1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
|
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
* University Research and Technology
* Corporation. All rights reserved.
* Copyright (c) 2004-2006 The University of Tennessee and The University
* of Tennessee Research Foundation. All rights
* reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include "coll_basic.h"
#include <stdlib.h>
#include "mpi.h"
#include "ompi/constants.h"
#include "ompi/datatype/ompi_datatype.h"
#include "ompi/mca/coll/coll.h"
#include "ompi/mca/pml/pml.h"
#include "ompi/mca/coll/base/coll_tags.h"
#include "coll_basic.h"
/*
* allgather_intra
*
* Function: - allgather using other MPI collections
* Accepts: - same as MPI_Allgather()
* Returns: - MPI_SUCCESS or error code
*/
int
mca_coll_basic_allgather_intra(void *sbuf, int scount,
struct ompi_datatype_t *sdtype, void *rbuf,
int rcount, struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
int err;
ptrdiff_t lb, extent;
/* Handle MPI_IN_PLACE (see explanantion in reduce.c for how to
allocate temp buffer) -- note that rank 0 can use IN_PLACE
natively, and we can just alias the right position in rbuf
as sbuf and avoid using a temporary buffer if gather is
implemented correctly */
if (MPI_IN_PLACE == sbuf && 0 != ompi_comm_rank(comm)) {
ompi_datatype_get_extent(rdtype, &lb, &extent);
sbuf = ((char*) rbuf) + (ompi_comm_rank(comm) * extent * rcount);
sdtype = rdtype;
scount = rcount;
}
/* Gather and broadcast. */
err = comm->c_coll.coll_gather(sbuf, scount, sdtype, rbuf, rcount,
rdtype, 0, comm, comm->c_coll.coll_gather_module);
if (MPI_SUCCESS == err) {
err = comm->c_coll.coll_bcast(rbuf, rcount * ompi_comm_size(comm),
rdtype, 0, comm, comm->c_coll.coll_bcast_module);
}
/* All done */
return err;
}
/*
* allgather_inter
*
* Function: - allgather using other MPI collections
* Accepts: - same as MPI_Allgather()
* Returns: - MPI_SUCCESS or error code
*/
int
mca_coll_basic_allgather_inter(void *sbuf, int scount,
struct ompi_datatype_t *sdtype,
void *rbuf, int rcount,
struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
int rank, root = 0, size, rsize, err, i;
char *tmpbuf = NULL, *ptmp;
ptrdiff_t rlb, slb, rextent, sextent, incr;
ompi_request_t *req;
mca_coll_basic_module_t *basic_module = (mca_coll_basic_module_t*) module;
ompi_request_t **reqs = basic_module->mccb_reqs;
rank = ompi_comm_rank(comm);
size = ompi_comm_size(comm);
rsize = ompi_comm_remote_size(comm);
/* Algorithm:
* - a gather to the root in remote group (simultaniously executed,
* thats why we cannot use coll_gather).
* - exchange the temp-results between two roots
* - inter-bcast (again simultanious).
*/
/* Step one: gather operations: */
if (rank != root) {
/* send your data to root */
err = MCA_PML_CALL(send(sbuf, scount, sdtype, root,
MCA_COLL_BASE_TAG_ALLGATHER,
MCA_PML_BASE_SEND_STANDARD, comm));
if (OMPI_SUCCESS != err) {
return err;
}
} else {
/* receive a msg. from all other procs. */
err = ompi_datatype_get_extent(rdtype, &rlb, &rextent);
if (OMPI_SUCCESS != err) {
return err;
}
err = ompi_datatype_get_extent(sdtype, &slb, &sextent);
if (OMPI_SUCCESS != err) {
return err;
}
/* Do a send-recv between the two root procs. to avoid deadlock */
err = MCA_PML_CALL(isend(sbuf, scount, sdtype, 0,
MCA_COLL_BASE_TAG_ALLGATHER,
MCA_PML_BASE_SEND_STANDARD,
comm, &reqs[rsize]));
if (OMPI_SUCCESS != err) {
return err;
}
err = MCA_PML_CALL(irecv(rbuf, rcount, rdtype, 0,
MCA_COLL_BASE_TAG_ALLGATHER, comm,
&reqs[0]));
if (OMPI_SUCCESS != err) {
return err;
}
incr = rextent * rcount;
ptmp = (char *) rbuf + incr;
for (i = 1; i < rsize; ++i, ptmp += incr) {
err = MCA_PML_CALL(irecv(ptmp, rcount, rdtype, i,
MCA_COLL_BASE_TAG_ALLGATHER,
comm, &reqs[i]));
if (MPI_SUCCESS != err) {
return err;
}
}
err = ompi_request_wait_all(rsize + 1, reqs, MPI_STATUSES_IGNORE);
if (OMPI_SUCCESS != err) {
return err;
}
/* Step 2: exchange the resuts between the root processes */
tmpbuf = (char *) malloc(scount * size * sextent);
if (NULL == tmpbuf) {
return err;
}
err = MCA_PML_CALL(isend(rbuf, rsize * rcount, rdtype, 0,
MCA_COLL_BASE_TAG_ALLGATHER,
MCA_PML_BASE_SEND_STANDARD, comm, &req));
if (OMPI_SUCCESS != err) {
goto exit;
}
err = MCA_PML_CALL(recv(tmpbuf, size * scount, sdtype, 0,
MCA_COLL_BASE_TAG_ALLGATHER, comm,
MPI_STATUS_IGNORE));
if (OMPI_SUCCESS != err) {
goto exit;
}
err = ompi_request_wait_all(1, &req, MPI_STATUS_IGNORE);
if (OMPI_SUCCESS != err) {
goto exit;
}
}
/* Step 3: bcast the data to the remote group. This
* happens in both groups simultaniously, thus we can
* not use coll_bcast (this would deadlock).
*/
if (rank != root) {
/* post the recv */
err = MCA_PML_CALL(recv(rbuf, rsize * rcount, rdtype, 0,
MCA_COLL_BASE_TAG_ALLGATHER, comm,
MPI_STATUS_IGNORE));
if (OMPI_SUCCESS != err) {
goto exit;
}
} else {
/* Send the data to every other process in the remote group
* except to rank zero. which has it already. */
for (i = 1; i < rsize; i++) {
err = MCA_PML_CALL(isend(tmpbuf, size * scount, sdtype, i,
MCA_COLL_BASE_TAG_ALLGATHER,
MCA_PML_BASE_SEND_STANDARD,
comm, &reqs[i - 1]));
if (OMPI_SUCCESS != err) {
goto exit;
}
}
err = ompi_request_wait_all(rsize - 1, reqs, MPI_STATUSES_IGNORE);
if (OMPI_SUCCESS != err) {
goto exit;
}
}
exit:
if (NULL != tmpbuf) {
free(tmpbuf);
}
return err;
}
|
