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.h"
9 static inline int MPIU_Mirror_permutation(unsigned int x, int bits)
11 /* a mask for the high order bits that should be copied as-is */
12 int high_mask = ~((0x1 << bits) - 1);
13 int retval = x & high_mask;
16 for (i = 0; i < bits; ++i) {
17 unsigned int bitval = (x & (0x1 << i)) >> i; /* 0x1 or 0x0 */
18 retval |= bitval << ((bits - i) - 1);
26 int Coll_reduce_scatter_mpich_pair::reduce_scatter(void *sendbuf, void *recvbuf, int recvcounts[],
27 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
29 int rank, comm_size, i;
30 MPI_Aint extent, true_extent, true_lb;
33 int mpi_errno = MPI_SUCCESS;
34 int total_count, dst, src;
36 comm_size = comm->size();
39 extent =datatype->get_extent();
40 datatype->extent(&true_lb, &true_extent);
42 if (op->is_commutative()) {
46 disps = (int*)xbt_malloc( comm_size * sizeof(int));
49 for (i=0; i<comm_size; i++) {
50 disps[i] = total_count;
51 total_count += recvcounts[i];
54 if (total_count == 0) {
59 if (sendbuf != MPI_IN_PLACE) {
60 /* copy local data into recvbuf */
61 Datatype::copy(((char *)sendbuf+disps[rank]*extent),
62 recvcounts[rank], datatype, recvbuf,
63 recvcounts[rank], datatype);
66 /* allocate temporary buffer to store incoming data */
67 tmp_recvbuf = (void*)smpi_get_tmp_recvbuffer(recvcounts[rank]*(MAX(true_extent,extent))+1);
68 /* adjust for potential negative lower bound in datatype */
69 tmp_recvbuf = (void *)((char*)tmp_recvbuf - true_lb);
71 for (i=1; i<comm_size; i++) {
72 src = (rank - i + comm_size) % comm_size;
73 dst = (rank + i) % comm_size;
75 /* send the data that dst needs. recv data that this process
76 needs from src into tmp_recvbuf */
77 if (sendbuf != MPI_IN_PLACE)
78 Request::sendrecv(((char *)sendbuf+disps[dst]*extent),
79 recvcounts[dst], datatype, dst,
80 COLL_TAG_SCATTER, tmp_recvbuf,
81 recvcounts[rank], datatype, src,
82 COLL_TAG_SCATTER, comm,
85 Request::sendrecv(((char *)recvbuf+disps[dst]*extent),
86 recvcounts[dst], datatype, dst,
87 COLL_TAG_SCATTER, tmp_recvbuf,
88 recvcounts[rank], datatype, src,
89 COLL_TAG_SCATTER, comm,
92 if (is_commutative || (src < rank)) {
93 if (sendbuf != MPI_IN_PLACE) {
94 if (op != MPI_OP_NULL)
95 op->apply(tmp_recvbuf, recvbuf, &recvcounts[rank], datatype);
98 if (op != MPI_OP_NULL)
99 op->apply(tmp_recvbuf, ((char*)recvbuf + disps[rank] * extent), &recvcounts[rank], datatype);
100 /* we can't store the result at the beginning of
101 recvbuf right here because there is useful data
102 there that other process/processes need. at the
103 end, we will copy back the result to the
104 beginning of recvbuf. */
108 if (sendbuf != MPI_IN_PLACE) {
109 if (op != MPI_OP_NULL)
110 op->apply(recvbuf, tmp_recvbuf, &recvcounts[rank], datatype);
111 /* copy result back into recvbuf */
113 Datatype::copy(tmp_recvbuf, recvcounts[rank], datatype, recvbuf, recvcounts[rank], datatype);
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);
129 /* if MPI_IN_PLACE, move output data to the beginning of
130 recvbuf. already done for rank 0. */
131 if ((sendbuf == MPI_IN_PLACE) && (rank != 0)) {
132 mpi_errno = Datatype::copy(((char *)recvbuf +
134 recvcounts[rank], datatype,
136 recvcounts[rank], datatype );
137 if (mpi_errno) return(mpi_errno);
141 smpi_free_tmp_buffer(tmp_recvbuf);
147 int Coll_reduce_scatter_mpich_noncomm::reduce_scatter(void *sendbuf, void *recvbuf, int recvcounts[],
148 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
150 int mpi_errno = MPI_SUCCESS;
151 int comm_size = comm->size() ;
152 int rank = comm->rank();
156 int recv_offset, send_offset;
157 int block_size, total_count, size;
158 MPI_Aint true_extent, true_lb;
164 datatype->extent(&true_lb, &true_extent);
168 while (pof2 < comm_size) {
173 /* begin error checking */
174 xbt_assert(pof2 == comm_size); /* FIXME this version only works for power of 2 procs */
176 for (i = 0; i < (comm_size - 1); ++i) {
177 xbt_assert(recvcounts[i] == recvcounts[i+1]);
179 /* end error checking */
181 /* size of a block (count of datatype per block, NOT bytes per block) */
182 block_size = recvcounts[0];
183 total_count = block_size * comm_size;
185 tmp_buf0=( void *)smpi_get_tmp_sendbuffer( true_extent * total_count);
186 tmp_buf1=( void *)smpi_get_tmp_recvbuffer( true_extent * total_count);
187 void *tmp_buf0_save=tmp_buf0;
188 void *tmp_buf1_save=tmp_buf1;
190 /* adjust for potential negative lower bound in datatype */
191 tmp_buf0 = (void *)((char*)tmp_buf0 - true_lb);
192 tmp_buf1 = (void *)((char*)tmp_buf1 - true_lb);
194 /* Copy our send data to tmp_buf0. We do this one block at a time and
195 permute the blocks as we go according to the mirror permutation. */
196 for (i = 0; i < comm_size; ++i) {
197 mpi_errno = Datatype::copy((char *)(sendbuf == MPI_IN_PLACE ? recvbuf : sendbuf) + (i * true_extent * block_size), block_size, datatype,
198 (char *)tmp_buf0 + (MPIU_Mirror_permutation(i, log2_comm_size) * true_extent * block_size), block_size, datatype);
199 if (mpi_errno) return(mpi_errno);
206 for (k = 0; k < log2_comm_size; ++k) {
207 /* use a double-buffering scheme to avoid local copies */
208 char *incoming_data = static_cast<char*>(buf0_was_inout ? tmp_buf1 : tmp_buf0);
209 char *outgoing_data = static_cast<char*>(buf0_was_inout ? tmp_buf0 : tmp_buf1);
210 int peer = rank ^ (0x1 << k);
214 /* we have the higher rank: send top half, recv bottom half */
218 /* we have the lower rank: recv top half, send bottom half */
222 Request::sendrecv(outgoing_data + send_offset*true_extent,
223 size, datatype, peer, COLL_TAG_SCATTER,
224 incoming_data + recv_offset*true_extent,
225 size, datatype, peer, COLL_TAG_SCATTER,
226 comm, MPI_STATUS_IGNORE);
227 /* always perform the reduction at recv_offset, the data at send_offset
228 is now our peer's responsibility */
230 /* higher ranked value so need to call op(received_data, my_data) */
231 if(op!=MPI_OP_NULL) op->apply(
232 incoming_data + recv_offset*true_extent,
233 outgoing_data + recv_offset*true_extent,
235 /* buf0_was_inout = buf0_was_inout; */
238 /* lower ranked value so need to call op(my_data, received_data) */
239 if (op != MPI_OP_NULL)
240 op->apply(outgoing_data + recv_offset * true_extent, incoming_data + recv_offset * true_extent, &size,
242 buf0_was_inout = not buf0_was_inout;
245 /* the next round of send/recv needs to happen within the block (of size
246 "size") that we just received and reduced */
247 send_offset = recv_offset;
250 xbt_assert(size == recvcounts[rank]);
252 /* copy the reduced data to the recvbuf */
253 result_ptr = (char *)(buf0_was_inout ? tmp_buf0 : tmp_buf1) + recv_offset * true_extent;
254 mpi_errno = Datatype::copy(result_ptr, size, datatype,
255 recvbuf, size, datatype);
256 smpi_free_tmp_buffer(tmp_buf0_save);
257 smpi_free_tmp_buffer(tmp_buf1_save);
258 if (mpi_errno) return(mpi_errno);
264 int Coll_reduce_scatter_mpich_rdb::reduce_scatter(void *sendbuf, void *recvbuf, int recvcounts[],
265 MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
267 int rank, comm_size, i;
268 MPI_Aint extent, true_extent, true_lb;
270 void *tmp_recvbuf, *tmp_results;
271 int mpi_errno = MPI_SUCCESS;
272 int dis[2], blklens[2], total_count, dst;
273 int mask, dst_tree_root, my_tree_root, j, k;
275 MPI_Datatype sendtype, recvtype;
276 int nprocs_completed, tmp_mask, tree_root, is_commutative=0;
277 comm_size = comm->size();
280 extent =datatype->get_extent();
281 datatype->extent(&true_lb, &true_extent);
283 if ((op==MPI_OP_NULL) || op->is_commutative()) {
287 disps = (int*)xbt_malloc( comm_size * sizeof(int));
290 for (i=0; i<comm_size; i++) {
291 disps[i] = total_count;
292 total_count += recvcounts[i];
295 /* noncommutative and (non-pof2 or block irregular), use recursive doubling. */
297 /* need to allocate temporary buffer to receive incoming data*/
298 tmp_recvbuf= (void *) smpi_get_tmp_recvbuffer( total_count*(MAX(true_extent,extent)));
299 /* adjust for potential negative lower bound in datatype */
300 tmp_recvbuf = (void *)((char*)tmp_recvbuf - true_lb);
302 /* need to allocate another temporary buffer to accumulate
304 tmp_results = (void *)smpi_get_tmp_sendbuffer( total_count*(MAX(true_extent,extent)));
305 /* adjust for potential negative lower bound in datatype */
306 tmp_results = (void *)((char*)tmp_results - true_lb);
308 /* copy sendbuf into tmp_results */
309 if (sendbuf != MPI_IN_PLACE)
310 mpi_errno = Datatype::copy(sendbuf, total_count, datatype,
311 tmp_results, total_count, datatype);
313 mpi_errno = Datatype::copy(recvbuf, total_count, datatype,
314 tmp_results, total_count, datatype);
316 if (mpi_errno) return(mpi_errno);
320 while (mask < comm_size) {
323 dst_tree_root = dst >> i;
326 my_tree_root = rank >> i;
329 /* At step 1, processes exchange (n-n/p) amount of
330 data; at step 2, (n-2n/p) amount of data; at step 3, (n-4n/p)
331 amount of data, and so forth. We use derived datatypes for this.
333 At each step, a process does not need to send data
334 indexed from my_tree_root to
335 my_tree_root+mask-1. Similarly, a process won't receive
336 data indexed from dst_tree_root to dst_tree_root+mask-1. */
338 /* calculate sendtype */
339 blklens[0] = blklens[1] = 0;
340 for (j=0; j<my_tree_root; j++)
341 blklens[0] += recvcounts[j];
342 for (j=my_tree_root+mask; j<comm_size; j++)
343 blklens[1] += recvcounts[j];
347 for (j=my_tree_root; (j<my_tree_root+mask) && (j<comm_size); j++)
348 dis[1] += recvcounts[j];
350 mpi_errno = Datatype::create_indexed(2, blklens, dis, datatype, &sendtype);
351 if (mpi_errno) return(mpi_errno);
355 /* calculate recvtype */
356 blklens[0] = blklens[1] = 0;
357 for (j=0; j<dst_tree_root && j<comm_size; j++)
358 blklens[0] += recvcounts[j];
359 for (j=dst_tree_root+mask; j<comm_size; j++)
360 blklens[1] += recvcounts[j];
364 for (j=dst_tree_root; (j<dst_tree_root+mask) && (j<comm_size); j++)
365 dis[1] += recvcounts[j];
367 mpi_errno = Datatype::create_indexed(2, blklens, dis, datatype, &recvtype);
368 if (mpi_errno) return(mpi_errno);
373 if (dst < comm_size) {
374 /* tmp_results contains data to be sent in each step. Data is
375 received in tmp_recvbuf and then accumulated into
376 tmp_results. accumulation is done later below. */
378 Request::sendrecv(tmp_results, 1, sendtype, dst,
380 tmp_recvbuf, 1, recvtype, dst,
381 COLL_TAG_SCATTER, comm,
386 /* if some processes in this process's subtree in this step
387 did not have any destination process to communicate with
388 because of non-power-of-two, we need to send them the
389 result. We use a logarithmic recursive-halfing algorithm
392 if (dst_tree_root + mask > comm_size) {
393 nprocs_completed = comm_size - my_tree_root - mask;
394 /* nprocs_completed is the number of processes in this
395 subtree that have all the data. Send data to others
396 in a tree fashion. First find root of current tree
397 that is being divided into two. k is the number of
398 least-significant bits in this process's rank that
399 must be zeroed out to find the rank of the root */
408 tmp_mask = mask >> 1;
410 dst = rank ^ tmp_mask;
412 tree_root = rank >> k;
415 /* send only if this proc has data and destination
416 doesn't have data. at any step, multiple processes
417 can send if they have the data */
419 (rank < tree_root + nprocs_completed)
420 && (dst >= tree_root + nprocs_completed)) {
421 /* send the current result */
422 Request::send(tmp_recvbuf, 1, recvtype,
423 dst, COLL_TAG_SCATTER,
426 /* recv only if this proc. doesn't have data and sender
428 else if ((dst < rank) &&
429 (dst < tree_root + nprocs_completed) &&
430 (rank >= tree_root + nprocs_completed)) {
431 Request::recv(tmp_recvbuf, 1, recvtype, dst,
433 comm, MPI_STATUS_IGNORE);
441 /* The following reduction is done here instead of after
442 the MPIC_Sendrecv_ft or MPIC_Recv_ft above. This is
443 because to do it above, in the noncommutative
444 case, we would need an extra temp buffer so as not to
445 overwrite temp_recvbuf, because temp_recvbuf may have
446 to be communicated to other processes in the
447 non-power-of-two case. To avoid that extra allocation,
448 we do the reduce here. */
450 if (is_commutative || (dst_tree_root < my_tree_root)) {
452 if (op != MPI_OP_NULL)
453 op->apply(tmp_recvbuf, tmp_results, &blklens[0], datatype);
454 if (op != MPI_OP_NULL)
455 op->apply(((char*)tmp_recvbuf + dis[1] * extent), ((char*)tmp_results + dis[1] * extent),
456 &blklens[1], datatype);
461 if (op != MPI_OP_NULL)
462 op->apply(tmp_results, tmp_recvbuf, &blklens[0], datatype);
463 if (op != MPI_OP_NULL)
464 op->apply(((char*)tmp_results + dis[1] * extent), ((char*)tmp_recvbuf + dis[1] * extent),
465 &blklens[1], datatype);
467 /* copy result back into tmp_results */
468 mpi_errno = Datatype::copy(tmp_recvbuf, 1, recvtype,
469 tmp_results, 1, recvtype);
470 if (mpi_errno) return(mpi_errno);
474 Datatype::unref(sendtype);
475 Datatype::unref(recvtype);
481 /* now copy final results from tmp_results to recvbuf */
482 mpi_errno = Datatype::copy(((char *)tmp_results+disps[rank]*extent),
483 recvcounts[rank], datatype, recvbuf,
484 recvcounts[rank], datatype);
485 if (mpi_errno) return(mpi_errno);
488 smpi_free_tmp_buffer(tmp_recvbuf);
489 smpi_free_tmp_buffer(tmp_results);