Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.8/dft/scalar/codelets/t2_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 d0c2a83c1364
children
rev   line source
Chris@82 1 /*
Chris@82 2 * Copyright (c) 2003, 2007-14 Matteo Frigo
Chris@82 3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
Chris@82 4 *
Chris@82 5 * This program is free software; you can redistribute it and/or modify
Chris@82 6 * it under the terms of the GNU General Public License as published by
Chris@82 7 * the Free Software Foundation; either version 2 of the License, or
Chris@82 8 * (at your option) any later version.
Chris@82 9 *
Chris@82 10 * This program is distributed in the hope that it will be useful,
Chris@82 11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
Chris@82 12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
Chris@82 13 * GNU General Public License for more details.
Chris@82 14 *
Chris@82 15 * You should have received a copy of the GNU General Public License
Chris@82 16 * along with this program; if not, write to the Free Software
Chris@82 17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
Chris@82 18 *
Chris@82 19 */
Chris@82 20
Chris@82 21 /* This file was automatically generated --- DO NOT EDIT */
Chris@82 22 /* Generated on Thu May 24 08:04:26 EDT 2018 */
Chris@82 23
Chris@82 24 #include "dft/codelet-dft.h"
Chris@82 25
Chris@82 26 #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA)
Chris@82 27
Chris@82 28 /* Generated by: ../../../genfft/gen_twiddle.native -fma -compact -variables 4 -pipeline-latency 4 -twiddle-log3 -precompute-twiddles -n 20 -name t2_20 -include dft/scalar/t.h */
Chris@82 29
Chris@82 30 /*
Chris@82 31 * This function contains 276 FP additions, 198 FP multiplications,
Chris@82 32 * (or, 136 additions, 58 multiplications, 140 fused multiply/add),
Chris@82 33 * 95 stack variables, 4 constants, and 80 memory accesses
Chris@82 34 */
Chris@82 35 #include "dft/scalar/t.h"
Chris@82 36
Chris@82 37 static void t2_20(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
Chris@82 38 {
Chris@82 39 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
Chris@82 40 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
Chris@82 41 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
Chris@82 42 DK(KP618033988, +0.618033988749894848204586834365638117720309180);
Chris@82 43 {
Chris@82 44 INT m;
Chris@82 45 for (m = mb, W = W + (mb * 8); m < me; m = m + 1, ri = ri + ms, ii = ii + ms, W = W + 8, MAKE_VOLATILE_STRIDE(40, rs)) {
Chris@82 46 E T2, Th, Tf, T6, T5, Ti, Tl, T1n, T3, Tt, Tv, T7, T17, T1L, T24;
Chris@82 47 E Tb, T13, T1P, T21, T1b, T1D, T1A, T1H, T1f, TA, Tw, Tq, Tm, TK, T1S;
Chris@82 48 E TO, T1p, T1q, T1u, T2n, T2k, T2h, T2d;
Chris@82 49 {
Chris@82 50 E Tk, Ta, T1e, T4, T1a, Tj, T12, T1G, T16, T1K, Tg, Tz;
Chris@82 51 T2 = W[0];
Chris@82 52 Th = W[3];
Chris@82 53 Tf = W[2];
Chris@82 54 Tg = T2 * Tf;
Chris@82 55 Tk = T2 * Th;
Chris@82 56 T6 = W[5];
Chris@82 57 Ta = T2 * T6;
Chris@82 58 T1e = Tf * T6;
Chris@82 59 T5 = W[1];
Chris@82 60 Ti = FNMS(T5, Th, Tg);
Chris@82 61 Tl = FMA(T5, Tf, Tk);
Chris@82 62 T1n = FMA(T5, Th, Tg);
Chris@82 63 T3 = W[4];
Chris@82 64 T4 = T2 * T3;
Chris@82 65 T1a = Tf * T3;
Chris@82 66 Tj = Ti * T3;
Chris@82 67 Tt = W[6];
Chris@82 68 T12 = Tf * Tt;
Chris@82 69 T1G = T2 * Tt;
Chris@82 70 Tv = W[7];
Chris@82 71 T16 = Tf * Tv;
Chris@82 72 T1K = T2 * Tv;
Chris@82 73 T7 = FNMS(T5, T6, T4);
Chris@82 74 T17 = FNMS(Th, Tt, T16);
Chris@82 75 T1L = FNMS(T5, Tt, T1K);
Chris@82 76 T24 = FMA(Th, T3, T1e);
Chris@82 77 Tb = FMA(T5, T3, Ta);
Chris@82 78 T13 = FMA(Th, Tv, T12);
Chris@82 79 T1P = FNMS(Tl, T6, Tj);
Chris@82 80 T21 = FNMS(Th, T6, T1a);
Chris@82 81 T1b = FMA(Th, T6, T1a);
Chris@82 82 T1D = FNMS(T5, T3, Ta);
Chris@82 83 T1A = FMA(T5, T6, T4);
Chris@82 84 T1H = FMA(T5, Tv, T1G);
Chris@82 85 T1f = FNMS(Th, T3, T1e);
Chris@82 86 Tz = Ti * Tv;
Chris@82 87 TA = FNMS(Tl, Tt, Tz);
Chris@82 88 {
Chris@82 89 E Tu, Tp, TJ, TN;
Chris@82 90 Tu = Ti * Tt;
Chris@82 91 Tw = FMA(Tl, Tv, Tu);
Chris@82 92 Tp = Ti * T6;
Chris@82 93 Tq = FNMS(Tl, T3, Tp);
Chris@82 94 Tm = FMA(Tl, T6, Tj);
Chris@82 95 TJ = Tm * Tt;
Chris@82 96 TN = Tm * Tv;
Chris@82 97 TK = FMA(Tq, Tv, TJ);
Chris@82 98 T1S = FMA(Tl, T3, Tp);
Chris@82 99 TO = FNMS(Tq, Tt, TN);
Chris@82 100 {
Chris@82 101 E T1o, T2g, T1t, T2c;
Chris@82 102 T1o = T1n * T3;
Chris@82 103 T2g = T1n * Tv;
Chris@82 104 T1t = T1n * T6;
Chris@82 105 T2c = T1n * Tt;
Chris@82 106 T1p = FNMS(T5, Tf, Tk);
Chris@82 107 T1q = FNMS(T1p, T6, T1o);
Chris@82 108 T1u = FMA(T1p, T3, T1t);
Chris@82 109 T2n = FNMS(T1p, T3, T1t);
Chris@82 110 T2k = FMA(T1p, T6, T1o);
Chris@82 111 T2h = FNMS(T1p, Tt, T2g);
Chris@82 112 T2d = FMA(T1p, Tv, T2c);
Chris@82 113 }
Chris@82 114 }
Chris@82 115 }
Chris@82 116 {
Chris@82 117 E Te, T2C, T4L, T57, TD, T58, T2H, T4H, T11, T2v, T4k, T4v, T2P, T3P, T3C;
Chris@82 118 E T3Z, T2r, T2z, T4g, T4z, T3b, T3T, T3u, T43, T20, T2y, T4d, T4y, T34, T3S;
Chris@82 119 E T3n, T42, T1y, T2w, T4n, T4w, T2W, T3Q, T3J, T40;
Chris@82 120 {
Chris@82 121 E T1, T4K, T8, T9, Tc, T4I, Td, T4J;
Chris@82 122 T1 = ri[0];
Chris@82 123 T4K = ii[0];
Chris@82 124 T8 = ri[WS(rs, 10)];
Chris@82 125 T9 = T7 * T8;
Chris@82 126 Tc = ii[WS(rs, 10)];
Chris@82 127 T4I = T7 * Tc;
Chris@82 128 Td = FMA(Tb, Tc, T9);
Chris@82 129 Te = T1 + Td;
Chris@82 130 T2C = T1 - Td;
Chris@82 131 T4J = FNMS(Tb, T8, T4I);
Chris@82 132 T4L = T4J + T4K;
Chris@82 133 T57 = T4K - T4J;
Chris@82 134 }
Chris@82 135 {
Chris@82 136 E Tn, To, Tr, T2D, Tx, Ty, TB, T2F;
Chris@82 137 Tn = ri[WS(rs, 5)];
Chris@82 138 To = Tm * Tn;
Chris@82 139 Tr = ii[WS(rs, 5)];
Chris@82 140 T2D = Tm * Tr;
Chris@82 141 Tx = ri[WS(rs, 15)];
Chris@82 142 Ty = Tw * Tx;
Chris@82 143 TB = ii[WS(rs, 15)];
Chris@82 144 T2F = Tw * TB;
Chris@82 145 {
Chris@82 146 E Ts, TC, T2E, T2G;
Chris@82 147 Ts = FMA(Tq, Tr, To);
Chris@82 148 TC = FMA(TA, TB, Ty);
Chris@82 149 TD = Ts + TC;
Chris@82 150 T58 = Ts - TC;
Chris@82 151 T2E = FNMS(Tq, Tn, T2D);
Chris@82 152 T2G = FNMS(TA, Tx, T2F);
Chris@82 153 T2H = T2E - T2G;
Chris@82 154 T4H = T2E + T2G;
Chris@82 155 }
Chris@82 156 }
Chris@82 157 {
Chris@82 158 E TI, T3x, TZ, T2N, TQ, T3z, TV, T2L;
Chris@82 159 {
Chris@82 160 E TF, TG, TH, T3w;
Chris@82 161 TF = ri[WS(rs, 4)];
Chris@82 162 TG = Ti * TF;
Chris@82 163 TH = ii[WS(rs, 4)];
Chris@82 164 T3w = Ti * TH;
Chris@82 165 TI = FMA(Tl, TH, TG);
Chris@82 166 T3x = FNMS(Tl, TF, T3w);
Chris@82 167 }
Chris@82 168 {
Chris@82 169 E TW, TX, TY, T2M;
Chris@82 170 TW = ri[WS(rs, 19)];
Chris@82 171 TX = Tt * TW;
Chris@82 172 TY = ii[WS(rs, 19)];
Chris@82 173 T2M = Tt * TY;
Chris@82 174 TZ = FMA(Tv, TY, TX);
Chris@82 175 T2N = FNMS(Tv, TW, T2M);
Chris@82 176 }
Chris@82 177 {
Chris@82 178 E TL, TM, TP, T3y;
Chris@82 179 TL = ri[WS(rs, 14)];
Chris@82 180 TM = TK * TL;
Chris@82 181 TP = ii[WS(rs, 14)];
Chris@82 182 T3y = TK * TP;
Chris@82 183 TQ = FMA(TO, TP, TM);
Chris@82 184 T3z = FNMS(TO, TL, T3y);
Chris@82 185 }
Chris@82 186 {
Chris@82 187 E TS, TT, TU, T2K;
Chris@82 188 TS = ri[WS(rs, 9)];
Chris@82 189 TT = T3 * TS;
Chris@82 190 TU = ii[WS(rs, 9)];
Chris@82 191 T2K = T3 * TU;
Chris@82 192 TV = FMA(T6, TU, TT);
Chris@82 193 T2L = FNMS(T6, TS, T2K);
Chris@82 194 }
Chris@82 195 {
Chris@82 196 E TR, T10, T4i, T4j;
Chris@82 197 TR = TI + TQ;
Chris@82 198 T10 = TV + TZ;
Chris@82 199 T11 = TR - T10;
Chris@82 200 T2v = TR + T10;
Chris@82 201 T4i = T3x + T3z;
Chris@82 202 T4j = T2L + T2N;
Chris@82 203 T4k = T4i - T4j;
Chris@82 204 T4v = T4i + T4j;
Chris@82 205 }
Chris@82 206 {
Chris@82 207 E T2J, T2O, T3A, T3B;
Chris@82 208 T2J = TI - TQ;
Chris@82 209 T2O = T2L - T2N;
Chris@82 210 T2P = T2J - T2O;
Chris@82 211 T3P = T2J + T2O;
Chris@82 212 T3A = T3x - T3z;
Chris@82 213 T3B = TV - TZ;
Chris@82 214 T3C = T3A + T3B;
Chris@82 215 T3Z = T3A - T3B;
Chris@82 216 }
Chris@82 217 }
Chris@82 218 {
Chris@82 219 E T26, T3p, T2p, T39, T2a, T3r, T2j, T37;
Chris@82 220 {
Chris@82 221 E T22, T23, T25, T3o;
Chris@82 222 T22 = ri[WS(rs, 12)];
Chris@82 223 T23 = T21 * T22;
Chris@82 224 T25 = ii[WS(rs, 12)];
Chris@82 225 T3o = T21 * T25;
Chris@82 226 T26 = FMA(T24, T25, T23);
Chris@82 227 T3p = FNMS(T24, T22, T3o);
Chris@82 228 }
Chris@82 229 {
Chris@82 230 E T2l, T2m, T2o, T38;
Chris@82 231 T2l = ri[WS(rs, 7)];
Chris@82 232 T2m = T2k * T2l;
Chris@82 233 T2o = ii[WS(rs, 7)];
Chris@82 234 T38 = T2k * T2o;
Chris@82 235 T2p = FMA(T2n, T2o, T2m);
Chris@82 236 T39 = FNMS(T2n, T2l, T38);
Chris@82 237 }
Chris@82 238 {
Chris@82 239 E T27, T28, T29, T3q;
Chris@82 240 T27 = ri[WS(rs, 2)];
Chris@82 241 T28 = T1n * T27;
Chris@82 242 T29 = ii[WS(rs, 2)];
Chris@82 243 T3q = T1n * T29;
Chris@82 244 T2a = FMA(T1p, T29, T28);
Chris@82 245 T3r = FNMS(T1p, T27, T3q);
Chris@82 246 }
Chris@82 247 {
Chris@82 248 E T2e, T2f, T2i, T36;
Chris@82 249 T2e = ri[WS(rs, 17)];
Chris@82 250 T2f = T2d * T2e;
Chris@82 251 T2i = ii[WS(rs, 17)];
Chris@82 252 T36 = T2d * T2i;
Chris@82 253 T2j = FMA(T2h, T2i, T2f);
Chris@82 254 T37 = FNMS(T2h, T2e, T36);
Chris@82 255 }
Chris@82 256 {
Chris@82 257 E T2b, T2q, T4e, T4f;
Chris@82 258 T2b = T26 + T2a;
Chris@82 259 T2q = T2j + T2p;
Chris@82 260 T2r = T2b - T2q;
Chris@82 261 T2z = T2b + T2q;
Chris@82 262 T4e = T3p + T3r;
Chris@82 263 T4f = T37 + T39;
Chris@82 264 T4g = T4e - T4f;
Chris@82 265 T4z = T4e + T4f;
Chris@82 266 }
Chris@82 267 {
Chris@82 268 E T35, T3a, T3s, T3t;
Chris@82 269 T35 = T26 - T2a;
Chris@82 270 T3a = T37 - T39;
Chris@82 271 T3b = T35 - T3a;
Chris@82 272 T3T = T35 + T3a;
Chris@82 273 T3s = T3p - T3r;
Chris@82 274 T3t = T2j - T2p;
Chris@82 275 T3u = T3s + T3t;
Chris@82 276 T43 = T3s - T3t;
Chris@82 277 }
Chris@82 278 }
Chris@82 279 {
Chris@82 280 E T1F, T3i, T1Y, T32, T1N, T3k, T1U, T30;
Chris@82 281 {
Chris@82 282 E T1B, T1C, T1E, T3h;
Chris@82 283 T1B = ri[WS(rs, 8)];
Chris@82 284 T1C = T1A * T1B;
Chris@82 285 T1E = ii[WS(rs, 8)];
Chris@82 286 T3h = T1A * T1E;
Chris@82 287 T1F = FMA(T1D, T1E, T1C);
Chris@82 288 T3i = FNMS(T1D, T1B, T3h);
Chris@82 289 }
Chris@82 290 {
Chris@82 291 E T1V, T1W, T1X, T31;
Chris@82 292 T1V = ri[WS(rs, 3)];
Chris@82 293 T1W = Tf * T1V;
Chris@82 294 T1X = ii[WS(rs, 3)];
Chris@82 295 T31 = Tf * T1X;
Chris@82 296 T1Y = FMA(Th, T1X, T1W);
Chris@82 297 T32 = FNMS(Th, T1V, T31);
Chris@82 298 }
Chris@82 299 {
Chris@82 300 E T1I, T1J, T1M, T3j;
Chris@82 301 T1I = ri[WS(rs, 18)];
Chris@82 302 T1J = T1H * T1I;
Chris@82 303 T1M = ii[WS(rs, 18)];
Chris@82 304 T3j = T1H * T1M;
Chris@82 305 T1N = FMA(T1L, T1M, T1J);
Chris@82 306 T3k = FNMS(T1L, T1I, T3j);
Chris@82 307 }
Chris@82 308 {
Chris@82 309 E T1Q, T1R, T1T, T2Z;
Chris@82 310 T1Q = ri[WS(rs, 13)];
Chris@82 311 T1R = T1P * T1Q;
Chris@82 312 T1T = ii[WS(rs, 13)];
Chris@82 313 T2Z = T1P * T1T;
Chris@82 314 T1U = FMA(T1S, T1T, T1R);
Chris@82 315 T30 = FNMS(T1S, T1Q, T2Z);
Chris@82 316 }
Chris@82 317 {
Chris@82 318 E T1O, T1Z, T4b, T4c;
Chris@82 319 T1O = T1F + T1N;
Chris@82 320 T1Z = T1U + T1Y;
Chris@82 321 T20 = T1O - T1Z;
Chris@82 322 T2y = T1O + T1Z;
Chris@82 323 T4b = T3i + T3k;
Chris@82 324 T4c = T30 + T32;
Chris@82 325 T4d = T4b - T4c;
Chris@82 326 T4y = T4b + T4c;
Chris@82 327 }
Chris@82 328 {
Chris@82 329 E T2Y, T33, T3l, T3m;
Chris@82 330 T2Y = T1F - T1N;
Chris@82 331 T33 = T30 - T32;
Chris@82 332 T34 = T2Y - T33;
Chris@82 333 T3S = T2Y + T33;
Chris@82 334 T3l = T3i - T3k;
Chris@82 335 T3m = T1U - T1Y;
Chris@82 336 T3n = T3l + T3m;
Chris@82 337 T42 = T3l - T3m;
Chris@82 338 }
Chris@82 339 }
Chris@82 340 {
Chris@82 341 E T19, T3E, T1w, T2U, T1h, T3G, T1m, T2S;
Chris@82 342 {
Chris@82 343 E T14, T15, T18, T3D;
Chris@82 344 T14 = ri[WS(rs, 16)];
Chris@82 345 T15 = T13 * T14;
Chris@82 346 T18 = ii[WS(rs, 16)];
Chris@82 347 T3D = T13 * T18;
Chris@82 348 T19 = FMA(T17, T18, T15);
Chris@82 349 T3E = FNMS(T17, T14, T3D);
Chris@82 350 }
Chris@82 351 {
Chris@82 352 E T1r, T1s, T1v, T2T;
Chris@82 353 T1r = ri[WS(rs, 11)];
Chris@82 354 T1s = T1q * T1r;
Chris@82 355 T1v = ii[WS(rs, 11)];
Chris@82 356 T2T = T1q * T1v;
Chris@82 357 T1w = FMA(T1u, T1v, T1s);
Chris@82 358 T2U = FNMS(T1u, T1r, T2T);
Chris@82 359 }
Chris@82 360 {
Chris@82 361 E T1c, T1d, T1g, T3F;
Chris@82 362 T1c = ri[WS(rs, 6)];
Chris@82 363 T1d = T1b * T1c;
Chris@82 364 T1g = ii[WS(rs, 6)];
Chris@82 365 T3F = T1b * T1g;
Chris@82 366 T1h = FMA(T1f, T1g, T1d);
Chris@82 367 T3G = FNMS(T1f, T1c, T3F);
Chris@82 368 }
Chris@82 369 {
Chris@82 370 E T1j, T1k, T1l, T2R;
Chris@82 371 T1j = ri[WS(rs, 1)];
Chris@82 372 T1k = T2 * T1j;
Chris@82 373 T1l = ii[WS(rs, 1)];
Chris@82 374 T2R = T2 * T1l;
Chris@82 375 T1m = FMA(T5, T1l, T1k);
Chris@82 376 T2S = FNMS(T5, T1j, T2R);
Chris@82 377 }
Chris@82 378 {
Chris@82 379 E T1i, T1x, T4l, T4m;
Chris@82 380 T1i = T19 + T1h;
Chris@82 381 T1x = T1m + T1w;
Chris@82 382 T1y = T1i - T1x;
Chris@82 383 T2w = T1i + T1x;
Chris@82 384 T4l = T3E + T3G;
Chris@82 385 T4m = T2S + T2U;
Chris@82 386 T4n = T4l - T4m;
Chris@82 387 T4w = T4l + T4m;
Chris@82 388 }
Chris@82 389 {
Chris@82 390 E T2Q, T2V, T3H, T3I;
Chris@82 391 T2Q = T19 - T1h;
Chris@82 392 T2V = T2S - T2U;
Chris@82 393 T2W = T2Q - T2V;
Chris@82 394 T3Q = T2Q + T2V;
Chris@82 395 T3H = T3E - T3G;
Chris@82 396 T3I = T1m - T1w;
Chris@82 397 T3J = T3H + T3I;
Chris@82 398 T40 = T3H - T3I;
Chris@82 399 }
Chris@82 400 }
Chris@82 401 {
Chris@82 402 E T4p, T4r, TE, T2t, T48, T49, T4q, T4a;
Chris@82 403 {
Chris@82 404 E T4h, T4o, T1z, T2s;
Chris@82 405 T4h = T4d - T4g;
Chris@82 406 T4o = T4k - T4n;
Chris@82 407 T4p = FNMS(KP618033988, T4o, T4h);
Chris@82 408 T4r = FMA(KP618033988, T4h, T4o);
Chris@82 409 TE = Te - TD;
Chris@82 410 T1z = T11 + T1y;
Chris@82 411 T2s = T20 + T2r;
Chris@82 412 T2t = T1z + T2s;
Chris@82 413 T48 = FNMS(KP250000000, T2t, TE);
Chris@82 414 T49 = T1z - T2s;
Chris@82 415 }
Chris@82 416 ri[WS(rs, 10)] = TE + T2t;
Chris@82 417 T4q = FMA(KP559016994, T49, T48);
Chris@82 418 ri[WS(rs, 14)] = FNMS(KP951056516, T4r, T4q);
Chris@82 419 ri[WS(rs, 6)] = FMA(KP951056516, T4r, T4q);
Chris@82 420 T4a = FNMS(KP559016994, T49, T48);
Chris@82 421 ri[WS(rs, 2)] = FNMS(KP951056516, T4p, T4a);
Chris@82 422 ri[WS(rs, 18)] = FMA(KP951056516, T4p, T4a);
Chris@82 423 }
Chris@82 424 {
Chris@82 425 E T54, T56, T4V, T4Y, T4Z, T50, T55, T51;
Chris@82 426 {
Chris@82 427 E T52, T53, T4W, T4X;
Chris@82 428 T52 = T20 - T2r;
Chris@82 429 T53 = T11 - T1y;
Chris@82 430 T54 = FNMS(KP618033988, T53, T52);
Chris@82 431 T56 = FMA(KP618033988, T52, T53);
Chris@82 432 T4V = T4L - T4H;
Chris@82 433 T4W = T4k + T4n;
Chris@82 434 T4X = T4d + T4g;
Chris@82 435 T4Y = T4W + T4X;
Chris@82 436 T4Z = FNMS(KP250000000, T4Y, T4V);
Chris@82 437 T50 = T4W - T4X;
Chris@82 438 }
Chris@82 439 ii[WS(rs, 10)] = T4Y + T4V;
Chris@82 440 T55 = FMA(KP559016994, T50, T4Z);
Chris@82 441 ii[WS(rs, 6)] = FNMS(KP951056516, T56, T55);
Chris@82 442 ii[WS(rs, 14)] = FMA(KP951056516, T56, T55);
Chris@82 443 T51 = FNMS(KP559016994, T50, T4Z);
Chris@82 444 ii[WS(rs, 2)] = FMA(KP951056516, T54, T51);
Chris@82 445 ii[WS(rs, 18)] = FNMS(KP951056516, T54, T51);
Chris@82 446 }
Chris@82 447 {
Chris@82 448 E T4B, T4D, T2u, T2B, T4s, T4t, T4C, T4u;
Chris@82 449 {
Chris@82 450 E T4x, T4A, T2x, T2A;
Chris@82 451 T4x = T4v - T4w;
Chris@82 452 T4A = T4y - T4z;
Chris@82 453 T4B = FMA(KP618033988, T4A, T4x);
Chris@82 454 T4D = FNMS(KP618033988, T4x, T4A);
Chris@82 455 T2u = Te + TD;
Chris@82 456 T2x = T2v + T2w;
Chris@82 457 T2A = T2y + T2z;
Chris@82 458 T2B = T2x + T2A;
Chris@82 459 T4s = FNMS(KP250000000, T2B, T2u);
Chris@82 460 T4t = T2x - T2A;
Chris@82 461 }
Chris@82 462 ri[0] = T2u + T2B;
Chris@82 463 T4C = FNMS(KP559016994, T4t, T4s);
Chris@82 464 ri[WS(rs, 12)] = FNMS(KP951056516, T4D, T4C);
Chris@82 465 ri[WS(rs, 8)] = FMA(KP951056516, T4D, T4C);
Chris@82 466 T4u = FMA(KP559016994, T4t, T4s);
Chris@82 467 ri[WS(rs, 4)] = FNMS(KP951056516, T4B, T4u);
Chris@82 468 ri[WS(rs, 16)] = FMA(KP951056516, T4B, T4u);
Chris@82 469 }
Chris@82 470 {
Chris@82 471 E T4S, T4U, T4M, T4G, T4N, T4O, T4T, T4P;
Chris@82 472 {
Chris@82 473 E T4Q, T4R, T4E, T4F;
Chris@82 474 T4Q = T2v - T2w;
Chris@82 475 T4R = T2y - T2z;
Chris@82 476 T4S = FMA(KP618033988, T4R, T4Q);
Chris@82 477 T4U = FNMS(KP618033988, T4Q, T4R);
Chris@82 478 T4M = T4H + T4L;
Chris@82 479 T4E = T4v + T4w;
Chris@82 480 T4F = T4y + T4z;
Chris@82 481 T4G = T4E + T4F;
Chris@82 482 T4N = FNMS(KP250000000, T4G, T4M);
Chris@82 483 T4O = T4E - T4F;
Chris@82 484 }
Chris@82 485 ii[0] = T4G + T4M;
Chris@82 486 T4T = FNMS(KP559016994, T4O, T4N);
Chris@82 487 ii[WS(rs, 8)] = FNMS(KP951056516, T4U, T4T);
Chris@82 488 ii[WS(rs, 12)] = FMA(KP951056516, T4U, T4T);
Chris@82 489 T4P = FMA(KP559016994, T4O, T4N);
Chris@82 490 ii[WS(rs, 4)] = FMA(KP951056516, T4S, T4P);
Chris@82 491 ii[WS(rs, 16)] = FNMS(KP951056516, T4S, T4P);
Chris@82 492 }
Chris@82 493 {
Chris@82 494 E T3L, T3N, T2I, T3d, T3e, T3f, T3M, T3g;
Chris@82 495 {
Chris@82 496 E T3v, T3K, T2X, T3c;
Chris@82 497 T3v = T3n - T3u;
Chris@82 498 T3K = T3C - T3J;
Chris@82 499 T3L = FNMS(KP618033988, T3K, T3v);
Chris@82 500 T3N = FMA(KP618033988, T3v, T3K);
Chris@82 501 T2I = T2C - T2H;
Chris@82 502 T2X = T2P + T2W;
Chris@82 503 T3c = T34 + T3b;
Chris@82 504 T3d = T2X + T3c;
Chris@82 505 T3e = FNMS(KP250000000, T3d, T2I);
Chris@82 506 T3f = T2X - T3c;
Chris@82 507 }
Chris@82 508 ri[WS(rs, 15)] = T2I + T3d;
Chris@82 509 T3M = FMA(KP559016994, T3f, T3e);
Chris@82 510 ri[WS(rs, 11)] = FMA(KP951056516, T3N, T3M);
Chris@82 511 ri[WS(rs, 19)] = FNMS(KP951056516, T3N, T3M);
Chris@82 512 T3g = FNMS(KP559016994, T3f, T3e);
Chris@82 513 ri[WS(rs, 3)] = FMA(KP951056516, T3L, T3g);
Chris@82 514 ri[WS(rs, 7)] = FNMS(KP951056516, T3L, T3g);
Chris@82 515 }
Chris@82 516 {
Chris@82 517 E T5u, T5w, T5l, T5o, T5p, T5q, T5v, T5r;
Chris@82 518 {
Chris@82 519 E T5s, T5t, T5m, T5n;
Chris@82 520 T5s = T34 - T3b;
Chris@82 521 T5t = T2P - T2W;
Chris@82 522 T5u = FNMS(KP618033988, T5t, T5s);
Chris@82 523 T5w = FMA(KP618033988, T5s, T5t);
Chris@82 524 T5l = T58 + T57;
Chris@82 525 T5m = T3C + T3J;
Chris@82 526 T5n = T3n + T3u;
Chris@82 527 T5o = T5m + T5n;
Chris@82 528 T5p = FNMS(KP250000000, T5o, T5l);
Chris@82 529 T5q = T5m - T5n;
Chris@82 530 }
Chris@82 531 ii[WS(rs, 15)] = T5o + T5l;
Chris@82 532 T5v = FMA(KP559016994, T5q, T5p);
Chris@82 533 ii[WS(rs, 11)] = FNMS(KP951056516, T5w, T5v);
Chris@82 534 ii[WS(rs, 19)] = FMA(KP951056516, T5w, T5v);
Chris@82 535 T5r = FNMS(KP559016994, T5q, T5p);
Chris@82 536 ii[WS(rs, 3)] = FNMS(KP951056516, T5u, T5r);
Chris@82 537 ii[WS(rs, 7)] = FMA(KP951056516, T5u, T5r);
Chris@82 538 }
Chris@82 539 {
Chris@82 540 E T45, T47, T3O, T3V, T3W, T3X, T46, T3Y;
Chris@82 541 {
Chris@82 542 E T41, T44, T3R, T3U;
Chris@82 543 T41 = T3Z - T40;
Chris@82 544 T44 = T42 - T43;
Chris@82 545 T45 = FMA(KP618033988, T44, T41);
Chris@82 546 T47 = FNMS(KP618033988, T41, T44);
Chris@82 547 T3O = T2C + T2H;
Chris@82 548 T3R = T3P + T3Q;
Chris@82 549 T3U = T3S + T3T;
Chris@82 550 T3V = T3R + T3U;
Chris@82 551 T3W = FNMS(KP250000000, T3V, T3O);
Chris@82 552 T3X = T3R - T3U;
Chris@82 553 }
Chris@82 554 ri[WS(rs, 5)] = T3O + T3V;
Chris@82 555 T46 = FNMS(KP559016994, T3X, T3W);
Chris@82 556 ri[WS(rs, 13)] = FMA(KP951056516, T47, T46);
Chris@82 557 ri[WS(rs, 17)] = FNMS(KP951056516, T47, T46);
Chris@82 558 T3Y = FMA(KP559016994, T3X, T3W);
Chris@82 559 ri[WS(rs, 1)] = FMA(KP951056516, T45, T3Y);
Chris@82 560 ri[WS(rs, 9)] = FNMS(KP951056516, T45, T3Y);
Chris@82 561 }
Chris@82 562 {
Chris@82 563 E T5i, T5k, T59, T5c, T5d, T5e, T5j, T5f;
Chris@82 564 {
Chris@82 565 E T5g, T5h, T5a, T5b;
Chris@82 566 T5g = T3P - T3Q;
Chris@82 567 T5h = T3S - T3T;
Chris@82 568 T5i = FMA(KP618033988, T5h, T5g);
Chris@82 569 T5k = FNMS(KP618033988, T5g, T5h);
Chris@82 570 T59 = T57 - T58;
Chris@82 571 T5a = T3Z + T40;
Chris@82 572 T5b = T42 + T43;
Chris@82 573 T5c = T5a + T5b;
Chris@82 574 T5d = FNMS(KP250000000, T5c, T59);
Chris@82 575 T5e = T5a - T5b;
Chris@82 576 }
Chris@82 577 ii[WS(rs, 5)] = T5c + T59;
Chris@82 578 T5j = FNMS(KP559016994, T5e, T5d);
Chris@82 579 ii[WS(rs, 13)] = FNMS(KP951056516, T5k, T5j);
Chris@82 580 ii[WS(rs, 17)] = FMA(KP951056516, T5k, T5j);
Chris@82 581 T5f = FMA(KP559016994, T5e, T5d);
Chris@82 582 ii[WS(rs, 1)] = FNMS(KP951056516, T5i, T5f);
Chris@82 583 ii[WS(rs, 9)] = FMA(KP951056516, T5i, T5f);
Chris@82 584 }
Chris@82 585 }
Chris@82 586 }
Chris@82 587 }
Chris@82 588 }
Chris@82 589
Chris@82 590 static const tw_instr twinstr[] = {
Chris@82 591 {TW_CEXP, 0, 1},
Chris@82 592 {TW_CEXP, 0, 3},
Chris@82 593 {TW_CEXP, 0, 9},
Chris@82 594 {TW_CEXP, 0, 19},
Chris@82 595 {TW_NEXT, 1, 0}
Chris@82 596 };
Chris@82 597
Chris@82 598 static const ct_desc desc = { 20, "t2_20", twinstr, &GENUS, {136, 58, 140, 0}, 0, 0, 0 };
Chris@82 599
Chris@82 600 void X(codelet_t2_20) (planner *p) {
Chris@82 601 X(kdft_dit_register) (p, t2_20, &desc);
Chris@82 602 }
Chris@82 603 #else
Chris@82 604
Chris@82 605 /* Generated by: ../../../genfft/gen_twiddle.native -compact -variables 4 -pipeline-latency 4 -twiddle-log3 -precompute-twiddles -n 20 -name t2_20 -include dft/scalar/t.h */
Chris@82 606
Chris@82 607 /*
Chris@82 608 * This function contains 276 FP additions, 164 FP multiplications,
Chris@82 609 * (or, 204 additions, 92 multiplications, 72 fused multiply/add),
Chris@82 610 * 123 stack variables, 4 constants, and 80 memory accesses
Chris@82 611 */
Chris@82 612 #include "dft/scalar/t.h"
Chris@82 613
Chris@82 614 static void t2_20(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
Chris@82 615 {
Chris@82 616 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
Chris@82 617 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
Chris@82 618 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
Chris@82 619 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
Chris@82 620 {
Chris@82 621 INT m;
Chris@82 622 for (m = mb, W = W + (mb * 8); m < me; m = m + 1, ri = ri + ms, ii = ii + ms, W = W + 8, MAKE_VOLATILE_STRIDE(40, rs)) {
Chris@82 623 E T2, T5, Tg, Ti, Tk, To, T1h, T1f, T6, T3, T8, T14, T1Q, Tc, T1O;
Chris@82 624 E T1v, T18, T1t, T1n, T24, T1j, T22, Tq, Tu, T1E, T1G, Tx, Ty, Tz, TJ;
Chris@82 625 E T1Z, TB, T1X, T1A, TZ, TL, T1y, TX;
Chris@82 626 {
Chris@82 627 E T7, T16, Ta, T13, T4, T17, Tb, T12;
Chris@82 628 {
Chris@82 629 E Th, Tn, Tj, Tm;
Chris@82 630 T2 = W[0];
Chris@82 631 T5 = W[1];
Chris@82 632 Tg = W[2];
Chris@82 633 Ti = W[3];
Chris@82 634 Th = T2 * Tg;
Chris@82 635 Tn = T5 * Tg;
Chris@82 636 Tj = T5 * Ti;
Chris@82 637 Tm = T2 * Ti;
Chris@82 638 Tk = Th - Tj;
Chris@82 639 To = Tm + Tn;
Chris@82 640 T1h = Tm - Tn;
Chris@82 641 T1f = Th + Tj;
Chris@82 642 T6 = W[5];
Chris@82 643 T7 = T5 * T6;
Chris@82 644 T16 = Tg * T6;
Chris@82 645 Ta = T2 * T6;
Chris@82 646 T13 = Ti * T6;
Chris@82 647 T3 = W[4];
Chris@82 648 T4 = T2 * T3;
Chris@82 649 T17 = Ti * T3;
Chris@82 650 Tb = T5 * T3;
Chris@82 651 T12 = Tg * T3;
Chris@82 652 }
Chris@82 653 T8 = T4 - T7;
Chris@82 654 T14 = T12 + T13;
Chris@82 655 T1Q = T16 + T17;
Chris@82 656 Tc = Ta + Tb;
Chris@82 657 T1O = T12 - T13;
Chris@82 658 T1v = Ta - Tb;
Chris@82 659 T18 = T16 - T17;
Chris@82 660 T1t = T4 + T7;
Chris@82 661 {
Chris@82 662 E T1l, T1m, T1g, T1i;
Chris@82 663 T1l = T1f * T6;
Chris@82 664 T1m = T1h * T3;
Chris@82 665 T1n = T1l + T1m;
Chris@82 666 T24 = T1l - T1m;
Chris@82 667 T1g = T1f * T3;
Chris@82 668 T1i = T1h * T6;
Chris@82 669 T1j = T1g - T1i;
Chris@82 670 T22 = T1g + T1i;
Chris@82 671 {
Chris@82 672 E Tl, Tp, Ts, Tt;
Chris@82 673 Tl = Tk * T3;
Chris@82 674 Tp = To * T6;
Chris@82 675 Tq = Tl + Tp;
Chris@82 676 Ts = Tk * T6;
Chris@82 677 Tt = To * T3;
Chris@82 678 Tu = Ts - Tt;
Chris@82 679 T1E = Tl - Tp;
Chris@82 680 T1G = Ts + Tt;
Chris@82 681 Tx = W[6];
Chris@82 682 Ty = W[7];
Chris@82 683 Tz = FMA(Tk, Tx, To * Ty);
Chris@82 684 TJ = FMA(Tq, Tx, Tu * Ty);
Chris@82 685 T1Z = FNMS(T1h, Tx, T1f * Ty);
Chris@82 686 TB = FNMS(To, Tx, Tk * Ty);
Chris@82 687 T1X = FMA(T1f, Tx, T1h * Ty);
Chris@82 688 T1A = FNMS(T5, Tx, T2 * Ty);
Chris@82 689 TZ = FNMS(Ti, Tx, Tg * Ty);
Chris@82 690 TL = FNMS(Tu, Tx, Tq * Ty);
Chris@82 691 T1y = FMA(T2, Tx, T5 * Ty);
Chris@82 692 TX = FMA(Tg, Tx, Ti * Ty);
Chris@82 693 }
Chris@82 694 }
Chris@82 695 }
Chris@82 696 {
Chris@82 697 E TF, T2b, T4A, T4J, T2K, T3r, T4a, T4m, T1N, T28, T29, T3C, T3F, T4o, T3X;
Chris@82 698 E T3Y, T44, T2f, T2g, T2h, T2n, T2s, T4L, T3g, T3h, T4w, T3n, T3o, T3p, T30;
Chris@82 699 E T35, T36, TW, T1r, T1s, T3J, T3M, T4n, T3U, T3V, T43, T2c, T2d, T2e, T2y;
Chris@82 700 E T2D, T4K, T3d, T3e, T4v, T3k, T3l, T3m, T2P, T2U, T2V;
Chris@82 701 {
Chris@82 702 E T1, T48, Te, T47, Tw, T2H, TD, T2I, T9, Td;
Chris@82 703 T1 = ri[0];
Chris@82 704 T48 = ii[0];
Chris@82 705 T9 = ri[WS(rs, 10)];
Chris@82 706 Td = ii[WS(rs, 10)];
Chris@82 707 Te = FMA(T8, T9, Tc * Td);
Chris@82 708 T47 = FNMS(Tc, T9, T8 * Td);
Chris@82 709 {
Chris@82 710 E Tr, Tv, TA, TC;
Chris@82 711 Tr = ri[WS(rs, 5)];
Chris@82 712 Tv = ii[WS(rs, 5)];
Chris@82 713 Tw = FMA(Tq, Tr, Tu * Tv);
Chris@82 714 T2H = FNMS(Tu, Tr, Tq * Tv);
Chris@82 715 TA = ri[WS(rs, 15)];
Chris@82 716 TC = ii[WS(rs, 15)];
Chris@82 717 TD = FMA(Tz, TA, TB * TC);
Chris@82 718 T2I = FNMS(TB, TA, Tz * TC);
Chris@82 719 }
Chris@82 720 {
Chris@82 721 E Tf, TE, T4y, T4z;
Chris@82 722 Tf = T1 + Te;
Chris@82 723 TE = Tw + TD;
Chris@82 724 TF = Tf - TE;
Chris@82 725 T2b = Tf + TE;
Chris@82 726 T4y = T48 - T47;
Chris@82 727 T4z = Tw - TD;
Chris@82 728 T4A = T4y - T4z;
Chris@82 729 T4J = T4z + T4y;
Chris@82 730 }
Chris@82 731 {
Chris@82 732 E T2G, T2J, T46, T49;
Chris@82 733 T2G = T1 - Te;
Chris@82 734 T2J = T2H - T2I;
Chris@82 735 T2K = T2G - T2J;
Chris@82 736 T3r = T2G + T2J;
Chris@82 737 T46 = T2H + T2I;
Chris@82 738 T49 = T47 + T48;
Chris@82 739 T4a = T46 + T49;
Chris@82 740 T4m = T49 - T46;
Chris@82 741 }
Chris@82 742 }
Chris@82 743 {
Chris@82 744 E T1D, T3A, T2l, T2W, T27, T3E, T2r, T34, T1M, T3B, T2m, T2Z, T1W, T3D, T2q;
Chris@82 745 E T31;
Chris@82 746 {
Chris@82 747 E T1x, T2j, T1C, T2k;
Chris@82 748 {
Chris@82 749 E T1u, T1w, T1z, T1B;
Chris@82 750 T1u = ri[WS(rs, 8)];
Chris@82 751 T1w = ii[WS(rs, 8)];
Chris@82 752 T1x = FMA(T1t, T1u, T1v * T1w);
Chris@82 753 T2j = FNMS(T1v, T1u, T1t * T1w);
Chris@82 754 T1z = ri[WS(rs, 18)];
Chris@82 755 T1B = ii[WS(rs, 18)];
Chris@82 756 T1C = FMA(T1y, T1z, T1A * T1B);
Chris@82 757 T2k = FNMS(T1A, T1z, T1y * T1B);
Chris@82 758 }
Chris@82 759 T1D = T1x + T1C;
Chris@82 760 T3A = T2j + T2k;
Chris@82 761 T2l = T2j - T2k;
Chris@82 762 T2W = T1x - T1C;
Chris@82 763 }
Chris@82 764 {
Chris@82 765 E T21, T32, T26, T33;
Chris@82 766 {
Chris@82 767 E T1Y, T20, T23, T25;
Chris@82 768 T1Y = ri[WS(rs, 17)];
Chris@82 769 T20 = ii[WS(rs, 17)];
Chris@82 770 T21 = FMA(T1X, T1Y, T1Z * T20);
Chris@82 771 T32 = FNMS(T1Z, T1Y, T1X * T20);
Chris@82 772 T23 = ri[WS(rs, 7)];
Chris@82 773 T25 = ii[WS(rs, 7)];
Chris@82 774 T26 = FMA(T22, T23, T24 * T25);
Chris@82 775 T33 = FNMS(T24, T23, T22 * T25);
Chris@82 776 }
Chris@82 777 T27 = T21 + T26;
Chris@82 778 T3E = T32 + T33;
Chris@82 779 T2r = T21 - T26;
Chris@82 780 T34 = T32 - T33;
Chris@82 781 }
Chris@82 782 {
Chris@82 783 E T1I, T2X, T1L, T2Y;
Chris@82 784 {
Chris@82 785 E T1F, T1H, T1J, T1K;
Chris@82 786 T1F = ri[WS(rs, 13)];
Chris@82 787 T1H = ii[WS(rs, 13)];
Chris@82 788 T1I = FMA(T1E, T1F, T1G * T1H);
Chris@82 789 T2X = FNMS(T1G, T1F, T1E * T1H);
Chris@82 790 T1J = ri[WS(rs, 3)];
Chris@82 791 T1K = ii[WS(rs, 3)];
Chris@82 792 T1L = FMA(Tg, T1J, Ti * T1K);
Chris@82 793 T2Y = FNMS(Ti, T1J, Tg * T1K);
Chris@82 794 }
Chris@82 795 T1M = T1I + T1L;
Chris@82 796 T3B = T2X + T2Y;
Chris@82 797 T2m = T1I - T1L;
Chris@82 798 T2Z = T2X - T2Y;
Chris@82 799 }
Chris@82 800 {
Chris@82 801 E T1S, T2o, T1V, T2p;
Chris@82 802 {
Chris@82 803 E T1P, T1R, T1T, T1U;
Chris@82 804 T1P = ri[WS(rs, 12)];
Chris@82 805 T1R = ii[WS(rs, 12)];
Chris@82 806 T1S = FMA(T1O, T1P, T1Q * T1R);
Chris@82 807 T2o = FNMS(T1Q, T1P, T1O * T1R);
Chris@82 808 T1T = ri[WS(rs, 2)];
Chris@82 809 T1U = ii[WS(rs, 2)];
Chris@82 810 T1V = FMA(T1f, T1T, T1h * T1U);
Chris@82 811 T2p = FNMS(T1h, T1T, T1f * T1U);
Chris@82 812 }
Chris@82 813 T1W = T1S + T1V;
Chris@82 814 T3D = T2o + T2p;
Chris@82 815 T2q = T2o - T2p;
Chris@82 816 T31 = T1S - T1V;
Chris@82 817 }
Chris@82 818 T1N = T1D - T1M;
Chris@82 819 T28 = T1W - T27;
Chris@82 820 T29 = T1N + T28;
Chris@82 821 T3C = T3A - T3B;
Chris@82 822 T3F = T3D - T3E;
Chris@82 823 T4o = T3C + T3F;
Chris@82 824 T3X = T3A + T3B;
Chris@82 825 T3Y = T3D + T3E;
Chris@82 826 T44 = T3X + T3Y;
Chris@82 827 T2f = T1D + T1M;
Chris@82 828 T2g = T1W + T27;
Chris@82 829 T2h = T2f + T2g;
Chris@82 830 T2n = T2l + T2m;
Chris@82 831 T2s = T2q + T2r;
Chris@82 832 T4L = T2n + T2s;
Chris@82 833 T3g = T2l - T2m;
Chris@82 834 T3h = T2q - T2r;
Chris@82 835 T4w = T3g + T3h;
Chris@82 836 T3n = T2W + T2Z;
Chris@82 837 T3o = T31 + T34;
Chris@82 838 T3p = T3n + T3o;
Chris@82 839 T30 = T2W - T2Z;
Chris@82 840 T35 = T31 - T34;
Chris@82 841 T36 = T30 + T35;
Chris@82 842 }
Chris@82 843 {
Chris@82 844 E TO, T3H, T2w, T2L, T1q, T3L, T2C, T2T, TV, T3I, T2x, T2O, T1b, T3K, T2B;
Chris@82 845 E T2Q;
Chris@82 846 {
Chris@82 847 E TI, T2u, TN, T2v;
Chris@82 848 {
Chris@82 849 E TG, TH, TK, TM;
Chris@82 850 TG = ri[WS(rs, 4)];
Chris@82 851 TH = ii[WS(rs, 4)];
Chris@82 852 TI = FMA(Tk, TG, To * TH);
Chris@82 853 T2u = FNMS(To, TG, Tk * TH);
Chris@82 854 TK = ri[WS(rs, 14)];
Chris@82 855 TM = ii[WS(rs, 14)];
Chris@82 856 TN = FMA(TJ, TK, TL * TM);
Chris@82 857 T2v = FNMS(TL, TK, TJ * TM);
Chris@82 858 }
Chris@82 859 TO = TI + TN;
Chris@82 860 T3H = T2u + T2v;
Chris@82 861 T2w = T2u - T2v;
Chris@82 862 T2L = TI - TN;
Chris@82 863 }
Chris@82 864 {
Chris@82 865 E T1e, T2R, T1p, T2S;
Chris@82 866 {
Chris@82 867 E T1c, T1d, T1k, T1o;
Chris@82 868 T1c = ri[WS(rs, 1)];
Chris@82 869 T1d = ii[WS(rs, 1)];
Chris@82 870 T1e = FMA(T2, T1c, T5 * T1d);
Chris@82 871 T2R = FNMS(T5, T1c, T2 * T1d);
Chris@82 872 T1k = ri[WS(rs, 11)];
Chris@82 873 T1o = ii[WS(rs, 11)];
Chris@82 874 T1p = FMA(T1j, T1k, T1n * T1o);
Chris@82 875 T2S = FNMS(T1n, T1k, T1j * T1o);
Chris@82 876 }
Chris@82 877 T1q = T1e + T1p;
Chris@82 878 T3L = T2R + T2S;
Chris@82 879 T2C = T1e - T1p;
Chris@82 880 T2T = T2R - T2S;
Chris@82 881 }
Chris@82 882 {
Chris@82 883 E TR, T2M, TU, T2N;
Chris@82 884 {
Chris@82 885 E TP, TQ, TS, TT;
Chris@82 886 TP = ri[WS(rs, 9)];
Chris@82 887 TQ = ii[WS(rs, 9)];
Chris@82 888 TR = FMA(T3, TP, T6 * TQ);
Chris@82 889 T2M = FNMS(T6, TP, T3 * TQ);
Chris@82 890 TS = ri[WS(rs, 19)];
Chris@82 891 TT = ii[WS(rs, 19)];
Chris@82 892 TU = FMA(Tx, TS, Ty * TT);
Chris@82 893 T2N = FNMS(Ty, TS, Tx * TT);
Chris@82 894 }
Chris@82 895 TV = TR + TU;
Chris@82 896 T3I = T2M + T2N;
Chris@82 897 T2x = TR - TU;
Chris@82 898 T2O = T2M - T2N;
Chris@82 899 }
Chris@82 900 {
Chris@82 901 E T11, T2z, T1a, T2A;
Chris@82 902 {
Chris@82 903 E TY, T10, T15, T19;
Chris@82 904 TY = ri[WS(rs, 16)];
Chris@82 905 T10 = ii[WS(rs, 16)];
Chris@82 906 T11 = FMA(TX, TY, TZ * T10);
Chris@82 907 T2z = FNMS(TZ, TY, TX * T10);
Chris@82 908 T15 = ri[WS(rs, 6)];
Chris@82 909 T19 = ii[WS(rs, 6)];
Chris@82 910 T1a = FMA(T14, T15, T18 * T19);
Chris@82 911 T2A = FNMS(T18, T15, T14 * T19);
Chris@82 912 }
Chris@82 913 T1b = T11 + T1a;
Chris@82 914 T3K = T2z + T2A;
Chris@82 915 T2B = T2z - T2A;
Chris@82 916 T2Q = T11 - T1a;
Chris@82 917 }
Chris@82 918 TW = TO - TV;
Chris@82 919 T1r = T1b - T1q;
Chris@82 920 T1s = TW + T1r;
Chris@82 921 T3J = T3H - T3I;
Chris@82 922 T3M = T3K - T3L;
Chris@82 923 T4n = T3J + T3M;
Chris@82 924 T3U = T3H + T3I;
Chris@82 925 T3V = T3K + T3L;
Chris@82 926 T43 = T3U + T3V;
Chris@82 927 T2c = TO + TV;
Chris@82 928 T2d = T1b + T1q;
Chris@82 929 T2e = T2c + T2d;
Chris@82 930 T2y = T2w + T2x;
Chris@82 931 T2D = T2B + T2C;
Chris@82 932 T4K = T2y + T2D;
Chris@82 933 T3d = T2w - T2x;
Chris@82 934 T3e = T2B - T2C;
Chris@82 935 T4v = T3d + T3e;
Chris@82 936 T3k = T2L + T2O;
Chris@82 937 T3l = T2Q + T2T;
Chris@82 938 T3m = T3k + T3l;
Chris@82 939 T2P = T2L - T2O;
Chris@82 940 T2U = T2Q - T2T;
Chris@82 941 T2V = T2P + T2U;
Chris@82 942 }
Chris@82 943 {
Chris@82 944 E T3y, T2a, T3x, T3O, T3Q, T3G, T3N, T3P, T3z;
Chris@82 945 T3y = KP559016994 * (T1s - T29);
Chris@82 946 T2a = T1s + T29;
Chris@82 947 T3x = FNMS(KP250000000, T2a, TF);
Chris@82 948 T3G = T3C - T3F;
Chris@82 949 T3N = T3J - T3M;
Chris@82 950 T3O = FNMS(KP587785252, T3N, KP951056516 * T3G);
Chris@82 951 T3Q = FMA(KP951056516, T3N, KP587785252 * T3G);
Chris@82 952 ri[WS(rs, 10)] = TF + T2a;
Chris@82 953 T3P = T3y + T3x;
Chris@82 954 ri[WS(rs, 14)] = T3P - T3Q;
Chris@82 955 ri[WS(rs, 6)] = T3P + T3Q;
Chris@82 956 T3z = T3x - T3y;
Chris@82 957 ri[WS(rs, 2)] = T3z - T3O;
Chris@82 958 ri[WS(rs, 18)] = T3z + T3O;
Chris@82 959 }
Chris@82 960 {
Chris@82 961 E T4r, T4p, T4q, T4l, T4u, T4j, T4k, T4t, T4s;
Chris@82 962 T4r = KP559016994 * (T4n - T4o);
Chris@82 963 T4p = T4n + T4o;
Chris@82 964 T4q = FNMS(KP250000000, T4p, T4m);
Chris@82 965 T4j = T1N - T28;
Chris@82 966 T4k = TW - T1r;
Chris@82 967 T4l = FNMS(KP587785252, T4k, KP951056516 * T4j);
Chris@82 968 T4u = FMA(KP951056516, T4k, KP587785252 * T4j);
Chris@82 969 ii[WS(rs, 10)] = T4p + T4m;
Chris@82 970 T4t = T4r + T4q;
Chris@82 971 ii[WS(rs, 6)] = T4t - T4u;
Chris@82 972 ii[WS(rs, 14)] = T4u + T4t;
Chris@82 973 T4s = T4q - T4r;
Chris@82 974 ii[WS(rs, 2)] = T4l + T4s;
Chris@82 975 ii[WS(rs, 18)] = T4s - T4l;
Chris@82 976 }
Chris@82 977 {
Chris@82 978 E T3R, T2i, T3S, T40, T42, T3W, T3Z, T41, T3T;
Chris@82 979 T3R = KP559016994 * (T2e - T2h);
Chris@82 980 T2i = T2e + T2h;
Chris@82 981 T3S = FNMS(KP250000000, T2i, T2b);
Chris@82 982 T3W = T3U - T3V;
Chris@82 983 T3Z = T3X - T3Y;
Chris@82 984 T40 = FMA(KP951056516, T3W, KP587785252 * T3Z);
Chris@82 985 T42 = FNMS(KP587785252, T3W, KP951056516 * T3Z);
Chris@82 986 ri[0] = T2b + T2i;
Chris@82 987 T41 = T3S - T3R;
Chris@82 988 ri[WS(rs, 12)] = T41 - T42;
Chris@82 989 ri[WS(rs, 8)] = T41 + T42;
Chris@82 990 T3T = T3R + T3S;
Chris@82 991 ri[WS(rs, 4)] = T3T - T40;
Chris@82 992 ri[WS(rs, 16)] = T3T + T40;
Chris@82 993 }
Chris@82 994 {
Chris@82 995 E T4e, T45, T4f, T4d, T4i, T4b, T4c, T4h, T4g;
Chris@82 996 T4e = KP559016994 * (T43 - T44);
Chris@82 997 T45 = T43 + T44;
Chris@82 998 T4f = FNMS(KP250000000, T45, T4a);
Chris@82 999 T4b = T2c - T2d;
Chris@82 1000 T4c = T2f - T2g;
Chris@82 1001 T4d = FMA(KP951056516, T4b, KP587785252 * T4c);
Chris@82 1002 T4i = FNMS(KP587785252, T4b, KP951056516 * T4c);
Chris@82 1003 ii[0] = T45 + T4a;
Chris@82 1004 T4h = T4f - T4e;
Chris@82 1005 ii[WS(rs, 8)] = T4h - T4i;
Chris@82 1006 ii[WS(rs, 12)] = T4i + T4h;
Chris@82 1007 T4g = T4e + T4f;
Chris@82 1008 ii[WS(rs, 4)] = T4d + T4g;
Chris@82 1009 ii[WS(rs, 16)] = T4g - T4d;
Chris@82 1010 }
Chris@82 1011 {
Chris@82 1012 E T39, T37, T38, T2F, T3b, T2t, T2E, T3c, T3a;
Chris@82 1013 T39 = KP559016994 * (T2V - T36);
Chris@82 1014 T37 = T2V + T36;
Chris@82 1015 T38 = FNMS(KP250000000, T37, T2K);
Chris@82 1016 T2t = T2n - T2s;
Chris@82 1017 T2E = T2y - T2D;
Chris@82 1018 T2F = FNMS(KP587785252, T2E, KP951056516 * T2t);
Chris@82 1019 T3b = FMA(KP951056516, T2E, KP587785252 * T2t);
Chris@82 1020 ri[WS(rs, 15)] = T2K + T37;
Chris@82 1021 T3c = T39 + T38;
Chris@82 1022 ri[WS(rs, 11)] = T3b + T3c;
Chris@82 1023 ri[WS(rs, 19)] = T3c - T3b;
Chris@82 1024 T3a = T38 - T39;
Chris@82 1025 ri[WS(rs, 3)] = T2F + T3a;
Chris@82 1026 ri[WS(rs, 7)] = T3a - T2F;
Chris@82 1027 }
Chris@82 1028 {
Chris@82 1029 E T4O, T4M, T4N, T4S, T4U, T4Q, T4R, T4T, T4P;
Chris@82 1030 T4O = KP559016994 * (T4K - T4L);
Chris@82 1031 T4M = T4K + T4L;
Chris@82 1032 T4N = FNMS(KP250000000, T4M, T4J);
Chris@82 1033 T4Q = T30 - T35;
Chris@82 1034 T4R = T2P - T2U;
Chris@82 1035 T4S = FNMS(KP587785252, T4R, KP951056516 * T4Q);
Chris@82 1036 T4U = FMA(KP951056516, T4R, KP587785252 * T4Q);
Chris@82 1037 ii[WS(rs, 15)] = T4M + T4J;
Chris@82 1038 T4T = T4O + T4N;
Chris@82 1039 ii[WS(rs, 11)] = T4T - T4U;
Chris@82 1040 ii[WS(rs, 19)] = T4U + T4T;
Chris@82 1041 T4P = T4N - T4O;
Chris@82 1042 ii[WS(rs, 3)] = T4P - T4S;
Chris@82 1043 ii[WS(rs, 7)] = T4S + T4P;
Chris@82 1044 }
Chris@82 1045 {
Chris@82 1046 E T3q, T3s, T3t, T3j, T3v, T3f, T3i, T3w, T3u;
Chris@82 1047 T3q = KP559016994 * (T3m - T3p);
Chris@82 1048 T3s = T3m + T3p;
Chris@82 1049 T3t = FNMS(KP250000000, T3s, T3r);
Chris@82 1050 T3f = T3d - T3e;
Chris@82 1051 T3i = T3g - T3h;
Chris@82 1052 T3j = FMA(KP951056516, T3f, KP587785252 * T3i);
Chris@82 1053 T3v = FNMS(KP587785252, T3f, KP951056516 * T3i);
Chris@82 1054 ri[WS(rs, 5)] = T3r + T3s;
Chris@82 1055 T3w = T3t - T3q;
Chris@82 1056 ri[WS(rs, 13)] = T3v + T3w;
Chris@82 1057 ri[WS(rs, 17)] = T3w - T3v;
Chris@82 1058 T3u = T3q + T3t;
Chris@82 1059 ri[WS(rs, 1)] = T3j + T3u;
Chris@82 1060 ri[WS(rs, 9)] = T3u - T3j;
Chris@82 1061 }
Chris@82 1062 {
Chris@82 1063 E T4x, T4B, T4C, T4G, T4I, T4E, T4F, T4H, T4D;
Chris@82 1064 T4x = KP559016994 * (T4v - T4w);
Chris@82 1065 T4B = T4v + T4w;
Chris@82 1066 T4C = FNMS(KP250000000, T4B, T4A);
Chris@82 1067 T4E = T3k - T3l;
Chris@82 1068 T4F = T3n - T3o;
Chris@82 1069 T4G = FMA(KP951056516, T4E, KP587785252 * T4F);
Chris@82 1070 T4I = FNMS(KP587785252, T4E, KP951056516 * T4F);
Chris@82 1071 ii[WS(rs, 5)] = T4B + T4A;
Chris@82 1072 T4H = T4C - T4x;
Chris@82 1073 ii[WS(rs, 13)] = T4H - T4I;
Chris@82 1074 ii[WS(rs, 17)] = T4I + T4H;
Chris@82 1075 T4D = T4x + T4C;
Chris@82 1076 ii[WS(rs, 1)] = T4D - T4G;
Chris@82 1077 ii[WS(rs, 9)] = T4G + T4D;
Chris@82 1078 }
Chris@82 1079 }
Chris@82 1080 }
Chris@82 1081 }
Chris@82 1082 }
Chris@82 1083
Chris@82 1084 static const tw_instr twinstr[] = {
Chris@82 1085 {TW_CEXP, 0, 1},
Chris@82 1086 {TW_CEXP, 0, 3},
Chris@82 1087 {TW_CEXP, 0, 9},
Chris@82 1088 {TW_CEXP, 0, 19},
Chris@82 1089 {TW_NEXT, 1, 0}
Chris@82 1090 };
Chris@82 1091
Chris@82 1092 static const ct_desc desc = { 20, "t2_20", twinstr, &GENUS, {204, 92, 72, 0}, 0, 0, 0 };
Chris@82 1093
Chris@82 1094 void X(codelet_t2_20) (planner *p) {
Chris@82 1095 X(kdft_dit_register) (p, t2_20, &desc);
Chris@82 1096 }
Chris@82 1097 #endif