Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.5/dft/scalar/codelets/t1_20.c @ 84:08ae793730bd

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Add null config files
author Chris Cannam
date Mon, 02 Mar 2020 14:03:47 +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 16:36:24 EDT 2016 */
Chris@42 23
Chris@42 24 #include "codelet-dft.h"
Chris@42 25
Chris@42 26 #ifdef HAVE_FMA
Chris@42 27
Chris@42 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 */
Chris@42 29
Chris@42 30 /*
Chris@42 31 * This function contains 246 FP additions, 148 FP multiplications,
Chris@42 32 * (or, 136 additions, 38 multiplications, 110 fused multiply/add),
Chris@42 33 * 97 stack variables, 4 constants, and 80 memory accesses
Chris@42 34 */
Chris@42 35 #include "t.h"
Chris@42 36
Chris@42 37 static void t1_20(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
Chris@42 38 {
Chris@42 39 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
Chris@42 40 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
Chris@42 41 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
Chris@42 42 DK(KP618033988, +0.618033988749894848204586834365638117720309180);
Chris@42 43 {
Chris@42 44 INT m;
Chris@42 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)) {
Chris@42 46 E T4P, T4Y, T50, T4U, T4S, T4T, T4Z, T4V;
Chris@42 47 {
Chris@42 48 E T4N, T4r, T8, T2i, T4n, T2n, T4O, Tl, T2v, T3v, T40, T4b, TN, T2b, T3F;
Chris@42 49 E T3i, T2R, T3z, T3W, T4f, T27, T2f, T3J, T3a, T2K, T3y, T3T, T4e, T1G, T2e;
Chris@42 50 E T3I, T33, T2C, T3w, T43, T4c, T1e, T2c, T3G, T3p;
Chris@42 51 {
Chris@42 52 E T1, T4q, T3, T6, T2, T5;
Chris@42 53 T1 = ri[0];
Chris@42 54 T4q = ii[0];
Chris@42 55 T3 = ri[WS(rs, 10)];
Chris@42 56 T6 = ii[WS(rs, 10)];
Chris@42 57 T2 = W[18];
Chris@42 58 T5 = W[19];
Chris@42 59 {
Chris@42 60 E Ta, Td, Tg, T2j, Tb, Tj, Tf, Tc, Ti;
Chris@42 61 {
Chris@42 62 E T4o, T4, T9, T4p, T7;
Chris@42 63 Ta = ri[WS(rs, 5)];
Chris@42 64 Td = ii[WS(rs, 5)];
Chris@42 65 T4o = T2 * T6;
Chris@42 66 T4 = T2 * T3;
Chris@42 67 T9 = W[8];
Chris@42 68 Tg = ri[WS(rs, 15)];
Chris@42 69 T4p = FNMS(T5, T3, T4o);
Chris@42 70 T7 = FMA(T5, T6, T4);
Chris@42 71 T2j = T9 * Td;
Chris@42 72 Tb = T9 * Ta;
Chris@42 73 T4N = T4q - T4p;
Chris@42 74 T4r = T4p + T4q;
Chris@42 75 T8 = T1 + T7;
Chris@42 76 T2i = T1 - T7;
Chris@42 77 Tj = ii[WS(rs, 15)];
Chris@42 78 Tf = W[28];
Chris@42 79 }
Chris@42 80 Tc = W[9];
Chris@42 81 Ti = W[29];
Chris@42 82 {
Chris@42 83 E T3d, Ts, T2t, TL, TB, TE, TD, T3f, Ty, T2q, TC;
Chris@42 84 {
Chris@42 85 E TH, TK, TJ, T2s, TI;
Chris@42 86 {
Chris@42 87 E To, Tr, Tp, T3c, Tq, TG;
Chris@42 88 {
Chris@42 89 E T2k, Te, T2m, Tk, T2l, Th, Tn;
Chris@42 90 To = ri[WS(rs, 4)];
Chris@42 91 T2l = Tf * Tj;
Chris@42 92 Th = Tf * Tg;
Chris@42 93 T2k = FNMS(Tc, Ta, T2j);
Chris@42 94 Te = FMA(Tc, Td, Tb);
Chris@42 95 T2m = FNMS(Ti, Tg, T2l);
Chris@42 96 Tk = FMA(Ti, Tj, Th);
Chris@42 97 Tr = ii[WS(rs, 4)];
Chris@42 98 Tn = W[6];
Chris@42 99 T4n = T2k + T2m;
Chris@42 100 T2n = T2k - T2m;
Chris@42 101 T4O = Te - Tk;
Chris@42 102 Tl = Te + Tk;
Chris@42 103 Tp = Tn * To;
Chris@42 104 T3c = Tn * Tr;
Chris@42 105 }
Chris@42 106 Tq = W[7];
Chris@42 107 TH = ri[WS(rs, 19)];
Chris@42 108 TK = ii[WS(rs, 19)];
Chris@42 109 TG = W[36];
Chris@42 110 T3d = FNMS(Tq, To, T3c);
Chris@42 111 Ts = FMA(Tq, Tr, Tp);
Chris@42 112 TJ = W[37];
Chris@42 113 T2s = TG * TK;
Chris@42 114 TI = TG * TH;
Chris@42 115 }
Chris@42 116 {
Chris@42 117 E Tu, Tx, Tt, Tw, T3e, Tv, TA;
Chris@42 118 Tu = ri[WS(rs, 14)];
Chris@42 119 Tx = ii[WS(rs, 14)];
Chris@42 120 T2t = FNMS(TJ, TH, T2s);
Chris@42 121 TL = FMA(TJ, TK, TI);
Chris@42 122 Tt = W[26];
Chris@42 123 Tw = W[27];
Chris@42 124 TB = ri[WS(rs, 9)];
Chris@42 125 TE = ii[WS(rs, 9)];
Chris@42 126 T3e = Tt * Tx;
Chris@42 127 Tv = Tt * Tu;
Chris@42 128 TA = W[16];
Chris@42 129 TD = W[17];
Chris@42 130 T3f = FNMS(Tw, Tu, T3e);
Chris@42 131 Ty = FMA(Tw, Tx, Tv);
Chris@42 132 T2q = TA * TE;
Chris@42 133 TC = TA * TB;
Chris@42 134 }
Chris@42 135 }
Chris@42 136 {
Chris@42 137 E T3g, T3Y, Tz, T2p, T2r, TF;
Chris@42 138 T3g = T3d - T3f;
Chris@42 139 T3Y = T3d + T3f;
Chris@42 140 Tz = Ts + Ty;
Chris@42 141 T2p = Ts - Ty;
Chris@42 142 T2r = FNMS(TD, TB, T2q);
Chris@42 143 TF = FMA(TD, TE, TC);
Chris@42 144 {
Chris@42 145 E T3Z, T2u, T3h, TM;
Chris@42 146 T3Z = T2r + T2t;
Chris@42 147 T2u = T2r - T2t;
Chris@42 148 T3h = TF - TL;
Chris@42 149 TM = TF + TL;
Chris@42 150 T2v = T2p - T2u;
Chris@42 151 T3v = T2p + T2u;
Chris@42 152 T40 = T3Y - T3Z;
Chris@42 153 T4b = T3Y + T3Z;
Chris@42 154 TN = Tz - TM;
Chris@42 155 T2b = Tz + TM;
Chris@42 156 T3F = T3g - T3h;
Chris@42 157 T3i = T3g + T3h;
Chris@42 158 }
Chris@42 159 }
Chris@42 160 }
Chris@42 161 }
Chris@42 162 }
Chris@42 163 {
Chris@42 164 E T35, T1M, T2P, T25, T1V, T1Y, T1X, T37, T1S, T2M, T1W;
Chris@42 165 {
Chris@42 166 E T21, T24, T23, T2O, T22;
Chris@42 167 {
Chris@42 168 E T1I, T1L, T1H, T1K, T34, T1J, T20;
Chris@42 169 T1I = ri[WS(rs, 12)];
Chris@42 170 T1L = ii[WS(rs, 12)];
Chris@42 171 T1H = W[22];
Chris@42 172 T1K = W[23];
Chris@42 173 T21 = ri[WS(rs, 7)];
Chris@42 174 T24 = ii[WS(rs, 7)];
Chris@42 175 T34 = T1H * T1L;
Chris@42 176 T1J = T1H * T1I;
Chris@42 177 T20 = W[12];
Chris@42 178 T23 = W[13];
Chris@42 179 T35 = FNMS(T1K, T1I, T34);
Chris@42 180 T1M = FMA(T1K, T1L, T1J);
Chris@42 181 T2O = T20 * T24;
Chris@42 182 T22 = T20 * T21;
Chris@42 183 }
Chris@42 184 {
Chris@42 185 E T1O, T1R, T1N, T1Q, T36, T1P, T1U;
Chris@42 186 T1O = ri[WS(rs, 2)];
Chris@42 187 T1R = ii[WS(rs, 2)];
Chris@42 188 T2P = FNMS(T23, T21, T2O);
Chris@42 189 T25 = FMA(T23, T24, T22);
Chris@42 190 T1N = W[2];
Chris@42 191 T1Q = W[3];
Chris@42 192 T1V = ri[WS(rs, 17)];
Chris@42 193 T1Y = ii[WS(rs, 17)];
Chris@42 194 T36 = T1N * T1R;
Chris@42 195 T1P = T1N * T1O;
Chris@42 196 T1U = W[32];
Chris@42 197 T1X = W[33];
Chris@42 198 T37 = FNMS(T1Q, T1O, T36);
Chris@42 199 T1S = FMA(T1Q, T1R, T1P);
Chris@42 200 T2M = T1U * T1Y;
Chris@42 201 T1W = T1U * T1V;
Chris@42 202 }
Chris@42 203 }
Chris@42 204 {
Chris@42 205 E T38, T3U, T1T, T2L, T2N, T1Z;
Chris@42 206 T38 = T35 - T37;
Chris@42 207 T3U = T35 + T37;
Chris@42 208 T1T = T1M + T1S;
Chris@42 209 T2L = T1M - T1S;
Chris@42 210 T2N = FNMS(T1X, T1V, T2M);
Chris@42 211 T1Z = FMA(T1X, T1Y, T1W);
Chris@42 212 {
Chris@42 213 E T3V, T2Q, T39, T26;
Chris@42 214 T3V = T2N + T2P;
Chris@42 215 T2Q = T2N - T2P;
Chris@42 216 T39 = T1Z - T25;
Chris@42 217 T26 = T1Z + T25;
Chris@42 218 T2R = T2L - T2Q;
Chris@42 219 T3z = T2L + T2Q;
Chris@42 220 T3W = T3U - T3V;
Chris@42 221 T4f = T3U + T3V;
Chris@42 222 T27 = T1T - T26;
Chris@42 223 T2f = T1T + T26;
Chris@42 224 T3J = T38 - T39;
Chris@42 225 T3a = T38 + T39;
Chris@42 226 }
Chris@42 227 }
Chris@42 228 }
Chris@42 229 {
Chris@42 230 E T2Y, T1l, T2I, T1E, T1u, T1x, T1w, T30, T1r, T2F, T1v;
Chris@42 231 {
Chris@42 232 E T1A, T1D, T1C, T2H, T1B;
Chris@42 233 {
Chris@42 234 E T1h, T1k, T1g, T1j, T2X, T1i, T1z;
Chris@42 235 T1h = ri[WS(rs, 8)];
Chris@42 236 T1k = ii[WS(rs, 8)];
Chris@42 237 T1g = W[14];
Chris@42 238 T1j = W[15];
Chris@42 239 T1A = ri[WS(rs, 3)];
Chris@42 240 T1D = ii[WS(rs, 3)];
Chris@42 241 T2X = T1g * T1k;
Chris@42 242 T1i = T1g * T1h;
Chris@42 243 T1z = W[4];
Chris@42 244 T1C = W[5];
Chris@42 245 T2Y = FNMS(T1j, T1h, T2X);
Chris@42 246 T1l = FMA(T1j, T1k, T1i);
Chris@42 247 T2H = T1z * T1D;
Chris@42 248 T1B = T1z * T1A;
Chris@42 249 }
Chris@42 250 {
Chris@42 251 E T1n, T1q, T1m, T1p, T2Z, T1o, T1t;
Chris@42 252 T1n = ri[WS(rs, 18)];
Chris@42 253 T1q = ii[WS(rs, 18)];
Chris@42 254 T2I = FNMS(T1C, T1A, T2H);
Chris@42 255 T1E = FMA(T1C, T1D, T1B);
Chris@42 256 T1m = W[34];
Chris@42 257 T1p = W[35];
Chris@42 258 T1u = ri[WS(rs, 13)];
Chris@42 259 T1x = ii[WS(rs, 13)];
Chris@42 260 T2Z = T1m * T1q;
Chris@42 261 T1o = T1m * T1n;
Chris@42 262 T1t = W[24];
Chris@42 263 T1w = W[25];
Chris@42 264 T30 = FNMS(T1p, T1n, T2Z);
Chris@42 265 T1r = FMA(T1p, T1q, T1o);
Chris@42 266 T2F = T1t * T1x;
Chris@42 267 T1v = T1t * T1u;
Chris@42 268 }
Chris@42 269 }
Chris@42 270 {
Chris@42 271 E T31, T3R, T1s, T2E, T2G, T1y;
Chris@42 272 T31 = T2Y - T30;
Chris@42 273 T3R = T2Y + T30;
Chris@42 274 T1s = T1l + T1r;
Chris@42 275 T2E = T1l - T1r;
Chris@42 276 T2G = FNMS(T1w, T1u, T2F);
Chris@42 277 T1y = FMA(T1w, T1x, T1v);
Chris@42 278 {
Chris@42 279 E T3S, T2J, T32, T1F;
Chris@42 280 T3S = T2G + T2I;
Chris@42 281 T2J = T2G - T2I;
Chris@42 282 T32 = T1y - T1E;
Chris@42 283 T1F = T1y + T1E;
Chris@42 284 T2K = T2E - T2J;
Chris@42 285 T3y = T2E + T2J;
Chris@42 286 T3T = T3R - T3S;
Chris@42 287 T4e = T3R + T3S;
Chris@42 288 T1G = T1s - T1F;
Chris@42 289 T2e = T1s + T1F;
Chris@42 290 T3I = T31 - T32;
Chris@42 291 T33 = T31 + T32;
Chris@42 292 }
Chris@42 293 }
Chris@42 294 }
Chris@42 295 {
Chris@42 296 E T3k, TT, T2A, T1c, T12, T15, T14, T3m, TZ, T2x, T13;
Chris@42 297 {
Chris@42 298 E T18, T1b, T1a, T2z, T19;
Chris@42 299 {
Chris@42 300 E TP, TS, TO, TR, T3j, TQ, T17;
Chris@42 301 TP = ri[WS(rs, 16)];
Chris@42 302 TS = ii[WS(rs, 16)];
Chris@42 303 TO = W[30];
Chris@42 304 TR = W[31];
Chris@42 305 T18 = ri[WS(rs, 11)];
Chris@42 306 T1b = ii[WS(rs, 11)];
Chris@42 307 T3j = TO * TS;
Chris@42 308 TQ = TO * TP;
Chris@42 309 T17 = W[20];
Chris@42 310 T1a = W[21];
Chris@42 311 T3k = FNMS(TR, TP, T3j);
Chris@42 312 TT = FMA(TR, TS, TQ);
Chris@42 313 T2z = T17 * T1b;
Chris@42 314 T19 = T17 * T18;
Chris@42 315 }
Chris@42 316 {
Chris@42 317 E TV, TY, TU, TX, T3l, TW, T11;
Chris@42 318 TV = ri[WS(rs, 6)];
Chris@42 319 TY = ii[WS(rs, 6)];
Chris@42 320 T2A = FNMS(T1a, T18, T2z);
Chris@42 321 T1c = FMA(T1a, T1b, T19);
Chris@42 322 TU = W[10];
Chris@42 323 TX = W[11];
Chris@42 324 T12 = ri[WS(rs, 1)];
Chris@42 325 T15 = ii[WS(rs, 1)];
Chris@42 326 T3l = TU * TY;
Chris@42 327 TW = TU * TV;
Chris@42 328 T11 = W[0];
Chris@42 329 T14 = W[1];
Chris@42 330 T3m = FNMS(TX, TV, T3l);
Chris@42 331 TZ = FMA(TX, TY, TW);
Chris@42 332 T2x = T11 * T15;
Chris@42 333 T13 = T11 * T12;
Chris@42 334 }
Chris@42 335 }
Chris@42 336 {
Chris@42 337 E T3n, T41, T10, T2w, T2y, T16;
Chris@42 338 T3n = T3k - T3m;
Chris@42 339 T41 = T3k + T3m;
Chris@42 340 T10 = TT + TZ;
Chris@42 341 T2w = TT - TZ;
Chris@42 342 T2y = FNMS(T14, T12, T2x);
Chris@42 343 T16 = FMA(T14, T15, T13);
Chris@42 344 {
Chris@42 345 E T42, T2B, T3o, T1d;
Chris@42 346 T42 = T2y + T2A;
Chris@42 347 T2B = T2y - T2A;
Chris@42 348 T3o = T16 - T1c;
Chris@42 349 T1d = T16 + T1c;
Chris@42 350 T2C = T2w - T2B;
Chris@42 351 T3w = T2w + T2B;
Chris@42 352 T43 = T41 - T42;
Chris@42 353 T4c = T41 + T42;
Chris@42 354 T1e = T10 - T1d;
Chris@42 355 T2c = T10 + T1d;
Chris@42 356 T3G = T3n - T3o;
Chris@42 357 T3p = T3n + T3o;
Chris@42 358 }
Chris@42 359 }
Chris@42 360 }
Chris@42 361 {
Chris@42 362 E T4s, T4k, T4l, T4h, T4j, T49, T4y, T4A, T48;
Chris@42 363 {
Chris@42 364 E T4D, T4C, T2a, T47, T45, T4B, T4M, T4K, T46, T3Q;
Chris@42 365 {
Chris@42 366 E Tm, T1f, T4J, T4I, T28, T3X, T44, T29, T3P, T3O;
Chris@42 367 T4D = T3T + T3W;
Chris@42 368 T3X = T3T - T3W;
Chris@42 369 T44 = T40 - T43;
Chris@42 370 T4C = T40 + T43;
Chris@42 371 T2a = T8 + Tl;
Chris@42 372 Tm = T8 - Tl;
Chris@42 373 T1f = TN + T1e;
Chris@42 374 T4J = TN - T1e;
Chris@42 375 T4I = T1G - T27;
Chris@42 376 T28 = T1G + T27;
Chris@42 377 T47 = FMA(KP618033988, T3X, T44);
Chris@42 378 T45 = FNMS(KP618033988, T44, T3X);
Chris@42 379 T29 = T1f + T28;
Chris@42 380 T3P = T1f - T28;
Chris@42 381 T4B = T4r - T4n;
Chris@42 382 T4s = T4n + T4r;
Chris@42 383 ri[WS(rs, 10)] = Tm + T29;
Chris@42 384 T3O = FNMS(KP250000000, T29, Tm);
Chris@42 385 T4M = FMA(KP618033988, T4I, T4J);
Chris@42 386 T4K = FNMS(KP618033988, T4J, T4I);
Chris@42 387 T46 = FMA(KP559016994, T3P, T3O);
Chris@42 388 T3Q = FNMS(KP559016994, T3P, T3O);
Chris@42 389 }
Chris@42 390 {
Chris@42 391 E T2d, T4w, T4x, T2g, T2h;
Chris@42 392 {
Chris@42 393 E T4d, T4G, T4F, T4g, T4E, T4L, T4H;
Chris@42 394 T4k = T4b + T4c;
Chris@42 395 T4d = T4b - T4c;
Chris@42 396 T4G = T4C - T4D;
Chris@42 397 T4E = T4C + T4D;
Chris@42 398 ri[WS(rs, 18)] = FMA(KP951056516, T45, T3Q);
Chris@42 399 ri[WS(rs, 2)] = FNMS(KP951056516, T45, T3Q);
Chris@42 400 ri[WS(rs, 6)] = FMA(KP951056516, T47, T46);
Chris@42 401 ri[WS(rs, 14)] = FNMS(KP951056516, T47, T46);
Chris@42 402 ii[WS(rs, 10)] = T4E + T4B;
Chris@42 403 T4F = FNMS(KP250000000, T4E, T4B);
Chris@42 404 T4g = T4e - T4f;
Chris@42 405 T4l = T4e + T4f;
Chris@42 406 T2d = T2b + T2c;
Chris@42 407 T4w = T2b - T2c;
Chris@42 408 T4L = FMA(KP559016994, T4G, T4F);
Chris@42 409 T4H = FNMS(KP559016994, T4G, T4F);
Chris@42 410 T4h = FMA(KP618033988, T4g, T4d);
Chris@42 411 T4j = FNMS(KP618033988, T4d, T4g);
Chris@42 412 ii[WS(rs, 18)] = FNMS(KP951056516, T4K, T4H);
Chris@42 413 ii[WS(rs, 2)] = FMA(KP951056516, T4K, T4H);
Chris@42 414 ii[WS(rs, 14)] = FMA(KP951056516, T4M, T4L);
Chris@42 415 ii[WS(rs, 6)] = FNMS(KP951056516, T4M, T4L);
Chris@42 416 T4x = T2e - T2f;
Chris@42 417 T2g = T2e + T2f;
Chris@42 418 }
Chris@42 419 T2h = T2d + T2g;
Chris@42 420 T49 = T2d - T2g;
Chris@42 421 T4y = FMA(KP618033988, T4x, T4w);
Chris@42 422 T4A = FNMS(KP618033988, T4w, T4x);
Chris@42 423 ri[0] = T2a + T2h;
Chris@42 424 T48 = FNMS(KP250000000, T2h, T2a);
Chris@42 425 }
Chris@42 426 }
Chris@42 427 {
Chris@42 428 E T3u, T51, T5a, T5c, T56, T54;
Chris@42 429 {
Chris@42 430 E T53, T52, T3t, T3r, T2o, T59, T58, T2T, T2V, T4u, T4t, T2U, T3s, T2W;
Chris@42 431 {
Chris@42 432 E T3b, T3q, T4i, T4a, T4m;
Chris@42 433 T53 = T33 + T3a;
Chris@42 434 T3b = T33 - T3a;
Chris@42 435 T3q = T3i - T3p;
Chris@42 436 T52 = T3i + T3p;
Chris@42 437 T4i = FNMS(KP559016994, T49, T48);
Chris@42 438 T4a = FMA(KP559016994, T49, T48);
Chris@42 439 T4m = T4k + T4l;
Chris@42 440 T4u = T4k - T4l;
Chris@42 441 ri[WS(rs, 16)] = FMA(KP951056516, T4h, T4a);
Chris@42 442 ri[WS(rs, 4)] = FNMS(KP951056516, T4h, T4a);
Chris@42 443 ri[WS(rs, 8)] = FMA(KP951056516, T4j, T4i);
Chris@42 444 ri[WS(rs, 12)] = FNMS(KP951056516, T4j, T4i);
Chris@42 445 ii[0] = T4m + T4s;
Chris@42 446 T4t = FNMS(KP250000000, T4m, T4s);
Chris@42 447 T3t = FMA(KP618033988, T3b, T3q);
Chris@42 448 T3r = FNMS(KP618033988, T3q, T3b);
Chris@42 449 }
Chris@42 450 T3u = T2i + T2n;
Chris@42 451 T2o = T2i - T2n;
Chris@42 452 {
Chris@42 453 E T4v, T4z, T2D, T2S;
Chris@42 454 T4v = FMA(KP559016994, T4u, T4t);
Chris@42 455 T4z = FNMS(KP559016994, T4u, T4t);
Chris@42 456 T2D = T2v + T2C;
Chris@42 457 T59 = T2v - T2C;
Chris@42 458 T58 = T2K - T2R;
Chris@42 459 T2S = T2K + T2R;
Chris@42 460 ii[WS(rs, 16)] = FNMS(KP951056516, T4y, T4v);
Chris@42 461 ii[WS(rs, 4)] = FMA(KP951056516, T4y, T4v);
Chris@42 462 ii[WS(rs, 12)] = FMA(KP951056516, T4A, T4z);
Chris@42 463 ii[WS(rs, 8)] = FNMS(KP951056516, T4A, T4z);
Chris@42 464 T2T = T2D + T2S;
Chris@42 465 T2V = T2D - T2S;
Chris@42 466 }
Chris@42 467 ri[WS(rs, 15)] = T2o + T2T;
Chris@42 468 T2U = FNMS(KP250000000, T2T, T2o);
Chris@42 469 T51 = T4O + T4N;
Chris@42 470 T4P = T4N - T4O;
Chris@42 471 T5a = FNMS(KP618033988, T59, T58);
Chris@42 472 T5c = FMA(KP618033988, T58, T59);
Chris@42 473 T3s = FMA(KP559016994, T2V, T2U);
Chris@42 474 T2W = FNMS(KP559016994, T2V, T2U);
Chris@42 475 ri[WS(rs, 7)] = FNMS(KP951056516, T3r, T2W);
Chris@42 476 ri[WS(rs, 3)] = FMA(KP951056516, T3r, T2W);
Chris@42 477 ri[WS(rs, 19)] = FNMS(KP951056516, T3t, T3s);
Chris@42 478 ri[WS(rs, 11)] = FMA(KP951056516, T3t, T3s);
Chris@42 479 T56 = T52 - T53;
Chris@42 480 T54 = T52 + T53;
Chris@42 481 }
Chris@42 482 {
Chris@42 483 E T4Q, T4R, T3N, T3L, T4W, T4X, T3B, T3D, T3H, T3K, T55, T3C, T3M, T3E;
Chris@42 484 T4Q = T3F + T3G;
Chris@42 485 T3H = T3F - T3G;
Chris@42 486 T3K = T3I - T3J;
Chris@42 487 T4R = T3I + T3J;
Chris@42 488 ii[WS(rs, 15)] = T54 + T51;
Chris@42 489 T55 = FNMS(KP250000000, T54, T51);
Chris@42 490 T3N = FNMS(KP618033988, T3H, T3K);
Chris@42 491 T3L = FMA(KP618033988, T3K, T3H);
Chris@42 492 {
Chris@42 493 E T57, T5b, T3x, T3A;
Chris@42 494 T57 = FNMS(KP559016994, T56, T55);
Chris@42 495 T5b = FMA(KP559016994, T56, T55);
Chris@42 496 T3x = T3v + T3w;
Chris@42 497 T4W = T3v - T3w;
Chris@42 498 T4X = T3y - T3z;
Chris@42 499 T3A = T3y + T3z;
Chris@42 500 ii[WS(rs, 7)] = FMA(KP951056516, T5a, T57);
Chris@42 501 ii[WS(rs, 3)] = FNMS(KP951056516, T5a, T57);
Chris@42 502 ii[WS(rs, 19)] = FMA(KP951056516, T5c, T5b);
Chris@42 503 ii[WS(rs, 11)] = FNMS(KP951056516, T5c, T5b);
Chris@42 504 T3B = T3x + T3A;
Chris@42 505 T3D = T3x - T3A;
Chris@42 506 }
Chris@42 507 ri[WS(rs, 5)] = T3u + T3B;
Chris@42 508 T3C = FNMS(KP250000000, T3B, T3u);
Chris@42 509 T4Y = FMA(KP618033988, T4X, T4W);
Chris@42 510 T50 = FNMS(KP618033988, T4W, T4X);
Chris@42 511 T3M = FNMS(KP559016994, T3D, T3C);
Chris@42 512 T3E = FMA(KP559016994, T3D, T3C);
Chris@42 513 ri[WS(rs, 9)] = FNMS(KP951056516, T3L, T3E);
Chris@42 514 ri[WS(rs, 1)] = FMA(KP951056516, T3L, T3E);
Chris@42 515 ri[WS(rs, 17)] = FNMS(KP951056516, T3N, T3M);
Chris@42 516 ri[WS(rs, 13)] = FMA(KP951056516, T3N, T3M);
Chris@42 517 T4U = T4Q - T4R;
Chris@42 518 T4S = T4Q + T4R;
Chris@42 519 }
Chris@42 520 }
Chris@42 521 }
Chris@42 522 }
Chris@42 523 ii[WS(rs, 5)] = T4S + T4P;
Chris@42 524 T4T = FNMS(KP250000000, T4S, T4P);
Chris@42 525 T4Z = FNMS(KP559016994, T4U, T4T);
Chris@42 526 T4V = FMA(KP559016994, T4U, T4T);
Chris@42 527 ii[WS(rs, 9)] = FMA(KP951056516, T4Y, T4V);
Chris@42 528 ii[WS(rs, 1)] = FNMS(KP951056516, T4Y, T4V);
Chris@42 529 ii[WS(rs, 17)] = FMA(KP951056516, T50, T4Z);
Chris@42 530 ii[WS(rs, 13)] = FNMS(KP951056516, T50, T4Z);
Chris@42 531 }
Chris@42 532 }
Chris@42 533 }
Chris@42 534
Chris@42 535 static const tw_instr twinstr[] = {
Chris@42 536 {TW_FULL, 0, 20},
Chris@42 537 {TW_NEXT, 1, 0}
Chris@42 538 };
Chris@42 539
Chris@42 540 static const ct_desc desc = { 20, "t1_20", twinstr, &GENUS, {136, 38, 110, 0}, 0, 0, 0 };
Chris@42 541
Chris@42 542 void X(codelet_t1_20) (planner *p) {
Chris@42 543 X(kdft_dit_register) (p, t1_20, &desc);
Chris@42 544 }
Chris@42 545 #else /* HAVE_FMA */
Chris@42 546
Chris@42 547 /* Generated by: ../../../genfft/gen_twiddle.native -compact -variables 4 -pipeline-latency 4 -n 20 -name t1_20 -include t.h */
Chris@42 548
Chris@42 549 /*
Chris@42 550 * This function contains 246 FP additions, 124 FP multiplications,
Chris@42 551 * (or, 184 additions, 62 multiplications, 62 fused multiply/add),
Chris@42 552 * 85 stack variables, 4 constants, and 80 memory accesses
Chris@42 553 */
Chris@42 554 #include "t.h"
Chris@42 555
Chris@42 556 static void t1_20(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
Chris@42 557 {
Chris@42 558 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
Chris@42 559 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
Chris@42 560 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
Chris@42 561 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
Chris@42 562 {
Chris@42 563 INT m;
Chris@42 564 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)) {
Chris@42 565 E Tj, T1R, T4g, T4p, T2q, T37, T3Q, T42, T1r, T1O, T1P, T3i, T3l, T44, T3D;
Chris@42 566 E T3E, T3K, T1V, T1W, T1X, T23, T28, T4r, T2W, T2X, T4c, T33, T34, T35, T2G;
Chris@42 567 E T2L, T2M, TG, T13, T14, T3p, T3s, T43, T3A, T3B, T3J, T1S, T1T, T1U, T2e;
Chris@42 568 E T2j, T4q, T2T, T2U, T4b, T30, T31, T32, T2v, T2A, T2B;
Chris@42 569 {
Chris@42 570 E T1, T3O, T6, T3N, Tc, T2n, Th, T2o;
Chris@42 571 T1 = ri[0];
Chris@42 572 T3O = ii[0];
Chris@42 573 {
Chris@42 574 E T3, T5, T2, T4;
Chris@42 575 T3 = ri[WS(rs, 10)];
Chris@42 576 T5 = ii[WS(rs, 10)];
Chris@42 577 T2 = W[18];
Chris@42 578 T4 = W[19];
Chris@42 579 T6 = FMA(T2, T3, T4 * T5);
Chris@42 580 T3N = FNMS(T4, T3, T2 * T5);
Chris@42 581 }
Chris@42 582 {
Chris@42 583 E T9, Tb, T8, Ta;
Chris@42 584 T9 = ri[WS(rs, 5)];
Chris@42 585 Tb = ii[WS(rs, 5)];
Chris@42 586 T8 = W[8];
Chris@42 587 Ta = W[9];
Chris@42 588 Tc = FMA(T8, T9, Ta * Tb);
Chris@42 589 T2n = FNMS(Ta, T9, T8 * Tb);
Chris@42 590 }
Chris@42 591 {
Chris@42 592 E Te, Tg, Td, Tf;
Chris@42 593 Te = ri[WS(rs, 15)];
Chris@42 594 Tg = ii[WS(rs, 15)];
Chris@42 595 Td = W[28];
Chris@42 596 Tf = W[29];
Chris@42 597 Th = FMA(Td, Te, Tf * Tg);
Chris@42 598 T2o = FNMS(Tf, Te, Td * Tg);
Chris@42 599 }
Chris@42 600 {
Chris@42 601 E T7, Ti, T4e, T4f;
Chris@42 602 T7 = T1 + T6;
Chris@42 603 Ti = Tc + Th;
Chris@42 604 Tj = T7 - Ti;
Chris@42 605 T1R = T7 + Ti;
Chris@42 606 T4e = T3O - T3N;
Chris@42 607 T4f = Tc - Th;
Chris@42 608 T4g = T4e - T4f;
Chris@42 609 T4p = T4f + T4e;
Chris@42 610 }
Chris@42 611 {
Chris@42 612 E T2m, T2p, T3M, T3P;
Chris@42 613 T2m = T1 - T6;
Chris@42 614 T2p = T2n - T2o;
Chris@42 615 T2q = T2m - T2p;
Chris@42 616 T37 = T2m + T2p;
Chris@42 617 T3M = T2n + T2o;
Chris@42 618 T3P = T3N + T3O;
Chris@42 619 T3Q = T3M + T3P;
Chris@42 620 T42 = T3P - T3M;
Chris@42 621 }
Chris@42 622 }
Chris@42 623 {
Chris@42 624 E T1f, T3g, T21, T2C, T1N, T3k, T27, T2K, T1q, T3h, T22, T2F, T1C, T3j, T26;
Chris@42 625 E T2H;
Chris@42 626 {
Chris@42 627 E T19, T1Z, T1e, T20;
Chris@42 628 {
Chris@42 629 E T16, T18, T15, T17;
Chris@42 630 T16 = ri[WS(rs, 8)];
Chris@42 631 T18 = ii[WS(rs, 8)];
Chris@42 632 T15 = W[14];
Chris@42 633 T17 = W[15];
Chris@42 634 T19 = FMA(T15, T16, T17 * T18);
Chris@42 635 T1Z = FNMS(T17, T16, T15 * T18);
Chris@42 636 }
Chris@42 637 {
Chris@42 638 E T1b, T1d, T1a, T1c;
Chris@42 639 T1b = ri[WS(rs, 18)];
Chris@42 640 T1d = ii[WS(rs, 18)];
Chris@42 641 T1a = W[34];
Chris@42 642 T1c = W[35];
Chris@42 643 T1e = FMA(T1a, T1b, T1c * T1d);
Chris@42 644 T20 = FNMS(T1c, T1b, T1a * T1d);
Chris@42 645 }
Chris@42 646 T1f = T19 + T1e;
Chris@42 647 T3g = T1Z + T20;
Chris@42 648 T21 = T1Z - T20;
Chris@42 649 T2C = T19 - T1e;
Chris@42 650 }
Chris@42 651 {
Chris@42 652 E T1H, T2I, T1M, T2J;
Chris@42 653 {
Chris@42 654 E T1E, T1G, T1D, T1F;
Chris@42 655 T1E = ri[WS(rs, 17)];
Chris@42 656 T1G = ii[WS(rs, 17)];
Chris@42 657 T1D = W[32];
Chris@42 658 T1F = W[33];
Chris@42 659 T1H = FMA(T1D, T1E, T1F * T1G);
Chris@42 660 T2I = FNMS(T1F, T1E, T1D * T1G);
Chris@42 661 }
Chris@42 662 {
Chris@42 663 E T1J, T1L, T1I, T1K;
Chris@42 664 T1J = ri[WS(rs, 7)];
Chris@42 665 T1L = ii[WS(rs, 7)];
Chris@42 666 T1I = W[12];
Chris@42 667 T1K = W[13];
Chris@42 668 T1M = FMA(T1I, T1J, T1K * T1L);
Chris@42 669 T2J = FNMS(T1K, T1J, T1I * T1L);
Chris@42 670 }
Chris@42 671 T1N = T1H + T1M;
Chris@42 672 T3k = T2I + T2J;
Chris@42 673 T27 = T1H - T1M;
Chris@42 674 T2K = T2I - T2J;
Chris@42 675 }
Chris@42 676 {
Chris@42 677 E T1k, T2D, T1p, T2E;
Chris@42 678 {
Chris@42 679 E T1h, T1j, T1g, T1i;
Chris@42 680 T1h = ri[WS(rs, 13)];
Chris@42 681 T1j = ii[WS(rs, 13)];
Chris@42 682 T1g = W[24];
Chris@42 683 T1i = W[25];
Chris@42 684 T1k = FMA(T1g, T1h, T1i * T1j);
Chris@42 685 T2D = FNMS(T1i, T1h, T1g * T1j);
Chris@42 686 }
Chris@42 687 {
Chris@42 688 E T1m, T1o, T1l, T1n;
Chris@42 689 T1m = ri[WS(rs, 3)];
Chris@42 690 T1o = ii[WS(rs, 3)];
Chris@42 691 T1l = W[4];
Chris@42 692 T1n = W[5];
Chris@42 693 T1p = FMA(T1l, T1m, T1n * T1o);
Chris@42 694 T2E = FNMS(T1n, T1m, T1l * T1o);
Chris@42 695 }
Chris@42 696 T1q = T1k + T1p;
Chris@42 697 T3h = T2D + T2E;
Chris@42 698 T22 = T1k - T1p;
Chris@42 699 T2F = T2D - T2E;
Chris@42 700 }
Chris@42 701 {
Chris@42 702 E T1w, T24, T1B, T25;
Chris@42 703 {
Chris@42 704 E T1t, T1v, T1s, T1u;
Chris@42 705 T1t = ri[WS(rs, 12)];
Chris@42 706 T1v = ii[WS(rs, 12)];
Chris@42 707 T1s = W[22];
Chris@42 708 T1u = W[23];
Chris@42 709 T1w = FMA(T1s, T1t, T1u * T1v);
Chris@42 710 T24 = FNMS(T1u, T1t, T1s * T1v);
Chris@42 711 }
Chris@42 712 {
Chris@42 713 E T1y, T1A, T1x, T1z;
Chris@42 714 T1y = ri[WS(rs, 2)];
Chris@42 715 T1A = ii[WS(rs, 2)];
Chris@42 716 T1x = W[2];
Chris@42 717 T1z = W[3];
Chris@42 718 T1B = FMA(T1x, T1y, T1z * T1A);
Chris@42 719 T25 = FNMS(T1z, T1y, T1x * T1A);
Chris@42 720 }
Chris@42 721 T1C = T1w + T1B;
Chris@42 722 T3j = T24 + T25;
Chris@42 723 T26 = T24 - T25;
Chris@42 724 T2H = T1w - T1B;
Chris@42 725 }
Chris@42 726 T1r = T1f - T1q;
Chris@42 727 T1O = T1C - T1N;
Chris@42 728 T1P = T1r + T1O;
Chris@42 729 T3i = T3g - T3h;
Chris@42 730 T3l = T3j - T3k;
Chris@42 731 T44 = T3i + T3l;
Chris@42 732 T3D = T3g + T3h;
Chris@42 733 T3E = T3j + T3k;
Chris@42 734 T3K = T3D + T3E;
Chris@42 735 T1V = T1f + T1q;
Chris@42 736 T1W = T1C + T1N;
Chris@42 737 T1X = T1V + T1W;
Chris@42 738 T23 = T21 + T22;
Chris@42 739 T28 = T26 + T27;
Chris@42 740 T4r = T23 + T28;
Chris@42 741 T2W = T21 - T22;
Chris@42 742 T2X = T26 - T27;
Chris@42 743 T4c = T2W + T2X;
Chris@42 744 T33 = T2C + T2F;
Chris@42 745 T34 = T2H + T2K;
Chris@42 746 T35 = T33 + T34;
Chris@42 747 T2G = T2C - T2F;
Chris@42 748 T2L = T2H - T2K;
Chris@42 749 T2M = T2G + T2L;
Chris@42 750 }
Chris@42 751 {
Chris@42 752 E Tu, T3n, T2c, T2r, T12, T3r, T2i, T2z, TF, T3o, T2d, T2u, TR, T3q, T2h;
Chris@42 753 E T2w;
Chris@42 754 {
Chris@42 755 E To, T2a, Tt, T2b;
Chris@42 756 {
Chris@42 757 E Tl, Tn, Tk, Tm;
Chris@42 758 Tl = ri[WS(rs, 4)];
Chris@42 759 Tn = ii[WS(rs, 4)];
Chris@42 760 Tk = W[6];
Chris@42 761 Tm = W[7];
Chris@42 762 To = FMA(Tk, Tl, Tm * Tn);
Chris@42 763 T2a = FNMS(Tm, Tl, Tk * Tn);
Chris@42 764 }
Chris@42 765 {
Chris@42 766 E Tq, Ts, Tp, Tr;
Chris@42 767 Tq = ri[WS(rs, 14)];
Chris@42 768 Ts = ii[WS(rs, 14)];
Chris@42 769 Tp = W[26];
Chris@42 770 Tr = W[27];
Chris@42 771 Tt = FMA(Tp, Tq, Tr * Ts);
Chris@42 772 T2b = FNMS(Tr, Tq, Tp * Ts);
Chris@42 773 }
Chris@42 774 Tu = To + Tt;
Chris@42 775 T3n = T2a + T2b;
Chris@42 776 T2c = T2a - T2b;
Chris@42 777 T2r = To - Tt;
Chris@42 778 }
Chris@42 779 {
Chris@42 780 E TW, T2x, T11, T2y;
Chris@42 781 {
Chris@42 782 E TT, TV, TS, TU;
Chris@42 783 TT = ri[WS(rs, 1)];
Chris@42 784 TV = ii[WS(rs, 1)];
Chris@42 785 TS = W[0];
Chris@42 786 TU = W[1];
Chris@42 787 TW = FMA(TS, TT, TU * TV);
Chris@42 788 T2x = FNMS(TU, TT, TS * TV);
Chris@42 789 }
Chris@42 790 {
Chris@42 791 E TY, T10, TX, TZ;
Chris@42 792 TY = ri[WS(rs, 11)];
Chris@42 793 T10 = ii[WS(rs, 11)];
Chris@42 794 TX = W[20];
Chris@42 795 TZ = W[21];
Chris@42 796 T11 = FMA(TX, TY, TZ * T10);
Chris@42 797 T2y = FNMS(TZ, TY, TX * T10);
Chris@42 798 }
Chris@42 799 T12 = TW + T11;
Chris@42 800 T3r = T2x + T2y;
Chris@42 801 T2i = TW - T11;
Chris@42 802 T2z = T2x - T2y;
Chris@42 803 }
Chris@42 804 {
Chris@42 805 E Tz, T2s, TE, T2t;
Chris@42 806 {
Chris@42 807 E Tw, Ty, Tv, Tx;
Chris@42 808 Tw = ri[WS(rs, 9)];
Chris@42 809 Ty = ii[WS(rs, 9)];
Chris@42 810 Tv = W[16];
Chris@42 811 Tx = W[17];
Chris@42 812 Tz = FMA(Tv, Tw, Tx * Ty);
Chris@42 813 T2s = FNMS(Tx, Tw, Tv * Ty);
Chris@42 814 }
Chris@42 815 {
Chris@42 816 E TB, TD, TA, TC;
Chris@42 817 TB = ri[WS(rs, 19)];
Chris@42 818 TD = ii[WS(rs, 19)];
Chris@42 819 TA = W[36];
Chris@42 820 TC = W[37];
Chris@42 821 TE = FMA(TA, TB, TC * TD);
Chris@42 822 T2t = FNMS(TC, TB, TA * TD);
Chris@42 823 }
Chris@42 824 TF = Tz + TE;
Chris@42 825 T3o = T2s + T2t;
Chris@42 826 T2d = Tz - TE;
Chris@42 827 T2u = T2s - T2t;
Chris@42 828 }
Chris@42 829 {
Chris@42 830 E TL, T2f, TQ, T2g;
Chris@42 831 {
Chris@42 832 E TI, TK, TH, TJ;
Chris@42 833 TI = ri[WS(rs, 16)];
Chris@42 834 TK = ii[WS(rs, 16)];
Chris@42 835 TH = W[30];
Chris@42 836 TJ = W[31];
Chris@42 837 TL = FMA(TH, TI, TJ * TK);
Chris@42 838 T2f = FNMS(TJ, TI, TH * TK);
Chris@42 839 }
Chris@42 840 {
Chris@42 841 E TN, TP, TM, TO;
Chris@42 842 TN = ri[WS(rs, 6)];
Chris@42 843 TP = ii[WS(rs, 6)];
Chris@42 844 TM = W[10];
Chris@42 845 TO = W[11];
Chris@42 846 TQ = FMA(TM, TN, TO * TP);
Chris@42 847 T2g = FNMS(TO, TN, TM * TP);
Chris@42 848 }
Chris@42 849 TR = TL + TQ;
Chris@42 850 T3q = T2f + T2g;
Chris@42 851 T2h = T2f - T2g;
Chris@42 852 T2w = TL - TQ;
Chris@42 853 }
Chris@42 854 TG = Tu - TF;
Chris@42 855 T13 = TR - T12;
Chris@42 856 T14 = TG + T13;
Chris@42 857 T3p = T3n - T3o;
Chris@42 858 T3s = T3q - T3r;
Chris@42 859 T43 = T3p + T3s;
Chris@42 860 T3A = T3n + T3o;
Chris@42 861 T3B = T3q + T3r;
Chris@42 862 T3J = T3A + T3B;
Chris@42 863 T1S = Tu + TF;
Chris@42 864 T1T = TR + T12;
Chris@42 865 T1U = T1S + T1T;
Chris@42 866 T2e = T2c + T2d;
Chris@42 867 T2j = T2h + T2i;
Chris@42 868 T4q = T2e + T2j;
Chris@42 869 T2T = T2c - T2d;
Chris@42 870 T2U = T2h - T2i;
Chris@42 871 T4b = T2T + T2U;
Chris@42 872 T30 = T2r + T2u;
Chris@42 873 T31 = T2w + T2z;
Chris@42 874 T32 = T30 + T31;
Chris@42 875 T2v = T2r - T2u;
Chris@42 876 T2A = T2w - T2z;
Chris@42 877 T2B = T2v + T2A;
Chris@42 878 }
Chris@42 879 {
Chris@42 880 E T3e, T1Q, T3d, T3u, T3w, T3m, T3t, T3v, T3f;
Chris@42 881 T3e = KP559016994 * (T14 - T1P);
Chris@42 882 T1Q = T14 + T1P;
Chris@42 883 T3d = FNMS(KP250000000, T1Q, Tj);
Chris@42 884 T3m = T3i - T3l;
Chris@42 885 T3t = T3p - T3s;
Chris@42 886 T3u = FNMS(KP587785252, T3t, KP951056516 * T3m);
Chris@42 887 T3w = FMA(KP951056516, T3t, KP587785252 * T3m);
Chris@42 888 ri[WS(rs, 10)] = Tj + T1Q;
Chris@42 889 T3v = T3e + T3d;
Chris@42 890 ri[WS(rs, 14)] = T3v - T3w;
Chris@42 891 ri[WS(rs, 6)] = T3v + T3w;
Chris@42 892 T3f = T3d - T3e;
Chris@42 893 ri[WS(rs, 2)] = T3f - T3u;
Chris@42 894 ri[WS(rs, 18)] = T3f + T3u;
Chris@42 895 }
Chris@42 896 {
Chris@42 897 E T47, T45, T46, T41, T4a, T3Z, T40, T49, T48;
Chris@42 898 T47 = KP559016994 * (T43 - T44);
Chris@42 899 T45 = T43 + T44;
Chris@42 900 T46 = FNMS(KP250000000, T45, T42);
Chris@42 901 T3Z = T1r - T1O;
Chris@42 902 T40 = TG - T13;
Chris@42 903 T41 = FNMS(KP587785252, T40, KP951056516 * T3Z);
Chris@42 904 T4a = FMA(KP951056516, T40, KP587785252 * T3Z);
Chris@42 905 ii[WS(rs, 10)] = T45 + T42;
Chris@42 906 T49 = T47 + T46;
Chris@42 907 ii[WS(rs, 6)] = T49 - T4a;
Chris@42 908 ii[WS(rs, 14)] = T4a + T49;
Chris@42 909 T48 = T46 - T47;
Chris@42 910 ii[WS(rs, 2)] = T41 + T48;
Chris@42 911 ii[WS(rs, 18)] = T48 - T41;
Chris@42 912 }
Chris@42 913 {
Chris@42 914 E T3x, T1Y, T3y, T3G, T3I, T3C, T3F, T3H, T3z;
Chris@42 915 T3x = KP559016994 * (T1U - T1X);
Chris@42 916 T1Y = T1U + T1X;
Chris@42 917 T3y = FNMS(KP250000000, T1Y, T1R);
Chris@42 918 T3C = T3A - T3B;
Chris@42 919 T3F = T3D - T3E;
Chris@42 920 T3G = FMA(KP951056516, T3C, KP587785252 * T3F);
Chris@42 921 T3I = FNMS(KP587785252, T3C, KP951056516 * T3F);
Chris@42 922 ri[0] = T1R + T1Y;
Chris@42 923 T3H = T3y - T3x;
Chris@42 924 ri[WS(rs, 12)] = T3H - T3I;
Chris@42 925 ri[WS(rs, 8)] = T3H + T3I;
Chris@42 926 T3z = T3x + T3y;
Chris@42 927 ri[WS(rs, 4)] = T3z - T3G;
Chris@42 928 ri[WS(rs, 16)] = T3z + T3G;
Chris@42 929 }
Chris@42 930 {
Chris@42 931 E T3U, T3L, T3V, T3T, T3Y, T3R, T3S, T3X, T3W;
Chris@42 932 T3U = KP559016994 * (T3J - T3K);
Chris@42 933 T3L = T3J + T3K;
Chris@42 934 T3V = FNMS(KP250000000, T3L, T3Q);
Chris@42 935 T3R = T1S - T1T;
Chris@42 936 T3S = T1V - T1W;
Chris@42 937 T3T = FMA(KP951056516, T3R, KP587785252 * T3S);
Chris@42 938 T3Y = FNMS(KP587785252, T3R, KP951056516 * T3S);
Chris@42 939 ii[0] = T3L + T3Q;
Chris@42 940 T3X = T3V - T3U;
Chris@42 941 ii[WS(rs, 8)] = T3X - T3Y;
Chris@42 942 ii[WS(rs, 12)] = T3Y + T3X;
Chris@42 943 T3W = T3U + T3V;
Chris@42 944 ii[WS(rs, 4)] = T3T + T3W;
Chris@42 945 ii[WS(rs, 16)] = T3W - T3T;
Chris@42 946 }
Chris@42 947 {
Chris@42 948 E T2P, T2N, T2O, T2l, T2R, T29, T2k, T2S, T2Q;
Chris@42 949 T2P = KP559016994 * (T2B - T2M);
Chris@42 950 T2N = T2B + T2M;
Chris@42 951 T2O = FNMS(KP250000000, T2N, T2q);
Chris@42 952 T29 = T23 - T28;
Chris@42 953 T2k = T2e - T2j;
Chris@42 954 T2l = FNMS(KP587785252, T2k, KP951056516 * T29);
Chris@42 955 T2R = FMA(KP951056516, T2k, KP587785252 * T29);
Chris@42 956 ri[WS(rs, 15)] = T2q + T2N;
Chris@42 957 T2S = T2P + T2O;
Chris@42 958 ri[WS(rs, 11)] = T2R + T2S;
Chris@42 959 ri[WS(rs, 19)] = T2S - T2R;
Chris@42 960 T2Q = T2O - T2P;
Chris@42 961 ri[WS(rs, 3)] = T2l + T2Q;
Chris@42 962 ri[WS(rs, 7)] = T2Q - T2l;
Chris@42 963 }
Chris@42 964 {
Chris@42 965 E T4u, T4s, T4t, T4y, T4A, T4w, T4x, T4z, T4v;
Chris@42 966 T4u = KP559016994 * (T4q - T4r);
Chris@42 967 T4s = T4q + T4r;
Chris@42 968 T4t = FNMS(KP250000000, T4s, T4p);
Chris@42 969 T4w = T2G - T2L;
Chris@42 970 T4x = T2v - T2A;
Chris@42 971 T4y = FNMS(KP587785252, T4x, KP951056516 * T4w);
Chris@42 972 T4A = FMA(KP951056516, T4x, KP587785252 * T4w);
Chris@42 973 ii[WS(rs, 15)] = T4s + T4p;
Chris@42 974 T4z = T4u + T4t;
Chris@42 975 ii[WS(rs, 11)] = T4z - T4A;
Chris@42 976 ii[WS(rs, 19)] = T4A + T4z;
Chris@42 977 T4v = T4t - T4u;
Chris@42 978 ii[WS(rs, 3)] = T4v - T4y;
Chris@42 979 ii[WS(rs, 7)] = T4y + T4v;
Chris@42 980 }
Chris@42 981 {
Chris@42 982 E T36, T38, T39, T2Z, T3b, T2V, T2Y, T3c, T3a;
Chris@42 983 T36 = KP559016994 * (T32 - T35);
Chris@42 984 T38 = T32 + T35;
Chris@42 985 T39 = FNMS(KP250000000, T38, T37);
Chris@42 986 T2V = T2T - T2U;
Chris@42 987 T2Y = T2W - T2X;
Chris@42 988 T2Z = FMA(KP951056516, T2V, KP587785252 * T2Y);
Chris@42 989 T3b = FNMS(KP587785252, T2V, KP951056516 * T2Y);
Chris@42 990 ri[WS(rs, 5)] = T37 + T38;
Chris@42 991 T3c = T39 - T36;
Chris@42 992 ri[WS(rs, 13)] = T3b + T3c;
Chris@42 993 ri[WS(rs, 17)] = T3c - T3b;
Chris@42 994 T3a = T36 + T39;
Chris@42 995 ri[WS(rs, 1)] = T2Z + T3a;
Chris@42 996 ri[WS(rs, 9)] = T3a - T2Z;
Chris@42 997 }
Chris@42 998 {
Chris@42 999 E T4d, T4h, T4i, T4m, T4o, T4k, T4l, T4n, T4j;
Chris@42 1000 T4d = KP559016994 * (T4b - T4c);
Chris@42 1001 T4h = T4b + T4c;
Chris@42 1002 T4i = FNMS(KP250000000, T4h, T4g);
Chris@42 1003 T4k = T30 - T31;
Chris@42 1004 T4l = T33 - T34;
Chris@42 1005 T4m = FMA(KP951056516, T4k, KP587785252 * T4l);
Chris@42 1006 T4o = FNMS(KP587785252, T4k, KP951056516 * T4l);
Chris@42 1007 ii[WS(rs, 5)] = T4h + T4g;
Chris@42 1008 T4n = T4i - T4d;
Chris@42 1009 ii[WS(rs, 13)] = T4n - T4o;
Chris@42 1010 ii[WS(rs, 17)] = T4o + T4n;
Chris@42 1011 T4j = T4d + T4i;
Chris@42 1012 ii[WS(rs, 1)] = T4j - T4m;
Chris@42 1013 ii[WS(rs, 9)] = T4m + T4j;
Chris@42 1014 }
Chris@42 1015 }
Chris@42 1016 }
Chris@42 1017 }
Chris@42 1018
Chris@42 1019 static const tw_instr twinstr[] = {
Chris@42 1020 {TW_FULL, 0, 20},
Chris@42 1021 {TW_NEXT, 1, 0}
Chris@42 1022 };
Chris@42 1023
Chris@42 1024 static const ct_desc desc = { 20, "t1_20", twinstr, &GENUS, {184, 62, 62, 0}, 0, 0, 0 };
Chris@42 1025
Chris@42 1026 void X(codelet_t1_20) (planner *p) {
Chris@42 1027 X(kdft_dit_register) (p, t1_20, &desc);
Chris@42 1028 }
Chris@42 1029 #endif /* HAVE_FMA */