annotate fft/fftw/fftw-3.3.4/reodft/reodft11e-r2hc-odd.c @ 40:223f770b5341 kissfft-double tip

Try a double-precision kissfft
author Chris Cannam
date Wed, 07 Sep 2016 10:40:32 +0100
parents 26056e866c29
children
rev   line source
Chris@19 1 /*
Chris@19 2 * Copyright (c) 2003, 2007-14 Matteo Frigo
Chris@19 3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
Chris@19 4 *
Chris@19 5 * This program is free software; you can redistribute it and/or modify
Chris@19 6 * it under the terms of the GNU General Public License as published by
Chris@19 7 * the Free Software Foundation; either version 2 of the License, or
Chris@19 8 * (at your option) any later version.
Chris@19 9 *
Chris@19 10 * This program is distributed in the hope that it will be useful,
Chris@19 11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
Chris@19 12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
Chris@19 13 * GNU General Public License for more details.
Chris@19 14 *
Chris@19 15 * You should have received a copy of the GNU General Public License
Chris@19 16 * along with this program; if not, write to the Free Software
Chris@19 17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
Chris@19 18 *
Chris@19 19 */
Chris@19 20
Chris@19 21
Chris@19 22 /* Do an R{E,O}DFT11 problem via an R2HC problem of the same *odd* size,
Chris@19 23 with some permutations and post-processing, as described in:
Chris@19 24
Chris@19 25 S. C. Chan and K. L. Ho, "Fast algorithms for computing the
Chris@19 26 discrete cosine transform," IEEE Trans. Circuits Systems II:
Chris@19 27 Analog & Digital Sig. Proc. 39 (3), 185--190 (1992).
Chris@19 28
Chris@19 29 (For even sizes, see reodft11e-radix2.c.)
Chris@19 30
Chris@19 31 This algorithm is related to the 8 x n prime-factor-algorithm (PFA)
Chris@19 32 decomposition of the size 8n "logical" DFT corresponding to the
Chris@19 33 R{EO}DFT11.
Chris@19 34
Chris@19 35 Aside from very confusing notation (several symbols are redefined
Chris@19 36 from one line to the next), be aware that this paper has some
Chris@19 37 errors. In particular, the signs are wrong in Eqs. (34-35). Also,
Chris@19 38 Eqs. (36-37) should be simply C(k) = C(2k + 1 mod N), and similarly
Chris@19 39 for S (or, equivalently, the second cases should have 2*N - 2*k - 1
Chris@19 40 instead of N - k - 1). Note also that in their definition of the
Chris@19 41 DFT, similarly to FFTW's, the exponent's sign is -1, but they
Chris@19 42 forgot to correspondingly multiply S (the sine terms) by -1.
Chris@19 43 */
Chris@19 44
Chris@19 45 #include "reodft.h"
Chris@19 46
Chris@19 47 typedef struct {
Chris@19 48 solver super;
Chris@19 49 } S;
Chris@19 50
Chris@19 51 typedef struct {
Chris@19 52 plan_rdft super;
Chris@19 53 plan *cld;
Chris@19 54 INT is, os;
Chris@19 55 INT n;
Chris@19 56 INT vl;
Chris@19 57 INT ivs, ovs;
Chris@19 58 rdft_kind kind;
Chris@19 59 } P;
Chris@19 60
Chris@19 61 static DK(SQRT2, +1.4142135623730950488016887242096980785696718753769);
Chris@19 62
Chris@19 63 #define SGN_SET(x, i) ((i) % 2 ? -(x) : (x))
Chris@19 64
Chris@19 65 static void apply_re11(const plan *ego_, R *I, R *O)
Chris@19 66 {
Chris@19 67 const P *ego = (const P *) ego_;
Chris@19 68 INT is = ego->is, os = ego->os;
Chris@19 69 INT i, n = ego->n, n2 = n/2;
Chris@19 70 INT iv, vl = ego->vl;
Chris@19 71 INT ivs = ego->ivs, ovs = ego->ovs;
Chris@19 72 R *buf;
Chris@19 73
Chris@19 74 buf = (R *) MALLOC(sizeof(R) * n, BUFFERS);
Chris@19 75
Chris@19 76 for (iv = 0; iv < vl; ++iv, I += ivs, O += ovs) {
Chris@19 77 {
Chris@19 78 INT m;
Chris@19 79 for (i = 0, m = n2; m < n; ++i, m += 4)
Chris@19 80 buf[i] = I[is * m];
Chris@19 81 for (; m < 2 * n; ++i, m += 4)
Chris@19 82 buf[i] = -I[is * (2*n - m - 1)];
Chris@19 83 for (; m < 3 * n; ++i, m += 4)
Chris@19 84 buf[i] = -I[is * (m - 2*n)];
Chris@19 85 for (; m < 4 * n; ++i, m += 4)
Chris@19 86 buf[i] = I[is * (4*n - m - 1)];
Chris@19 87 m -= 4 * n;
Chris@19 88 for (; i < n; ++i, m += 4)
Chris@19 89 buf[i] = I[is * m];
Chris@19 90 }
Chris@19 91
Chris@19 92 { /* child plan: R2HC of size n */
Chris@19 93 plan_rdft *cld = (plan_rdft *) ego->cld;
Chris@19 94 cld->apply((plan *) cld, buf, buf);
Chris@19 95 }
Chris@19 96
Chris@19 97 /* FIXME: strength-reduce loop by 4 to eliminate ugly sgn_set? */
Chris@19 98 for (i = 0; i + i + 1 < n2; ++i) {
Chris@19 99 INT k = i + i + 1;
Chris@19 100 E c1, s1;
Chris@19 101 E c2, s2;
Chris@19 102 c1 = buf[k];
Chris@19 103 c2 = buf[k + 1];
Chris@19 104 s2 = buf[n - (k + 1)];
Chris@19 105 s1 = buf[n - k];
Chris@19 106
Chris@19 107 O[os * i] = SQRT2 * (SGN_SET(c1, (i+1)/2) +
Chris@19 108 SGN_SET(s1, i/2));
Chris@19 109 O[os * (n - (i+1))] = SQRT2 * (SGN_SET(c1, (n-i)/2) -
Chris@19 110 SGN_SET(s1, (n-(i+1))/2));
Chris@19 111
Chris@19 112 O[os * (n2 - (i+1))] = SQRT2 * (SGN_SET(c2, (n2-i)/2) -
Chris@19 113 SGN_SET(s2, (n2-(i+1))/2));
Chris@19 114 O[os * (n2 + (i+1))] = SQRT2 * (SGN_SET(c2, (n2+i+2)/2) +
Chris@19 115 SGN_SET(s2, (n2+(i+1))/2));
Chris@19 116 }
Chris@19 117 if (i + i + 1 == n2) {
Chris@19 118 E c, s;
Chris@19 119 c = buf[n2];
Chris@19 120 s = buf[n - n2];
Chris@19 121 O[os * i] = SQRT2 * (SGN_SET(c, (i+1)/2) +
Chris@19 122 SGN_SET(s, i/2));
Chris@19 123 O[os * (n - (i+1))] = SQRT2 * (SGN_SET(c, (i+2)/2) +
Chris@19 124 SGN_SET(s, (i+1)/2));
Chris@19 125 }
Chris@19 126 O[os * n2] = SQRT2 * SGN_SET(buf[0], (n2+1)/2);
Chris@19 127 }
Chris@19 128
Chris@19 129 X(ifree)(buf);
Chris@19 130 }
Chris@19 131
Chris@19 132 /* like for rodft01, rodft11 is obtained from redft11 by
Chris@19 133 reversing the input and flipping the sign of every other output. */
Chris@19 134 static void apply_ro11(const plan *ego_, R *I, R *O)
Chris@19 135 {
Chris@19 136 const P *ego = (const P *) ego_;
Chris@19 137 INT is = ego->is, os = ego->os;
Chris@19 138 INT i, n = ego->n, n2 = n/2;
Chris@19 139 INT iv, vl = ego->vl;
Chris@19 140 INT ivs = ego->ivs, ovs = ego->ovs;
Chris@19 141 R *buf;
Chris@19 142
Chris@19 143 buf = (R *) MALLOC(sizeof(R) * n, BUFFERS);
Chris@19 144
Chris@19 145 for (iv = 0; iv < vl; ++iv, I += ivs, O += ovs) {
Chris@19 146 {
Chris@19 147 INT m;
Chris@19 148 for (i = 0, m = n2; m < n; ++i, m += 4)
Chris@19 149 buf[i] = I[is * (n - 1 - m)];
Chris@19 150 for (; m < 2 * n; ++i, m += 4)
Chris@19 151 buf[i] = -I[is * (m - n)];
Chris@19 152 for (; m < 3 * n; ++i, m += 4)
Chris@19 153 buf[i] = -I[is * (3*n - 1 - m)];
Chris@19 154 for (; m < 4 * n; ++i, m += 4)
Chris@19 155 buf[i] = I[is * (m - 3*n)];
Chris@19 156 m -= 4 * n;
Chris@19 157 for (; i < n; ++i, m += 4)
Chris@19 158 buf[i] = I[is * (n - 1 - m)];
Chris@19 159 }
Chris@19 160
Chris@19 161 { /* child plan: R2HC of size n */
Chris@19 162 plan_rdft *cld = (plan_rdft *) ego->cld;
Chris@19 163 cld->apply((plan *) cld, buf, buf);
Chris@19 164 }
Chris@19 165
Chris@19 166 /* FIXME: strength-reduce loop by 4 to eliminate ugly sgn_set? */
Chris@19 167 for (i = 0; i + i + 1 < n2; ++i) {
Chris@19 168 INT k = i + i + 1;
Chris@19 169 INT j;
Chris@19 170 E c1, s1;
Chris@19 171 E c2, s2;
Chris@19 172 c1 = buf[k];
Chris@19 173 c2 = buf[k + 1];
Chris@19 174 s2 = buf[n - (k + 1)];
Chris@19 175 s1 = buf[n - k];
Chris@19 176
Chris@19 177 O[os * i] = SQRT2 * (SGN_SET(c1, (i+1)/2 + i) +
Chris@19 178 SGN_SET(s1, i/2 + i));
Chris@19 179 O[os * (n - (i+1))] = SQRT2 * (SGN_SET(c1, (n-i)/2 + i) -
Chris@19 180 SGN_SET(s1, (n-(i+1))/2 + i));
Chris@19 181
Chris@19 182 j = n2 - (i+1);
Chris@19 183 O[os * j] = SQRT2 * (SGN_SET(c2, (n2-i)/2 + j) -
Chris@19 184 SGN_SET(s2, (n2-(i+1))/2 + j));
Chris@19 185 O[os * (n2 + (i+1))] = SQRT2 * (SGN_SET(c2, (n2+i+2)/2 + j) +
Chris@19 186 SGN_SET(s2, (n2+(i+1))/2 + j));
Chris@19 187 }
Chris@19 188 if (i + i + 1 == n2) {
Chris@19 189 E c, s;
Chris@19 190 c = buf[n2];
Chris@19 191 s = buf[n - n2];
Chris@19 192 O[os * i] = SQRT2 * (SGN_SET(c, (i+1)/2 + i) +
Chris@19 193 SGN_SET(s, i/2 + i));
Chris@19 194 O[os * (n - (i+1))] = SQRT2 * (SGN_SET(c, (i+2)/2 + i) +
Chris@19 195 SGN_SET(s, (i+1)/2 + i));
Chris@19 196 }
Chris@19 197 O[os * n2] = SQRT2 * SGN_SET(buf[0], (n2+1)/2 + n2);
Chris@19 198 }
Chris@19 199
Chris@19 200 X(ifree)(buf);
Chris@19 201 }
Chris@19 202
Chris@19 203 static void awake(plan *ego_, enum wakefulness wakefulness)
Chris@19 204 {
Chris@19 205 P *ego = (P *) ego_;
Chris@19 206 X(plan_awake)(ego->cld, wakefulness);
Chris@19 207 }
Chris@19 208
Chris@19 209 static void destroy(plan *ego_)
Chris@19 210 {
Chris@19 211 P *ego = (P *) ego_;
Chris@19 212 X(plan_destroy_internal)(ego->cld);
Chris@19 213 }
Chris@19 214
Chris@19 215 static void print(const plan *ego_, printer *p)
Chris@19 216 {
Chris@19 217 const P *ego = (const P *) ego_;
Chris@19 218 p->print(p, "(%se-r2hc-odd-%D%v%(%p%))",
Chris@19 219 X(rdft_kind_str)(ego->kind), ego->n, ego->vl, ego->cld);
Chris@19 220 }
Chris@19 221
Chris@19 222 static int applicable0(const solver *ego_, const problem *p_)
Chris@19 223 {
Chris@19 224 const problem_rdft *p = (const problem_rdft *) p_;
Chris@19 225 UNUSED(ego_);
Chris@19 226
Chris@19 227 return (1
Chris@19 228 && p->sz->rnk == 1
Chris@19 229 && p->vecsz->rnk <= 1
Chris@19 230 && p->sz->dims[0].n % 2 == 1
Chris@19 231 && (p->kind[0] == REDFT11 || p->kind[0] == RODFT11)
Chris@19 232 );
Chris@19 233 }
Chris@19 234
Chris@19 235 static int applicable(const solver *ego, const problem *p, const planner *plnr)
Chris@19 236 {
Chris@19 237 return (!NO_SLOWP(plnr) && applicable0(ego, p));
Chris@19 238 }
Chris@19 239
Chris@19 240 static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
Chris@19 241 {
Chris@19 242 P *pln;
Chris@19 243 const problem_rdft *p;
Chris@19 244 plan *cld;
Chris@19 245 R *buf;
Chris@19 246 INT n;
Chris@19 247 opcnt ops;
Chris@19 248
Chris@19 249 static const plan_adt padt = {
Chris@19 250 X(rdft_solve), awake, print, destroy
Chris@19 251 };
Chris@19 252
Chris@19 253 if (!applicable(ego_, p_, plnr))
Chris@19 254 return (plan *)0;
Chris@19 255
Chris@19 256 p = (const problem_rdft *) p_;
Chris@19 257
Chris@19 258 n = p->sz->dims[0].n;
Chris@19 259 buf = (R *) MALLOC(sizeof(R) * n, BUFFERS);
Chris@19 260
Chris@19 261 cld = X(mkplan_d)(plnr, X(mkproblem_rdft_1_d)(X(mktensor_1d)(n, 1, 1),
Chris@19 262 X(mktensor_0d)(),
Chris@19 263 buf, buf, R2HC));
Chris@19 264 X(ifree)(buf);
Chris@19 265 if (!cld)
Chris@19 266 return (plan *)0;
Chris@19 267
Chris@19 268 pln = MKPLAN_RDFT(P, &padt, p->kind[0]==REDFT11 ? apply_re11:apply_ro11);
Chris@19 269 pln->n = n;
Chris@19 270 pln->is = p->sz->dims[0].is;
Chris@19 271 pln->os = p->sz->dims[0].os;
Chris@19 272 pln->cld = cld;
Chris@19 273 pln->kind = p->kind[0];
Chris@19 274
Chris@19 275 X(tensor_tornk1)(p->vecsz, &pln->vl, &pln->ivs, &pln->ovs);
Chris@19 276
Chris@19 277 X(ops_zero)(&ops);
Chris@19 278 ops.add = n - 1;
Chris@19 279 ops.mul = n;
Chris@19 280 ops.other = 4*n;
Chris@19 281
Chris@19 282 X(ops_zero)(&pln->super.super.ops);
Chris@19 283 X(ops_madd2)(pln->vl, &ops, &pln->super.super.ops);
Chris@19 284 X(ops_madd2)(pln->vl, &cld->ops, &pln->super.super.ops);
Chris@19 285
Chris@19 286 return &(pln->super.super);
Chris@19 287 }
Chris@19 288
Chris@19 289 /* constructor */
Chris@19 290 static solver *mksolver(void)
Chris@19 291 {
Chris@19 292 static const solver_adt sadt = { PROBLEM_RDFT, mkplan, 0 };
Chris@19 293 S *slv = MKSOLVER(S, &sadt);
Chris@19 294 return &(slv->super);
Chris@19 295 }
Chris@19 296
Chris@19 297 void X(reodft11e_r2hc_odd_register)(planner *p)
Chris@19 298 {
Chris@19 299 REGISTER_SOLVER(p, mksolver());
Chris@19 300 }