Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.5/dft/scalar/codelets/q1_5.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 16:37:57 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_twidsq.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -reload-twiddle -dif -n 5 -name q1_5 -include q.h */
Chris@42 29
Chris@42 30 /*
Chris@42 31 * This function contains 200 FP additions, 170 FP multiplications,
Chris@42 32 * (or, 70 additions, 40 multiplications, 130 fused multiply/add),
Chris@42 33 * 104 stack variables, 4 constants, and 100 memory accesses
Chris@42 34 */
Chris@42 35 #include "q.h"
Chris@42 36
Chris@42 37 static void q1_5(R *rio, R *iio, const R *W, stride rs, stride vs, 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 * 8); m < me; m = m + 1, rio = rio + ms, iio = iio + ms, W = W + 8, MAKE_VOLATILE_STRIDE(10, rs), MAKE_VOLATILE_STRIDE(0, vs)) {
Chris@42 46 E T1x, T1w, T1v;
Chris@42 47 {
Chris@42 48 E T1, Tn, TM, Tw, Tb, T8, Ta, TV, Tq, Ts, TH, Tj, Tr, T1h, T1q;
Chris@42 49 E T1G, T12, T15, T1P, T14, T1k, T1m, T1B, T1d, T1l, T2b, T2k, T2A, T1W, T1Z;
Chris@42 50 E T3Z, T1Y, T2e, T2g, T2v, T27, T2f, T3D, T42, T44, T4j, T3V, T43, T2J, T48;
Chris@42 51 E T4o, T3K, T3N, T35, T3M, T2V, T3e, T3u, T2Q, T2T, T37, T30, T2S, T2W;
Chris@42 52 {
Chris@42 53 E T1Q, T2j, T1V, T1R;
Chris@42 54 {
Chris@42 55 E Tp, Ti, Td, Te;
Chris@42 56 {
Chris@42 57 E T5, T6, T2, T3, T7, Tv;
Chris@42 58 T1 = rio[0];
Chris@42 59 T5 = rio[WS(rs, 2)];
Chris@42 60 T6 = rio[WS(rs, 3)];
Chris@42 61 T2 = rio[WS(rs, 1)];
Chris@42 62 T3 = rio[WS(rs, 4)];
Chris@42 63 Tn = iio[0];
Chris@42 64 T7 = T5 + T6;
Chris@42 65 Tv = T5 - T6;
Chris@42 66 {
Chris@42 67 E T4, Tu, Tg, Th;
Chris@42 68 T4 = T2 + T3;
Chris@42 69 Tu = T2 - T3;
Chris@42 70 Tg = iio[WS(rs, 2)];
Chris@42 71 Th = iio[WS(rs, 3)];
Chris@42 72 TM = FNMS(KP618033988, Tu, Tv);
Chris@42 73 Tw = FMA(KP618033988, Tv, Tu);
Chris@42 74 Tb = T4 - T7;
Chris@42 75 T8 = T4 + T7;
Chris@42 76 Tp = Tg + Th;
Chris@42 77 Ti = Tg - Th;
Chris@42 78 Ta = FNMS(KP250000000, T8, T1);
Chris@42 79 Td = iio[WS(rs, 1)];
Chris@42 80 Te = iio[WS(rs, 4)];
Chris@42 81 }
Chris@42 82 }
Chris@42 83 {
Chris@42 84 E TW, T1p, T11, TX;
Chris@42 85 TV = rio[WS(vs, 1)];
Chris@42 86 {
Chris@42 87 E TZ, T10, Tf, To;
Chris@42 88 TZ = rio[WS(vs, 1) + WS(rs, 2)];
Chris@42 89 T10 = rio[WS(vs, 1) + WS(rs, 3)];
Chris@42 90 Tf = Td - Te;
Chris@42 91 To = Td + Te;
Chris@42 92 TW = rio[WS(vs, 1) + WS(rs, 1)];
Chris@42 93 T1p = TZ - T10;
Chris@42 94 T11 = TZ + T10;
Chris@42 95 Tq = To + Tp;
Chris@42 96 Ts = To - Tp;
Chris@42 97 TH = FNMS(KP618033988, Tf, Ti);
Chris@42 98 Tj = FMA(KP618033988, Ti, Tf);
Chris@42 99 Tr = FNMS(KP250000000, Tq, Tn);
Chris@42 100 TX = rio[WS(vs, 1) + WS(rs, 4)];
Chris@42 101 }
Chris@42 102 {
Chris@42 103 E T17, T1j, T1c, T18;
Chris@42 104 T1h = iio[WS(vs, 1)];
Chris@42 105 {
Chris@42 106 E T1a, T1b, TY, T1o;
Chris@42 107 T1a = iio[WS(vs, 1) + WS(rs, 2)];
Chris@42 108 T1b = iio[WS(vs, 1) + WS(rs, 3)];
Chris@42 109 TY = TW + TX;
Chris@42 110 T1o = TW - TX;
Chris@42 111 T17 = iio[WS(vs, 1) + WS(rs, 1)];
Chris@42 112 T1j = T1a + T1b;
Chris@42 113 T1c = T1a - T1b;
Chris@42 114 T1q = FMA(KP618033988, T1p, T1o);
Chris@42 115 T1G = FNMS(KP618033988, T1o, T1p);
Chris@42 116 T12 = TY + T11;
Chris@42 117 T15 = TY - T11;
Chris@42 118 T18 = iio[WS(vs, 1) + WS(rs, 4)];
Chris@42 119 }
Chris@42 120 T1P = rio[WS(vs, 2)];
Chris@42 121 T14 = FNMS(KP250000000, T12, TV);
Chris@42 122 {
Chris@42 123 E T1T, T1i, T19, T1U;
Chris@42 124 T1T = rio[WS(vs, 2) + WS(rs, 2)];
Chris@42 125 T1i = T17 + T18;
Chris@42 126 T19 = T17 - T18;
Chris@42 127 T1U = rio[WS(vs, 2) + WS(rs, 3)];
Chris@42 128 T1Q = rio[WS(vs, 2) + WS(rs, 1)];
Chris@42 129 T1k = T1i + T1j;
Chris@42 130 T1m = T1i - T1j;
Chris@42 131 T1B = FNMS(KP618033988, T19, T1c);
Chris@42 132 T1d = FMA(KP618033988, T1c, T19);
Chris@42 133 T2j = T1T - T1U;
Chris@42 134 T1V = T1T + T1U;
Chris@42 135 T1l = FNMS(KP250000000, T1k, T1h);
Chris@42 136 T1R = rio[WS(vs, 2) + WS(rs, 4)];
Chris@42 137 }
Chris@42 138 }
Chris@42 139 }
Chris@42 140 }
Chris@42 141 {
Chris@42 142 E T3P, T41, T3U, T3Q;
Chris@42 143 {
Chris@42 144 E T21, T2d, T26, T22;
Chris@42 145 T2b = iio[WS(vs, 2)];
Chris@42 146 {
Chris@42 147 E T24, T25, T1S, T2i;
Chris@42 148 T24 = iio[WS(vs, 2) + WS(rs, 2)];
Chris@42 149 T25 = iio[WS(vs, 2) + WS(rs, 3)];
Chris@42 150 T1S = T1Q + T1R;
Chris@42 151 T2i = T1Q - T1R;
Chris@42 152 T21 = iio[WS(vs, 2) + WS(rs, 1)];
Chris@42 153 T2d = T24 + T25;
Chris@42 154 T26 = T24 - T25;
Chris@42 155 T2k = FMA(KP618033988, T2j, T2i);
Chris@42 156 T2A = FNMS(KP618033988, T2i, T2j);
Chris@42 157 T1W = T1S + T1V;
Chris@42 158 T1Z = T1S - T1V;
Chris@42 159 T22 = iio[WS(vs, 2) + WS(rs, 4)];
Chris@42 160 }
Chris@42 161 T3Z = iio[WS(vs, 4)];
Chris@42 162 T1Y = FNMS(KP250000000, T1W, T1P);
Chris@42 163 {
Chris@42 164 E T3S, T2c, T23, T3T;
Chris@42 165 T3S = iio[WS(vs, 4) + WS(rs, 2)];
Chris@42 166 T2c = T21 + T22;
Chris@42 167 T23 = T21 - T22;
Chris@42 168 T3T = iio[WS(vs, 4) + WS(rs, 3)];
Chris@42 169 T3P = iio[WS(vs, 4) + WS(rs, 1)];
Chris@42 170 T2e = T2c + T2d;
Chris@42 171 T2g = T2c - T2d;
Chris@42 172 T2v = FNMS(KP618033988, T23, T26);
Chris@42 173 T27 = FMA(KP618033988, T26, T23);
Chris@42 174 T41 = T3S + T3T;
Chris@42 175 T3U = T3S - T3T;
Chris@42 176 T2f = FNMS(KP250000000, T2e, T2b);
Chris@42 177 T3Q = iio[WS(vs, 4) + WS(rs, 4)];
Chris@42 178 }
Chris@42 179 }
Chris@42 180 {
Chris@42 181 E T3E, T47, T3J, T3F;
Chris@42 182 T3D = rio[WS(vs, 4)];
Chris@42 183 {
Chris@42 184 E T3H, T3I, T3R, T40;
Chris@42 185 T3H = rio[WS(vs, 4) + WS(rs, 2)];
Chris@42 186 T3I = rio[WS(vs, 4) + WS(rs, 3)];
Chris@42 187 T3R = T3P - T3Q;
Chris@42 188 T40 = T3P + T3Q;
Chris@42 189 T3E = rio[WS(vs, 4) + WS(rs, 1)];
Chris@42 190 T47 = T3H - T3I;
Chris@42 191 T3J = T3H + T3I;
Chris@42 192 T42 = T40 + T41;
Chris@42 193 T44 = T40 - T41;
Chris@42 194 T4j = FNMS(KP618033988, T3R, T3U);
Chris@42 195 T3V = FMA(KP618033988, T3U, T3R);
Chris@42 196 T43 = FNMS(KP250000000, T42, T3Z);
Chris@42 197 T3F = rio[WS(vs, 4) + WS(rs, 4)];
Chris@42 198 }
Chris@42 199 {
Chris@42 200 E T2K, T3d, T2P, T2L;
Chris@42 201 T2J = rio[WS(vs, 3)];
Chris@42 202 {
Chris@42 203 E T2N, T2O, T3G, T46;
Chris@42 204 T2N = rio[WS(vs, 3) + WS(rs, 2)];
Chris@42 205 T2O = rio[WS(vs, 3) + WS(rs, 3)];
Chris@42 206 T3G = T3E + T3F;
Chris@42 207 T46 = T3E - T3F;
Chris@42 208 T2K = rio[WS(vs, 3) + WS(rs, 1)];
Chris@42 209 T3d = T2N - T2O;
Chris@42 210 T2P = T2N + T2O;
Chris@42 211 T48 = FMA(KP618033988, T47, T46);
Chris@42 212 T4o = FNMS(KP618033988, T46, T47);
Chris@42 213 T3K = T3G + T3J;
Chris@42 214 T3N = T3G - T3J;
Chris@42 215 T2L = rio[WS(vs, 3) + WS(rs, 4)];
Chris@42 216 }
Chris@42 217 T35 = iio[WS(vs, 3)];
Chris@42 218 T3M = FNMS(KP250000000, T3K, T3D);
Chris@42 219 {
Chris@42 220 E T2Y, T3c, T2M, T2Z;
Chris@42 221 T2Y = iio[WS(vs, 3) + WS(rs, 2)];
Chris@42 222 T3c = T2K - T2L;
Chris@42 223 T2M = T2K + T2L;
Chris@42 224 T2Z = iio[WS(vs, 3) + WS(rs, 3)];
Chris@42 225 T2V = iio[WS(vs, 3) + WS(rs, 1)];
Chris@42 226 T3e = FMA(KP618033988, T3d, T3c);
Chris@42 227 T3u = FNMS(KP618033988, T3c, T3d);
Chris@42 228 T2Q = T2M + T2P;
Chris@42 229 T2T = T2M - T2P;
Chris@42 230 T37 = T2Y + T2Z;
Chris@42 231 T30 = T2Y - T2Z;
Chris@42 232 T2S = FNMS(KP250000000, T2Q, T2J);
Chris@42 233 T2W = iio[WS(vs, 3) + WS(rs, 4)];
Chris@42 234 }
Chris@42 235 }
Chris@42 236 }
Chris@42 237 }
Chris@42 238 }
Chris@42 239 {
Chris@42 240 E T3a, T31, T3p, T39, T2X, T36, T38;
Chris@42 241 rio[0] = T1 + T8;
Chris@42 242 iio[0] = Tn + Tq;
Chris@42 243 rio[WS(rs, 1)] = TV + T12;
Chris@42 244 T2X = T2V - T2W;
Chris@42 245 T36 = T2V + T2W;
Chris@42 246 iio[WS(rs, 1)] = T1h + T1k;
Chris@42 247 rio[WS(rs, 2)] = T1P + T1W;
Chris@42 248 T3a = T36 - T37;
Chris@42 249 T38 = T36 + T37;
Chris@42 250 T31 = FMA(KP618033988, T30, T2X);
Chris@42 251 T3p = FNMS(KP618033988, T2X, T30);
Chris@42 252 T39 = FNMS(KP250000000, T38, T35);
Chris@42 253 iio[WS(rs, 2)] = T2b + T2e;
Chris@42 254 iio[WS(rs, 4)] = T3Z + T42;
Chris@42 255 rio[WS(rs, 4)] = T3D + T3K;
Chris@42 256 rio[WS(rs, 3)] = T2J + T2Q;
Chris@42 257 iio[WS(rs, 3)] = T35 + T38;
Chris@42 258 {
Chris@42 259 E T3O, T45, T2r, T2q, T2p, TT, TS, TR;
Chris@42 260 {
Chris@42 261 E TG, TL, TD, TC, TB, Tc, Tt;
Chris@42 262 TG = FNMS(KP559016994, Tb, Ta);
Chris@42 263 Tc = FMA(KP559016994, Tb, Ta);
Chris@42 264 Tt = FMA(KP559016994, Ts, Tr);
Chris@42 265 TL = FNMS(KP559016994, Ts, Tr);
Chris@42 266 {
Chris@42 267 E T9, Tm, Tk, TA, Tx;
Chris@42 268 T9 = W[0];
Chris@42 269 Tm = W[1];
Chris@42 270 Tk = FMA(KP951056516, Tj, Tc);
Chris@42 271 TA = FNMS(KP951056516, Tj, Tc);
Chris@42 272 Tx = FNMS(KP951056516, Tw, Tt);
Chris@42 273 TD = FMA(KP951056516, Tw, Tt);
Chris@42 274 {
Chris@42 275 E Tz, Tl, Ty, TE;
Chris@42 276 Tz = W[6];
Chris@42 277 Tl = T9 * Tk;
Chris@42 278 TC = W[7];
Chris@42 279 Ty = T9 * Tx;
Chris@42 280 TE = Tz * TD;
Chris@42 281 TB = Tz * TA;
Chris@42 282 rio[WS(vs, 1)] = FMA(Tm, Tx, Tl);
Chris@42 283 iio[WS(vs, 1)] = FNMS(Tm, Tk, Ty);
Chris@42 284 iio[WS(vs, 4)] = FNMS(TC, TA, TE);
Chris@42 285 }
Chris@42 286 }
Chris@42 287 rio[WS(vs, 4)] = FMA(TC, TD, TB);
Chris@42 288 {
Chris@42 289 E TF, TK, TI, TQ, TN;
Chris@42 290 TF = W[2];
Chris@42 291 TK = W[3];
Chris@42 292 TI = FNMS(KP951056516, TH, TG);
Chris@42 293 TQ = FMA(KP951056516, TH, TG);
Chris@42 294 TN = FMA(KP951056516, TM, TL);
Chris@42 295 TT = FNMS(KP951056516, TM, TL);
Chris@42 296 {
Chris@42 297 E TP, TJ, TO, TU;
Chris@42 298 TP = W[4];
Chris@42 299 TJ = TF * TI;
Chris@42 300 TS = W[5];
Chris@42 301 TO = TF * TN;
Chris@42 302 TU = TP * TT;
Chris@42 303 TR = TP * TQ;
Chris@42 304 rio[WS(vs, 2)] = FMA(TK, TN, TJ);
Chris@42 305 iio[WS(vs, 2)] = FNMS(TK, TI, TO);
Chris@42 306 iio[WS(vs, 3)] = FNMS(TS, TQ, TU);
Chris@42 307 }
Chris@42 308 }
Chris@42 309 }
Chris@42 310 rio[WS(vs, 3)] = FMA(TS, TT, TR);
Chris@42 311 {
Chris@42 312 E T20, T2h, T2H, T2G, T2F, T2u, T2z;
Chris@42 313 T20 = FMA(KP559016994, T1Z, T1Y);
Chris@42 314 T2u = FNMS(KP559016994, T1Z, T1Y);
Chris@42 315 T2z = FNMS(KP559016994, T2g, T2f);
Chris@42 316 T2h = FMA(KP559016994, T2g, T2f);
Chris@42 317 {
Chris@42 318 E T2t, T2y, T2w, T2E, T2B;
Chris@42 319 T2t = W[2];
Chris@42 320 T2y = W[3];
Chris@42 321 T2w = FNMS(KP951056516, T2v, T2u);
Chris@42 322 T2E = FMA(KP951056516, T2v, T2u);
Chris@42 323 T2B = FMA(KP951056516, T2A, T2z);
Chris@42 324 T2H = FNMS(KP951056516, T2A, T2z);
Chris@42 325 {
Chris@42 326 E T2D, T2x, T2C, T2I;
Chris@42 327 T2D = W[4];
Chris@42 328 T2x = T2t * T2w;
Chris@42 329 T2G = W[5];
Chris@42 330 T2C = T2t * T2B;
Chris@42 331 T2I = T2D * T2H;
Chris@42 332 T2F = T2D * T2E;
Chris@42 333 rio[WS(vs, 2) + WS(rs, 2)] = FMA(T2y, T2B, T2x);
Chris@42 334 iio[WS(vs, 2) + WS(rs, 2)] = FNMS(T2y, T2w, T2C);
Chris@42 335 iio[WS(vs, 3) + WS(rs, 2)] = FNMS(T2G, T2E, T2I);
Chris@42 336 }
Chris@42 337 }
Chris@42 338 rio[WS(vs, 3) + WS(rs, 2)] = FMA(T2G, T2H, T2F);
Chris@42 339 {
Chris@42 340 E T4v, T4u, T4t, T4i, T4n;
Chris@42 341 T3O = FMA(KP559016994, T3N, T3M);
Chris@42 342 T4i = FNMS(KP559016994, T3N, T3M);
Chris@42 343 T4n = FNMS(KP559016994, T44, T43);
Chris@42 344 T45 = FMA(KP559016994, T44, T43);
Chris@42 345 {
Chris@42 346 E T4h, T4m, T4k, T4s, T4p;
Chris@42 347 T4h = W[2];
Chris@42 348 T4m = W[3];
Chris@42 349 T4k = FNMS(KP951056516, T4j, T4i);
Chris@42 350 T4s = FMA(KP951056516, T4j, T4i);
Chris@42 351 T4p = FMA(KP951056516, T4o, T4n);
Chris@42 352 T4v = FNMS(KP951056516, T4o, T4n);
Chris@42 353 {
Chris@42 354 E T4r, T4l, T4q, T4w;
Chris@42 355 T4r = W[4];
Chris@42 356 T4l = T4h * T4k;
Chris@42 357 T4u = W[5];
Chris@42 358 T4q = T4h * T4p;
Chris@42 359 T4w = T4r * T4v;
Chris@42 360 T4t = T4r * T4s;
Chris@42 361 rio[WS(vs, 2) + WS(rs, 4)] = FMA(T4m, T4p, T4l);
Chris@42 362 iio[WS(vs, 2) + WS(rs, 4)] = FNMS(T4m, T4k, T4q);
Chris@42 363 iio[WS(vs, 3) + WS(rs, 4)] = FNMS(T4u, T4s, T4w);
Chris@42 364 }
Chris@42 365 }
Chris@42 366 rio[WS(vs, 3) + WS(rs, 4)] = FMA(T4u, T4v, T4t);
Chris@42 367 {
Chris@42 368 E T1X, T2a, T28, T2o, T2l;
Chris@42 369 T1X = W[0];
Chris@42 370 T2a = W[1];
Chris@42 371 T28 = FMA(KP951056516, T27, T20);
Chris@42 372 T2o = FNMS(KP951056516, T27, T20);
Chris@42 373 T2l = FNMS(KP951056516, T2k, T2h);
Chris@42 374 T2r = FMA(KP951056516, T2k, T2h);
Chris@42 375 {
Chris@42 376 E T2n, T29, T2m, T2s;
Chris@42 377 T2n = W[6];
Chris@42 378 T29 = T1X * T28;
Chris@42 379 T2q = W[7];
Chris@42 380 T2m = T1X * T2l;
Chris@42 381 T2s = T2n * T2r;
Chris@42 382 T2p = T2n * T2o;
Chris@42 383 rio[WS(vs, 1) + WS(rs, 2)] = FMA(T2a, T2l, T29);
Chris@42 384 iio[WS(vs, 1) + WS(rs, 2)] = FNMS(T2a, T28, T2m);
Chris@42 385 iio[WS(vs, 4) + WS(rs, 2)] = FNMS(T2q, T2o, T2s);
Chris@42 386 }
Chris@42 387 }
Chris@42 388 }
Chris@42 389 }
Chris@42 390 rio[WS(vs, 4) + WS(rs, 2)] = FMA(T2q, T2r, T2p);
Chris@42 391 {
Chris@42 392 E T3B, T3A, T3z, T4f, T4e, T4d;
Chris@42 393 {
Chris@42 394 E T3o, T3t, T3l, T3k, T3j, T2U, T3b;
Chris@42 395 T3o = FNMS(KP559016994, T2T, T2S);
Chris@42 396 T2U = FMA(KP559016994, T2T, T2S);
Chris@42 397 T3b = FMA(KP559016994, T3a, T39);
Chris@42 398 T3t = FNMS(KP559016994, T3a, T39);
Chris@42 399 {
Chris@42 400 E T2R, T34, T32, T3i, T3f;
Chris@42 401 T2R = W[0];
Chris@42 402 T34 = W[1];
Chris@42 403 T32 = FMA(KP951056516, T31, T2U);
Chris@42 404 T3i = FNMS(KP951056516, T31, T2U);
Chris@42 405 T3f = FNMS(KP951056516, T3e, T3b);
Chris@42 406 T3l = FMA(KP951056516, T3e, T3b);
Chris@42 407 {
Chris@42 408 E T3h, T33, T3g, T3m;
Chris@42 409 T3h = W[6];
Chris@42 410 T33 = T2R * T32;
Chris@42 411 T3k = W[7];
Chris@42 412 T3g = T2R * T3f;
Chris@42 413 T3m = T3h * T3l;
Chris@42 414 T3j = T3h * T3i;
Chris@42 415 rio[WS(vs, 1) + WS(rs, 3)] = FMA(T34, T3f, T33);
Chris@42 416 iio[WS(vs, 1) + WS(rs, 3)] = FNMS(T34, T32, T3g);
Chris@42 417 iio[WS(vs, 4) + WS(rs, 3)] = FNMS(T3k, T3i, T3m);
Chris@42 418 }
Chris@42 419 }
Chris@42 420 rio[WS(vs, 4) + WS(rs, 3)] = FMA(T3k, T3l, T3j);
Chris@42 421 {
Chris@42 422 E T3n, T3s, T3q, T3y, T3v;
Chris@42 423 T3n = W[2];
Chris@42 424 T3s = W[3];
Chris@42 425 T3q = FNMS(KP951056516, T3p, T3o);
Chris@42 426 T3y = FMA(KP951056516, T3p, T3o);
Chris@42 427 T3v = FMA(KP951056516, T3u, T3t);
Chris@42 428 T3B = FNMS(KP951056516, T3u, T3t);
Chris@42 429 {
Chris@42 430 E T3x, T3r, T3w, T3C;
Chris@42 431 T3x = W[4];
Chris@42 432 T3r = T3n * T3q;
Chris@42 433 T3A = W[5];
Chris@42 434 T3w = T3n * T3v;
Chris@42 435 T3C = T3x * T3B;
Chris@42 436 T3z = T3x * T3y;
Chris@42 437 rio[WS(vs, 2) + WS(rs, 3)] = FMA(T3s, T3v, T3r);
Chris@42 438 iio[WS(vs, 2) + WS(rs, 3)] = FNMS(T3s, T3q, T3w);
Chris@42 439 iio[WS(vs, 3) + WS(rs, 3)] = FNMS(T3A, T3y, T3C);
Chris@42 440 }
Chris@42 441 }
Chris@42 442 }
Chris@42 443 rio[WS(vs, 3) + WS(rs, 3)] = FMA(T3A, T3B, T3z);
Chris@42 444 {
Chris@42 445 E T3L, T3Y, T3W, T4c, T49;
Chris@42 446 T3L = W[0];
Chris@42 447 T3Y = W[1];
Chris@42 448 T3W = FMA(KP951056516, T3V, T3O);
Chris@42 449 T4c = FNMS(KP951056516, T3V, T3O);
Chris@42 450 T49 = FNMS(KP951056516, T48, T45);
Chris@42 451 T4f = FMA(KP951056516, T48, T45);
Chris@42 452 {
Chris@42 453 E T4b, T3X, T4a, T4g;
Chris@42 454 T4b = W[6];
Chris@42 455 T3X = T3L * T3W;
Chris@42 456 T4e = W[7];
Chris@42 457 T4a = T3L * T49;
Chris@42 458 T4g = T4b * T4f;
Chris@42 459 T4d = T4b * T4c;
Chris@42 460 rio[WS(vs, 1) + WS(rs, 4)] = FMA(T3Y, T49, T3X);
Chris@42 461 iio[WS(vs, 1) + WS(rs, 4)] = FNMS(T3Y, T3W, T4a);
Chris@42 462 iio[WS(vs, 4) + WS(rs, 4)] = FNMS(T4e, T4c, T4g);
Chris@42 463 }
Chris@42 464 }
Chris@42 465 rio[WS(vs, 4) + WS(rs, 4)] = FMA(T4e, T4f, T4d);
Chris@42 466 {
Chris@42 467 E T16, T1n, T1N, T1M, T1L, T1A, T1F;
Chris@42 468 T16 = FMA(KP559016994, T15, T14);
Chris@42 469 T1A = FNMS(KP559016994, T15, T14);
Chris@42 470 T1F = FNMS(KP559016994, T1m, T1l);
Chris@42 471 T1n = FMA(KP559016994, T1m, T1l);
Chris@42 472 {
Chris@42 473 E T1z, T1E, T1C, T1K, T1H;
Chris@42 474 T1z = W[2];
Chris@42 475 T1E = W[3];
Chris@42 476 T1C = FNMS(KP951056516, T1B, T1A);
Chris@42 477 T1K = FMA(KP951056516, T1B, T1A);
Chris@42 478 T1H = FMA(KP951056516, T1G, T1F);
Chris@42 479 T1N = FNMS(KP951056516, T1G, T1F);
Chris@42 480 {
Chris@42 481 E T1J, T1D, T1I, T1O;
Chris@42 482 T1J = W[4];
Chris@42 483 T1D = T1z * T1C;
Chris@42 484 T1M = W[5];
Chris@42 485 T1I = T1z * T1H;
Chris@42 486 T1O = T1J * T1N;
Chris@42 487 T1L = T1J * T1K;
Chris@42 488 rio[WS(vs, 2) + WS(rs, 1)] = FMA(T1E, T1H, T1D);
Chris@42 489 iio[WS(vs, 2) + WS(rs, 1)] = FNMS(T1E, T1C, T1I);
Chris@42 490 iio[WS(vs, 3) + WS(rs, 1)] = FNMS(T1M, T1K, T1O);
Chris@42 491 }
Chris@42 492 }
Chris@42 493 rio[WS(vs, 3) + WS(rs, 1)] = FMA(T1M, T1N, T1L);
Chris@42 494 {
Chris@42 495 E T13, T1g, T1e, T1u, T1r;
Chris@42 496 T13 = W[0];
Chris@42 497 T1g = W[1];
Chris@42 498 T1e = FMA(KP951056516, T1d, T16);
Chris@42 499 T1u = FNMS(KP951056516, T1d, T16);
Chris@42 500 T1r = FNMS(KP951056516, T1q, T1n);
Chris@42 501 T1x = FMA(KP951056516, T1q, T1n);
Chris@42 502 {
Chris@42 503 E T1t, T1f, T1s, T1y;
Chris@42 504 T1t = W[6];
Chris@42 505 T1f = T13 * T1e;
Chris@42 506 T1w = W[7];
Chris@42 507 T1s = T13 * T1r;
Chris@42 508 T1y = T1t * T1x;
Chris@42 509 T1v = T1t * T1u;
Chris@42 510 rio[WS(vs, 1) + WS(rs, 1)] = FMA(T1g, T1r, T1f);
Chris@42 511 iio[WS(vs, 1) + WS(rs, 1)] = FNMS(T1g, T1e, T1s);
Chris@42 512 iio[WS(vs, 4) + WS(rs, 1)] = FNMS(T1w, T1u, T1y);
Chris@42 513 }
Chris@42 514 }
Chris@42 515 }
Chris@42 516 }
Chris@42 517 }
Chris@42 518 }
Chris@42 519 }
Chris@42 520 rio[WS(vs, 4) + WS(rs, 1)] = FMA(T1w, T1x, T1v);
Chris@42 521 }
Chris@42 522 }
Chris@42 523 }
Chris@42 524
Chris@42 525 static const tw_instr twinstr[] = {
Chris@42 526 {TW_FULL, 0, 5},
Chris@42 527 {TW_NEXT, 1, 0}
Chris@42 528 };
Chris@42 529
Chris@42 530 static const ct_desc desc = { 5, "q1_5", twinstr, &GENUS, {70, 40, 130, 0}, 0, 0, 0 };
Chris@42 531
Chris@42 532 void X(codelet_q1_5) (planner *p) {
Chris@42 533 X(kdft_difsq_register) (p, q1_5, &desc);
Chris@42 534 }
Chris@42 535 #else /* HAVE_FMA */
Chris@42 536
Chris@42 537 /* Generated by: ../../../genfft/gen_twidsq.native -compact -variables 4 -pipeline-latency 4 -reload-twiddle -dif -n 5 -name q1_5 -include q.h */
Chris@42 538
Chris@42 539 /*
Chris@42 540 * This function contains 200 FP additions, 140 FP multiplications,
Chris@42 541 * (or, 130 additions, 70 multiplications, 70 fused multiply/add),
Chris@42 542 * 75 stack variables, 4 constants, and 100 memory accesses
Chris@42 543 */
Chris@42 544 #include "q.h"
Chris@42 545
Chris@42 546 static void q1_5(R *rio, R *iio, const R *W, stride rs, stride vs, INT mb, INT me, INT ms)
Chris@42 547 {
Chris@42 548 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
Chris@42 549 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
Chris@42 550 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
Chris@42 551 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
Chris@42 552 {
Chris@42 553 INT m;
Chris@42 554 for (m = mb, W = W + (mb * 8); m < me; m = m + 1, rio = rio + ms, iio = iio + ms, W = W + 8, MAKE_VOLATILE_STRIDE(10, rs), MAKE_VOLATILE_STRIDE(0, vs)) {
Chris@42 555 E T1, Ta, TG, Tv, T8, Tb, Tp, Tj, TD, To, Tq, Tr, TN, TW, T1s;
Chris@42 556 E T1h, TU, TX, T1b, T15, T1p, T1a, T1c, T1d, T1z, T1I, T2e, T23, T1G, T1J;
Chris@42 557 E T1X, T1R, T2b, T1W, T1Y, T1Z, T3v, T3p, T3J, T3u, T3w, T3x, T37, T3g, T3M;
Chris@42 558 E T3B, T3e, T3h, T2l, T2u, T30, T2P, T2s, T2v, T2J, T2D, T2X, T2I, T2K, T2L;
Chris@42 559 {
Chris@42 560 E T7, Tu, T4, Tt;
Chris@42 561 T1 = rio[0];
Chris@42 562 {
Chris@42 563 E T5, T6, T2, T3;
Chris@42 564 T5 = rio[WS(rs, 2)];
Chris@42 565 T6 = rio[WS(rs, 3)];
Chris@42 566 T7 = T5 + T6;
Chris@42 567 Tu = T5 - T6;
Chris@42 568 T2 = rio[WS(rs, 1)];
Chris@42 569 T3 = rio[WS(rs, 4)];
Chris@42 570 T4 = T2 + T3;
Chris@42 571 Tt = T2 - T3;
Chris@42 572 }
Chris@42 573 Ta = KP559016994 * (T4 - T7);
Chris@42 574 TG = FNMS(KP587785252, Tt, KP951056516 * Tu);
Chris@42 575 Tv = FMA(KP951056516, Tt, KP587785252 * Tu);
Chris@42 576 T8 = T4 + T7;
Chris@42 577 Tb = FNMS(KP250000000, T8, T1);
Chris@42 578 }
Chris@42 579 {
Chris@42 580 E Ti, Tn, Tf, Tm;
Chris@42 581 Tp = iio[0];
Chris@42 582 {
Chris@42 583 E Tg, Th, Td, Te;
Chris@42 584 Tg = iio[WS(rs, 2)];
Chris@42 585 Th = iio[WS(rs, 3)];
Chris@42 586 Ti = Tg - Th;
Chris@42 587 Tn = Tg + Th;
Chris@42 588 Td = iio[WS(rs, 1)];
Chris@42 589 Te = iio[WS(rs, 4)];
Chris@42 590 Tf = Td - Te;
Chris@42 591 Tm = Td + Te;
Chris@42 592 }
Chris@42 593 Tj = FMA(KP951056516, Tf, KP587785252 * Ti);
Chris@42 594 TD = FNMS(KP587785252, Tf, KP951056516 * Ti);
Chris@42 595 To = KP559016994 * (Tm - Tn);
Chris@42 596 Tq = Tm + Tn;
Chris@42 597 Tr = FNMS(KP250000000, Tq, Tp);
Chris@42 598 }
Chris@42 599 {
Chris@42 600 E TT, T1g, TQ, T1f;
Chris@42 601 TN = rio[WS(vs, 1)];
Chris@42 602 {
Chris@42 603 E TR, TS, TO, TP;
Chris@42 604 TR = rio[WS(vs, 1) + WS(rs, 2)];
Chris@42 605 TS = rio[WS(vs, 1) + WS(rs, 3)];
Chris@42 606 TT = TR + TS;
Chris@42 607 T1g = TR - TS;
Chris@42 608 TO = rio[WS(vs, 1) + WS(rs, 1)];
Chris@42 609 TP = rio[WS(vs, 1) + WS(rs, 4)];
Chris@42 610 TQ = TO + TP;
Chris@42 611 T1f = TO - TP;
Chris@42 612 }
Chris@42 613 TW = KP559016994 * (TQ - TT);
Chris@42 614 T1s = FNMS(KP587785252, T1f, KP951056516 * T1g);
Chris@42 615 T1h = FMA(KP951056516, T1f, KP587785252 * T1g);
Chris@42 616 TU = TQ + TT;
Chris@42 617 TX = FNMS(KP250000000, TU, TN);
Chris@42 618 }
Chris@42 619 {
Chris@42 620 E T14, T19, T11, T18;
Chris@42 621 T1b = iio[WS(vs, 1)];
Chris@42 622 {
Chris@42 623 E T12, T13, TZ, T10;
Chris@42 624 T12 = iio[WS(vs, 1) + WS(rs, 2)];
Chris@42 625 T13 = iio[WS(vs, 1) + WS(rs, 3)];
Chris@42 626 T14 = T12 - T13;
Chris@42 627 T19 = T12 + T13;
Chris@42 628 TZ = iio[WS(vs, 1) + WS(rs, 1)];
Chris@42 629 T10 = iio[WS(vs, 1) + WS(rs, 4)];
Chris@42 630 T11 = TZ - T10;
Chris@42 631 T18 = TZ + T10;
Chris@42 632 }
Chris@42 633 T15 = FMA(KP951056516, T11, KP587785252 * T14);
Chris@42 634 T1p = FNMS(KP587785252, T11, KP951056516 * T14);
Chris@42 635 T1a = KP559016994 * (T18 - T19);
Chris@42 636 T1c = T18 + T19;
Chris@42 637 T1d = FNMS(KP250000000, T1c, T1b);
Chris@42 638 }
Chris@42 639 {
Chris@42 640 E T1F, T22, T1C, T21;
Chris@42 641 T1z = rio[WS(vs, 2)];
Chris@42 642 {
Chris@42 643 E T1D, T1E, T1A, T1B;
Chris@42 644 T1D = rio[WS(vs, 2) + WS(rs, 2)];
Chris@42 645 T1E = rio[WS(vs, 2) + WS(rs, 3)];
Chris@42 646 T1F = T1D + T1E;
Chris@42 647 T22 = T1D - T1E;
Chris@42 648 T1A = rio[WS(vs, 2) + WS(rs, 1)];
Chris@42 649 T1B = rio[WS(vs, 2) + WS(rs, 4)];
Chris@42 650 T1C = T1A + T1B;
Chris@42 651 T21 = T1A - T1B;
Chris@42 652 }
Chris@42 653 T1I = KP559016994 * (T1C - T1F);
Chris@42 654 T2e = FNMS(KP587785252, T21, KP951056516 * T22);
Chris@42 655 T23 = FMA(KP951056516, T21, KP587785252 * T22);
Chris@42 656 T1G = T1C + T1F;
Chris@42 657 T1J = FNMS(KP250000000, T1G, T1z);
Chris@42 658 }
Chris@42 659 {
Chris@42 660 E T1Q, T1V, T1N, T1U;
Chris@42 661 T1X = iio[WS(vs, 2)];
Chris@42 662 {
Chris@42 663 E T1O, T1P, T1L, T1M;
Chris@42 664 T1O = iio[WS(vs, 2) + WS(rs, 2)];
Chris@42 665 T1P = iio[WS(vs, 2) + WS(rs, 3)];
Chris@42 666 T1Q = T1O - T1P;
Chris@42 667 T1V = T1O + T1P;
Chris@42 668 T1L = iio[WS(vs, 2) + WS(rs, 1)];
Chris@42 669 T1M = iio[WS(vs, 2) + WS(rs, 4)];
Chris@42 670 T1N = T1L - T1M;
Chris@42 671 T1U = T1L + T1M;
Chris@42 672 }
Chris@42 673 T1R = FMA(KP951056516, T1N, KP587785252 * T1Q);
Chris@42 674 T2b = FNMS(KP587785252, T1N, KP951056516 * T1Q);
Chris@42 675 T1W = KP559016994 * (T1U - T1V);
Chris@42 676 T1Y = T1U + T1V;
Chris@42 677 T1Z = FNMS(KP250000000, T1Y, T1X);
Chris@42 678 }
Chris@42 679 {
Chris@42 680 E T3o, T3t, T3l, T3s;
Chris@42 681 T3v = iio[WS(vs, 4)];
Chris@42 682 {
Chris@42 683 E T3m, T3n, T3j, T3k;
Chris@42 684 T3m = iio[WS(vs, 4) + WS(rs, 2)];
Chris@42 685 T3n = iio[WS(vs, 4) + WS(rs, 3)];
Chris@42 686 T3o = T3m - T3n;
Chris@42 687 T3t = T3m + T3n;
Chris@42 688 T3j = iio[WS(vs, 4) + WS(rs, 1)];
Chris@42 689 T3k = iio[WS(vs, 4) + WS(rs, 4)];
Chris@42 690 T3l = T3j - T3k;
Chris@42 691 T3s = T3j + T3k;
Chris@42 692 }
Chris@42 693 T3p = FMA(KP951056516, T3l, KP587785252 * T3o);
Chris@42 694 T3J = FNMS(KP587785252, T3l, KP951056516 * T3o);
Chris@42 695 T3u = KP559016994 * (T3s - T3t);
Chris@42 696 T3w = T3s + T3t;
Chris@42 697 T3x = FNMS(KP250000000, T3w, T3v);
Chris@42 698 }
Chris@42 699 {
Chris@42 700 E T3d, T3A, T3a, T3z;
Chris@42 701 T37 = rio[WS(vs, 4)];
Chris@42 702 {
Chris@42 703 E T3b, T3c, T38, T39;
Chris@42 704 T3b = rio[WS(vs, 4) + WS(rs, 2)];
Chris@42 705 T3c = rio[WS(vs, 4) + WS(rs, 3)];
Chris@42 706 T3d = T3b + T3c;
Chris@42 707 T3A = T3b - T3c;
Chris@42 708 T38 = rio[WS(vs, 4) + WS(rs, 1)];
Chris@42 709 T39 = rio[WS(vs, 4) + WS(rs, 4)];
Chris@42 710 T3a = T38 + T39;
Chris@42 711 T3z = T38 - T39;
Chris@42 712 }
Chris@42 713 T3g = KP559016994 * (T3a - T3d);
Chris@42 714 T3M = FNMS(KP587785252, T3z, KP951056516 * T3A);
Chris@42 715 T3B = FMA(KP951056516, T3z, KP587785252 * T3A);
Chris@42 716 T3e = T3a + T3d;
Chris@42 717 T3h = FNMS(KP250000000, T3e, T37);
Chris@42 718 }
Chris@42 719 {
Chris@42 720 E T2r, T2O, T2o, T2N;
Chris@42 721 T2l = rio[WS(vs, 3)];
Chris@42 722 {
Chris@42 723 E T2p, T2q, T2m, T2n;
Chris@42 724 T2p = rio[WS(vs, 3) + WS(rs, 2)];
Chris@42 725 T2q = rio[WS(vs, 3) + WS(rs, 3)];
Chris@42 726 T2r = T2p + T2q;
Chris@42 727 T2O = T2p - T2q;
Chris@42 728 T2m = rio[WS(vs, 3) + WS(rs, 1)];
Chris@42 729 T2n = rio[WS(vs, 3) + WS(rs, 4)];
Chris@42 730 T2o = T2m + T2n;
Chris@42 731 T2N = T2m - T2n;
Chris@42 732 }
Chris@42 733 T2u = KP559016994 * (T2o - T2r);
Chris@42 734 T30 = FNMS(KP587785252, T2N, KP951056516 * T2O);
Chris@42 735 T2P = FMA(KP951056516, T2N, KP587785252 * T2O);
Chris@42 736 T2s = T2o + T2r;
Chris@42 737 T2v = FNMS(KP250000000, T2s, T2l);
Chris@42 738 }
Chris@42 739 {
Chris@42 740 E T2C, T2H, T2z, T2G;
Chris@42 741 T2J = iio[WS(vs, 3)];
Chris@42 742 {
Chris@42 743 E T2A, T2B, T2x, T2y;
Chris@42 744 T2A = iio[WS(vs, 3) + WS(rs, 2)];
Chris@42 745 T2B = iio[WS(vs, 3) + WS(rs, 3)];
Chris@42 746 T2C = T2A - T2B;
Chris@42 747 T2H = T2A + T2B;
Chris@42 748 T2x = iio[WS(vs, 3) + WS(rs, 1)];
Chris@42 749 T2y = iio[WS(vs, 3) + WS(rs, 4)];
Chris@42 750 T2z = T2x - T2y;
Chris@42 751 T2G = T2x + T2y;
Chris@42 752 }
Chris@42 753 T2D = FMA(KP951056516, T2z, KP587785252 * T2C);
Chris@42 754 T2X = FNMS(KP587785252, T2z, KP951056516 * T2C);
Chris@42 755 T2I = KP559016994 * (T2G - T2H);
Chris@42 756 T2K = T2G + T2H;
Chris@42 757 T2L = FNMS(KP250000000, T2K, T2J);
Chris@42 758 }
Chris@42 759 rio[0] = T1 + T8;
Chris@42 760 iio[0] = Tp + Tq;
Chris@42 761 rio[WS(rs, 1)] = TN + TU;
Chris@42 762 iio[WS(rs, 1)] = T1b + T1c;
Chris@42 763 rio[WS(rs, 2)] = T1z + T1G;
Chris@42 764 iio[WS(rs, 2)] = T1X + T1Y;
Chris@42 765 iio[WS(rs, 4)] = T3v + T3w;
Chris@42 766 rio[WS(rs, 4)] = T37 + T3e;
Chris@42 767 rio[WS(rs, 3)] = T2l + T2s;
Chris@42 768 iio[WS(rs, 3)] = T2J + T2K;
Chris@42 769 {
Chris@42 770 E Tk, Ty, Tw, TA, Tc, Ts;
Chris@42 771 Tc = Ta + Tb;
Chris@42 772 Tk = Tc + Tj;
Chris@42 773 Ty = Tc - Tj;
Chris@42 774 Ts = To + Tr;
Chris@42 775 Tw = Ts - Tv;
Chris@42 776 TA = Tv + Ts;
Chris@42 777 {
Chris@42 778 E T9, Tl, Tx, Tz;
Chris@42 779 T9 = W[0];
Chris@42 780 Tl = W[1];
Chris@42 781 rio[WS(vs, 1)] = FMA(T9, Tk, Tl * Tw);
Chris@42 782 iio[WS(vs, 1)] = FNMS(Tl, Tk, T9 * Tw);
Chris@42 783 Tx = W[6];
Chris@42 784 Tz = W[7];
Chris@42 785 rio[WS(vs, 4)] = FMA(Tx, Ty, Tz * TA);
Chris@42 786 iio[WS(vs, 4)] = FNMS(Tz, Ty, Tx * TA);
Chris@42 787 }
Chris@42 788 }
Chris@42 789 {
Chris@42 790 E TE, TK, TI, TM, TC, TH;
Chris@42 791 TC = Tb - Ta;
Chris@42 792 TE = TC - TD;
Chris@42 793 TK = TC + TD;
Chris@42 794 TH = Tr - To;
Chris@42 795 TI = TG + TH;
Chris@42 796 TM = TH - TG;
Chris@42 797 {
Chris@42 798 E TB, TF, TJ, TL;
Chris@42 799 TB = W[2];
Chris@42 800 TF = W[3];
Chris@42 801 rio[WS(vs, 2)] = FMA(TB, TE, TF * TI);
Chris@42 802 iio[WS(vs, 2)] = FNMS(TF, TE, TB * TI);
Chris@42 803 TJ = W[4];
Chris@42 804 TL = W[5];
Chris@42 805 rio[WS(vs, 3)] = FMA(TJ, TK, TL * TM);
Chris@42 806 iio[WS(vs, 3)] = FNMS(TL, TK, TJ * TM);
Chris@42 807 }
Chris@42 808 }
Chris@42 809 {
Chris@42 810 E T2c, T2i, T2g, T2k, T2a, T2f;
Chris@42 811 T2a = T1J - T1I;
Chris@42 812 T2c = T2a - T2b;
Chris@42 813 T2i = T2a + T2b;
Chris@42 814 T2f = T1Z - T1W;
Chris@42 815 T2g = T2e + T2f;
Chris@42 816 T2k = T2f - T2e;
Chris@42 817 {
Chris@42 818 E T29, T2d, T2h, T2j;
Chris@42 819 T29 = W[2];
Chris@42 820 T2d = W[3];
Chris@42 821 rio[WS(vs, 2) + WS(rs, 2)] = FMA(T29, T2c, T2d * T2g);
Chris@42 822 iio[WS(vs, 2) + WS(rs, 2)] = FNMS(T2d, T2c, T29 * T2g);
Chris@42 823 T2h = W[4];
Chris@42 824 T2j = W[5];
Chris@42 825 rio[WS(vs, 3) + WS(rs, 2)] = FMA(T2h, T2i, T2j * T2k);
Chris@42 826 iio[WS(vs, 3) + WS(rs, 2)] = FNMS(T2j, T2i, T2h * T2k);
Chris@42 827 }
Chris@42 828 }
Chris@42 829 {
Chris@42 830 E T3K, T3Q, T3O, T3S, T3I, T3N;
Chris@42 831 T3I = T3h - T3g;
Chris@42 832 T3K = T3I - T3J;
Chris@42 833 T3Q = T3I + T3J;
Chris@42 834 T3N = T3x - T3u;
Chris@42 835 T3O = T3M + T3N;
Chris@42 836 T3S = T3N - T3M;
Chris@42 837 {
Chris@42 838 E T3H, T3L, T3P, T3R;
Chris@42 839 T3H = W[2];
Chris@42 840 T3L = W[3];
Chris@42 841 rio[WS(vs, 2) + WS(rs, 4)] = FMA(T3H, T3K, T3L * T3O);
Chris@42 842 iio[WS(vs, 2) + WS(rs, 4)] = FNMS(T3L, T3K, T3H * T3O);
Chris@42 843 T3P = W[4];
Chris@42 844 T3R = W[5];
Chris@42 845 rio[WS(vs, 3) + WS(rs, 4)] = FMA(T3P, T3Q, T3R * T3S);
Chris@42 846 iio[WS(vs, 3) + WS(rs, 4)] = FNMS(T3R, T3Q, T3P * T3S);
Chris@42 847 }
Chris@42 848 }
Chris@42 849 {
Chris@42 850 E T1S, T26, T24, T28, T1K, T20;
Chris@42 851 T1K = T1I + T1J;
Chris@42 852 T1S = T1K + T1R;
Chris@42 853 T26 = T1K - T1R;
Chris@42 854 T20 = T1W + T1Z;
Chris@42 855 T24 = T20 - T23;
Chris@42 856 T28 = T23 + T20;
Chris@42 857 {
Chris@42 858 E T1H, T1T, T25, T27;
Chris@42 859 T1H = W[0];
Chris@42 860 T1T = W[1];
Chris@42 861 rio[WS(vs, 1) + WS(rs, 2)] = FMA(T1H, T1S, T1T * T24);
Chris@42 862 iio[WS(vs, 1) + WS(rs, 2)] = FNMS(T1T, T1S, T1H * T24);
Chris@42 863 T25 = W[6];
Chris@42 864 T27 = W[7];
Chris@42 865 rio[WS(vs, 4) + WS(rs, 2)] = FMA(T25, T26, T27 * T28);
Chris@42 866 iio[WS(vs, 4) + WS(rs, 2)] = FNMS(T27, T26, T25 * T28);
Chris@42 867 }
Chris@42 868 }
Chris@42 869 {
Chris@42 870 E T2E, T2S, T2Q, T2U, T2w, T2M;
Chris@42 871 T2w = T2u + T2v;
Chris@42 872 T2E = T2w + T2D;
Chris@42 873 T2S = T2w - T2D;
Chris@42 874 T2M = T2I + T2L;
Chris@42 875 T2Q = T2M - T2P;
Chris@42 876 T2U = T2P + T2M;
Chris@42 877 {
Chris@42 878 E T2t, T2F, T2R, T2T;
Chris@42 879 T2t = W[0];
Chris@42 880 T2F = W[1];
Chris@42 881 rio[WS(vs, 1) + WS(rs, 3)] = FMA(T2t, T2E, T2F * T2Q);
Chris@42 882 iio[WS(vs, 1) + WS(rs, 3)] = FNMS(T2F, T2E, T2t * T2Q);
Chris@42 883 T2R = W[6];
Chris@42 884 T2T = W[7];
Chris@42 885 rio[WS(vs, 4) + WS(rs, 3)] = FMA(T2R, T2S, T2T * T2U);
Chris@42 886 iio[WS(vs, 4) + WS(rs, 3)] = FNMS(T2T, T2S, T2R * T2U);
Chris@42 887 }
Chris@42 888 }
Chris@42 889 {
Chris@42 890 E T2Y, T34, T32, T36, T2W, T31;
Chris@42 891 T2W = T2v - T2u;
Chris@42 892 T2Y = T2W - T2X;
Chris@42 893 T34 = T2W + T2X;
Chris@42 894 T31 = T2L - T2I;
Chris@42 895 T32 = T30 + T31;
Chris@42 896 T36 = T31 - T30;
Chris@42 897 {
Chris@42 898 E T2V, T2Z, T33, T35;
Chris@42 899 T2V = W[2];
Chris@42 900 T2Z = W[3];
Chris@42 901 rio[WS(vs, 2) + WS(rs, 3)] = FMA(T2V, T2Y, T2Z * T32);
Chris@42 902 iio[WS(vs, 2) + WS(rs, 3)] = FNMS(T2Z, T2Y, T2V * T32);
Chris@42 903 T33 = W[4];
Chris@42 904 T35 = W[5];
Chris@42 905 rio[WS(vs, 3) + WS(rs, 3)] = FMA(T33, T34, T35 * T36);
Chris@42 906 iio[WS(vs, 3) + WS(rs, 3)] = FNMS(T35, T34, T33 * T36);
Chris@42 907 }
Chris@42 908 }
Chris@42 909 {
Chris@42 910 E T3q, T3E, T3C, T3G, T3i, T3y;
Chris@42 911 T3i = T3g + T3h;
Chris@42 912 T3q = T3i + T3p;
Chris@42 913 T3E = T3i - T3p;
Chris@42 914 T3y = T3u + T3x;
Chris@42 915 T3C = T3y - T3B;
Chris@42 916 T3G = T3B + T3y;
Chris@42 917 {
Chris@42 918 E T3f, T3r, T3D, T3F;
Chris@42 919 T3f = W[0];
Chris@42 920 T3r = W[1];
Chris@42 921 rio[WS(vs, 1) + WS(rs, 4)] = FMA(T3f, T3q, T3r * T3C);
Chris@42 922 iio[WS(vs, 1) + WS(rs, 4)] = FNMS(T3r, T3q, T3f * T3C);
Chris@42 923 T3D = W[6];
Chris@42 924 T3F = W[7];
Chris@42 925 rio[WS(vs, 4) + WS(rs, 4)] = FMA(T3D, T3E, T3F * T3G);
Chris@42 926 iio[WS(vs, 4) + WS(rs, 4)] = FNMS(T3F, T3E, T3D * T3G);
Chris@42 927 }
Chris@42 928 }
Chris@42 929 {
Chris@42 930 E T1q, T1w, T1u, T1y, T1o, T1t;
Chris@42 931 T1o = TX - TW;
Chris@42 932 T1q = T1o - T1p;
Chris@42 933 T1w = T1o + T1p;
Chris@42 934 T1t = T1d - T1a;
Chris@42 935 T1u = T1s + T1t;
Chris@42 936 T1y = T1t - T1s;
Chris@42 937 {
Chris@42 938 E T1n, T1r, T1v, T1x;
Chris@42 939 T1n = W[2];
Chris@42 940 T1r = W[3];
Chris@42 941 rio[WS(vs, 2) + WS(rs, 1)] = FMA(T1n, T1q, T1r * T1u);
Chris@42 942 iio[WS(vs, 2) + WS(rs, 1)] = FNMS(T1r, T1q, T1n * T1u);
Chris@42 943 T1v = W[4];
Chris@42 944 T1x = W[5];
Chris@42 945 rio[WS(vs, 3) + WS(rs, 1)] = FMA(T1v, T1w, T1x * T1y);
Chris@42 946 iio[WS(vs, 3) + WS(rs, 1)] = FNMS(T1x, T1w, T1v * T1y);
Chris@42 947 }
Chris@42 948 }
Chris@42 949 {
Chris@42 950 E T16, T1k, T1i, T1m, TY, T1e;
Chris@42 951 TY = TW + TX;
Chris@42 952 T16 = TY + T15;
Chris@42 953 T1k = TY - T15;
Chris@42 954 T1e = T1a + T1d;
Chris@42 955 T1i = T1e - T1h;
Chris@42 956 T1m = T1h + T1e;
Chris@42 957 {
Chris@42 958 E TV, T17, T1j, T1l;
Chris@42 959 TV = W[0];
Chris@42 960 T17 = W[1];
Chris@42 961 rio[WS(vs, 1) + WS(rs, 1)] = FMA(TV, T16, T17 * T1i);
Chris@42 962 iio[WS(vs, 1) + WS(rs, 1)] = FNMS(T17, T16, TV * T1i);
Chris@42 963 T1j = W[6];
Chris@42 964 T1l = W[7];
Chris@42 965 rio[WS(vs, 4) + WS(rs, 1)] = FMA(T1j, T1k, T1l * T1m);
Chris@42 966 iio[WS(vs, 4) + WS(rs, 1)] = FNMS(T1l, T1k, T1j * T1m);
Chris@42 967 }
Chris@42 968 }
Chris@42 969 }
Chris@42 970 }
Chris@42 971 }
Chris@42 972
Chris@42 973 static const tw_instr twinstr[] = {
Chris@42 974 {TW_FULL, 0, 5},
Chris@42 975 {TW_NEXT, 1, 0}
Chris@42 976 };
Chris@42 977
Chris@42 978 static const ct_desc desc = { 5, "q1_5", twinstr, &GENUS, {130, 70, 70, 0}, 0, 0, 0 };
Chris@42 979
Chris@42 980 void X(codelet_q1_5) (planner *p) {
Chris@42 981 X(kdft_difsq_register) (p, q1_5, &desc);
Chris@42 982 }
Chris@42 983 #endif /* HAVE_FMA */