Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.5/rdft/scalar/r2cf/hc2cf_16.c @ 56:af97cad61ff0

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Add updated build of PortAudio for OSX
author Chris Cannam <cannam@all-day-breakfast.com>
date Tue, 03 Jan 2017 15:10:52 +0000
parents 2cd0e3b3e1fd
children
rev   line source
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 /* This file was automatically generated --- DO NOT EDIT */
Chris@42 22 /* Generated on Sat Jul 30 16:48:09 EDT 2016 */
Chris@42 23
Chris@42 24 #include "codelet-rdft.h"
Chris@42 25
Chris@42 26 #ifdef HAVE_FMA
Chris@42 27
Chris@42 28 /* Generated by: ../../../genfft/gen_hc2c.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 16 -dit -name hc2cf_16 -include hc2cf.h */
Chris@42 29
Chris@42 30 /*
Chris@42 31 * This function contains 174 FP additions, 100 FP multiplications,
Chris@42 32 * (or, 104 additions, 30 multiplications, 70 fused multiply/add),
Chris@42 33 * 97 stack variables, 3 constants, and 64 memory accesses
Chris@42 34 */
Chris@42 35 #include "hc2cf.h"
Chris@42 36
Chris@42 37 static void hc2cf_16(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
Chris@42 38 {
Chris@42 39 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
Chris@42 40 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
Chris@42 41 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
Chris@42 42 {
Chris@42 43 INT m;
Chris@42 44 for (m = mb, W = W + ((mb - 1) * 30); m < me; m = m + 1, Rp = Rp + ms, Ip = Ip + ms, Rm = Rm - ms, Im = Im - ms, W = W + 30, MAKE_VOLATILE_STRIDE(64, rs)) {
Chris@42 45 E T3G, T3F;
Chris@42 46 {
Chris@42 47 E T3z, T3o, T8, T1I, T2p, T35, T2r, T1s, T2w, T36, T2k, T1F, T3k, T1N, T3A;
Chris@42 48 E Tl, T1T, T2V, T1U, Tz, T29, T30, T2c, T11, TB, TE, T2h, T31, T2a, T1e;
Chris@42 49 E TC, T1X, TH, TK, TG, TD, TJ;
Chris@42 50 {
Chris@42 51 E Ta, Td, Tb, T1J, Tg, Tj, Tf, Tc, Ti;
Chris@42 52 {
Chris@42 53 E T1h, T1k, T1n, T2l, T1i, T1q, T1m, T1j, T1p;
Chris@42 54 {
Chris@42 55 E T1, T3n, T3, T6, T2, T5;
Chris@42 56 T1 = Rp[0];
Chris@42 57 T3n = Rm[0];
Chris@42 58 T3 = Rp[WS(rs, 4)];
Chris@42 59 T6 = Rm[WS(rs, 4)];
Chris@42 60 T2 = W[14];
Chris@42 61 T5 = W[15];
Chris@42 62 {
Chris@42 63 E T3l, T4, T1g, T3m, T7;
Chris@42 64 T1h = Ip[WS(rs, 7)];
Chris@42 65 T1k = Im[WS(rs, 7)];
Chris@42 66 T3l = T2 * T6;
Chris@42 67 T4 = T2 * T3;
Chris@42 68 T1g = W[28];
Chris@42 69 T1n = Ip[WS(rs, 3)];
Chris@42 70 T3m = FNMS(T5, T3, T3l);
Chris@42 71 T7 = FMA(T5, T6, T4);
Chris@42 72 T2l = T1g * T1k;
Chris@42 73 T1i = T1g * T1h;
Chris@42 74 T3z = T3n - T3m;
Chris@42 75 T3o = T3m + T3n;
Chris@42 76 T8 = T1 + T7;
Chris@42 77 T1I = T1 - T7;
Chris@42 78 T1q = Im[WS(rs, 3)];
Chris@42 79 T1m = W[12];
Chris@42 80 }
Chris@42 81 T1j = W[29];
Chris@42 82 T1p = W[13];
Chris@42 83 }
Chris@42 84 {
Chris@42 85 E T1u, T1x, T1v, T2s, T1A, T1D, T1z, T1w, T1C;
Chris@42 86 {
Chris@42 87 E T2m, T1l, T2o, T1r, T2n, T1o, T1t;
Chris@42 88 T1u = Ip[WS(rs, 1)];
Chris@42 89 T2n = T1m * T1q;
Chris@42 90 T1o = T1m * T1n;
Chris@42 91 T2m = FNMS(T1j, T1h, T2l);
Chris@42 92 T1l = FMA(T1j, T1k, T1i);
Chris@42 93 T2o = FNMS(T1p, T1n, T2n);
Chris@42 94 T1r = FMA(T1p, T1q, T1o);
Chris@42 95 T1x = Im[WS(rs, 1)];
Chris@42 96 T1t = W[4];
Chris@42 97 T2p = T2m - T2o;
Chris@42 98 T35 = T2m + T2o;
Chris@42 99 T2r = T1l - T1r;
Chris@42 100 T1s = T1l + T1r;
Chris@42 101 T1v = T1t * T1u;
Chris@42 102 T2s = T1t * T1x;
Chris@42 103 }
Chris@42 104 T1A = Ip[WS(rs, 5)];
Chris@42 105 T1D = Im[WS(rs, 5)];
Chris@42 106 T1z = W[20];
Chris@42 107 T1w = W[5];
Chris@42 108 T1C = W[21];
Chris@42 109 {
Chris@42 110 E T2t, T1y, T2v, T1E, T2u, T1B, T9;
Chris@42 111 Ta = Rp[WS(rs, 2)];
Chris@42 112 T2u = T1z * T1D;
Chris@42 113 T1B = T1z * T1A;
Chris@42 114 T2t = FNMS(T1w, T1u, T2s);
Chris@42 115 T1y = FMA(T1w, T1x, T1v);
Chris@42 116 T2v = FNMS(T1C, T1A, T2u);
Chris@42 117 T1E = FMA(T1C, T1D, T1B);
Chris@42 118 Td = Rm[WS(rs, 2)];
Chris@42 119 T9 = W[6];
Chris@42 120 T2w = T2t - T2v;
Chris@42 121 T36 = T2t + T2v;
Chris@42 122 T2k = T1E - T1y;
Chris@42 123 T1F = T1y + T1E;
Chris@42 124 Tb = T9 * Ta;
Chris@42 125 T1J = T9 * Td;
Chris@42 126 }
Chris@42 127 Tg = Rp[WS(rs, 6)];
Chris@42 128 Tj = Rm[WS(rs, 6)];
Chris@42 129 Tf = W[22];
Chris@42 130 Tc = W[7];
Chris@42 131 Ti = W[23];
Chris@42 132 }
Chris@42 133 }
Chris@42 134 {
Chris@42 135 E TQ, TT, TR, T25, TW, TZ, TV, TS, TY;
Chris@42 136 {
Chris@42 137 E To, Tr, Tp, T1P, Tu, Tx, Tt, Tq, Tw;
Chris@42 138 {
Chris@42 139 E T1K, Te, T1M, Tk, T1L, Th, Tn;
Chris@42 140 To = Rp[WS(rs, 1)];
Chris@42 141 T1L = Tf * Tj;
Chris@42 142 Th = Tf * Tg;
Chris@42 143 T1K = FNMS(Tc, Ta, T1J);
Chris@42 144 Te = FMA(Tc, Td, Tb);
Chris@42 145 T1M = FNMS(Ti, Tg, T1L);
Chris@42 146 Tk = FMA(Ti, Tj, Th);
Chris@42 147 Tr = Rm[WS(rs, 1)];
Chris@42 148 Tn = W[2];
Chris@42 149 T3k = T1K + T1M;
Chris@42 150 T1N = T1K - T1M;
Chris@42 151 T3A = Te - Tk;
Chris@42 152 Tl = Te + Tk;
Chris@42 153 Tp = Tn * To;
Chris@42 154 T1P = Tn * Tr;
Chris@42 155 }
Chris@42 156 Tu = Rp[WS(rs, 5)];
Chris@42 157 Tx = Rm[WS(rs, 5)];
Chris@42 158 Tt = W[18];
Chris@42 159 Tq = W[3];
Chris@42 160 Tw = W[19];
Chris@42 161 {
Chris@42 162 E T1Q, Ts, T1S, Ty, T1R, Tv, TP;
Chris@42 163 TQ = Ip[0];
Chris@42 164 T1R = Tt * Tx;
Chris@42 165 Tv = Tt * Tu;
Chris@42 166 T1Q = FNMS(Tq, To, T1P);
Chris@42 167 Ts = FMA(Tq, Tr, Tp);
Chris@42 168 T1S = FNMS(Tw, Tu, T1R);
Chris@42 169 Ty = FMA(Tw, Tx, Tv);
Chris@42 170 TT = Im[0];
Chris@42 171 TP = W[0];
Chris@42 172 T1T = T1Q - T1S;
Chris@42 173 T2V = T1Q + T1S;
Chris@42 174 T1U = Ts - Ty;
Chris@42 175 Tz = Ts + Ty;
Chris@42 176 TR = TP * TQ;
Chris@42 177 T25 = TP * TT;
Chris@42 178 }
Chris@42 179 TW = Ip[WS(rs, 4)];
Chris@42 180 TZ = Im[WS(rs, 4)];
Chris@42 181 TV = W[16];
Chris@42 182 TS = W[1];
Chris@42 183 TY = W[17];
Chris@42 184 }
Chris@42 185 {
Chris@42 186 E T13, T16, T14, T2d, T19, T1c, T18, T15, T1b;
Chris@42 187 {
Chris@42 188 E T26, TU, T28, T10, T27, TX, T12;
Chris@42 189 T13 = Ip[WS(rs, 2)];
Chris@42 190 T27 = TV * TZ;
Chris@42 191 TX = TV * TW;
Chris@42 192 T26 = FNMS(TS, TQ, T25);
Chris@42 193 TU = FMA(TS, TT, TR);
Chris@42 194 T28 = FNMS(TY, TW, T27);
Chris@42 195 T10 = FMA(TY, TZ, TX);
Chris@42 196 T16 = Im[WS(rs, 2)];
Chris@42 197 T12 = W[8];
Chris@42 198 T29 = T26 - T28;
Chris@42 199 T30 = T26 + T28;
Chris@42 200 T2c = TU - T10;
Chris@42 201 T11 = TU + T10;
Chris@42 202 T14 = T12 * T13;
Chris@42 203 T2d = T12 * T16;
Chris@42 204 }
Chris@42 205 T19 = Ip[WS(rs, 6)];
Chris@42 206 T1c = Im[WS(rs, 6)];
Chris@42 207 T18 = W[24];
Chris@42 208 T15 = W[9];
Chris@42 209 T1b = W[25];
Chris@42 210 {
Chris@42 211 E T2e, T17, T2g, T1d, T2f, T1a, TA;
Chris@42 212 TB = Rp[WS(rs, 7)];
Chris@42 213 T2f = T18 * T1c;
Chris@42 214 T1a = T18 * T19;
Chris@42 215 T2e = FNMS(T15, T13, T2d);
Chris@42 216 T17 = FMA(T15, T16, T14);
Chris@42 217 T2g = FNMS(T1b, T19, T2f);
Chris@42 218 T1d = FMA(T1b, T1c, T1a);
Chris@42 219 TE = Rm[WS(rs, 7)];
Chris@42 220 TA = W[26];
Chris@42 221 T2h = T2e - T2g;
Chris@42 222 T31 = T2e + T2g;
Chris@42 223 T2a = T17 - T1d;
Chris@42 224 T1e = T17 + T1d;
Chris@42 225 TC = TA * TB;
Chris@42 226 T1X = TA * TE;
Chris@42 227 }
Chris@42 228 TH = Rp[WS(rs, 3)];
Chris@42 229 TK = Rm[WS(rs, 3)];
Chris@42 230 TG = W[10];
Chris@42 231 TD = W[27];
Chris@42 232 TJ = W[11];
Chris@42 233 }
Chris@42 234 }
Chris@42 235 }
Chris@42 236 {
Chris@42 237 E T2U, T3u, T2Z, T21, T1W, T34, T2X, T3f, T32, T3t, T1H, T3q, T3e, TO, T3g;
Chris@42 238 E T37, T3r, T3s, T3h, T3i;
Chris@42 239 {
Chris@42 240 E Tm, T1Y, TF, T20, TL, T3p, T1Z, TI;
Chris@42 241 T2U = T8 - Tl;
Chris@42 242 Tm = T8 + Tl;
Chris@42 243 T1Z = TG * TK;
Chris@42 244 TI = TG * TH;
Chris@42 245 T1Y = FNMS(TD, TB, T1X);
Chris@42 246 TF = FMA(TD, TE, TC);
Chris@42 247 T20 = FNMS(TJ, TH, T1Z);
Chris@42 248 TL = FMA(TJ, TK, TI);
Chris@42 249 T3p = T3k + T3o;
Chris@42 250 T3u = T3o - T3k;
Chris@42 251 {
Chris@42 252 E T1f, TM, T1G, T3j, T2W, TN;
Chris@42 253 T2Z = T11 - T1e;
Chris@42 254 T1f = T11 + T1e;
Chris@42 255 T21 = T1Y - T20;
Chris@42 256 T2W = T1Y + T20;
Chris@42 257 T1W = TF - TL;
Chris@42 258 TM = TF + TL;
Chris@42 259 T1G = T1s + T1F;
Chris@42 260 T34 = T1s - T1F;
Chris@42 261 T2X = T2V - T2W;
Chris@42 262 T3j = T2V + T2W;
Chris@42 263 T3f = T30 + T31;
Chris@42 264 T32 = T30 - T31;
Chris@42 265 T3t = TM - Tz;
Chris@42 266 TN = Tz + TM;
Chris@42 267 T3r = T1G - T1f;
Chris@42 268 T1H = T1f + T1G;
Chris@42 269 T3s = T3p - T3j;
Chris@42 270 T3q = T3j + T3p;
Chris@42 271 T3e = Tm - TN;
Chris@42 272 TO = Tm + TN;
Chris@42 273 T3g = T35 + T36;
Chris@42 274 T37 = T35 - T36;
Chris@42 275 }
Chris@42 276 }
Chris@42 277 Im[WS(rs, 3)] = T3r - T3s;
Chris@42 278 Ip[WS(rs, 4)] = T3r + T3s;
Chris@42 279 Rp[0] = TO + T1H;
Chris@42 280 Rm[WS(rs, 7)] = TO - T1H;
Chris@42 281 T3h = T3f - T3g;
Chris@42 282 T3i = T3f + T3g;
Chris@42 283 {
Chris@42 284 E T3a, T2Y, T3x, T3v, T3b, T33;
Chris@42 285 Ip[0] = T3i + T3q;
Chris@42 286 Im[WS(rs, 7)] = T3i - T3q;
Chris@42 287 Rp[WS(rs, 4)] = T3e + T3h;
Chris@42 288 Rm[WS(rs, 3)] = T3e - T3h;
Chris@42 289 T3a = T2U - T2X;
Chris@42 290 T2Y = T2U + T2X;
Chris@42 291 T3x = T3u - T3t;
Chris@42 292 T3v = T3t + T3u;
Chris@42 293 T3b = T32 - T2Z;
Chris@42 294 T33 = T2Z + T32;
Chris@42 295 {
Chris@42 296 E T2E, T1O, T3B, T3H, T2x, T2q, T3C, T23, T2S, T2O, T2K, T2J, T3I, T2H, T2B;
Chris@42 297 E T2j;
Chris@42 298 {
Chris@42 299 E T2F, T1V, T22, T2G, T3c, T38;
Chris@42 300 T2E = T1I + T1N;
Chris@42 301 T1O = T1I - T1N;
Chris@42 302 T3B = T3z - T3A;
Chris@42 303 T3H = T3A + T3z;
Chris@42 304 T3c = T34 + T37;
Chris@42 305 T38 = T34 - T37;
Chris@42 306 T2F = T1U + T1T;
Chris@42 307 T1V = T1T - T1U;
Chris@42 308 {
Chris@42 309 E T3d, T3w, T3y, T39;
Chris@42 310 T3d = T3b - T3c;
Chris@42 311 T3w = T3b + T3c;
Chris@42 312 T3y = T38 - T33;
Chris@42 313 T39 = T33 + T38;
Chris@42 314 Rp[WS(rs, 6)] = FMA(KP707106781, T3d, T3a);
Chris@42 315 Rm[WS(rs, 1)] = FNMS(KP707106781, T3d, T3a);
Chris@42 316 Ip[WS(rs, 2)] = FMA(KP707106781, T3w, T3v);
Chris@42 317 Im[WS(rs, 5)] = FMS(KP707106781, T3w, T3v);
Chris@42 318 Ip[WS(rs, 6)] = FMA(KP707106781, T3y, T3x);
Chris@42 319 Im[WS(rs, 1)] = FMS(KP707106781, T3y, T3x);
Chris@42 320 Rp[WS(rs, 2)] = FMA(KP707106781, T39, T2Y);
Chris@42 321 Rm[WS(rs, 5)] = FNMS(KP707106781, T39, T2Y);
Chris@42 322 T22 = T1W + T21;
Chris@42 323 T2G = T1W - T21;
Chris@42 324 }
Chris@42 325 {
Chris@42 326 E T2M, T2N, T2b, T2i;
Chris@42 327 T2x = T2r - T2w;
Chris@42 328 T2M = T2r + T2w;
Chris@42 329 T2N = T2p + T2k;
Chris@42 330 T2q = T2k - T2p;
Chris@42 331 T3C = T1V + T22;
Chris@42 332 T23 = T1V - T22;
Chris@42 333 T2S = FMA(KP414213562, T2M, T2N);
Chris@42 334 T2O = FNMS(KP414213562, T2N, T2M);
Chris@42 335 T2K = T29 - T2a;
Chris@42 336 T2b = T29 + T2a;
Chris@42 337 T2i = T2c - T2h;
Chris@42 338 T2J = T2c + T2h;
Chris@42 339 T3I = T2G - T2F;
Chris@42 340 T2H = T2F + T2G;
Chris@42 341 T2B = FNMS(KP414213562, T2b, T2i);
Chris@42 342 T2j = FMA(KP414213562, T2i, T2b);
Chris@42 343 }
Chris@42 344 }
Chris@42 345 {
Chris@42 346 E T2R, T2L, T3L, T3M;
Chris@42 347 {
Chris@42 348 E T2A, T24, T2C, T2y, T3J, T3K, T2D, T2z;
Chris@42 349 T2A = FNMS(KP707106781, T23, T1O);
Chris@42 350 T24 = FMA(KP707106781, T23, T1O);
Chris@42 351 T2R = FNMS(KP414213562, T2J, T2K);
Chris@42 352 T2L = FMA(KP414213562, T2K, T2J);
Chris@42 353 T2C = FNMS(KP414213562, T2q, T2x);
Chris@42 354 T2y = FMA(KP414213562, T2x, T2q);
Chris@42 355 T3J = FMA(KP707106781, T3I, T3H);
Chris@42 356 T3L = FNMS(KP707106781, T3I, T3H);
Chris@42 357 T3K = T2C - T2B;
Chris@42 358 T2D = T2B + T2C;
Chris@42 359 T3M = T2y - T2j;
Chris@42 360 T2z = T2j + T2y;
Chris@42 361 Ip[WS(rs, 3)] = FMA(KP923879532, T3K, T3J);
Chris@42 362 Im[WS(rs, 4)] = FMS(KP923879532, T3K, T3J);
Chris@42 363 Rp[WS(rs, 3)] = FMA(KP923879532, T2z, T24);
Chris@42 364 Rm[WS(rs, 4)] = FNMS(KP923879532, T2z, T24);
Chris@42 365 Rm[0] = FMA(KP923879532, T2D, T2A);
Chris@42 366 Rp[WS(rs, 7)] = FNMS(KP923879532, T2D, T2A);
Chris@42 367 }
Chris@42 368 {
Chris@42 369 E T2Q, T3D, T3E, T2T, T2I, T2P;
Chris@42 370 T2Q = FNMS(KP707106781, T2H, T2E);
Chris@42 371 T2I = FMA(KP707106781, T2H, T2E);
Chris@42 372 T2P = T2L + T2O;
Chris@42 373 T3G = T2O - T2L;
Chris@42 374 T3F = FNMS(KP707106781, T3C, T3B);
Chris@42 375 T3D = FMA(KP707106781, T3C, T3B);
Chris@42 376 Ip[WS(rs, 7)] = FMA(KP923879532, T3M, T3L);
Chris@42 377 Im[0] = FMS(KP923879532, T3M, T3L);
Chris@42 378 Rp[WS(rs, 1)] = FMA(KP923879532, T2P, T2I);
Chris@42 379 Rm[WS(rs, 6)] = FNMS(KP923879532, T2P, T2I);
Chris@42 380 T3E = T2R + T2S;
Chris@42 381 T2T = T2R - T2S;
Chris@42 382 Ip[WS(rs, 1)] = FMA(KP923879532, T3E, T3D);
Chris@42 383 Im[WS(rs, 6)] = FMS(KP923879532, T3E, T3D);
Chris@42 384 Rp[WS(rs, 5)] = FMA(KP923879532, T2T, T2Q);
Chris@42 385 Rm[WS(rs, 2)] = FNMS(KP923879532, T2T, T2Q);
Chris@42 386 }
Chris@42 387 }
Chris@42 388 }
Chris@42 389 }
Chris@42 390 }
Chris@42 391 }
Chris@42 392 Ip[WS(rs, 5)] = FMA(KP923879532, T3G, T3F);
Chris@42 393 Im[WS(rs, 2)] = FMS(KP923879532, T3G, T3F);
Chris@42 394 }
Chris@42 395 }
Chris@42 396 }
Chris@42 397
Chris@42 398 static const tw_instr twinstr[] = {
Chris@42 399 {TW_FULL, 1, 16},
Chris@42 400 {TW_NEXT, 1, 0}
Chris@42 401 };
Chris@42 402
Chris@42 403 static const hc2c_desc desc = { 16, "hc2cf_16", twinstr, &GENUS, {104, 30, 70, 0} };
Chris@42 404
Chris@42 405 void X(codelet_hc2cf_16) (planner *p) {
Chris@42 406 X(khc2c_register) (p, hc2cf_16, &desc, HC2C_VIA_RDFT);
Chris@42 407 }
Chris@42 408 #else /* HAVE_FMA */
Chris@42 409
Chris@42 410 /* Generated by: ../../../genfft/gen_hc2c.native -compact -variables 4 -pipeline-latency 4 -n 16 -dit -name hc2cf_16 -include hc2cf.h */
Chris@42 411
Chris@42 412 /*
Chris@42 413 * This function contains 174 FP additions, 84 FP multiplications,
Chris@42 414 * (or, 136 additions, 46 multiplications, 38 fused multiply/add),
Chris@42 415 * 52 stack variables, 3 constants, and 64 memory accesses
Chris@42 416 */
Chris@42 417 #include "hc2cf.h"
Chris@42 418
Chris@42 419 static void hc2cf_16(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
Chris@42 420 {
Chris@42 421 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
Chris@42 422 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
Chris@42 423 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
Chris@42 424 {
Chris@42 425 INT m;
Chris@42 426 for (m = mb, W = W + ((mb - 1) * 30); m < me; m = m + 1, Rp = Rp + ms, Ip = Ip + ms, Rm = Rm - ms, Im = Im - ms, W = W + 30, MAKE_VOLATILE_STRIDE(64, rs)) {
Chris@42 427 E T7, T37, T1t, T2U, Ti, T38, T1w, T2R, Tu, T2s, T1C, T2c, TF, T2t, T1H;
Chris@42 428 E T2d, T1f, T1q, T2B, T2C, T2D, T2E, T1Z, T2j, T24, T2k, TS, T13, T2w, T2x;
Chris@42 429 E T2y, T2z, T1O, T2g, T1T, T2h;
Chris@42 430 {
Chris@42 431 E T1, T2T, T6, T2S;
Chris@42 432 T1 = Rp[0];
Chris@42 433 T2T = Rm[0];
Chris@42 434 {
Chris@42 435 E T3, T5, T2, T4;
Chris@42 436 T3 = Rp[WS(rs, 4)];
Chris@42 437 T5 = Rm[WS(rs, 4)];
Chris@42 438 T2 = W[14];
Chris@42 439 T4 = W[15];
Chris@42 440 T6 = FMA(T2, T3, T4 * T5);
Chris@42 441 T2S = FNMS(T4, T3, T2 * T5);
Chris@42 442 }
Chris@42 443 T7 = T1 + T6;
Chris@42 444 T37 = T2T - T2S;
Chris@42 445 T1t = T1 - T6;
Chris@42 446 T2U = T2S + T2T;
Chris@42 447 }
Chris@42 448 {
Chris@42 449 E Tc, T1u, Th, T1v;
Chris@42 450 {
Chris@42 451 E T9, Tb, T8, Ta;
Chris@42 452 T9 = Rp[WS(rs, 2)];
Chris@42 453 Tb = Rm[WS(rs, 2)];
Chris@42 454 T8 = W[6];
Chris@42 455 Ta = W[7];
Chris@42 456 Tc = FMA(T8, T9, Ta * Tb);
Chris@42 457 T1u = FNMS(Ta, T9, T8 * Tb);
Chris@42 458 }
Chris@42 459 {
Chris@42 460 E Te, Tg, Td, Tf;
Chris@42 461 Te = Rp[WS(rs, 6)];
Chris@42 462 Tg = Rm[WS(rs, 6)];
Chris@42 463 Td = W[22];
Chris@42 464 Tf = W[23];
Chris@42 465 Th = FMA(Td, Te, Tf * Tg);
Chris@42 466 T1v = FNMS(Tf, Te, Td * Tg);
Chris@42 467 }
Chris@42 468 Ti = Tc + Th;
Chris@42 469 T38 = Tc - Th;
Chris@42 470 T1w = T1u - T1v;
Chris@42 471 T2R = T1u + T1v;
Chris@42 472 }
Chris@42 473 {
Chris@42 474 E To, T1y, Tt, T1z, T1A, T1B;
Chris@42 475 {
Chris@42 476 E Tl, Tn, Tk, Tm;
Chris@42 477 Tl = Rp[WS(rs, 1)];
Chris@42 478 Tn = Rm[WS(rs, 1)];
Chris@42 479 Tk = W[2];
Chris@42 480 Tm = W[3];
Chris@42 481 To = FMA(Tk, Tl, Tm * Tn);
Chris@42 482 T1y = FNMS(Tm, Tl, Tk * Tn);
Chris@42 483 }
Chris@42 484 {
Chris@42 485 E Tq, Ts, Tp, Tr;
Chris@42 486 Tq = Rp[WS(rs, 5)];
Chris@42 487 Ts = Rm[WS(rs, 5)];
Chris@42 488 Tp = W[18];
Chris@42 489 Tr = W[19];
Chris@42 490 Tt = FMA(Tp, Tq, Tr * Ts);
Chris@42 491 T1z = FNMS(Tr, Tq, Tp * Ts);
Chris@42 492 }
Chris@42 493 Tu = To + Tt;
Chris@42 494 T2s = T1y + T1z;
Chris@42 495 T1A = T1y - T1z;
Chris@42 496 T1B = To - Tt;
Chris@42 497 T1C = T1A - T1B;
Chris@42 498 T2c = T1B + T1A;
Chris@42 499 }
Chris@42 500 {
Chris@42 501 E Tz, T1E, TE, T1F, T1D, T1G;
Chris@42 502 {
Chris@42 503 E Tw, Ty, Tv, Tx;
Chris@42 504 Tw = Rp[WS(rs, 7)];
Chris@42 505 Ty = Rm[WS(rs, 7)];
Chris@42 506 Tv = W[26];
Chris@42 507 Tx = W[27];
Chris@42 508 Tz = FMA(Tv, Tw, Tx * Ty);
Chris@42 509 T1E = FNMS(Tx, Tw, Tv * Ty);
Chris@42 510 }
Chris@42 511 {
Chris@42 512 E TB, TD, TA, TC;
Chris@42 513 TB = Rp[WS(rs, 3)];
Chris@42 514 TD = Rm[WS(rs, 3)];
Chris@42 515 TA = W[10];
Chris@42 516 TC = W[11];
Chris@42 517 TE = FMA(TA, TB, TC * TD);
Chris@42 518 T1F = FNMS(TC, TB, TA * TD);
Chris@42 519 }
Chris@42 520 TF = Tz + TE;
Chris@42 521 T2t = T1E + T1F;
Chris@42 522 T1D = Tz - TE;
Chris@42 523 T1G = T1E - T1F;
Chris@42 524 T1H = T1D + T1G;
Chris@42 525 T2d = T1D - T1G;
Chris@42 526 }
Chris@42 527 {
Chris@42 528 E T19, T20, T1p, T1X, T1e, T21, T1k, T1W;
Chris@42 529 {
Chris@42 530 E T16, T18, T15, T17;
Chris@42 531 T16 = Ip[WS(rs, 7)];
Chris@42 532 T18 = Im[WS(rs, 7)];
Chris@42 533 T15 = W[28];
Chris@42 534 T17 = W[29];
Chris@42 535 T19 = FMA(T15, T16, T17 * T18);
Chris@42 536 T20 = FNMS(T17, T16, T15 * T18);
Chris@42 537 }
Chris@42 538 {
Chris@42 539 E T1m, T1o, T1l, T1n;
Chris@42 540 T1m = Ip[WS(rs, 5)];
Chris@42 541 T1o = Im[WS(rs, 5)];
Chris@42 542 T1l = W[20];
Chris@42 543 T1n = W[21];
Chris@42 544 T1p = FMA(T1l, T1m, T1n * T1o);
Chris@42 545 T1X = FNMS(T1n, T1m, T1l * T1o);
Chris@42 546 }
Chris@42 547 {
Chris@42 548 E T1b, T1d, T1a, T1c;
Chris@42 549 T1b = Ip[WS(rs, 3)];
Chris@42 550 T1d = Im[WS(rs, 3)];
Chris@42 551 T1a = W[12];
Chris@42 552 T1c = W[13];
Chris@42 553 T1e = FMA(T1a, T1b, T1c * T1d);
Chris@42 554 T21 = FNMS(T1c, T1b, T1a * T1d);
Chris@42 555 }
Chris@42 556 {
Chris@42 557 E T1h, T1j, T1g, T1i;
Chris@42 558 T1h = Ip[WS(rs, 1)];
Chris@42 559 T1j = Im[WS(rs, 1)];
Chris@42 560 T1g = W[4];
Chris@42 561 T1i = W[5];
Chris@42 562 T1k = FMA(T1g, T1h, T1i * T1j);
Chris@42 563 T1W = FNMS(T1i, T1h, T1g * T1j);
Chris@42 564 }
Chris@42 565 T1f = T19 + T1e;
Chris@42 566 T1q = T1k + T1p;
Chris@42 567 T2B = T1f - T1q;
Chris@42 568 T2C = T20 + T21;
Chris@42 569 T2D = T1W + T1X;
Chris@42 570 T2E = T2C - T2D;
Chris@42 571 {
Chris@42 572 E T1V, T1Y, T22, T23;
Chris@42 573 T1V = T19 - T1e;
Chris@42 574 T1Y = T1W - T1X;
Chris@42 575 T1Z = T1V - T1Y;
Chris@42 576 T2j = T1V + T1Y;
Chris@42 577 T22 = T20 - T21;
Chris@42 578 T23 = T1k - T1p;
Chris@42 579 T24 = T22 + T23;
Chris@42 580 T2k = T22 - T23;
Chris@42 581 }
Chris@42 582 }
Chris@42 583 {
Chris@42 584 E TM, T1K, T12, T1R, TR, T1L, TX, T1Q;
Chris@42 585 {
Chris@42 586 E TJ, TL, TI, TK;
Chris@42 587 TJ = Ip[0];
Chris@42 588 TL = Im[0];
Chris@42 589 TI = W[0];
Chris@42 590 TK = W[1];
Chris@42 591 TM = FMA(TI, TJ, TK * TL);
Chris@42 592 T1K = FNMS(TK, TJ, TI * TL);
Chris@42 593 }
Chris@42 594 {
Chris@42 595 E TZ, T11, TY, T10;
Chris@42 596 TZ = Ip[WS(rs, 6)];
Chris@42 597 T11 = Im[WS(rs, 6)];
Chris@42 598 TY = W[24];
Chris@42 599 T10 = W[25];
Chris@42 600 T12 = FMA(TY, TZ, T10 * T11);
Chris@42 601 T1R = FNMS(T10, TZ, TY * T11);
Chris@42 602 }
Chris@42 603 {
Chris@42 604 E TO, TQ, TN, TP;
Chris@42 605 TO = Ip[WS(rs, 4)];
Chris@42 606 TQ = Im[WS(rs, 4)];
Chris@42 607 TN = W[16];
Chris@42 608 TP = W[17];
Chris@42 609 TR = FMA(TN, TO, TP * TQ);
Chris@42 610 T1L = FNMS(TP, TO, TN * TQ);
Chris@42 611 }
Chris@42 612 {
Chris@42 613 E TU, TW, TT, TV;
Chris@42 614 TU = Ip[WS(rs, 2)];
Chris@42 615 TW = Im[WS(rs, 2)];
Chris@42 616 TT = W[8];
Chris@42 617 TV = W[9];
Chris@42 618 TX = FMA(TT, TU, TV * TW);
Chris@42 619 T1Q = FNMS(TV, TU, TT * TW);
Chris@42 620 }
Chris@42 621 TS = TM + TR;
Chris@42 622 T13 = TX + T12;
Chris@42 623 T2w = TS - T13;
Chris@42 624 T2x = T1K + T1L;
Chris@42 625 T2y = T1Q + T1R;
Chris@42 626 T2z = T2x - T2y;
Chris@42 627 {
Chris@42 628 E T1M, T1N, T1P, T1S;
Chris@42 629 T1M = T1K - T1L;
Chris@42 630 T1N = TX - T12;
Chris@42 631 T1O = T1M + T1N;
Chris@42 632 T2g = T1M - T1N;
Chris@42 633 T1P = TM - TR;
Chris@42 634 T1S = T1Q - T1R;
Chris@42 635 T1T = T1P - T1S;
Chris@42 636 T2h = T1P + T1S;
Chris@42 637 }
Chris@42 638 }
Chris@42 639 {
Chris@42 640 E T1J, T27, T3g, T3i, T26, T3h, T2a, T3d;
Chris@42 641 {
Chris@42 642 E T1x, T1I, T3e, T3f;
Chris@42 643 T1x = T1t - T1w;
Chris@42 644 T1I = KP707106781 * (T1C - T1H);
Chris@42 645 T1J = T1x + T1I;
Chris@42 646 T27 = T1x - T1I;
Chris@42 647 T3e = KP707106781 * (T2d - T2c);
Chris@42 648 T3f = T38 + T37;
Chris@42 649 T3g = T3e + T3f;
Chris@42 650 T3i = T3f - T3e;
Chris@42 651 }
Chris@42 652 {
Chris@42 653 E T1U, T25, T28, T29;
Chris@42 654 T1U = FMA(KP923879532, T1O, KP382683432 * T1T);
Chris@42 655 T25 = FNMS(KP923879532, T24, KP382683432 * T1Z);
Chris@42 656 T26 = T1U + T25;
Chris@42 657 T3h = T25 - T1U;
Chris@42 658 T28 = FNMS(KP923879532, T1T, KP382683432 * T1O);
Chris@42 659 T29 = FMA(KP382683432, T24, KP923879532 * T1Z);
Chris@42 660 T2a = T28 - T29;
Chris@42 661 T3d = T28 + T29;
Chris@42 662 }
Chris@42 663 Rm[WS(rs, 4)] = T1J - T26;
Chris@42 664 Im[WS(rs, 4)] = T3d - T3g;
Chris@42 665 Rp[WS(rs, 3)] = T1J + T26;
Chris@42 666 Ip[WS(rs, 3)] = T3d + T3g;
Chris@42 667 Rm[0] = T27 - T2a;
Chris@42 668 Im[0] = T3h - T3i;
Chris@42 669 Rp[WS(rs, 7)] = T27 + T2a;
Chris@42 670 Ip[WS(rs, 7)] = T3h + T3i;
Chris@42 671 }
Chris@42 672 {
Chris@42 673 E T2v, T2H, T32, T34, T2G, T33, T2K, T2Z;
Chris@42 674 {
Chris@42 675 E T2r, T2u, T30, T31;
Chris@42 676 T2r = T7 - Ti;
Chris@42 677 T2u = T2s - T2t;
Chris@42 678 T2v = T2r + T2u;
Chris@42 679 T2H = T2r - T2u;
Chris@42 680 T30 = TF - Tu;
Chris@42 681 T31 = T2U - T2R;
Chris@42 682 T32 = T30 + T31;
Chris@42 683 T34 = T31 - T30;
Chris@42 684 }
Chris@42 685 {
Chris@42 686 E T2A, T2F, T2I, T2J;
Chris@42 687 T2A = T2w + T2z;
Chris@42 688 T2F = T2B - T2E;
Chris@42 689 T2G = KP707106781 * (T2A + T2F);
Chris@42 690 T33 = KP707106781 * (T2F - T2A);
Chris@42 691 T2I = T2z - T2w;
Chris@42 692 T2J = T2B + T2E;
Chris@42 693 T2K = KP707106781 * (T2I - T2J);
Chris@42 694 T2Z = KP707106781 * (T2I + T2J);
Chris@42 695 }
Chris@42 696 Rm[WS(rs, 5)] = T2v - T2G;
Chris@42 697 Im[WS(rs, 5)] = T2Z - T32;
Chris@42 698 Rp[WS(rs, 2)] = T2v + T2G;
Chris@42 699 Ip[WS(rs, 2)] = T2Z + T32;
Chris@42 700 Rm[WS(rs, 1)] = T2H - T2K;
Chris@42 701 Im[WS(rs, 1)] = T33 - T34;
Chris@42 702 Rp[WS(rs, 6)] = T2H + T2K;
Chris@42 703 Ip[WS(rs, 6)] = T33 + T34;
Chris@42 704 }
Chris@42 705 {
Chris@42 706 E T2f, T2n, T3a, T3c, T2m, T3b, T2q, T35;
Chris@42 707 {
Chris@42 708 E T2b, T2e, T36, T39;
Chris@42 709 T2b = T1t + T1w;
Chris@42 710 T2e = KP707106781 * (T2c + T2d);
Chris@42 711 T2f = T2b + T2e;
Chris@42 712 T2n = T2b - T2e;
Chris@42 713 T36 = KP707106781 * (T1C + T1H);
Chris@42 714 T39 = T37 - T38;
Chris@42 715 T3a = T36 + T39;
Chris@42 716 T3c = T39 - T36;
Chris@42 717 }
Chris@42 718 {
Chris@42 719 E T2i, T2l, T2o, T2p;
Chris@42 720 T2i = FMA(KP382683432, T2g, KP923879532 * T2h);
Chris@42 721 T2l = FNMS(KP382683432, T2k, KP923879532 * T2j);
Chris@42 722 T2m = T2i + T2l;
Chris@42 723 T3b = T2l - T2i;
Chris@42 724 T2o = FNMS(KP382683432, T2h, KP923879532 * T2g);
Chris@42 725 T2p = FMA(KP923879532, T2k, KP382683432 * T2j);
Chris@42 726 T2q = T2o - T2p;
Chris@42 727 T35 = T2o + T2p;
Chris@42 728 }
Chris@42 729 Rm[WS(rs, 6)] = T2f - T2m;
Chris@42 730 Im[WS(rs, 6)] = T35 - T3a;
Chris@42 731 Rp[WS(rs, 1)] = T2f + T2m;
Chris@42 732 Ip[WS(rs, 1)] = T35 + T3a;
Chris@42 733 Rm[WS(rs, 2)] = T2n - T2q;
Chris@42 734 Im[WS(rs, 2)] = T3b - T3c;
Chris@42 735 Rp[WS(rs, 5)] = T2n + T2q;
Chris@42 736 Ip[WS(rs, 5)] = T3b + T3c;
Chris@42 737 }
Chris@42 738 {
Chris@42 739 E TH, T2L, T2W, T2Y, T1s, T2X, T2O, T2P;
Chris@42 740 {
Chris@42 741 E Tj, TG, T2Q, T2V;
Chris@42 742 Tj = T7 + Ti;
Chris@42 743 TG = Tu + TF;
Chris@42 744 TH = Tj + TG;
Chris@42 745 T2L = Tj - TG;
Chris@42 746 T2Q = T2s + T2t;
Chris@42 747 T2V = T2R + T2U;
Chris@42 748 T2W = T2Q + T2V;
Chris@42 749 T2Y = T2V - T2Q;
Chris@42 750 }
Chris@42 751 {
Chris@42 752 E T14, T1r, T2M, T2N;
Chris@42 753 T14 = TS + T13;
Chris@42 754 T1r = T1f + T1q;
Chris@42 755 T1s = T14 + T1r;
Chris@42 756 T2X = T1r - T14;
Chris@42 757 T2M = T2x + T2y;
Chris@42 758 T2N = T2C + T2D;
Chris@42 759 T2O = T2M - T2N;
Chris@42 760 T2P = T2M + T2N;
Chris@42 761 }
Chris@42 762 Rm[WS(rs, 7)] = TH - T1s;
Chris@42 763 Im[WS(rs, 7)] = T2P - T2W;
Chris@42 764 Rp[0] = TH + T1s;
Chris@42 765 Ip[0] = T2P + T2W;
Chris@42 766 Rm[WS(rs, 3)] = T2L - T2O;
Chris@42 767 Im[WS(rs, 3)] = T2X - T2Y;
Chris@42 768 Rp[WS(rs, 4)] = T2L + T2O;
Chris@42 769 Ip[WS(rs, 4)] = T2X + T2Y;
Chris@42 770 }
Chris@42 771 }
Chris@42 772 }
Chris@42 773 }
Chris@42 774
Chris@42 775 static const tw_instr twinstr[] = {
Chris@42 776 {TW_FULL, 1, 16},
Chris@42 777 {TW_NEXT, 1, 0}
Chris@42 778 };
Chris@42 779
Chris@42 780 static const hc2c_desc desc = { 16, "hc2cf_16", twinstr, &GENUS, {136, 46, 38, 0} };
Chris@42 781
Chris@42 782 void X(codelet_hc2cf_16) (planner *p) {
Chris@42 783 X(khc2c_register) (p, hc2cf_16, &desc, HC2C_VIA_RDFT);
Chris@42 784 }
Chris@42 785 #endif /* HAVE_FMA */