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
|
/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
/*
* Copyright (c) 2018 Siberian State University of Telecommunications
* and Information Science. All rights reserved.
* Copyright (c) 2018 Research Organization for Information Science
* and Technology (RIST). All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include "mpi.h"
#include "ompi/constants.h"
#include "ompi/datatype/ompi_datatype.h"
#include "ompi/communicator/communicator.h"
#include "ompi/mca/coll/coll.h"
#include "ompi/mca/coll/base/coll_base_functions.h"
#include "ompi/mca/coll/base/coll_tags.h"
#include "ompi/mca/coll/base/coll_base_util.h"
#include "ompi/mca/pml/pml.h"
#include "ompi/op/op.h"
/*
* ompi_coll_base_exscan_intra_linear
*
* Function: Linear algorithm for exclusive scan.
* Accepts: Same as MPI_Exscan
* Returns: MPI_SUCCESS or error code
*/
int
ompi_coll_base_exscan_intra_linear(const void *sbuf, void *rbuf, int count,
struct ompi_datatype_t *dtype,
struct ompi_op_t *op,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
int size, rank, err;
ptrdiff_t dsize, gap;
char *free_buffer = NULL;
char *reduce_buffer = NULL;
rank = ompi_comm_rank(comm);
size = ompi_comm_size(comm);
/* For MPI_IN_PLACE, just adjust send buffer to point to
* receive buffer. */
if (MPI_IN_PLACE == sbuf) {
sbuf = rbuf;
}
/* If we're rank 0, then just send our sbuf to the next rank, and
* we are done. */
if (0 == rank) {
return MCA_PML_CALL(send(sbuf, count, dtype, rank + 1,
MCA_COLL_BASE_TAG_EXSCAN,
MCA_PML_BASE_SEND_STANDARD, comm));
}
/* If we're the last rank, then just receive the result from the
* prior rank, and we are done. */
else if ((size - 1) == rank) {
return MCA_PML_CALL(recv(rbuf, count, dtype, rank - 1,
MCA_COLL_BASE_TAG_EXSCAN, comm,
MPI_STATUS_IGNORE));
}
/* Otherwise, get the result from the prior rank, combine it with my
* data, and send it to the next rank */
/* Get a temporary buffer to perform the reduction into. Rationale
* for malloc'ing this size is provided in coll_basic_reduce.c. */
dsize = opal_datatype_span(&dtype->super, count, &gap);
free_buffer = (char*)malloc(dsize);
if (NULL == free_buffer) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
reduce_buffer = free_buffer - gap;
err = ompi_datatype_copy_content_same_ddt(dtype, count,
reduce_buffer, (char*)sbuf);
/* Receive the reduced value from the prior rank */
err = MCA_PML_CALL(recv(rbuf, count, dtype, rank - 1,
MCA_COLL_BASE_TAG_EXSCAN, comm, MPI_STATUS_IGNORE));
if (MPI_SUCCESS != err) {
goto error;
}
/* Now reduce the prior rank's result with my source buffer. The source
* buffer had been previously copied into the temporary reduce_buffer. */
ompi_op_reduce(op, rbuf, reduce_buffer, count, dtype);
/* Send my result off to the next rank */
err = MCA_PML_CALL(send(reduce_buffer, count, dtype, rank + 1,
MCA_COLL_BASE_TAG_EXSCAN,
MCA_PML_BASE_SEND_STANDARD, comm));
/* Error */
error:
free(free_buffer);
/* All done */
return err;
}
/*
* ompi_coll_base_exscan_intra_recursivedoubling
*
* Function: Recursive doubling algorithm for exclusive scan.
* Accepts: Same as MPI_Exscan
* Returns: MPI_SUCCESS or error code
*
* Description: Implements recursive doubling algorithm for MPI_Exscan.
* The algorithm preserves order of operations so it can
* be used both by commutative and non-commutative operations.
*
* Example for 5 processes and commutative operation MPI_SUM:
* Process: 0 1 2 3 4
* recvbuf: - - - - -
* psend: [0] [1] [2] [3] [4]
*
* Step 1:
* recvbuf: - [0] - [2] -
* psend: [1+0] [0+1] [3+2] [2+3] [4]
*
* Step 2:
* recvbuf: - [0] [1+0] [(0+1)+2] -
* psend: [(3+2)+(1+0)] [(2+3)+(0+1)] [(1+0)+(3+2)] [(1+0)+(2+3)] [4]
*
* Step 3:
* recvbuf: - [0] [1+0] [(0+1)+2] [(3+2)+(1+0)]
* psend: [4+((3+2)+(1+0))] [((3+2)+(1+0))+4]
*
* Time complexity (worst case): \ceil(\log_2(p))(2\alpha + 2m\beta + 2m\gamma)
* Memory requirements (per process): 2 * count * typesize = O(count)
* Limitations: intra-communicators only
*/
int ompi_coll_base_exscan_intra_recursivedoubling(
const void *sendbuf, void *recvbuf, int count, struct ompi_datatype_t *datatype,
struct ompi_op_t *op, struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
int err = MPI_SUCCESS;
char *tmpsend_raw = NULL, *tmprecv_raw = NULL;
int comm_size = ompi_comm_size(comm);
int rank = ompi_comm_rank(comm);
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, "coll:base:exscan_intra_recursivedoubling: rank %d/%d",
rank, comm_size));
if (count == 0)
return MPI_SUCCESS;
if (comm_size < 2)
return MPI_SUCCESS;
ptrdiff_t dsize, gap;
dsize = opal_datatype_span(&datatype->super, count, &gap);
tmpsend_raw = malloc(dsize);
tmprecv_raw = malloc(dsize);
if (NULL == tmpsend_raw || NULL == tmprecv_raw) {
err = OMPI_ERR_OUT_OF_RESOURCE;
goto cleanup_and_return;
}
char *psend = tmpsend_raw - gap;
char *precv = tmprecv_raw - gap;
if (sendbuf != MPI_IN_PLACE) {
err = ompi_datatype_copy_content_same_ddt(datatype, count, psend, (char *)sendbuf);
if (MPI_SUCCESS != err) { goto cleanup_and_return; }
} else {
err = ompi_datatype_copy_content_same_ddt(datatype, count, psend, recvbuf);
if (MPI_SUCCESS != err) { goto cleanup_and_return; }
}
int is_commute = ompi_op_is_commute(op);
int is_first_block = 1;
for (int mask = 1; mask < comm_size; mask <<= 1) {
int remote = rank ^ mask;
if (remote < comm_size) {
err = ompi_coll_base_sendrecv(psend, count, datatype, remote,
MCA_COLL_BASE_TAG_EXSCAN,
precv, count, datatype, remote,
MCA_COLL_BASE_TAG_EXSCAN, comm,
MPI_STATUS_IGNORE, rank);
if (MPI_SUCCESS != err) { goto cleanup_and_return; }
if (rank > remote) {
/* Assertion: rank > 0 and rbuf is valid */
if (is_first_block) {
err = ompi_datatype_copy_content_same_ddt(datatype, count,
recvbuf, precv);
if (MPI_SUCCESS != err) { goto cleanup_and_return; }
is_first_block = 0;
} else {
/* Accumulate prefix reduction: recvbuf = precv <op> recvbuf */
ompi_op_reduce(op, precv, recvbuf, count, datatype);
}
/* Partial result: psend = precv <op> psend */
ompi_op_reduce(op, precv, psend, count, datatype);
} else {
if (is_commute) {
/* psend = precv <op> psend */
ompi_op_reduce(op, precv, psend, count, datatype);
} else {
/* precv = psend <op> precv */
ompi_op_reduce(op, psend, precv, count, datatype);
char *tmp = psend;
psend = precv;
precv = tmp;
}
}
}
}
cleanup_and_return:
if (NULL != tmpsend_raw)
free(tmpsend_raw);
if (NULL != tmprecv_raw)
free(tmprecv_raw);
return err;
}
|
