Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.8/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 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 08:04:30 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_twidsq.native -fma -compact -variables 4 -pipeline-latency 4 -reload-twiddle -dif -n 5 -name q1_5 -include dft/scalar/q.h */
Chris@82 29
Chris@82 30 /*
Chris@82 31 * This function contains 200 FP additions, 170 FP multiplications,
Chris@82 32 * (or, 70 additions, 40 multiplications, 130 fused multiply/add),
Chris@82 33 * 75 stack variables, 4 constants, and 100 memory accesses
Chris@82 34 */
Chris@82 35 #include "dft/scalar/q.h"
Chris@82 36
Chris@82 37 static void q1_5(R *rio, R *iio, const R *W, stride rs, stride vs, 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, rio = rio + ms, iio = iio + ms, W = W + 8, MAKE_VOLATILE_STRIDE(10, rs), MAKE_VOLATILE_STRIDE(0, vs)) {
Chris@82 46 E T1, Tb, TM, Tw, T8, Ta, Tn, Tj, TH, Ts, Tq, Tr, TV, T15, T1G;
Chris@82 47 E T1q, T12, T14, T1h, T1d, T1B, T1m, T1k, T1l, T1P, T1Z, T2A, T2k, T1W, T1Y;
Chris@82 48 E T2b, T27, T2v, T2g, T2e, T2f, T3Z, T3V, T4j, T44, T42, T43, T3D, T3N, T4o;
Chris@82 49 E T48, T3K, T3M, T2J, T2T, T3u, T3e, T2Q, T2S, T35, T31, T3p, T3a, T38, T39;
Chris@82 50 {
Chris@82 51 E T7, Tv, T4, Tu;
Chris@82 52 T1 = rio[0];
Chris@82 53 {
Chris@82 54 E T5, T6, T2, T3;
Chris@82 55 T5 = rio[WS(rs, 2)];
Chris@82 56 T6 = rio[WS(rs, 3)];
Chris@82 57 T7 = T5 + T6;
Chris@82 58 Tv = T5 - T6;
Chris@82 59 T2 = rio[WS(rs, 1)];
Chris@82 60 T3 = rio[WS(rs, 4)];
Chris@82 61 T4 = T2 + T3;
Chris@82 62 Tu = T2 - T3;
Chris@82 63 }
Chris@82 64 Tb = T4 - T7;
Chris@82 65 TM = FNMS(KP618033988, Tu, Tv);
Chris@82 66 Tw = FMA(KP618033988, Tv, Tu);
Chris@82 67 T8 = T4 + T7;
Chris@82 68 Ta = FNMS(KP250000000, T8, T1);
Chris@82 69 }
Chris@82 70 {
Chris@82 71 E Ti, Tp, Tf, To;
Chris@82 72 Tn = iio[0];
Chris@82 73 {
Chris@82 74 E Tg, Th, Td, Te;
Chris@82 75 Tg = iio[WS(rs, 2)];
Chris@82 76 Th = iio[WS(rs, 3)];
Chris@82 77 Ti = Tg - Th;
Chris@82 78 Tp = Tg + Th;
Chris@82 79 Td = iio[WS(rs, 1)];
Chris@82 80 Te = iio[WS(rs, 4)];
Chris@82 81 Tf = Td - Te;
Chris@82 82 To = Td + Te;
Chris@82 83 }
Chris@82 84 Tj = FMA(KP618033988, Ti, Tf);
Chris@82 85 TH = FNMS(KP618033988, Tf, Ti);
Chris@82 86 Ts = To - Tp;
Chris@82 87 Tq = To + Tp;
Chris@82 88 Tr = FNMS(KP250000000, Tq, Tn);
Chris@82 89 }
Chris@82 90 {
Chris@82 91 E T11, T1p, TY, T1o;
Chris@82 92 TV = rio[WS(vs, 1)];
Chris@82 93 {
Chris@82 94 E TZ, T10, TW, TX;
Chris@82 95 TZ = rio[WS(vs, 1) + WS(rs, 2)];
Chris@82 96 T10 = rio[WS(vs, 1) + WS(rs, 3)];
Chris@82 97 T11 = TZ + T10;
Chris@82 98 T1p = TZ - T10;
Chris@82 99 TW = rio[WS(vs, 1) + WS(rs, 1)];
Chris@82 100 TX = rio[WS(vs, 1) + WS(rs, 4)];
Chris@82 101 TY = TW + TX;
Chris@82 102 T1o = TW - TX;
Chris@82 103 }
Chris@82 104 T15 = TY - T11;
Chris@82 105 T1G = FNMS(KP618033988, T1o, T1p);
Chris@82 106 T1q = FMA(KP618033988, T1p, T1o);
Chris@82 107 T12 = TY + T11;
Chris@82 108 T14 = FNMS(KP250000000, T12, TV);
Chris@82 109 }
Chris@82 110 {
Chris@82 111 E T1c, T1j, T19, T1i;
Chris@82 112 T1h = iio[WS(vs, 1)];
Chris@82 113 {
Chris@82 114 E T1a, T1b, T17, T18;
Chris@82 115 T1a = iio[WS(vs, 1) + WS(rs, 2)];
Chris@82 116 T1b = iio[WS(vs, 1) + WS(rs, 3)];
Chris@82 117 T1c = T1a - T1b;
Chris@82 118 T1j = T1a + T1b;
Chris@82 119 T17 = iio[WS(vs, 1) + WS(rs, 1)];
Chris@82 120 T18 = iio[WS(vs, 1) + WS(rs, 4)];
Chris@82 121 T19 = T17 - T18;
Chris@82 122 T1i = T17 + T18;
Chris@82 123 }
Chris@82 124 T1d = FMA(KP618033988, T1c, T19);
Chris@82 125 T1B = FNMS(KP618033988, T19, T1c);
Chris@82 126 T1m = T1i - T1j;
Chris@82 127 T1k = T1i + T1j;
Chris@82 128 T1l = FNMS(KP250000000, T1k, T1h);
Chris@82 129 }
Chris@82 130 {
Chris@82 131 E T1V, T2j, T1S, T2i;
Chris@82 132 T1P = rio[WS(vs, 2)];
Chris@82 133 {
Chris@82 134 E T1T, T1U, T1Q, T1R;
Chris@82 135 T1T = rio[WS(vs, 2) + WS(rs, 2)];
Chris@82 136 T1U = rio[WS(vs, 2) + WS(rs, 3)];
Chris@82 137 T1V = T1T + T1U;
Chris@82 138 T2j = T1T - T1U;
Chris@82 139 T1Q = rio[WS(vs, 2) + WS(rs, 1)];
Chris@82 140 T1R = rio[WS(vs, 2) + WS(rs, 4)];
Chris@82 141 T1S = T1Q + T1R;
Chris@82 142 T2i = T1Q - T1R;
Chris@82 143 }
Chris@82 144 T1Z = T1S - T1V;
Chris@82 145 T2A = FNMS(KP618033988, T2i, T2j);
Chris@82 146 T2k = FMA(KP618033988, T2j, T2i);
Chris@82 147 T1W = T1S + T1V;
Chris@82 148 T1Y = FNMS(KP250000000, T1W, T1P);
Chris@82 149 }
Chris@82 150 {
Chris@82 151 E T26, T2d, T23, T2c;
Chris@82 152 T2b = iio[WS(vs, 2)];
Chris@82 153 {
Chris@82 154 E T24, T25, T21, T22;
Chris@82 155 T24 = iio[WS(vs, 2) + WS(rs, 2)];
Chris@82 156 T25 = iio[WS(vs, 2) + WS(rs, 3)];
Chris@82 157 T26 = T24 - T25;
Chris@82 158 T2d = T24 + T25;
Chris@82 159 T21 = iio[WS(vs, 2) + WS(rs, 1)];
Chris@82 160 T22 = iio[WS(vs, 2) + WS(rs, 4)];
Chris@82 161 T23 = T21 - T22;
Chris@82 162 T2c = T21 + T22;
Chris@82 163 }
Chris@82 164 T27 = FMA(KP618033988, T26, T23);
Chris@82 165 T2v = FNMS(KP618033988, T23, T26);
Chris@82 166 T2g = T2c - T2d;
Chris@82 167 T2e = T2c + T2d;
Chris@82 168 T2f = FNMS(KP250000000, T2e, T2b);
Chris@82 169 }
Chris@82 170 {
Chris@82 171 E T3U, T41, T3R, T40;
Chris@82 172 T3Z = iio[WS(vs, 4)];
Chris@82 173 {
Chris@82 174 E T3S, T3T, T3P, T3Q;
Chris@82 175 T3S = iio[WS(vs, 4) + WS(rs, 2)];
Chris@82 176 T3T = iio[WS(vs, 4) + WS(rs, 3)];
Chris@82 177 T3U = T3S - T3T;
Chris@82 178 T41 = T3S + T3T;
Chris@82 179 T3P = iio[WS(vs, 4) + WS(rs, 1)];
Chris@82 180 T3Q = iio[WS(vs, 4) + WS(rs, 4)];
Chris@82 181 T3R = T3P - T3Q;
Chris@82 182 T40 = T3P + T3Q;
Chris@82 183 }
Chris@82 184 T3V = FMA(KP618033988, T3U, T3R);
Chris@82 185 T4j = FNMS(KP618033988, T3R, T3U);
Chris@82 186 T44 = T40 - T41;
Chris@82 187 T42 = T40 + T41;
Chris@82 188 T43 = FNMS(KP250000000, T42, T3Z);
Chris@82 189 }
Chris@82 190 {
Chris@82 191 E T3J, T47, T3G, T46;
Chris@82 192 T3D = rio[WS(vs, 4)];
Chris@82 193 {
Chris@82 194 E T3H, T3I, T3E, T3F;
Chris@82 195 T3H = rio[WS(vs, 4) + WS(rs, 2)];
Chris@82 196 T3I = rio[WS(vs, 4) + WS(rs, 3)];
Chris@82 197 T3J = T3H + T3I;
Chris@82 198 T47 = T3H - T3I;
Chris@82 199 T3E = rio[WS(vs, 4) + WS(rs, 1)];
Chris@82 200 T3F = rio[WS(vs, 4) + WS(rs, 4)];
Chris@82 201 T3G = T3E + T3F;
Chris@82 202 T46 = T3E - T3F;
Chris@82 203 }
Chris@82 204 T3N = T3G - T3J;
Chris@82 205 T4o = FNMS(KP618033988, T46, T47);
Chris@82 206 T48 = FMA(KP618033988, T47, T46);
Chris@82 207 T3K = T3G + T3J;
Chris@82 208 T3M = FNMS(KP250000000, T3K, T3D);
Chris@82 209 }
Chris@82 210 {
Chris@82 211 E T2P, T3d, T2M, T3c;
Chris@82 212 T2J = rio[WS(vs, 3)];
Chris@82 213 {
Chris@82 214 E T2N, T2O, T2K, T2L;
Chris@82 215 T2N = rio[WS(vs, 3) + WS(rs, 2)];
Chris@82 216 T2O = rio[WS(vs, 3) + WS(rs, 3)];
Chris@82 217 T2P = T2N + T2O;
Chris@82 218 T3d = T2N - T2O;
Chris@82 219 T2K = rio[WS(vs, 3) + WS(rs, 1)];
Chris@82 220 T2L = rio[WS(vs, 3) + WS(rs, 4)];
Chris@82 221 T2M = T2K + T2L;
Chris@82 222 T3c = T2K - T2L;
Chris@82 223 }
Chris@82 224 T2T = T2M - T2P;
Chris@82 225 T3u = FNMS(KP618033988, T3c, T3d);
Chris@82 226 T3e = FMA(KP618033988, T3d, T3c);
Chris@82 227 T2Q = T2M + T2P;
Chris@82 228 T2S = FNMS(KP250000000, T2Q, T2J);
Chris@82 229 }
Chris@82 230 {
Chris@82 231 E T30, T37, T2X, T36;
Chris@82 232 T35 = iio[WS(vs, 3)];
Chris@82 233 {
Chris@82 234 E T2Y, T2Z, T2V, T2W;
Chris@82 235 T2Y = iio[WS(vs, 3) + WS(rs, 2)];
Chris@82 236 T2Z = iio[WS(vs, 3) + WS(rs, 3)];
Chris@82 237 T30 = T2Y - T2Z;
Chris@82 238 T37 = T2Y + T2Z;
Chris@82 239 T2V = iio[WS(vs, 3) + WS(rs, 1)];
Chris@82 240 T2W = iio[WS(vs, 3) + WS(rs, 4)];
Chris@82 241 T2X = T2V - T2W;
Chris@82 242 T36 = T2V + T2W;
Chris@82 243 }
Chris@82 244 T31 = FMA(KP618033988, T30, T2X);
Chris@82 245 T3p = FNMS(KP618033988, T2X, T30);
Chris@82 246 T3a = T36 - T37;
Chris@82 247 T38 = T36 + T37;
Chris@82 248 T39 = FNMS(KP250000000, T38, T35);
Chris@82 249 }
Chris@82 250 rio[0] = T1 + T8;
Chris@82 251 iio[0] = Tn + Tq;
Chris@82 252 rio[WS(rs, 1)] = TV + T12;
Chris@82 253 iio[WS(rs, 1)] = T1h + T1k;
Chris@82 254 rio[WS(rs, 2)] = T1P + T1W;
Chris@82 255 iio[WS(rs, 2)] = T2b + T2e;
Chris@82 256 iio[WS(rs, 4)] = T3Z + T42;
Chris@82 257 rio[WS(rs, 4)] = T3D + T3K;
Chris@82 258 rio[WS(rs, 3)] = T2J + T2Q;
Chris@82 259 iio[WS(rs, 3)] = T35 + T38;
Chris@82 260 {
Chris@82 261 E Tk, TA, Tx, TD, Tc, Tt;
Chris@82 262 Tc = FMA(KP559016994, Tb, Ta);
Chris@82 263 Tk = FMA(KP951056516, Tj, Tc);
Chris@82 264 TA = FNMS(KP951056516, Tj, Tc);
Chris@82 265 Tt = FMA(KP559016994, Ts, Tr);
Chris@82 266 Tx = FNMS(KP951056516, Tw, Tt);
Chris@82 267 TD = FMA(KP951056516, Tw, Tt);
Chris@82 268 {
Chris@82 269 E Tl, Ty, T9, Tm;
Chris@82 270 T9 = W[0];
Chris@82 271 Tl = T9 * Tk;
Chris@82 272 Ty = T9 * Tx;
Chris@82 273 Tm = W[1];
Chris@82 274 rio[WS(vs, 1)] = FMA(Tm, Tx, Tl);
Chris@82 275 iio[WS(vs, 1)] = FNMS(Tm, Tk, Ty);
Chris@82 276 }
Chris@82 277 {
Chris@82 278 E TB, TE, Tz, TC;
Chris@82 279 Tz = W[6];
Chris@82 280 TB = Tz * TA;
Chris@82 281 TE = Tz * TD;
Chris@82 282 TC = W[7];
Chris@82 283 rio[WS(vs, 4)] = FMA(TC, TD, TB);
Chris@82 284 iio[WS(vs, 4)] = FNMS(TC, TA, TE);
Chris@82 285 }
Chris@82 286 }
Chris@82 287 {
Chris@82 288 E TI, TQ, TN, TT, TG, TL;
Chris@82 289 TG = FNMS(KP559016994, Tb, Ta);
Chris@82 290 TI = FNMS(KP951056516, TH, TG);
Chris@82 291 TQ = FMA(KP951056516, TH, TG);
Chris@82 292 TL = FNMS(KP559016994, Ts, Tr);
Chris@82 293 TN = FMA(KP951056516, TM, TL);
Chris@82 294 TT = FNMS(KP951056516, TM, TL);
Chris@82 295 {
Chris@82 296 E TJ, TO, TF, TK;
Chris@82 297 TF = W[2];
Chris@82 298 TJ = TF * TI;
Chris@82 299 TO = TF * TN;
Chris@82 300 TK = W[3];
Chris@82 301 rio[WS(vs, 2)] = FMA(TK, TN, TJ);
Chris@82 302 iio[WS(vs, 2)] = FNMS(TK, TI, TO);
Chris@82 303 }
Chris@82 304 {
Chris@82 305 E TR, TU, TP, TS;
Chris@82 306 TP = W[4];
Chris@82 307 TR = TP * TQ;
Chris@82 308 TU = TP * TT;
Chris@82 309 TS = W[5];
Chris@82 310 rio[WS(vs, 3)] = FMA(TS, TT, TR);
Chris@82 311 iio[WS(vs, 3)] = FNMS(TS, TQ, TU);
Chris@82 312 }
Chris@82 313 }
Chris@82 314 {
Chris@82 315 E T2w, T2E, T2B, T2H, T2u, T2z;
Chris@82 316 T2u = FNMS(KP559016994, T1Z, T1Y);
Chris@82 317 T2w = FNMS(KP951056516, T2v, T2u);
Chris@82 318 T2E = FMA(KP951056516, T2v, T2u);
Chris@82 319 T2z = FNMS(KP559016994, T2g, T2f);
Chris@82 320 T2B = FMA(KP951056516, T2A, T2z);
Chris@82 321 T2H = FNMS(KP951056516, T2A, T2z);
Chris@82 322 {
Chris@82 323 E T2x, T2C, T2t, T2y;
Chris@82 324 T2t = W[2];
Chris@82 325 T2x = T2t * T2w;
Chris@82 326 T2C = T2t * T2B;
Chris@82 327 T2y = W[3];
Chris@82 328 rio[WS(vs, 2) + WS(rs, 2)] = FMA(T2y, T2B, T2x);
Chris@82 329 iio[WS(vs, 2) + WS(rs, 2)] = FNMS(T2y, T2w, T2C);
Chris@82 330 }
Chris@82 331 {
Chris@82 332 E T2F, T2I, T2D, T2G;
Chris@82 333 T2D = W[4];
Chris@82 334 T2F = T2D * T2E;
Chris@82 335 T2I = T2D * T2H;
Chris@82 336 T2G = W[5];
Chris@82 337 rio[WS(vs, 3) + WS(rs, 2)] = FMA(T2G, T2H, T2F);
Chris@82 338 iio[WS(vs, 3) + WS(rs, 2)] = FNMS(T2G, T2E, T2I);
Chris@82 339 }
Chris@82 340 }
Chris@82 341 {
Chris@82 342 E T4k, T4s, T4p, T4v, T4i, T4n;
Chris@82 343 T4i = FNMS(KP559016994, T3N, T3M);
Chris@82 344 T4k = FNMS(KP951056516, T4j, T4i);
Chris@82 345 T4s = FMA(KP951056516, T4j, T4i);
Chris@82 346 T4n = FNMS(KP559016994, T44, T43);
Chris@82 347 T4p = FMA(KP951056516, T4o, T4n);
Chris@82 348 T4v = FNMS(KP951056516, T4o, T4n);
Chris@82 349 {
Chris@82 350 E T4l, T4q, T4h, T4m;
Chris@82 351 T4h = W[2];
Chris@82 352 T4l = T4h * T4k;
Chris@82 353 T4q = T4h * T4p;
Chris@82 354 T4m = W[3];
Chris@82 355 rio[WS(vs, 2) + WS(rs, 4)] = FMA(T4m, T4p, T4l);
Chris@82 356 iio[WS(vs, 2) + WS(rs, 4)] = FNMS(T4m, T4k, T4q);
Chris@82 357 }
Chris@82 358 {
Chris@82 359 E T4t, T4w, T4r, T4u;
Chris@82 360 T4r = W[4];
Chris@82 361 T4t = T4r * T4s;
Chris@82 362 T4w = T4r * T4v;
Chris@82 363 T4u = W[5];
Chris@82 364 rio[WS(vs, 3) + WS(rs, 4)] = FMA(T4u, T4v, T4t);
Chris@82 365 iio[WS(vs, 3) + WS(rs, 4)] = FNMS(T4u, T4s, T4w);
Chris@82 366 }
Chris@82 367 }
Chris@82 368 {
Chris@82 369 E T28, T2o, T2l, T2r, T20, T2h;
Chris@82 370 T20 = FMA(KP559016994, T1Z, T1Y);
Chris@82 371 T28 = FMA(KP951056516, T27, T20);
Chris@82 372 T2o = FNMS(KP951056516, T27, T20);
Chris@82 373 T2h = FMA(KP559016994, T2g, T2f);
Chris@82 374 T2l = FNMS(KP951056516, T2k, T2h);
Chris@82 375 T2r = FMA(KP951056516, T2k, T2h);
Chris@82 376 {
Chris@82 377 E T29, T2m, T1X, T2a;
Chris@82 378 T1X = W[0];
Chris@82 379 T29 = T1X * T28;
Chris@82 380 T2m = T1X * T2l;
Chris@82 381 T2a = W[1];
Chris@82 382 rio[WS(vs, 1) + WS(rs, 2)] = FMA(T2a, T2l, T29);
Chris@82 383 iio[WS(vs, 1) + WS(rs, 2)] = FNMS(T2a, T28, T2m);
Chris@82 384 }
Chris@82 385 {
Chris@82 386 E T2p, T2s, T2n, T2q;
Chris@82 387 T2n = W[6];
Chris@82 388 T2p = T2n * T2o;
Chris@82 389 T2s = T2n * T2r;
Chris@82 390 T2q = W[7];
Chris@82 391 rio[WS(vs, 4) + WS(rs, 2)] = FMA(T2q, T2r, T2p);
Chris@82 392 iio[WS(vs, 4) + WS(rs, 2)] = FNMS(T2q, T2o, T2s);
Chris@82 393 }
Chris@82 394 }
Chris@82 395 {
Chris@82 396 E T32, T3i, T3f, T3l, T2U, T3b;
Chris@82 397 T2U = FMA(KP559016994, T2T, T2S);
Chris@82 398 T32 = FMA(KP951056516, T31, T2U);
Chris@82 399 T3i = FNMS(KP951056516, T31, T2U);
Chris@82 400 T3b = FMA(KP559016994, T3a, T39);
Chris@82 401 T3f = FNMS(KP951056516, T3e, T3b);
Chris@82 402 T3l = FMA(KP951056516, T3e, T3b);
Chris@82 403 {
Chris@82 404 E T33, T3g, T2R, T34;
Chris@82 405 T2R = W[0];
Chris@82 406 T33 = T2R * T32;
Chris@82 407 T3g = T2R * T3f;
Chris@82 408 T34 = W[1];
Chris@82 409 rio[WS(vs, 1) + WS(rs, 3)] = FMA(T34, T3f, T33);
Chris@82 410 iio[WS(vs, 1) + WS(rs, 3)] = FNMS(T34, T32, T3g);
Chris@82 411 }
Chris@82 412 {
Chris@82 413 E T3j, T3m, T3h, T3k;
Chris@82 414 T3h = W[6];
Chris@82 415 T3j = T3h * T3i;
Chris@82 416 T3m = T3h * T3l;
Chris@82 417 T3k = W[7];
Chris@82 418 rio[WS(vs, 4) + WS(rs, 3)] = FMA(T3k, T3l, T3j);
Chris@82 419 iio[WS(vs, 4) + WS(rs, 3)] = FNMS(T3k, T3i, T3m);
Chris@82 420 }
Chris@82 421 }
Chris@82 422 {
Chris@82 423 E T3q, T3y, T3v, T3B, T3o, T3t;
Chris@82 424 T3o = FNMS(KP559016994, T2T, T2S);
Chris@82 425 T3q = FNMS(KP951056516, T3p, T3o);
Chris@82 426 T3y = FMA(KP951056516, T3p, T3o);
Chris@82 427 T3t = FNMS(KP559016994, T3a, T39);
Chris@82 428 T3v = FMA(KP951056516, T3u, T3t);
Chris@82 429 T3B = FNMS(KP951056516, T3u, T3t);
Chris@82 430 {
Chris@82 431 E T3r, T3w, T3n, T3s;
Chris@82 432 T3n = W[2];
Chris@82 433 T3r = T3n * T3q;
Chris@82 434 T3w = T3n * T3v;
Chris@82 435 T3s = W[3];
Chris@82 436 rio[WS(vs, 2) + WS(rs, 3)] = FMA(T3s, T3v, T3r);
Chris@82 437 iio[WS(vs, 2) + WS(rs, 3)] = FNMS(T3s, T3q, T3w);
Chris@82 438 }
Chris@82 439 {
Chris@82 440 E T3z, T3C, T3x, T3A;
Chris@82 441 T3x = W[4];
Chris@82 442 T3z = T3x * T3y;
Chris@82 443 T3C = T3x * T3B;
Chris@82 444 T3A = W[5];
Chris@82 445 rio[WS(vs, 3) + WS(rs, 3)] = FMA(T3A, T3B, T3z);
Chris@82 446 iio[WS(vs, 3) + WS(rs, 3)] = FNMS(T3A, T3y, T3C);
Chris@82 447 }
Chris@82 448 }
Chris@82 449 {
Chris@82 450 E T3W, T4c, T49, T4f, T3O, T45;
Chris@82 451 T3O = FMA(KP559016994, T3N, T3M);
Chris@82 452 T3W = FMA(KP951056516, T3V, T3O);
Chris@82 453 T4c = FNMS(KP951056516, T3V, T3O);
Chris@82 454 T45 = FMA(KP559016994, T44, T43);
Chris@82 455 T49 = FNMS(KP951056516, T48, T45);
Chris@82 456 T4f = FMA(KP951056516, T48, T45);
Chris@82 457 {
Chris@82 458 E T3X, T4a, T3L, T3Y;
Chris@82 459 T3L = W[0];
Chris@82 460 T3X = T3L * T3W;
Chris@82 461 T4a = T3L * T49;
Chris@82 462 T3Y = W[1];
Chris@82 463 rio[WS(vs, 1) + WS(rs, 4)] = FMA(T3Y, T49, T3X);
Chris@82 464 iio[WS(vs, 1) + WS(rs, 4)] = FNMS(T3Y, T3W, T4a);
Chris@82 465 }
Chris@82 466 {
Chris@82 467 E T4d, T4g, T4b, T4e;
Chris@82 468 T4b = W[6];
Chris@82 469 T4d = T4b * T4c;
Chris@82 470 T4g = T4b * T4f;
Chris@82 471 T4e = W[7];
Chris@82 472 rio[WS(vs, 4) + WS(rs, 4)] = FMA(T4e, T4f, T4d);
Chris@82 473 iio[WS(vs, 4) + WS(rs, 4)] = FNMS(T4e, T4c, T4g);
Chris@82 474 }
Chris@82 475 }
Chris@82 476 {
Chris@82 477 E T1C, T1K, T1H, T1N, T1A, T1F;
Chris@82 478 T1A = FNMS(KP559016994, T15, T14);
Chris@82 479 T1C = FNMS(KP951056516, T1B, T1A);
Chris@82 480 T1K = FMA(KP951056516, T1B, T1A);
Chris@82 481 T1F = FNMS(KP559016994, T1m, T1l);
Chris@82 482 T1H = FMA(KP951056516, T1G, T1F);
Chris@82 483 T1N = FNMS(KP951056516, T1G, T1F);
Chris@82 484 {
Chris@82 485 E T1D, T1I, T1z, T1E;
Chris@82 486 T1z = W[2];
Chris@82 487 T1D = T1z * T1C;
Chris@82 488 T1I = T1z * T1H;
Chris@82 489 T1E = W[3];
Chris@82 490 rio[WS(vs, 2) + WS(rs, 1)] = FMA(T1E, T1H, T1D);
Chris@82 491 iio[WS(vs, 2) + WS(rs, 1)] = FNMS(T1E, T1C, T1I);
Chris@82 492 }
Chris@82 493 {
Chris@82 494 E T1L, T1O, T1J, T1M;
Chris@82 495 T1J = W[4];
Chris@82 496 T1L = T1J * T1K;
Chris@82 497 T1O = T1J * T1N;
Chris@82 498 T1M = W[5];
Chris@82 499 rio[WS(vs, 3) + WS(rs, 1)] = FMA(T1M, T1N, T1L);
Chris@82 500 iio[WS(vs, 3) + WS(rs, 1)] = FNMS(T1M, T1K, T1O);
Chris@82 501 }
Chris@82 502 }
Chris@82 503 {
Chris@82 504 E T1e, T1u, T1r, T1x, T16, T1n;
Chris@82 505 T16 = FMA(KP559016994, T15, T14);
Chris@82 506 T1e = FMA(KP951056516, T1d, T16);
Chris@82 507 T1u = FNMS(KP951056516, T1d, T16);
Chris@82 508 T1n = FMA(KP559016994, T1m, T1l);
Chris@82 509 T1r = FNMS(KP951056516, T1q, T1n);
Chris@82 510 T1x = FMA(KP951056516, T1q, T1n);
Chris@82 511 {
Chris@82 512 E T1f, T1s, T13, T1g;
Chris@82 513 T13 = W[0];
Chris@82 514 T1f = T13 * T1e;
Chris@82 515 T1s = T13 * T1r;
Chris@82 516 T1g = W[1];
Chris@82 517 rio[WS(vs, 1) + WS(rs, 1)] = FMA(T1g, T1r, T1f);
Chris@82 518 iio[WS(vs, 1) + WS(rs, 1)] = FNMS(T1g, T1e, T1s);
Chris@82 519 }
Chris@82 520 {
Chris@82 521 E T1v, T1y, T1t, T1w;
Chris@82 522 T1t = W[6];
Chris@82 523 T1v = T1t * T1u;
Chris@82 524 T1y = T1t * T1x;
Chris@82 525 T1w = W[7];
Chris@82 526 rio[WS(vs, 4) + WS(rs, 1)] = FMA(T1w, T1x, T1v);
Chris@82 527 iio[WS(vs, 4) + WS(rs, 1)] = FNMS(T1w, T1u, T1y);
Chris@82 528 }
Chris@82 529 }
Chris@82 530 }
Chris@82 531 }
Chris@82 532 }
Chris@82 533
Chris@82 534 static const tw_instr twinstr[] = {
Chris@82 535 {TW_FULL, 0, 5},
Chris@82 536 {TW_NEXT, 1, 0}
Chris@82 537 };
Chris@82 538
Chris@82 539 static const ct_desc desc = { 5, "q1_5", twinstr, &GENUS, {70, 40, 130, 0}, 0, 0, 0 };
Chris@82 540
Chris@82 541 void X(codelet_q1_5) (planner *p) {
Chris@82 542 X(kdft_difsq_register) (p, q1_5, &desc);
Chris@82 543 }
Chris@82 544 #else
Chris@82 545
Chris@82 546 /* Generated by: ../../../genfft/gen_twidsq.native -compact -variables 4 -pipeline-latency 4 -reload-twiddle -dif -n 5 -name q1_5 -include dft/scalar/q.h */
Chris@82 547
Chris@82 548 /*
Chris@82 549 * This function contains 200 FP additions, 140 FP multiplications,
Chris@82 550 * (or, 130 additions, 70 multiplications, 70 fused multiply/add),
Chris@82 551 * 75 stack variables, 4 constants, and 100 memory accesses
Chris@82 552 */
Chris@82 553 #include "dft/scalar/q.h"
Chris@82 554
Chris@82 555 static void q1_5(R *rio, R *iio, const R *W, stride rs, stride vs, INT mb, INT me, INT ms)
Chris@82 556 {
Chris@82 557 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
Chris@82 558 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
Chris@82 559 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
Chris@82 560 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
Chris@82 561 {
Chris@82 562 INT m;
Chris@82 563 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@82 564 E T1, Ta, TG, Tv, T8, Tb, Tp, Tj, TD, To, Tq, Tr, TN, TW, T1s;
Chris@82 565 E T1h, TU, TX, T1b, T15, T1p, T1a, T1c, T1d, T1z, T1I, T2e, T23, T1G, T1J;
Chris@82 566 E T1X, T1R, T2b, T1W, T1Y, T1Z, T3v, T3p, T3J, T3u, T3w, T3x, T37, T3g, T3M;
Chris@82 567 E T3B, T3e, T3h, T2l, T2u, T30, T2P, T2s, T2v, T2J, T2D, T2X, T2I, T2K, T2L;
Chris@82 568 {
Chris@82 569 E T7, Tu, T4, Tt;
Chris@82 570 T1 = rio[0];
Chris@82 571 {
Chris@82 572 E T5, T6, T2, T3;
Chris@82 573 T5 = rio[WS(rs, 2)];
Chris@82 574 T6 = rio[WS(rs, 3)];
Chris@82 575 T7 = T5 + T6;
Chris@82 576 Tu = T5 - T6;
Chris@82 577 T2 = rio[WS(rs, 1)];
Chris@82 578 T3 = rio[WS(rs, 4)];
Chris@82 579 T4 = T2 + T3;
Chris@82 580 Tt = T2 - T3;
Chris@82 581 }
Chris@82 582 Ta = KP559016994 * (T4 - T7);
Chris@82 583 TG = FNMS(KP587785252, Tt, KP951056516 * Tu);
Chris@82 584 Tv = FMA(KP951056516, Tt, KP587785252 * Tu);
Chris@82 585 T8 = T4 + T7;
Chris@82 586 Tb = FNMS(KP250000000, T8, T1);
Chris@82 587 }
Chris@82 588 {
Chris@82 589 E Ti, Tn, Tf, Tm;
Chris@82 590 Tp = iio[0];
Chris@82 591 {
Chris@82 592 E Tg, Th, Td, Te;
Chris@82 593 Tg = iio[WS(rs, 2)];
Chris@82 594 Th = iio[WS(rs, 3)];
Chris@82 595 Ti = Tg - Th;
Chris@82 596 Tn = Tg + Th;
Chris@82 597 Td = iio[WS(rs, 1)];
Chris@82 598 Te = iio[WS(rs, 4)];
Chris@82 599 Tf = Td - Te;
Chris@82 600 Tm = Td + Te;
Chris@82 601 }
Chris@82 602 Tj = FMA(KP951056516, Tf, KP587785252 * Ti);
Chris@82 603 TD = FNMS(KP587785252, Tf, KP951056516 * Ti);
Chris@82 604 To = KP559016994 * (Tm - Tn);
Chris@82 605 Tq = Tm + Tn;
Chris@82 606 Tr = FNMS(KP250000000, Tq, Tp);
Chris@82 607 }
Chris@82 608 {
Chris@82 609 E TT, T1g, TQ, T1f;
Chris@82 610 TN = rio[WS(vs, 1)];
Chris@82 611 {
Chris@82 612 E TR, TS, TO, TP;
Chris@82 613 TR = rio[WS(vs, 1) + WS(rs, 2)];
Chris@82 614 TS = rio[WS(vs, 1) + WS(rs, 3)];
Chris@82 615 TT = TR + TS;
Chris@82 616 T1g = TR - TS;
Chris@82 617 TO = rio[WS(vs, 1) + WS(rs, 1)];
Chris@82 618 TP = rio[WS(vs, 1) + WS(rs, 4)];
Chris@82 619 TQ = TO + TP;
Chris@82 620 T1f = TO - TP;
Chris@82 621 }
Chris@82 622 TW = KP559016994 * (TQ - TT);
Chris@82 623 T1s = FNMS(KP587785252, T1f, KP951056516 * T1g);
Chris@82 624 T1h = FMA(KP951056516, T1f, KP587785252 * T1g);
Chris@82 625 TU = TQ + TT;
Chris@82 626 TX = FNMS(KP250000000, TU, TN);
Chris@82 627 }
Chris@82 628 {
Chris@82 629 E T14, T19, T11, T18;
Chris@82 630 T1b = iio[WS(vs, 1)];
Chris@82 631 {
Chris@82 632 E T12, T13, TZ, T10;
Chris@82 633 T12 = iio[WS(vs, 1) + WS(rs, 2)];
Chris@82 634 T13 = iio[WS(vs, 1) + WS(rs, 3)];
Chris@82 635 T14 = T12 - T13;
Chris@82 636 T19 = T12 + T13;
Chris@82 637 TZ = iio[WS(vs, 1) + WS(rs, 1)];
Chris@82 638 T10 = iio[WS(vs, 1) + WS(rs, 4)];
Chris@82 639 T11 = TZ - T10;
Chris@82 640 T18 = TZ + T10;
Chris@82 641 }
Chris@82 642 T15 = FMA(KP951056516, T11, KP587785252 * T14);
Chris@82 643 T1p = FNMS(KP587785252, T11, KP951056516 * T14);
Chris@82 644 T1a = KP559016994 * (T18 - T19);
Chris@82 645 T1c = T18 + T19;
Chris@82 646 T1d = FNMS(KP250000000, T1c, T1b);
Chris@82 647 }
Chris@82 648 {
Chris@82 649 E T1F, T22, T1C, T21;
Chris@82 650 T1z = rio[WS(vs, 2)];
Chris@82 651 {
Chris@82 652 E T1D, T1E, T1A, T1B;
Chris@82 653 T1D = rio[WS(vs, 2) + WS(rs, 2)];
Chris@82 654 T1E = rio[WS(vs, 2) + WS(rs, 3)];
Chris@82 655 T1F = T1D + T1E;
Chris@82 656 T22 = T1D - T1E;
Chris@82 657 T1A = rio[WS(vs, 2) + WS(rs, 1)];
Chris@82 658 T1B = rio[WS(vs, 2) + WS(rs, 4)];
Chris@82 659 T1C = T1A + T1B;
Chris@82 660 T21 = T1A - T1B;
Chris@82 661 }
Chris@82 662 T1I = KP559016994 * (T1C - T1F);
Chris@82 663 T2e = FNMS(KP587785252, T21, KP951056516 * T22);
Chris@82 664 T23 = FMA(KP951056516, T21, KP587785252 * T22);
Chris@82 665 T1G = T1C + T1F;
Chris@82 666 T1J = FNMS(KP250000000, T1G, T1z);
Chris@82 667 }
Chris@82 668 {
Chris@82 669 E T1Q, T1V, T1N, T1U;
Chris@82 670 T1X = iio[WS(vs, 2)];
Chris@82 671 {
Chris@82 672 E T1O, T1P, T1L, T1M;
Chris@82 673 T1O = iio[WS(vs, 2) + WS(rs, 2)];
Chris@82 674 T1P = iio[WS(vs, 2) + WS(rs, 3)];
Chris@82 675 T1Q = T1O - T1P;
Chris@82 676 T1V = T1O + T1P;
Chris@82 677 T1L = iio[WS(vs, 2) + WS(rs, 1)];
Chris@82 678 T1M = iio[WS(vs, 2) + WS(rs, 4)];
Chris@82 679 T1N = T1L - T1M;
Chris@82 680 T1U = T1L + T1M;
Chris@82 681 }
Chris@82 682 T1R = FMA(KP951056516, T1N, KP587785252 * T1Q);
Chris@82 683 T2b = FNMS(KP587785252, T1N, KP951056516 * T1Q);
Chris@82 684 T1W = KP559016994 * (T1U - T1V);
Chris@82 685 T1Y = T1U + T1V;
Chris@82 686 T1Z = FNMS(KP250000000, T1Y, T1X);
Chris@82 687 }
Chris@82 688 {
Chris@82 689 E T3o, T3t, T3l, T3s;
Chris@82 690 T3v = iio[WS(vs, 4)];
Chris@82 691 {
Chris@82 692 E T3m, T3n, T3j, T3k;
Chris@82 693 T3m = iio[WS(vs, 4) + WS(rs, 2)];
Chris@82 694 T3n = iio[WS(vs, 4) + WS(rs, 3)];
Chris@82 695 T3o = T3m - T3n;
Chris@82 696 T3t = T3m + T3n;
Chris@82 697 T3j = iio[WS(vs, 4) + WS(rs, 1)];
Chris@82 698 T3k = iio[WS(vs, 4) + WS(rs, 4)];
Chris@82 699 T3l = T3j - T3k;
Chris@82 700 T3s = T3j + T3k;
Chris@82 701 }
Chris@82 702 T3p = FMA(KP951056516, T3l, KP587785252 * T3o);
Chris@82 703 T3J = FNMS(KP587785252, T3l, KP951056516 * T3o);
Chris@82 704 T3u = KP559016994 * (T3s - T3t);
Chris@82 705 T3w = T3s + T3t;
Chris@82 706 T3x = FNMS(KP250000000, T3w, T3v);
Chris@82 707 }
Chris@82 708 {
Chris@82 709 E T3d, T3A, T3a, T3z;
Chris@82 710 T37 = rio[WS(vs, 4)];
Chris@82 711 {
Chris@82 712 E T3b, T3c, T38, T39;
Chris@82 713 T3b = rio[WS(vs, 4) + WS(rs, 2)];
Chris@82 714 T3c = rio[WS(vs, 4) + WS(rs, 3)];
Chris@82 715 T3d = T3b + T3c;
Chris@82 716 T3A = T3b - T3c;
Chris@82 717 T38 = rio[WS(vs, 4) + WS(rs, 1)];
Chris@82 718 T39 = rio[WS(vs, 4) + WS(rs, 4)];
Chris@82 719 T3a = T38 + T39;
Chris@82 720 T3z = T38 - T39;
Chris@82 721 }
Chris@82 722 T3g = KP559016994 * (T3a - T3d);
Chris@82 723 T3M = FNMS(KP587785252, T3z, KP951056516 * T3A);
Chris@82 724 T3B = FMA(KP951056516, T3z, KP587785252 * T3A);
Chris@82 725 T3e = T3a + T3d;
Chris@82 726 T3h = FNMS(KP250000000, T3e, T37);
Chris@82 727 }
Chris@82 728 {
Chris@82 729 E T2r, T2O, T2o, T2N;
Chris@82 730 T2l = rio[WS(vs, 3)];
Chris@82 731 {
Chris@82 732 E T2p, T2q, T2m, T2n;
Chris@82 733 T2p = rio[WS(vs, 3) + WS(rs, 2)];
Chris@82 734 T2q = rio[WS(vs, 3) + WS(rs, 3)];
Chris@82 735 T2r = T2p + T2q;
Chris@82 736 T2O = T2p - T2q;
Chris@82 737 T2m = rio[WS(vs, 3) + WS(rs, 1)];
Chris@82 738 T2n = rio[WS(vs, 3) + WS(rs, 4)];
Chris@82 739 T2o = T2m + T2n;
Chris@82 740 T2N = T2m - T2n;
Chris@82 741 }
Chris@82 742 T2u = KP559016994 * (T2o - T2r);
Chris@82 743 T30 = FNMS(KP587785252, T2N, KP951056516 * T2O);
Chris@82 744 T2P = FMA(KP951056516, T2N, KP587785252 * T2O);
Chris@82 745 T2s = T2o + T2r;
Chris@82 746 T2v = FNMS(KP250000000, T2s, T2l);
Chris@82 747 }
Chris@82 748 {
Chris@82 749 E T2C, T2H, T2z, T2G;
Chris@82 750 T2J = iio[WS(vs, 3)];
Chris@82 751 {
Chris@82 752 E T2A, T2B, T2x, T2y;
Chris@82 753 T2A = iio[WS(vs, 3) + WS(rs, 2)];
Chris@82 754 T2B = iio[WS(vs, 3) + WS(rs, 3)];
Chris@82 755 T2C = T2A - T2B;
Chris@82 756 T2H = T2A + T2B;
Chris@82 757 T2x = iio[WS(vs, 3) + WS(rs, 1)];
Chris@82 758 T2y = iio[WS(vs, 3) + WS(rs, 4)];
Chris@82 759 T2z = T2x - T2y;
Chris@82 760 T2G = T2x + T2y;
Chris@82 761 }
Chris@82 762 T2D = FMA(KP951056516, T2z, KP587785252 * T2C);
Chris@82 763 T2X = FNMS(KP587785252, T2z, KP951056516 * T2C);
Chris@82 764 T2I = KP559016994 * (T2G - T2H);
Chris@82 765 T2K = T2G + T2H;
Chris@82 766 T2L = FNMS(KP250000000, T2K, T2J);
Chris@82 767 }
Chris@82 768 rio[0] = T1 + T8;
Chris@82 769 iio[0] = Tp + Tq;
Chris@82 770 rio[WS(rs, 1)] = TN + TU;
Chris@82 771 iio[WS(rs, 1)] = T1b + T1c;
Chris@82 772 rio[WS(rs, 2)] = T1z + T1G;
Chris@82 773 iio[WS(rs, 2)] = T1X + T1Y;
Chris@82 774 iio[WS(rs, 4)] = T3v + T3w;
Chris@82 775 rio[WS(rs, 4)] = T37 + T3e;
Chris@82 776 rio[WS(rs, 3)] = T2l + T2s;
Chris@82 777 iio[WS(rs, 3)] = T2J + T2K;
Chris@82 778 {
Chris@82 779 E Tk, Ty, Tw, TA, Tc, Ts;
Chris@82 780 Tc = Ta + Tb;
Chris@82 781 Tk = Tc + Tj;
Chris@82 782 Ty = Tc - Tj;
Chris@82 783 Ts = To + Tr;
Chris@82 784 Tw = Ts - Tv;
Chris@82 785 TA = Tv + Ts;
Chris@82 786 {
Chris@82 787 E T9, Tl, Tx, Tz;
Chris@82 788 T9 = W[0];
Chris@82 789 Tl = W[1];
Chris@82 790 rio[WS(vs, 1)] = FMA(T9, Tk, Tl * Tw);
Chris@82 791 iio[WS(vs, 1)] = FNMS(Tl, Tk, T9 * Tw);
Chris@82 792 Tx = W[6];
Chris@82 793 Tz = W[7];
Chris@82 794 rio[WS(vs, 4)] = FMA(Tx, Ty, Tz * TA);
Chris@82 795 iio[WS(vs, 4)] = FNMS(Tz, Ty, Tx * TA);
Chris@82 796 }
Chris@82 797 }
Chris@82 798 {
Chris@82 799 E TE, TK, TI, TM, TC, TH;
Chris@82 800 TC = Tb - Ta;
Chris@82 801 TE = TC - TD;
Chris@82 802 TK = TC + TD;
Chris@82 803 TH = Tr - To;
Chris@82 804 TI = TG + TH;
Chris@82 805 TM = TH - TG;
Chris@82 806 {
Chris@82 807 E TB, TF, TJ, TL;
Chris@82 808 TB = W[2];
Chris@82 809 TF = W[3];
Chris@82 810 rio[WS(vs, 2)] = FMA(TB, TE, TF * TI);
Chris@82 811 iio[WS(vs, 2)] = FNMS(TF, TE, TB * TI);
Chris@82 812 TJ = W[4];
Chris@82 813 TL = W[5];
Chris@82 814 rio[WS(vs, 3)] = FMA(TJ, TK, TL * TM);
Chris@82 815 iio[WS(vs, 3)] = FNMS(TL, TK, TJ * TM);
Chris@82 816 }
Chris@82 817 }
Chris@82 818 {
Chris@82 819 E T2c, T2i, T2g, T2k, T2a, T2f;
Chris@82 820 T2a = T1J - T1I;
Chris@82 821 T2c = T2a - T2b;
Chris@82 822 T2i = T2a + T2b;
Chris@82 823 T2f = T1Z - T1W;
Chris@82 824 T2g = T2e + T2f;
Chris@82 825 T2k = T2f - T2e;
Chris@82 826 {
Chris@82 827 E T29, T2d, T2h, T2j;
Chris@82 828 T29 = W[2];
Chris@82 829 T2d = W[3];
Chris@82 830 rio[WS(vs, 2) + WS(rs, 2)] = FMA(T29, T2c, T2d * T2g);
Chris@82 831 iio[WS(vs, 2) + WS(rs, 2)] = FNMS(T2d, T2c, T29 * T2g);
Chris@82 832 T2h = W[4];
Chris@82 833 T2j = W[5];
Chris@82 834 rio[WS(vs, 3) + WS(rs, 2)] = FMA(T2h, T2i, T2j * T2k);
Chris@82 835 iio[WS(vs, 3) + WS(rs, 2)] = FNMS(T2j, T2i, T2h * T2k);
Chris@82 836 }
Chris@82 837 }
Chris@82 838 {
Chris@82 839 E T3K, T3Q, T3O, T3S, T3I, T3N;
Chris@82 840 T3I = T3h - T3g;
Chris@82 841 T3K = T3I - T3J;
Chris@82 842 T3Q = T3I + T3J;
Chris@82 843 T3N = T3x - T3u;
Chris@82 844 T3O = T3M + T3N;
Chris@82 845 T3S = T3N - T3M;
Chris@82 846 {
Chris@82 847 E T3H, T3L, T3P, T3R;
Chris@82 848 T3H = W[2];
Chris@82 849 T3L = W[3];
Chris@82 850 rio[WS(vs, 2) + WS(rs, 4)] = FMA(T3H, T3K, T3L * T3O);
Chris@82 851 iio[WS(vs, 2) + WS(rs, 4)] = FNMS(T3L, T3K, T3H * T3O);
Chris@82 852 T3P = W[4];
Chris@82 853 T3R = W[5];
Chris@82 854 rio[WS(vs, 3) + WS(rs, 4)] = FMA(T3P, T3Q, T3R * T3S);
Chris@82 855 iio[WS(vs, 3) + WS(rs, 4)] = FNMS(T3R, T3Q, T3P * T3S);
Chris@82 856 }
Chris@82 857 }
Chris@82 858 {
Chris@82 859 E T1S, T26, T24, T28, T1K, T20;
Chris@82 860 T1K = T1I + T1J;
Chris@82 861 T1S = T1K + T1R;
Chris@82 862 T26 = T1K - T1R;
Chris@82 863 T20 = T1W + T1Z;
Chris@82 864 T24 = T20 - T23;
Chris@82 865 T28 = T23 + T20;
Chris@82 866 {
Chris@82 867 E T1H, T1T, T25, T27;
Chris@82 868 T1H = W[0];
Chris@82 869 T1T = W[1];
Chris@82 870 rio[WS(vs, 1) + WS(rs, 2)] = FMA(T1H, T1S, T1T * T24);
Chris@82 871 iio[WS(vs, 1) + WS(rs, 2)] = FNMS(T1T, T1S, T1H * T24);
Chris@82 872 T25 = W[6];
Chris@82 873 T27 = W[7];
Chris@82 874 rio[WS(vs, 4) + WS(rs, 2)] = FMA(T25, T26, T27 * T28);
Chris@82 875 iio[WS(vs, 4) + WS(rs, 2)] = FNMS(T27, T26, T25 * T28);
Chris@82 876 }
Chris@82 877 }
Chris@82 878 {
Chris@82 879 E T2E, T2S, T2Q, T2U, T2w, T2M;
Chris@82 880 T2w = T2u + T2v;
Chris@82 881 T2E = T2w + T2D;
Chris@82 882 T2S = T2w - T2D;
Chris@82 883 T2M = T2I + T2L;
Chris@82 884 T2Q = T2M - T2P;
Chris@82 885 T2U = T2P + T2M;
Chris@82 886 {
Chris@82 887 E T2t, T2F, T2R, T2T;
Chris@82 888 T2t = W[0];
Chris@82 889 T2F = W[1];
Chris@82 890 rio[WS(vs, 1) + WS(rs, 3)] = FMA(T2t, T2E, T2F * T2Q);
Chris@82 891 iio[WS(vs, 1) + WS(rs, 3)] = FNMS(T2F, T2E, T2t * T2Q);
Chris@82 892 T2R = W[6];
Chris@82 893 T2T = W[7];
Chris@82 894 rio[WS(vs, 4) + WS(rs, 3)] = FMA(T2R, T2S, T2T * T2U);
Chris@82 895 iio[WS(vs, 4) + WS(rs, 3)] = FNMS(T2T, T2S, T2R * T2U);
Chris@82 896 }
Chris@82 897 }
Chris@82 898 {
Chris@82 899 E T2Y, T34, T32, T36, T2W, T31;
Chris@82 900 T2W = T2v - T2u;
Chris@82 901 T2Y = T2W - T2X;
Chris@82 902 T34 = T2W + T2X;
Chris@82 903 T31 = T2L - T2I;
Chris@82 904 T32 = T30 + T31;
Chris@82 905 T36 = T31 - T30;
Chris@82 906 {
Chris@82 907 E T2V, T2Z, T33, T35;
Chris@82 908 T2V = W[2];
Chris@82 909 T2Z = W[3];
Chris@82 910 rio[WS(vs, 2) + WS(rs, 3)] = FMA(T2V, T2Y, T2Z * T32);
Chris@82 911 iio[WS(vs, 2) + WS(rs, 3)] = FNMS(T2Z, T2Y, T2V * T32);
Chris@82 912 T33 = W[4];
Chris@82 913 T35 = W[5];
Chris@82 914 rio[WS(vs, 3) + WS(rs, 3)] = FMA(T33, T34, T35 * T36);
Chris@82 915 iio[WS(vs, 3) + WS(rs, 3)] = FNMS(T35, T34, T33 * T36);
Chris@82 916 }
Chris@82 917 }
Chris@82 918 {
Chris@82 919 E T3q, T3E, T3C, T3G, T3i, T3y;
Chris@82 920 T3i = T3g + T3h;
Chris@82 921 T3q = T3i + T3p;
Chris@82 922 T3E = T3i - T3p;
Chris@82 923 T3y = T3u + T3x;
Chris@82 924 T3C = T3y - T3B;
Chris@82 925 T3G = T3B + T3y;
Chris@82 926 {
Chris@82 927 E T3f, T3r, T3D, T3F;
Chris@82 928 T3f = W[0];
Chris@82 929 T3r = W[1];
Chris@82 930 rio[WS(vs, 1) + WS(rs, 4)] = FMA(T3f, T3q, T3r * T3C);
Chris@82 931 iio[WS(vs, 1) + WS(rs, 4)] = FNMS(T3r, T3q, T3f * T3C);
Chris@82 932 T3D = W[6];
Chris@82 933 T3F = W[7];
Chris@82 934 rio[WS(vs, 4) + WS(rs, 4)] = FMA(T3D, T3E, T3F * T3G);
Chris@82 935 iio[WS(vs, 4) + WS(rs, 4)] = FNMS(T3F, T3E, T3D * T3G);
Chris@82 936 }
Chris@82 937 }
Chris@82 938 {
Chris@82 939 E T1q, T1w, T1u, T1y, T1o, T1t;
Chris@82 940 T1o = TX - TW;
Chris@82 941 T1q = T1o - T1p;
Chris@82 942 T1w = T1o + T1p;
Chris@82 943 T1t = T1d - T1a;
Chris@82 944 T1u = T1s + T1t;
Chris@82 945 T1y = T1t - T1s;
Chris@82 946 {
Chris@82 947 E T1n, T1r, T1v, T1x;
Chris@82 948 T1n = W[2];
Chris@82 949 T1r = W[3];
Chris@82 950 rio[WS(vs, 2) + WS(rs, 1)] = FMA(T1n, T1q, T1r * T1u);
Chris@82 951 iio[WS(vs, 2) + WS(rs, 1)] = FNMS(T1r, T1q, T1n * T1u);
Chris@82 952 T1v = W[4];
Chris@82 953 T1x = W[5];
Chris@82 954 rio[WS(vs, 3) + WS(rs, 1)] = FMA(T1v, T1w, T1x * T1y);
Chris@82 955 iio[WS(vs, 3) + WS(rs, 1)] = FNMS(T1x, T1w, T1v * T1y);
Chris@82 956 }
Chris@82 957 }
Chris@82 958 {
Chris@82 959 E T16, T1k, T1i, T1m, TY, T1e;
Chris@82 960 TY = TW + TX;
Chris@82 961 T16 = TY + T15;
Chris@82 962 T1k = TY - T15;
Chris@82 963 T1e = T1a + T1d;
Chris@82 964 T1i = T1e - T1h;
Chris@82 965 T1m = T1h + T1e;
Chris@82 966 {
Chris@82 967 E TV, T17, T1j, T1l;
Chris@82 968 TV = W[0];
Chris@82 969 T17 = W[1];
Chris@82 970 rio[WS(vs, 1) + WS(rs, 1)] = FMA(TV, T16, T17 * T1i);
Chris@82 971 iio[WS(vs, 1) + WS(rs, 1)] = FNMS(T17, T16, TV * T1i);
Chris@82 972 T1j = W[6];
Chris@82 973 T1l = W[7];
Chris@82 974 rio[WS(vs, 4) + WS(rs, 1)] = FMA(T1j, T1k, T1l * T1m);
Chris@82 975 iio[WS(vs, 4) + WS(rs, 1)] = FNMS(T1l, T1k, T1j * T1m);
Chris@82 976 }
Chris@82 977 }
Chris@82 978 }
Chris@82 979 }
Chris@82 980 }
Chris@82 981
Chris@82 982 static const tw_instr twinstr[] = {
Chris@82 983 {TW_FULL, 0, 5},
Chris@82 984 {TW_NEXT, 1, 0}
Chris@82 985 };
Chris@82 986
Chris@82 987 static const ct_desc desc = { 5, "q1_5", twinstr, &GENUS, {130, 70, 70, 0}, 0, 0, 0 };
Chris@82 988
Chris@82 989 void X(codelet_q1_5) (planner *p) {
Chris@82 990 X(kdft_difsq_register) (p, q1_5, &desc);
Chris@82 991 }
Chris@82 992 #endif