1 /* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
3 * Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
4 * University Research and Technology
5 * Corporation. All rights reserved.
6 * Copyright (c) 2004-2012 The University of Tennessee and The University
7 * of Tennessee Research Foundation. All rights
9 * Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
10 * University of Stuttgart. All rights reserved.
11 * Copyright (c) 2004-2005 The Regents of the University of California.
12 * All rights reserved.
13 * Copyright (c) 2008 Sun Microsystems, Inc. All rights reserved.
14 * Copyright (c) 2009 University of Houston. All rights reserved.
17 * Additional copyrights may follow
22 #include "colls_private.h"
23 #include "coll_tuned_topo.h"
25 #define MCA_COLL_BASE_TAG_REDUCE_SCATTER 222
27 * Recursive-halving function is (*mostly*) copied from the BASIC coll module.
28 * I have removed the part which handles "large" message sizes
29 * (non-overlapping version of reduce_Scatter).
32 /* copied function (with appropriate renaming) starts here */
35 * reduce_scatter_ompi_basic_recursivehalving
37 * Function: - reduce scatter implementation using recursive-halving
39 * Accepts: - same as MPI_Reduce_scatter()
40 * Returns: - MPI_SUCCESS or error code
41 * Limitation: - Works only for commutative operations.
44 smpi_coll_tuned_reduce_scatter_ompi_basic_recursivehalving(void *sbuf,
52 int i, rank, size, count, err = MPI_SUCCESS;
53 int tmp_size=1, remain = 0, tmp_rank, *disps = NULL;
54 ptrdiff_t true_lb, true_extent, lb, extent, buf_size;
55 char *recv_buf = NULL, *recv_buf_free = NULL;
56 char *result_buf = NULL, *result_buf_free = NULL;
59 rank = smpi_comm_rank(comm);
60 size = smpi_comm_size(comm);
62 XBT_DEBUG("coll:tuned:reduce_scatter_ompi_basic_recursivehalving, rank %d", rank);
64 /* Find displacements and the like */
65 disps = (int*) xbt_malloc(sizeof(int) * size);
66 if (NULL == disps) return MPI_ERR_OTHER;
69 for (i = 0; i < (size - 1); ++i) {
70 disps[i + 1] = disps[i] + rcounts[i];
72 count = disps[size - 1] + rcounts[size - 1];
74 /* short cut the trivial case */
80 /* get datatype information */
81 smpi_datatype_extent(dtype, &lb, &extent);
82 smpi_datatype_extent(dtype, &true_lb, &true_extent);
83 buf_size = true_extent + (ptrdiff_t)(count - 1) * extent;
85 /* Handle MPI_IN_PLACE */
86 if (MPI_IN_PLACE == sbuf) {
90 /* Allocate temporary receive buffer. */
91 recv_buf_free = (char*) xbt_malloc(buf_size);
92 recv_buf = recv_buf_free - lb;
93 if (NULL == recv_buf_free) {
98 /* allocate temporary buffer for results */
99 result_buf_free = (char*) xbt_malloc(buf_size);
100 result_buf = result_buf_free - lb;
102 /* copy local buffer into the temporary results */
103 err =smpi_datatype_copy(sbuf, count, dtype, result_buf, count, dtype);
104 if (MPI_SUCCESS != err) goto cleanup;
106 /* figure out power of two mapping: grow until larger than
107 comm size, then go back one, to get the largest power of
108 two less than comm size */
109 while (tmp_size <= size) tmp_size <<= 1;
111 remain = size - tmp_size;
113 /* If comm size is not a power of two, have the first "remain"
114 procs with an even rank send to rank + 1, leaving a power of
115 two procs to do the rest of the algorithm */
116 if (rank < 2 * remain) {
117 if ((rank & 1) == 0) {
118 smpi_mpi_send(result_buf, count, dtype, rank + 1,
119 MCA_COLL_BASE_TAG_REDUCE_SCATTER,
121 /* we don't participate from here on out */
124 smpi_mpi_recv(recv_buf, count, dtype, rank - 1,
125 MCA_COLL_BASE_TAG_REDUCE_SCATTER,
126 comm, MPI_STATUS_IGNORE);
128 /* integrate their results into our temp results */
129 smpi_op_apply(op, recv_buf, result_buf, &count, &dtype);
131 /* adjust rank to be the bottom "remain" ranks */
135 /* just need to adjust rank to show that the bottom "even
136 remain" ranks dropped out */
137 tmp_rank = rank - remain;
140 /* For ranks not kicked out by the above code, perform the
143 int *tmp_disps = NULL, *tmp_rcounts = NULL;
144 int mask, send_index, recv_index, last_index;
146 /* recalculate disps and rcounts to account for the
147 special "remainder" processes that are no longer doing
149 tmp_rcounts = (int*) xbt_malloc(tmp_size * sizeof(int));
150 if (NULL == tmp_rcounts) {
154 tmp_disps = (int*) xbt_malloc(tmp_size * sizeof(int));
155 if (NULL == tmp_disps) {
156 xbt_free(tmp_rcounts);
161 for (i = 0 ; i < tmp_size ; ++i) {
163 /* need to include old neighbor as well */
164 tmp_rcounts[i] = rcounts[i * 2 + 1] + rcounts[i * 2];
166 tmp_rcounts[i] = rcounts[i + remain];
171 for (i = 0; i < tmp_size - 1; ++i) {
172 tmp_disps[i + 1] = tmp_disps[i] + tmp_rcounts[i];
175 /* do the recursive halving communication. Don't use the
176 dimension information on the communicator because I
177 think the information is invalidated by our "shrinking"
178 of the communicator */
179 mask = tmp_size >> 1;
180 send_index = recv_index = 0;
181 last_index = tmp_size;
183 int tmp_peer, peer, send_count, recv_count;
186 tmp_peer = tmp_rank ^ mask;
187 peer = (tmp_peer < remain) ? tmp_peer * 2 + 1 : tmp_peer + remain;
189 /* figure out if we're sending, receiving, or both */
190 send_count = recv_count = 0;
191 if (tmp_rank < tmp_peer) {
192 send_index = recv_index + mask;
193 for (i = send_index ; i < last_index ; ++i) {
194 send_count += tmp_rcounts[i];
196 for (i = recv_index ; i < send_index ; ++i) {
197 recv_count += tmp_rcounts[i];
200 recv_index = send_index + mask;
201 for (i = send_index ; i < recv_index ; ++i) {
202 send_count += tmp_rcounts[i];
204 for (i = recv_index ; i < last_index ; ++i) {
205 recv_count += tmp_rcounts[i];
209 /* actual data transfer. Send from result_buf,
210 receive into recv_buf */
211 if (send_count > 0 && recv_count != 0) {
212 request=smpi_mpi_irecv(recv_buf + (ptrdiff_t)tmp_disps[recv_index] * extent,
213 recv_count, dtype, peer,
214 MCA_COLL_BASE_TAG_REDUCE_SCATTER,
216 if (MPI_SUCCESS != err) {
217 xbt_free(tmp_rcounts);
222 if (recv_count > 0 && send_count != 0) {
223 smpi_mpi_send(result_buf + (ptrdiff_t)tmp_disps[send_index] * extent,
224 send_count, dtype, peer,
225 MCA_COLL_BASE_TAG_REDUCE_SCATTER,
227 if (MPI_SUCCESS != err) {
228 xbt_free(tmp_rcounts);
233 if (send_count > 0 && recv_count != 0) {
234 smpi_mpi_wait(&request, MPI_STATUS_IGNORE);
237 /* if we received something on this step, push it into
238 the results buffer */
239 if (recv_count > 0) {
241 recv_buf + (ptrdiff_t)tmp_disps[recv_index] * extent,
242 result_buf + (ptrdiff_t)tmp_disps[recv_index] * extent,
243 &recv_count, &dtype);
246 /* update for next iteration */
247 send_index = recv_index;
248 last_index = recv_index + mask;
252 /* copy local results from results buffer into real receive buffer */
253 if (0 != rcounts[rank]) {
254 err = smpi_datatype_copy(result_buf + disps[rank] * extent,
255 rcounts[rank], dtype,
256 rbuf, rcounts[rank], dtype);
257 if (MPI_SUCCESS != err) {
258 xbt_free(tmp_rcounts);
264 xbt_free(tmp_rcounts);
268 /* Now fix up the non-power of two case, by having the odd
269 procs send the even procs the proper results */
270 if (rank < (2 * remain)) {
271 if ((rank & 1) == 0) {
273 smpi_mpi_recv(rbuf, rcounts[rank], dtype, rank + 1,
274 MCA_COLL_BASE_TAG_REDUCE_SCATTER,
275 comm, MPI_STATUS_IGNORE);
278 if (rcounts[rank - 1]) {
279 smpi_mpi_send(result_buf + disps[rank - 1] * extent,
280 rcounts[rank - 1], dtype, rank - 1,
281 MCA_COLL_BASE_TAG_REDUCE_SCATTER,
288 if (NULL != disps) xbt_free(disps);
289 if (NULL != recv_buf_free) xbt_free(recv_buf_free);
290 if (NULL != result_buf_free) xbt_free(result_buf_free);
295 /* copied function (with appropriate renaming) ends here */
299 * smpi_coll_tuned_reduce_scatter_ompi_ring
301 * Function: Ring algorithm for reduce_scatter operation
302 * Accepts: Same as MPI_Reduce_scatter()
303 * Returns: MPI_SUCCESS or error code
305 * Description: Implements ring algorithm for reduce_scatter:
306 * the block sizes defined in rcounts are exchanged and
307 8 updated until they reach proper destination.
308 * Algorithm requires 2 * max(rcounts) extra buffering
310 * Limitations: The algorithm DOES NOT preserve order of operations so it
311 * can be used only for commutative operations.
312 * Example on 5 nodes:
315 * [00] [10] -> [20] [30] [40]
316 * [01] [11] [21] -> [31] [41]
317 * [02] [12] [22] [32] -> [42]
318 * -> [03] [13] [23] [33] [43] --> ..
319 * [04] -> [14] [24] [34] [44]
322 * Step 0: rank r sends block (r-1) to rank (r+1) and
323 * receives block (r+1) from rank (r-1) [with wraparound].
325 * [00] [10] [10+20] -> [30] [40]
326 * [01] [11] [21] [21+31] -> [41]
327 * -> [02] [12] [22] [32] [32+42] -->..
328 * [43+03] -> [13] [23] [33] [43]
329 * [04] [04+14] -> [24] [34] [44]
333 * [00] [10] [10+20] [10+20+30] -> [40]
334 * -> [01] [11] [21] [21+31] [21+31+41] ->
335 * [32+42+02] -> [12] [22] [32] [32+42]
336 * [03] [43+03+13] -> [23] [33] [43]
337 * [04] [04+14] [04+14+24] -> [34] [44]
341 * -> [00] [10] [10+20] [10+20+30] [10+20+30+40] ->
342 * [21+31+41+01]-> [11] [21] [21+31] [21+31+41]
343 * [32+42+02] [32+42+02+12]-> [22] [32] [32+42]
344 * [03] [43+03+13] [43+03+13+23]-> [33] [43]
345 * [04] [04+14] [04+14+24] [04+14+24+34] -> [44]
349 * [10+20+30+40+00] [10] [10+20] [10+20+30] [10+20+30+40]
350 * [21+31+41+01] [21+31+41+01+11] [21] [21+31] [21+31+41]
351 * [32+42+02] [32+42+02+12] [32+42+02+12+22] [32] [32+42]
352 * [03] [43+03+13] [43+03+13+23] [43+03+13+23+33] [43]
353 * [04] [04+14] [04+14+24] [04+14+24+34] [04+14+24+34+44]
358 smpi_coll_tuned_reduce_scatter_ompi_ring(void *sbuf, void *rbuf, int *rcounts,
364 int ret, line, rank, size, i, k, recv_from, send_to, total_count, max_block_count;
365 int inbi, *displs = NULL;
366 char *tmpsend = NULL, *tmprecv = NULL, *accumbuf = NULL, *accumbuf_free = NULL;
367 char *inbuf_free[2] = {NULL, NULL}, *inbuf[2] = {NULL, NULL};
368 ptrdiff_t true_lb, true_extent, lb, extent, max_real_segsize;
369 MPI_Request reqs[2] = {NULL, NULL};
371 size = smpi_comm_size(comm);
372 rank = smpi_comm_rank(comm);
374 XBT_DEBUG( "coll:tuned:reduce_scatter_ompi_ring rank %d, size %d",
377 /* Determine the maximum number of elements per node,
378 corresponding block size, and displacements array.
380 displs = (int*) xbt_malloc(size * sizeof(int));
381 if (NULL == displs) { ret = -1; line = __LINE__; goto error_hndl; }
383 total_count = rcounts[0];
384 max_block_count = rcounts[0];
385 for (i = 1; i < size; i++) {
386 displs[i] = total_count;
387 total_count += rcounts[i];
388 if (max_block_count < rcounts[i]) max_block_count = rcounts[i];
391 /* Special case for size == 1 */
393 if (MPI_IN_PLACE != sbuf) {
394 ret = smpi_datatype_copy((char*)sbuf, total_count, dtype, (char*)rbuf, total_count, dtype);
395 if (ret < 0) { line = __LINE__; goto error_hndl; }
401 /* Allocate and initialize temporary buffers, we need:
402 - a temporary buffer to perform reduction (size total_count) since
403 rbuf can be of rcounts[rank] size.
404 - up to two temporary buffers used for communication/computation overlap.
406 smpi_datatype_extent(dtype, &lb, &extent);
407 smpi_datatype_extent(dtype, &true_lb, &true_extent);
409 max_real_segsize = true_extent + (ptrdiff_t)(max_block_count - 1) * extent;
411 accumbuf_free = (char*)xbt_malloc(true_extent + (ptrdiff_t)(total_count - 1) * extent);
412 if (NULL == accumbuf_free) { ret = -1; line = __LINE__; goto error_hndl; }
413 accumbuf = accumbuf_free - lb;
415 inbuf_free[0] = (char*)xbt_malloc(max_real_segsize);
416 if (NULL == inbuf_free[0]) { ret = -1; line = __LINE__; goto error_hndl; }
417 inbuf[0] = inbuf_free[0] - lb;
419 inbuf_free[1] = (char*)xbt_malloc(max_real_segsize);
420 if (NULL == inbuf_free[1]) { ret = -1; line = __LINE__; goto error_hndl; }
421 inbuf[1] = inbuf_free[1] - lb;
424 /* Handle MPI_IN_PLACE for size > 1 */
425 if (MPI_IN_PLACE == sbuf) {
429 ret = smpi_datatype_copy((char*)sbuf, total_count, dtype, accumbuf, total_count, dtype);
430 if (ret < 0) { line = __LINE__; goto error_hndl; }
432 /* Computation loop */
435 For each of the remote nodes:
436 - post irecv for block (r-2) from (r-1) with wrap around
437 - send block (r-1) to (r+1)
438 - in loop for every step k = 2 .. n
439 - post irecv for block (r - 1 + n - k) % n
440 - wait on block (r + n - k) % n to arrive
441 - compute on block (r + n - k ) % n
442 - send block (r + n - k) % n
444 - compute on block (r)
445 - copy block (r) to rbuf
446 Note that we must be careful when computing the begining of buffers and
447 for send operations and computation we must compute the exact block size.
449 send_to = (rank + 1) % size;
450 recv_from = (rank + size - 1) % size;
453 /* Initialize first receive from the neighbor on the left */
454 reqs[inbi]=smpi_mpi_irecv(inbuf[inbi], max_block_count, dtype, recv_from,
455 MCA_COLL_BASE_TAG_REDUCE_SCATTER, comm
457 tmpsend = accumbuf + (ptrdiff_t)displs[recv_from] * extent;
458 smpi_mpi_send(tmpsend, rcounts[recv_from], dtype, send_to,
459 MCA_COLL_BASE_TAG_REDUCE_SCATTER,
462 for (k = 2; k < size; k++) {
463 const int prevblock = (rank + size - k) % size;
467 /* Post irecv for the current block */
468 reqs[inbi]=smpi_mpi_irecv(inbuf[inbi], max_block_count, dtype, recv_from,
469 MCA_COLL_BASE_TAG_REDUCE_SCATTER, comm
472 /* Wait on previous block to arrive */
473 smpi_mpi_wait(&reqs[inbi ^ 0x1], MPI_STATUS_IGNORE);
475 /* Apply operation on previous block: result goes to rbuf
476 rbuf[prevblock] = inbuf[inbi ^ 0x1] (op) rbuf[prevblock]
478 tmprecv = accumbuf + (ptrdiff_t)displs[prevblock] * extent;
479 smpi_op_apply(op, inbuf[inbi ^ 0x1], tmprecv, &(rcounts[prevblock]), &dtype);
481 /* send previous block to send_to */
482 smpi_mpi_send(tmprecv, rcounts[prevblock], dtype, send_to,
483 MCA_COLL_BASE_TAG_REDUCE_SCATTER,
487 /* Wait on the last block to arrive */
488 smpi_mpi_wait(&reqs[inbi], MPI_STATUS_IGNORE);
490 /* Apply operation on the last block (my block)
491 rbuf[rank] = inbuf[inbi] (op) rbuf[rank] */
492 tmprecv = accumbuf + (ptrdiff_t)displs[rank] * extent;
493 smpi_op_apply(op, inbuf[inbi], tmprecv, &(rcounts[rank]), &dtype);
495 /* Copy result from tmprecv to rbuf */
496 ret = smpi_datatype_copy(tmprecv, rcounts[rank], dtype, (char*)rbuf, rcounts[rank], dtype);
497 if (ret < 0) { line = __LINE__; goto error_hndl; }
499 if (NULL != displs) xbt_free(displs);
500 if (NULL != accumbuf_free) xbt_free(accumbuf_free);
501 if (NULL != inbuf_free[0]) xbt_free(inbuf_free[0]);
502 if (NULL != inbuf_free[1]) xbt_free(inbuf_free[1]);
507 XBT_DEBUG( "%s:%4d\tRank %d Error occurred %d\n",
508 __FILE__, line, rank, ret);
509 if (NULL != displs) xbt_free(displs);
510 if (NULL != accumbuf_free) xbt_free(accumbuf_free);
511 if (NULL != inbuf_free[0]) xbt_free(inbuf_free[0]);
512 if (NULL != inbuf_free[1]) xbt_free(inbuf_free[1]);