1 /*
2 * Copyright (c) 2003, 2007-14 Matteo Frigo
3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
4 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
9 *
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
14 *
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
18 *
19 */
20
21 #include "mpi-rdft.h"
22
destroy(problem * ego_)23 static void destroy(problem *ego_)
24 {
25 problem_mpi_rdft *ego = (problem_mpi_rdft *) ego_;
26 XM(dtensor_destroy)(ego->sz);
27 MPI_Comm_free(&ego->comm);
28 #if !defined(STRUCT_HACK_C99) && !defined(STRUCT_HACK_KR)
29 X(ifree0)(ego->kind);
30 #endif
31 X(ifree)(ego_);
32 }
33
hash(const problem * p_,md5 * m)34 static void hash(const problem *p_, md5 *m)
35 {
36 const problem_mpi_rdft *p = (const problem_mpi_rdft *) p_;
37 int i;
38 X(md5puts)(m, "mpi-dft");
39 X(md5int)(m, p->I == p->O);
40 /* don't include alignment -- may differ between processes
41 X(md5int)(m, X(ialignment_of)(p->I));
42 X(md5int)(m, X(ialignment_of)(p->O));
43 ... note that applicability of MPI plans does not depend
44 on alignment (although optimality may, in principle). */
45 XM(dtensor_md5)(m, p->sz);
46 X(md5INT)(m, p->vn);
47 for (i = 0; i < p->sz->rnk; ++i)
48 X(md5int)(m, p->kind[i]);
49 X(md5int)(m, p->flags);
50 MPI_Comm_size(p->comm, &i); X(md5int)(m, i);
51 A(XM(md5_equal)(*m, p->comm));
52 }
53
print(const problem * ego_,printer * p)54 static void print(const problem *ego_, printer *p)
55 {
56 const problem_mpi_rdft *ego = (const problem_mpi_rdft *) ego_;
57 int i;
58 p->print(p, "(mpi-rdft %d %d %d ",
59 ego->I == ego->O,
60 X(ialignment_of)(ego->I),
61 X(ialignment_of)(ego->O));
62 XM(dtensor_print)(ego->sz, p);
63 for (i = 0; i < ego->sz->rnk; ++i)
64 p->print(p, " %d", (int)ego->kind[i]);
65 p->print(p, " %D %d", ego->vn, ego->flags);
66 MPI_Comm_size(ego->comm, &i); p->print(p, " %d)", i);
67 }
68
zero(const problem * ego_)69 static void zero(const problem *ego_)
70 {
71 const problem_mpi_rdft *ego = (const problem_mpi_rdft *) ego_;
72 R *I = ego->I;
73 INT i, N;
74 int my_pe;
75
76 MPI_Comm_rank(ego->comm, &my_pe);
77 N = ego->vn * XM(total_block)(ego->sz, IB, my_pe);
78 for (i = 0; i < N; ++i) I[i] = K(0.0);
79 }
80
81 static const problem_adt padt =
82 {
83 PROBLEM_MPI_RDFT,
84 hash,
85 zero,
86 print,
87 destroy
88 };
89
XM(mkproblem_rdft)90 problem *XM(mkproblem_rdft)(const dtensor *sz, INT vn,
91 R *I, R *O,
92 MPI_Comm comm,
93 const rdft_kind *kind, unsigned flags)
94 {
95 problem_mpi_rdft *ego;
96 int i, rnk = sz->rnk;
97 int n_pes;
98
99 A(XM(dtensor_validp)(sz) && FINITE_RNK(sz->rnk));
100 MPI_Comm_size(comm, &n_pes);
101 A(n_pes >= XM(num_blocks_total)(sz, IB)
102 && n_pes >= XM(num_blocks_total)(sz, OB));
103 A(vn >= 0);
104
105 #if defined(STRUCT_HACK_KR)
106 ego = (problem_mpi_rdft *) X(mkproblem)(sizeof(problem_mpi_rdft)
107 + sizeof(rdft_kind)
108 * (rnk > 0 ? rnk - 1 : 0), &padt);
109 #elif defined(STRUCT_HACK_C99)
110 ego = (problem_mpi_rdft *) X(mkproblem)(sizeof(problem_mpi_rdft)
111 + sizeof(rdft_kind) * rnk, &padt);
112 #else
113 ego = (problem_mpi_rdft *) X(mkproblem)(sizeof(problem_mpi_rdft), &padt);
114 ego->kind = (rdft_kind *) MALLOC(sizeof(rdft_kind) * rnk, PROBLEMS);
115 #endif
116
117 /* enforce pointer equality if untainted pointers are equal */
118 if (UNTAINT(I) == UNTAINT(O))
119 I = O = JOIN_TAINT(I, O);
120
121 ego->sz = XM(dtensor_canonical)(sz, 0);
122 ego->vn = vn;
123 ego->I = I;
124 ego->O = O;
125 for (i = 0; i< ego->sz->rnk; ++i)
126 ego->kind[i] = kind[i];
127
128 /* canonicalize: replace TRANSPOSED_IN with TRANSPOSED_OUT by
129 swapping the first two dimensions (for rnk > 1) */
130 if ((flags & TRANSPOSED_IN) && ego->sz->rnk > 1) {
131 rdft_kind k = ego->kind[0];
132 ddim dim0 = ego->sz->dims[0];
133 ego->sz->dims[0] = ego->sz->dims[1];
134 ego->sz->dims[1] = dim0;
135 ego->kind[0] = ego->kind[1];
136 ego->kind[1] = k;
137 flags &= ~TRANSPOSED_IN;
138 flags ^= TRANSPOSED_OUT;
139 }
140 ego->flags = flags;
141
142 MPI_Comm_dup(comm, &ego->comm);
143
144 return &(ego->super);
145 }
146
XM(mkproblem_rdft_d)147 problem *XM(mkproblem_rdft_d)(dtensor *sz, INT vn,
148 R *I, R *O,
149 MPI_Comm comm,
150 const rdft_kind *kind, unsigned flags)
151 {
152 problem *p = XM(mkproblem_rdft)(sz, vn, I, O, comm, kind, flags);
153 XM(dtensor_destroy)(sz);
154 return p;
155 }
156