Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.8/dft/scalar/codelets/t1_15.c @ 82:d0c2a83c1364

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Add FFTW 3.3.8 source, and a Linux build
author Chris Cannam
date Tue, 19 Nov 2019 14:52:55 +0000
parents
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 08:04:14 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 -n 15 -name t1_15 -include dft/scalar/t.h */
Chris@82 29
Chris@82 30 /*
Chris@82 31 * This function contains 184 FP additions, 140 FP multiplications,
Chris@82 32 * (or, 72 additions, 28 multiplications, 112 fused multiply/add),
Chris@82 33 * 51 stack variables, 6 constants, and 60 memory accesses
Chris@82 34 */
Chris@82 35 #include "dft/scalar/t.h"
Chris@82 36
Chris@82 37 static void t1_15(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 DK(KP866025403, +0.866025403784438646763723170752936183471402627);
Chris@82 44 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
Chris@82 45 {
Chris@82 46 INT m;
Chris@82 47 for (m = mb, W = W + (mb * 28); m < me; m = m + 1, ri = ri + ms, ii = ii + ms, W = W + 28, MAKE_VOLATILE_STRIDE(30, rs)) {
Chris@82 48 E T1, T3j, T1G, T3u, Te, T1B, T3i, T3t, T1y, T2i, T2a, T2M, T37, T2V, Tz;
Chris@82 49 E T2e, T1O, T2t, T39, T2X, TT, T2f, T1V, T2z, T3a, T2Y, T1e, T2h, T23, T2G;
Chris@82 50 E T36, T2U;
Chris@82 51 {
Chris@82 52 E T7, T1D, Td, T1F;
Chris@82 53 T1 = ri[0];
Chris@82 54 T3j = ii[0];
Chris@82 55 {
Chris@82 56 E T3, T6, T4, T1C, T2, T5;
Chris@82 57 T3 = ri[WS(rs, 5)];
Chris@82 58 T6 = ii[WS(rs, 5)];
Chris@82 59 T2 = W[8];
Chris@82 60 T4 = T2 * T3;
Chris@82 61 T1C = T2 * T6;
Chris@82 62 T5 = W[9];
Chris@82 63 T7 = FMA(T5, T6, T4);
Chris@82 64 T1D = FNMS(T5, T3, T1C);
Chris@82 65 }
Chris@82 66 {
Chris@82 67 E T9, Tc, Ta, T1E, T8, Tb;
Chris@82 68 T9 = ri[WS(rs, 10)];
Chris@82 69 Tc = ii[WS(rs, 10)];
Chris@82 70 T8 = W[18];
Chris@82 71 Ta = T8 * T9;
Chris@82 72 T1E = T8 * Tc;
Chris@82 73 Tb = W[19];
Chris@82 74 Td = FMA(Tb, Tc, Ta);
Chris@82 75 T1F = FNMS(Tb, T9, T1E);
Chris@82 76 }
Chris@82 77 T1G = T1D - T1F;
Chris@82 78 T3u = Td - T7;
Chris@82 79 Te = T7 + Td;
Chris@82 80 T1B = FNMS(KP500000000, Te, T1);
Chris@82 81 T3i = T1D + T1F;
Chris@82 82 T3t = FNMS(KP500000000, T3i, T3j);
Chris@82 83 }
Chris@82 84 {
Chris@82 85 E T1k, T2I, T1w, T28, T1q, T26;
Chris@82 86 {
Chris@82 87 E T1g, T1j, T1h, T2H, T1f, T1i;
Chris@82 88 T1g = ri[WS(rs, 9)];
Chris@82 89 T1j = ii[WS(rs, 9)];
Chris@82 90 T1f = W[16];
Chris@82 91 T1h = T1f * T1g;
Chris@82 92 T2H = T1f * T1j;
Chris@82 93 T1i = W[17];
Chris@82 94 T1k = FMA(T1i, T1j, T1h);
Chris@82 95 T2I = FNMS(T1i, T1g, T2H);
Chris@82 96 }
Chris@82 97 {
Chris@82 98 E T1s, T1v, T1t, T27, T1r, T1u;
Chris@82 99 T1s = ri[WS(rs, 4)];
Chris@82 100 T1v = ii[WS(rs, 4)];
Chris@82 101 T1r = W[6];
Chris@82 102 T1t = T1r * T1s;
Chris@82 103 T27 = T1r * T1v;
Chris@82 104 T1u = W[7];
Chris@82 105 T1w = FMA(T1u, T1v, T1t);
Chris@82 106 T28 = FNMS(T1u, T1s, T27);
Chris@82 107 }
Chris@82 108 {
Chris@82 109 E T1m, T1p, T1n, T25, T1l, T1o;
Chris@82 110 T1m = ri[WS(rs, 14)];
Chris@82 111 T1p = ii[WS(rs, 14)];
Chris@82 112 T1l = W[26];
Chris@82 113 T1n = T1l * T1m;
Chris@82 114 T25 = T1l * T1p;
Chris@82 115 T1o = W[27];
Chris@82 116 T1q = FMA(T1o, T1p, T1n);
Chris@82 117 T26 = FNMS(T1o, T1m, T25);
Chris@82 118 }
Chris@82 119 {
Chris@82 120 E T29, T1x, T24, T2L, T2J, T2K;
Chris@82 121 T29 = T26 - T28;
Chris@82 122 T1x = T1q + T1w;
Chris@82 123 T24 = FNMS(KP500000000, T1x, T1k);
Chris@82 124 T1y = T1k + T1x;
Chris@82 125 T2i = FMA(KP866025403, T29, T24);
Chris@82 126 T2a = FNMS(KP866025403, T29, T24);
Chris@82 127 T2L = T1w - T1q;
Chris@82 128 T2J = T26 + T28;
Chris@82 129 T2K = FNMS(KP500000000, T2J, T2I);
Chris@82 130 T2M = FMA(KP866025403, T2L, T2K);
Chris@82 131 T37 = T2I + T2J;
Chris@82 132 T2V = FNMS(KP866025403, T2L, T2K);
Chris@82 133 }
Chris@82 134 }
Chris@82 135 {
Chris@82 136 E Tl, T2p, Tx, T1M, Tr, T1K;
Chris@82 137 {
Chris@82 138 E Th, Tk, Ti, T2o, Tg, Tj;
Chris@82 139 Th = ri[WS(rs, 3)];
Chris@82 140 Tk = ii[WS(rs, 3)];
Chris@82 141 Tg = W[4];
Chris@82 142 Ti = Tg * Th;
Chris@82 143 T2o = Tg * Tk;
Chris@82 144 Tj = W[5];
Chris@82 145 Tl = FMA(Tj, Tk, Ti);
Chris@82 146 T2p = FNMS(Tj, Th, T2o);
Chris@82 147 }
Chris@82 148 {
Chris@82 149 E Tt, Tw, Tu, T1L, Ts, Tv;
Chris@82 150 Tt = ri[WS(rs, 13)];
Chris@82 151 Tw = ii[WS(rs, 13)];
Chris@82 152 Ts = W[24];
Chris@82 153 Tu = Ts * Tt;
Chris@82 154 T1L = Ts * Tw;
Chris@82 155 Tv = W[25];
Chris@82 156 Tx = FMA(Tv, Tw, Tu);
Chris@82 157 T1M = FNMS(Tv, Tt, T1L);
Chris@82 158 }
Chris@82 159 {
Chris@82 160 E Tn, Tq, To, T1J, Tm, Tp;
Chris@82 161 Tn = ri[WS(rs, 8)];
Chris@82 162 Tq = ii[WS(rs, 8)];
Chris@82 163 Tm = W[14];
Chris@82 164 To = Tm * Tn;
Chris@82 165 T1J = Tm * Tq;
Chris@82 166 Tp = W[15];
Chris@82 167 Tr = FMA(Tp, Tq, To);
Chris@82 168 T1K = FNMS(Tp, Tn, T1J);
Chris@82 169 }
Chris@82 170 {
Chris@82 171 E T1N, Ty, T1I, T2s, T2q, T2r;
Chris@82 172 T1N = T1K - T1M;
Chris@82 173 Ty = Tr + Tx;
Chris@82 174 T1I = FNMS(KP500000000, Ty, Tl);
Chris@82 175 Tz = Tl + Ty;
Chris@82 176 T2e = FMA(KP866025403, T1N, T1I);
Chris@82 177 T1O = FNMS(KP866025403, T1N, T1I);
Chris@82 178 T2s = Tx - Tr;
Chris@82 179 T2q = T1K + T1M;
Chris@82 180 T2r = FNMS(KP500000000, T2q, T2p);
Chris@82 181 T2t = FMA(KP866025403, T2s, T2r);
Chris@82 182 T39 = T2p + T2q;
Chris@82 183 T2X = FNMS(KP866025403, T2s, T2r);
Chris@82 184 }
Chris@82 185 }
Chris@82 186 {
Chris@82 187 E TF, T2v, TR, T1T, TL, T1R;
Chris@82 188 {
Chris@82 189 E TB, TE, TC, T2u, TA, TD;
Chris@82 190 TB = ri[WS(rs, 12)];
Chris@82 191 TE = ii[WS(rs, 12)];
Chris@82 192 TA = W[22];
Chris@82 193 TC = TA * TB;
Chris@82 194 T2u = TA * TE;
Chris@82 195 TD = W[23];
Chris@82 196 TF = FMA(TD, TE, TC);
Chris@82 197 T2v = FNMS(TD, TB, T2u);
Chris@82 198 }
Chris@82 199 {
Chris@82 200 E TN, TQ, TO, T1S, TM, TP;
Chris@82 201 TN = ri[WS(rs, 7)];
Chris@82 202 TQ = ii[WS(rs, 7)];
Chris@82 203 TM = W[12];
Chris@82 204 TO = TM * TN;
Chris@82 205 T1S = TM * TQ;
Chris@82 206 TP = W[13];
Chris@82 207 TR = FMA(TP, TQ, TO);
Chris@82 208 T1T = FNMS(TP, TN, T1S);
Chris@82 209 }
Chris@82 210 {
Chris@82 211 E TH, TK, TI, T1Q, TG, TJ;
Chris@82 212 TH = ri[WS(rs, 2)];
Chris@82 213 TK = ii[WS(rs, 2)];
Chris@82 214 TG = W[2];
Chris@82 215 TI = TG * TH;
Chris@82 216 T1Q = TG * TK;
Chris@82 217 TJ = W[3];
Chris@82 218 TL = FMA(TJ, TK, TI);
Chris@82 219 T1R = FNMS(TJ, TH, T1Q);
Chris@82 220 }
Chris@82 221 {
Chris@82 222 E T1U, TS, T1P, T2y, T2w, T2x;
Chris@82 223 T1U = T1R - T1T;
Chris@82 224 TS = TL + TR;
Chris@82 225 T1P = FNMS(KP500000000, TS, TF);
Chris@82 226 TT = TF + TS;
Chris@82 227 T2f = FMA(KP866025403, T1U, T1P);
Chris@82 228 T1V = FNMS(KP866025403, T1U, T1P);
Chris@82 229 T2y = TR - TL;
Chris@82 230 T2w = T1R + T1T;
Chris@82 231 T2x = FNMS(KP500000000, T2w, T2v);
Chris@82 232 T2z = FMA(KP866025403, T2y, T2x);
Chris@82 233 T3a = T2v + T2w;
Chris@82 234 T2Y = FNMS(KP866025403, T2y, T2x);
Chris@82 235 }
Chris@82 236 }
Chris@82 237 {
Chris@82 238 E T10, T2C, T1c, T21, T16, T1Z;
Chris@82 239 {
Chris@82 240 E TW, TZ, TX, T2B, TV, TY;
Chris@82 241 TW = ri[WS(rs, 6)];
Chris@82 242 TZ = ii[WS(rs, 6)];
Chris@82 243 TV = W[10];
Chris@82 244 TX = TV * TW;
Chris@82 245 T2B = TV * TZ;
Chris@82 246 TY = W[11];
Chris@82 247 T10 = FMA(TY, TZ, TX);
Chris@82 248 T2C = FNMS(TY, TW, T2B);
Chris@82 249 }
Chris@82 250 {
Chris@82 251 E T18, T1b, T19, T20, T17, T1a;
Chris@82 252 T18 = ri[WS(rs, 1)];
Chris@82 253 T1b = ii[WS(rs, 1)];
Chris@82 254 T17 = W[0];
Chris@82 255 T19 = T17 * T18;
Chris@82 256 T20 = T17 * T1b;
Chris@82 257 T1a = W[1];
Chris@82 258 T1c = FMA(T1a, T1b, T19);
Chris@82 259 T21 = FNMS(T1a, T18, T20);
Chris@82 260 }
Chris@82 261 {
Chris@82 262 E T12, T15, T13, T1Y, T11, T14;
Chris@82 263 T12 = ri[WS(rs, 11)];
Chris@82 264 T15 = ii[WS(rs, 11)];
Chris@82 265 T11 = W[20];
Chris@82 266 T13 = T11 * T12;
Chris@82 267 T1Y = T11 * T15;
Chris@82 268 T14 = W[21];
Chris@82 269 T16 = FMA(T14, T15, T13);
Chris@82 270 T1Z = FNMS(T14, T12, T1Y);
Chris@82 271 }
Chris@82 272 {
Chris@82 273 E T22, T1d, T1X, T2F, T2D, T2E;
Chris@82 274 T22 = T1Z - T21;
Chris@82 275 T1d = T16 + T1c;
Chris@82 276 T1X = FNMS(KP500000000, T1d, T10);
Chris@82 277 T1e = T10 + T1d;
Chris@82 278 T2h = FMA(KP866025403, T22, T1X);
Chris@82 279 T23 = FNMS(KP866025403, T22, T1X);
Chris@82 280 T2F = T1c - T16;
Chris@82 281 T2D = T1Z + T21;
Chris@82 282 T2E = FNMS(KP500000000, T2D, T2C);
Chris@82 283 T2G = FMA(KP866025403, T2F, T2E);
Chris@82 284 T36 = T2C + T2D;
Chris@82 285 T2U = FNMS(KP866025403, T2F, T2E);
Chris@82 286 }
Chris@82 287 }
Chris@82 288 {
Chris@82 289 E T3c, T3e, Tf, T1A, T33, T34, T3d, T35;
Chris@82 290 {
Chris@82 291 E T38, T3b, TU, T1z;
Chris@82 292 T38 = T36 - T37;
Chris@82 293 T3b = T39 - T3a;
Chris@82 294 T3c = FNMS(KP618033988, T3b, T38);
Chris@82 295 T3e = FMA(KP618033988, T38, T3b);
Chris@82 296 Tf = T1 + Te;
Chris@82 297 TU = Tz + TT;
Chris@82 298 T1z = T1e + T1y;
Chris@82 299 T1A = TU + T1z;
Chris@82 300 T33 = FNMS(KP250000000, T1A, Tf);
Chris@82 301 T34 = TU - T1z;
Chris@82 302 }
Chris@82 303 ri[0] = Tf + T1A;
Chris@82 304 T3d = FMA(KP559016994, T34, T33);
Chris@82 305 ri[WS(rs, 9)] = FNMS(KP951056516, T3e, T3d);
Chris@82 306 ri[WS(rs, 6)] = FMA(KP951056516, T3e, T3d);
Chris@82 307 T35 = FNMS(KP559016994, T34, T33);
Chris@82 308 ri[WS(rs, 12)] = FNMS(KP951056516, T3c, T35);
Chris@82 309 ri[WS(rs, 3)] = FMA(KP951056516, T3c, T35);
Chris@82 310 }
Chris@82 311 {
Chris@82 312 E T3q, T3s, T3k, T3h, T3l, T3m, T3r, T3n;
Chris@82 313 {
Chris@82 314 E T3o, T3p, T3f, T3g;
Chris@82 315 T3o = T1e - T1y;
Chris@82 316 T3p = Tz - TT;
Chris@82 317 T3q = FNMS(KP618033988, T3p, T3o);
Chris@82 318 T3s = FMA(KP618033988, T3o, T3p);
Chris@82 319 T3k = T3i + T3j;
Chris@82 320 T3f = T39 + T3a;
Chris@82 321 T3g = T36 + T37;
Chris@82 322 T3h = T3f + T3g;
Chris@82 323 T3l = FNMS(KP250000000, T3h, T3k);
Chris@82 324 T3m = T3f - T3g;
Chris@82 325 }
Chris@82 326 ii[0] = T3h + T3k;
Chris@82 327 T3r = FMA(KP559016994, T3m, T3l);
Chris@82 328 ii[WS(rs, 6)] = FNMS(KP951056516, T3s, T3r);
Chris@82 329 ii[WS(rs, 9)] = FMA(KP951056516, T3s, T3r);
Chris@82 330 T3n = FNMS(KP559016994, T3m, T3l);
Chris@82 331 ii[WS(rs, 3)] = FNMS(KP951056516, T3q, T3n);
Chris@82 332 ii[WS(rs, 12)] = FMA(KP951056516, T3q, T3n);
Chris@82 333 }
Chris@82 334 {
Chris@82 335 E T30, T32, T1H, T2c, T2R, T2S, T31, T2T;
Chris@82 336 {
Chris@82 337 E T2W, T2Z, T1W, T2b;
Chris@82 338 T2W = T2U - T2V;
Chris@82 339 T2Z = T2X - T2Y;
Chris@82 340 T30 = FNMS(KP618033988, T2Z, T2W);
Chris@82 341 T32 = FMA(KP618033988, T2W, T2Z);
Chris@82 342 T1H = FNMS(KP866025403, T1G, T1B);
Chris@82 343 T1W = T1O + T1V;
Chris@82 344 T2b = T23 + T2a;
Chris@82 345 T2c = T1W + T2b;
Chris@82 346 T2R = FNMS(KP250000000, T2c, T1H);
Chris@82 347 T2S = T1W - T2b;
Chris@82 348 }
Chris@82 349 ri[WS(rs, 5)] = T1H + T2c;
Chris@82 350 T31 = FMA(KP559016994, T2S, T2R);
Chris@82 351 ri[WS(rs, 14)] = FNMS(KP951056516, T32, T31);
Chris@82 352 ri[WS(rs, 11)] = FMA(KP951056516, T32, T31);
Chris@82 353 T2T = FNMS(KP559016994, T2S, T2R);
Chris@82 354 ri[WS(rs, 2)] = FNMS(KP951056516, T30, T2T);
Chris@82 355 ri[WS(rs, 8)] = FMA(KP951056516, T30, T2T);
Chris@82 356 }
Chris@82 357 {
Chris@82 358 E T3Q, T3S, T3H, T3K, T3L, T3M, T3R, T3N;
Chris@82 359 {
Chris@82 360 E T3O, T3P, T3I, T3J;
Chris@82 361 T3O = T23 - T2a;
Chris@82 362 T3P = T1O - T1V;
Chris@82 363 T3Q = FNMS(KP618033988, T3P, T3O);
Chris@82 364 T3S = FMA(KP618033988, T3O, T3P);
Chris@82 365 T3H = FNMS(KP866025403, T3u, T3t);
Chris@82 366 T3I = T2X + T2Y;
Chris@82 367 T3J = T2U + T2V;
Chris@82 368 T3K = T3I + T3J;
Chris@82 369 T3L = FNMS(KP250000000, T3K, T3H);
Chris@82 370 T3M = T3I - T3J;
Chris@82 371 }
Chris@82 372 ii[WS(rs, 5)] = T3K + T3H;
Chris@82 373 T3R = FMA(KP559016994, T3M, T3L);
Chris@82 374 ii[WS(rs, 11)] = FNMS(KP951056516, T3S, T3R);
Chris@82 375 ii[WS(rs, 14)] = FMA(KP951056516, T3S, T3R);
Chris@82 376 T3N = FNMS(KP559016994, T3M, T3L);
Chris@82 377 ii[WS(rs, 2)] = FMA(KP951056516, T3Q, T3N);
Chris@82 378 ii[WS(rs, 8)] = FNMS(KP951056516, T3Q, T3N);
Chris@82 379 }
Chris@82 380 {
Chris@82 381 E T3E, T3G, T3v, T3y, T3z, T3A, T3F, T3B;
Chris@82 382 {
Chris@82 383 E T3C, T3D, T3w, T3x;
Chris@82 384 T3C = T2e - T2f;
Chris@82 385 T3D = T2h - T2i;
Chris@82 386 T3E = FMA(KP618033988, T3D, T3C);
Chris@82 387 T3G = FNMS(KP618033988, T3C, T3D);
Chris@82 388 T3v = FMA(KP866025403, T3u, T3t);
Chris@82 389 T3w = T2t + T2z;
Chris@82 390 T3x = T2G + T2M;
Chris@82 391 T3y = T3w + T3x;
Chris@82 392 T3z = FNMS(KP250000000, T3y, T3v);
Chris@82 393 T3A = T3w - T3x;
Chris@82 394 }
Chris@82 395 ii[WS(rs, 10)] = T3y + T3v;
Chris@82 396 T3F = FNMS(KP559016994, T3A, T3z);
Chris@82 397 ii[WS(rs, 7)] = FMA(KP951056516, T3G, T3F);
Chris@82 398 ii[WS(rs, 13)] = FNMS(KP951056516, T3G, T3F);
Chris@82 399 T3B = FMA(KP559016994, T3A, T3z);
Chris@82 400 ii[WS(rs, 1)] = FNMS(KP951056516, T3E, T3B);
Chris@82 401 ii[WS(rs, 4)] = FMA(KP951056516, T3E, T3B);
Chris@82 402 }
Chris@82 403 {
Chris@82 404 E T2O, T2Q, T2d, T2k, T2l, T2m, T2P, T2n;
Chris@82 405 {
Chris@82 406 E T2A, T2N, T2g, T2j;
Chris@82 407 T2A = T2t - T2z;
Chris@82 408 T2N = T2G - T2M;
Chris@82 409 T2O = FMA(KP618033988, T2N, T2A);
Chris@82 410 T2Q = FNMS(KP618033988, T2A, T2N);
Chris@82 411 T2d = FMA(KP866025403, T1G, T1B);
Chris@82 412 T2g = T2e + T2f;
Chris@82 413 T2j = T2h + T2i;
Chris@82 414 T2k = T2g + T2j;
Chris@82 415 T2l = FNMS(KP250000000, T2k, T2d);
Chris@82 416 T2m = T2g - T2j;
Chris@82 417 }
Chris@82 418 ri[WS(rs, 10)] = T2d + T2k;
Chris@82 419 T2P = FNMS(KP559016994, T2m, T2l);
Chris@82 420 ri[WS(rs, 7)] = FNMS(KP951056516, T2Q, T2P);
Chris@82 421 ri[WS(rs, 13)] = FMA(KP951056516, T2Q, T2P);
Chris@82 422 T2n = FMA(KP559016994, T2m, T2l);
Chris@82 423 ri[WS(rs, 4)] = FNMS(KP951056516, T2O, T2n);
Chris@82 424 ri[WS(rs, 1)] = FMA(KP951056516, T2O, T2n);
Chris@82 425 }
Chris@82 426 }
Chris@82 427 }
Chris@82 428 }
Chris@82 429
Chris@82 430 static const tw_instr twinstr[] = {
Chris@82 431 {TW_FULL, 0, 15},
Chris@82 432 {TW_NEXT, 1, 0}
Chris@82 433 };
Chris@82 434
Chris@82 435 static const ct_desc desc = { 15, "t1_15", twinstr, &GENUS, {72, 28, 112, 0}, 0, 0, 0 };
Chris@82 436
Chris@82 437 void X(codelet_t1_15) (planner *p) {
Chris@82 438 X(kdft_dit_register) (p, t1_15, &desc);
Chris@82 439 }
Chris@82 440 #else
Chris@82 441
Chris@82 442 /* Generated by: ../../../genfft/gen_twiddle.native -compact -variables 4 -pipeline-latency 4 -n 15 -name t1_15 -include dft/scalar/t.h */
Chris@82 443
Chris@82 444 /*
Chris@82 445 * This function contains 184 FP additions, 112 FP multiplications,
Chris@82 446 * (or, 128 additions, 56 multiplications, 56 fused multiply/add),
Chris@82 447 * 65 stack variables, 6 constants, and 60 memory accesses
Chris@82 448 */
Chris@82 449 #include "dft/scalar/t.h"
Chris@82 450
Chris@82 451 static void t1_15(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
Chris@82 452 {
Chris@82 453 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
Chris@82 454 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
Chris@82 455 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
Chris@82 456 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
Chris@82 457 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
Chris@82 458 DK(KP866025403, +0.866025403784438646763723170752936183471402627);
Chris@82 459 {
Chris@82 460 INT m;
Chris@82 461 for (m = mb, W = W + (mb * 28); m < me; m = m + 1, ri = ri + ms, ii = ii + ms, W = W + 28, MAKE_VOLATILE_STRIDE(30, rs)) {
Chris@82 462 E T1q, T34, Td, T1n, T2S, T35, T13, T1k, T1l, T2E, T2F, T2O, T1H, T1T, T2k;
Chris@82 463 E T2t, T2f, T2s, T1M, T1U, Tu, TL, TM, T2H, T2I, T2N, T1w, T1Q, T29, T2w;
Chris@82 464 E T24, T2v, T1B, T1R;
Chris@82 465 {
Chris@82 466 E T1, T2R, T6, T1o, Tb, T1p, Tc, T2Q;
Chris@82 467 T1 = ri[0];
Chris@82 468 T2R = ii[0];
Chris@82 469 {
Chris@82 470 E T3, T5, T2, T4;
Chris@82 471 T3 = ri[WS(rs, 5)];
Chris@82 472 T5 = ii[WS(rs, 5)];
Chris@82 473 T2 = W[8];
Chris@82 474 T4 = W[9];
Chris@82 475 T6 = FMA(T2, T3, T4 * T5);
Chris@82 476 T1o = FNMS(T4, T3, T2 * T5);
Chris@82 477 }
Chris@82 478 {
Chris@82 479 E T8, Ta, T7, T9;
Chris@82 480 T8 = ri[WS(rs, 10)];
Chris@82 481 Ta = ii[WS(rs, 10)];
Chris@82 482 T7 = W[18];
Chris@82 483 T9 = W[19];
Chris@82 484 Tb = FMA(T7, T8, T9 * Ta);
Chris@82 485 T1p = FNMS(T9, T8, T7 * Ta);
Chris@82 486 }
Chris@82 487 T1q = KP866025403 * (T1o - T1p);
Chris@82 488 T34 = KP866025403 * (Tb - T6);
Chris@82 489 Tc = T6 + Tb;
Chris@82 490 Td = T1 + Tc;
Chris@82 491 T1n = FNMS(KP500000000, Tc, T1);
Chris@82 492 T2Q = T1o + T1p;
Chris@82 493 T2S = T2Q + T2R;
Chris@82 494 T35 = FNMS(KP500000000, T2Q, T2R);
Chris@82 495 }
Chris@82 496 {
Chris@82 497 E TR, T2c, T18, T2h, TW, T1E, T11, T1F, T12, T2d, T1d, T1J, T1i, T1K, T1j;
Chris@82 498 E T2i;
Chris@82 499 {
Chris@82 500 E TO, TQ, TN, TP;
Chris@82 501 TO = ri[WS(rs, 6)];
Chris@82 502 TQ = ii[WS(rs, 6)];
Chris@82 503 TN = W[10];
Chris@82 504 TP = W[11];
Chris@82 505 TR = FMA(TN, TO, TP * TQ);
Chris@82 506 T2c = FNMS(TP, TO, TN * TQ);
Chris@82 507 }
Chris@82 508 {
Chris@82 509 E T15, T17, T14, T16;
Chris@82 510 T15 = ri[WS(rs, 9)];
Chris@82 511 T17 = ii[WS(rs, 9)];
Chris@82 512 T14 = W[16];
Chris@82 513 T16 = W[17];
Chris@82 514 T18 = FMA(T14, T15, T16 * T17);
Chris@82 515 T2h = FNMS(T16, T15, T14 * T17);
Chris@82 516 }
Chris@82 517 {
Chris@82 518 E TT, TV, TS, TU;
Chris@82 519 TT = ri[WS(rs, 11)];
Chris@82 520 TV = ii[WS(rs, 11)];
Chris@82 521 TS = W[20];
Chris@82 522 TU = W[21];
Chris@82 523 TW = FMA(TS, TT, TU * TV);
Chris@82 524 T1E = FNMS(TU, TT, TS * TV);
Chris@82 525 }
Chris@82 526 {
Chris@82 527 E TY, T10, TX, TZ;
Chris@82 528 TY = ri[WS(rs, 1)];
Chris@82 529 T10 = ii[WS(rs, 1)];
Chris@82 530 TX = W[0];
Chris@82 531 TZ = W[1];
Chris@82 532 T11 = FMA(TX, TY, TZ * T10);
Chris@82 533 T1F = FNMS(TZ, TY, TX * T10);
Chris@82 534 }
Chris@82 535 T12 = TW + T11;
Chris@82 536 T2d = T1E + T1F;
Chris@82 537 {
Chris@82 538 E T1a, T1c, T19, T1b;
Chris@82 539 T1a = ri[WS(rs, 14)];
Chris@82 540 T1c = ii[WS(rs, 14)];
Chris@82 541 T19 = W[26];
Chris@82 542 T1b = W[27];
Chris@82 543 T1d = FMA(T19, T1a, T1b * T1c);
Chris@82 544 T1J = FNMS(T1b, T1a, T19 * T1c);
Chris@82 545 }
Chris@82 546 {
Chris@82 547 E T1f, T1h, T1e, T1g;
Chris@82 548 T1f = ri[WS(rs, 4)];
Chris@82 549 T1h = ii[WS(rs, 4)];
Chris@82 550 T1e = W[6];
Chris@82 551 T1g = W[7];
Chris@82 552 T1i = FMA(T1e, T1f, T1g * T1h);
Chris@82 553 T1K = FNMS(T1g, T1f, T1e * T1h);
Chris@82 554 }
Chris@82 555 T1j = T1d + T1i;
Chris@82 556 T2i = T1J + T1K;
Chris@82 557 {
Chris@82 558 E T1D, T1G, T2g, T2j;
Chris@82 559 T13 = TR + T12;
Chris@82 560 T1k = T18 + T1j;
Chris@82 561 T1l = T13 + T1k;
Chris@82 562 T2E = T2c + T2d;
Chris@82 563 T2F = T2h + T2i;
Chris@82 564 T2O = T2E + T2F;
Chris@82 565 T1D = FNMS(KP500000000, T12, TR);
Chris@82 566 T1G = KP866025403 * (T1E - T1F);
Chris@82 567 T1H = T1D - T1G;
Chris@82 568 T1T = T1D + T1G;
Chris@82 569 T2g = KP866025403 * (T1i - T1d);
Chris@82 570 T2j = FNMS(KP500000000, T2i, T2h);
Chris@82 571 T2k = T2g + T2j;
Chris@82 572 T2t = T2j - T2g;
Chris@82 573 {
Chris@82 574 E T2b, T2e, T1I, T1L;
Chris@82 575 T2b = KP866025403 * (T11 - TW);
Chris@82 576 T2e = FNMS(KP500000000, T2d, T2c);
Chris@82 577 T2f = T2b + T2e;
Chris@82 578 T2s = T2e - T2b;
Chris@82 579 T1I = FNMS(KP500000000, T1j, T18);
Chris@82 580 T1L = KP866025403 * (T1J - T1K);
Chris@82 581 T1M = T1I - T1L;
Chris@82 582 T1U = T1I + T1L;
Chris@82 583 }
Chris@82 584 }
Chris@82 585 }
Chris@82 586 {
Chris@82 587 E Ti, T21, Tz, T26, Tn, T1t, Ts, T1u, Tt, T22, TE, T1y, TJ, T1z, TK;
Chris@82 588 E T27;
Chris@82 589 {
Chris@82 590 E Tf, Th, Te, Tg;
Chris@82 591 Tf = ri[WS(rs, 3)];
Chris@82 592 Th = ii[WS(rs, 3)];
Chris@82 593 Te = W[4];
Chris@82 594 Tg = W[5];
Chris@82 595 Ti = FMA(Te, Tf, Tg * Th);
Chris@82 596 T21 = FNMS(Tg, Tf, Te * Th);
Chris@82 597 }
Chris@82 598 {
Chris@82 599 E Tw, Ty, Tv, Tx;
Chris@82 600 Tw = ri[WS(rs, 12)];
Chris@82 601 Ty = ii[WS(rs, 12)];
Chris@82 602 Tv = W[22];
Chris@82 603 Tx = W[23];
Chris@82 604 Tz = FMA(Tv, Tw, Tx * Ty);
Chris@82 605 T26 = FNMS(Tx, Tw, Tv * Ty);
Chris@82 606 }
Chris@82 607 {
Chris@82 608 E Tk, Tm, Tj, Tl;
Chris@82 609 Tk = ri[WS(rs, 8)];
Chris@82 610 Tm = ii[WS(rs, 8)];
Chris@82 611 Tj = W[14];
Chris@82 612 Tl = W[15];
Chris@82 613 Tn = FMA(Tj, Tk, Tl * Tm);
Chris@82 614 T1t = FNMS(Tl, Tk, Tj * Tm);
Chris@82 615 }
Chris@82 616 {
Chris@82 617 E Tp, Tr, To, Tq;
Chris@82 618 Tp = ri[WS(rs, 13)];
Chris@82 619 Tr = ii[WS(rs, 13)];
Chris@82 620 To = W[24];
Chris@82 621 Tq = W[25];
Chris@82 622 Ts = FMA(To, Tp, Tq * Tr);
Chris@82 623 T1u = FNMS(Tq, Tp, To * Tr);
Chris@82 624 }
Chris@82 625 Tt = Tn + Ts;
Chris@82 626 T22 = T1t + T1u;
Chris@82 627 {
Chris@82 628 E TB, TD, TA, TC;
Chris@82 629 TB = ri[WS(rs, 2)];
Chris@82 630 TD = ii[WS(rs, 2)];
Chris@82 631 TA = W[2];
Chris@82 632 TC = W[3];
Chris@82 633 TE = FMA(TA, TB, TC * TD);
Chris@82 634 T1y = FNMS(TC, TB, TA * TD);
Chris@82 635 }
Chris@82 636 {
Chris@82 637 E TG, TI, TF, TH;
Chris@82 638 TG = ri[WS(rs, 7)];
Chris@82 639 TI = ii[WS(rs, 7)];
Chris@82 640 TF = W[12];
Chris@82 641 TH = W[13];
Chris@82 642 TJ = FMA(TF, TG, TH * TI);
Chris@82 643 T1z = FNMS(TH, TG, TF * TI);
Chris@82 644 }
Chris@82 645 TK = TE + TJ;
Chris@82 646 T27 = T1y + T1z;
Chris@82 647 {
Chris@82 648 E T1s, T1v, T25, T28;
Chris@82 649 Tu = Ti + Tt;
Chris@82 650 TL = Tz + TK;
Chris@82 651 TM = Tu + TL;
Chris@82 652 T2H = T21 + T22;
Chris@82 653 T2I = T26 + T27;
Chris@82 654 T2N = T2H + T2I;
Chris@82 655 T1s = FNMS(KP500000000, Tt, Ti);
Chris@82 656 T1v = KP866025403 * (T1t - T1u);
Chris@82 657 T1w = T1s - T1v;
Chris@82 658 T1Q = T1s + T1v;
Chris@82 659 T25 = KP866025403 * (TJ - TE);
Chris@82 660 T28 = FNMS(KP500000000, T27, T26);
Chris@82 661 T29 = T25 + T28;
Chris@82 662 T2w = T28 - T25;
Chris@82 663 {
Chris@82 664 E T20, T23, T1x, T1A;
Chris@82 665 T20 = KP866025403 * (Ts - Tn);
Chris@82 666 T23 = FNMS(KP500000000, T22, T21);
Chris@82 667 T24 = T20 + T23;
Chris@82 668 T2v = T23 - T20;
Chris@82 669 T1x = FNMS(KP500000000, TK, Tz);
Chris@82 670 T1A = KP866025403 * (T1y - T1z);
Chris@82 671 T1B = T1x - T1A;
Chris@82 672 T1R = T1x + T1A;
Chris@82 673 }
Chris@82 674 }
Chris@82 675 }
Chris@82 676 {
Chris@82 677 E T2C, T1m, T2B, T2K, T2M, T2G, T2J, T2L, T2D;
Chris@82 678 T2C = KP559016994 * (TM - T1l);
Chris@82 679 T1m = TM + T1l;
Chris@82 680 T2B = FNMS(KP250000000, T1m, Td);
Chris@82 681 T2G = T2E - T2F;
Chris@82 682 T2J = T2H - T2I;
Chris@82 683 T2K = FNMS(KP587785252, T2J, KP951056516 * T2G);
Chris@82 684 T2M = FMA(KP951056516, T2J, KP587785252 * T2G);
Chris@82 685 ri[0] = Td + T1m;
Chris@82 686 T2L = T2C + T2B;
Chris@82 687 ri[WS(rs, 9)] = T2L - T2M;
Chris@82 688 ri[WS(rs, 6)] = T2L + T2M;
Chris@82 689 T2D = T2B - T2C;
Chris@82 690 ri[WS(rs, 12)] = T2D - T2K;
Chris@82 691 ri[WS(rs, 3)] = T2D + T2K;
Chris@82 692 }
Chris@82 693 {
Chris@82 694 E T2U, T2P, T2T, T2Y, T30, T2W, T2X, T2Z, T2V;
Chris@82 695 T2U = KP559016994 * (T2N - T2O);
Chris@82 696 T2P = T2N + T2O;
Chris@82 697 T2T = FNMS(KP250000000, T2P, T2S);
Chris@82 698 T2W = T13 - T1k;
Chris@82 699 T2X = Tu - TL;
Chris@82 700 T2Y = FNMS(KP587785252, T2X, KP951056516 * T2W);
Chris@82 701 T30 = FMA(KP951056516, T2X, KP587785252 * T2W);
Chris@82 702 ii[0] = T2P + T2S;
Chris@82 703 T2Z = T2U + T2T;
Chris@82 704 ii[WS(rs, 6)] = T2Z - T30;
Chris@82 705 ii[WS(rs, 9)] = T30 + T2Z;
Chris@82 706 T2V = T2T - T2U;
Chris@82 707 ii[WS(rs, 3)] = T2V - T2Y;
Chris@82 708 ii[WS(rs, 12)] = T2Y + T2V;
Chris@82 709 }
Chris@82 710 {
Chris@82 711 E T2y, T2A, T1r, T1O, T2p, T2q, T2z, T2r;
Chris@82 712 {
Chris@82 713 E T2u, T2x, T1C, T1N;
Chris@82 714 T2u = T2s - T2t;
Chris@82 715 T2x = T2v - T2w;
Chris@82 716 T2y = FNMS(KP587785252, T2x, KP951056516 * T2u);
Chris@82 717 T2A = FMA(KP951056516, T2x, KP587785252 * T2u);
Chris@82 718 T1r = T1n - T1q;
Chris@82 719 T1C = T1w + T1B;
Chris@82 720 T1N = T1H + T1M;
Chris@82 721 T1O = T1C + T1N;
Chris@82 722 T2p = FNMS(KP250000000, T1O, T1r);
Chris@82 723 T2q = KP559016994 * (T1C - T1N);
Chris@82 724 }
Chris@82 725 ri[WS(rs, 5)] = T1r + T1O;
Chris@82 726 T2z = T2q + T2p;
Chris@82 727 ri[WS(rs, 14)] = T2z - T2A;
Chris@82 728 ri[WS(rs, 11)] = T2z + T2A;
Chris@82 729 T2r = T2p - T2q;
Chris@82 730 ri[WS(rs, 2)] = T2r - T2y;
Chris@82 731 ri[WS(rs, 8)] = T2r + T2y;
Chris@82 732 }
Chris@82 733 {
Chris@82 734 E T3h, T3q, T3i, T3l, T3m, T3n, T3p, T3o;
Chris@82 735 {
Chris@82 736 E T3f, T3g, T3j, T3k;
Chris@82 737 T3f = T1H - T1M;
Chris@82 738 T3g = T1w - T1B;
Chris@82 739 T3h = FNMS(KP587785252, T3g, KP951056516 * T3f);
Chris@82 740 T3q = FMA(KP951056516, T3g, KP587785252 * T3f);
Chris@82 741 T3i = T35 - T34;
Chris@82 742 T3j = T2v + T2w;
Chris@82 743 T3k = T2s + T2t;
Chris@82 744 T3l = T3j + T3k;
Chris@82 745 T3m = FNMS(KP250000000, T3l, T3i);
Chris@82 746 T3n = KP559016994 * (T3j - T3k);
Chris@82 747 }
Chris@82 748 ii[WS(rs, 5)] = T3l + T3i;
Chris@82 749 T3p = T3n + T3m;
Chris@82 750 ii[WS(rs, 11)] = T3p - T3q;
Chris@82 751 ii[WS(rs, 14)] = T3q + T3p;
Chris@82 752 T3o = T3m - T3n;
Chris@82 753 ii[WS(rs, 2)] = T3h + T3o;
Chris@82 754 ii[WS(rs, 8)] = T3o - T3h;
Chris@82 755 }
Chris@82 756 {
Chris@82 757 E T3c, T3d, T36, T37, T33, T38, T3e, T39;
Chris@82 758 {
Chris@82 759 E T3a, T3b, T31, T32;
Chris@82 760 T3a = T1Q - T1R;
Chris@82 761 T3b = T1T - T1U;
Chris@82 762 T3c = FMA(KP951056516, T3a, KP587785252 * T3b);
Chris@82 763 T3d = FNMS(KP587785252, T3a, KP951056516 * T3b);
Chris@82 764 T36 = T34 + T35;
Chris@82 765 T31 = T24 + T29;
Chris@82 766 T32 = T2f + T2k;
Chris@82 767 T37 = T31 + T32;
Chris@82 768 T33 = KP559016994 * (T31 - T32);
Chris@82 769 T38 = FNMS(KP250000000, T37, T36);
Chris@82 770 }
Chris@82 771 ii[WS(rs, 10)] = T37 + T36;
Chris@82 772 T3e = T38 - T33;
Chris@82 773 ii[WS(rs, 7)] = T3d + T3e;
Chris@82 774 ii[WS(rs, 13)] = T3e - T3d;
Chris@82 775 T39 = T33 + T38;
Chris@82 776 ii[WS(rs, 1)] = T39 - T3c;
Chris@82 777 ii[WS(rs, 4)] = T3c + T39;
Chris@82 778 }
Chris@82 779 {
Chris@82 780 E T2m, T2o, T1P, T1W, T1X, T1Y, T2n, T1Z;
Chris@82 781 {
Chris@82 782 E T2a, T2l, T1S, T1V;
Chris@82 783 T2a = T24 - T29;
Chris@82 784 T2l = T2f - T2k;
Chris@82 785 T2m = FMA(KP951056516, T2a, KP587785252 * T2l);
Chris@82 786 T2o = FNMS(KP587785252, T2a, KP951056516 * T2l);
Chris@82 787 T1P = T1n + T1q;
Chris@82 788 T1S = T1Q + T1R;
Chris@82 789 T1V = T1T + T1U;
Chris@82 790 T1W = T1S + T1V;
Chris@82 791 T1X = KP559016994 * (T1S - T1V);
Chris@82 792 T1Y = FNMS(KP250000000, T1W, T1P);
Chris@82 793 }
Chris@82 794 ri[WS(rs, 10)] = T1P + T1W;
Chris@82 795 T2n = T1Y - T1X;
Chris@82 796 ri[WS(rs, 7)] = T2n - T2o;
Chris@82 797 ri[WS(rs, 13)] = T2n + T2o;
Chris@82 798 T1Z = T1X + T1Y;
Chris@82 799 ri[WS(rs, 4)] = T1Z - T2m;
Chris@82 800 ri[WS(rs, 1)] = T1Z + T2m;
Chris@82 801 }
Chris@82 802 }
Chris@82 803 }
Chris@82 804 }
Chris@82 805
Chris@82 806 static const tw_instr twinstr[] = {
Chris@82 807 {TW_FULL, 0, 15},
Chris@82 808 {TW_NEXT, 1, 0}
Chris@82 809 };
Chris@82 810
Chris@82 811 static const ct_desc desc = { 15, "t1_15", twinstr, &GENUS, {128, 56, 56, 0}, 0, 0, 0 };
Chris@82 812
Chris@82 813 void X(codelet_t1_15) (planner *p) {
Chris@82 814 X(kdft_dit_register) (p, t1_15, &desc);
Chris@82 815 }
Chris@82 816 #endif