annotate src/fftw-3.3.5/dft/scalar/codelets/q1_5.c @ 127:7867fa7e1b6b

Current fftw source
author Chris Cannam <cannam@all-day-breakfast.com>
date Tue, 18 Oct 2016 13:40:26 +0100
parents
children
rev   line source
cannam@127 1 /*
cannam@127 2 * Copyright (c) 2003, 2007-14 Matteo Frigo
cannam@127 3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
cannam@127 4 *
cannam@127 5 * This program is free software; you can redistribute it and/or modify
cannam@127 6 * it under the terms of the GNU General Public License as published by
cannam@127 7 * the Free Software Foundation; either version 2 of the License, or
cannam@127 8 * (at your option) any later version.
cannam@127 9 *
cannam@127 10 * This program is distributed in the hope that it will be useful,
cannam@127 11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
cannam@127 12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
cannam@127 13 * GNU General Public License for more details.
cannam@127 14 *
cannam@127 15 * You should have received a copy of the GNU General Public License
cannam@127 16 * along with this program; if not, write to the Free Software
cannam@127 17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
cannam@127 18 *
cannam@127 19 */
cannam@127 20
cannam@127 21 /* This file was automatically generated --- DO NOT EDIT */
cannam@127 22 /* Generated on Sat Jul 30 16:37:57 EDT 2016 */
cannam@127 23
cannam@127 24 #include "codelet-dft.h"
cannam@127 25
cannam@127 26 #ifdef HAVE_FMA
cannam@127 27
cannam@127 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 */
cannam@127 29
cannam@127 30 /*
cannam@127 31 * This function contains 200 FP additions, 170 FP multiplications,
cannam@127 32 * (or, 70 additions, 40 multiplications, 130 fused multiply/add),
cannam@127 33 * 104 stack variables, 4 constants, and 100 memory accesses
cannam@127 34 */
cannam@127 35 #include "q.h"
cannam@127 36
cannam@127 37 static void q1_5(R *rio, R *iio, const R *W, stride rs, stride vs, INT mb, INT me, INT ms)
cannam@127 38 {
cannam@127 39 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
cannam@127 40 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
cannam@127 41 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
cannam@127 42 DK(KP618033988, +0.618033988749894848204586834365638117720309180);
cannam@127 43 {
cannam@127 44 INT m;
cannam@127 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)) {
cannam@127 46 E T1x, T1w, T1v;
cannam@127 47 {
cannam@127 48 E T1, Tn, TM, Tw, Tb, T8, Ta, TV, Tq, Ts, TH, Tj, Tr, T1h, T1q;
cannam@127 49 E T1G, T12, T15, T1P, T14, T1k, T1m, T1B, T1d, T1l, T2b, T2k, T2A, T1W, T1Z;
cannam@127 50 E T3Z, T1Y, T2e, T2g, T2v, T27, T2f, T3D, T42, T44, T4j, T3V, T43, T2J, T48;
cannam@127 51 E T4o, T3K, T3N, T35, T3M, T2V, T3e, T3u, T2Q, T2T, T37, T30, T2S, T2W;
cannam@127 52 {
cannam@127 53 E T1Q, T2j, T1V, T1R;
cannam@127 54 {
cannam@127 55 E Tp, Ti, Td, Te;
cannam@127 56 {
cannam@127 57 E T5, T6, T2, T3, T7, Tv;
cannam@127 58 T1 = rio[0];
cannam@127 59 T5 = rio[WS(rs, 2)];
cannam@127 60 T6 = rio[WS(rs, 3)];
cannam@127 61 T2 = rio[WS(rs, 1)];
cannam@127 62 T3 = rio[WS(rs, 4)];
cannam@127 63 Tn = iio[0];
cannam@127 64 T7 = T5 + T6;
cannam@127 65 Tv = T5 - T6;
cannam@127 66 {
cannam@127 67 E T4, Tu, Tg, Th;
cannam@127 68 T4 = T2 + T3;
cannam@127 69 Tu = T2 - T3;
cannam@127 70 Tg = iio[WS(rs, 2)];
cannam@127 71 Th = iio[WS(rs, 3)];
cannam@127 72 TM = FNMS(KP618033988, Tu, Tv);
cannam@127 73 Tw = FMA(KP618033988, Tv, Tu);
cannam@127 74 Tb = T4 - T7;
cannam@127 75 T8 = T4 + T7;
cannam@127 76 Tp = Tg + Th;
cannam@127 77 Ti = Tg - Th;
cannam@127 78 Ta = FNMS(KP250000000, T8, T1);
cannam@127 79 Td = iio[WS(rs, 1)];
cannam@127 80 Te = iio[WS(rs, 4)];
cannam@127 81 }
cannam@127 82 }
cannam@127 83 {
cannam@127 84 E TW, T1p, T11, TX;
cannam@127 85 TV = rio[WS(vs, 1)];
cannam@127 86 {
cannam@127 87 E TZ, T10, Tf, To;
cannam@127 88 TZ = rio[WS(vs, 1) + WS(rs, 2)];
cannam@127 89 T10 = rio[WS(vs, 1) + WS(rs, 3)];
cannam@127 90 Tf = Td - Te;
cannam@127 91 To = Td + Te;
cannam@127 92 TW = rio[WS(vs, 1) + WS(rs, 1)];
cannam@127 93 T1p = TZ - T10;
cannam@127 94 T11 = TZ + T10;
cannam@127 95 Tq = To + Tp;
cannam@127 96 Ts = To - Tp;
cannam@127 97 TH = FNMS(KP618033988, Tf, Ti);
cannam@127 98 Tj = FMA(KP618033988, Ti, Tf);
cannam@127 99 Tr = FNMS(KP250000000, Tq, Tn);
cannam@127 100 TX = rio[WS(vs, 1) + WS(rs, 4)];
cannam@127 101 }
cannam@127 102 {
cannam@127 103 E T17, T1j, T1c, T18;
cannam@127 104 T1h = iio[WS(vs, 1)];
cannam@127 105 {
cannam@127 106 E T1a, T1b, TY, T1o;
cannam@127 107 T1a = iio[WS(vs, 1) + WS(rs, 2)];
cannam@127 108 T1b = iio[WS(vs, 1) + WS(rs, 3)];
cannam@127 109 TY = TW + TX;
cannam@127 110 T1o = TW - TX;
cannam@127 111 T17 = iio[WS(vs, 1) + WS(rs, 1)];
cannam@127 112 T1j = T1a + T1b;
cannam@127 113 T1c = T1a - T1b;
cannam@127 114 T1q = FMA(KP618033988, T1p, T1o);
cannam@127 115 T1G = FNMS(KP618033988, T1o, T1p);
cannam@127 116 T12 = TY + T11;
cannam@127 117 T15 = TY - T11;
cannam@127 118 T18 = iio[WS(vs, 1) + WS(rs, 4)];
cannam@127 119 }
cannam@127 120 T1P = rio[WS(vs, 2)];
cannam@127 121 T14 = FNMS(KP250000000, T12, TV);
cannam@127 122 {
cannam@127 123 E T1T, T1i, T19, T1U;
cannam@127 124 T1T = rio[WS(vs, 2) + WS(rs, 2)];
cannam@127 125 T1i = T17 + T18;
cannam@127 126 T19 = T17 - T18;
cannam@127 127 T1U = rio[WS(vs, 2) + WS(rs, 3)];
cannam@127 128 T1Q = rio[WS(vs, 2) + WS(rs, 1)];
cannam@127 129 T1k = T1i + T1j;
cannam@127 130 T1m = T1i - T1j;
cannam@127 131 T1B = FNMS(KP618033988, T19, T1c);
cannam@127 132 T1d = FMA(KP618033988, T1c, T19);
cannam@127 133 T2j = T1T - T1U;
cannam@127 134 T1V = T1T + T1U;
cannam@127 135 T1l = FNMS(KP250000000, T1k, T1h);
cannam@127 136 T1R = rio[WS(vs, 2) + WS(rs, 4)];
cannam@127 137 }
cannam@127 138 }
cannam@127 139 }
cannam@127 140 }
cannam@127 141 {
cannam@127 142 E T3P, T41, T3U, T3Q;
cannam@127 143 {
cannam@127 144 E T21, T2d, T26, T22;
cannam@127 145 T2b = iio[WS(vs, 2)];
cannam@127 146 {
cannam@127 147 E T24, T25, T1S, T2i;
cannam@127 148 T24 = iio[WS(vs, 2) + WS(rs, 2)];
cannam@127 149 T25 = iio[WS(vs, 2) + WS(rs, 3)];
cannam@127 150 T1S = T1Q + T1R;
cannam@127 151 T2i = T1Q - T1R;
cannam@127 152 T21 = iio[WS(vs, 2) + WS(rs, 1)];
cannam@127 153 T2d = T24 + T25;
cannam@127 154 T26 = T24 - T25;
cannam@127 155 T2k = FMA(KP618033988, T2j, T2i);
cannam@127 156 T2A = FNMS(KP618033988, T2i, T2j);
cannam@127 157 T1W = T1S + T1V;
cannam@127 158 T1Z = T1S - T1V;
cannam@127 159 T22 = iio[WS(vs, 2) + WS(rs, 4)];
cannam@127 160 }
cannam@127 161 T3Z = iio[WS(vs, 4)];
cannam@127 162 T1Y = FNMS(KP250000000, T1W, T1P);
cannam@127 163 {
cannam@127 164 E T3S, T2c, T23, T3T;
cannam@127 165 T3S = iio[WS(vs, 4) + WS(rs, 2)];
cannam@127 166 T2c = T21 + T22;
cannam@127 167 T23 = T21 - T22;
cannam@127 168 T3T = iio[WS(vs, 4) + WS(rs, 3)];
cannam@127 169 T3P = iio[WS(vs, 4) + WS(rs, 1)];
cannam@127 170 T2e = T2c + T2d;
cannam@127 171 T2g = T2c - T2d;
cannam@127 172 T2v = FNMS(KP618033988, T23, T26);
cannam@127 173 T27 = FMA(KP618033988, T26, T23);
cannam@127 174 T41 = T3S + T3T;
cannam@127 175 T3U = T3S - T3T;
cannam@127 176 T2f = FNMS(KP250000000, T2e, T2b);
cannam@127 177 T3Q = iio[WS(vs, 4) + WS(rs, 4)];
cannam@127 178 }
cannam@127 179 }
cannam@127 180 {
cannam@127 181 E T3E, T47, T3J, T3F;
cannam@127 182 T3D = rio[WS(vs, 4)];
cannam@127 183 {
cannam@127 184 E T3H, T3I, T3R, T40;
cannam@127 185 T3H = rio[WS(vs, 4) + WS(rs, 2)];
cannam@127 186 T3I = rio[WS(vs, 4) + WS(rs, 3)];
cannam@127 187 T3R = T3P - T3Q;
cannam@127 188 T40 = T3P + T3Q;
cannam@127 189 T3E = rio[WS(vs, 4) + WS(rs, 1)];
cannam@127 190 T47 = T3H - T3I;
cannam@127 191 T3J = T3H + T3I;
cannam@127 192 T42 = T40 + T41;
cannam@127 193 T44 = T40 - T41;
cannam@127 194 T4j = FNMS(KP618033988, T3R, T3U);
cannam@127 195 T3V = FMA(KP618033988, T3U, T3R);
cannam@127 196 T43 = FNMS(KP250000000, T42, T3Z);
cannam@127 197 T3F = rio[WS(vs, 4) + WS(rs, 4)];
cannam@127 198 }
cannam@127 199 {
cannam@127 200 E T2K, T3d, T2P, T2L;
cannam@127 201 T2J = rio[WS(vs, 3)];
cannam@127 202 {
cannam@127 203 E T2N, T2O, T3G, T46;
cannam@127 204 T2N = rio[WS(vs, 3) + WS(rs, 2)];
cannam@127 205 T2O = rio[WS(vs, 3) + WS(rs, 3)];
cannam@127 206 T3G = T3E + T3F;
cannam@127 207 T46 = T3E - T3F;
cannam@127 208 T2K = rio[WS(vs, 3) + WS(rs, 1)];
cannam@127 209 T3d = T2N - T2O;
cannam@127 210 T2P = T2N + T2O;
cannam@127 211 T48 = FMA(KP618033988, T47, T46);
cannam@127 212 T4o = FNMS(KP618033988, T46, T47);
cannam@127 213 T3K = T3G + T3J;
cannam@127 214 T3N = T3G - T3J;
cannam@127 215 T2L = rio[WS(vs, 3) + WS(rs, 4)];
cannam@127 216 }
cannam@127 217 T35 = iio[WS(vs, 3)];
cannam@127 218 T3M = FNMS(KP250000000, T3K, T3D);
cannam@127 219 {
cannam@127 220 E T2Y, T3c, T2M, T2Z;
cannam@127 221 T2Y = iio[WS(vs, 3) + WS(rs, 2)];
cannam@127 222 T3c = T2K - T2L;
cannam@127 223 T2M = T2K + T2L;
cannam@127 224 T2Z = iio[WS(vs, 3) + WS(rs, 3)];
cannam@127 225 T2V = iio[WS(vs, 3) + WS(rs, 1)];
cannam@127 226 T3e = FMA(KP618033988, T3d, T3c);
cannam@127 227 T3u = FNMS(KP618033988, T3c, T3d);
cannam@127 228 T2Q = T2M + T2P;
cannam@127 229 T2T = T2M - T2P;
cannam@127 230 T37 = T2Y + T2Z;
cannam@127 231 T30 = T2Y - T2Z;
cannam@127 232 T2S = FNMS(KP250000000, T2Q, T2J);
cannam@127 233 T2W = iio[WS(vs, 3) + WS(rs, 4)];
cannam@127 234 }
cannam@127 235 }
cannam@127 236 }
cannam@127 237 }
cannam@127 238 }
cannam@127 239 {
cannam@127 240 E T3a, T31, T3p, T39, T2X, T36, T38;
cannam@127 241 rio[0] = T1 + T8;
cannam@127 242 iio[0] = Tn + Tq;
cannam@127 243 rio[WS(rs, 1)] = TV + T12;
cannam@127 244 T2X = T2V - T2W;
cannam@127 245 T36 = T2V + T2W;
cannam@127 246 iio[WS(rs, 1)] = T1h + T1k;
cannam@127 247 rio[WS(rs, 2)] = T1P + T1W;
cannam@127 248 T3a = T36 - T37;
cannam@127 249 T38 = T36 + T37;
cannam@127 250 T31 = FMA(KP618033988, T30, T2X);
cannam@127 251 T3p = FNMS(KP618033988, T2X, T30);
cannam@127 252 T39 = FNMS(KP250000000, T38, T35);
cannam@127 253 iio[WS(rs, 2)] = T2b + T2e;
cannam@127 254 iio[WS(rs, 4)] = T3Z + T42;
cannam@127 255 rio[WS(rs, 4)] = T3D + T3K;
cannam@127 256 rio[WS(rs, 3)] = T2J + T2Q;
cannam@127 257 iio[WS(rs, 3)] = T35 + T38;
cannam@127 258 {
cannam@127 259 E T3O, T45, T2r, T2q, T2p, TT, TS, TR;
cannam@127 260 {
cannam@127 261 E TG, TL, TD, TC, TB, Tc, Tt;
cannam@127 262 TG = FNMS(KP559016994, Tb, Ta);
cannam@127 263 Tc = FMA(KP559016994, Tb, Ta);
cannam@127 264 Tt = FMA(KP559016994, Ts, Tr);
cannam@127 265 TL = FNMS(KP559016994, Ts, Tr);
cannam@127 266 {
cannam@127 267 E T9, Tm, Tk, TA, Tx;
cannam@127 268 T9 = W[0];
cannam@127 269 Tm = W[1];
cannam@127 270 Tk = FMA(KP951056516, Tj, Tc);
cannam@127 271 TA = FNMS(KP951056516, Tj, Tc);
cannam@127 272 Tx = FNMS(KP951056516, Tw, Tt);
cannam@127 273 TD = FMA(KP951056516, Tw, Tt);
cannam@127 274 {
cannam@127 275 E Tz, Tl, Ty, TE;
cannam@127 276 Tz = W[6];
cannam@127 277 Tl = T9 * Tk;
cannam@127 278 TC = W[7];
cannam@127 279 Ty = T9 * Tx;
cannam@127 280 TE = Tz * TD;
cannam@127 281 TB = Tz * TA;
cannam@127 282 rio[WS(vs, 1)] = FMA(Tm, Tx, Tl);
cannam@127 283 iio[WS(vs, 1)] = FNMS(Tm, Tk, Ty);
cannam@127 284 iio[WS(vs, 4)] = FNMS(TC, TA, TE);
cannam@127 285 }
cannam@127 286 }
cannam@127 287 rio[WS(vs, 4)] = FMA(TC, TD, TB);
cannam@127 288 {
cannam@127 289 E TF, TK, TI, TQ, TN;
cannam@127 290 TF = W[2];
cannam@127 291 TK = W[3];
cannam@127 292 TI = FNMS(KP951056516, TH, TG);
cannam@127 293 TQ = FMA(KP951056516, TH, TG);
cannam@127 294 TN = FMA(KP951056516, TM, TL);
cannam@127 295 TT = FNMS(KP951056516, TM, TL);
cannam@127 296 {
cannam@127 297 E TP, TJ, TO, TU;
cannam@127 298 TP = W[4];
cannam@127 299 TJ = TF * TI;
cannam@127 300 TS = W[5];
cannam@127 301 TO = TF * TN;
cannam@127 302 TU = TP * TT;
cannam@127 303 TR = TP * TQ;
cannam@127 304 rio[WS(vs, 2)] = FMA(TK, TN, TJ);
cannam@127 305 iio[WS(vs, 2)] = FNMS(TK, TI, TO);
cannam@127 306 iio[WS(vs, 3)] = FNMS(TS, TQ, TU);
cannam@127 307 }
cannam@127 308 }
cannam@127 309 }
cannam@127 310 rio[WS(vs, 3)] = FMA(TS, TT, TR);
cannam@127 311 {
cannam@127 312 E T20, T2h, T2H, T2G, T2F, T2u, T2z;
cannam@127 313 T20 = FMA(KP559016994, T1Z, T1Y);
cannam@127 314 T2u = FNMS(KP559016994, T1Z, T1Y);
cannam@127 315 T2z = FNMS(KP559016994, T2g, T2f);
cannam@127 316 T2h = FMA(KP559016994, T2g, T2f);
cannam@127 317 {
cannam@127 318 E T2t, T2y, T2w, T2E, T2B;
cannam@127 319 T2t = W[2];
cannam@127 320 T2y = W[3];
cannam@127 321 T2w = FNMS(KP951056516, T2v, T2u);
cannam@127 322 T2E = FMA(KP951056516, T2v, T2u);
cannam@127 323 T2B = FMA(KP951056516, T2A, T2z);
cannam@127 324 T2H = FNMS(KP951056516, T2A, T2z);
cannam@127 325 {
cannam@127 326 E T2D, T2x, T2C, T2I;
cannam@127 327 T2D = W[4];
cannam@127 328 T2x = T2t * T2w;
cannam@127 329 T2G = W[5];
cannam@127 330 T2C = T2t * T2B;
cannam@127 331 T2I = T2D * T2H;
cannam@127 332 T2F = T2D * T2E;
cannam@127 333 rio[WS(vs, 2) + WS(rs, 2)] = FMA(T2y, T2B, T2x);
cannam@127 334 iio[WS(vs, 2) + WS(rs, 2)] = FNMS(T2y, T2w, T2C);
cannam@127 335 iio[WS(vs, 3) + WS(rs, 2)] = FNMS(T2G, T2E, T2I);
cannam@127 336 }
cannam@127 337 }
cannam@127 338 rio[WS(vs, 3) + WS(rs, 2)] = FMA(T2G, T2H, T2F);
cannam@127 339 {
cannam@127 340 E T4v, T4u, T4t, T4i, T4n;
cannam@127 341 T3O = FMA(KP559016994, T3N, T3M);
cannam@127 342 T4i = FNMS(KP559016994, T3N, T3M);
cannam@127 343 T4n = FNMS(KP559016994, T44, T43);
cannam@127 344 T45 = FMA(KP559016994, T44, T43);
cannam@127 345 {
cannam@127 346 E T4h, T4m, T4k, T4s, T4p;
cannam@127 347 T4h = W[2];
cannam@127 348 T4m = W[3];
cannam@127 349 T4k = FNMS(KP951056516, T4j, T4i);
cannam@127 350 T4s = FMA(KP951056516, T4j, T4i);
cannam@127 351 T4p = FMA(KP951056516, T4o, T4n);
cannam@127 352 T4v = FNMS(KP951056516, T4o, T4n);
cannam@127 353 {
cannam@127 354 E T4r, T4l, T4q, T4w;
cannam@127 355 T4r = W[4];
cannam@127 356 T4l = T4h * T4k;
cannam@127 357 T4u = W[5];
cannam@127 358 T4q = T4h * T4p;
cannam@127 359 T4w = T4r * T4v;
cannam@127 360 T4t = T4r * T4s;
cannam@127 361 rio[WS(vs, 2) + WS(rs, 4)] = FMA(T4m, T4p, T4l);
cannam@127 362 iio[WS(vs, 2) + WS(rs, 4)] = FNMS(T4m, T4k, T4q);
cannam@127 363 iio[WS(vs, 3) + WS(rs, 4)] = FNMS(T4u, T4s, T4w);
cannam@127 364 }
cannam@127 365 }
cannam@127 366 rio[WS(vs, 3) + WS(rs, 4)] = FMA(T4u, T4v, T4t);
cannam@127 367 {
cannam@127 368 E T1X, T2a, T28, T2o, T2l;
cannam@127 369 T1X = W[0];
cannam@127 370 T2a = W[1];
cannam@127 371 T28 = FMA(KP951056516, T27, T20);
cannam@127 372 T2o = FNMS(KP951056516, T27, T20);
cannam@127 373 T2l = FNMS(KP951056516, T2k, T2h);
cannam@127 374 T2r = FMA(KP951056516, T2k, T2h);
cannam@127 375 {
cannam@127 376 E T2n, T29, T2m, T2s;
cannam@127 377 T2n = W[6];
cannam@127 378 T29 = T1X * T28;
cannam@127 379 T2q = W[7];
cannam@127 380 T2m = T1X * T2l;
cannam@127 381 T2s = T2n * T2r;
cannam@127 382 T2p = T2n * T2o;
cannam@127 383 rio[WS(vs, 1) + WS(rs, 2)] = FMA(T2a, T2l, T29);
cannam@127 384 iio[WS(vs, 1) + WS(rs, 2)] = FNMS(T2a, T28, T2m);
cannam@127 385 iio[WS(vs, 4) + WS(rs, 2)] = FNMS(T2q, T2o, T2s);
cannam@127 386 }
cannam@127 387 }
cannam@127 388 }
cannam@127 389 }
cannam@127 390 rio[WS(vs, 4) + WS(rs, 2)] = FMA(T2q, T2r, T2p);
cannam@127 391 {
cannam@127 392 E T3B, T3A, T3z, T4f, T4e, T4d;
cannam@127 393 {
cannam@127 394 E T3o, T3t, T3l, T3k, T3j, T2U, T3b;
cannam@127 395 T3o = FNMS(KP559016994, T2T, T2S);
cannam@127 396 T2U = FMA(KP559016994, T2T, T2S);
cannam@127 397 T3b = FMA(KP559016994, T3a, T39);
cannam@127 398 T3t = FNMS(KP559016994, T3a, T39);
cannam@127 399 {
cannam@127 400 E T2R, T34, T32, T3i, T3f;
cannam@127 401 T2R = W[0];
cannam@127 402 T34 = W[1];
cannam@127 403 T32 = FMA(KP951056516, T31, T2U);
cannam@127 404 T3i = FNMS(KP951056516, T31, T2U);
cannam@127 405 T3f = FNMS(KP951056516, T3e, T3b);
cannam@127 406 T3l = FMA(KP951056516, T3e, T3b);
cannam@127 407 {
cannam@127 408 E T3h, T33, T3g, T3m;
cannam@127 409 T3h = W[6];
cannam@127 410 T33 = T2R * T32;
cannam@127 411 T3k = W[7];
cannam@127 412 T3g = T2R * T3f;
cannam@127 413 T3m = T3h * T3l;
cannam@127 414 T3j = T3h * T3i;
cannam@127 415 rio[WS(vs, 1) + WS(rs, 3)] = FMA(T34, T3f, T33);
cannam@127 416 iio[WS(vs, 1) + WS(rs, 3)] = FNMS(T34, T32, T3g);
cannam@127 417 iio[WS(vs, 4) + WS(rs, 3)] = FNMS(T3k, T3i, T3m);
cannam@127 418 }
cannam@127 419 }
cannam@127 420 rio[WS(vs, 4) + WS(rs, 3)] = FMA(T3k, T3l, T3j);
cannam@127 421 {
cannam@127 422 E T3n, T3s, T3q, T3y, T3v;
cannam@127 423 T3n = W[2];
cannam@127 424 T3s = W[3];
cannam@127 425 T3q = FNMS(KP951056516, T3p, T3o);
cannam@127 426 T3y = FMA(KP951056516, T3p, T3o);
cannam@127 427 T3v = FMA(KP951056516, T3u, T3t);
cannam@127 428 T3B = FNMS(KP951056516, T3u, T3t);
cannam@127 429 {
cannam@127 430 E T3x, T3r, T3w, T3C;
cannam@127 431 T3x = W[4];
cannam@127 432 T3r = T3n * T3q;
cannam@127 433 T3A = W[5];
cannam@127 434 T3w = T3n * T3v;
cannam@127 435 T3C = T3x * T3B;
cannam@127 436 T3z = T3x * T3y;
cannam@127 437 rio[WS(vs, 2) + WS(rs, 3)] = FMA(T3s, T3v, T3r);
cannam@127 438 iio[WS(vs, 2) + WS(rs, 3)] = FNMS(T3s, T3q, T3w);
cannam@127 439 iio[WS(vs, 3) + WS(rs, 3)] = FNMS(T3A, T3y, T3C);
cannam@127 440 }
cannam@127 441 }
cannam@127 442 }
cannam@127 443 rio[WS(vs, 3) + WS(rs, 3)] = FMA(T3A, T3B, T3z);
cannam@127 444 {
cannam@127 445 E T3L, T3Y, T3W, T4c, T49;
cannam@127 446 T3L = W[0];
cannam@127 447 T3Y = W[1];
cannam@127 448 T3W = FMA(KP951056516, T3V, T3O);
cannam@127 449 T4c = FNMS(KP951056516, T3V, T3O);
cannam@127 450 T49 = FNMS(KP951056516, T48, T45);
cannam@127 451 T4f = FMA(KP951056516, T48, T45);
cannam@127 452 {
cannam@127 453 E T4b, T3X, T4a, T4g;
cannam@127 454 T4b = W[6];
cannam@127 455 T3X = T3L * T3W;
cannam@127 456 T4e = W[7];
cannam@127 457 T4a = T3L * T49;
cannam@127 458 T4g = T4b * T4f;
cannam@127 459 T4d = T4b * T4c;
cannam@127 460 rio[WS(vs, 1) + WS(rs, 4)] = FMA(T3Y, T49, T3X);
cannam@127 461 iio[WS(vs, 1) + WS(rs, 4)] = FNMS(T3Y, T3W, T4a);
cannam@127 462 iio[WS(vs, 4) + WS(rs, 4)] = FNMS(T4e, T4c, T4g);
cannam@127 463 }
cannam@127 464 }
cannam@127 465 rio[WS(vs, 4) + WS(rs, 4)] = FMA(T4e, T4f, T4d);
cannam@127 466 {
cannam@127 467 E T16, T1n, T1N, T1M, T1L, T1A, T1F;
cannam@127 468 T16 = FMA(KP559016994, T15, T14);
cannam@127 469 T1A = FNMS(KP559016994, T15, T14);
cannam@127 470 T1F = FNMS(KP559016994, T1m, T1l);
cannam@127 471 T1n = FMA(KP559016994, T1m, T1l);
cannam@127 472 {
cannam@127 473 E T1z, T1E, T1C, T1K, T1H;
cannam@127 474 T1z = W[2];
cannam@127 475 T1E = W[3];
cannam@127 476 T1C = FNMS(KP951056516, T1B, T1A);
cannam@127 477 T1K = FMA(KP951056516, T1B, T1A);
cannam@127 478 T1H = FMA(KP951056516, T1G, T1F);
cannam@127 479 T1N = FNMS(KP951056516, T1G, T1F);
cannam@127 480 {
cannam@127 481 E T1J, T1D, T1I, T1O;
cannam@127 482 T1J = W[4];
cannam@127 483 T1D = T1z * T1C;
cannam@127 484 T1M = W[5];
cannam@127 485 T1I = T1z * T1H;
cannam@127 486 T1O = T1J * T1N;
cannam@127 487 T1L = T1J * T1K;
cannam@127 488 rio[WS(vs, 2) + WS(rs, 1)] = FMA(T1E, T1H, T1D);
cannam@127 489 iio[WS(vs, 2) + WS(rs, 1)] = FNMS(T1E, T1C, T1I);
cannam@127 490 iio[WS(vs, 3) + WS(rs, 1)] = FNMS(T1M, T1K, T1O);
cannam@127 491 }
cannam@127 492 }
cannam@127 493 rio[WS(vs, 3) + WS(rs, 1)] = FMA(T1M, T1N, T1L);
cannam@127 494 {
cannam@127 495 E T13, T1g, T1e, T1u, T1r;
cannam@127 496 T13 = W[0];
cannam@127 497 T1g = W[1];
cannam@127 498 T1e = FMA(KP951056516, T1d, T16);
cannam@127 499 T1u = FNMS(KP951056516, T1d, T16);
cannam@127 500 T1r = FNMS(KP951056516, T1q, T1n);
cannam@127 501 T1x = FMA(KP951056516, T1q, T1n);
cannam@127 502 {
cannam@127 503 E T1t, T1f, T1s, T1y;
cannam@127 504 T1t = W[6];
cannam@127 505 T1f = T13 * T1e;
cannam@127 506 T1w = W[7];
cannam@127 507 T1s = T13 * T1r;
cannam@127 508 T1y = T1t * T1x;
cannam@127 509 T1v = T1t * T1u;
cannam@127 510 rio[WS(vs, 1) + WS(rs, 1)] = FMA(T1g, T1r, T1f);
cannam@127 511 iio[WS(vs, 1) + WS(rs, 1)] = FNMS(T1g, T1e, T1s);
cannam@127 512 iio[WS(vs, 4) + WS(rs, 1)] = FNMS(T1w, T1u, T1y);
cannam@127 513 }
cannam@127 514 }
cannam@127 515 }
cannam@127 516 }
cannam@127 517 }
cannam@127 518 }
cannam@127 519 }
cannam@127 520 rio[WS(vs, 4) + WS(rs, 1)] = FMA(T1w, T1x, T1v);
cannam@127 521 }
cannam@127 522 }
cannam@127 523 }
cannam@127 524
cannam@127 525 static const tw_instr twinstr[] = {
cannam@127 526 {TW_FULL, 0, 5},
cannam@127 527 {TW_NEXT, 1, 0}
cannam@127 528 };
cannam@127 529
cannam@127 530 static const ct_desc desc = { 5, "q1_5", twinstr, &GENUS, {70, 40, 130, 0}, 0, 0, 0 };
cannam@127 531
cannam@127 532 void X(codelet_q1_5) (planner *p) {
cannam@127 533 X(kdft_difsq_register) (p, q1_5, &desc);
cannam@127 534 }
cannam@127 535 #else /* HAVE_FMA */
cannam@127 536
cannam@127 537 /* Generated by: ../../../genfft/gen_twidsq.native -compact -variables 4 -pipeline-latency 4 -reload-twiddle -dif -n 5 -name q1_5 -include q.h */
cannam@127 538
cannam@127 539 /*
cannam@127 540 * This function contains 200 FP additions, 140 FP multiplications,
cannam@127 541 * (or, 130 additions, 70 multiplications, 70 fused multiply/add),
cannam@127 542 * 75 stack variables, 4 constants, and 100 memory accesses
cannam@127 543 */
cannam@127 544 #include "q.h"
cannam@127 545
cannam@127 546 static void q1_5(R *rio, R *iio, const R *W, stride rs, stride vs, INT mb, INT me, INT ms)
cannam@127 547 {
cannam@127 548 DK(KP250000000, +0.250000000000000000000000000000000000000000000);
cannam@127 549 DK(KP587785252, +0.587785252292473129168705954639072768597652438);
cannam@127 550 DK(KP951056516, +0.951056516295153572116439333379382143405698634);
cannam@127 551 DK(KP559016994, +0.559016994374947424102293417182819058860154590);
cannam@127 552 {
cannam@127 553 INT m;
cannam@127 554 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)) {
cannam@127 555 E T1, Ta, TG, Tv, T8, Tb, Tp, Tj, TD, To, Tq, Tr, TN, TW, T1s;
cannam@127 556 E T1h, TU, TX, T1b, T15, T1p, T1a, T1c, T1d, T1z, T1I, T2e, T23, T1G, T1J;
cannam@127 557 E T1X, T1R, T2b, T1W, T1Y, T1Z, T3v, T3p, T3J, T3u, T3w, T3x, T37, T3g, T3M;
cannam@127 558 E T3B, T3e, T3h, T2l, T2u, T30, T2P, T2s, T2v, T2J, T2D, T2X, T2I, T2K, T2L;
cannam@127 559 {
cannam@127 560 E T7, Tu, T4, Tt;
cannam@127 561 T1 = rio[0];
cannam@127 562 {
cannam@127 563 E T5, T6, T2, T3;
cannam@127 564 T5 = rio[WS(rs, 2)];
cannam@127 565 T6 = rio[WS(rs, 3)];
cannam@127 566 T7 = T5 + T6;
cannam@127 567 Tu = T5 - T6;
cannam@127 568 T2 = rio[WS(rs, 1)];
cannam@127 569 T3 = rio[WS(rs, 4)];
cannam@127 570 T4 = T2 + T3;
cannam@127 571 Tt = T2 - T3;
cannam@127 572 }
cannam@127 573 Ta = KP559016994 * (T4 - T7);
cannam@127 574 TG = FNMS(KP587785252, Tt, KP951056516 * Tu);
cannam@127 575 Tv = FMA(KP951056516, Tt, KP587785252 * Tu);
cannam@127 576 T8 = T4 + T7;
cannam@127 577 Tb = FNMS(KP250000000, T8, T1);
cannam@127 578 }
cannam@127 579 {
cannam@127 580 E Ti, Tn, Tf, Tm;
cannam@127 581 Tp = iio[0];
cannam@127 582 {
cannam@127 583 E Tg, Th, Td, Te;
cannam@127 584 Tg = iio[WS(rs, 2)];
cannam@127 585 Th = iio[WS(rs, 3)];
cannam@127 586 Ti = Tg - Th;
cannam@127 587 Tn = Tg + Th;
cannam@127 588 Td = iio[WS(rs, 1)];
cannam@127 589 Te = iio[WS(rs, 4)];
cannam@127 590 Tf = Td - Te;
cannam@127 591 Tm = Td + Te;
cannam@127 592 }
cannam@127 593 Tj = FMA(KP951056516, Tf, KP587785252 * Ti);
cannam@127 594 TD = FNMS(KP587785252, Tf, KP951056516 * Ti);
cannam@127 595 To = KP559016994 * (Tm - Tn);
cannam@127 596 Tq = Tm + Tn;
cannam@127 597 Tr = FNMS(KP250000000, Tq, Tp);
cannam@127 598 }
cannam@127 599 {
cannam@127 600 E TT, T1g, TQ, T1f;
cannam@127 601 TN = rio[WS(vs, 1)];
cannam@127 602 {
cannam@127 603 E TR, TS, TO, TP;
cannam@127 604 TR = rio[WS(vs, 1) + WS(rs, 2)];
cannam@127 605 TS = rio[WS(vs, 1) + WS(rs, 3)];
cannam@127 606 TT = TR + TS;
cannam@127 607 T1g = TR - TS;
cannam@127 608 TO = rio[WS(vs, 1) + WS(rs, 1)];
cannam@127 609 TP = rio[WS(vs, 1) + WS(rs, 4)];
cannam@127 610 TQ = TO + TP;
cannam@127 611 T1f = TO - TP;
cannam@127 612 }
cannam@127 613 TW = KP559016994 * (TQ - TT);
cannam@127 614 T1s = FNMS(KP587785252, T1f, KP951056516 * T1g);
cannam@127 615 T1h = FMA(KP951056516, T1f, KP587785252 * T1g);
cannam@127 616 TU = TQ + TT;
cannam@127 617 TX = FNMS(KP250000000, TU, TN);
cannam@127 618 }
cannam@127 619 {
cannam@127 620 E T14, T19, T11, T18;
cannam@127 621 T1b = iio[WS(vs, 1)];
cannam@127 622 {
cannam@127 623 E T12, T13, TZ, T10;
cannam@127 624 T12 = iio[WS(vs, 1) + WS(rs, 2)];
cannam@127 625 T13 = iio[WS(vs, 1) + WS(rs, 3)];
cannam@127 626 T14 = T12 - T13;
cannam@127 627 T19 = T12 + T13;
cannam@127 628 TZ = iio[WS(vs, 1) + WS(rs, 1)];
cannam@127 629 T10 = iio[WS(vs, 1) + WS(rs, 4)];
cannam@127 630 T11 = TZ - T10;
cannam@127 631 T18 = TZ + T10;
cannam@127 632 }
cannam@127 633 T15 = FMA(KP951056516, T11, KP587785252 * T14);
cannam@127 634 T1p = FNMS(KP587785252, T11, KP951056516 * T14);
cannam@127 635 T1a = KP559016994 * (T18 - T19);
cannam@127 636 T1c = T18 + T19;
cannam@127 637 T1d = FNMS(KP250000000, T1c, T1b);
cannam@127 638 }
cannam@127 639 {
cannam@127 640 E T1F, T22, T1C, T21;
cannam@127 641 T1z = rio[WS(vs, 2)];
cannam@127 642 {
cannam@127 643 E T1D, T1E, T1A, T1B;
cannam@127 644 T1D = rio[WS(vs, 2) + WS(rs, 2)];
cannam@127 645 T1E = rio[WS(vs, 2) + WS(rs, 3)];
cannam@127 646 T1F = T1D + T1E;
cannam@127 647 T22 = T1D - T1E;
cannam@127 648 T1A = rio[WS(vs, 2) + WS(rs, 1)];
cannam@127 649 T1B = rio[WS(vs, 2) + WS(rs, 4)];
cannam@127 650 T1C = T1A + T1B;
cannam@127 651 T21 = T1A - T1B;
cannam@127 652 }
cannam@127 653 T1I = KP559016994 * (T1C - T1F);
cannam@127 654 T2e = FNMS(KP587785252, T21, KP951056516 * T22);
cannam@127 655 T23 = FMA(KP951056516, T21, KP587785252 * T22);
cannam@127 656 T1G = T1C + T1F;
cannam@127 657 T1J = FNMS(KP250000000, T1G, T1z);
cannam@127 658 }
cannam@127 659 {
cannam@127 660 E T1Q, T1V, T1N, T1U;
cannam@127 661 T1X = iio[WS(vs, 2)];
cannam@127 662 {
cannam@127 663 E T1O, T1P, T1L, T1M;
cannam@127 664 T1O = iio[WS(vs, 2) + WS(rs, 2)];
cannam@127 665 T1P = iio[WS(vs, 2) + WS(rs, 3)];
cannam@127 666 T1Q = T1O - T1P;
cannam@127 667 T1V = T1O + T1P;
cannam@127 668 T1L = iio[WS(vs, 2) + WS(rs, 1)];
cannam@127 669 T1M = iio[WS(vs, 2) + WS(rs, 4)];
cannam@127 670 T1N = T1L - T1M;
cannam@127 671 T1U = T1L + T1M;
cannam@127 672 }
cannam@127 673 T1R = FMA(KP951056516, T1N, KP587785252 * T1Q);
cannam@127 674 T2b = FNMS(KP587785252, T1N, KP951056516 * T1Q);
cannam@127 675 T1W = KP559016994 * (T1U - T1V);
cannam@127 676 T1Y = T1U + T1V;
cannam@127 677 T1Z = FNMS(KP250000000, T1Y, T1X);
cannam@127 678 }
cannam@127 679 {
cannam@127 680 E T3o, T3t, T3l, T3s;
cannam@127 681 T3v = iio[WS(vs, 4)];
cannam@127 682 {
cannam@127 683 E T3m, T3n, T3j, T3k;
cannam@127 684 T3m = iio[WS(vs, 4) + WS(rs, 2)];
cannam@127 685 T3n = iio[WS(vs, 4) + WS(rs, 3)];
cannam@127 686 T3o = T3m - T3n;
cannam@127 687 T3t = T3m + T3n;
cannam@127 688 T3j = iio[WS(vs, 4) + WS(rs, 1)];
cannam@127 689 T3k = iio[WS(vs, 4) + WS(rs, 4)];
cannam@127 690 T3l = T3j - T3k;
cannam@127 691 T3s = T3j + T3k;
cannam@127 692 }
cannam@127 693 T3p = FMA(KP951056516, T3l, KP587785252 * T3o);
cannam@127 694 T3J = FNMS(KP587785252, T3l, KP951056516 * T3o);
cannam@127 695 T3u = KP559016994 * (T3s - T3t);
cannam@127 696 T3w = T3s + T3t;
cannam@127 697 T3x = FNMS(KP250000000, T3w, T3v);
cannam@127 698 }
cannam@127 699 {
cannam@127 700 E T3d, T3A, T3a, T3z;
cannam@127 701 T37 = rio[WS(vs, 4)];
cannam@127 702 {
cannam@127 703 E T3b, T3c, T38, T39;
cannam@127 704 T3b = rio[WS(vs, 4) + WS(rs, 2)];
cannam@127 705 T3c = rio[WS(vs, 4) + WS(rs, 3)];
cannam@127 706 T3d = T3b + T3c;
cannam@127 707 T3A = T3b - T3c;
cannam@127 708 T38 = rio[WS(vs, 4) + WS(rs, 1)];
cannam@127 709 T39 = rio[WS(vs, 4) + WS(rs, 4)];
cannam@127 710 T3a = T38 + T39;
cannam@127 711 T3z = T38 - T39;
cannam@127 712 }
cannam@127 713 T3g = KP559016994 * (T3a - T3d);
cannam@127 714 T3M = FNMS(KP587785252, T3z, KP951056516 * T3A);
cannam@127 715 T3B = FMA(KP951056516, T3z, KP587785252 * T3A);
cannam@127 716 T3e = T3a + T3d;
cannam@127 717 T3h = FNMS(KP250000000, T3e, T37);
cannam@127 718 }
cannam@127 719 {
cannam@127 720 E T2r, T2O, T2o, T2N;
cannam@127 721 T2l = rio[WS(vs, 3)];
cannam@127 722 {
cannam@127 723 E T2p, T2q, T2m, T2n;
cannam@127 724 T2p = rio[WS(vs, 3) + WS(rs, 2)];
cannam@127 725 T2q = rio[WS(vs, 3) + WS(rs, 3)];
cannam@127 726 T2r = T2p + T2q;
cannam@127 727 T2O = T2p - T2q;
cannam@127 728 T2m = rio[WS(vs, 3) + WS(rs, 1)];
cannam@127 729 T2n = rio[WS(vs, 3) + WS(rs, 4)];
cannam@127 730 T2o = T2m + T2n;
cannam@127 731 T2N = T2m - T2n;
cannam@127 732 }
cannam@127 733 T2u = KP559016994 * (T2o - T2r);
cannam@127 734 T30 = FNMS(KP587785252, T2N, KP951056516 * T2O);
cannam@127 735 T2P = FMA(KP951056516, T2N, KP587785252 * T2O);
cannam@127 736 T2s = T2o + T2r;
cannam@127 737 T2v = FNMS(KP250000000, T2s, T2l);
cannam@127 738 }
cannam@127 739 {
cannam@127 740 E T2C, T2H, T2z, T2G;
cannam@127 741 T2J = iio[WS(vs, 3)];
cannam@127 742 {
cannam@127 743 E T2A, T2B, T2x, T2y;
cannam@127 744 T2A = iio[WS(vs, 3) + WS(rs, 2)];
cannam@127 745 T2B = iio[WS(vs, 3) + WS(rs, 3)];
cannam@127 746 T2C = T2A - T2B;
cannam@127 747 T2H = T2A + T2B;
cannam@127 748 T2x = iio[WS(vs, 3) + WS(rs, 1)];
cannam@127 749 T2y = iio[WS(vs, 3) + WS(rs, 4)];
cannam@127 750 T2z = T2x - T2y;
cannam@127 751 T2G = T2x + T2y;
cannam@127 752 }
cannam@127 753 T2D = FMA(KP951056516, T2z, KP587785252 * T2C);
cannam@127 754 T2X = FNMS(KP587785252, T2z, KP951056516 * T2C);
cannam@127 755 T2I = KP559016994 * (T2G - T2H);
cannam@127 756 T2K = T2G + T2H;
cannam@127 757 T2L = FNMS(KP250000000, T2K, T2J);
cannam@127 758 }
cannam@127 759 rio[0] = T1 + T8;
cannam@127 760 iio[0] = Tp + Tq;
cannam@127 761 rio[WS(rs, 1)] = TN + TU;
cannam@127 762 iio[WS(rs, 1)] = T1b + T1c;
cannam@127 763 rio[WS(rs, 2)] = T1z + T1G;
cannam@127 764 iio[WS(rs, 2)] = T1X + T1Y;
cannam@127 765 iio[WS(rs, 4)] = T3v + T3w;
cannam@127 766 rio[WS(rs, 4)] = T37 + T3e;
cannam@127 767 rio[WS(rs, 3)] = T2l + T2s;
cannam@127 768 iio[WS(rs, 3)] = T2J + T2K;
cannam@127 769 {
cannam@127 770 E Tk, Ty, Tw, TA, Tc, Ts;
cannam@127 771 Tc = Ta + Tb;
cannam@127 772 Tk = Tc + Tj;
cannam@127 773 Ty = Tc - Tj;
cannam@127 774 Ts = To + Tr;
cannam@127 775 Tw = Ts - Tv;
cannam@127 776 TA = Tv + Ts;
cannam@127 777 {
cannam@127 778 E T9, Tl, Tx, Tz;
cannam@127 779 T9 = W[0];
cannam@127 780 Tl = W[1];
cannam@127 781 rio[WS(vs, 1)] = FMA(T9, Tk, Tl * Tw);
cannam@127 782 iio[WS(vs, 1)] = FNMS(Tl, Tk, T9 * Tw);
cannam@127 783 Tx = W[6];
cannam@127 784 Tz = W[7];
cannam@127 785 rio[WS(vs, 4)] = FMA(Tx, Ty, Tz * TA);
cannam@127 786 iio[WS(vs, 4)] = FNMS(Tz, Ty, Tx * TA);
cannam@127 787 }
cannam@127 788 }
cannam@127 789 {
cannam@127 790 E TE, TK, TI, TM, TC, TH;
cannam@127 791 TC = Tb - Ta;
cannam@127 792 TE = TC - TD;
cannam@127 793 TK = TC + TD;
cannam@127 794 TH = Tr - To;
cannam@127 795 TI = TG + TH;
cannam@127 796 TM = TH - TG;
cannam@127 797 {
cannam@127 798 E TB, TF, TJ, TL;
cannam@127 799 TB = W[2];
cannam@127 800 TF = W[3];
cannam@127 801 rio[WS(vs, 2)] = FMA(TB, TE, TF * TI);
cannam@127 802 iio[WS(vs, 2)] = FNMS(TF, TE, TB * TI);
cannam@127 803 TJ = W[4];
cannam@127 804 TL = W[5];
cannam@127 805 rio[WS(vs, 3)] = FMA(TJ, TK, TL * TM);
cannam@127 806 iio[WS(vs, 3)] = FNMS(TL, TK, TJ * TM);
cannam@127 807 }
cannam@127 808 }
cannam@127 809 {
cannam@127 810 E T2c, T2i, T2g, T2k, T2a, T2f;
cannam@127 811 T2a = T1J - T1I;
cannam@127 812 T2c = T2a - T2b;
cannam@127 813 T2i = T2a + T2b;
cannam@127 814 T2f = T1Z - T1W;
cannam@127 815 T2g = T2e + T2f;
cannam@127 816 T2k = T2f - T2e;
cannam@127 817 {
cannam@127 818 E T29, T2d, T2h, T2j;
cannam@127 819 T29 = W[2];
cannam@127 820 T2d = W[3];
cannam@127 821 rio[WS(vs, 2) + WS(rs, 2)] = FMA(T29, T2c, T2d * T2g);
cannam@127 822 iio[WS(vs, 2) + WS(rs, 2)] = FNMS(T2d, T2c, T29 * T2g);
cannam@127 823 T2h = W[4];
cannam@127 824 T2j = W[5];
cannam@127 825 rio[WS(vs, 3) + WS(rs, 2)] = FMA(T2h, T2i, T2j * T2k);
cannam@127 826 iio[WS(vs, 3) + WS(rs, 2)] = FNMS(T2j, T2i, T2h * T2k);
cannam@127 827 }
cannam@127 828 }
cannam@127 829 {
cannam@127 830 E T3K, T3Q, T3O, T3S, T3I, T3N;
cannam@127 831 T3I = T3h - T3g;
cannam@127 832 T3K = T3I - T3J;
cannam@127 833 T3Q = T3I + T3J;
cannam@127 834 T3N = T3x - T3u;
cannam@127 835 T3O = T3M + T3N;
cannam@127 836 T3S = T3N - T3M;
cannam@127 837 {
cannam@127 838 E T3H, T3L, T3P, T3R;
cannam@127 839 T3H = W[2];
cannam@127 840 T3L = W[3];
cannam@127 841 rio[WS(vs, 2) + WS(rs, 4)] = FMA(T3H, T3K, T3L * T3O);
cannam@127 842 iio[WS(vs, 2) + WS(rs, 4)] = FNMS(T3L, T3K, T3H * T3O);
cannam@127 843 T3P = W[4];
cannam@127 844 T3R = W[5];
cannam@127 845 rio[WS(vs, 3) + WS(rs, 4)] = FMA(T3P, T3Q, T3R * T3S);
cannam@127 846 iio[WS(vs, 3) + WS(rs, 4)] = FNMS(T3R, T3Q, T3P * T3S);
cannam@127 847 }
cannam@127 848 }
cannam@127 849 {
cannam@127 850 E T1S, T26, T24, T28, T1K, T20;
cannam@127 851 T1K = T1I + T1J;
cannam@127 852 T1S = T1K + T1R;
cannam@127 853 T26 = T1K - T1R;
cannam@127 854 T20 = T1W + T1Z;
cannam@127 855 T24 = T20 - T23;
cannam@127 856 T28 = T23 + T20;
cannam@127 857 {
cannam@127 858 E T1H, T1T, T25, T27;
cannam@127 859 T1H = W[0];
cannam@127 860 T1T = W[1];
cannam@127 861 rio[WS(vs, 1) + WS(rs, 2)] = FMA(T1H, T1S, T1T * T24);
cannam@127 862 iio[WS(vs, 1) + WS(rs, 2)] = FNMS(T1T, T1S, T1H * T24);
cannam@127 863 T25 = W[6];
cannam@127 864 T27 = W[7];
cannam@127 865 rio[WS(vs, 4) + WS(rs, 2)] = FMA(T25, T26, T27 * T28);
cannam@127 866 iio[WS(vs, 4) + WS(rs, 2)] = FNMS(T27, T26, T25 * T28);
cannam@127 867 }
cannam@127 868 }
cannam@127 869 {
cannam@127 870 E T2E, T2S, T2Q, T2U, T2w, T2M;
cannam@127 871 T2w = T2u + T2v;
cannam@127 872 T2E = T2w + T2D;
cannam@127 873 T2S = T2w - T2D;
cannam@127 874 T2M = T2I + T2L;
cannam@127 875 T2Q = T2M - T2P;
cannam@127 876 T2U = T2P + T2M;
cannam@127 877 {
cannam@127 878 E T2t, T2F, T2R, T2T;
cannam@127 879 T2t = W[0];
cannam@127 880 T2F = W[1];
cannam@127 881 rio[WS(vs, 1) + WS(rs, 3)] = FMA(T2t, T2E, T2F * T2Q);
cannam@127 882 iio[WS(vs, 1) + WS(rs, 3)] = FNMS(T2F, T2E, T2t * T2Q);
cannam@127 883 T2R = W[6];
cannam@127 884 T2T = W[7];
cannam@127 885 rio[WS(vs, 4) + WS(rs, 3)] = FMA(T2R, T2S, T2T * T2U);
cannam@127 886 iio[WS(vs, 4) + WS(rs, 3)] = FNMS(T2T, T2S, T2R * T2U);
cannam@127 887 }
cannam@127 888 }
cannam@127 889 {
cannam@127 890 E T2Y, T34, T32, T36, T2W, T31;
cannam@127 891 T2W = T2v - T2u;
cannam@127 892 T2Y = T2W - T2X;
cannam@127 893 T34 = T2W + T2X;
cannam@127 894 T31 = T2L - T2I;
cannam@127 895 T32 = T30 + T31;
cannam@127 896 T36 = T31 - T30;
cannam@127 897 {
cannam@127 898 E T2V, T2Z, T33, T35;
cannam@127 899 T2V = W[2];
cannam@127 900 T2Z = W[3];
cannam@127 901 rio[WS(vs, 2) + WS(rs, 3)] = FMA(T2V, T2Y, T2Z * T32);
cannam@127 902 iio[WS(vs, 2) + WS(rs, 3)] = FNMS(T2Z, T2Y, T2V * T32);
cannam@127 903 T33 = W[4];
cannam@127 904 T35 = W[5];
cannam@127 905 rio[WS(vs, 3) + WS(rs, 3)] = FMA(T33, T34, T35 * T36);
cannam@127 906 iio[WS(vs, 3) + WS(rs, 3)] = FNMS(T35, T34, T33 * T36);
cannam@127 907 }
cannam@127 908 }
cannam@127 909 {
cannam@127 910 E T3q, T3E, T3C, T3G, T3i, T3y;
cannam@127 911 T3i = T3g + T3h;
cannam@127 912 T3q = T3i + T3p;
cannam@127 913 T3E = T3i - T3p;
cannam@127 914 T3y = T3u + T3x;
cannam@127 915 T3C = T3y - T3B;
cannam@127 916 T3G = T3B + T3y;
cannam@127 917 {
cannam@127 918 E T3f, T3r, T3D, T3F;
cannam@127 919 T3f = W[0];
cannam@127 920 T3r = W[1];
cannam@127 921 rio[WS(vs, 1) + WS(rs, 4)] = FMA(T3f, T3q, T3r * T3C);
cannam@127 922 iio[WS(vs, 1) + WS(rs, 4)] = FNMS(T3r, T3q, T3f * T3C);
cannam@127 923 T3D = W[6];
cannam@127 924 T3F = W[7];
cannam@127 925 rio[WS(vs, 4) + WS(rs, 4)] = FMA(T3D, T3E, T3F * T3G);
cannam@127 926 iio[WS(vs, 4) + WS(rs, 4)] = FNMS(T3F, T3E, T3D * T3G);
cannam@127 927 }
cannam@127 928 }
cannam@127 929 {
cannam@127 930 E T1q, T1w, T1u, T1y, T1o, T1t;
cannam@127 931 T1o = TX - TW;
cannam@127 932 T1q = T1o - T1p;
cannam@127 933 T1w = T1o + T1p;
cannam@127 934 T1t = T1d - T1a;
cannam@127 935 T1u = T1s + T1t;
cannam@127 936 T1y = T1t - T1s;
cannam@127 937 {
cannam@127 938 E T1n, T1r, T1v, T1x;
cannam@127 939 T1n = W[2];
cannam@127 940 T1r = W[3];
cannam@127 941 rio[WS(vs, 2) + WS(rs, 1)] = FMA(T1n, T1q, T1r * T1u);
cannam@127 942 iio[WS(vs, 2) + WS(rs, 1)] = FNMS(T1r, T1q, T1n * T1u);
cannam@127 943 T1v = W[4];
cannam@127 944 T1x = W[5];
cannam@127 945 rio[WS(vs, 3) + WS(rs, 1)] = FMA(T1v, T1w, T1x * T1y);
cannam@127 946 iio[WS(vs, 3) + WS(rs, 1)] = FNMS(T1x, T1w, T1v * T1y);
cannam@127 947 }
cannam@127 948 }
cannam@127 949 {
cannam@127 950 E T16, T1k, T1i, T1m, TY, T1e;
cannam@127 951 TY = TW + TX;
cannam@127 952 T16 = TY + T15;
cannam@127 953 T1k = TY - T15;
cannam@127 954 T1e = T1a + T1d;
cannam@127 955 T1i = T1e - T1h;
cannam@127 956 T1m = T1h + T1e;
cannam@127 957 {
cannam@127 958 E TV, T17, T1j, T1l;
cannam@127 959 TV = W[0];
cannam@127 960 T17 = W[1];
cannam@127 961 rio[WS(vs, 1) + WS(rs, 1)] = FMA(TV, T16, T17 * T1i);
cannam@127 962 iio[WS(vs, 1) + WS(rs, 1)] = FNMS(T17, T16, TV * T1i);
cannam@127 963 T1j = W[6];
cannam@127 964 T1l = W[7];
cannam@127 965 rio[WS(vs, 4) + WS(rs, 1)] = FMA(T1j, T1k, T1l * T1m);
cannam@127 966 iio[WS(vs, 4) + WS(rs, 1)] = FNMS(T1l, T1k, T1j * T1m);
cannam@127 967 }
cannam@127 968 }
cannam@127 969 }
cannam@127 970 }
cannam@127 971 }
cannam@127 972
cannam@127 973 static const tw_instr twinstr[] = {
cannam@127 974 {TW_FULL, 0, 5},
cannam@127 975 {TW_NEXT, 1, 0}
cannam@127 976 };
cannam@127 977
cannam@127 978 static const ct_desc desc = { 5, "q1_5", twinstr, &GENUS, {130, 70, 70, 0}, 0, 0, 0 };
cannam@127 979
cannam@127 980 void X(codelet_q1_5) (planner *p) {
cannam@127 981 X(kdft_difsq_register) (p, q1_5, &desc);
cannam@127 982 }
cannam@127 983 #endif /* HAVE_FMA */