1 /* Copyright (c) 2013-2022. 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 reduce_scatter__mpich_pair(const void *sendbuf, void *recvbuf, const 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;
32 unsigned char* tmp_recvbuf;
33 int mpi_errno = MPI_SUCCESS;
34 int total_count, dst, src;
35 comm_size = comm->size();
38 extent =datatype->get_extent();
39 datatype->extent(&true_lb, &true_extent);
41 bool is_commutative = (op == MPI_OP_NULL || op->is_commutative());
43 int* disps = new int[comm_size];
46 for (i=0; i<comm_size; i++) {
47 disps[i] = total_count;
48 total_count += recvcounts[i];
51 if (total_count == 0) {
56 if (sendbuf != MPI_IN_PLACE) {
57 /* copy local data into recvbuf */
58 Datatype::copy(((char *)sendbuf+disps[rank]*extent),
59 recvcounts[rank], datatype, recvbuf,
60 recvcounts[rank], datatype);
63 /* allocate temporary buffer to store incoming data */
64 tmp_recvbuf = smpi_get_tmp_recvbuffer(recvcounts[rank] * std::max(true_extent, extent) + 1);
65 /* adjust for potential negative lower bound in datatype */
66 tmp_recvbuf = tmp_recvbuf - true_lb;
68 for (i=1; i<comm_size; i++) {
69 src = (rank - i + comm_size) % comm_size;
70 dst = (rank + i) % comm_size;
72 /* send the data that dst needs. recv data that this process
73 needs from src into tmp_recvbuf */
74 if (sendbuf != MPI_IN_PLACE)
75 Request::sendrecv(((char *)sendbuf+disps[dst]*extent),
76 recvcounts[dst], datatype, dst,
77 COLL_TAG_REDUCE_SCATTER, tmp_recvbuf,
78 recvcounts[rank], datatype, src,
79 COLL_TAG_REDUCE_SCATTER, comm,
82 Request::sendrecv(((char *)recvbuf+disps[dst]*extent),
83 recvcounts[dst], datatype, dst,
84 COLL_TAG_REDUCE_SCATTER, tmp_recvbuf,
85 recvcounts[rank], datatype, src,
86 COLL_TAG_REDUCE_SCATTER, comm,
89 if (is_commutative || (src < rank)) {
90 if (sendbuf != MPI_IN_PLACE) {
91 if (op != MPI_OP_NULL)
92 op->apply(tmp_recvbuf, recvbuf, &recvcounts[rank], datatype);
95 if (op != MPI_OP_NULL)
96 op->apply(tmp_recvbuf, ((char*)recvbuf + disps[rank] * extent), &recvcounts[rank], datatype);
97 /* we can't store the result at the beginning of
98 recvbuf right here because there is useful data
99 there that other process/processes need. at the
100 end, we will copy back the result to the
101 beginning of recvbuf. */
105 if (sendbuf != MPI_IN_PLACE) {
106 if (op != MPI_OP_NULL)
107 op->apply(recvbuf, tmp_recvbuf, &recvcounts[rank], datatype);
108 /* copy result back into recvbuf */
110 Datatype::copy(tmp_recvbuf, recvcounts[rank], datatype, recvbuf, recvcounts[rank], datatype);
113 smpi_free_tmp_buffer(tmp_recvbuf);
118 if (op != MPI_OP_NULL)
119 op->apply(((char*)recvbuf + disps[rank] * extent), tmp_recvbuf, &recvcounts[rank], datatype);
120 /* copy result back into recvbuf */
121 mpi_errno = Datatype::copy(tmp_recvbuf, recvcounts[rank], datatype,
122 ((char*)recvbuf + disps[rank] * extent), recvcounts[rank], datatype);
125 smpi_free_tmp_buffer(tmp_recvbuf);
132 /* if MPI_IN_PLACE, move output data to the beginning of
133 recvbuf. already done for rank 0. */
134 if ((sendbuf == MPI_IN_PLACE) && (rank != 0)) {
135 mpi_errno = Datatype::copy(((char *)recvbuf +
137 recvcounts[rank], datatype,
139 recvcounts[rank], datatype );
142 smpi_free_tmp_buffer(tmp_recvbuf);
148 smpi_free_tmp_buffer(tmp_recvbuf);
154 int reduce_scatter__mpich_noncomm(const void *sendbuf, void *recvbuf, const int recvcounts[],
155 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
157 int mpi_errno = MPI_SUCCESS;
158 int comm_size = comm->size() ;
159 int rank = comm->rank();
163 int recv_offset, send_offset;
164 int block_size, total_count, size;
165 MPI_Aint true_extent, true_lb;
167 unsigned char* tmp_buf0;
168 unsigned char* tmp_buf1;
169 unsigned char* result_ptr;
171 datatype->extent(&true_lb, &true_extent);
175 while (pof2 < comm_size) {
180 /* begin error checking */
181 xbt_assert(pof2 == comm_size); /* FIXME this version only works for power of 2 procs */
183 for (i = 0; i < (comm_size - 1); ++i) {
184 xbt_assert(recvcounts[i] == recvcounts[i+1]);
186 /* end error checking */
188 /* size of a block (count of datatype per block, NOT bytes per block) */
189 block_size = recvcounts[0];
190 total_count = block_size * comm_size;
192 tmp_buf0 = smpi_get_tmp_sendbuffer(true_extent * total_count);
193 tmp_buf1 = smpi_get_tmp_recvbuffer(true_extent * total_count);
194 unsigned char* tmp_buf0_save = tmp_buf0;
195 unsigned char* tmp_buf1_save = tmp_buf1;
197 /* adjust for potential negative lower bound in datatype */
198 tmp_buf0 = tmp_buf0 - true_lb;
199 tmp_buf1 = tmp_buf1 - true_lb;
201 /* Copy our send data to tmp_buf0. We do this one block at a time and
202 permute the blocks as we go according to the mirror permutation. */
203 for (i = 0; i < comm_size; ++i) {
204 mpi_errno = Datatype::copy(
205 static_cast<const char*>(sendbuf == MPI_IN_PLACE ? recvbuf : sendbuf) + (i * true_extent * block_size), block_size,
206 datatype, tmp_buf0 + (MPIU_Mirror_permutation(i, log2_comm_size) * true_extent * block_size), block_size,
216 for (k = 0; k < log2_comm_size; ++k) {
217 /* use a double-buffering scheme to avoid local copies */
218 unsigned char* incoming_data = buf0_was_inout ? tmp_buf1 : tmp_buf0;
219 unsigned char* outgoing_data = buf0_was_inout ? tmp_buf0 : tmp_buf1;
220 int peer = rank ^ (0x1 << k);
224 /* we have the higher rank: send top half, recv bottom half */
228 /* we have the lower rank: recv top half, send bottom half */
232 Request::sendrecv(outgoing_data + send_offset*true_extent,
233 size, datatype, peer, COLL_TAG_REDUCE_SCATTER,
234 incoming_data + recv_offset*true_extent,
235 size, datatype, peer, COLL_TAG_REDUCE_SCATTER,
236 comm, MPI_STATUS_IGNORE);
237 /* always perform the reduction at recv_offset, the data at send_offset
238 is now our peer's responsibility */
240 /* higher ranked value so need to call op(received_data, my_data) */
241 if(op!=MPI_OP_NULL) op->apply(
242 incoming_data + recv_offset*true_extent,
243 outgoing_data + recv_offset*true_extent,
245 /* buf0_was_inout = buf0_was_inout; */
248 /* lower ranked value so need to call op(my_data, received_data) */
249 if (op != MPI_OP_NULL)
250 op->apply(outgoing_data + recv_offset * true_extent, incoming_data + recv_offset * true_extent, &size,
252 buf0_was_inout = not buf0_was_inout;
255 /* the next round of send/recv needs to happen within the block (of size
256 "size") that we just received and reduced */
257 send_offset = recv_offset;
260 xbt_assert(size == recvcounts[rank]);
262 /* copy the reduced data to the recvbuf */
263 result_ptr = (buf0_was_inout ? tmp_buf0 : tmp_buf1) + recv_offset * true_extent;
264 mpi_errno = Datatype::copy(result_ptr, size, datatype,
265 recvbuf, size, datatype);
266 smpi_free_tmp_buffer(tmp_buf0_save);
267 smpi_free_tmp_buffer(tmp_buf1_save);
268 if (mpi_errno) return(mpi_errno);
274 int reduce_scatter__mpich_rdb(const void *sendbuf, void *recvbuf, const int recvcounts[],
275 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
277 int rank, comm_size, i;
278 MPI_Aint extent, true_extent, true_lb;
279 int mpi_errno = MPI_SUCCESS;
280 int dis[2], blklens[2], total_count, dst;
281 int mask, dst_tree_root, my_tree_root, j, k;
283 MPI_Datatype sendtype, recvtype;
284 int nprocs_completed, tmp_mask, tree_root;
285 comm_size = comm->size();
288 extent =datatype->get_extent();
289 datatype->extent(&true_lb, &true_extent);
291 bool is_commutative = (op == MPI_OP_NULL || op->is_commutative());
293 int* disps = new int[comm_size];
296 for (i=0; i<comm_size; i++) {
297 disps[i] = total_count;
298 total_count += recvcounts[i];
301 /* noncommutative and (non-pof2 or block irregular), use recursive doubling. */
303 /* need to allocate temporary buffer to receive incoming data*/
304 unsigned char* tmp_recvbuf = smpi_get_tmp_recvbuffer(total_count * std::max(true_extent, extent));
305 /* adjust for potential negative lower bound in datatype */
306 tmp_recvbuf = tmp_recvbuf - true_lb;
308 /* need to allocate another temporary buffer to accumulate
310 unsigned char* tmp_results = smpi_get_tmp_sendbuffer(total_count * std::max(true_extent, extent));
311 /* adjust for potential negative lower bound in datatype */
312 tmp_results = tmp_results - true_lb;
314 /* copy sendbuf into tmp_results */
315 if (sendbuf != MPI_IN_PLACE)
316 mpi_errno = Datatype::copy(sendbuf, total_count, datatype, tmp_results, total_count, datatype);
318 mpi_errno = Datatype::copy(recvbuf, total_count, datatype, tmp_results, total_count, datatype);
327 while (mask < comm_size) {
330 dst_tree_root = dst >> i;
333 my_tree_root = rank >> i;
336 /* At step 1, processes exchange (n-n/p) amount of
337 data; at step 2, (n-2n/p) amount of data; at step 3, (n-4n/p)
338 amount of data, and so forth. We use derived datatypes for this.
340 At each step, a process does not need to send data
341 indexed from my_tree_root to
342 my_tree_root+mask-1. Similarly, a process won't receive
343 data indexed from dst_tree_root to dst_tree_root+mask-1. */
345 /* calculate sendtype */
346 blklens[0] = blklens[1] = 0;
347 for (j = 0; j < my_tree_root; j++)
348 blklens[0] += recvcounts[j];
349 for (j = my_tree_root + mask; j < comm_size; j++)
350 blklens[1] += recvcounts[j];
354 for (j = my_tree_root; (j < my_tree_root + mask) && (j < comm_size); j++)
355 dis[1] += recvcounts[j];
357 mpi_errno = Datatype::create_indexed(2, blklens, dis, datatype, &sendtype);
363 /* calculate recvtype */
364 blklens[0] = blklens[1] = 0;
365 for (j = 0; j < dst_tree_root && j < comm_size; j++)
366 blklens[0] += recvcounts[j];
367 for (j = dst_tree_root + mask; j < comm_size; j++)
368 blklens[1] += recvcounts[j];
372 for (j = dst_tree_root; (j < dst_tree_root + mask) && (j < comm_size); j++)
373 dis[1] += recvcounts[j];
375 mpi_errno = Datatype::create_indexed(2, blklens, dis, datatype, &recvtype);
382 if (dst < comm_size) {
383 /* tmp_results contains data to be sent in each step. Data is
384 received in tmp_recvbuf and then accumulated into
385 tmp_results. accumulation is done later below. */
387 Request::sendrecv(tmp_results, 1, sendtype, dst, COLL_TAG_REDUCE_SCATTER, tmp_recvbuf, 1, recvtype, dst,
388 COLL_TAG_REDUCE_SCATTER, comm, MPI_STATUS_IGNORE);
392 /* if some processes in this process's subtree in this step
393 did not have any destination process to communicate with
394 because of non-power-of-two, we need to send them the
395 result. We use a logarithmic recursive-halfing algorithm
398 if (dst_tree_root + mask > comm_size) {
399 nprocs_completed = comm_size - my_tree_root - mask;
400 /* nprocs_completed is the number of processes in this
401 subtree that have all the data. Send data to others
402 in a tree fashion. First find root of current tree
403 that is being divided into two. k is the number of
404 least-significant bits in this process's rank that
405 must be zeroed out to find the rank of the root */
414 tmp_mask = mask >> 1;
416 dst = rank ^ tmp_mask;
418 tree_root = rank >> k;
421 /* send only if this proc has data and destination
422 doesn't have data. at any step, multiple processes
423 can send if they have the data */
424 if ((dst > rank) && (rank < tree_root + nprocs_completed) && (dst >= tree_root + nprocs_completed)) {
425 /* send the current result */
426 Request::send(tmp_recvbuf, 1, recvtype, dst, COLL_TAG_REDUCE_SCATTER, comm);
428 /* recv only if this proc. doesn't have data and sender
430 else if ((dst < rank) && (dst < tree_root + nprocs_completed) && (rank >= tree_root + nprocs_completed)) {
431 Request::recv(tmp_recvbuf, 1, recvtype, dst, COLL_TAG_REDUCE_SCATTER, comm, MPI_STATUS_IGNORE);
439 /* The following reduction is done here instead of after
440 the MPIC_Sendrecv_ft or MPIC_Recv_ft above. This is
441 because to do it above, in the noncommutative
442 case, we would need an extra temp buffer so as not to
443 overwrite temp_recvbuf, because temp_recvbuf may have
444 to be communicated to other processes in the
445 non-power-of-two case. To avoid that extra allocation,
446 we do the reduce here. */
448 if (is_commutative || (dst_tree_root < my_tree_root)) {
450 if (op != MPI_OP_NULL)
451 op->apply(tmp_recvbuf, tmp_results, &blklens[0], datatype);
452 if (op != MPI_OP_NULL)
453 op->apply(tmp_recvbuf + dis[1] * extent, tmp_results + dis[1] * extent, &blklens[1], datatype);
457 if (op != MPI_OP_NULL)
458 op->apply(tmp_results, tmp_recvbuf, &blklens[0], datatype);
459 if (op != MPI_OP_NULL)
460 op->apply(tmp_results + dis[1] * extent, tmp_recvbuf + dis[1] * extent, &blklens[1], datatype);
462 /* copy result back into tmp_results */
463 mpi_errno = Datatype::copy(tmp_recvbuf, 1, recvtype, tmp_results, 1, recvtype);
469 Datatype::unref(sendtype);
470 Datatype::unref(recvtype);
476 /* now copy final results from tmp_results to recvbuf */
477 mpi_errno = Datatype::copy(tmp_results + disps[rank] * extent, recvcounts[rank], datatype, recvbuf,
478 recvcounts[rank], datatype);
479 if (mpi_errno) return(mpi_errno);
482 smpi_free_tmp_buffer(tmp_recvbuf);
483 smpi_free_tmp_buffer(tmp_results);