annotate src/fftw-3.3.3/dft/scalar/codelets/t1_20.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 07:35:53 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 20 -name t1_20 -include t.h */
cannam@95 29
cannam@95 30 /*
cannam@95 31 * This function contains 246 FP additions, 148 FP multiplications,
cannam@95 32 * (or, 136 additions, 38 multiplications, 110 fused multiply/add),
cannam@95 33 * 97 stack variables, 4 constants, and 80 memory accesses
cannam@95 34 */
cannam@95 35 #include "t.h"
cannam@95 36
cannam@95 37 static void t1_20(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 {
cannam@95 44 INT m;
cannam@95 45 for (m = mb, W = W + (mb * 38); m < me; m = m + 1, ri = ri + ms, ii = ii + ms, W = W + 38, MAKE_VOLATILE_STRIDE(40, rs)) {
cannam@95 46 E T4P, T4Y, T50, T4U, T4S, T4T, T4Z, T4V;
cannam@95 47 {
cannam@95 48 E T4N, T4r, T8, T2i, T4n, T2n, T4O, Tl, T2v, T3v, T40, T4b, TN, T2b, T3F;
cannam@95 49 E T3i, T2R, T3z, T3W, T4f, T27, T2f, T3J, T3a, T2K, T3y, T3T, T4e, T1G, T2e;
cannam@95 50 E T3I, T33, T2C, T3w, T43, T4c, T1e, T2c, T3G, T3p;
cannam@95 51 {
cannam@95 52 E T1, T4q, T3, T6, T2, T5;
cannam@95 53 T1 = ri[0];
cannam@95 54 T4q = ii[0];
cannam@95 55 T3 = ri[WS(rs, 10)];
cannam@95 56 T6 = ii[WS(rs, 10)];
cannam@95 57 T2 = W[18];
cannam@95 58 T5 = W[19];
cannam@95 59 {
cannam@95 60 E Ta, Td, Tg, T2j, Tb, Tj, Tf, Tc, Ti;
cannam@95 61 {
cannam@95 62 E T4o, T4, T9, T4p, T7;
cannam@95 63 Ta = ri[WS(rs, 5)];
cannam@95 64 Td = ii[WS(rs, 5)];
cannam@95 65 T4o = T2 * T6;
cannam@95 66 T4 = T2 * T3;
cannam@95 67 T9 = W[8];
cannam@95 68 Tg = ri[WS(rs, 15)];
cannam@95 69 T4p = FNMS(T5, T3, T4o);
cannam@95 70 T7 = FMA(T5, T6, T4);
cannam@95 71 T2j = T9 * Td;
cannam@95 72 Tb = T9 * Ta;
cannam@95 73 T4N = T4q - T4p;
cannam@95 74 T4r = T4p + T4q;
cannam@95 75 T8 = T1 + T7;
cannam@95 76 T2i = T1 - T7;
cannam@95 77 Tj = ii[WS(rs, 15)];
cannam@95 78 Tf = W[28];
cannam@95 79 }
cannam@95 80 Tc = W[9];
cannam@95 81 Ti = W[29];
cannam@95 82 {
cannam@95 83 E T3d, Ts, T2t, TL, TB, TE, TD, T3f, Ty, T2q, TC;
cannam@95 84 {
cannam@95 85 E TH, TK, TJ, T2s, TI;
cannam@95 86 {
cannam@95 87 E To, Tr, Tp, T3c, Tq, TG;
cannam@95 88 {
cannam@95 89 E T2k, Te, T2m, Tk, T2l, Th, Tn;
cannam@95 90 To = ri[WS(rs, 4)];
cannam@95 91 T2l = Tf * Tj;
cannam@95 92 Th = Tf * Tg;
cannam@95 93 T2k = FNMS(Tc, Ta, T2j);
cannam@95 94 Te = FMA(Tc, Td, Tb);
cannam@95 95 T2m = FNMS(Ti, Tg, T2l);
cannam@95 96 Tk = FMA(Ti, Tj, Th);
cannam@95 97 Tr = ii[WS(rs, 4)];
cannam@95 98 Tn = W[6];
cannam@95 99 T4n = T2k + T2m;
cannam@95 100 T2n = T2k - T2m;
cannam@95 101 T4O = Te - Tk;
cannam@95 102 Tl = Te + Tk;
cannam@95 103 Tp = Tn * To;
cannam@95 104 T3c = Tn * Tr;
cannam@95 105 }
cannam@95 106 Tq = W[7];
cannam@95 107 TH = ri[WS(rs, 19)];
cannam@95 108 TK = ii[WS(rs, 19)];
cannam@95 109 TG = W[36];
cannam@95 110 T3d = FNMS(Tq, To, T3c);
cannam@95 111 Ts = FMA(Tq, Tr, Tp);
cannam@95 112 TJ = W[37];
cannam@95 113 T2s = TG * TK;
cannam@95 114 TI = TG * TH;
cannam@95 115 }
cannam@95 116 {
cannam@95 117 E Tu, Tx, Tt, Tw, T3e, Tv, TA;
cannam@95 118 Tu = ri[WS(rs, 14)];
cannam@95 119 Tx = ii[WS(rs, 14)];
cannam@95 120 T2t = FNMS(TJ, TH, T2s);
cannam@95 121 TL = FMA(TJ, TK, TI);
cannam@95 122 Tt = W[26];
cannam@95 123 Tw = W[27];
cannam@95 124 TB = ri[WS(rs, 9)];
cannam@95 125 TE = ii[WS(rs, 9)];
cannam@95 126 T3e = Tt * Tx;
cannam@95 127 Tv = Tt * Tu;
cannam@95 128 TA = W[16];
cannam@95 129 TD = W[17];
cannam@95 130 T3f = FNMS(Tw, Tu, T3e);
cannam@95 131 Ty = FMA(Tw, Tx, Tv);
cannam@95 132 T2q = TA * TE;
cannam@95 133 TC = TA * TB;
cannam@95 134 }
cannam@95 135 }
cannam@95 136 {
cannam@95 137 E T3g, T3Y, Tz, T2p, T2r, TF;
cannam@95 138 T3g = T3d - T3f;
cannam@95 139 T3Y = T3d + T3f;
cannam@95 140 Tz = Ts + Ty;
cannam@95 141 T2p = Ts - Ty;
cannam@95 142 T2r = FNMS(TD, TB, T2q);
cannam@95 143 TF = FMA(TD, TE, TC);
cannam@95 144 {
cannam@95 145 E T3Z, T2u, T3h, TM;
cannam@95 146 T3Z = T2r + T2t;
cannam@95 147 T2u = T2r - T2t;
cannam@95 148 T3h = TF - TL;
cannam@95 149 TM = TF + TL;
cannam@95 150 T2v = T2p - T2u;
cannam@95 151 T3v = T2p + T2u;
cannam@95 152 T40 = T3Y - T3Z;
cannam@95 153 T4b = T3Y + T3Z;
cannam@95 154 TN = Tz - TM;
cannam@95 155 T2b = Tz + TM;
cannam@95 156 T3F = T3g - T3h;
cannam@95 157 T3i = T3g + T3h;
cannam@95 158 }
cannam@95 159 }
cannam@95 160 }
cannam@95 161 }
cannam@95 162 }
cannam@95 163 {
cannam@95 164 E T35, T1M, T2P, T25, T1V, T1Y, T1X, T37, T1S, T2M, T1W;
cannam@95 165 {
cannam@95 166 E T21, T24, T23, T2O, T22;
cannam@95 167 {
cannam@95 168 E T1I, T1L, T1H, T1K, T34, T1J, T20;
cannam@95 169 T1I = ri[WS(rs, 12)];
cannam@95 170 T1L = ii[WS(rs, 12)];
cannam@95 171 T1H = W[22];
cannam@95 172 T1K = W[23];
cannam@95 173 T21 = ri[WS(rs, 7)];
cannam@95 174 T24 = ii[WS(rs, 7)];
cannam@95 175 T34 = T1H * T1L;
cannam@95 176 T1J = T1H * T1I;
cannam@95 177 T20 = W[12];
cannam@95 178 T23 = W[13];
cannam@95 179 T35 = FNMS(T1K, T1I, T34);
cannam@95 180 T1M = FMA(T1K, T1L, T1J);
cannam@95 181 T2O = T20 * T24;
cannam@95 182 T22 = T20 * T21;
cannam@95 183 }
cannam@95 184 {
cannam@95 185 E T1O, T1R, T1N, T1Q, T36, T1P, T1U;
cannam@95 186 T1O = ri[WS(rs, 2)];
cannam@95 187 T1R = ii[WS(rs, 2)];
cannam@95 188 T2P = FNMS(T23, T21, T2O);
cannam@95 189 T25 = FMA(T23, T24, T22);
cannam@95 190 T1N = W[2];
cannam@95 191 T1Q = W[3];
cannam@95 192 T1V = ri[WS(rs, 17)];
cannam@95 193 T1Y = ii[WS(rs, 17)];
cannam@95 194 T36 = T1N * T1R;
cannam@95 195 T1P = T1N * T1O;
cannam@95 196 T1U = W[32];
cannam@95 197 T1X = W[33];
cannam@95 198 T37 = FNMS(T1Q, T1O, T36);
cannam@95 199 T1S = FMA(T1Q, T1R, T1P);
cannam@95 200 T2M = T1U * T1Y;
cannam@95 201 T1W = T1U * T1V;
cannam@95 202 }
cannam@95 203 }
cannam@95 204 {
cannam@95 205 E T38, T3U, T1T, T2L, T2N, T1Z;
cannam@95 206 T38 = T35 - T37;
cannam@95 207 T3U = T35 + T37;
cannam@95 208 T1T = T1M + T1S;
cannam@95 209 T2L = T1M - T1S;
cannam@95 210 T2N = FNMS(T1X, T1V, T2M);
cannam@95 211 T1Z = FMA(T1X, T1Y, T1W);
cannam@95 212 {
cannam@95 213 E T3V, T2Q, T39, T26;
cannam@95 214 T3V = T2N + T2P;
cannam@95 215 T2Q = T2N - T2P;
cannam@95 216 T39 = T1Z - T25;
cannam@95 217 T26 = T1Z + T25;
cannam@95 218 T2R = T2L - T2Q;
cannam@95 219 T3z = T2L + T2Q;
cannam@95 220 T3W = T3U - T3V;
cannam@95 221 T4f = T3U + T3V;
cannam@95 222 T27 = T1T - T26;
cannam@95 223 T2f = T1T + T26;
cannam@95 224 T3J = T38 - T39;
cannam@95 225 T3a = T38 + T39;
cannam@95 226 }
cannam@95 227 }
cannam@95 228 }
cannam@95 229 {
cannam@95 230 E T2Y, T1l, T2I, T1E, T1u, T1x, T1w, T30, T1r, T2F, T1v;
cannam@95 231 {
cannam@95 232 E T1A, T1D, T1C, T2H, T1B;
cannam@95 233 {
cannam@95 234 E T1h, T1k, T1g, T1j, T2X, T1i, T1z;
cannam@95 235 T1h = ri[WS(rs, 8)];
cannam@95 236 T1k = ii[WS(rs, 8)];
cannam@95 237 T1g = W[14];
cannam@95 238 T1j = W[15];
cannam@95 239 T1A = ri[WS(rs, 3)];
cannam@95 240 T1D = ii[WS(rs, 3)];
cannam@95 241 T2X = T1g * T1k;
cannam@95 242 T1i = T1g * T1h;
cannam@95 243 T1z = W[4];
cannam@95 244 T1C = W[5];
cannam@95 245 T2Y = FNMS(T1j, T1h, T2X);
cannam@95 246 T1l = FMA(T1j, T1k, T1i);
cannam@95 247 T2H = T1z * T1D;
cannam@95 248 T1B = T1z * T1A;
cannam@95 249 }
cannam@95 250 {
cannam@95 251 E T1n, T1q, T1m, T1p, T2Z, T1o, T1t;
cannam@95 252 T1n = ri[WS(rs, 18)];
cannam@95 253 T1q = ii[WS(rs, 18)];
cannam@95 254 T2I = FNMS(T1C, T1A, T2H);
cannam@95 255 T1E = FMA(T1C, T1D, T1B);
cannam@95 256 T1m = W[34];
cannam@95 257 T1p = W[35];
cannam@95 258 T1u = ri[WS(rs, 13)];
cannam@95 259 T1x = ii[WS(rs, 13)];
cannam@95 260 T2Z = T1m * T1q;
cannam@95 261 T1o = T1m * T1n;
cannam@95 262 T1t = W[24];
cannam@95 263 T1w = W[25];
cannam@95 264 T30 = FNMS(T1p, T1n, T2Z);
cannam@95 265 T1r = FMA(T1p, T1q, T1o);
cannam@95 266 T2F = T1t * T1x;
cannam@95 267 T1v = T1t * T1u;
cannam@95 268 }
cannam@95 269 }
cannam@95 270 {
cannam@95 271 E T31, T3R, T1s, T2E, T2G, T1y;
cannam@95 272 T31 = T2Y - T30;
cannam@95 273 T3R = T2Y + T30;
cannam@95 274 T1s = T1l + T1r;
cannam@95 275 T2E = T1l - T1r;
cannam@95 276 T2G = FNMS(T1w, T1u, T2F);
cannam@95 277 T1y = FMA(T1w, T1x, T1v);
cannam@95 278 {
cannam@95 279 E T3S, T2J, T32, T1F;
cannam@95 280 T3S = T2G + T2I;
cannam@95 281 T2J = T2G - T2I;
cannam@95 282 T32 = T1y - T1E;
cannam@95 283 T1F = T1y + T1E;
cannam@95 284 T2K = T2E - T2J;
cannam@95 285 T3y = T2E + T2J;
cannam@95 286 T3T = T3R - T3S;
cannam@95 287 T4e = T3R + T3S;
cannam@95 288 T1G = T1s - T1F;
cannam@95 289 T2e = T1s + T1F;
cannam@95 290 T3I = T31 - T32;
cannam@95 291 T33 = T31 + T32;
cannam@95 292 }
cannam@95 293 }
cannam@95 294 }
cannam@95 295 {
cannam@95 296 E T3k, TT, T2A, T1c, T12, T15, T14, T3m, TZ, T2x, T13;
cannam@95 297 {
cannam@95 298 E T18, T1b, T1a, T2z, T19;
cannam@95 299 {
cannam@95 300 E TP, TS, TO, TR, T3j, TQ, T17;
cannam@95 301 TP = ri[WS(rs, 16)];
cannam@95 302 TS = ii[WS(rs, 16)];
cannam@95 303 TO = W[30];
cannam@95 304 TR = W[31];
cannam@95 305 T18 = ri[WS(rs, 11)];
cannam@95 306 T1b = ii[WS(rs, 11)];
cannam@95 307 T3j = TO * TS;
cannam@95 308 TQ = TO * TP;
cannam@95 309 T17 = W[20];
cannam@95 310 T1a = W[21];
cannam@95 311 T3k = FNMS(TR, TP, T3j);
cannam@95 312 TT = FMA(TR, TS, TQ);
cannam@95 313 T2z = T17 * T1b;
cannam@95 314 T19 = T17 * T18;
cannam@95 315 }
cannam@95 316 {
cannam@95 317 E TV, TY, TU, TX, T3l, TW, T11;
cannam@95 318 TV = ri[WS(rs, 6)];
cannam@95 319 TY = ii[WS(rs, 6)];
cannam@95 320 T2A = FNMS(T1a, T18, T2z);
cannam@95 321 T1c = FMA(T1a, T1b, T19);
cannam@95 322 TU = W[10];
cannam@95 323 TX = W[11];
cannam@95 324 T12 = ri[WS(rs, 1)];
cannam@95 325 T15 = ii[WS(rs, 1)];
cannam@95 326 T3l = TU * TY;
cannam@95 327 TW = TU * TV;
cannam@95 328 T11 = W[0];
cannam@95 329 T14 = W[1];
cannam@95 330 T3m = FNMS(TX, TV, T3l);
cannam@95 331 TZ = FMA(TX, TY, TW);
cannam@95 332 T2x = T11 * T15;
cannam@95 333 T13 = T11 * T12;
cannam@95 334 }
cannam@95 335 }
cannam@95 336 {
cannam@95 337 E T3n, T41, T10, T2w, T2y, T16;
cannam@95 338 T3n = T3k - T3m;
cannam@95 339 T41 = T3k + T3m;
cannam@95 340 T10 = TT + TZ;
cannam@95 341 T2w = TT - TZ;
cannam@95 342 T2y = FNMS(T14, T12, T2x);
cannam@95 343 T16 = FMA(T14, T15, T13);
cannam@95 344 {
cannam@95 345 E T42, T2B, T3o, T1d;
cannam@95 346 T42 = T2y + T2A;
cannam@95 347 T2B = T2y - T2A;
cannam@95 348 T3o = T16 - T1c;
cannam@95 349 T1d = T16 + T1c;
cannam@95 350 T2C = T2w - T2B;
cannam@95 351 T3w = T2w + T2B;
cannam@95 352 T43 = T41 - T42;
cannam@95 353 T4c = T41 + T42;
cannam@95 354 T1e = T10 - T1d;
cannam@95 355 T2c = T10 + T1d;
cannam@95 356 T3G = T3n - T3o;
cannam@95 357 T3p = T3n + T3o;
cannam@95 358 }
cannam@95 359 }
cannam@95 360 }
cannam@95 361 {
cannam@95 362 E T4s, T4k, T4l, T4h, T4j, T49, T4y, T4A, T48;
cannam@95 363 {
cannam@95 364 E T4D, T4C, T2a, T47, T45, T4B, T4M, T4K, T46, T3Q;
cannam@95 365 {
cannam@95 366 E Tm, T1f, T4J, T4I, T28, T3X, T44, T29, T3P, T3O;
cannam@95 367 T4D = T3T + T3W;
cannam@95 368 T3X = T3T - T3W;
cannam@95 369 T44 = T40 - T43;
cannam@95 370 T4C = T40 + T43;
cannam@95 371 T2a = T8 + Tl;
cannam@95 372 Tm = T8 - Tl;
cannam@95 373 T1f = TN + T1e;
cannam@95 374 T4J = TN - T1e;
cannam@95 375 T4I = T1G - T27;
cannam@95 376 T28 = T1G + T27;
cannam@95 377 T47 = FMA(KP618033988, T3X, T44);
cannam@95 378 T45 = FNMS(KP618033988, T44, T3X);
cannam@95 379 T29 = T1f + T28;
cannam@95 380 T3P = T1f - T28;
cannam@95 381 T4B = T4r - T4n;
cannam@95 382 T4s = T4n + T4r;
cannam@95 383 ri[WS(rs, 10)] = Tm + T29;
cannam@95 384 T3O = FNMS(KP250000000, T29, Tm);
cannam@95 385 T4M = FMA(KP618033988, T4I, T4J);
cannam@95 386 T4K = FNMS(KP618033988, T4J, T4I);
cannam@95 387 T46 = FMA(KP559016994, T3P, T3O);
cannam@95 388 T3Q = FNMS(KP559016994, T3P, T3O);
cannam@95 389 }
cannam@95 390 {
cannam@95 391 E T2d, T4w, T4x, T2g, T2h;
cannam@95 392 {
cannam@95 393 E T4d, T4G, T4F, T4g, T4E, T4L, T4H;
cannam@95 394 T4k = T4b + T4c;
cannam@95 395 T4d = T4b - T4c;
cannam@95 396 T4G = T4C - T4D;
cannam@95 397 T4E = T4C + T4D;
cannam@95 398 ri[WS(rs, 18)] = FMA(KP951056516, T45, T3Q);
cannam@95 399 ri[WS(rs, 2)] = FNMS(KP951056516, T45, T3Q);
cannam@95 400 ri[WS(rs, 6)] = FMA(KP951056516, T47, T46);
cannam@95 401 ri[WS(rs, 14)] = FNMS(KP951056516, T47, T46);
cannam@95 402 ii[WS(rs, 10)] = T4E + T4B;
cannam@95 403 T4F = FNMS(KP250000000, T4E, T4B);
cannam@95 404 T4g = T4e - T4f;
cannam@95 405 T4l = T4e + T4f;
cannam@95 406 T2d = T2b + T2c;
cannam@95 407 T4w = T2b - T2c;
cannam@95 408 T4L = FMA(KP559016994, T4G, T4F);
cannam@95 409 T4H = FNMS(KP559016994, T4G, T4F);
cannam@95 410 T4h = FMA(KP618033988, T4g, T4d);
cannam@95 411 T4j = FNMS(KP618033988, T4d, T4g);
cannam@95 412 ii[WS(rs, 18)] = FNMS(KP951056516, T4K, T4H);
cannam@95 413 ii[WS(rs, 2)] = FMA(KP951056516, T4K, T4H);
cannam@95 414 ii[WS(rs, 14)] = FMA(KP951056516, T4M, T4L);
cannam@95 415 ii[WS(rs, 6)] = FNMS(KP951056516, T4M, T4L);
cannam@95 416 T4x = T2e - T2f;
cannam@95 417 T2g = T2e + T2f;
cannam@95 418 }
cannam@95 419 T2h = T2d + T2g;
cannam@95 420 T49 = T2d - T2g;
cannam@95 421 T4y = FMA(KP618033988, T4x, T4w);
cannam@95 422 T4A = FNMS(KP618033988, T4w, T4x);
cannam@95 423 ri[0] = T2a + T2h;
cannam@95 424 T48 = FNMS(KP250000000, T2h, T2a);
cannam@95 425 }
cannam@95 426 }
cannam@95 427 {
cannam@95 428 E T3u, T51, T5a, T5c, T56, T54;
cannam@95 429 {
cannam@95 430 E T53, T52, T3t, T3r, T2o, T59, T58, T2T, T2V, T4u, T4t, T2U, T3s, T2W;
cannam@95 431 {
cannam@95 432 E T3b, T3q, T4i, T4a, T4m;
cannam@95 433 T53 = T33 + T3a;
cannam@95 434 T3b = T33 - T3a;
cannam@95 435 T3q = T3i - T3p;
cannam@95 436 T52 = T3i + T3p;
cannam@95 437 T4i = FNMS(KP559016994, T49, T48);
cannam@95 438 T4a = FMA(KP559016994, T49, T48);
cannam@95 439 T4m = T4k + T4l;
cannam@95 440 T4u = T4k - T4l;
cannam@95 441 ri[WS(rs, 16)] = FMA(KP951056516, T4h, T4a);
cannam@95 442 ri[WS(rs, 4)] = FNMS(KP951056516, T4h, T4a);
cannam@95 443 ri[WS(rs, 8)] = FMA(KP951056516, T4j, T4i);
cannam@95 444 ri[WS(rs, 12)] = FNMS(KP951056516, T4j, T4i);
cannam@95 445 ii[0] = T4m + T4s;
cannam@95 446 T4t = FNMS(KP250000000, T4m, T4s);
cannam@95 447 T3t = FMA(KP618033988, T3b, T3q);
cannam@95 448 T3r = FNMS(KP618033988, T3q, T3b);
cannam@95 449 }
cannam@95 450 T3u = T2i + T2n;
cannam@95 451 T2o = T2i - T2n;
cannam@95 452 {
cannam@95 453 E T4v, T4z, T2D, T2S;
cannam@95 454 T4v = FMA(KP559016994, T4u, T4t);
cannam@95 455 T4z = FNMS(KP559016994, T4u, T4t);
cannam@95 456 T2D = T2v + T2C;
cannam@95 457 T59 = T2v - T2C;
cannam@95 458 T58 = T2K - T2R;
cannam@95 459 T2S = T2K + T2R;
cannam@95 460 ii[WS(rs, 16)] = FNMS(KP951056516, T4y, T4v);
cannam@95 461 ii[WS(rs, 4)] = FMA(KP951056516, T4y, T4v);
cannam@95 462 ii[WS(rs, 12)] = FMA(KP951056516, T4A, T4z);
cannam@95 463 ii[WS(rs, 8)] = FNMS(KP951056516, T4A, T4z);
cannam@95 464 T2T = T2D + T2S;
cannam@95 465 T2V = T2D - T2S;
cannam@95 466 }
cannam@95 467 ri[WS(rs, 15)] = T2o + T2T;
cannam@95 468 T2U = FNMS(KP250000000, T2T, T2o);
cannam@95 469 T51 = T4O + T4N;
cannam@95 470 T4P = T4N - T4O;
cannam@95 471 T5a = FNMS(KP618033988, T59, T58);
cannam@95 472 T5c = FMA(KP618033988, T58, T59);
cannam@95 473 T3s = FMA(KP559016994, T2V, T2U);
cannam@95 474 T2W = FNMS(KP559016994, T2V, T2U);
cannam@95 475 ri[WS(rs, 7)] = FNMS(KP951056516, T3r, T2W);
cannam@95 476 ri[WS(rs, 3)] = FMA(KP951056516, T3r, T2W);
cannam@95 477 ri[WS(rs, 19)] = FNMS(KP951056516, T3t, T3s);
cannam@95 478 ri[WS(rs, 11)] = FMA(KP951056516, T3t, T3s);
cannam@95 479 T56 = T52 - T53;
cannam@95 480 T54 = T52 + T53;
cannam@95 481 }
cannam@95 482 {
cannam@95 483 E T4Q, T4R, T3N, T3L, T4W, T4X, T3B, T3D, T3H, T3K, T55, T3C, T3M, T3E;
cannam@95 484 T4Q = T3F + T3G;
cannam@95 485 T3H = T3F - T3G;
cannam@95 486 T3K = T3I - T3J;
cannam@95 487 T4R = T3I + T3J;
cannam@95 488 ii[WS(rs, 15)] = T54 + T51;
cannam@95 489 T55 = FNMS(KP250000000, T54, T51);
cannam@95 490 T3N = FNMS(KP618033988, T3H, T3K);
cannam@95 491 T3L = FMA(KP618033988, T3K, T3H);
cannam@95 492 {
cannam@95 493 E T57, T5b, T3x, T3A;
cannam@95 494 T57 = FNMS(KP559016994, T56, T55);
cannam@95 495 T5b = FMA(KP559016994, T56, T55);
cannam@95 496 T3x = T3v + T3w;
cannam@95 497 T4W = T3v - T3w;
cannam@95 498 T4X = T3y - T3z;
cannam@95 499 T3A = T3y + T3z;
cannam@95 500 ii[WS(rs, 7)] = FMA(KP951056516, T5a, T57);
cannam@95 501 ii[WS(rs, 3)] = FNMS(KP951056516, T5a, T57);
cannam@95 502 ii[WS(rs, 19)] = FMA(KP951056516, T5c, T5b);
cannam@95 503 ii[WS(rs, 11)] = FNMS(KP951056516, T5c, T5b);
cannam@95 504 T3B = T3x + T3A;
cannam@95 505 T3D = T3x - T3A;
cannam@95 506 }
cannam@95 507 ri[WS(rs, 5)] = T3u + T3B;
cannam@95 508 T3C = FNMS(KP250000000, T3B, T3u);
cannam@95 509 T4Y = FMA(KP618033988, T4X, T4W);
cannam@95 510 T50 = FNMS(KP618033988, T4W, T4X);
cannam@95 511 T3M = FNMS(KP559016994, T3D, T3C);
cannam@95 512 T3E = FMA(KP559016994, T3D, T3C);
cannam@95 513 ri[WS(rs, 9)] = FNMS(KP951056516, T3L, T3E);
cannam@95 514 ri[WS(rs, 1)] = FMA(KP951056516, T3L, T3E);
cannam@95 515 ri[WS(rs, 17)] = FNMS(KP951056516, T3N, T3M);
cannam@95 516 ri[WS(rs, 13)] = FMA(KP951056516, T3N, T3M);
cannam@95 517 T4U = T4Q - T4R;
cannam@95 518 T4S = T4Q + T4R;
cannam@95 519 }
cannam@95 520 }
cannam@95 521 }
cannam@95 522 }
cannam@95 523 ii[WS(rs, 5)] = T4S + T4P;
cannam@95 524 T4T = FNMS(KP250000000, T4S, T4P);
cannam@95 525 T4Z = FNMS(KP559016994, T4U, T4T);
cannam@95 526 T4V = FMA(KP559016994, T4U, T4T);
cannam@95 527 ii[WS(rs, 9)] = FMA(KP951056516, T4Y, T4V);
cannam@95 528 ii[WS(rs, 1)] = FNMS(KP951056516, T4Y, T4V);
cannam@95 529 ii[WS(rs, 17)] = FMA(KP951056516, T50, T4Z);
cannam@95 530 ii[WS(rs, 13)] = FNMS(KP951056516, T50, T4Z);
cannam@95 531 }
cannam@95 532 }
cannam@95 533 }
cannam@95 534
cannam@95 535 static const tw_instr twinstr[] = {
cannam@95 536 {TW_FULL, 0, 20},
cannam@95 537 {TW_NEXT, 1, 0}
cannam@95 538 };
cannam@95 539
cannam@95 540 static const ct_desc desc = { 20, "t1_20", twinstr, &GENUS, {136, 38, 110, 0}, 0, 0, 0 };
cannam@95 541
cannam@95 542 void X(codelet_t1_20) (planner *p) {
cannam@95 543 X(kdft_dit_register) (p, t1_20, &desc);
cannam@95 544 }
cannam@95 545 #else /* HAVE_FMA */
cannam@95 546
cannam@95 547 /* Generated by: ../../../genfft/gen_twiddle.native -compact -variables 4 -pipeline-latency 4 -n 20 -name t1_20 -include t.h */
cannam@95 548
cannam@95 549 /*
cannam@95 550 * This function contains 246 FP additions, 124 FP multiplications,
cannam@95 551 * (or, 184 additions, 62 multiplications, 62 fused multiply/add),
cannam@95 552 * 85 stack variables, 4 constants, and 80 memory accesses
cannam@95 553 */
cannam@95 554 #include "t.h"
cannam@95 555
cannam@95 556 static void t1_20(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
cannam@95 557 {
cannam@95 558 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
cannam@95 559 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
cannam@95 560 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
cannam@95 561 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
cannam@95 562 {
cannam@95 563 INT m;
cannam@95 564 for (m = mb, W = W + (mb * 38); m < me; m = m + 1, ri = ri + ms, ii = ii + ms, W = W + 38, MAKE_VOLATILE_STRIDE(40, rs)) {
cannam@95 565 E Tj, T1R, T4g, T4p, T2q, T37, T3Q, T42, T1r, T1O, T1P, T3i, T3l, T44, T3D;
cannam@95 566 E T3E, T3K, T1V, T1W, T1X, T23, T28, T4r, T2W, T2X, T4c, T33, T34, T35, T2G;
cannam@95 567 E T2L, T2M, TG, T13, T14, T3p, T3s, T43, T3A, T3B, T3J, T1S, T1T, T1U, T2e;
cannam@95 568 E T2j, T4q, T2T, T2U, T4b, T30, T31, T32, T2v, T2A, T2B;
cannam@95 569 {
cannam@95 570 E T1, T3O, T6, T3N, Tc, T2n, Th, T2o;
cannam@95 571 T1 = ri[0];
cannam@95 572 T3O = ii[0];
cannam@95 573 {
cannam@95 574 E T3, T5, T2, T4;
cannam@95 575 T3 = ri[WS(rs, 10)];
cannam@95 576 T5 = ii[WS(rs, 10)];
cannam@95 577 T2 = W[18];
cannam@95 578 T4 = W[19];
cannam@95 579 T6 = FMA(T2, T3, T4 * T5);
cannam@95 580 T3N = FNMS(T4, T3, T2 * T5);
cannam@95 581 }
cannam@95 582 {
cannam@95 583 E T9, Tb, T8, Ta;
cannam@95 584 T9 = ri[WS(rs, 5)];
cannam@95 585 Tb = ii[WS(rs, 5)];
cannam@95 586 T8 = W[8];
cannam@95 587 Ta = W[9];
cannam@95 588 Tc = FMA(T8, T9, Ta * Tb);
cannam@95 589 T2n = FNMS(Ta, T9, T8 * Tb);
cannam@95 590 }
cannam@95 591 {
cannam@95 592 E Te, Tg, Td, Tf;
cannam@95 593 Te = ri[WS(rs, 15)];
cannam@95 594 Tg = ii[WS(rs, 15)];
cannam@95 595 Td = W[28];
cannam@95 596 Tf = W[29];
cannam@95 597 Th = FMA(Td, Te, Tf * Tg);
cannam@95 598 T2o = FNMS(Tf, Te, Td * Tg);
cannam@95 599 }
cannam@95 600 {
cannam@95 601 E T7, Ti, T4e, T4f;
cannam@95 602 T7 = T1 + T6;
cannam@95 603 Ti = Tc + Th;
cannam@95 604 Tj = T7 - Ti;
cannam@95 605 T1R = T7 + Ti;
cannam@95 606 T4e = T3O - T3N;
cannam@95 607 T4f = Tc - Th;
cannam@95 608 T4g = T4e - T4f;
cannam@95 609 T4p = T4f + T4e;
cannam@95 610 }
cannam@95 611 {
cannam@95 612 E T2m, T2p, T3M, T3P;
cannam@95 613 T2m = T1 - T6;
cannam@95 614 T2p = T2n - T2o;
cannam@95 615 T2q = T2m - T2p;
cannam@95 616 T37 = T2m + T2p;
cannam@95 617 T3M = T2n + T2o;
cannam@95 618 T3P = T3N + T3O;
cannam@95 619 T3Q = T3M + T3P;
cannam@95 620 T42 = T3P - T3M;
cannam@95 621 }
cannam@95 622 }
cannam@95 623 {
cannam@95 624 E T1f, T3g, T21, T2C, T1N, T3k, T27, T2K, T1q, T3h, T22, T2F, T1C, T3j, T26;
cannam@95 625 E T2H;
cannam@95 626 {
cannam@95 627 E T19, T1Z, T1e, T20;
cannam@95 628 {
cannam@95 629 E T16, T18, T15, T17;
cannam@95 630 T16 = ri[WS(rs, 8)];
cannam@95 631 T18 = ii[WS(rs, 8)];
cannam@95 632 T15 = W[14];
cannam@95 633 T17 = W[15];
cannam@95 634 T19 = FMA(T15, T16, T17 * T18);
cannam@95 635 T1Z = FNMS(T17, T16, T15 * T18);
cannam@95 636 }
cannam@95 637 {
cannam@95 638 E T1b, T1d, T1a, T1c;
cannam@95 639 T1b = ri[WS(rs, 18)];
cannam@95 640 T1d = ii[WS(rs, 18)];
cannam@95 641 T1a = W[34];
cannam@95 642 T1c = W[35];
cannam@95 643 T1e = FMA(T1a, T1b, T1c * T1d);
cannam@95 644 T20 = FNMS(T1c, T1b, T1a * T1d);
cannam@95 645 }
cannam@95 646 T1f = T19 + T1e;
cannam@95 647 T3g = T1Z + T20;
cannam@95 648 T21 = T1Z - T20;
cannam@95 649 T2C = T19 - T1e;
cannam@95 650 }
cannam@95 651 {
cannam@95 652 E T1H, T2I, T1M, T2J;
cannam@95 653 {
cannam@95 654 E T1E, T1G, T1D, T1F;
cannam@95 655 T1E = ri[WS(rs, 17)];
cannam@95 656 T1G = ii[WS(rs, 17)];
cannam@95 657 T1D = W[32];
cannam@95 658 T1F = W[33];
cannam@95 659 T1H = FMA(T1D, T1E, T1F * T1G);
cannam@95 660 T2I = FNMS(T1F, T1E, T1D * T1G);
cannam@95 661 }
cannam@95 662 {
cannam@95 663 E T1J, T1L, T1I, T1K;
cannam@95 664 T1J = ri[WS(rs, 7)];
cannam@95 665 T1L = ii[WS(rs, 7)];
cannam@95 666 T1I = W[12];
cannam@95 667 T1K = W[13];
cannam@95 668 T1M = FMA(T1I, T1J, T1K * T1L);
cannam@95 669 T2J = FNMS(T1K, T1J, T1I * T1L);
cannam@95 670 }
cannam@95 671 T1N = T1H + T1M;
cannam@95 672 T3k = T2I + T2J;
cannam@95 673 T27 = T1H - T1M;
cannam@95 674 T2K = T2I - T2J;
cannam@95 675 }
cannam@95 676 {
cannam@95 677 E T1k, T2D, T1p, T2E;
cannam@95 678 {
cannam@95 679 E T1h, T1j, T1g, T1i;
cannam@95 680 T1h = ri[WS(rs, 13)];
cannam@95 681 T1j = ii[WS(rs, 13)];
cannam@95 682 T1g = W[24];
cannam@95 683 T1i = W[25];
cannam@95 684 T1k = FMA(T1g, T1h, T1i * T1j);
cannam@95 685 T2D = FNMS(T1i, T1h, T1g * T1j);
cannam@95 686 }
cannam@95 687 {
cannam@95 688 E T1m, T1o, T1l, T1n;
cannam@95 689 T1m = ri[WS(rs, 3)];
cannam@95 690 T1o = ii[WS(rs, 3)];
cannam@95 691 T1l = W[4];
cannam@95 692 T1n = W[5];
cannam@95 693 T1p = FMA(T1l, T1m, T1n * T1o);
cannam@95 694 T2E = FNMS(T1n, T1m, T1l * T1o);
cannam@95 695 }
cannam@95 696 T1q = T1k + T1p;
cannam@95 697 T3h = T2D + T2E;
cannam@95 698 T22 = T1k - T1p;
cannam@95 699 T2F = T2D - T2E;
cannam@95 700 }
cannam@95 701 {
cannam@95 702 E T1w, T24, T1B, T25;
cannam@95 703 {
cannam@95 704 E T1t, T1v, T1s, T1u;
cannam@95 705 T1t = ri[WS(rs, 12)];
cannam@95 706 T1v = ii[WS(rs, 12)];
cannam@95 707 T1s = W[22];
cannam@95 708 T1u = W[23];
cannam@95 709 T1w = FMA(T1s, T1t, T1u * T1v);
cannam@95 710 T24 = FNMS(T1u, T1t, T1s * T1v);
cannam@95 711 }
cannam@95 712 {
cannam@95 713 E T1y, T1A, T1x, T1z;
cannam@95 714 T1y = ri[WS(rs, 2)];
cannam@95 715 T1A = ii[WS(rs, 2)];
cannam@95 716 T1x = W[2];
cannam@95 717 T1z = W[3];
cannam@95 718 T1B = FMA(T1x, T1y, T1z * T1A);
cannam@95 719 T25 = FNMS(T1z, T1y, T1x * T1A);
cannam@95 720 }
cannam@95 721 T1C = T1w + T1B;
cannam@95 722 T3j = T24 + T25;
cannam@95 723 T26 = T24 - T25;
cannam@95 724 T2H = T1w - T1B;
cannam@95 725 }
cannam@95 726 T1r = T1f - T1q;
cannam@95 727 T1O = T1C - T1N;
cannam@95 728 T1P = T1r + T1O;
cannam@95 729 T3i = T3g - T3h;
cannam@95 730 T3l = T3j - T3k;
cannam@95 731 T44 = T3i + T3l;
cannam@95 732 T3D = T3g + T3h;
cannam@95 733 T3E = T3j + T3k;
cannam@95 734 T3K = T3D + T3E;
cannam@95 735 T1V = T1f + T1q;
cannam@95 736 T1W = T1C + T1N;
cannam@95 737 T1X = T1V + T1W;
cannam@95 738 T23 = T21 + T22;
cannam@95 739 T28 = T26 + T27;
cannam@95 740 T4r = T23 + T28;
cannam@95 741 T2W = T21 - T22;
cannam@95 742 T2X = T26 - T27;
cannam@95 743 T4c = T2W + T2X;
cannam@95 744 T33 = T2C + T2F;
cannam@95 745 T34 = T2H + T2K;
cannam@95 746 T35 = T33 + T34;
cannam@95 747 T2G = T2C - T2F;
cannam@95 748 T2L = T2H - T2K;
cannam@95 749 T2M = T2G + T2L;
cannam@95 750 }
cannam@95 751 {
cannam@95 752 E Tu, T3n, T2c, T2r, T12, T3r, T2i, T2z, TF, T3o, T2d, T2u, TR, T3q, T2h;
cannam@95 753 E T2w;
cannam@95 754 {
cannam@95 755 E To, T2a, Tt, T2b;
cannam@95 756 {
cannam@95 757 E Tl, Tn, Tk, Tm;
cannam@95 758 Tl = ri[WS(rs, 4)];
cannam@95 759 Tn = ii[WS(rs, 4)];
cannam@95 760 Tk = W[6];
cannam@95 761 Tm = W[7];
cannam@95 762 To = FMA(Tk, Tl, Tm * Tn);
cannam@95 763 T2a = FNMS(Tm, Tl, Tk * Tn);
cannam@95 764 }
cannam@95 765 {
cannam@95 766 E Tq, Ts, Tp, Tr;
cannam@95 767 Tq = ri[WS(rs, 14)];
cannam@95 768 Ts = ii[WS(rs, 14)];
cannam@95 769 Tp = W[26];
cannam@95 770 Tr = W[27];
cannam@95 771 Tt = FMA(Tp, Tq, Tr * Ts);
cannam@95 772 T2b = FNMS(Tr, Tq, Tp * Ts);
cannam@95 773 }
cannam@95 774 Tu = To + Tt;
cannam@95 775 T3n = T2a + T2b;
cannam@95 776 T2c = T2a - T2b;
cannam@95 777 T2r = To - Tt;
cannam@95 778 }
cannam@95 779 {
cannam@95 780 E TW, T2x, T11, T2y;
cannam@95 781 {
cannam@95 782 E TT, TV, TS, TU;
cannam@95 783 TT = ri[WS(rs, 1)];
cannam@95 784 TV = ii[WS(rs, 1)];
cannam@95 785 TS = W[0];
cannam@95 786 TU = W[1];
cannam@95 787 TW = FMA(TS, TT, TU * TV);
cannam@95 788 T2x = FNMS(TU, TT, TS * TV);
cannam@95 789 }
cannam@95 790 {
cannam@95 791 E TY, T10, TX, TZ;
cannam@95 792 TY = ri[WS(rs, 11)];
cannam@95 793 T10 = ii[WS(rs, 11)];
cannam@95 794 TX = W[20];
cannam@95 795 TZ = W[21];
cannam@95 796 T11 = FMA(TX, TY, TZ * T10);
cannam@95 797 T2y = FNMS(TZ, TY, TX * T10);
cannam@95 798 }
cannam@95 799 T12 = TW + T11;
cannam@95 800 T3r = T2x + T2y;
cannam@95 801 T2i = TW - T11;
cannam@95 802 T2z = T2x - T2y;
cannam@95 803 }
cannam@95 804 {
cannam@95 805 E Tz, T2s, TE, T2t;
cannam@95 806 {
cannam@95 807 E Tw, Ty, Tv, Tx;
cannam@95 808 Tw = ri[WS(rs, 9)];
cannam@95 809 Ty = ii[WS(rs, 9)];
cannam@95 810 Tv = W[16];
cannam@95 811 Tx = W[17];
cannam@95 812 Tz = FMA(Tv, Tw, Tx * Ty);
cannam@95 813 T2s = FNMS(Tx, Tw, Tv * Ty);
cannam@95 814 }
cannam@95 815 {
cannam@95 816 E TB, TD, TA, TC;
cannam@95 817 TB = ri[WS(rs, 19)];
cannam@95 818 TD = ii[WS(rs, 19)];
cannam@95 819 TA = W[36];
cannam@95 820 TC = W[37];
cannam@95 821 TE = FMA(TA, TB, TC * TD);
cannam@95 822 T2t = FNMS(TC, TB, TA * TD);
cannam@95 823 }
cannam@95 824 TF = Tz + TE;
cannam@95 825 T3o = T2s + T2t;
cannam@95 826 T2d = Tz - TE;
cannam@95 827 T2u = T2s - T2t;
cannam@95 828 }
cannam@95 829 {
cannam@95 830 E TL, T2f, TQ, T2g;
cannam@95 831 {
cannam@95 832 E TI, TK, TH, TJ;
cannam@95 833 TI = ri[WS(rs, 16)];
cannam@95 834 TK = ii[WS(rs, 16)];
cannam@95 835 TH = W[30];
cannam@95 836 TJ = W[31];
cannam@95 837 TL = FMA(TH, TI, TJ * TK);
cannam@95 838 T2f = FNMS(TJ, TI, TH * TK);
cannam@95 839 }
cannam@95 840 {
cannam@95 841 E TN, TP, TM, TO;
cannam@95 842 TN = ri[WS(rs, 6)];
cannam@95 843 TP = ii[WS(rs, 6)];
cannam@95 844 TM = W[10];
cannam@95 845 TO = W[11];
cannam@95 846 TQ = FMA(TM, TN, TO * TP);
cannam@95 847 T2g = FNMS(TO, TN, TM * TP);
cannam@95 848 }
cannam@95 849 TR = TL + TQ;
cannam@95 850 T3q = T2f + T2g;
cannam@95 851 T2h = T2f - T2g;
cannam@95 852 T2w = TL - TQ;
cannam@95 853 }
cannam@95 854 TG = Tu - TF;
cannam@95 855 T13 = TR - T12;
cannam@95 856 T14 = TG + T13;
cannam@95 857 T3p = T3n - T3o;
cannam@95 858 T3s = T3q - T3r;
cannam@95 859 T43 = T3p + T3s;
cannam@95 860 T3A = T3n + T3o;
cannam@95 861 T3B = T3q + T3r;
cannam@95 862 T3J = T3A + T3B;
cannam@95 863 T1S = Tu + TF;
cannam@95 864 T1T = TR + T12;
cannam@95 865 T1U = T1S + T1T;
cannam@95 866 T2e = T2c + T2d;
cannam@95 867 T2j = T2h + T2i;
cannam@95 868 T4q = T2e + T2j;
cannam@95 869 T2T = T2c - T2d;
cannam@95 870 T2U = T2h - T2i;
cannam@95 871 T4b = T2T + T2U;
cannam@95 872 T30 = T2r + T2u;
cannam@95 873 T31 = T2w + T2z;
cannam@95 874 T32 = T30 + T31;
cannam@95 875 T2v = T2r - T2u;
cannam@95 876 T2A = T2w - T2z;
cannam@95 877 T2B = T2v + T2A;
cannam@95 878 }
cannam@95 879 {
cannam@95 880 E T3e, T1Q, T3d, T3u, T3w, T3m, T3t, T3v, T3f;
cannam@95 881 T3e = KP559016994 * (T14 - T1P);
cannam@95 882 T1Q = T14 + T1P;
cannam@95 883 T3d = FNMS(KP250000000, T1Q, Tj);
cannam@95 884 T3m = T3i - T3l;
cannam@95 885 T3t = T3p - T3s;
cannam@95 886 T3u = FNMS(KP587785252, T3t, KP951056516 * T3m);
cannam@95 887 T3w = FMA(KP951056516, T3t, KP587785252 * T3m);
cannam@95 888 ri[WS(rs, 10)] = Tj + T1Q;
cannam@95 889 T3v = T3e + T3d;
cannam@95 890 ri[WS(rs, 14)] = T3v - T3w;
cannam@95 891 ri[WS(rs, 6)] = T3v + T3w;
cannam@95 892 T3f = T3d - T3e;
cannam@95 893 ri[WS(rs, 2)] = T3f - T3u;
cannam@95 894 ri[WS(rs, 18)] = T3f + T3u;
cannam@95 895 }
cannam@95 896 {
cannam@95 897 E T47, T45, T46, T41, T4a, T3Z, T40, T49, T48;
cannam@95 898 T47 = KP559016994 * (T43 - T44);
cannam@95 899 T45 = T43 + T44;
cannam@95 900 T46 = FNMS(KP250000000, T45, T42);
cannam@95 901 T3Z = T1r - T1O;
cannam@95 902 T40 = TG - T13;
cannam@95 903 T41 = FNMS(KP587785252, T40, KP951056516 * T3Z);
cannam@95 904 T4a = FMA(KP951056516, T40, KP587785252 * T3Z);
cannam@95 905 ii[WS(rs, 10)] = T45 + T42;
cannam@95 906 T49 = T47 + T46;
cannam@95 907 ii[WS(rs, 6)] = T49 - T4a;
cannam@95 908 ii[WS(rs, 14)] = T4a + T49;
cannam@95 909 T48 = T46 - T47;
cannam@95 910 ii[WS(rs, 2)] = T41 + T48;
cannam@95 911 ii[WS(rs, 18)] = T48 - T41;
cannam@95 912 }
cannam@95 913 {
cannam@95 914 E T3x, T1Y, T3y, T3G, T3I, T3C, T3F, T3H, T3z;
cannam@95 915 T3x = KP559016994 * (T1U - T1X);
cannam@95 916 T1Y = T1U + T1X;
cannam@95 917 T3y = FNMS(KP250000000, T1Y, T1R);
cannam@95 918 T3C = T3A - T3B;
cannam@95 919 T3F = T3D - T3E;
cannam@95 920 T3G = FMA(KP951056516, T3C, KP587785252 * T3F);
cannam@95 921 T3I = FNMS(KP587785252, T3C, KP951056516 * T3F);
cannam@95 922 ri[0] = T1R + T1Y;
cannam@95 923 T3H = T3y - T3x;
cannam@95 924 ri[WS(rs, 12)] = T3H - T3I;
cannam@95 925 ri[WS(rs, 8)] = T3H + T3I;
cannam@95 926 T3z = T3x + T3y;
cannam@95 927 ri[WS(rs, 4)] = T3z - T3G;
cannam@95 928 ri[WS(rs, 16)] = T3z + T3G;
cannam@95 929 }
cannam@95 930 {
cannam@95 931 E T3U, T3L, T3V, T3T, T3Y, T3R, T3S, T3X, T3W;
cannam@95 932 T3U = KP559016994 * (T3J - T3K);
cannam@95 933 T3L = T3J + T3K;
cannam@95 934 T3V = FNMS(KP250000000, T3L, T3Q);
cannam@95 935 T3R = T1S - T1T;
cannam@95 936 T3S = T1V - T1W;
cannam@95 937 T3T = FMA(KP951056516, T3R, KP587785252 * T3S);
cannam@95 938 T3Y = FNMS(KP587785252, T3R, KP951056516 * T3S);
cannam@95 939 ii[0] = T3L + T3Q;
cannam@95 940 T3X = T3V - T3U;
cannam@95 941 ii[WS(rs, 8)] = T3X - T3Y;
cannam@95 942 ii[WS(rs, 12)] = T3Y + T3X;
cannam@95 943 T3W = T3U + T3V;
cannam@95 944 ii[WS(rs, 4)] = T3T + T3W;
cannam@95 945 ii[WS(rs, 16)] = T3W - T3T;
cannam@95 946 }
cannam@95 947 {
cannam@95 948 E T2P, T2N, T2O, T2l, T2R, T29, T2k, T2S, T2Q;
cannam@95 949 T2P = KP559016994 * (T2B - T2M);
cannam@95 950 T2N = T2B + T2M;
cannam@95 951 T2O = FNMS(KP250000000, T2N, T2q);
cannam@95 952 T29 = T23 - T28;
cannam@95 953 T2k = T2e - T2j;
cannam@95 954 T2l = FNMS(KP587785252, T2k, KP951056516 * T29);
cannam@95 955 T2R = FMA(KP951056516, T2k, KP587785252 * T29);
cannam@95 956 ri[WS(rs, 15)] = T2q + T2N;
cannam@95 957 T2S = T2P + T2O;
cannam@95 958 ri[WS(rs, 11)] = T2R + T2S;
cannam@95 959 ri[WS(rs, 19)] = T2S - T2R;
cannam@95 960 T2Q = T2O - T2P;
cannam@95 961 ri[WS(rs, 3)] = T2l + T2Q;
cannam@95 962 ri[WS(rs, 7)] = T2Q - T2l;
cannam@95 963 }
cannam@95 964 {
cannam@95 965 E T4u, T4s, T4t, T4y, T4A, T4w, T4x, T4z, T4v;
cannam@95 966 T4u = KP559016994 * (T4q - T4r);
cannam@95 967 T4s = T4q + T4r;
cannam@95 968 T4t = FNMS(KP250000000, T4s, T4p);
cannam@95 969 T4w = T2G - T2L;
cannam@95 970 T4x = T2v - T2A;
cannam@95 971 T4y = FNMS(KP587785252, T4x, KP951056516 * T4w);
cannam@95 972 T4A = FMA(KP951056516, T4x, KP587785252 * T4w);
cannam@95 973 ii[WS(rs, 15)] = T4s + T4p;
cannam@95 974 T4z = T4u + T4t;
cannam@95 975 ii[WS(rs, 11)] = T4z - T4A;
cannam@95 976 ii[WS(rs, 19)] = T4A + T4z;
cannam@95 977 T4v = T4t - T4u;
cannam@95 978 ii[WS(rs, 3)] = T4v - T4y;
cannam@95 979 ii[WS(rs, 7)] = T4y + T4v;
cannam@95 980 }
cannam@95 981 {
cannam@95 982 E T36, T38, T39, T2Z, T3b, T2V, T2Y, T3c, T3a;
cannam@95 983 T36 = KP559016994 * (T32 - T35);
cannam@95 984 T38 = T32 + T35;
cannam@95 985 T39 = FNMS(KP250000000, T38, T37);
cannam@95 986 T2V = T2T - T2U;
cannam@95 987 T2Y = T2W - T2X;
cannam@95 988 T2Z = FMA(KP951056516, T2V, KP587785252 * T2Y);
cannam@95 989 T3b = FNMS(KP587785252, T2V, KP951056516 * T2Y);
cannam@95 990 ri[WS(rs, 5)] = T37 + T38;
cannam@95 991 T3c = T39 - T36;
cannam@95 992 ri[WS(rs, 13)] = T3b + T3c;
cannam@95 993 ri[WS(rs, 17)] = T3c - T3b;
cannam@95 994 T3a = T36 + T39;
cannam@95 995 ri[WS(rs, 1)] = T2Z + T3a;
cannam@95 996 ri[WS(rs, 9)] = T3a - T2Z;
cannam@95 997 }
cannam@95 998 {
cannam@95 999 E T4d, T4h, T4i, T4m, T4o, T4k, T4l, T4n, T4j;
cannam@95 1000 T4d = KP559016994 * (T4b - T4c);
cannam@95 1001 T4h = T4b + T4c;
cannam@95 1002 T4i = FNMS(KP250000000, T4h, T4g);
cannam@95 1003 T4k = T30 - T31;
cannam@95 1004 T4l = T33 - T34;
cannam@95 1005 T4m = FMA(KP951056516, T4k, KP587785252 * T4l);
cannam@95 1006 T4o = FNMS(KP587785252, T4k, KP951056516 * T4l);
cannam@95 1007 ii[WS(rs, 5)] = T4h + T4g;
cannam@95 1008 T4n = T4i - T4d;
cannam@95 1009 ii[WS(rs, 13)] = T4n - T4o;
cannam@95 1010 ii[WS(rs, 17)] = T4o + T4n;
cannam@95 1011 T4j = T4d + T4i;
cannam@95 1012 ii[WS(rs, 1)] = T4j - T4m;
cannam@95 1013 ii[WS(rs, 9)] = T4m + T4j;
cannam@95 1014 }
cannam@95 1015 }
cannam@95 1016 }
cannam@95 1017 }
cannam@95 1018
cannam@95 1019 static const tw_instr twinstr[] = {
cannam@95 1020 {TW_FULL, 0, 20},
cannam@95 1021 {TW_NEXT, 1, 0}
cannam@95 1022 };
cannam@95 1023
cannam@95 1024 static const ct_desc desc = { 20, "t1_20", twinstr, &GENUS, {184, 62, 62, 0}, 0, 0, 0 };
cannam@95 1025
cannam@95 1026 void X(codelet_t1_20) (planner *p) {
cannam@95 1027 X(kdft_dit_register) (p, t1_20, &desc);
cannam@95 1028 }
cannam@95 1029 #endif /* HAVE_FMA */