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
|
/*
* Copyright (c) 2003 Matteo Frigo
* Copyright (c) 2003 Massachusetts Institute of Technology
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2 of the License, or
* (at your option) any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, write to the Free Software
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
*
*/
/* $Id: rrank-geq2.c,v 1.1 2008/10/17 06:11:29 scuri Exp $ */
/* plans for RDFT of rank >= 2 (multidimensional) */
/* FIXME: this solver cannot strictly be applied to multidimensional
DHTs, since the latter are not separable...up to rnk-1 additional
post-processing passes may be required. See also:
R. N. Bracewell, O. Buneman, H. Hao, and J. Villasenor, "Fast
two-dimensional Hartley transform," Proc. IEEE 74, 1282-1283 (1986).
H. Hao and R. N. Bracewell, "A three-dimensional DFT algorithm
using the fast Hartley transform," Proc. IEEE 75(2), 264-266 (1987).
*/
#include "rdft.h"
typedef struct {
solver super;
int spltrnk;
const int *buddies;
int nbuddies;
} S;
typedef struct {
plan_rdft super;
plan *cld1, *cld2;
const S *solver;
} P;
/* Compute multi-dimensional RDFT by applying the two cld plans
(lower-rnk RDFTs). */
static void apply(const plan *ego_, R *I, R *O)
{
const P *ego = (const P *) ego_;
plan_rdft *cld1, *cld2;
cld1 = (plan_rdft *) ego->cld1;
cld1->apply(ego->cld1, I, O);
cld2 = (plan_rdft *) ego->cld2;
cld2->apply(ego->cld2, O, O);
}
static void awake(plan *ego_, int flg)
{
P *ego = (P *) ego_;
AWAKE(ego->cld1, flg);
AWAKE(ego->cld2, flg);
}
static void destroy(plan *ego_)
{
P *ego = (P *) ego_;
X(plan_destroy_internal)(ego->cld2);
X(plan_destroy_internal)(ego->cld1);
}
static void print(const plan *ego_, printer *p)
{
const P *ego = (const P *) ego_;
const S *s = ego->solver;
p->print(p, "(rdft-rank>=2/%d%(%p%)%(%p%))",
s->spltrnk, ego->cld1, ego->cld2);
}
static int picksplit(const S *ego, const tensor *sz, int *rp)
{
A(sz->rnk > 1); /* cannot split rnk <= 1 */
if (!X(pickdim)(ego->spltrnk, ego->buddies, ego->nbuddies, sz, 1, rp))
return 0;
*rp += 1; /* convert from dim. index to rank */
if (*rp >= sz->rnk) /* split must reduce rank */
return 0;
return 1;
}
static int applicable0(const solver *ego_, const problem *p_, int *rp)
{
if (RDFTP(p_)) {
const problem_rdft *p = (const problem_rdft *) p_;
const S *ego = (const S *)ego_;
return (1
&& p->sz->rnk >= 2
&& picksplit(ego, p->sz, rp)
);
}
return 0;
}
/* TODO: revise this. */
static int applicable(const solver *ego_, const problem *p_,
const planner *plnr, int *rp)
{
const S *ego = (const S *)ego_;
if (!applicable0(ego_, p_, rp)) return 0;
/* fixed spltrnk (unlike fftw2's spltrnk=1, default buddies[0] is
spltrnk=0, which is an asymptotic "theoretical optimum" for
an ideal cache; it's equivalent to spltrnk=1 for rnk < 4). */
if (NO_RANK_SPLITSP(plnr) && (ego->spltrnk != ego->buddies[0]))
return 0;
if (NO_UGLYP(plnr)) {
/* Heuristic: if the vector stride is greater than the transform
sz, don't use (prefer to do the vector loop first with a
vrank-geq1 plan). */
const problem_rdft *p = (const problem_rdft *) p_;
if (p->vecsz->rnk > 0 &&
X(tensor_min_stride)(p->vecsz) > X(tensor_max_index)(p->sz))
return 0;
}
return 1;
}
static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
{
const S *ego = (const S *) ego_;
const problem_rdft *p;
P *pln;
plan *cld1 = 0, *cld2 = 0;
tensor *sz1, *sz2, *vecszi, *sz2i;
int spltrnk;
static const plan_adt padt = {
X(rdft_solve), awake, print, destroy
};
if (!applicable(ego_, p_, plnr, &spltrnk))
return (plan *) 0;
p = (const problem_rdft *) p_;
X(tensor_split)(p->sz, &sz1, spltrnk, &sz2);
vecszi = X(tensor_copy_inplace)(p->vecsz, INPLACE_OS);
sz2i = X(tensor_copy_inplace)(sz2, INPLACE_OS);
cld1 = X(mkplan_d)(plnr,
X(mkproblem_rdft_d)(X(tensor_copy)(sz2),
X(tensor_append)(p->vecsz, sz1),
p->I, p->O, p->kind + spltrnk));
if (!cld1) goto nada;
cld2 = X(mkplan_d)(plnr,
X(mkproblem_rdft_d)(
X(tensor_copy_inplace)(sz1, INPLACE_OS),
X(tensor_append)(vecszi, sz2i),
p->O, p->O, p->kind));
if (!cld2) goto nada;
pln = MKPLAN_RDFT(P, &padt, apply);
pln->cld1 = cld1;
pln->cld2 = cld2;
pln->solver = ego;
X(ops_add)(&cld1->ops, &cld2->ops, &pln->super.super.ops);
X(tensor_destroy4)(sz2, sz1, vecszi, sz2i);
return &(pln->super.super);
nada:
X(plan_destroy_internal)(cld2);
X(plan_destroy_internal)(cld1);
X(tensor_destroy4)(sz2, sz1, vecszi, sz2i);
return (plan *) 0;
}
static solver *mksolver(int spltrnk, const int *buddies, int nbuddies)
{
static const solver_adt sadt = { mkplan };
S *slv = MKSOLVER(S, &sadt);
slv->spltrnk = spltrnk;
slv->buddies = buddies;
slv->nbuddies = nbuddies;
return &(slv->super);
}
void X(rdft_rank_geq2_register)(planner *p)
{
int i;
static const int buddies[] = { 0, 1, -2 };
const int nbuddies = sizeof(buddies) / sizeof(buddies[0]);
for (i = 0; i < nbuddies; ++i)
REGISTER_SOLVER(p, mksolver(buddies[i], buddies, nbuddies));
/* FIXME: Should we try more buddies? See also dft/rank-geq2. */
}
|
