1 /* Copyright (c) 2011-2019. The SimGrid Team. All rights reserved. */
3 /* This program is free software; you can redistribute it and/or modify it
4 * under the terms of the license (GNU LGPL) which comes with this package. */
6 #include "../colls_private.hpp"
7 #include "smpi_status.hpp"
12 static int scatter_for_bcast(
19 int rank, comm_size, src, dst;
20 int relative_rank, mask;
21 int mpi_errno = MPI_SUCCESS;
22 int scatter_size, curr_size, recv_size = 0, send_size;
24 comm_size = comm->size();
26 relative_rank = (rank >= root) ? rank - root : rank - root + comm_size;
28 /* use long message algorithm: binomial tree scatter followed by an allgather */
30 /* The scatter algorithm divides the buffer into nprocs pieces and
31 scatters them among the processes. Root gets the first piece,
32 root+1 gets the second piece, and so forth. Uses the same binomial
33 tree algorithm as above. Ceiling division
34 is used to compute the size of each piece. This means some
35 processes may not get any data. For example if bufsize = 97 and
36 nprocs = 16, ranks 15 and 16 will get 0 data. On each process, the
37 scattered data is stored at the same offset in the buffer as it is
38 on the root process. */
40 scatter_size = (nbytes + comm_size - 1)/comm_size; /* ceiling division */
41 curr_size = (rank == root) ? nbytes : 0; /* root starts with all the
45 while (mask < comm_size)
47 if (relative_rank & mask)
50 if (src < 0) src += comm_size;
51 recv_size = nbytes - relative_rank*scatter_size;
52 /* recv_size is larger than what might actually be sent by the
53 sender. We don't need compute the exact value because MPI
54 allows you to post a larger recv.*/
57 curr_size = 0; /* this process doesn't receive any data
58 because of uneven division */
62 Request::recv(((char *)tmp_buf +
63 relative_rank*scatter_size),
64 recv_size, MPI_BYTE, src,
65 COLL_TAG_BCAST, comm, &status);
66 /* query actual size of data received */
67 curr_size=Status::get_count(&status, MPI_BYTE);
74 /* This process is responsible for all processes that have bits
75 set from the LSB upto (but not including) mask. Because of
76 the "not including", we start by shifting mask back down
82 if (relative_rank + mask < comm_size)
84 send_size = curr_size - scatter_size * mask;
85 /* mask is also the size of this process's subtree */
90 if (dst >= comm_size) dst -= comm_size;
91 Request::send(((char *)tmp_buf +
92 scatter_size*(relative_rank+mask)),
93 send_size, MPI_BYTE, dst,
94 COLL_TAG_BCAST, comm);
95 curr_size -= send_size;
105 int bcast__scatter_rdb_allgather(
108 MPI_Datatype datatype,
113 int rank, comm_size, dst;
114 int relative_rank, mask;
115 int mpi_errno = MPI_SUCCESS;
116 int scatter_size, curr_size, recv_size = 0;
117 int j, k, i, tmp_mask, is_contig, is_homogeneous;
118 MPI_Aint type_size = 0, nbytes = 0;
119 int relative_dst, dst_tree_root, my_tree_root, send_offset;
120 int recv_offset, tree_root, nprocs_completed, offset;
122 MPI_Aint true_extent, true_lb;
125 comm_size = comm->size();
127 relative_rank = (rank >= root) ? rank - root : rank - root + comm_size;
129 /* If there is only one process, return */
130 if (comm_size == 1) goto fn_exit;
132 //if (HANDLE_GET_KIND(datatype) == HANDLE_KIND_BUILTIN)
133 if(datatype->flags() & DT_FLAG_CONTIGUOUS)
141 /* MPI_Type_size() might not give the accurate size of the packed
142 * datatype for heterogeneous systems (because of padding, encoding,
143 * etc). On the other hand, MPI_Pack_size() can become very
144 * expensive, depending on the implementation, especially for
145 * heterogeneous systems. We want to use MPI_Type_size() wherever
146 * possible, and MPI_Pack_size() in other places.
149 type_size=datatype->size();
151 nbytes = type_size * count;
153 goto fn_exit; /* nothing to do */
155 if (is_contig && is_homogeneous)
157 /* contiguous and homogeneous. no need to pack. */
158 datatype->extent(&true_lb, &true_extent);
160 tmp_buf = (char *) buffer + true_lb;
164 tmp_buf = new unsigned char[nbytes];
166 /* TODO: Pipeline the packing and communication */
169 mpi_errno = datatype->pack(buffer, count, tmp_buf, nbytes, &position, comm);
171 xbt_die("crash while packing %d", mpi_errno);
176 scatter_size = (nbytes + comm_size - 1)/comm_size; /* ceiling division */
178 mpi_errno = scatter_for_bcast(root, comm,
181 xbt_die("crash while scattering %d", mpi_errno);
184 /* curr_size is the amount of data that this process now has stored in
185 * buffer at byte offset (relative_rank*scatter_size) */
186 curr_size = scatter_size < (nbytes - (relative_rank * scatter_size)) ? scatter_size : (nbytes - (relative_rank * scatter_size)) ;
190 /* medium size allgather and pof2 comm_size. use recurive doubling. */
194 while (mask < comm_size)
196 relative_dst = relative_rank ^ mask;
198 dst = (relative_dst + root) % comm_size;
200 /* find offset into send and recv buffers.
201 zero out the least significant "i" bits of relative_rank and
202 relative_dst to find root of src and dst
203 subtrees. Use ranks of roots as index to send from
204 and recv into buffer */
206 dst_tree_root = relative_dst >> i;
209 my_tree_root = relative_rank >> i;
212 send_offset = my_tree_root * scatter_size;
213 recv_offset = dst_tree_root * scatter_size;
215 if (relative_dst < comm_size)
217 Request::sendrecv(((char *)tmp_buf + send_offset),
218 curr_size, MPI_BYTE, dst, COLL_TAG_BCAST,
219 ((char *)tmp_buf + recv_offset),
220 (nbytes-recv_offset < 0 ? 0 : nbytes-recv_offset),
221 MPI_BYTE, dst, COLL_TAG_BCAST, comm, &status);
222 recv_size=Status::get_count(&status, MPI_BYTE);
223 curr_size += recv_size;
226 /* if some processes in this process's subtree in this step
227 did not have any destination process to communicate with
228 because of non-power-of-two, we need to send them the
229 data that they would normally have received from those
230 processes. That is, the haves in this subtree must send to
231 the havenots. We use a logarithmic recursive-halfing algorithm
234 /* This part of the code will not currently be
235 executed because we are not using recursive
236 doubling for non power of two. Mark it as experimental
237 so that it doesn't show up as red in the coverage tests. */
239 /* --BEGIN EXPERIMENTAL-- */
240 if (dst_tree_root + mask > comm_size)
242 nprocs_completed = comm_size - my_tree_root - mask;
243 /* nprocs_completed is the number of processes in this
244 subtree that have all the data. Send data to others
245 in a tree fashion. First find root of current tree
246 that is being divided into two. k is the number of
247 least-significant bits in this process's rank that
248 must be zeroed out to find the rank of the root */
258 offset = scatter_size * (my_tree_root + mask);
259 tmp_mask = mask >> 1;
263 relative_dst = relative_rank ^ tmp_mask;
264 dst = (relative_dst + root) % comm_size;
266 tree_root = relative_rank >> k;
269 /* send only if this proc has data and destination
270 doesn't have data. */
273 printf("rank %d, dst %d, root %d, nprocs_completed %d\n", relative_rank, relative_dst, tree_root, nprocs_completed);
277 if ((relative_dst > relative_rank) &&
278 (relative_rank < tree_root + nprocs_completed)
279 && (relative_dst >= tree_root + nprocs_completed))
282 /* printf("Rank %d, send to %d, offset %d, size %d\n", rank, dst, offset, recv_size);
284 Request::send(((char *)tmp_buf + offset),
285 recv_size, MPI_BYTE, dst,
286 COLL_TAG_BCAST, comm);
287 /* recv_size was set in the previous
288 receive. that's the amount of data to be
291 /* recv only if this proc. doesn't have data and sender
293 else if ((relative_dst < relative_rank) &&
294 (relative_dst < tree_root + nprocs_completed) &&
295 (relative_rank >= tree_root + nprocs_completed))
297 /* printf("Rank %d waiting to recv from rank %d\n",
298 relative_rank, dst); */
299 Request::recv(((char *)tmp_buf + offset),
301 MPI_BYTE, dst, COLL_TAG_BCAST,
303 /* nprocs_completed is also equal to the no. of processes
304 whose data we don't have */
305 recv_size=Status::get_count(&status, MPI_BYTE);
306 curr_size += recv_size;
307 /* printf("Rank %d, recv from %d, offset %d, size %d\n", rank, dst, offset, recv_size);
314 /* --END EXPERIMENTAL-- */
320 /* check that we received as much as we expected */
321 /* recvd_size may not be accurate for packed heterogeneous data */
322 if (is_homogeneous && curr_size != nbytes) {
323 xbt_die("we didn't receive enough !");
326 if (not is_contig || not is_homogeneous) {
329 mpi_errno = MPI_Unpack(tmp_buf, nbytes, &position, buffer, count, datatype, comm);
331 xbt_die("error when unpacking %d", mpi_errno);
336 /* delete[] static_cast<unsigned char*>(tmp_buf); */