annotate src/fftw-3.3.3/reodft/reodft11e-r2hc-odd.c @ 23:619f715526df sv_v2.1

Update Vamp plugin SDK to 2.5
author Chris Cannam
date Thu, 09 May 2013 10:52:46 +0100
parents 37bf6b4a2645
children
rev   line source
Chris@10 1 /*
Chris@10 2 * Copyright (c) 2003, 2007-11 Matteo Frigo
Chris@10 3 * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
Chris@10 4 *
Chris@10 5 * This program is free software; you can redistribute it and/or modify
Chris@10 6 * it under the terms of the GNU General Public License as published by
Chris@10 7 * the Free Software Foundation; either version 2 of the License, or
Chris@10 8 * (at your option) any later version.
Chris@10 9 *
Chris@10 10 * This program is distributed in the hope that it will be useful,
Chris@10 11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
Chris@10 12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
Chris@10 13 * GNU General Public License for more details.
Chris@10 14 *
Chris@10 15 * You should have received a copy of the GNU General Public License
Chris@10 16 * along with this program; if not, write to the Free Software
Chris@10 17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
Chris@10 18 *
Chris@10 19 */
Chris@10 20
Chris@10 21
Chris@10 22 /* Do an R{E,O}DFT11 problem via an R2HC problem of the same *odd* size,
Chris@10 23 with some permutations and post-processing, as described in:
Chris@10 24
Chris@10 25 S. C. Chan and K. L. Ho, "Fast algorithms for computing the
Chris@10 26 discrete cosine transform," IEEE Trans. Circuits Systems II:
Chris@10 27 Analog & Digital Sig. Proc. 39 (3), 185--190 (1992).
Chris@10 28
Chris@10 29 (For even sizes, see reodft11e-radix2.c.)
Chris@10 30
Chris@10 31 This algorithm is related to the 8 x n prime-factor-algorithm (PFA)
Chris@10 32 decomposition of the size 8n "logical" DFT corresponding to the
Chris@10 33 R{EO}DFT11.
Chris@10 34
Chris@10 35 Aside from very confusing notation (several symbols are redefined
Chris@10 36 from one line to the next), be aware that this paper has some
Chris@10 37 errors. In particular, the signs are wrong in Eqs. (34-35). Also,
Chris@10 38 Eqs. (36-37) should be simply C(k) = C(2k + 1 mod N), and similarly
Chris@10 39 for S (or, equivalently, the second cases should have 2*N - 2*k - 1
Chris@10 40 instead of N - k - 1). Note also that in their definition of the
Chris@10 41 DFT, similarly to FFTW's, the exponent's sign is -1, but they
Chris@10 42 forgot to correspondingly multiply S (the sine terms) by -1.
Chris@10 43 */
Chris@10 44
Chris@10 45 #include "reodft.h"
Chris@10 46
Chris@10 47 typedef struct {
Chris@10 48 solver super;
Chris@10 49 } S;
Chris@10 50
Chris@10 51 typedef struct {
Chris@10 52 plan_rdft super;
Chris@10 53 plan *cld;
Chris@10 54 INT is, os;
Chris@10 55 INT n;
Chris@10 56 INT vl;
Chris@10 57 INT ivs, ovs;
Chris@10 58 rdft_kind kind;
Chris@10 59 } P;
Chris@10 60
Chris@10 61 static DK(SQRT2, +1.4142135623730950488016887242096980785696718753769);
Chris@10 62
Chris@10 63 #define SGN_SET(x, i) ((i) % 2 ? -(x) : (x))
Chris@10 64
Chris@10 65 static void apply_re11(const plan *ego_, R *I, R *O)
Chris@10 66 {
Chris@10 67 const P *ego = (const P *) ego_;
Chris@10 68 INT is = ego->is, os = ego->os;
Chris@10 69 INT i, n = ego->n, n2 = n/2;
Chris@10 70 INT iv, vl = ego->vl;
Chris@10 71 INT ivs = ego->ivs, ovs = ego->ovs;
Chris@10 72 R *buf;
Chris@10 73
Chris@10 74 buf = (R *) MALLOC(sizeof(R) * n, BUFFERS);
Chris@10 75
Chris@10 76 for (iv = 0; iv < vl; ++iv, I += ivs, O += ovs) {
Chris@10 77 {
Chris@10 78 INT m;
Chris@10 79 for (i = 0, m = n2; m < n; ++i, m += 4)
Chris@10 80 buf[i] = I[is * m];
Chris@10 81 for (; m < 2 * n; ++i, m += 4)
Chris@10 82 buf[i] = -I[is * (2*n - m - 1)];
Chris@10 83 for (; m < 3 * n; ++i, m += 4)
Chris@10 84 buf[i] = -I[is * (m - 2*n)];
Chris@10 85 for (; m < 4 * n; ++i, m += 4)
Chris@10 86 buf[i] = I[is * (4*n - m - 1)];
Chris@10 87 m -= 4 * n;
Chris@10 88 for (; i < n; ++i, m += 4)
Chris@10 89 buf[i] = I[is * m];
Chris@10 90 }
Chris@10 91
Chris@10 92 { /* child plan: R2HC of size n */
Chris@10 93 plan_rdft *cld = (plan_rdft *) ego->cld;
Chris@10 94 cld->apply((plan *) cld, buf, buf);
Chris@10 95 }
Chris@10 96
Chris@10 97 /* FIXME: strength-reduce loop by 4 to eliminate ugly sgn_set? */
Chris@10 98 for (i = 0; i + i + 1 < n2; ++i) {
Chris@10 99 INT k = i + i + 1;
Chris@10 100 E c1, s1;
Chris@10 101 E c2, s2;
Chris@10 102 c1 = buf[k];
Chris@10 103 c2 = buf[k + 1];
Chris@10 104 s2 = buf[n - (k + 1)];
Chris@10 105 s1 = buf[n - k];
Chris@10 106
Chris@10 107 O[os * i] = SQRT2 * (SGN_SET(c1, (i+1)/2) +
Chris@10 108 SGN_SET(s1, i/2));
Chris@10 109 O[os * (n - (i+1))] = SQRT2 * (SGN_SET(c1, (n-i)/2) -
Chris@10 110 SGN_SET(s1, (n-(i+1))/2));
Chris@10 111
Chris@10 112 O[os * (n2 - (i+1))] = SQRT2 * (SGN_SET(c2, (n2-i)/2) -
Chris@10 113 SGN_SET(s2, (n2-(i+1))/2));
Chris@10 114 O[os * (n2 + (i+1))] = SQRT2 * (SGN_SET(c2, (n2+i+2)/2) +
Chris@10 115 SGN_SET(s2, (n2+(i+1))/2));
Chris@10 116 }
Chris@10 117 if (i + i + 1 == n2) {
Chris@10 118 E c, s;
Chris@10 119 c = buf[n2];
Chris@10 120 s = buf[n - n2];
Chris@10 121 O[os * i] = SQRT2 * (SGN_SET(c, (i+1)/2) +
Chris@10 122 SGN_SET(s, i/2));
Chris@10 123 O[os * (n - (i+1))] = SQRT2 * (SGN_SET(c, (i+2)/2) +
Chris@10 124 SGN_SET(s, (i+1)/2));
Chris@10 125 }
Chris@10 126 O[os * n2] = SQRT2 * SGN_SET(buf[0], (n2+1)/2);
Chris@10 127 }
Chris@10 128
Chris@10 129 X(ifree)(buf);
Chris@10 130 }
Chris@10 131
Chris@10 132 /* like for rodft01, rodft11 is obtained from redft11 by
Chris@10 133 reversing the input and flipping the sign of every other output. */
Chris@10 134 static void apply_ro11(const plan *ego_, R *I, R *O)
Chris@10 135 {
Chris@10 136 const P *ego = (const P *) ego_;
Chris@10 137 INT is = ego->is, os = ego->os;
Chris@10 138 INT i, n = ego->n, n2 = n/2;
Chris@10 139 INT iv, vl = ego->vl;
Chris@10 140 INT ivs = ego->ivs, ovs = ego->ovs;
Chris@10 141 R *buf;
Chris@10 142
Chris@10 143 buf = (R *) MALLOC(sizeof(R) * n, BUFFERS);
Chris@10 144
Chris@10 145 for (iv = 0; iv < vl; ++iv, I += ivs, O += ovs) {
Chris@10 146 {
Chris@10 147 INT m;
Chris@10 148 for (i = 0, m = n2; m < n; ++i, m += 4)
Chris@10 149 buf[i] = I[is * (n - 1 - m)];
Chris@10 150 for (; m < 2 * n; ++i, m += 4)
Chris@10 151 buf[i] = -I[is * (m - n)];
Chris@10 152 for (; m < 3 * n; ++i, m += 4)
Chris@10 153 buf[i] = -I[is * (3*n - 1 - m)];
Chris@10 154 for (; m < 4 * n; ++i, m += 4)
Chris@10 155 buf[i] = I[is * (m - 3*n)];
Chris@10 156 m -= 4 * n;
Chris@10 157 for (; i < n; ++i, m += 4)
Chris@10 158 buf[i] = I[is * (n - 1 - m)];
Chris@10 159 }
Chris@10 160
Chris@10 161 { /* child plan: R2HC of size n */
Chris@10 162 plan_rdft *cld = (plan_rdft *) ego->cld;
Chris@10 163 cld->apply((plan *) cld, buf, buf);
Chris@10 164 }
Chris@10 165
Chris@10 166 /* FIXME: strength-reduce loop by 4 to eliminate ugly sgn_set? */
Chris@10 167 for (i = 0; i + i + 1 < n2; ++i) {
Chris@10 168 INT k = i + i + 1;
Chris@10 169 INT j;
Chris@10 170 E c1, s1;
Chris@10 171 E c2, s2;
Chris@10 172 c1 = buf[k];
Chris@10 173 c2 = buf[k + 1];
Chris@10 174 s2 = buf[n - (k + 1)];
Chris@10 175 s1 = buf[n - k];
Chris@10 176
Chris@10 177 O[os * i] = SQRT2 * (SGN_SET(c1, (i+1)/2 + i) +
Chris@10 178 SGN_SET(s1, i/2 + i));
Chris@10 179 O[os * (n - (i+1))] = SQRT2 * (SGN_SET(c1, (n-i)/2 + i) -
Chris@10 180 SGN_SET(s1, (n-(i+1))/2 + i));
Chris@10 181
Chris@10 182 j = n2 - (i+1);
Chris@10 183 O[os * j] = SQRT2 * (SGN_SET(c2, (n2-i)/2 + j) -
Chris@10 184 SGN_SET(s2, (n2-(i+1))/2 + j));
Chris@10 185 O[os * (n2 + (i+1))] = SQRT2 * (SGN_SET(c2, (n2+i+2)/2 + j) +
Chris@10 186 SGN_SET(s2, (n2+(i+1))/2 + j));
Chris@10 187 }
Chris@10 188 if (i + i + 1 == n2) {
Chris@10 189 E c, s;
Chris@10 190 c = buf[n2];
Chris@10 191 s = buf[n - n2];
Chris@10 192 O[os * i] = SQRT2 * (SGN_SET(c, (i+1)/2 + i) +
Chris@10 193 SGN_SET(s, i/2 + i));
Chris@10 194 O[os * (n - (i+1))] = SQRT2 * (SGN_SET(c, (i+2)/2 + i) +
Chris@10 195 SGN_SET(s, (i+1)/2 + i));
Chris@10 196 }
Chris@10 197 O[os * n2] = SQRT2 * SGN_SET(buf[0], (n2+1)/2 + n2);
Chris@10 198 }
Chris@10 199
Chris@10 200 X(ifree)(buf);
Chris@10 201 }
Chris@10 202
Chris@10 203 static void awake(plan *ego_, enum wakefulness wakefulness)
Chris@10 204 {
Chris@10 205 P *ego = (P *) ego_;
Chris@10 206 X(plan_awake)(ego->cld, wakefulness);
Chris@10 207 }
Chris@10 208
Chris@10 209 static void destroy(plan *ego_)
Chris@10 210 {
Chris@10 211 P *ego = (P *) ego_;
Chris@10 212 X(plan_destroy_internal)(ego->cld);
Chris@10 213 }
Chris@10 214
Chris@10 215 static void print(const plan *ego_, printer *p)
Chris@10 216 {
Chris@10 217 const P *ego = (const P *) ego_;
Chris@10 218 p->print(p, "(%se-r2hc-odd-%D%v%(%p%))",
Chris@10 219 X(rdft_kind_str)(ego->kind), ego->n, ego->vl, ego->cld);
Chris@10 220 }
Chris@10 221
Chris@10 222 static int applicable0(const solver *ego_, const problem *p_)
Chris@10 223 {
Chris@10 224 const problem_rdft *p = (const problem_rdft *) p_;
Chris@10 225 UNUSED(ego_);
Chris@10 226
Chris@10 227 return (1
Chris@10 228 && p->sz->rnk == 1
Chris@10 229 && p->vecsz->rnk <= 1
Chris@10 230 && p->sz->dims[0].n % 2 == 1
Chris@10 231 && (p->kind[0] == REDFT11 || p->kind[0] == RODFT11)
Chris@10 232 );
Chris@10 233 }
Chris@10 234
Chris@10 235 static int applicable(const solver *ego, const problem *p, const planner *plnr)
Chris@10 236 {
Chris@10 237 return (!NO_SLOWP(plnr) && applicable0(ego, p));
Chris@10 238 }
Chris@10 239
Chris@10 240 static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
Chris@10 241 {
Chris@10 242 P *pln;
Chris@10 243 const problem_rdft *p;
Chris@10 244 plan *cld;
Chris@10 245 R *buf;
Chris@10 246 INT n;
Chris@10 247 opcnt ops;
Chris@10 248
Chris@10 249 static const plan_adt padt = {
Chris@10 250 X(rdft_solve), awake, print, destroy
Chris@10 251 };
Chris@10 252
Chris@10 253 if (!applicable(ego_, p_, plnr))
Chris@10 254 return (plan *)0;
Chris@10 255
Chris@10 256 p = (const problem_rdft *) p_;
Chris@10 257
Chris@10 258 n = p->sz->dims[0].n;
Chris@10 259 buf = (R *) MALLOC(sizeof(R) * n, BUFFERS);
Chris@10 260
Chris@10 261 cld = X(mkplan_d)(plnr, X(mkproblem_rdft_1_d)(X(mktensor_1d)(n, 1, 1),
Chris@10 262 X(mktensor_0d)(),
Chris@10 263 buf, buf, R2HC));
Chris@10 264 X(ifree)(buf);
Chris@10 265 if (!cld)
Chris@10 266 return (plan *)0;
Chris@10 267
Chris@10 268 pln = MKPLAN_RDFT(P, &padt, p->kind[0]==REDFT11 ? apply_re11:apply_ro11);
Chris@10 269 pln->n = n;
Chris@10 270 pln->is = p->sz->dims[0].is;
Chris@10 271 pln->os = p->sz->dims[0].os;
Chris@10 272 pln->cld = cld;
Chris@10 273 pln->kind = p->kind[0];
Chris@10 274
Chris@10 275 X(tensor_tornk1)(p->vecsz, &pln->vl, &pln->ivs, &pln->ovs);
Chris@10 276
Chris@10 277 X(ops_zero)(&ops);
Chris@10 278 ops.add = n - 1;
Chris@10 279 ops.mul = n;
Chris@10 280 ops.other = 4*n;
Chris@10 281
Chris@10 282 X(ops_zero)(&pln->super.super.ops);
Chris@10 283 X(ops_madd2)(pln->vl, &ops, &pln->super.super.ops);
Chris@10 284 X(ops_madd2)(pln->vl, &cld->ops, &pln->super.super.ops);
Chris@10 285
Chris@10 286 return &(pln->super.super);
Chris@10 287 }
Chris@10 288
Chris@10 289 /* constructor */
Chris@10 290 static solver *mksolver(void)
Chris@10 291 {
Chris@10 292 static const solver_adt sadt = { PROBLEM_RDFT, mkplan, 0 };
Chris@10 293 S *slv = MKSOLVER(S, &sadt);
Chris@10 294 return &(slv->super);
Chris@10 295 }
Chris@10 296
Chris@10 297 void X(reodft11e_r2hc_odd_register)(planner *p)
Chris@10 298 {
Chris@10 299 REGISTER_SOLVER(p, mksolver());
Chris@10 300 }