annotate src/fftw-3.3.3/dft/scalar/codelets/t1_15.c @ 95:89f5e221ed7b

Add FFTW3
author Chris Cannam <cannam@all-day-breakfast.com>
date Wed, 20 Mar 2013 15:35:50 +0000
parents
children
rev   line source
cannam@95 1 /*
cannam@95 2 * Copyright (c) 2003, 2007-11 Matteo Frigo
cannam@95 3 * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
cannam@95 4 *
cannam@95 5 * This program is free software; you can redistribute it and/or modify
cannam@95 6 * it under the terms of the GNU General Public License as published by
cannam@95 7 * the Free Software Foundation; either version 2 of the License, or
cannam@95 8 * (at your option) any later version.
cannam@95 9 *
cannam@95 10 * This program is distributed in the hope that it will be useful,
cannam@95 11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
cannam@95 12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
cannam@95 13 * GNU General Public License for more details.
cannam@95 14 *
cannam@95 15 * You should have received a copy of the GNU General Public License
cannam@95 16 * along with this program; if not, write to the Free Software
cannam@95 17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
cannam@95 18 *
cannam@95 19 */
cannam@95 20
cannam@95 21 /* This file was automatically generated --- DO NOT EDIT */
cannam@95 22 /* Generated on Sun Nov 25 07:35:50 EST 2012 */
cannam@95 23
cannam@95 24 #include "codelet-dft.h"
cannam@95 25
cannam@95 26 #ifdef HAVE_FMA
cannam@95 27
cannam@95 28 /* Generated by: ../../../genfft/gen_twiddle.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -n 15 -name t1_15 -include t.h */
cannam@95 29
cannam@95 30 /*
cannam@95 31 * This function contains 184 FP additions, 140 FP multiplications,
cannam@95 32 * (or, 72 additions, 28 multiplications, 112 fused multiply/add),
cannam@95 33 * 89 stack variables, 6 constants, and 60 memory accesses
cannam@95 34 */
cannam@95 35 #include "t.h"
cannam@95 36
cannam@95 37 static void t1_15(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
cannam@95 38 {
cannam@95 39 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
cannam@95 40 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
cannam@95 41 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
cannam@95 42 DK(KP618033988, +0.618033988749894848204586834365638117720309180);
cannam@95 43 DK(KP866025403, +0.866025403784438646763723170752936183471402627);
cannam@95 44 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
cannam@95 45 {
cannam@95 46 INT m;
cannam@95 47 for (m = mb, W = W + (mb * 28); m < me; m = m + 1, ri = ri + ms, ii = ii + ms, W = W + 28, MAKE_VOLATILE_STRIDE(30, rs)) {
cannam@95 48 E T2d, T2O, T2Q, T2m, T2k, T2l, T2P, T2n;
cannam@95 49 {
cannam@95 50 E T1G, T3u, T3k, T3t, T1B, Tf, T37, T1y, T2V, T2M, T2a, T2i, T39, Tz, T2X;
cannam@95 51 E T2t, T1O, T2e, T3a, TT, T10, T2Y, T2z, T1V, T2f, T2C, T12, T15, T14, T21;
cannam@95 52 E T1c, T1Y, T13;
cannam@95 53 {
cannam@95 54 E T2I, T1k, T1m, T1p, T1o, T28, T1w, T25, T1n;
cannam@95 55 {
cannam@95 56 E T1, T3j, T9, Tc, Tb, T1D, T7, T1E, Ta, T1j, T1i, T1h;
cannam@95 57 T1 = ri[0];
cannam@95 58 T3j = ii[0];
cannam@95 59 {
cannam@95 60 E T3, T6, T2, T5, T1C, T4, T8;
cannam@95 61 T3 = ri[WS(rs, 5)];
cannam@95 62 T6 = ii[WS(rs, 5)];
cannam@95 63 T2 = W[8];
cannam@95 64 T5 = W[9];
cannam@95 65 T9 = ri[WS(rs, 10)];
cannam@95 66 Tc = ii[WS(rs, 10)];
cannam@95 67 T1C = T2 * T6;
cannam@95 68 T4 = T2 * T3;
cannam@95 69 T8 = W[18];
cannam@95 70 Tb = W[19];
cannam@95 71 T1D = FNMS(T5, T3, T1C);
cannam@95 72 T7 = FMA(T5, T6, T4);
cannam@95 73 T1E = T8 * Tc;
cannam@95 74 Ta = T8 * T9;
cannam@95 75 }
cannam@95 76 {
cannam@95 77 E T1g, T1F, Td, T1f, T3i, Te, T2H;
cannam@95 78 T1g = ri[WS(rs, 9)];
cannam@95 79 T1j = ii[WS(rs, 9)];
cannam@95 80 T1F = FNMS(Tb, T9, T1E);
cannam@95 81 Td = FMA(Tb, Tc, Ta);
cannam@95 82 T1f = W[16];
cannam@95 83 T1i = W[17];
cannam@95 84 T1G = T1D - T1F;
cannam@95 85 T3i = T1D + T1F;
cannam@95 86 T3u = Td - T7;
cannam@95 87 Te = T7 + Td;
cannam@95 88 T2H = T1f * T1j;
cannam@95 89 T1h = T1f * T1g;
cannam@95 90 T3k = T3i + T3j;
cannam@95 91 T3t = FNMS(KP500000000, T3i, T3j);
cannam@95 92 T1B = FNMS(KP500000000, Te, T1);
cannam@95 93 Tf = T1 + Te;
cannam@95 94 T2I = FNMS(T1i, T1g, T2H);
cannam@95 95 }
cannam@95 96 T1k = FMA(T1i, T1j, T1h);
cannam@95 97 {
cannam@95 98 E T1s, T1v, T1r, T1u, T27, T1t, T1l;
cannam@95 99 T1s = ri[WS(rs, 4)];
cannam@95 100 T1v = ii[WS(rs, 4)];
cannam@95 101 T1r = W[6];
cannam@95 102 T1u = W[7];
cannam@95 103 T1m = ri[WS(rs, 14)];
cannam@95 104 T1p = ii[WS(rs, 14)];
cannam@95 105 T27 = T1r * T1v;
cannam@95 106 T1t = T1r * T1s;
cannam@95 107 T1l = W[26];
cannam@95 108 T1o = W[27];
cannam@95 109 T28 = FNMS(T1u, T1s, T27);
cannam@95 110 T1w = FMA(T1u, T1v, T1t);
cannam@95 111 T25 = T1l * T1p;
cannam@95 112 T1n = T1l * T1m;
cannam@95 113 }
cannam@95 114 }
cannam@95 115 {
cannam@95 116 E Tl, T2p, Tn, Tq, Tp, T1M, Tx, T1J, To;
cannam@95 117 {
cannam@95 118 E Th, Tk, T26, T1q, Tg, Tj;
cannam@95 119 Th = ri[WS(rs, 3)];
cannam@95 120 Tk = ii[WS(rs, 3)];
cannam@95 121 T26 = FNMS(T1o, T1m, T25);
cannam@95 122 T1q = FMA(T1o, T1p, T1n);
cannam@95 123 Tg = W[4];
cannam@95 124 Tj = W[5];
cannam@95 125 {
cannam@95 126 E T29, T2J, T1x, T2L;
cannam@95 127 T29 = T26 - T28;
cannam@95 128 T2J = T26 + T28;
cannam@95 129 T1x = T1q + T1w;
cannam@95 130 T2L = T1w - T1q;
cannam@95 131 {
cannam@95 132 E T2o, Ti, T2K, T24;
cannam@95 133 T2o = Tg * Tk;
cannam@95 134 Ti = Tg * Th;
cannam@95 135 T2K = FNMS(KP500000000, T2J, T2I);
cannam@95 136 T37 = T2I + T2J;
cannam@95 137 T24 = FNMS(KP500000000, T1x, T1k);
cannam@95 138 T1y = T1k + T1x;
cannam@95 139 Tl = FMA(Tj, Tk, Ti);
cannam@95 140 T2V = FNMS(KP866025403, T2L, T2K);
cannam@95 141 T2M = FMA(KP866025403, T2L, T2K);
cannam@95 142 T2a = FNMS(KP866025403, T29, T24);
cannam@95 143 T2i = FMA(KP866025403, T29, T24);
cannam@95 144 T2p = FNMS(Tj, Th, T2o);
cannam@95 145 }
cannam@95 146 }
cannam@95 147 }
cannam@95 148 {
cannam@95 149 E Tt, Tw, Ts, Tv, T1L, Tu, Tm;
cannam@95 150 Tt = ri[WS(rs, 13)];
cannam@95 151 Tw = ii[WS(rs, 13)];
cannam@95 152 Ts = W[24];
cannam@95 153 Tv = W[25];
cannam@95 154 Tn = ri[WS(rs, 8)];
cannam@95 155 Tq = ii[WS(rs, 8)];
cannam@95 156 T1L = Ts * Tw;
cannam@95 157 Tu = Ts * Tt;
cannam@95 158 Tm = W[14];
cannam@95 159 Tp = W[15];
cannam@95 160 T1M = FNMS(Tv, Tt, T1L);
cannam@95 161 Tx = FMA(Tv, Tw, Tu);
cannam@95 162 T1J = Tm * Tq;
cannam@95 163 To = Tm * Tn;
cannam@95 164 }
cannam@95 165 {
cannam@95 166 E TF, T2v, TH, TK, TJ, T1T, TR, T1Q, TI;
cannam@95 167 {
cannam@95 168 E TB, TE, T1K, Tr, TA, TD;
cannam@95 169 TB = ri[WS(rs, 12)];
cannam@95 170 TE = ii[WS(rs, 12)];
cannam@95 171 T1K = FNMS(Tp, Tn, T1J);
cannam@95 172 Tr = FMA(Tp, Tq, To);
cannam@95 173 TA = W[22];
cannam@95 174 TD = W[23];
cannam@95 175 {
cannam@95 176 E T1N, T2q, Ty, T2s;
cannam@95 177 T1N = T1K - T1M;
cannam@95 178 T2q = T1K + T1M;
cannam@95 179 Ty = Tr + Tx;
cannam@95 180 T2s = Tx - Tr;
cannam@95 181 {
cannam@95 182 E T2u, TC, T2r, T1I;
cannam@95 183 T2u = TA * TE;
cannam@95 184 TC = TA * TB;
cannam@95 185 T2r = FNMS(KP500000000, T2q, T2p);
cannam@95 186 T39 = T2p + T2q;
cannam@95 187 T1I = FNMS(KP500000000, Ty, Tl);
cannam@95 188 Tz = Tl + Ty;
cannam@95 189 TF = FMA(TD, TE, TC);
cannam@95 190 T2X = FNMS(KP866025403, T2s, T2r);
cannam@95 191 T2t = FMA(KP866025403, T2s, T2r);
cannam@95 192 T1O = FNMS(KP866025403, T1N, T1I);
cannam@95 193 T2e = FMA(KP866025403, T1N, T1I);
cannam@95 194 T2v = FNMS(TD, TB, T2u);
cannam@95 195 }
cannam@95 196 }
cannam@95 197 }
cannam@95 198 {
cannam@95 199 E TN, TQ, TM, TP, T1S, TO, TG;
cannam@95 200 TN = ri[WS(rs, 7)];
cannam@95 201 TQ = ii[WS(rs, 7)];
cannam@95 202 TM = W[12];
cannam@95 203 TP = W[13];
cannam@95 204 TH = ri[WS(rs, 2)];
cannam@95 205 TK = ii[WS(rs, 2)];
cannam@95 206 T1S = TM * TQ;
cannam@95 207 TO = TM * TN;
cannam@95 208 TG = W[2];
cannam@95 209 TJ = W[3];
cannam@95 210 T1T = FNMS(TP, TN, T1S);
cannam@95 211 TR = FMA(TP, TQ, TO);
cannam@95 212 T1Q = TG * TK;
cannam@95 213 TI = TG * TH;
cannam@95 214 }
cannam@95 215 {
cannam@95 216 E TW, TZ, T1R, TL, TV, TY;
cannam@95 217 TW = ri[WS(rs, 6)];
cannam@95 218 TZ = ii[WS(rs, 6)];
cannam@95 219 T1R = FNMS(TJ, TH, T1Q);
cannam@95 220 TL = FMA(TJ, TK, TI);
cannam@95 221 TV = W[10];
cannam@95 222 TY = W[11];
cannam@95 223 {
cannam@95 224 E T1U, T2w, TS, T2y;
cannam@95 225 T1U = T1R - T1T;
cannam@95 226 T2w = T1R + T1T;
cannam@95 227 TS = TL + TR;
cannam@95 228 T2y = TR - TL;
cannam@95 229 {
cannam@95 230 E T2B, TX, T2x, T1P;
cannam@95 231 T2B = TV * TZ;
cannam@95 232 TX = TV * TW;
cannam@95 233 T2x = FNMS(KP500000000, T2w, T2v);
cannam@95 234 T3a = T2v + T2w;
cannam@95 235 T1P = FNMS(KP500000000, TS, TF);
cannam@95 236 TT = TF + TS;
cannam@95 237 T10 = FMA(TY, TZ, TX);
cannam@95 238 T2Y = FNMS(KP866025403, T2y, T2x);
cannam@95 239 T2z = FMA(KP866025403, T2y, T2x);
cannam@95 240 T1V = FNMS(KP866025403, T1U, T1P);
cannam@95 241 T2f = FMA(KP866025403, T1U, T1P);
cannam@95 242 T2C = FNMS(TY, TW, T2B);
cannam@95 243 }
cannam@95 244 }
cannam@95 245 }
cannam@95 246 {
cannam@95 247 E T18, T1b, T17, T1a, T20, T19, T11;
cannam@95 248 T18 = ri[WS(rs, 1)];
cannam@95 249 T1b = ii[WS(rs, 1)];
cannam@95 250 T17 = W[0];
cannam@95 251 T1a = W[1];
cannam@95 252 T12 = ri[WS(rs, 11)];
cannam@95 253 T15 = ii[WS(rs, 11)];
cannam@95 254 T20 = T17 * T1b;
cannam@95 255 T19 = T17 * T18;
cannam@95 256 T11 = W[20];
cannam@95 257 T14 = W[21];
cannam@95 258 T21 = FNMS(T1a, T18, T20);
cannam@95 259 T1c = FMA(T1a, T1b, T19);
cannam@95 260 T1Y = T11 * T15;
cannam@95 261 T13 = T11 * T12;
cannam@95 262 }
cannam@95 263 }
cannam@95 264 }
cannam@95 265 }
cannam@95 266 {
cannam@95 267 E T2G, T2h, T3J, T3I, T32, T30, T1H, T1W, T3P, T3O, T2b;
cannam@95 268 {
cannam@95 269 E T3f, T3b, T1Z, T16, T3p, TU;
cannam@95 270 T3f = T39 + T3a;
cannam@95 271 T3b = T39 - T3a;
cannam@95 272 T1Z = FNMS(T14, T12, T1Y);
cannam@95 273 T16 = FMA(T14, T15, T13);
cannam@95 274 T3p = Tz - TT;
cannam@95 275 TU = Tz + TT;
cannam@95 276 {
cannam@95 277 E T3g, T2U, T23, T3c, T3e, T3q, T3s, T1A, T34, T3r, T3n;
cannam@95 278 {
cannam@95 279 E T22, T1d, T2F, T2E, T36, T2D;
cannam@95 280 T22 = T1Z - T21;
cannam@95 281 T2D = T1Z + T21;
cannam@95 282 T1d = T16 + T1c;
cannam@95 283 T2F = T1c - T16;
cannam@95 284 T2E = FNMS(KP500000000, T2D, T2C);
cannam@95 285 T36 = T2C + T2D;
cannam@95 286 {
cannam@95 287 E T1e, T1X, T38, T1z, T3o;
cannam@95 288 T1e = T10 + T1d;
cannam@95 289 T1X = FNMS(KP500000000, T1d, T10);
cannam@95 290 T38 = T36 - T37;
cannam@95 291 T3g = T36 + T37;
cannam@95 292 T2G = FMA(KP866025403, T2F, T2E);
cannam@95 293 T2U = FNMS(KP866025403, T2F, T2E);
cannam@95 294 T1z = T1e + T1y;
cannam@95 295 T3o = T1e - T1y;
cannam@95 296 T2h = FMA(KP866025403, T22, T1X);
cannam@95 297 T23 = FNMS(KP866025403, T22, T1X);
cannam@95 298 T3c = FNMS(KP618033988, T3b, T38);
cannam@95 299 T3e = FMA(KP618033988, T38, T3b);
cannam@95 300 T3q = FNMS(KP618033988, T3p, T3o);
cannam@95 301 T3s = FMA(KP618033988, T3o, T3p);
cannam@95 302 T1A = TU + T1z;
cannam@95 303 T34 = TU - T1z;
cannam@95 304 }
cannam@95 305 }
cannam@95 306 {
cannam@95 307 E T2W, T33, T3m, T3h, T2Z, T3d, T35, T3l;
cannam@95 308 T3J = T2U + T2V;
cannam@95 309 T2W = T2U - T2V;
cannam@95 310 ri[0] = Tf + T1A;
cannam@95 311 T33 = FNMS(KP250000000, T1A, Tf);
cannam@95 312 T3m = T3f - T3g;
cannam@95 313 T3h = T3f + T3g;
cannam@95 314 T2Z = T2X - T2Y;
cannam@95 315 T3I = T2X + T2Y;
cannam@95 316 T3d = FMA(KP559016994, T34, T33);
cannam@95 317 T35 = FNMS(KP559016994, T34, T33);
cannam@95 318 ii[0] = T3h + T3k;
cannam@95 319 T3l = FNMS(KP250000000, T3h, T3k);
cannam@95 320 ri[WS(rs, 3)] = FMA(KP951056516, T3c, T35);
cannam@95 321 ri[WS(rs, 12)] = FNMS(KP951056516, T3c, T35);
cannam@95 322 ri[WS(rs, 6)] = FMA(KP951056516, T3e, T3d);
cannam@95 323 ri[WS(rs, 9)] = FNMS(KP951056516, T3e, T3d);
cannam@95 324 T3r = FMA(KP559016994, T3m, T3l);
cannam@95 325 T3n = FNMS(KP559016994, T3m, T3l);
cannam@95 326 T32 = FMA(KP618033988, T2W, T2Z);
cannam@95 327 T30 = FNMS(KP618033988, T2Z, T2W);
cannam@95 328 }
cannam@95 329 ii[WS(rs, 12)] = FMA(KP951056516, T3q, T3n);
cannam@95 330 ii[WS(rs, 3)] = FNMS(KP951056516, T3q, T3n);
cannam@95 331 ii[WS(rs, 9)] = FMA(KP951056516, T3s, T3r);
cannam@95 332 ii[WS(rs, 6)] = FNMS(KP951056516, T3s, T3r);
cannam@95 333 T2d = FMA(KP866025403, T1G, T1B);
cannam@95 334 T1H = FNMS(KP866025403, T1G, T1B);
cannam@95 335 T1W = T1O + T1V;
cannam@95 336 T3P = T1O - T1V;
cannam@95 337 T3O = T23 - T2a;
cannam@95 338 T2b = T23 + T2a;
cannam@95 339 }
cannam@95 340 }
cannam@95 341 {
cannam@95 342 E T3H, T3v, T2S, T3Q, T3S, T2R, T2c;
cannam@95 343 T3H = FNMS(KP866025403, T3u, T3t);
cannam@95 344 T3v = FMA(KP866025403, T3u, T3t);
cannam@95 345 T2c = T1W + T2b;
cannam@95 346 T2S = T1W - T2b;
cannam@95 347 T3Q = FNMS(KP618033988, T3P, T3O);
cannam@95 348 T3S = FMA(KP618033988, T3O, T3P);
cannam@95 349 ri[WS(rs, 5)] = T1H + T2c;
cannam@95 350 T2R = FNMS(KP250000000, T2c, T1H);
cannam@95 351 {
cannam@95 352 E T2g, T2j, T3G, T3E, T2A, T2N, T3y, T3A, T3M, T3L, T3z, T3F, T3B;
cannam@95 353 {
cannam@95 354 E T3C, T3D, T31, T2T, T3K;
cannam@95 355 T2g = T2e + T2f;
cannam@95 356 T3C = T2e - T2f;
cannam@95 357 T3D = T2h - T2i;
cannam@95 358 T2j = T2h + T2i;
cannam@95 359 T31 = FMA(KP559016994, T2S, T2R);
cannam@95 360 T2T = FNMS(KP559016994, T2S, T2R);
cannam@95 361 T3K = T3I + T3J;
cannam@95 362 T3M = T3I - T3J;
cannam@95 363 ri[WS(rs, 8)] = FMA(KP951056516, T30, T2T);
cannam@95 364 ri[WS(rs, 2)] = FNMS(KP951056516, T30, T2T);
cannam@95 365 ri[WS(rs, 11)] = FMA(KP951056516, T32, T31);
cannam@95 366 ri[WS(rs, 14)] = FNMS(KP951056516, T32, T31);
cannam@95 367 ii[WS(rs, 5)] = T3K + T3H;
cannam@95 368 T3L = FNMS(KP250000000, T3K, T3H);
cannam@95 369 T3G = FNMS(KP618033988, T3C, T3D);
cannam@95 370 T3E = FMA(KP618033988, T3D, T3C);
cannam@95 371 }
cannam@95 372 {
cannam@95 373 E T3N, T3R, T3w, T3x;
cannam@95 374 T3N = FNMS(KP559016994, T3M, T3L);
cannam@95 375 T3R = FMA(KP559016994, T3M, T3L);
cannam@95 376 T3w = T2t + T2z;
cannam@95 377 T2A = T2t - T2z;
cannam@95 378 T2N = T2G - T2M;
cannam@95 379 T3x = T2G + T2M;
cannam@95 380 ii[WS(rs, 8)] = FNMS(KP951056516, T3Q, T3N);
cannam@95 381 ii[WS(rs, 2)] = FMA(KP951056516, T3Q, T3N);
cannam@95 382 ii[WS(rs, 14)] = FMA(KP951056516, T3S, T3R);
cannam@95 383 ii[WS(rs, 11)] = FNMS(KP951056516, T3S, T3R);
cannam@95 384 T3y = T3w + T3x;
cannam@95 385 T3A = T3w - T3x;
cannam@95 386 }
cannam@95 387 ii[WS(rs, 10)] = T3y + T3v;
cannam@95 388 T3z = FNMS(KP250000000, T3y, T3v);
cannam@95 389 T2O = FMA(KP618033988, T2N, T2A);
cannam@95 390 T2Q = FNMS(KP618033988, T2A, T2N);
cannam@95 391 T3F = FNMS(KP559016994, T3A, T3z);
cannam@95 392 T3B = FMA(KP559016994, T3A, T3z);
cannam@95 393 ii[WS(rs, 4)] = FMA(KP951056516, T3E, T3B);
cannam@95 394 ii[WS(rs, 1)] = FNMS(KP951056516, T3E, T3B);
cannam@95 395 ii[WS(rs, 13)] = FNMS(KP951056516, T3G, T3F);
cannam@95 396 ii[WS(rs, 7)] = FMA(KP951056516, T3G, T3F);
cannam@95 397 T2m = T2g - T2j;
cannam@95 398 T2k = T2g + T2j;
cannam@95 399 }
cannam@95 400 }
cannam@95 401 }
cannam@95 402 }
cannam@95 403 ri[WS(rs, 10)] = T2d + T2k;
cannam@95 404 T2l = FNMS(KP250000000, T2k, T2d);
cannam@95 405 T2P = FNMS(KP559016994, T2m, T2l);
cannam@95 406 T2n = FMA(KP559016994, T2m, T2l);
cannam@95 407 ri[WS(rs, 1)] = FMA(KP951056516, T2O, T2n);
cannam@95 408 ri[WS(rs, 4)] = FNMS(KP951056516, T2O, T2n);
cannam@95 409 ri[WS(rs, 13)] = FMA(KP951056516, T2Q, T2P);
cannam@95 410 ri[WS(rs, 7)] = FNMS(KP951056516, T2Q, T2P);
cannam@95 411 }
cannam@95 412 }
cannam@95 413 }
cannam@95 414
cannam@95 415 static const tw_instr twinstr[] = {
cannam@95 416 {TW_FULL, 0, 15},
cannam@95 417 {TW_NEXT, 1, 0}
cannam@95 418 };
cannam@95 419
cannam@95 420 static const ct_desc desc = { 15, "t1_15", twinstr, &GENUS, {72, 28, 112, 0}, 0, 0, 0 };
cannam@95 421
cannam@95 422 void X(codelet_t1_15) (planner *p) {
cannam@95 423 X(kdft_dit_register) (p, t1_15, &desc);
cannam@95 424 }
cannam@95 425 #else /* HAVE_FMA */
cannam@95 426
cannam@95 427 /* Generated by: ../../../genfft/gen_twiddle.native -compact -variables 4 -pipeline-latency 4 -n 15 -name t1_15 -include t.h */
cannam@95 428
cannam@95 429 /*
cannam@95 430 * This function contains 184 FP additions, 112 FP multiplications,
cannam@95 431 * (or, 128 additions, 56 multiplications, 56 fused multiply/add),
cannam@95 432 * 65 stack variables, 6 constants, and 60 memory accesses
cannam@95 433 */
cannam@95 434 #include "t.h"
cannam@95 435
cannam@95 436 static void t1_15(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
cannam@95 437 {
cannam@95 438 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
cannam@95 439 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
cannam@95 440 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
cannam@95 441 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
cannam@95 442 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
cannam@95 443 DK(KP866025403, +0.866025403784438646763723170752936183471402627);
cannam@95 444 {
cannam@95 445 INT m;
cannam@95 446 for (m = mb, W = W + (mb * 28); m < me; m = m + 1, ri = ri + ms, ii = ii + ms, W = W + 28, MAKE_VOLATILE_STRIDE(30, rs)) {
cannam@95 447 E T1q, T34, Td, T1n, T2S, T35, T13, T1k, T1l, T2E, T2F, T2O, T1H, T1T, T2k;
cannam@95 448 E T2t, T2f, T2s, T1M, T1U, Tu, TL, TM, T2H, T2I, T2N, T1w, T1Q, T29, T2w;
cannam@95 449 E T24, T2v, T1B, T1R;
cannam@95 450 {
cannam@95 451 E T1, T2R, T6, T1o, Tb, T1p, Tc, T2Q;
cannam@95 452 T1 = ri[0];
cannam@95 453 T2R = ii[0];
cannam@95 454 {
cannam@95 455 E T3, T5, T2, T4;
cannam@95 456 T3 = ri[WS(rs, 5)];
cannam@95 457 T5 = ii[WS(rs, 5)];
cannam@95 458 T2 = W[8];
cannam@95 459 T4 = W[9];
cannam@95 460 T6 = FMA(T2, T3, T4 * T5);
cannam@95 461 T1o = FNMS(T4, T3, T2 * T5);
cannam@95 462 }
cannam@95 463 {
cannam@95 464 E T8, Ta, T7, T9;
cannam@95 465 T8 = ri[WS(rs, 10)];
cannam@95 466 Ta = ii[WS(rs, 10)];
cannam@95 467 T7 = W[18];
cannam@95 468 T9 = W[19];
cannam@95 469 Tb = FMA(T7, T8, T9 * Ta);
cannam@95 470 T1p = FNMS(T9, T8, T7 * Ta);
cannam@95 471 }
cannam@95 472 T1q = KP866025403 * (T1o - T1p);
cannam@95 473 T34 = KP866025403 * (Tb - T6);
cannam@95 474 Tc = T6 + Tb;
cannam@95 475 Td = T1 + Tc;
cannam@95 476 T1n = FNMS(KP500000000, Tc, T1);
cannam@95 477 T2Q = T1o + T1p;
cannam@95 478 T2S = T2Q + T2R;
cannam@95 479 T35 = FNMS(KP500000000, T2Q, T2R);
cannam@95 480 }
cannam@95 481 {
cannam@95 482 E TR, T2c, T18, T2h, TW, T1E, T11, T1F, T12, T2d, T1d, T1J, T1i, T1K, T1j;
cannam@95 483 E T2i;
cannam@95 484 {
cannam@95 485 E TO, TQ, TN, TP;
cannam@95 486 TO = ri[WS(rs, 6)];
cannam@95 487 TQ = ii[WS(rs, 6)];
cannam@95 488 TN = W[10];
cannam@95 489 TP = W[11];
cannam@95 490 TR = FMA(TN, TO, TP * TQ);
cannam@95 491 T2c = FNMS(TP, TO, TN * TQ);
cannam@95 492 }
cannam@95 493 {
cannam@95 494 E T15, T17, T14, T16;
cannam@95 495 T15 = ri[WS(rs, 9)];
cannam@95 496 T17 = ii[WS(rs, 9)];
cannam@95 497 T14 = W[16];
cannam@95 498 T16 = W[17];
cannam@95 499 T18 = FMA(T14, T15, T16 * T17);
cannam@95 500 T2h = FNMS(T16, T15, T14 * T17);
cannam@95 501 }
cannam@95 502 {
cannam@95 503 E TT, TV, TS, TU;
cannam@95 504 TT = ri[WS(rs, 11)];
cannam@95 505 TV = ii[WS(rs, 11)];
cannam@95 506 TS = W[20];
cannam@95 507 TU = W[21];
cannam@95 508 TW = FMA(TS, TT, TU * TV);
cannam@95 509 T1E = FNMS(TU, TT, TS * TV);
cannam@95 510 }
cannam@95 511 {
cannam@95 512 E TY, T10, TX, TZ;
cannam@95 513 TY = ri[WS(rs, 1)];
cannam@95 514 T10 = ii[WS(rs, 1)];
cannam@95 515 TX = W[0];
cannam@95 516 TZ = W[1];
cannam@95 517 T11 = FMA(TX, TY, TZ * T10);
cannam@95 518 T1F = FNMS(TZ, TY, TX * T10);
cannam@95 519 }
cannam@95 520 T12 = TW + T11;
cannam@95 521 T2d = T1E + T1F;
cannam@95 522 {
cannam@95 523 E T1a, T1c, T19, T1b;
cannam@95 524 T1a = ri[WS(rs, 14)];
cannam@95 525 T1c = ii[WS(rs, 14)];
cannam@95 526 T19 = W[26];
cannam@95 527 T1b = W[27];
cannam@95 528 T1d = FMA(T19, T1a, T1b * T1c);
cannam@95 529 T1J = FNMS(T1b, T1a, T19 * T1c);
cannam@95 530 }
cannam@95 531 {
cannam@95 532 E T1f, T1h, T1e, T1g;
cannam@95 533 T1f = ri[WS(rs, 4)];
cannam@95 534 T1h = ii[WS(rs, 4)];
cannam@95 535 T1e = W[6];
cannam@95 536 T1g = W[7];
cannam@95 537 T1i = FMA(T1e, T1f, T1g * T1h);
cannam@95 538 T1K = FNMS(T1g, T1f, T1e * T1h);
cannam@95 539 }
cannam@95 540 T1j = T1d + T1i;
cannam@95 541 T2i = T1J + T1K;
cannam@95 542 {
cannam@95 543 E T1D, T1G, T2g, T2j;
cannam@95 544 T13 = TR + T12;
cannam@95 545 T1k = T18 + T1j;
cannam@95 546 T1l = T13 + T1k;
cannam@95 547 T2E = T2c + T2d;
cannam@95 548 T2F = T2h + T2i;
cannam@95 549 T2O = T2E + T2F;
cannam@95 550 T1D = FNMS(KP500000000, T12, TR);
cannam@95 551 T1G = KP866025403 * (T1E - T1F);
cannam@95 552 T1H = T1D - T1G;
cannam@95 553 T1T = T1D + T1G;
cannam@95 554 T2g = KP866025403 * (T1i - T1d);
cannam@95 555 T2j = FNMS(KP500000000, T2i, T2h);
cannam@95 556 T2k = T2g + T2j;
cannam@95 557 T2t = T2j - T2g;
cannam@95 558 {
cannam@95 559 E T2b, T2e, T1I, T1L;
cannam@95 560 T2b = KP866025403 * (T11 - TW);
cannam@95 561 T2e = FNMS(KP500000000, T2d, T2c);
cannam@95 562 T2f = T2b + T2e;
cannam@95 563 T2s = T2e - T2b;
cannam@95 564 T1I = FNMS(KP500000000, T1j, T18);
cannam@95 565 T1L = KP866025403 * (T1J - T1K);
cannam@95 566 T1M = T1I - T1L;
cannam@95 567 T1U = T1I + T1L;
cannam@95 568 }
cannam@95 569 }
cannam@95 570 }
cannam@95 571 {
cannam@95 572 E Ti, T21, Tz, T26, Tn, T1t, Ts, T1u, Tt, T22, TE, T1y, TJ, T1z, TK;
cannam@95 573 E T27;
cannam@95 574 {
cannam@95 575 E Tf, Th, Te, Tg;
cannam@95 576 Tf = ri[WS(rs, 3)];
cannam@95 577 Th = ii[WS(rs, 3)];
cannam@95 578 Te = W[4];
cannam@95 579 Tg = W[5];
cannam@95 580 Ti = FMA(Te, Tf, Tg * Th);
cannam@95 581 T21 = FNMS(Tg, Tf, Te * Th);
cannam@95 582 }
cannam@95 583 {
cannam@95 584 E Tw, Ty, Tv, Tx;
cannam@95 585 Tw = ri[WS(rs, 12)];
cannam@95 586 Ty = ii[WS(rs, 12)];
cannam@95 587 Tv = W[22];
cannam@95 588 Tx = W[23];
cannam@95 589 Tz = FMA(Tv, Tw, Tx * Ty);
cannam@95 590 T26 = FNMS(Tx, Tw, Tv * Ty);
cannam@95 591 }
cannam@95 592 {
cannam@95 593 E Tk, Tm, Tj, Tl;
cannam@95 594 Tk = ri[WS(rs, 8)];
cannam@95 595 Tm = ii[WS(rs, 8)];
cannam@95 596 Tj = W[14];
cannam@95 597 Tl = W[15];
cannam@95 598 Tn = FMA(Tj, Tk, Tl * Tm);
cannam@95 599 T1t = FNMS(Tl, Tk, Tj * Tm);
cannam@95 600 }
cannam@95 601 {
cannam@95 602 E Tp, Tr, To, Tq;
cannam@95 603 Tp = ri[WS(rs, 13)];
cannam@95 604 Tr = ii[WS(rs, 13)];
cannam@95 605 To = W[24];
cannam@95 606 Tq = W[25];
cannam@95 607 Ts = FMA(To, Tp, Tq * Tr);
cannam@95 608 T1u = FNMS(Tq, Tp, To * Tr);
cannam@95 609 }
cannam@95 610 Tt = Tn + Ts;
cannam@95 611 T22 = T1t + T1u;
cannam@95 612 {
cannam@95 613 E TB, TD, TA, TC;
cannam@95 614 TB = ri[WS(rs, 2)];
cannam@95 615 TD = ii[WS(rs, 2)];
cannam@95 616 TA = W[2];
cannam@95 617 TC = W[3];
cannam@95 618 TE = FMA(TA, TB, TC * TD);
cannam@95 619 T1y = FNMS(TC, TB, TA * TD);
cannam@95 620 }
cannam@95 621 {
cannam@95 622 E TG, TI, TF, TH;
cannam@95 623 TG = ri[WS(rs, 7)];
cannam@95 624 TI = ii[WS(rs, 7)];
cannam@95 625 TF = W[12];
cannam@95 626 TH = W[13];
cannam@95 627 TJ = FMA(TF, TG, TH * TI);
cannam@95 628 T1z = FNMS(TH, TG, TF * TI);
cannam@95 629 }
cannam@95 630 TK = TE + TJ;
cannam@95 631 T27 = T1y + T1z;
cannam@95 632 {
cannam@95 633 E T1s, T1v, T25, T28;
cannam@95 634 Tu = Ti + Tt;
cannam@95 635 TL = Tz + TK;
cannam@95 636 TM = Tu + TL;
cannam@95 637 T2H = T21 + T22;
cannam@95 638 T2I = T26 + T27;
cannam@95 639 T2N = T2H + T2I;
cannam@95 640 T1s = FNMS(KP500000000, Tt, Ti);
cannam@95 641 T1v = KP866025403 * (T1t - T1u);
cannam@95 642 T1w = T1s - T1v;
cannam@95 643 T1Q = T1s + T1v;
cannam@95 644 T25 = KP866025403 * (TJ - TE);
cannam@95 645 T28 = FNMS(KP500000000, T27, T26);
cannam@95 646 T29 = T25 + T28;
cannam@95 647 T2w = T28 - T25;
cannam@95 648 {
cannam@95 649 E T20, T23, T1x, T1A;
cannam@95 650 T20 = KP866025403 * (Ts - Tn);
cannam@95 651 T23 = FNMS(KP500000000, T22, T21);
cannam@95 652 T24 = T20 + T23;
cannam@95 653 T2v = T23 - T20;
cannam@95 654 T1x = FNMS(KP500000000, TK, Tz);
cannam@95 655 T1A = KP866025403 * (T1y - T1z);
cannam@95 656 T1B = T1x - T1A;
cannam@95 657 T1R = T1x + T1A;
cannam@95 658 }
cannam@95 659 }
cannam@95 660 }
cannam@95 661 {
cannam@95 662 E T2C, T1m, T2B, T2K, T2M, T2G, T2J, T2L, T2D;
cannam@95 663 T2C = KP559016994 * (TM - T1l);
cannam@95 664 T1m = TM + T1l;
cannam@95 665 T2B = FNMS(KP250000000, T1m, Td);
cannam@95 666 T2G = T2E - T2F;
cannam@95 667 T2J = T2H - T2I;
cannam@95 668 T2K = FNMS(KP587785252, T2J, KP951056516 * T2G);
cannam@95 669 T2M = FMA(KP951056516, T2J, KP587785252 * T2G);
cannam@95 670 ri[0] = Td + T1m;
cannam@95 671 T2L = T2C + T2B;
cannam@95 672 ri[WS(rs, 9)] = T2L - T2M;
cannam@95 673 ri[WS(rs, 6)] = T2L + T2M;
cannam@95 674 T2D = T2B - T2C;
cannam@95 675 ri[WS(rs, 12)] = T2D - T2K;
cannam@95 676 ri[WS(rs, 3)] = T2D + T2K;
cannam@95 677 }
cannam@95 678 {
cannam@95 679 E T2U, T2P, T2T, T2Y, T30, T2W, T2X, T2Z, T2V;
cannam@95 680 T2U = KP559016994 * (T2N - T2O);
cannam@95 681 T2P = T2N + T2O;
cannam@95 682 T2T = FNMS(KP250000000, T2P, T2S);
cannam@95 683 T2W = T13 - T1k;
cannam@95 684 T2X = Tu - TL;
cannam@95 685 T2Y = FNMS(KP587785252, T2X, KP951056516 * T2W);
cannam@95 686 T30 = FMA(KP951056516, T2X, KP587785252 * T2W);
cannam@95 687 ii[0] = T2P + T2S;
cannam@95 688 T2Z = T2U + T2T;
cannam@95 689 ii[WS(rs, 6)] = T2Z - T30;
cannam@95 690 ii[WS(rs, 9)] = T30 + T2Z;
cannam@95 691 T2V = T2T - T2U;
cannam@95 692 ii[WS(rs, 3)] = T2V - T2Y;
cannam@95 693 ii[WS(rs, 12)] = T2Y + T2V;
cannam@95 694 }
cannam@95 695 {
cannam@95 696 E T2y, T2A, T1r, T1O, T2p, T2q, T2z, T2r;
cannam@95 697 {
cannam@95 698 E T2u, T2x, T1C, T1N;
cannam@95 699 T2u = T2s - T2t;
cannam@95 700 T2x = T2v - T2w;
cannam@95 701 T2y = FNMS(KP587785252, T2x, KP951056516 * T2u);
cannam@95 702 T2A = FMA(KP951056516, T2x, KP587785252 * T2u);
cannam@95 703 T1r = T1n - T1q;
cannam@95 704 T1C = T1w + T1B;
cannam@95 705 T1N = T1H + T1M;
cannam@95 706 T1O = T1C + T1N;
cannam@95 707 T2p = FNMS(KP250000000, T1O, T1r);
cannam@95 708 T2q = KP559016994 * (T1C - T1N);
cannam@95 709 }
cannam@95 710 ri[WS(rs, 5)] = T1r + T1O;
cannam@95 711 T2z = T2q + T2p;
cannam@95 712 ri[WS(rs, 14)] = T2z - T2A;
cannam@95 713 ri[WS(rs, 11)] = T2z + T2A;
cannam@95 714 T2r = T2p - T2q;
cannam@95 715 ri[WS(rs, 2)] = T2r - T2y;
cannam@95 716 ri[WS(rs, 8)] = T2r + T2y;
cannam@95 717 }
cannam@95 718 {
cannam@95 719 E T3h, T3q, T3i, T3l, T3m, T3n, T3p, T3o;
cannam@95 720 {
cannam@95 721 E T3f, T3g, T3j, T3k;
cannam@95 722 T3f = T1H - T1M;
cannam@95 723 T3g = T1w - T1B;
cannam@95 724 T3h = FNMS(KP587785252, T3g, KP951056516 * T3f);
cannam@95 725 T3q = FMA(KP951056516, T3g, KP587785252 * T3f);
cannam@95 726 T3i = T35 - T34;
cannam@95 727 T3j = T2v + T2w;
cannam@95 728 T3k = T2s + T2t;
cannam@95 729 T3l = T3j + T3k;
cannam@95 730 T3m = FNMS(KP250000000, T3l, T3i);
cannam@95 731 T3n = KP559016994 * (T3j - T3k);
cannam@95 732 }
cannam@95 733 ii[WS(rs, 5)] = T3l + T3i;
cannam@95 734 T3p = T3n + T3m;
cannam@95 735 ii[WS(rs, 11)] = T3p - T3q;
cannam@95 736 ii[WS(rs, 14)] = T3q + T3p;
cannam@95 737 T3o = T3m - T3n;
cannam@95 738 ii[WS(rs, 2)] = T3h + T3o;
cannam@95 739 ii[WS(rs, 8)] = T3o - T3h;
cannam@95 740 }
cannam@95 741 {
cannam@95 742 E T3c, T3d, T36, T37, T33, T38, T3e, T39;
cannam@95 743 {
cannam@95 744 E T3a, T3b, T31, T32;
cannam@95 745 T3a = T1Q - T1R;
cannam@95 746 T3b = T1T - T1U;
cannam@95 747 T3c = FMA(KP951056516, T3a, KP587785252 * T3b);
cannam@95 748 T3d = FNMS(KP587785252, T3a, KP951056516 * T3b);
cannam@95 749 T36 = T34 + T35;
cannam@95 750 T31 = T24 + T29;
cannam@95 751 T32 = T2f + T2k;
cannam@95 752 T37 = T31 + T32;
cannam@95 753 T33 = KP559016994 * (T31 - T32);
cannam@95 754 T38 = FNMS(KP250000000, T37, T36);
cannam@95 755 }
cannam@95 756 ii[WS(rs, 10)] = T37 + T36;
cannam@95 757 T3e = T38 - T33;
cannam@95 758 ii[WS(rs, 7)] = T3d + T3e;
cannam@95 759 ii[WS(rs, 13)] = T3e - T3d;
cannam@95 760 T39 = T33 + T38;
cannam@95 761 ii[WS(rs, 1)] = T39 - T3c;
cannam@95 762 ii[WS(rs, 4)] = T3c + T39;
cannam@95 763 }
cannam@95 764 {
cannam@95 765 E T2m, T2o, T1P, T1W, T1X, T1Y, T2n, T1Z;
cannam@95 766 {
cannam@95 767 E T2a, T2l, T1S, T1V;
cannam@95 768 T2a = T24 - T29;
cannam@95 769 T2l = T2f - T2k;
cannam@95 770 T2m = FMA(KP951056516, T2a, KP587785252 * T2l);
cannam@95 771 T2o = FNMS(KP587785252, T2a, KP951056516 * T2l);
cannam@95 772 T1P = T1n + T1q;
cannam@95 773 T1S = T1Q + T1R;
cannam@95 774 T1V = T1T + T1U;
cannam@95 775 T1W = T1S + T1V;
cannam@95 776 T1X = KP559016994 * (T1S - T1V);
cannam@95 777 T1Y = FNMS(KP250000000, T1W, T1P);
cannam@95 778 }
cannam@95 779 ri[WS(rs, 10)] = T1P + T1W;
cannam@95 780 T2n = T1Y - T1X;
cannam@95 781 ri[WS(rs, 7)] = T2n - T2o;
cannam@95 782 ri[WS(rs, 13)] = T2n + T2o;
cannam@95 783 T1Z = T1X + T1Y;
cannam@95 784 ri[WS(rs, 4)] = T1Z - T2m;
cannam@95 785 ri[WS(rs, 1)] = T1Z + T2m;
cannam@95 786 }
cannam@95 787 }
cannam@95 788 }
cannam@95 789 }
cannam@95 790
cannam@95 791 static const tw_instr twinstr[] = {
cannam@95 792 {TW_FULL, 0, 15},
cannam@95 793 {TW_NEXT, 1, 0}
cannam@95 794 };
cannam@95 795
cannam@95 796 static const ct_desc desc = { 15, "t1_15", twinstr, &GENUS, {128, 56, 56, 0}, 0, 0, 0 };
cannam@95 797
cannam@95 798 void X(codelet_t1_15) (planner *p) {
cannam@95 799 X(kdft_dit_register) (p, t1_15, &desc);
cannam@95 800 }
cannam@95 801 #endif /* HAVE_FMA */