Chris@42
|
1 /*
|
Chris@42
|
2 * Copyright (c) 2003, 2007-14 Matteo Frigo
|
Chris@42
|
3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
|
Chris@42
|
4 *
|
Chris@42
|
5 * This program is free software; you can redistribute it and/or modify
|
Chris@42
|
6 * it under the terms of the GNU General Public License as published by
|
Chris@42
|
7 * the Free Software Foundation; either version 2 of the License, or
|
Chris@42
|
8 * (at your option) any later version.
|
Chris@42
|
9 *
|
Chris@42
|
10 * This program is distributed in the hope that it will be useful,
|
Chris@42
|
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
Chris@42
|
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
Chris@42
|
13 * GNU General Public License for more details.
|
Chris@42
|
14 *
|
Chris@42
|
15 * You should have received a copy of the GNU General Public License
|
Chris@42
|
16 * along with this program; if not, write to the Free Software
|
Chris@42
|
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
Chris@42
|
18 *
|
Chris@42
|
19 */
|
Chris@42
|
20
|
Chris@42
|
21 #include "ct-hc2c.h"
|
Chris@42
|
22 #include "dft.h"
|
Chris@42
|
23
|
Chris@42
|
24 typedef struct {
|
Chris@42
|
25 plan_rdft2 super;
|
Chris@42
|
26 plan *cld;
|
Chris@42
|
27 plan *cldw;
|
Chris@42
|
28 INT r;
|
Chris@42
|
29 } P;
|
Chris@42
|
30
|
Chris@42
|
31 static void apply_dit(const plan *ego_, R *r0, R *r1, R *cr, R *ci)
|
Chris@42
|
32 {
|
Chris@42
|
33 const P *ego = (const P *) ego_;
|
Chris@42
|
34 plan_rdft *cld;
|
Chris@42
|
35 plan_hc2c *cldw;
|
Chris@42
|
36 UNUSED(r1);
|
Chris@42
|
37
|
Chris@42
|
38 cld = (plan_rdft *) ego->cld;
|
Chris@42
|
39 cld->apply(ego->cld, r0, cr);
|
Chris@42
|
40
|
Chris@42
|
41 cldw = (plan_hc2c *) ego->cldw;
|
Chris@42
|
42 cldw->apply(ego->cldw, cr, ci);
|
Chris@42
|
43 }
|
Chris@42
|
44
|
Chris@42
|
45 static void apply_dif(const plan *ego_, R *r0, R *r1, R *cr, R *ci)
|
Chris@42
|
46 {
|
Chris@42
|
47 const P *ego = (const P *) ego_;
|
Chris@42
|
48 plan_rdft *cld;
|
Chris@42
|
49 plan_hc2c *cldw;
|
Chris@42
|
50 UNUSED(r1);
|
Chris@42
|
51
|
Chris@42
|
52 cldw = (plan_hc2c *) ego->cldw;
|
Chris@42
|
53 cldw->apply(ego->cldw, cr, ci);
|
Chris@42
|
54
|
Chris@42
|
55 cld = (plan_rdft *) ego->cld;
|
Chris@42
|
56 cld->apply(ego->cld, cr, r0);
|
Chris@42
|
57 }
|
Chris@42
|
58
|
Chris@42
|
59 static void apply_dit_dft(const plan *ego_, R *r0, R *r1, R *cr, R *ci)
|
Chris@42
|
60 {
|
Chris@42
|
61 const P *ego = (const P *) ego_;
|
Chris@42
|
62 plan_dft *cld;
|
Chris@42
|
63 plan_hc2c *cldw;
|
Chris@42
|
64
|
Chris@42
|
65 cld = (plan_dft *) ego->cld;
|
Chris@42
|
66 cld->apply(ego->cld, r0, r1, cr, ci);
|
Chris@42
|
67
|
Chris@42
|
68 cldw = (plan_hc2c *) ego->cldw;
|
Chris@42
|
69 cldw->apply(ego->cldw, cr, ci);
|
Chris@42
|
70 }
|
Chris@42
|
71
|
Chris@42
|
72 static void apply_dif_dft(const plan *ego_, R *r0, R *r1, R *cr, R *ci)
|
Chris@42
|
73 {
|
Chris@42
|
74 const P *ego = (const P *) ego_;
|
Chris@42
|
75 plan_dft *cld;
|
Chris@42
|
76 plan_hc2c *cldw;
|
Chris@42
|
77
|
Chris@42
|
78 cldw = (plan_hc2c *) ego->cldw;
|
Chris@42
|
79 cldw->apply(ego->cldw, cr, ci);
|
Chris@42
|
80
|
Chris@42
|
81 cld = (plan_dft *) ego->cld;
|
Chris@42
|
82 cld->apply(ego->cld, ci, cr, r1, r0);
|
Chris@42
|
83 }
|
Chris@42
|
84
|
Chris@42
|
85 static void awake(plan *ego_, enum wakefulness wakefulness)
|
Chris@42
|
86 {
|
Chris@42
|
87 P *ego = (P *) ego_;
|
Chris@42
|
88 X(plan_awake)(ego->cld, wakefulness);
|
Chris@42
|
89 X(plan_awake)(ego->cldw, wakefulness);
|
Chris@42
|
90 }
|
Chris@42
|
91
|
Chris@42
|
92 static void destroy(plan *ego_)
|
Chris@42
|
93 {
|
Chris@42
|
94 P *ego = (P *) ego_;
|
Chris@42
|
95 X(plan_destroy_internal)(ego->cldw);
|
Chris@42
|
96 X(plan_destroy_internal)(ego->cld);
|
Chris@42
|
97 }
|
Chris@42
|
98
|
Chris@42
|
99 static void print(const plan *ego_, printer *p)
|
Chris@42
|
100 {
|
Chris@42
|
101 const P *ego = (const P *) ego_;
|
Chris@42
|
102 p->print(p, "(rdft2-ct-%s/%D%(%p%)%(%p%))",
|
Chris@42
|
103 (ego->super.apply == apply_dit ||
|
Chris@42
|
104 ego->super.apply == apply_dit_dft)
|
Chris@42
|
105 ? "dit" : "dif",
|
Chris@42
|
106 ego->r, ego->cldw, ego->cld);
|
Chris@42
|
107 }
|
Chris@42
|
108
|
Chris@42
|
109 static int applicable0(const hc2c_solver *ego, const problem *p_, planner *plnr)
|
Chris@42
|
110 {
|
Chris@42
|
111 const problem_rdft2 *p = (const problem_rdft2 *) p_;
|
Chris@42
|
112 INT r;
|
Chris@42
|
113
|
Chris@42
|
114 return (1
|
Chris@42
|
115 && p->sz->rnk == 1
|
Chris@42
|
116 && p->vecsz->rnk <= 1
|
Chris@42
|
117
|
Chris@42
|
118 && (/* either the problem is R2HC, which is solved by DIT */
|
Chris@42
|
119 (p->kind == R2HC)
|
Chris@42
|
120 ||
|
Chris@42
|
121 /* or the problem is HC2R, in which case it is solved
|
Chris@42
|
122 by DIF, which destroys the input */
|
Chris@42
|
123 (p->kind == HC2R &&
|
Chris@42
|
124 (p->r0 == p->cr || !NO_DESTROY_INPUTP(plnr))))
|
Chris@42
|
125
|
Chris@42
|
126 && ((r = X(choose_radix)(ego->r, p->sz->dims[0].n)) > 0)
|
Chris@42
|
127 && p->sz->dims[0].n > r);
|
Chris@42
|
128 }
|
Chris@42
|
129
|
Chris@42
|
130 static int hc2c_applicable(const hc2c_solver *ego, const problem *p_,
|
Chris@42
|
131 planner *plnr)
|
Chris@42
|
132 {
|
Chris@42
|
133 const problem_rdft2 *p;
|
Chris@42
|
134
|
Chris@42
|
135 if (!applicable0(ego, p_, plnr))
|
Chris@42
|
136 return 0;
|
Chris@42
|
137
|
Chris@42
|
138 p = (const problem_rdft2 *) p_;
|
Chris@42
|
139
|
Chris@42
|
140 return (0
|
Chris@42
|
141 || p->vecsz->rnk == 0
|
Chris@42
|
142 || !NO_VRECURSEP(plnr)
|
Chris@42
|
143 );
|
Chris@42
|
144 }
|
Chris@42
|
145
|
Chris@42
|
146 static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
|
Chris@42
|
147 {
|
Chris@42
|
148 const hc2c_solver *ego = (const hc2c_solver *) ego_;
|
Chris@42
|
149 const problem_rdft2 *p;
|
Chris@42
|
150 P *pln = 0;
|
Chris@42
|
151 plan *cld = 0, *cldw = 0;
|
Chris@42
|
152 INT n, r, m, v, ivs, ovs;
|
Chris@42
|
153 iodim *d;
|
Chris@42
|
154
|
Chris@42
|
155 static const plan_adt padt = {
|
Chris@42
|
156 X(rdft2_solve), awake, print, destroy
|
Chris@42
|
157 };
|
Chris@42
|
158
|
Chris@42
|
159 if (!hc2c_applicable(ego, p_, plnr))
|
Chris@42
|
160 return (plan *) 0;
|
Chris@42
|
161
|
Chris@42
|
162 p = (const problem_rdft2 *) p_;
|
Chris@42
|
163 d = p->sz->dims;
|
Chris@42
|
164 n = d[0].n;
|
Chris@42
|
165 r = X(choose_radix)(ego->r, n);
|
Chris@42
|
166 A((r % 2) == 0);
|
Chris@42
|
167 m = n / r;
|
Chris@42
|
168
|
Chris@42
|
169 X(tensor_tornk1)(p->vecsz, &v, &ivs, &ovs);
|
Chris@42
|
170
|
Chris@42
|
171 switch (p->kind) {
|
Chris@42
|
172 case R2HC:
|
Chris@42
|
173 cldw = ego->mkcldw(ego, R2HC,
|
Chris@42
|
174 r, m * d[0].os,
|
Chris@42
|
175 m, d[0].os,
|
Chris@42
|
176 v, ovs,
|
Chris@42
|
177 p->cr, p->ci, plnr);
|
Chris@42
|
178 if (!cldw) goto nada;
|
Chris@42
|
179
|
Chris@42
|
180 switch (ego->hc2ckind) {
|
Chris@42
|
181 case HC2C_VIA_RDFT:
|
Chris@42
|
182 cld = X(mkplan_d)(
|
Chris@42
|
183 plnr,
|
Chris@42
|
184 X(mkproblem_rdft_1_d)(
|
Chris@42
|
185 X(mktensor_1d)(m, (r/2)*d[0].is, d[0].os),
|
Chris@42
|
186 X(mktensor_3d)(
|
Chris@42
|
187 2, p->r1 - p->r0, p->ci - p->cr,
|
Chris@42
|
188 r / 2, d[0].is, m * d[0].os,
|
Chris@42
|
189 v, ivs, ovs),
|
Chris@42
|
190 p->r0, p->cr, R2HC)
|
Chris@42
|
191 );
|
Chris@42
|
192 if (!cld) goto nada;
|
Chris@42
|
193
|
Chris@42
|
194 pln = MKPLAN_RDFT2(P, &padt, apply_dit);
|
Chris@42
|
195 break;
|
Chris@42
|
196
|
Chris@42
|
197 case HC2C_VIA_DFT:
|
Chris@42
|
198 cld = X(mkplan_d)(
|
Chris@42
|
199 plnr,
|
Chris@42
|
200 X(mkproblem_dft_d)(
|
Chris@42
|
201 X(mktensor_1d)(m, (r/2)*d[0].is, d[0].os),
|
Chris@42
|
202 X(mktensor_2d)(
|
Chris@42
|
203 r / 2, d[0].is, m * d[0].os,
|
Chris@42
|
204 v, ivs, ovs),
|
Chris@42
|
205 p->r0, p->r1, p->cr, p->ci)
|
Chris@42
|
206 );
|
Chris@42
|
207 if (!cld) goto nada;
|
Chris@42
|
208
|
Chris@42
|
209 pln = MKPLAN_RDFT2(P, &padt, apply_dit_dft);
|
Chris@42
|
210 break;
|
Chris@42
|
211 }
|
Chris@42
|
212 break;
|
Chris@42
|
213
|
Chris@42
|
214 case HC2R:
|
Chris@42
|
215 cldw = ego->mkcldw(ego, HC2R,
|
Chris@42
|
216 r, m * d[0].is,
|
Chris@42
|
217 m, d[0].is,
|
Chris@42
|
218 v, ivs,
|
Chris@42
|
219 p->cr, p->ci, plnr);
|
Chris@42
|
220 if (!cldw) goto nada;
|
Chris@42
|
221
|
Chris@42
|
222 switch (ego->hc2ckind) {
|
Chris@42
|
223 case HC2C_VIA_RDFT:
|
Chris@42
|
224 cld = X(mkplan_d)(
|
Chris@42
|
225 plnr,
|
Chris@42
|
226 X(mkproblem_rdft_1_d)(
|
Chris@42
|
227 X(mktensor_1d)(m, d[0].is, (r/2)*d[0].os),
|
Chris@42
|
228 X(mktensor_3d)(
|
Chris@42
|
229 2, p->ci - p->cr, p->r1 - p->r0,
|
Chris@42
|
230 r / 2, m * d[0].is, d[0].os,
|
Chris@42
|
231 v, ivs, ovs),
|
Chris@42
|
232 p->cr, p->r0, HC2R)
|
Chris@42
|
233 );
|
Chris@42
|
234 if (!cld) goto nada;
|
Chris@42
|
235
|
Chris@42
|
236 pln = MKPLAN_RDFT2(P, &padt, apply_dif);
|
Chris@42
|
237 break;
|
Chris@42
|
238
|
Chris@42
|
239 case HC2C_VIA_DFT:
|
Chris@42
|
240 cld = X(mkplan_d)(
|
Chris@42
|
241 plnr,
|
Chris@42
|
242 X(mkproblem_dft_d)(
|
Chris@42
|
243 X(mktensor_1d)(m, d[0].is, (r/2)*d[0].os),
|
Chris@42
|
244 X(mktensor_2d)(
|
Chris@42
|
245 r / 2, m * d[0].is, d[0].os,
|
Chris@42
|
246 v, ivs, ovs),
|
Chris@42
|
247 p->ci, p->cr, p->r1, p->r0)
|
Chris@42
|
248 );
|
Chris@42
|
249 if (!cld) goto nada;
|
Chris@42
|
250
|
Chris@42
|
251 pln = MKPLAN_RDFT2(P, &padt, apply_dif_dft);
|
Chris@42
|
252 break;
|
Chris@42
|
253 }
|
Chris@42
|
254 break;
|
Chris@42
|
255
|
Chris@42
|
256 default:
|
Chris@42
|
257 A(0);
|
Chris@42
|
258 }
|
Chris@42
|
259
|
Chris@42
|
260 pln->cld = cld;
|
Chris@42
|
261 pln->cldw = cldw;
|
Chris@42
|
262 pln->r = r;
|
Chris@42
|
263 X(ops_add)(&cld->ops, &cldw->ops, &pln->super.super.ops);
|
Chris@42
|
264
|
Chris@42
|
265 /* inherit could_prune_now_p attribute from cldw */
|
Chris@42
|
266 pln->super.super.could_prune_now_p = cldw->could_prune_now_p;
|
Chris@42
|
267
|
Chris@42
|
268 return &(pln->super.super);
|
Chris@42
|
269
|
Chris@42
|
270 nada:
|
Chris@42
|
271 X(plan_destroy_internal)(cldw);
|
Chris@42
|
272 X(plan_destroy_internal)(cld);
|
Chris@42
|
273 return (plan *) 0;
|
Chris@42
|
274 }
|
Chris@42
|
275
|
Chris@42
|
276 hc2c_solver *X(mksolver_hc2c)(size_t size, INT r,
|
Chris@42
|
277 hc2c_kind hc2ckind,
|
Chris@42
|
278 hc2c_mkinferior mkcldw)
|
Chris@42
|
279 {
|
Chris@42
|
280 static const solver_adt sadt = { PROBLEM_RDFT2, mkplan, 0 };
|
Chris@42
|
281 hc2c_solver *slv = (hc2c_solver *)X(mksolver)(size, &sadt);
|
Chris@42
|
282 slv->r = r;
|
Chris@42
|
283 slv->hc2ckind = hc2ckind;
|
Chris@42
|
284 slv->mkcldw = mkcldw;
|
Chris@42
|
285 return slv;
|
Chris@42
|
286 }
|
Chris@42
|
287
|
Chris@42
|
288 plan *X(mkplan_hc2c)(size_t size, const plan_adt *adt, hc2capply apply)
|
Chris@42
|
289 {
|
Chris@42
|
290 plan_hc2c *ego;
|
Chris@42
|
291
|
Chris@42
|
292 ego = (plan_hc2c *) X(mkplan)(size, adt);
|
Chris@42
|
293 ego->apply = apply;
|
Chris@42
|
294
|
Chris@42
|
295 return &(ego->super);
|
Chris@42
|
296 }
|