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
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
|
/* Copyright (C) 2003-2007 CAMP
* Copyright (C) 2007-2009 CAMd
* Copyright (C) 2005 CSC - IT Center for Science Ltd.
* Please see the accompanying LICENSE file for further information. */
// Copyright (C) 2003 CAMP
// Please see the accompanying LICENSE file for further information.
#include "extensions.h"
#include "bc.h"
#include <string.h>
#include <assert.h>
#include <stdio.h>
#include <stdlib.h>
boundary_conditions* bc_init(const long size1[3],
const long padding[3][2],
const long npadding[3][2],
const long neighbors[3][2],
MPI_Comm comm, bool real, bool cfd)
{
boundary_conditions* bc = GPAW_MALLOC(boundary_conditions, 1);
for (int i = 0; i < 3; i++)
{
bc->size1[i] = size1[i];
bc->size2[i] = size1[i] + padding[i][0] + padding[i][1];
bc->padding[i] = padding[i][0];
}
bc->comm = comm;
bc->ndouble = (real ? 1 : 2);
bc->cfd = cfd;
int rank = 0;
if (comm != MPI_COMM_NULL)
MPI_Comm_rank(comm, &rank);
int start[3];
int size[3];
for (int i = 0; i < 3; i++)
{
start[i] = padding[i][0];
size[i] = size1[i];
}
for (int i = 0; i < 3; i++)
{
int n = bc->ndouble;
for (int j = 0; j < 3; j++)
if (j != i)
n *= size[j];
for (int d = 0; d < 2; d++)
{
int ds = npadding[i][d];
int dr = padding[i][d];
for (int j = 0; j < 3; j++)
{
bc->sendstart[i][d][j] = start[j];
bc->sendsize[i][d][j] = size[j];
bc->recvstart[i][d][j] = start[j];
bc->recvsize[i][d][j] = size[j];
}
if (d == 0)
{
bc->sendstart[i][d][i] = dr;
bc->recvstart[i][d][i] = 0;
}
else
{
bc->sendstart[i][d][i] = padding[i][0] + size1[i] - ds;
bc->recvstart[i][d][i] = padding[i][0] + size1[i];
}
bc->sendsize[i][d][i] = ds;
bc->recvsize[i][d][i] = dr;
bc->sendproc[i][d] = DO_NOTHING;
bc->recvproc[i][d] = DO_NOTHING;
bc->nsend[i][d] = 0;
bc->nrecv[i][d] = 0;
int p = neighbors[i][d];
if (p == rank)
{
if (ds > 0)
bc->sendproc[i][d] = COPY_DATA;
if (dr > 0)
bc->recvproc[i][d] = COPY_DATA;
}
else if (p >= 0)
{
// Communication required:
if (ds > 0)
{
bc->sendproc[i][d] = p;
bc->nsend[i][d] = n * ds;
}
if (dr > 0)
{
bc->recvproc[i][d] = p;
bc->nrecv[i][d] = n * dr;
}
}
}
if (cfd == 0)
{
start[i] = 0;
size[i] = bc->size2[i];
}
// If the two neighboring processors along the
// i'th axis are the same, then we join the two communications
// into one:
#ifndef GPAW_GPU
bc->rjoin[i] = ((bc->recvproc[i][0] == bc->recvproc[i][1]) &&
bc->recvproc[i][0] >= 0);
bc->sjoin[i] = ((bc->sendproc[i][0] == bc->sendproc[i][1]) &&
bc->sendproc[i][0] >= 0);
#else
bc->rjoin[i] = 0;
bc->sjoin[i] = 0;
#endif
}
bc->maxsend = 0;
bc->maxrecv = 0;
for (int i = 0; i < 3; i++)
{
int n = bc->nsend[i][0] + bc->nsend[i][1];
if (n > bc->maxsend)
bc->maxsend = n;
n = bc->nrecv[i][0] + bc->nrecv[i][1];
if (n > bc->maxrecv)
bc->maxrecv = n;
}
#ifdef GPAW_GPU
bc_init_gpu(bc);
#endif
return bc;
}
void bc_unpack1(const boundary_conditions* bc,
const double* aa1, double* aa2, int i,
MPI_Request recvreq[2],
MPI_Request sendreq[2],
double* rbuff, double* sbuff,
const double_complex phases[2], int thd, int nin)
{
int ng = bc->ndouble * bc->size1[0] * bc->size1[1] * bc->size1[2];
int ng2 = bc->ndouble * bc->size2[0] * bc->size2[1] * bc->size2[2];
bool real = (bc->ndouble == 1);
for (int m = 0; m < nin; m++)
// Copy data:
if (i == 0)
{
// Zero all of a2 array. We should only zero the bounaries
// that are not periodic, but it's simpler to zero everything!
// XXX
memset(aa2 + m * ng2, 0, ng2 * sizeof(double));
// Copy data from a1 to central part of a2:
if (real)
bmgs_paste(aa1 + m * ng, bc->size1, aa2 + m * ng2,
bc->size2, bc->sendstart[0][0]);
else
bmgs_pastez((const double_complex*)(aa1 + m * ng), bc->size1,
(double_complex*)(aa2 + m * ng2),
bc->size2, bc->sendstart[0][0]);
}
#ifdef PARALLEL
// Start receiving.
for (int d = 0; d < 2; d++)
{
int p = bc->recvproc[i][d];
if (p >= 0)
{
if (bc->rjoin[i])
{
if (d == 0)
MPI_Irecv(rbuff, (bc->nrecv[i][0] + bc->nrecv[i][1]) * nin,
MPI_DOUBLE, p,
10 * thd + 1000 * i + 100000,
bc->comm, &recvreq[0]);
}
else
{
MPI_Irecv(rbuff, bc->nrecv[i][d] * nin, MPI_DOUBLE, p,
d + 10 * thd + 1000 * i,
bc->comm, &recvreq[d]);
rbuff += bc->nrecv[i][d] * nin;
}
}
}
// Prepare send-buffers and start sending:
double* sbuf = sbuff;
double* sbuf0 = sbuff;
for (int d = 0; d < 2; d++)
{
sendreq[d] = 0;
int p = bc->sendproc[i][d];
if (p >= 0)
{
const int* start = bc->sendstart[i][d];
const int* size = bc->sendsize[i][d];
for (int m = 0; m < nin; m++)
if (real)
bmgs_cut(aa2 + m * ng2, bc->size2, start,
sbuf + m * bc->nsend[i][d],
size);
else
bmgs_cutmz((const double_complex*)(aa2 + m * ng2),
bc->size2, start,
(double_complex*)(sbuf + m * bc->nsend[i][d]),
size, phases[d]);
if (bc->sjoin[i])
{
if (d == 1)
{
MPI_Isend(sbuf0, (bc->nsend[i][0] + bc->nsend[i][1]) * nin,
MPI_DOUBLE, p,
10 * thd + 1000 * i + 100000,
bc->comm, &sendreq[0]);
}
}
else
{
MPI_Isend(sbuf, bc->nsend[i][d] * nin, MPI_DOUBLE, p,
1 - d + 10 * thd + 1000 * i, bc->comm, &sendreq[d]);
}
sbuf += bc->nsend[i][d] * nin;
}
}
#endif // Parallel
for (int m = 0; m < nin; m++)
{
// Copy data for periodic boundary conditions:
for (int d = 0; d < 2; d++)
if (bc->sendproc[i][d] == COPY_DATA)
{
if (real)
bmgs_translate(aa2 + m * ng2, bc->size2, bc->sendsize[i][d],
bc->sendstart[i][d], bc->recvstart[i][1 - d]);
else
bmgs_translatemz((double_complex*)(aa2 + m * ng2), bc->size2,
bc->sendsize[i][d],
bc->sendstart[i][d], bc->recvstart[i][1 - d],
phases[d]);
}
}
}
void bc_unpack2(const boundary_conditions* bc,
double* a2, int i,
MPI_Request recvreq[2],
MPI_Request sendreq[2],
double* rbuf, int nin)
{
#ifdef PARALLEL
int ng2 = bc->ndouble * bc->size2[0] * bc->size2[1] * bc->size2[2];
// Store data from receive-buffer:
bool real = (bc->ndouble == 1);
double* rbuf0 = rbuf;
for (int d = 0; d < 2; d++)
if (bc->recvproc[i][d] >= 0)
{
if (bc->rjoin[i])
{
if (d == 0)
{
MPI_Wait(&recvreq[0], MPI_STATUS_IGNORE);
rbuf += bc->nrecv[i][1] * nin;
}
else
rbuf = rbuf0;
}
else
MPI_Wait(&recvreq[d], MPI_STATUS_IGNORE);
for (int m = 0; m < nin; m++)
if (real)
bmgs_paste(rbuf + m * bc->nrecv[i][d], bc->recvsize[i][d],
a2 + m * ng2, bc->size2, bc->recvstart[i][d]);
else
bmgs_pastez((const double_complex*)(rbuf +
m * bc->nrecv[i][d]),
bc->recvsize[i][d],
(double_complex*)(a2 + m * ng2),
bc->size2, bc->recvstart[i][d]);
rbuf += bc->nrecv[i][d] * nin;
}
// This does not work on the ibm with gcc! We do a blocking send instead.
for (int d = 0; d < 2; d++)
if (sendreq[d] != 0)
MPI_Wait(&sendreq[d], MPI_STATUS_IGNORE);
#endif // PARALLEL
}
|
