank], datatype,
123 ((char*)recvbuf + disps[rank] * extent), recvcounts[rank], datatype);
130 /* if MPI_IN_PLACE, move output data to the beginning of
131 recvbuf. already done for rank 0. */
132 if ((sendbuf == MPI_IN_PLACE) && (rank != 0)) {
133 mpi_errno = Datatype::copy(((char *)recvbuf +
135 recvcounts[rank], datatype,
137 recvcounts[rank], datatype );
138 if (mpi_errno) return(mpi_errno);
142 smpi_free_tmp_buffer(tmp_recvbuf);
148 int Coll_reduce_scatter_mpich_noncomm::reduce_scatter(void *sendbuf, void *recvbuf, int recvcounts[],
149 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
151 int mpi_errno = MPI_SUCCESS;
152 int comm_size = comm->size() ;
153 int rank = comm->rank();
157 int recv_offset, send_offset;
158 int block_size, total_count, size;
159 MPI_Aint true_extent, true_lb;
165 datatype->extent(&true_lb, &true_extent);
169 while (pof2 < comm_size) {
174 /* begin error checking */
175 xbt_assert(pof2 == comm_size); /* FIXME this version only works for power of 2 procs */
177 for (i = 0; i < (comm_size - 1); ++i) {
178 xbt_assert(recvcounts[i] == recvcounts[i+1]);
180 /* end error checking */
182 /* size of a block (count of datatype per block, NOT bytes per block) */
183 block_size = recvcounts[0];
184 total_count = block_size * comm_size;
186 tmp_buf0=( void *)smpi_get_tmp_sendbuffer( true_extent * total_count);
187 tmp_buf1=( void *)smpi_get_tmp_recvbuffer( true_extent * total_count);
188 void *tmp_buf0_save=tmp_buf0;
189 void *tmp_buf1_save=tmp_buf1;
191 /* adjust for potential negative lower bound in datatype */
192 tmp_buf0 = (void *)((char*)tmp_buf0 - true_lb);
193 tmp_buf1 = (void *)((char*)tmp_buf1 - true_lb);
195 /* Copy our send data to tmp_buf0. We do this one block at a time and
196 permute the blocks as we go according to the mirror permutation. */
197 for (i = 0; i < comm_size; ++i) {
198 mpi_errno = Datatype::copy((char *)(sendbuf == MPI_IN_PLACE ? recvbuf : sendbuf) + (i * true_extent * block_size), block_size, datatype,
199 (char *)tmp_buf0 + (MPIU_Mirror_permutation(i, log2_comm_size) * true_extent * block_size), block_size, datatype);
200 if (mpi_errno) return(mpi_errno);
207 for (k = 0; k < log2_comm_size; ++k) {
208 /* use a double-buffering scheme to avoid local copies */
209 char *incoming_data = static_cast<char*>(buf0_was_inout ? tmp_buf1 : tmp_buf0);
210 char *outgoing_data = static_cast<char*>(buf0_was_inout ? tmp_buf0 : tmp_buf1);
211 int peer = rank ^ (0x1 << k);
215 /* we have the higher rank: send top half, recv bottom half */
219 /* we have the lower rank: recv top half, send bottom half */
223 Request::sendrecv(outgoing_data + send_offset*true_extent,
224 size, datatype, peer, COLL_TAG_SCATTER,
225 incoming_data + recv_offset*true_extent,
226 size, datatype, peer, COLL_TAG_SCATTER,
227 comm, MPI_STATUS_IGNORE);
228 /* always perform the reduction at recv_offset, the data at send_offset
229 is now our peer's responsibility */
231 /* higher ranked value so need to call op(received_data, my_data) */
232 if(op!=MPI_OP_NULL) op->apply(
233 incoming_data + recv_offset*true_extent,
234 outgoing_data + recv_offset*true_extent,
236 /* buf0_was_inout = buf0_was_inout; */
239 /* lower ranked value so need to call op(my_data, received_data) */
240 if (op != MPI_OP_NULL)
241 op->apply(outgoing_data + recv_offset * true_extent, incoming_data + recv_offset * true_extent, &size,
243 buf0_was_inout = not buf0_was_inout;
246 /* the next round of send/recv needs to happen within the block (of size
247 "size") that we just received and reduced */
248 send_offset = recv_offset;
251 xbt_assert(size == recvcounts[rank]);
253 /* copy the reduced data to the recvbuf */
254 result_ptr = (char *)(buf0_was_inout ? tmp_buf0 : tmp_buf1) + recv_offset * true_extent;
255 mpi_errno = Datatype::copy(result_ptr, size, datatype,
256 recvbuf, size, datatype);
257 smpi_free_tmp_buffer(tmp_buf0_save);
258 smpi_free_tmp_buffer(tmp_buf1_save);
259 if (mpi_errno) return(mpi_errno);
265 int Coll_reduce_scatter_mpich_rdb::reduce_scatter(void *sendbuf, void *recvbuf, int recvcounts[],
266 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
268 int rank, comm_size, i;
269 MPI_Aint extent, true_extent, true_lb;
271 void *tmp_recvbuf, *tmp_results;
272 int mpi_errno = MPI_SUCCESS;
273 int dis[2], blklens[2], total_count, dst;
274 int mask, dst_tree_root, my_tree_root, j, k;
276 MPI_Datatype sendtype, recvtype;
277 int nprocs_completed, tmp_mask, tree_root, is_commutative=0;
278 comm_size = comm->size();
281 extent =datatype->get_extent();
282 datatype->extent(&true_lb, &true_extent);
284 if ((op==MPI_OP_NULL) || op->is_commutative()) {
288 disps = (int*)xbt_malloc( comm_size * sizeof(int));
291 for (i=0; i<comm_size; i++) {
292 disps[i] = total_count;
293 total_count += recvcounts[i];
296 /* noncommutative and (non-pof2 or block irregular), use recursive doubling. */
298 /* need to allocate temporary buffer to receive incoming data*/
299 tmp_recvbuf= (void*)smpi_get_tmp_recvbuffer(total_count * std::max(true_extent, extent));
300 /* adjust for potential negative lower bound in datatype */
301 tmp_recvbuf = (void *)((char*)tmp_recvbuf - true_lb);
303 /* need to allocate another temporary buffer to accumulate
305 tmp_results = (void*)smpi_get_tmp_sendbuffer(total_count * std::max(true_extent, extent));
306 /* adjust for potential negative lower bound in datatype */
307 tmp_results = (void *)((char*)tmp_results - true_lb);
309 /* copy sendbuf into tmp_results */
310 if (sendbuf != MPI_IN_PLACE)
311 mpi_errno = Datatype::copy(sendbuf, total_count, datatype,
312 tmp_results, total_count, datatype);
314 mpi_errno = Datatype::copy(recvbuf, total_count, datatype,
315 tmp_results, total_count, datatype);
317 if (mpi_errno) return(mpi_errno);
321 while (mask < comm_size) {
324 dst_tree_root = dst >> i;
327 my_tree_root = rank >> i;
330 /* At step 1, processes exchange (n-n/p) amount of
331 data; at step 2, (n-2n/p) amount of data; at step 3, (n-4n/p)
332 amount of data, and so forth. We use derived datatypes for this.
334 At each step, a process does not need to send data
335 indexed from my_tree_root to
336 my_tree_root+mask-1. Similarly, a process won't receive
337 data indexed from dst_tree_root to dst_tree_root+mask-1. */
339 /* calculate sendtype */
340 blklens[0] = blklens[1] = 0;
341 for (j=0; j<my_tree_root; j++)
342 blklens[0] += recvcounts[j];
343 for (j=my_tree_root+mask; j<comm_size; j++)
344 blklens[1] += recvcounts[j];
348 for (j=my_tree_root; (j<my_tree_root+mask) && (j<comm_size); j++)
349 dis[1] += recvcounts[j];
351 mpi_errno = Datatype::create_indexed(2, blklens, dis, datatype, &sendtype);
352 if (mpi_errno) return(mpi_errno);
356 /* calculate recvtype */
357 blklens[0] = blklens[1] = 0;
358 for (j=0; j<dst_tree_root && j<comm_size; j++)
359 blklens[0] += recvcounts[j];
360 for (j=dst_tree_root+mask; j<comm_size; j++)
361 blklens[1] += recvcounts[j];
365 for (j=dst_tree_root; (j<dst_tree_root+mask) && (j<comm_size); j++)
366 dis[1] += recvcounts[j];
368 mpi_errno = Datatype::create_indexed(2, blklens, dis, datatype, &recvtype);
369 if (mpi_errno) return(mpi_errno);
374 if (dst < comm_size) {
375 /* tmp_results contains data to be sent in each step. Data is
376 received in tmp_recvbuf and then accumulated into
377 tmp_results. accumulation is done later below. */
379 Request::sendrecv(tmp_results, 1, sendtype, dst,
381 tmp_recvbuf, 1, recvtype, dst,
382 COLL_TAG_SCATTER, comm,
387 /* if some processes in this process's subtree in this step
388 did not have any destination process to communicate with
389 because of non-power-of-two, we need to send them the
390 result. We use a logarithmic recursive-halfing algorithm
393 if (dst_tree_root + mask > comm_size) {
394 nprocs_completed = comm_size - my_tree_root - mask;
395 /* nprocs_completed is the number of processes in this
396 subtree that have all the data. Send data to others
397 in a tree fashion. First find root of current tree
398 that is being divided into two. k is the number of
399 least-significant bits in this process's rank that
400 must be zeroed out to find the rank of the root */
409 tmp_mask = mask >> 1;
411 dst = rank ^ tmp_mask;
413 tree_root = rank >> k;
416 /* send only if this proc has data and destination
417 doesn't have data. at any step, multiple processes
418 can send if they have the data */
420 (rank < tree_root + nprocs_completed)
421 && (dst >= tree_root + nprocs_completed)) {
422 /* send the current result */
423 Request::send(tmp_recvbuf, 1, recvtype,
424 dst, COLL_TAG_SCATTER,
427 /* recv only if this proc. doesn't have data and sender
429 else if ((dst < rank) &&
430 (dst < tree_root + nprocs_completed) &&
431 (rank >= tree_root + nprocs_completed)) {
432 Request::recv(tmp_recvbuf, 1, recvtype, dst,
434 comm, MPI_STATUS_IGNORE);
442 /* The following reduction is done here instead of after
443 the MPIC_Sendrecv_ft or MPIC_Recv_ft above. This is
444 because to do it above, in the noncommutative
445 case, we would need an extra temp buffer so as not to
446 overwrite temp_recvbuf, because temp_recvbuf may have
447 to be communicated to other processes in the
448 non-power-of-two case. To avoid that extra allocation,
449 we do the reduce here. */
451 if (is_commutative || (dst_tree_root < my_tree_root)) {
453 if (op != MPI_OP_NULL)
454 op->apply(tmp_recvbuf, tmp_results, &blklens[0], datatype);
455 if (op != MPI_OP_NULL)
456 op->apply(((char*)tmp_recvbuf + dis[1] * extent), ((char*)tmp_results + dis[1] * extent),
457 &blklens[1], datatype);
462 if (op != MPI_OP_NULL)
463 op->apply(tmp_results, tmp_recvbuf, &blklens[0], datatype);
464 if (op != MPI_OP_NULL)
465 op->apply(((char*)tmp_results + dis[1] * extent), ((char*)tmp_recvbuf + dis[1] * extent),
466 &blklens[1], datatype);
468 /* copy result back into tmp_results */
469 mpi_errno = Datatype::copy(tmp_recvbuf, 1, recvtype,
470 tmp_results, 1, recvtype);
471 if (mpi_errno) return(mpi_errno);
475 Datatype::unref(sendtype);
476 Datatype::unref(recvtype);
482 /* now copy final results from tmp_results to recvbuf */
483 mpi_errno = Datatype::copy(((char *)tmp_results+disps[rank]*extent),
484 recvcounts[rank], datatype, recvbuf,
485 recvcounts[rank], datatype);
486 if (mpi_errno) return(mpi_errno);
489 smpi_free_tmp_buffer(tmp_recvbuf);
490 smpi_free_tmp_buffer(tmp_results);