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
|
/*
* 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: rodft00e-r2hc.c,v 1.1 2008/10/17 06:13:18 scuri Exp $ */
/* Do a RODFT00 problem via an R2HC problem, with some pre/post-processing.
This code uses the trick from FFTPACK, also documented in a similar
form by Numerical Recipes. Unfortunately, this algorithm seems to
have intrinsic numerical problems (similar to those in
reodft11e-r2hc.c), possibly due to the fact that it multiplies its
input by a sine, causing a loss of precision near the zero. For
transforms of 16k points, it has already lost three or four decimal
places of accuracy, which we deem unacceptable.
So, we have abandoned this algorithm in favor of the one in
rodft00-r2hc-pad.c, which unfortunately sacrifices 30-50% in speed.
The only other alternative in the literature that does not have
similar numerical difficulties seems to be the direct adaptation of
the Cooley-Tukey decomposition for antisymmetric data, but this
would require a whole new set of codelets and it's not clear that
it's worth it at this point. */
#include "reodft.h"
typedef struct {
solver super;
} S;
typedef struct {
plan_rdft super;
plan *cld;
twid *td;
int is, os;
int n;
int vl;
int ivs, ovs;
} P;
static void apply(const plan *ego_, R *I, R *O)
{
const P *ego = (const P *) ego_;
int is = ego->is, os = ego->os;
int i, n = ego->n;
int iv, vl = ego->vl;
int ivs = ego->ivs, ovs = ego->ovs;
R *W = ego->td->W;
R *buf;
buf = (R *) MALLOC(sizeof(R) * n, BUFFERS);
for (iv = 0; iv < vl; ++iv, I += ivs, O += ovs) {
buf[0] = 0;
for (i = 1; i < n - i; ++i) {
E a, b, apb, amb;
a = I[is * (i - 1)];
b = I[is * ((n - i) - 1)];
apb = K(2.0) * W[i] * (a + b);
amb = (a - b);
buf[i] = apb + amb;
buf[n - i] = apb - amb;
}
if (i == n - i) {
buf[i] = K(4.0) * I[is * (i - 1)];
}
{
plan_rdft *cld = (plan_rdft *) ego->cld;
cld->apply((plan *) cld, buf, buf);
}
/* FIXME: use recursive/cascade summation for better stability? */
O[0] = buf[0] * 0.5;
for (i = 1; i + i < n - 1; ++i) {
int k = i + i;
O[os * (k - 1)] = -buf[n - i];
O[os * k] = O[os * (k - 2)] + buf[i];
}
if (i + i == n - 1) {
O[os * (n - 2)] = -buf[n - i];
}
}
X(ifree)(buf);
}
static void awake(plan *ego_, int flg)
{
P *ego = (P *) ego_;
static const tw_instr rodft00e_tw[] = {
{ TW_SIN, 0, 1 },
{ TW_NEXT, 1, 0 }
};
AWAKE(ego->cld, flg);
X(twiddle_awake)(flg, &ego->td, rodft00e_tw, 2*ego->n, 1, (ego->n+1)/2);
}
static void destroy(plan *ego_)
{
P *ego = (P *) ego_;
X(plan_destroy_internal)(ego->cld);
}
static void print(const plan *ego_, printer *p)
{
const P *ego = (const P *) ego_;
p->print(p, "(rodft00e-r2hc-%d%v%(%p%))", ego->n - 1, ego->vl, ego->cld);
}
static int applicable0(const solver *ego_, const problem *p_)
{
UNUSED(ego_);
if (RDFTP(p_)) {
const problem_rdft *p = (const problem_rdft *) p_;
return (1
&& p->sz->rnk == 1
&& p->vecsz->rnk <= 1
&& p->kind[0] == RODFT00
);
}
return 0;
}
static int applicable(const solver *ego, const problem *p, const planner *plnr)
{
return (!NO_UGLYP(plnr) && applicable0(ego, p));
}
static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
{
P *pln;
const problem_rdft *p;
plan *cld;
R *buf;
int n;
opcnt ops;
static const plan_adt padt = {
X(rdft_solve), awake, print, destroy
};
if (!applicable(ego_, p_, plnr))
return (plan *)0;
p = (const problem_rdft *) p_;
n = p->sz->dims[0].n + 1;
buf = (R *) MALLOC(sizeof(R) * n, BUFFERS);
cld = X(mkplan_d)(plnr, X(mkproblem_rdft_1_d)(X(mktensor_1d)(n, 1, 1),
X(mktensor_0d)(),
buf, buf, R2HC));
X(ifree)(buf);
if (!cld)
return (plan *)0;
pln = MKPLAN_RDFT(P, &padt, apply);
pln->n = n;
pln->is = p->sz->dims[0].is;
pln->os = p->sz->dims[0].os;
pln->cld = cld;
pln->td = 0;
X(tensor_tornk1)(p->vecsz, &pln->vl, &pln->ivs, &pln->ovs);
X(ops_zero)(&ops);
ops.other = 4 + (n-1)/2 * 5 + (n-2)/2 * 5;
ops.add = (n-1)/2 * 4 + (n-2)/2 * 1;
ops.mul = 1 + (n-1)/2 * 2;
if (n % 2 == 0)
ops.mul += 1;
X(ops_zero)(&pln->super.super.ops);
X(ops_madd2)(pln->vl, &ops, &pln->super.super.ops);
X(ops_madd2)(pln->vl, &cld->ops, &pln->super.super.ops);
return &(pln->super.super);
}
/* constructor */
static solver *mksolver(void)
{
static const solver_adt sadt = { mkplan };
S *slv = MKSOLVER(S, &sadt);
return &(slv->super);
}
void X(rodft00e_r2hc_register)(planner *p)
{
REGISTER_SOLVER(p, mksolver());
}
|
