annotate src/fftw-3.3.5/dft/scalar/codelets/q1_6.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 16:37:59 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 6 -name q1_6 -include q.h */
cannam@127 29
cannam@127 30 /*
cannam@127 31 * This function contains 276 FP additions, 192 FP multiplications,
cannam@127 32 * (or, 144 additions, 60 multiplications, 132 fused multiply/add),
cannam@127 33 * 129 stack variables, 2 constants, and 144 memory accesses
cannam@127 34 */
cannam@127 35 #include "q.h"
cannam@127 36
cannam@127 37 static void q1_6(R *rio, R *iio, const R *W, stride rs, stride vs, INT mb, INT me, INT ms)
cannam@127 38 {
cannam@127 39 DK(KP866025403, +0.866025403784438646763723170752936183471402627);
cannam@127 40 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
cannam@127 41 {
cannam@127 42 INT m;
cannam@127 43 for (m = mb, W = W + (mb * 10); m < me; m = m + 1, rio = rio + ms, iio = iio + ms, W = W + 10, MAKE_VOLATILE_STRIDE(12, rs), MAKE_VOLATILE_STRIDE(0, vs)) {
cannam@127 44 E T4c, T4f, T4e, T4g, T4d;
cannam@127 45 {
cannam@127 46 E T3, Tw, Ta, TW, Tg, TG, TM, TT, TU, TP, Tn, T17, TV, TJ, Tv;
cannam@127 47 E T1A, T1e, T20, T1k, T1K, T1Q, T1X, T1Y, T1T, T1r, T1Z, T1N, T1z, T31, T32;
cannam@127 48 E T2X, T2v, T2b, T33, T2R, T2D, T2E, T2i, T34, T3f, T2o, T2O, T2U, T3I, T3m;
cannam@127 49 E T48, T3s, T3S, T3Y, T45, T46, T41, T3z, T4j, T47, T3V, T3H, T4M, T4q, T5c;
cannam@127 50 E T4w, T4W, T52, T59, T5a, T55, T4D, T5b, T4Z, T4L, T6d, T5r, T6e, T69, T5H;
cannam@127 51 E T5w, T5n, T6f, T63, T5P, T5s, T5o, T5p;
cannam@127 52 {
cannam@127 53 E T2f, T2k, T2g, T2c, T2d;
cannam@127 54 {
cannam@127 55 E T1b, T1g, T1c, T18, T19;
cannam@127 56 {
cannam@127 57 E T4, Tc, Te, T9, T5;
cannam@127 58 {
cannam@127 59 E T1, T2, T7, T8;
cannam@127 60 T1 = rio[0];
cannam@127 61 T2 = rio[WS(rs, 3)];
cannam@127 62 T7 = rio[WS(rs, 4)];
cannam@127 63 T8 = rio[WS(rs, 1)];
cannam@127 64 T4 = rio[WS(rs, 2)];
cannam@127 65 Tc = T1 - T2;
cannam@127 66 T3 = T1 + T2;
cannam@127 67 Te = T7 - T8;
cannam@127 68 T9 = T7 + T8;
cannam@127 69 T5 = rio[WS(rs, 5)];
cannam@127 70 }
cannam@127 71 {
cannam@127 72 E TN, Tj, Tk, Tl, Tt, Th, Ti;
cannam@127 73 Th = iio[WS(rs, 2)];
cannam@127 74 Ti = iio[WS(rs, 5)];
cannam@127 75 {
cannam@127 76 E Tr, Ts, Td, T6, Tf;
cannam@127 77 Tr = iio[0];
cannam@127 78 Td = T4 - T5;
cannam@127 79 T6 = T4 + T5;
cannam@127 80 TN = Th + Ti;
cannam@127 81 Tj = Th - Ti;
cannam@127 82 Tf = Td + Te;
cannam@127 83 Tw = Te - Td;
cannam@127 84 Ta = T6 + T9;
cannam@127 85 TW = T9 - T6;
cannam@127 86 Tg = FNMS(KP500000000, Tf, Tc);
cannam@127 87 TG = Tc + Tf;
cannam@127 88 Ts = iio[WS(rs, 3)];
cannam@127 89 TM = FNMS(KP500000000, Ta, T3);
cannam@127 90 Tk = iio[WS(rs, 4)];
cannam@127 91 Tl = iio[WS(rs, 1)];
cannam@127 92 Tt = Tr - Ts;
cannam@127 93 TT = Tr + Ts;
cannam@127 94 }
cannam@127 95 {
cannam@127 96 E T15, TO, Tm, T16, Tu;
cannam@127 97 T15 = rio[WS(vs, 1)];
cannam@127 98 TO = Tk + Tl;
cannam@127 99 Tm = Tk - Tl;
cannam@127 100 T16 = rio[WS(vs, 1) + WS(rs, 3)];
cannam@127 101 T1b = rio[WS(vs, 1) + WS(rs, 4)];
cannam@127 102 TU = TN + TO;
cannam@127 103 TP = TN - TO;
cannam@127 104 Tu = Tj + Tm;
cannam@127 105 Tn = Tj - Tm;
cannam@127 106 T1g = T15 - T16;
cannam@127 107 T17 = T15 + T16;
cannam@127 108 TV = FNMS(KP500000000, TU, TT);
cannam@127 109 TJ = Tt + Tu;
cannam@127 110 Tv = FNMS(KP500000000, Tu, Tt);
cannam@127 111 T1c = rio[WS(vs, 1) + WS(rs, 1)];
cannam@127 112 T18 = rio[WS(vs, 1) + WS(rs, 2)];
cannam@127 113 T19 = rio[WS(vs, 1) + WS(rs, 5)];
cannam@127 114 }
cannam@127 115 }
cannam@127 116 }
cannam@127 117 {
cannam@127 118 E T1v, T1R, T1n, T1w, T1o, T1p;
cannam@127 119 {
cannam@127 120 E T1l, T1i, T1d, T1h, T1a, T1m, T1j;
cannam@127 121 T1l = iio[WS(vs, 1) + WS(rs, 2)];
cannam@127 122 T1i = T1b - T1c;
cannam@127 123 T1d = T1b + T1c;
cannam@127 124 T1h = T18 - T19;
cannam@127 125 T1a = T18 + T19;
cannam@127 126 T1m = iio[WS(vs, 1) + WS(rs, 5)];
cannam@127 127 T1v = iio[WS(vs, 1)];
cannam@127 128 T1j = T1h + T1i;
cannam@127 129 T1A = T1i - T1h;
cannam@127 130 T1e = T1a + T1d;
cannam@127 131 T20 = T1d - T1a;
cannam@127 132 T1R = T1l + T1m;
cannam@127 133 T1n = T1l - T1m;
cannam@127 134 T1k = FNMS(KP500000000, T1j, T1g);
cannam@127 135 T1K = T1g + T1j;
cannam@127 136 T1Q = FNMS(KP500000000, T1e, T17);
cannam@127 137 T1w = iio[WS(vs, 1) + WS(rs, 3)];
cannam@127 138 T1o = iio[WS(vs, 1) + WS(rs, 4)];
cannam@127 139 T1p = iio[WS(vs, 1) + WS(rs, 1)];
cannam@127 140 }
cannam@127 141 {
cannam@127 142 E T2z, T2V, T2r, T2A, T2s, T2t;
cannam@127 143 {
cannam@127 144 E T2p, T1x, T1S, T1q, T2q, T1y;
cannam@127 145 T2p = iio[WS(vs, 2) + WS(rs, 2)];
cannam@127 146 T1X = T1v + T1w;
cannam@127 147 T1x = T1v - T1w;
cannam@127 148 T1S = T1o + T1p;
cannam@127 149 T1q = T1o - T1p;
cannam@127 150 T2q = iio[WS(vs, 2) + WS(rs, 5)];
cannam@127 151 T2z = iio[WS(vs, 2)];
cannam@127 152 T1Y = T1R + T1S;
cannam@127 153 T1T = T1R - T1S;
cannam@127 154 T1y = T1n + T1q;
cannam@127 155 T1r = T1n - T1q;
cannam@127 156 T2V = T2p + T2q;
cannam@127 157 T2r = T2p - T2q;
cannam@127 158 T1Z = FNMS(KP500000000, T1Y, T1X);
cannam@127 159 T1N = T1x + T1y;
cannam@127 160 T1z = FNMS(KP500000000, T1y, T1x);
cannam@127 161 T2A = iio[WS(vs, 2) + WS(rs, 3)];
cannam@127 162 T2s = iio[WS(vs, 2) + WS(rs, 4)];
cannam@127 163 T2t = iio[WS(vs, 2) + WS(rs, 1)];
cannam@127 164 }
cannam@127 165 {
cannam@127 166 E T29, T2B, T2W, T2u, T2a, T2C;
cannam@127 167 T29 = rio[WS(vs, 2)];
cannam@127 168 T31 = T2z + T2A;
cannam@127 169 T2B = T2z - T2A;
cannam@127 170 T2W = T2s + T2t;
cannam@127 171 T2u = T2s - T2t;
cannam@127 172 T2a = rio[WS(vs, 2) + WS(rs, 3)];
cannam@127 173 T2f = rio[WS(vs, 2) + WS(rs, 4)];
cannam@127 174 T32 = T2V + T2W;
cannam@127 175 T2X = T2V - T2W;
cannam@127 176 T2C = T2r + T2u;
cannam@127 177 T2v = T2r - T2u;
cannam@127 178 T2k = T29 - T2a;
cannam@127 179 T2b = T29 + T2a;
cannam@127 180 T33 = FNMS(KP500000000, T32, T31);
cannam@127 181 T2R = T2B + T2C;
cannam@127 182 T2D = FNMS(KP500000000, T2C, T2B);
cannam@127 183 T2g = rio[WS(vs, 2) + WS(rs, 1)];
cannam@127 184 T2c = rio[WS(vs, 2) + WS(rs, 2)];
cannam@127 185 T2d = rio[WS(vs, 2) + WS(rs, 5)];
cannam@127 186 }
cannam@127 187 }
cannam@127 188 }
cannam@127 189 }
cannam@127 190 {
cannam@127 191 E T4n, T4s, T4o, T4k, T4l;
cannam@127 192 {
cannam@127 193 E T3j, T3o, T3k, T3g, T3h;
cannam@127 194 {
cannam@127 195 E T3d, T2m, T2h, T2l, T2e, T3e, T2n;
cannam@127 196 T3d = rio[WS(vs, 3)];
cannam@127 197 T2m = T2f - T2g;
cannam@127 198 T2h = T2f + T2g;
cannam@127 199 T2l = T2c - T2d;
cannam@127 200 T2e = T2c + T2d;
cannam@127 201 T3e = rio[WS(vs, 3) + WS(rs, 3)];
cannam@127 202 T3j = rio[WS(vs, 3) + WS(rs, 4)];
cannam@127 203 T2n = T2l + T2m;
cannam@127 204 T2E = T2m - T2l;
cannam@127 205 T2i = T2e + T2h;
cannam@127 206 T34 = T2h - T2e;
cannam@127 207 T3o = T3d - T3e;
cannam@127 208 T3f = T3d + T3e;
cannam@127 209 T2o = FNMS(KP500000000, T2n, T2k);
cannam@127 210 T2O = T2k + T2n;
cannam@127 211 T2U = FNMS(KP500000000, T2i, T2b);
cannam@127 212 T3k = rio[WS(vs, 3) + WS(rs, 1)];
cannam@127 213 T3g = rio[WS(vs, 3) + WS(rs, 2)];
cannam@127 214 T3h = rio[WS(vs, 3) + WS(rs, 5)];
cannam@127 215 }
cannam@127 216 {
cannam@127 217 E T3D, T3Z, T3v, T3E, T3w, T3x;
cannam@127 218 {
cannam@127 219 E T3t, T3q, T3l, T3p, T3i, T3u, T3r;
cannam@127 220 T3t = iio[WS(vs, 3) + WS(rs, 2)];
cannam@127 221 T3q = T3j - T3k;
cannam@127 222 T3l = T3j + T3k;
cannam@127 223 T3p = T3g - T3h;
cannam@127 224 T3i = T3g + T3h;
cannam@127 225 T3u = iio[WS(vs, 3) + WS(rs, 5)];
cannam@127 226 T3D = iio[WS(vs, 3)];
cannam@127 227 T3r = T3p + T3q;
cannam@127 228 T3I = T3q - T3p;
cannam@127 229 T3m = T3i + T3l;
cannam@127 230 T48 = T3l - T3i;
cannam@127 231 T3Z = T3t + T3u;
cannam@127 232 T3v = T3t - T3u;
cannam@127 233 T3s = FNMS(KP500000000, T3r, T3o);
cannam@127 234 T3S = T3o + T3r;
cannam@127 235 T3Y = FNMS(KP500000000, T3m, T3f);
cannam@127 236 T3E = iio[WS(vs, 3) + WS(rs, 3)];
cannam@127 237 T3w = iio[WS(vs, 3) + WS(rs, 4)];
cannam@127 238 T3x = iio[WS(vs, 3) + WS(rs, 1)];
cannam@127 239 }
cannam@127 240 {
cannam@127 241 E T4h, T3F, T40, T3y, T4i, T3G;
cannam@127 242 T4h = rio[WS(vs, 4)];
cannam@127 243 T45 = T3D + T3E;
cannam@127 244 T3F = T3D - T3E;
cannam@127 245 T40 = T3w + T3x;
cannam@127 246 T3y = T3w - T3x;
cannam@127 247 T4i = rio[WS(vs, 4) + WS(rs, 3)];
cannam@127 248 T4n = rio[WS(vs, 4) + WS(rs, 4)];
cannam@127 249 T46 = T3Z + T40;
cannam@127 250 T41 = T3Z - T40;
cannam@127 251 T3G = T3v + T3y;
cannam@127 252 T3z = T3v - T3y;
cannam@127 253 T4s = T4h - T4i;
cannam@127 254 T4j = T4h + T4i;
cannam@127 255 T47 = FNMS(KP500000000, T46, T45);
cannam@127 256 T3V = T3F + T3G;
cannam@127 257 T3H = FNMS(KP500000000, T3G, T3F);
cannam@127 258 T4o = rio[WS(vs, 4) + WS(rs, 1)];
cannam@127 259 T4k = rio[WS(vs, 4) + WS(rs, 2)];
cannam@127 260 T4l = rio[WS(vs, 4) + WS(rs, 5)];
cannam@127 261 }
cannam@127 262 }
cannam@127 263 }
cannam@127 264 {
cannam@127 265 E T4H, T53, T4z, T4I, T4A, T4B;
cannam@127 266 {
cannam@127 267 E T4x, T4u, T4p, T4t, T4m, T4y, T4v;
cannam@127 268 T4x = iio[WS(vs, 4) + WS(rs, 2)];
cannam@127 269 T4u = T4n - T4o;
cannam@127 270 T4p = T4n + T4o;
cannam@127 271 T4t = T4k - T4l;
cannam@127 272 T4m = T4k + T4l;
cannam@127 273 T4y = iio[WS(vs, 4) + WS(rs, 5)];
cannam@127 274 T4H = iio[WS(vs, 4)];
cannam@127 275 T4v = T4t + T4u;
cannam@127 276 T4M = T4u - T4t;
cannam@127 277 T4q = T4m + T4p;
cannam@127 278 T5c = T4p - T4m;
cannam@127 279 T53 = T4x + T4y;
cannam@127 280 T4z = T4x - T4y;
cannam@127 281 T4w = FNMS(KP500000000, T4v, T4s);
cannam@127 282 T4W = T4s + T4v;
cannam@127 283 T52 = FNMS(KP500000000, T4q, T4j);
cannam@127 284 T4I = iio[WS(vs, 4) + WS(rs, 3)];
cannam@127 285 T4A = iio[WS(vs, 4) + WS(rs, 4)];
cannam@127 286 T4B = iio[WS(vs, 4) + WS(rs, 1)];
cannam@127 287 }
cannam@127 288 {
cannam@127 289 E T5L, T67, T5D, T5M, T5E, T5F;
cannam@127 290 {
cannam@127 291 E T5B, T4J, T54, T4C, T5C, T4K;
cannam@127 292 T5B = iio[WS(vs, 5) + WS(rs, 2)];
cannam@127 293 T59 = T4H + T4I;
cannam@127 294 T4J = T4H - T4I;
cannam@127 295 T54 = T4A + T4B;
cannam@127 296 T4C = T4A - T4B;
cannam@127 297 T5C = iio[WS(vs, 5) + WS(rs, 5)];
cannam@127 298 T5L = iio[WS(vs, 5)];
cannam@127 299 T5a = T53 + T54;
cannam@127 300 T55 = T53 - T54;
cannam@127 301 T4K = T4z + T4C;
cannam@127 302 T4D = T4z - T4C;
cannam@127 303 T67 = T5B + T5C;
cannam@127 304 T5D = T5B - T5C;
cannam@127 305 T5b = FNMS(KP500000000, T5a, T59);
cannam@127 306 T4Z = T4J + T4K;
cannam@127 307 T4L = FNMS(KP500000000, T4K, T4J);
cannam@127 308 T5M = iio[WS(vs, 5) + WS(rs, 3)];
cannam@127 309 T5E = iio[WS(vs, 5) + WS(rs, 4)];
cannam@127 310 T5F = iio[WS(vs, 5) + WS(rs, 1)];
cannam@127 311 }
cannam@127 312 {
cannam@127 313 E T5l, T5N, T68, T5G, T5m, T5O;
cannam@127 314 T5l = rio[WS(vs, 5)];
cannam@127 315 T6d = T5L + T5M;
cannam@127 316 T5N = T5L - T5M;
cannam@127 317 T68 = T5E + T5F;
cannam@127 318 T5G = T5E - T5F;
cannam@127 319 T5m = rio[WS(vs, 5) + WS(rs, 3)];
cannam@127 320 T5r = rio[WS(vs, 5) + WS(rs, 4)];
cannam@127 321 T6e = T67 + T68;
cannam@127 322 T69 = T67 - T68;
cannam@127 323 T5O = T5D + T5G;
cannam@127 324 T5H = T5D - T5G;
cannam@127 325 T5w = T5l - T5m;
cannam@127 326 T5n = T5l + T5m;
cannam@127 327 T6f = FNMS(KP500000000, T6e, T6d);
cannam@127 328 T63 = T5N + T5O;
cannam@127 329 T5P = FNMS(KP500000000, T5O, T5N);
cannam@127 330 T5s = rio[WS(vs, 5) + WS(rs, 1)];
cannam@127 331 T5o = rio[WS(vs, 5) + WS(rs, 2)];
cannam@127 332 T5p = rio[WS(vs, 5) + WS(rs, 5)];
cannam@127 333 }
cannam@127 334 }
cannam@127 335 }
cannam@127 336 }
cannam@127 337 }
cannam@127 338 {
cannam@127 339 E T6a, T6h, T5I, T5R, T65, T6c;
cannam@127 340 {
cannam@127 341 E T5Q, T5u, T6g, T5A, T60, T66;
cannam@127 342 {
cannam@127 343 E T5y, T5t, T5x, T5q, T5z;
cannam@127 344 rio[0] = T3 + Ta;
cannam@127 345 T5y = T5r - T5s;
cannam@127 346 T5t = T5r + T5s;
cannam@127 347 T5x = T5o - T5p;
cannam@127 348 T5q = T5o + T5p;
cannam@127 349 iio[0] = TT + TU;
cannam@127 350 rio[WS(rs, 1)] = T17 + T1e;
cannam@127 351 T5z = T5x + T5y;
cannam@127 352 T5Q = T5y - T5x;
cannam@127 353 T5u = T5q + T5t;
cannam@127 354 T6g = T5t - T5q;
cannam@127 355 T5A = FNMS(KP500000000, T5z, T5w);
cannam@127 356 T60 = T5w + T5z;
cannam@127 357 iio[WS(rs, 1)] = T1X + T1Y;
cannam@127 358 T66 = FNMS(KP500000000, T5u, T5n);
cannam@127 359 rio[WS(rs, 2)] = T2b + T2i;
cannam@127 360 }
cannam@127 361 iio[WS(rs, 2)] = T31 + T32;
cannam@127 362 iio[WS(rs, 4)] = T59 + T5a;
cannam@127 363 rio[WS(rs, 4)] = T4j + T4q;
cannam@127 364 rio[WS(rs, 3)] = T3f + T3m;
cannam@127 365 iio[WS(rs, 3)] = T45 + T46;
cannam@127 366 {
cannam@127 367 E TA, TD, TQ, T10, T13, TX, TZ, T12;
cannam@127 368 rio[WS(rs, 5)] = T5n + T5u;
cannam@127 369 iio[WS(rs, 5)] = T6d + T6e;
cannam@127 370 {
cannam@127 371 E To, Tx, Tb, Tq;
cannam@127 372 TA = FNMS(KP866025403, Tn, Tg);
cannam@127 373 To = FMA(KP866025403, Tn, Tg);
cannam@127 374 Tx = FMA(KP866025403, Tw, Tv);
cannam@127 375 TD = FNMS(KP866025403, Tw, Tv);
cannam@127 376 Tb = W[0];
cannam@127 377 Tq = W[1];
cannam@127 378 {
cannam@127 379 E TI, TK, TH, Ty, Tp, TF;
cannam@127 380 Ty = Tb * Tx;
cannam@127 381 Tp = Tb * To;
cannam@127 382 TF = W[4];
cannam@127 383 TI = W[5];
cannam@127 384 iio[WS(vs, 1)] = FNMS(Tq, To, Ty);
cannam@127 385 rio[WS(vs, 1)] = FMA(Tq, Tx, Tp);
cannam@127 386 TK = TF * TJ;
cannam@127 387 TH = TF * TG;
cannam@127 388 TQ = FNMS(KP866025403, TP, TM);
cannam@127 389 T10 = FMA(KP866025403, TP, TM);
cannam@127 390 T13 = FMA(KP866025403, TW, TV);
cannam@127 391 TX = FNMS(KP866025403, TW, TV);
cannam@127 392 iio[WS(vs, 3)] = FNMS(TI, TG, TK);
cannam@127 393 rio[WS(vs, 3)] = FMA(TI, TJ, TH);
cannam@127 394 TZ = W[6];
cannam@127 395 T12 = W[7];
cannam@127 396 }
cannam@127 397 }
cannam@127 398 {
cannam@127 399 E TC, TE, TB, TL, TS;
cannam@127 400 {
cannam@127 401 E T62, T64, T61, T14, T11, T5Z;
cannam@127 402 T14 = TZ * T13;
cannam@127 403 T11 = TZ * T10;
cannam@127 404 T5Z = W[4];
cannam@127 405 T62 = W[5];
cannam@127 406 iio[WS(vs, 4)] = FNMS(T12, T10, T14);
cannam@127 407 rio[WS(vs, 4)] = FMA(T12, T13, T11);
cannam@127 408 T64 = T5Z * T63;
cannam@127 409 T61 = T5Z * T60;
cannam@127 410 {
cannam@127 411 E T6k, T6n, T6j, T6m, T6o, T6l, Tz;
cannam@127 412 T6a = FNMS(KP866025403, T69, T66);
cannam@127 413 T6k = FMA(KP866025403, T69, T66);
cannam@127 414 T6n = FMA(KP866025403, T6g, T6f);
cannam@127 415 T6h = FNMS(KP866025403, T6g, T6f);
cannam@127 416 iio[WS(vs, 3) + WS(rs, 5)] = FNMS(T62, T60, T64);
cannam@127 417 rio[WS(vs, 3) + WS(rs, 5)] = FMA(T62, T63, T61);
cannam@127 418 T6j = W[6];
cannam@127 419 T6m = W[7];
cannam@127 420 T6o = T6j * T6n;
cannam@127 421 T6l = T6j * T6k;
cannam@127 422 Tz = W[8];
cannam@127 423 TC = W[9];
cannam@127 424 iio[WS(vs, 4) + WS(rs, 5)] = FNMS(T6m, T6k, T6o);
cannam@127 425 rio[WS(vs, 4) + WS(rs, 5)] = FMA(T6m, T6n, T6l);
cannam@127 426 TE = Tz * TD;
cannam@127 427 TB = Tz * TA;
cannam@127 428 }
cannam@127 429 }
cannam@127 430 iio[WS(vs, 5)] = FNMS(TC, TA, TE);
cannam@127 431 rio[WS(vs, 5)] = FMA(TC, TD, TB);
cannam@127 432 TL = W[2];
cannam@127 433 TS = W[3];
cannam@127 434 {
cannam@127 435 E T5U, T5X, T5W, T5Y, T5V, TY, TR, T5T;
cannam@127 436 T5I = FMA(KP866025403, T5H, T5A);
cannam@127 437 T5U = FNMS(KP866025403, T5H, T5A);
cannam@127 438 T5X = FNMS(KP866025403, T5Q, T5P);
cannam@127 439 T5R = FMA(KP866025403, T5Q, T5P);
cannam@127 440 TY = TL * TX;
cannam@127 441 TR = TL * TQ;
cannam@127 442 T5T = W[8];
cannam@127 443 T5W = W[9];
cannam@127 444 iio[WS(vs, 2)] = FNMS(TS, TQ, TY);
cannam@127 445 rio[WS(vs, 2)] = FMA(TS, TX, TR);
cannam@127 446 T5Y = T5T * T5X;
cannam@127 447 T5V = T5T * T5U;
cannam@127 448 iio[WS(vs, 5) + WS(rs, 5)] = FNMS(T5W, T5U, T5Y);
cannam@127 449 rio[WS(vs, 5) + WS(rs, 5)] = FMA(T5W, T5X, T5V);
cannam@127 450 T65 = W[2];
cannam@127 451 T6c = W[3];
cannam@127 452 }
cannam@127 453 }
cannam@127 454 }
cannam@127 455 }
cannam@127 456 {
cannam@127 457 E T5g, T5j, T5f, T5i;
cannam@127 458 {
cannam@127 459 E T1E, T1H, T3M, T3P, T56, T5d, T58, T5e, T57;
cannam@127 460 {
cannam@127 461 E T1s, T1B, T1f, T1u;
cannam@127 462 {
cannam@127 463 E T5K, T5S, T5J, T6i, T6b, T5v;
cannam@127 464 T6i = T65 * T6h;
cannam@127 465 T6b = T65 * T6a;
cannam@127 466 T5v = W[0];
cannam@127 467 T5K = W[1];
cannam@127 468 iio[WS(vs, 2) + WS(rs, 5)] = FNMS(T6c, T6a, T6i);
cannam@127 469 rio[WS(vs, 2) + WS(rs, 5)] = FMA(T6c, T6h, T6b);
cannam@127 470 T5S = T5v * T5R;
cannam@127 471 T5J = T5v * T5I;
cannam@127 472 T1E = FNMS(KP866025403, T1r, T1k);
cannam@127 473 T1s = FMA(KP866025403, T1r, T1k);
cannam@127 474 T1B = FMA(KP866025403, T1A, T1z);
cannam@127 475 T1H = FNMS(KP866025403, T1A, T1z);
cannam@127 476 iio[WS(vs, 1) + WS(rs, 5)] = FNMS(T5K, T5I, T5S);
cannam@127 477 rio[WS(vs, 1) + WS(rs, 5)] = FMA(T5K, T5R, T5J);
cannam@127 478 T1f = W[0];
cannam@127 479 T1u = W[1];
cannam@127 480 }
cannam@127 481 {
cannam@127 482 E T3U, T3W, T3T, T1C, T1t, T3R;
cannam@127 483 T1C = T1f * T1B;
cannam@127 484 T1t = T1f * T1s;
cannam@127 485 T3R = W[4];
cannam@127 486 T3U = W[5];
cannam@127 487 iio[WS(vs, 1) + WS(rs, 1)] = FNMS(T1u, T1s, T1C);
cannam@127 488 rio[WS(vs, 1) + WS(rs, 1)] = FMA(T1u, T1B, T1t);
cannam@127 489 T3W = T3R * T3V;
cannam@127 490 T3T = T3R * T3S;
cannam@127 491 {
cannam@127 492 E T3A, T3J, T3n, T3C, T3K, T3B, T51;
cannam@127 493 T3M = FNMS(KP866025403, T3z, T3s);
cannam@127 494 T3A = FMA(KP866025403, T3z, T3s);
cannam@127 495 T3J = FMA(KP866025403, T3I, T3H);
cannam@127 496 T3P = FNMS(KP866025403, T3I, T3H);
cannam@127 497 iio[WS(vs, 3) + WS(rs, 3)] = FNMS(T3U, T3S, T3W);
cannam@127 498 rio[WS(vs, 3) + WS(rs, 3)] = FMA(T3U, T3V, T3T);
cannam@127 499 T3n = W[0];
cannam@127 500 T3C = W[1];
cannam@127 501 T5g = FMA(KP866025403, T55, T52);
cannam@127 502 T56 = FNMS(KP866025403, T55, T52);
cannam@127 503 T5d = FNMS(KP866025403, T5c, T5b);
cannam@127 504 T5j = FMA(KP866025403, T5c, T5b);
cannam@127 505 T3K = T3n * T3J;
cannam@127 506 T3B = T3n * T3A;
cannam@127 507 T51 = W[2];
cannam@127 508 T58 = W[3];
cannam@127 509 iio[WS(vs, 1) + WS(rs, 3)] = FNMS(T3C, T3A, T3K);
cannam@127 510 rio[WS(vs, 1) + WS(rs, 3)] = FMA(T3C, T3J, T3B);
cannam@127 511 T5e = T51 * T5d;
cannam@127 512 T57 = T51 * T56;
cannam@127 513 }
cannam@127 514 }
cannam@127 515 }
cannam@127 516 {
cannam@127 517 E T38, T3b, T3O, T3Q, T3N, T37, T3a;
cannam@127 518 {
cannam@127 519 E T2Y, T35, T2T, T30, T36, T2Z, T3L;
cannam@127 520 T38 = FMA(KP866025403, T2X, T2U);
cannam@127 521 T2Y = FNMS(KP866025403, T2X, T2U);
cannam@127 522 T35 = FNMS(KP866025403, T34, T33);
cannam@127 523 T3b = FMA(KP866025403, T34, T33);
cannam@127 524 iio[WS(vs, 2) + WS(rs, 4)] = FNMS(T58, T56, T5e);
cannam@127 525 rio[WS(vs, 2) + WS(rs, 4)] = FMA(T58, T5d, T57);
cannam@127 526 T2T = W[2];
cannam@127 527 T30 = W[3];
cannam@127 528 T36 = T2T * T35;
cannam@127 529 T2Z = T2T * T2Y;
cannam@127 530 T3L = W[8];
cannam@127 531 T3O = W[9];
cannam@127 532 iio[WS(vs, 2) + WS(rs, 2)] = FNMS(T30, T2Y, T36);
cannam@127 533 rio[WS(vs, 2) + WS(rs, 2)] = FMA(T30, T35, T2Z);
cannam@127 534 T3Q = T3L * T3P;
cannam@127 535 T3N = T3L * T3M;
cannam@127 536 }
cannam@127 537 iio[WS(vs, 5) + WS(rs, 3)] = FNMS(T3O, T3M, T3Q);
cannam@127 538 rio[WS(vs, 5) + WS(rs, 3)] = FMA(T3O, T3P, T3N);
cannam@127 539 T37 = W[6];
cannam@127 540 T3a = W[7];
cannam@127 541 {
cannam@127 542 E T1G, T1I, T1F, T3c, T39, T1D;
cannam@127 543 T3c = T37 * T3b;
cannam@127 544 T39 = T37 * T38;
cannam@127 545 T1D = W[8];
cannam@127 546 T1G = W[9];
cannam@127 547 iio[WS(vs, 4) + WS(rs, 2)] = FNMS(T3a, T38, T3c);
cannam@127 548 rio[WS(vs, 4) + WS(rs, 2)] = FMA(T3a, T3b, T39);
cannam@127 549 T1I = T1D * T1H;
cannam@127 550 T1F = T1D * T1E;
cannam@127 551 iio[WS(vs, 5) + WS(rs, 1)] = FNMS(T1G, T1E, T1I);
cannam@127 552 rio[WS(vs, 5) + WS(rs, 1)] = FMA(T1G, T1H, T1F);
cannam@127 553 T5f = W[6];
cannam@127 554 T5i = W[7];
cannam@127 555 }
cannam@127 556 }
cannam@127 557 }
cannam@127 558 {
cannam@127 559 E T4Q, T4T, T2I, T2w, T2F, T2L, T2y, T2G, T2x, T4V, T4Y;
cannam@127 560 {
cannam@127 561 E T1M, T1O, T1L, T5k, T5h, T1J;
cannam@127 562 T5k = T5f * T5j;
cannam@127 563 T5h = T5f * T5g;
cannam@127 564 T1J = W[4];
cannam@127 565 T1M = W[5];
cannam@127 566 iio[WS(vs, 4) + WS(rs, 4)] = FNMS(T5i, T5g, T5k);
cannam@127 567 rio[WS(vs, 4) + WS(rs, 4)] = FMA(T5i, T5j, T5h);
cannam@127 568 T1O = T1J * T1N;
cannam@127 569 T1L = T1J * T1K;
cannam@127 570 iio[WS(vs, 3) + WS(rs, 1)] = FNMS(T1M, T1K, T1O);
cannam@127 571 rio[WS(vs, 3) + WS(rs, 1)] = FMA(T1M, T1N, T1L);
cannam@127 572 T4V = W[4];
cannam@127 573 T4Y = W[5];
cannam@127 574 }
cannam@127 575 {
cannam@127 576 E T4E, T4N, T4G, T4O, T4F, T50, T4X, T4r;
cannam@127 577 T4Q = FNMS(KP866025403, T4D, T4w);
cannam@127 578 T4E = FMA(KP866025403, T4D, T4w);
cannam@127 579 T4N = FMA(KP866025403, T4M, T4L);
cannam@127 580 T4T = FNMS(KP866025403, T4M, T4L);
cannam@127 581 T50 = T4V * T4Z;
cannam@127 582 T4X = T4V * T4W;
cannam@127 583 T4r = W[0];
cannam@127 584 T4G = W[1];
cannam@127 585 iio[WS(vs, 3) + WS(rs, 4)] = FNMS(T4Y, T4W, T50);
cannam@127 586 rio[WS(vs, 3) + WS(rs, 4)] = FMA(T4Y, T4Z, T4X);
cannam@127 587 T4O = T4r * T4N;
cannam@127 588 T4F = T4r * T4E;
cannam@127 589 {
cannam@127 590 E T2N, T2Q, T2S, T2P, T2j;
cannam@127 591 iio[WS(vs, 1) + WS(rs, 4)] = FNMS(T4G, T4E, T4O);
cannam@127 592 rio[WS(vs, 1) + WS(rs, 4)] = FMA(T4G, T4N, T4F);
cannam@127 593 T2N = W[4];
cannam@127 594 T2Q = W[5];
cannam@127 595 T2I = FNMS(KP866025403, T2v, T2o);
cannam@127 596 T2w = FMA(KP866025403, T2v, T2o);
cannam@127 597 T2F = FMA(KP866025403, T2E, T2D);
cannam@127 598 T2L = FNMS(KP866025403, T2E, T2D);
cannam@127 599 T2S = T2N * T2R;
cannam@127 600 T2P = T2N * T2O;
cannam@127 601 T2j = W[0];
cannam@127 602 T2y = W[1];
cannam@127 603 iio[WS(vs, 3) + WS(rs, 2)] = FNMS(T2Q, T2O, T2S);
cannam@127 604 rio[WS(vs, 3) + WS(rs, 2)] = FMA(T2Q, T2R, T2P);
cannam@127 605 T2G = T2j * T2F;
cannam@127 606 T2x = T2j * T2w;
cannam@127 607 }
cannam@127 608 }
cannam@127 609 {
cannam@127 610 E T1U, T21, T2H, T2K;
cannam@127 611 {
cannam@127 612 E T24, T27, T23, T26;
cannam@127 613 T1U = FNMS(KP866025403, T1T, T1Q);
cannam@127 614 T24 = FMA(KP866025403, T1T, T1Q);
cannam@127 615 T27 = FMA(KP866025403, T20, T1Z);
cannam@127 616 T21 = FNMS(KP866025403, T20, T1Z);
cannam@127 617 iio[WS(vs, 1) + WS(rs, 2)] = FNMS(T2y, T2w, T2G);
cannam@127 618 rio[WS(vs, 1) + WS(rs, 2)] = FMA(T2y, T2F, T2x);
cannam@127 619 T23 = W[6];
cannam@127 620 T26 = W[7];
cannam@127 621 {
cannam@127 622 E T42, T49, T44, T4a, T43, T28, T25, T3X;
cannam@127 623 T4c = FMA(KP866025403, T41, T3Y);
cannam@127 624 T42 = FNMS(KP866025403, T41, T3Y);
cannam@127 625 T49 = FNMS(KP866025403, T48, T47);
cannam@127 626 T4f = FMA(KP866025403, T48, T47);
cannam@127 627 T28 = T23 * T27;
cannam@127 628 T25 = T23 * T24;
cannam@127 629 T3X = W[2];
cannam@127 630 T44 = W[3];
cannam@127 631 iio[WS(vs, 4) + WS(rs, 1)] = FNMS(T26, T24, T28);
cannam@127 632 rio[WS(vs, 4) + WS(rs, 1)] = FMA(T26, T27, T25);
cannam@127 633 T4a = T3X * T49;
cannam@127 634 T43 = T3X * T42;
cannam@127 635 iio[WS(vs, 2) + WS(rs, 3)] = FNMS(T44, T42, T4a);
cannam@127 636 rio[WS(vs, 2) + WS(rs, 3)] = FMA(T44, T49, T43);
cannam@127 637 T2H = W[8];
cannam@127 638 T2K = W[9];
cannam@127 639 }
cannam@127 640 }
cannam@127 641 {
cannam@127 642 E T4S, T4U, T4R, T2M, T2J, T4P;
cannam@127 643 T2M = T2H * T2L;
cannam@127 644 T2J = T2H * T2I;
cannam@127 645 T4P = W[8];
cannam@127 646 T4S = W[9];
cannam@127 647 iio[WS(vs, 5) + WS(rs, 2)] = FNMS(T2K, T2I, T2M);
cannam@127 648 rio[WS(vs, 5) + WS(rs, 2)] = FMA(T2K, T2L, T2J);
cannam@127 649 T4U = T4P * T4T;
cannam@127 650 T4R = T4P * T4Q;
cannam@127 651 {
cannam@127 652 E T1P, T1W, T22, T1V, T4b;
cannam@127 653 iio[WS(vs, 5) + WS(rs, 4)] = FNMS(T4S, T4Q, T4U);
cannam@127 654 rio[WS(vs, 5) + WS(rs, 4)] = FMA(T4S, T4T, T4R);
cannam@127 655 T1P = W[2];
cannam@127 656 T1W = W[3];
cannam@127 657 T22 = T1P * T21;
cannam@127 658 T1V = T1P * T1U;
cannam@127 659 T4b = W[6];
cannam@127 660 T4e = W[7];
cannam@127 661 iio[WS(vs, 2) + WS(rs, 1)] = FNMS(T1W, T1U, T22);
cannam@127 662 rio[WS(vs, 2) + WS(rs, 1)] = FMA(T1W, T21, T1V);
cannam@127 663 T4g = T4b * T4f;
cannam@127 664 T4d = T4b * T4c;
cannam@127 665 }
cannam@127 666 }
cannam@127 667 }
cannam@127 668 }
cannam@127 669 }
cannam@127 670 }
cannam@127 671 }
cannam@127 672 iio[WS(vs, 4) + WS(rs, 3)] = FNMS(T4e, T4c, T4g);
cannam@127 673 rio[WS(vs, 4) + WS(rs, 3)] = FMA(T4e, T4f, T4d);
cannam@127 674 }
cannam@127 675 }
cannam@127 676 }
cannam@127 677
cannam@127 678 static const tw_instr twinstr[] = {
cannam@127 679 {TW_FULL, 0, 6},
cannam@127 680 {TW_NEXT, 1, 0}
cannam@127 681 };
cannam@127 682
cannam@127 683 static const ct_desc desc = { 6, "q1_6", twinstr, &GENUS, {144, 60, 132, 0}, 0, 0, 0 };
cannam@127 684
cannam@127 685 void X(codelet_q1_6) (planner *p) {
cannam@127 686 X(kdft_difsq_register) (p, q1_6, &desc);
cannam@127 687 }
cannam@127 688 #else /* HAVE_FMA */
cannam@127 689
cannam@127 690 /* Generated by: ../../../genfft/gen_twidsq.native -compact -variables 4 -pipeline-latency 4 -reload-twiddle -dif -n 6 -name q1_6 -include q.h */
cannam@127 691
cannam@127 692 /*
cannam@127 693 * This function contains 276 FP additions, 168 FP multiplications,
cannam@127 694 * (or, 192 additions, 84 multiplications, 84 fused multiply/add),
cannam@127 695 * 85 stack variables, 2 constants, and 144 memory accesses
cannam@127 696 */
cannam@127 697 #include "q.h"
cannam@127 698
cannam@127 699 static void q1_6(R *rio, R *iio, const R *W, stride rs, stride vs, INT mb, INT me, INT ms)
cannam@127 700 {
cannam@127 701 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
cannam@127 702 DK(KP866025403, +0.866025403784438646763723170752936183471402627);
cannam@127 703 {
cannam@127 704 INT m;
cannam@127 705 for (m = mb, W = W + (mb * 10); m < me; m = m + 1, rio = rio + ms, iio = iio + ms, W = W + 10, MAKE_VOLATILE_STRIDE(12, rs), MAKE_VOLATILE_STRIDE(0, vs)) {
cannam@127 706 E T3, Tc, Tt, TM, TX, T16, T1n, T1G, T2h, T2A, T1R, T20, T2L, T2U, T3b;
cannam@127 707 E T3u, T3F, T3O, T45, T4o, T4Z, T5i, T4z, T4I, Ta, TP, Tf, Tq, Tn, TN;
cannam@127 708 E Tu, TJ, T14, T1J, T19, T1k, T1h, T1H, T1o, T1D, T2b, T2B, T2i, T2x, T1Y;
cannam@127 709 E T2D, T23, T2e, T2S, T3x, T2X, T38, T35, T3v, T3c, T3r, T3M, T4r, T3R, T42;
cannam@127 710 E T3Z, T4p, T46, T4l, T4T, T5j, T50, T5f, T4G, T5l, T4L, T4W;
cannam@127 711 {
cannam@127 712 E T1, T2, T1l, T1m;
cannam@127 713 T1 = rio[0];
cannam@127 714 T2 = rio[WS(rs, 3)];
cannam@127 715 T3 = T1 + T2;
cannam@127 716 Tc = T1 - T2;
cannam@127 717 {
cannam@127 718 E Tr, Ts, TV, TW;
cannam@127 719 Tr = iio[0];
cannam@127 720 Ts = iio[WS(rs, 3)];
cannam@127 721 Tt = Tr - Ts;
cannam@127 722 TM = Tr + Ts;
cannam@127 723 TV = rio[WS(vs, 1)];
cannam@127 724 TW = rio[WS(vs, 1) + WS(rs, 3)];
cannam@127 725 TX = TV + TW;
cannam@127 726 T16 = TV - TW;
cannam@127 727 }
cannam@127 728 T1l = iio[WS(vs, 1)];
cannam@127 729 T1m = iio[WS(vs, 1) + WS(rs, 3)];
cannam@127 730 T1n = T1l - T1m;
cannam@127 731 T1G = T1l + T1m;
cannam@127 732 {
cannam@127 733 E T2f, T2g, T1P, T1Q;
cannam@127 734 T2f = iio[WS(vs, 2)];
cannam@127 735 T2g = iio[WS(vs, 2) + WS(rs, 3)];
cannam@127 736 T2h = T2f - T2g;
cannam@127 737 T2A = T2f + T2g;
cannam@127 738 T1P = rio[WS(vs, 2)];
cannam@127 739 T1Q = rio[WS(vs, 2) + WS(rs, 3)];
cannam@127 740 T1R = T1P + T1Q;
cannam@127 741 T20 = T1P - T1Q;
cannam@127 742 }
cannam@127 743 }
cannam@127 744 {
cannam@127 745 E T2J, T2K, T43, T44;
cannam@127 746 T2J = rio[WS(vs, 3)];
cannam@127 747 T2K = rio[WS(vs, 3) + WS(rs, 3)];
cannam@127 748 T2L = T2J + T2K;
cannam@127 749 T2U = T2J - T2K;
cannam@127 750 {
cannam@127 751 E T39, T3a, T3D, T3E;
cannam@127 752 T39 = iio[WS(vs, 3)];
cannam@127 753 T3a = iio[WS(vs, 3) + WS(rs, 3)];
cannam@127 754 T3b = T39 - T3a;
cannam@127 755 T3u = T39 + T3a;
cannam@127 756 T3D = rio[WS(vs, 4)];
cannam@127 757 T3E = rio[WS(vs, 4) + WS(rs, 3)];
cannam@127 758 T3F = T3D + T3E;
cannam@127 759 T3O = T3D - T3E;
cannam@127 760 }
cannam@127 761 T43 = iio[WS(vs, 4)];
cannam@127 762 T44 = iio[WS(vs, 4) + WS(rs, 3)];
cannam@127 763 T45 = T43 - T44;
cannam@127 764 T4o = T43 + T44;
cannam@127 765 {
cannam@127 766 E T4X, T4Y, T4x, T4y;
cannam@127 767 T4X = iio[WS(vs, 5)];
cannam@127 768 T4Y = iio[WS(vs, 5) + WS(rs, 3)];
cannam@127 769 T4Z = T4X - T4Y;
cannam@127 770 T5i = T4X + T4Y;
cannam@127 771 T4x = rio[WS(vs, 5)];
cannam@127 772 T4y = rio[WS(vs, 5) + WS(rs, 3)];
cannam@127 773 T4z = T4x + T4y;
cannam@127 774 T4I = T4x - T4y;
cannam@127 775 }
cannam@127 776 }
cannam@127 777 {
cannam@127 778 E T6, Td, T9, Te;
cannam@127 779 {
cannam@127 780 E T4, T5, T7, T8;
cannam@127 781 T4 = rio[WS(rs, 2)];
cannam@127 782 T5 = rio[WS(rs, 5)];
cannam@127 783 T6 = T4 + T5;
cannam@127 784 Td = T4 - T5;
cannam@127 785 T7 = rio[WS(rs, 4)];
cannam@127 786 T8 = rio[WS(rs, 1)];
cannam@127 787 T9 = T7 + T8;
cannam@127 788 Te = T7 - T8;
cannam@127 789 }
cannam@127 790 Ta = T6 + T9;
cannam@127 791 TP = KP866025403 * (T9 - T6);
cannam@127 792 Tf = Td + Te;
cannam@127 793 Tq = KP866025403 * (Te - Td);
cannam@127 794 }
cannam@127 795 {
cannam@127 796 E Tj, TH, Tm, TI;
cannam@127 797 {
cannam@127 798 E Th, Ti, Tk, Tl;
cannam@127 799 Th = iio[WS(rs, 2)];
cannam@127 800 Ti = iio[WS(rs, 5)];
cannam@127 801 Tj = Th - Ti;
cannam@127 802 TH = Th + Ti;
cannam@127 803 Tk = iio[WS(rs, 4)];
cannam@127 804 Tl = iio[WS(rs, 1)];
cannam@127 805 Tm = Tk - Tl;
cannam@127 806 TI = Tk + Tl;
cannam@127 807 }
cannam@127 808 Tn = KP866025403 * (Tj - Tm);
cannam@127 809 TN = TH + TI;
cannam@127 810 Tu = Tj + Tm;
cannam@127 811 TJ = KP866025403 * (TH - TI);
cannam@127 812 }
cannam@127 813 {
cannam@127 814 E T10, T17, T13, T18;
cannam@127 815 {
cannam@127 816 E TY, TZ, T11, T12;
cannam@127 817 TY = rio[WS(vs, 1) + WS(rs, 2)];
cannam@127 818 TZ = rio[WS(vs, 1) + WS(rs, 5)];
cannam@127 819 T10 = TY + TZ;
cannam@127 820 T17 = TY - TZ;
cannam@127 821 T11 = rio[WS(vs, 1) + WS(rs, 4)];
cannam@127 822 T12 = rio[WS(vs, 1) + WS(rs, 1)];
cannam@127 823 T13 = T11 + T12;
cannam@127 824 T18 = T11 - T12;
cannam@127 825 }
cannam@127 826 T14 = T10 + T13;
cannam@127 827 T1J = KP866025403 * (T13 - T10);
cannam@127 828 T19 = T17 + T18;
cannam@127 829 T1k = KP866025403 * (T18 - T17);
cannam@127 830 }
cannam@127 831 {
cannam@127 832 E T1d, T1B, T1g, T1C;
cannam@127 833 {
cannam@127 834 E T1b, T1c, T1e, T1f;
cannam@127 835 T1b = iio[WS(vs, 1) + WS(rs, 2)];
cannam@127 836 T1c = iio[WS(vs, 1) + WS(rs, 5)];
cannam@127 837 T1d = T1b - T1c;
cannam@127 838 T1B = T1b + T1c;
cannam@127 839 T1e = iio[WS(vs, 1) + WS(rs, 4)];
cannam@127 840 T1f = iio[WS(vs, 1) + WS(rs, 1)];
cannam@127 841 T1g = T1e - T1f;
cannam@127 842 T1C = T1e + T1f;
cannam@127 843 }
cannam@127 844 T1h = KP866025403 * (T1d - T1g);
cannam@127 845 T1H = T1B + T1C;
cannam@127 846 T1o = T1d + T1g;
cannam@127 847 T1D = KP866025403 * (T1B - T1C);
cannam@127 848 }
cannam@127 849 {
cannam@127 850 E T27, T2v, T2a, T2w;
cannam@127 851 {
cannam@127 852 E T25, T26, T28, T29;
cannam@127 853 T25 = iio[WS(vs, 2) + WS(rs, 2)];
cannam@127 854 T26 = iio[WS(vs, 2) + WS(rs, 5)];
cannam@127 855 T27 = T25 - T26;
cannam@127 856 T2v = T25 + T26;
cannam@127 857 T28 = iio[WS(vs, 2) + WS(rs, 4)];
cannam@127 858 T29 = iio[WS(vs, 2) + WS(rs, 1)];
cannam@127 859 T2a = T28 - T29;
cannam@127 860 T2w = T28 + T29;
cannam@127 861 }
cannam@127 862 T2b = KP866025403 * (T27 - T2a);
cannam@127 863 T2B = T2v + T2w;
cannam@127 864 T2i = T27 + T2a;
cannam@127 865 T2x = KP866025403 * (T2v - T2w);
cannam@127 866 }
cannam@127 867 {
cannam@127 868 E T1U, T21, T1X, T22;
cannam@127 869 {
cannam@127 870 E T1S, T1T, T1V, T1W;
cannam@127 871 T1S = rio[WS(vs, 2) + WS(rs, 2)];
cannam@127 872 T1T = rio[WS(vs, 2) + WS(rs, 5)];
cannam@127 873 T1U = T1S + T1T;
cannam@127 874 T21 = T1S - T1T;
cannam@127 875 T1V = rio[WS(vs, 2) + WS(rs, 4)];
cannam@127 876 T1W = rio[WS(vs, 2) + WS(rs, 1)];
cannam@127 877 T1X = T1V + T1W;
cannam@127 878 T22 = T1V - T1W;
cannam@127 879 }
cannam@127 880 T1Y = T1U + T1X;
cannam@127 881 T2D = KP866025403 * (T1X - T1U);
cannam@127 882 T23 = T21 + T22;
cannam@127 883 T2e = KP866025403 * (T22 - T21);
cannam@127 884 }
cannam@127 885 {
cannam@127 886 E T2O, T2V, T2R, T2W;
cannam@127 887 {
cannam@127 888 E T2M, T2N, T2P, T2Q;
cannam@127 889 T2M = rio[WS(vs, 3) + WS(rs, 2)];
cannam@127 890 T2N = rio[WS(vs, 3) + WS(rs, 5)];
cannam@127 891 T2O = T2M + T2N;
cannam@127 892 T2V = T2M - T2N;
cannam@127 893 T2P = rio[WS(vs, 3) + WS(rs, 4)];
cannam@127 894 T2Q = rio[WS(vs, 3) + WS(rs, 1)];
cannam@127 895 T2R = T2P + T2Q;
cannam@127 896 T2W = T2P - T2Q;
cannam@127 897 }
cannam@127 898 T2S = T2O + T2R;
cannam@127 899 T3x = KP866025403 * (T2R - T2O);
cannam@127 900 T2X = T2V + T2W;
cannam@127 901 T38 = KP866025403 * (T2W - T2V);
cannam@127 902 }
cannam@127 903 {
cannam@127 904 E T31, T3p, T34, T3q;
cannam@127 905 {
cannam@127 906 E T2Z, T30, T32, T33;
cannam@127 907 T2Z = iio[WS(vs, 3) + WS(rs, 2)];
cannam@127 908 T30 = iio[WS(vs, 3) + WS(rs, 5)];
cannam@127 909 T31 = T2Z - T30;
cannam@127 910 T3p = T2Z + T30;
cannam@127 911 T32 = iio[WS(vs, 3) + WS(rs, 4)];
cannam@127 912 T33 = iio[WS(vs, 3) + WS(rs, 1)];
cannam@127 913 T34 = T32 - T33;
cannam@127 914 T3q = T32 + T33;
cannam@127 915 }
cannam@127 916 T35 = KP866025403 * (T31 - T34);
cannam@127 917 T3v = T3p + T3q;
cannam@127 918 T3c = T31 + T34;
cannam@127 919 T3r = KP866025403 * (T3p - T3q);
cannam@127 920 }
cannam@127 921 {
cannam@127 922 E T3I, T3P, T3L, T3Q;
cannam@127 923 {
cannam@127 924 E T3G, T3H, T3J, T3K;
cannam@127 925 T3G = rio[WS(vs, 4) + WS(rs, 2)];
cannam@127 926 T3H = rio[WS(vs, 4) + WS(rs, 5)];
cannam@127 927 T3I = T3G + T3H;
cannam@127 928 T3P = T3G - T3H;
cannam@127 929 T3J = rio[WS(vs, 4) + WS(rs, 4)];
cannam@127 930 T3K = rio[WS(vs, 4) + WS(rs, 1)];
cannam@127 931 T3L = T3J + T3K;
cannam@127 932 T3Q = T3J - T3K;
cannam@127 933 }
cannam@127 934 T3M = T3I + T3L;
cannam@127 935 T4r = KP866025403 * (T3L - T3I);
cannam@127 936 T3R = T3P + T3Q;
cannam@127 937 T42 = KP866025403 * (T3Q - T3P);
cannam@127 938 }
cannam@127 939 {
cannam@127 940 E T3V, T4j, T3Y, T4k;
cannam@127 941 {
cannam@127 942 E T3T, T3U, T3W, T3X;
cannam@127 943 T3T = iio[WS(vs, 4) + WS(rs, 2)];
cannam@127 944 T3U = iio[WS(vs, 4) + WS(rs, 5)];
cannam@127 945 T3V = T3T - T3U;
cannam@127 946 T4j = T3T + T3U;
cannam@127 947 T3W = iio[WS(vs, 4) + WS(rs, 4)];
cannam@127 948 T3X = iio[WS(vs, 4) + WS(rs, 1)];
cannam@127 949 T3Y = T3W - T3X;
cannam@127 950 T4k = T3W + T3X;
cannam@127 951 }
cannam@127 952 T3Z = KP866025403 * (T3V - T3Y);
cannam@127 953 T4p = T4j + T4k;
cannam@127 954 T46 = T3V + T3Y;
cannam@127 955 T4l = KP866025403 * (T4j - T4k);
cannam@127 956 }
cannam@127 957 {
cannam@127 958 E T4P, T5d, T4S, T5e;
cannam@127 959 {
cannam@127 960 E T4N, T4O, T4Q, T4R;
cannam@127 961 T4N = iio[WS(vs, 5) + WS(rs, 2)];
cannam@127 962 T4O = iio[WS(vs, 5) + WS(rs, 5)];
cannam@127 963 T4P = T4N - T4O;
cannam@127 964 T5d = T4N + T4O;
cannam@127 965 T4Q = iio[WS(vs, 5) + WS(rs, 4)];
cannam@127 966 T4R = iio[WS(vs, 5) + WS(rs, 1)];
cannam@127 967 T4S = T4Q - T4R;
cannam@127 968 T5e = T4Q + T4R;
cannam@127 969 }
cannam@127 970 T4T = KP866025403 * (T4P - T4S);
cannam@127 971 T5j = T5d + T5e;
cannam@127 972 T50 = T4P + T4S;
cannam@127 973 T5f = KP866025403 * (T5d - T5e);
cannam@127 974 }
cannam@127 975 {
cannam@127 976 E T4C, T4J, T4F, T4K;
cannam@127 977 {
cannam@127 978 E T4A, T4B, T4D, T4E;
cannam@127 979 T4A = rio[WS(vs, 5) + WS(rs, 2)];
cannam@127 980 T4B = rio[WS(vs, 5) + WS(rs, 5)];
cannam@127 981 T4C = T4A + T4B;
cannam@127 982 T4J = T4A - T4B;
cannam@127 983 T4D = rio[WS(vs, 5) + WS(rs, 4)];
cannam@127 984 T4E = rio[WS(vs, 5) + WS(rs, 1)];
cannam@127 985 T4F = T4D + T4E;
cannam@127 986 T4K = T4D - T4E;
cannam@127 987 }
cannam@127 988 T4G = T4C + T4F;
cannam@127 989 T5l = KP866025403 * (T4F - T4C);
cannam@127 990 T4L = T4J + T4K;
cannam@127 991 T4W = KP866025403 * (T4K - T4J);
cannam@127 992 }
cannam@127 993 rio[0] = T3 + Ta;
cannam@127 994 iio[0] = TM + TN;
cannam@127 995 rio[WS(rs, 1)] = TX + T14;
cannam@127 996 iio[WS(rs, 1)] = T1G + T1H;
cannam@127 997 rio[WS(rs, 3)] = T2L + T2S;
cannam@127 998 rio[WS(rs, 2)] = T1R + T1Y;
cannam@127 999 iio[WS(rs, 2)] = T2A + T2B;
cannam@127 1000 iio[WS(rs, 3)] = T3u + T3v;
cannam@127 1001 iio[WS(rs, 4)] = T4o + T4p;
cannam@127 1002 iio[WS(rs, 5)] = T5i + T5j;
cannam@127 1003 rio[WS(rs, 5)] = T4z + T4G;
cannam@127 1004 rio[WS(rs, 4)] = T3F + T3M;
cannam@127 1005 {
cannam@127 1006 E T1w, T1y, T1v, T1x;
cannam@127 1007 T1w = T16 + T19;
cannam@127 1008 T1y = T1n + T1o;
cannam@127 1009 T1v = W[4];
cannam@127 1010 T1x = W[5];
cannam@127 1011 rio[WS(vs, 3) + WS(rs, 1)] = FMA(T1v, T1w, T1x * T1y);
cannam@127 1012 iio[WS(vs, 3) + WS(rs, 1)] = FNMS(T1x, T1w, T1v * T1y);
cannam@127 1013 }
cannam@127 1014 {
cannam@127 1015 E T58, T5a, T57, T59;
cannam@127 1016 T58 = T4I + T4L;
cannam@127 1017 T5a = T4Z + T50;
cannam@127 1018 T57 = W[4];
cannam@127 1019 T59 = W[5];
cannam@127 1020 rio[WS(vs, 3) + WS(rs, 5)] = FMA(T57, T58, T59 * T5a);
cannam@127 1021 iio[WS(vs, 3) + WS(rs, 5)] = FNMS(T59, T58, T57 * T5a);
cannam@127 1022 }
cannam@127 1023 {
cannam@127 1024 E TC, TE, TB, TD;
cannam@127 1025 TC = Tc + Tf;
cannam@127 1026 TE = Tt + Tu;
cannam@127 1027 TB = W[4];
cannam@127 1028 TD = W[5];
cannam@127 1029 rio[WS(vs, 3)] = FMA(TB, TC, TD * TE);
cannam@127 1030 iio[WS(vs, 3)] = FNMS(TD, TC, TB * TE);
cannam@127 1031 }
cannam@127 1032 {
cannam@127 1033 E T4e, T4g, T4d, T4f;
cannam@127 1034 T4e = T3O + T3R;
cannam@127 1035 T4g = T45 + T46;
cannam@127 1036 T4d = W[4];
cannam@127 1037 T4f = W[5];
cannam@127 1038 rio[WS(vs, 3) + WS(rs, 4)] = FMA(T4d, T4e, T4f * T4g);
cannam@127 1039 iio[WS(vs, 3) + WS(rs, 4)] = FNMS(T4f, T4e, T4d * T4g);
cannam@127 1040 }
cannam@127 1041 {
cannam@127 1042 E T3k, T3m, T3j, T3l;
cannam@127 1043 T3k = T2U + T2X;
cannam@127 1044 T3m = T3b + T3c;
cannam@127 1045 T3j = W[4];
cannam@127 1046 T3l = W[5];
cannam@127 1047 rio[WS(vs, 3) + WS(rs, 3)] = FMA(T3j, T3k, T3l * T3m);
cannam@127 1048 iio[WS(vs, 3) + WS(rs, 3)] = FNMS(T3l, T3k, T3j * T3m);
cannam@127 1049 }
cannam@127 1050 {
cannam@127 1051 E T2q, T2s, T2p, T2r;
cannam@127 1052 T2q = T20 + T23;
cannam@127 1053 T2s = T2h + T2i;
cannam@127 1054 T2p = W[4];
cannam@127 1055 T2r = W[5];
cannam@127 1056 rio[WS(vs, 3) + WS(rs, 2)] = FMA(T2p, T2q, T2r * T2s);
cannam@127 1057 iio[WS(vs, 3) + WS(rs, 2)] = FNMS(T2r, T2q, T2p * T2s);
cannam@127 1058 }
cannam@127 1059 {
cannam@127 1060 E T5g, T5o, T5m, T5q, T5c, T5k;
cannam@127 1061 T5c = FNMS(KP500000000, T4G, T4z);
cannam@127 1062 T5g = T5c - T5f;
cannam@127 1063 T5o = T5c + T5f;
cannam@127 1064 T5k = FNMS(KP500000000, T5j, T5i);
cannam@127 1065 T5m = T5k - T5l;
cannam@127 1066 T5q = T5l + T5k;
cannam@127 1067 {
cannam@127 1068 E T5b, T5h, T5n, T5p;
cannam@127 1069 T5b = W[2];
cannam@127 1070 T5h = W[3];
cannam@127 1071 rio[WS(vs, 2) + WS(rs, 5)] = FMA(T5b, T5g, T5h * T5m);
cannam@127 1072 iio[WS(vs, 2) + WS(rs, 5)] = FNMS(T5h, T5g, T5b * T5m);
cannam@127 1073 T5n = W[6];
cannam@127 1074 T5p = W[7];
cannam@127 1075 rio[WS(vs, 4) + WS(rs, 5)] = FMA(T5n, T5o, T5p * T5q);
cannam@127 1076 iio[WS(vs, 4) + WS(rs, 5)] = FNMS(T5p, T5o, T5n * T5q);
cannam@127 1077 }
cannam@127 1078 }
cannam@127 1079 {
cannam@127 1080 E To, Ty, Tw, TA, Tg, Tv;
cannam@127 1081 Tg = FNMS(KP500000000, Tf, Tc);
cannam@127 1082 To = Tg + Tn;
cannam@127 1083 Ty = Tg - Tn;
cannam@127 1084 Tv = FNMS(KP500000000, Tu, Tt);
cannam@127 1085 Tw = Tq + Tv;
cannam@127 1086 TA = Tv - Tq;
cannam@127 1087 {
cannam@127 1088 E Tb, Tp, Tx, Tz;
cannam@127 1089 Tb = W[0];
cannam@127 1090 Tp = W[1];
cannam@127 1091 rio[WS(vs, 1)] = FMA(Tb, To, Tp * Tw);
cannam@127 1092 iio[WS(vs, 1)] = FNMS(Tp, To, Tb * Tw);
cannam@127 1093 Tx = W[8];
cannam@127 1094 Tz = W[9];
cannam@127 1095 rio[WS(vs, 5)] = FMA(Tx, Ty, Tz * TA);
cannam@127 1096 iio[WS(vs, 5)] = FNMS(Tz, Ty, Tx * TA);
cannam@127 1097 }
cannam@127 1098 }
cannam@127 1099 {
cannam@127 1100 E T36, T3g, T3e, T3i, T2Y, T3d;
cannam@127 1101 T2Y = FNMS(KP500000000, T2X, T2U);
cannam@127 1102 T36 = T2Y + T35;
cannam@127 1103 T3g = T2Y - T35;
cannam@127 1104 T3d = FNMS(KP500000000, T3c, T3b);
cannam@127 1105 T3e = T38 + T3d;
cannam@127 1106 T3i = T3d - T38;
cannam@127 1107 {
cannam@127 1108 E T2T, T37, T3f, T3h;
cannam@127 1109 T2T = W[0];
cannam@127 1110 T37 = W[1];
cannam@127 1111 rio[WS(vs, 1) + WS(rs, 3)] = FMA(T2T, T36, T37 * T3e);
cannam@127 1112 iio[WS(vs, 1) + WS(rs, 3)] = FNMS(T37, T36, T2T * T3e);
cannam@127 1113 T3f = W[8];
cannam@127 1114 T3h = W[9];
cannam@127 1115 rio[WS(vs, 5) + WS(rs, 3)] = FMA(T3f, T3g, T3h * T3i);
cannam@127 1116 iio[WS(vs, 5) + WS(rs, 3)] = FNMS(T3h, T3g, T3f * T3i);
cannam@127 1117 }
cannam@127 1118 }
cannam@127 1119 {
cannam@127 1120 E T2y, T2G, T2E, T2I, T2u, T2C;
cannam@127 1121 T2u = FNMS(KP500000000, T1Y, T1R);
cannam@127 1122 T2y = T2u - T2x;
cannam@127 1123 T2G = T2u + T2x;
cannam@127 1124 T2C = FNMS(KP500000000, T2B, T2A);
cannam@127 1125 T2E = T2C - T2D;
cannam@127 1126 T2I = T2D + T2C;
cannam@127 1127 {
cannam@127 1128 E T2t, T2z, T2F, T2H;
cannam@127 1129 T2t = W[2];
cannam@127 1130 T2z = W[3];
cannam@127 1131 rio[WS(vs, 2) + WS(rs, 2)] = FMA(T2t, T2y, T2z * T2E);
cannam@127 1132 iio[WS(vs, 2) + WS(rs, 2)] = FNMS(T2z, T2y, T2t * T2E);
cannam@127 1133 T2F = W[6];
cannam@127 1134 T2H = W[7];
cannam@127 1135 rio[WS(vs, 4) + WS(rs, 2)] = FMA(T2F, T2G, T2H * T2I);
cannam@127 1136 iio[WS(vs, 4) + WS(rs, 2)] = FNMS(T2H, T2G, T2F * T2I);
cannam@127 1137 }
cannam@127 1138 }
cannam@127 1139 {
cannam@127 1140 E T3s, T3A, T3y, T3C, T3o, T3w;
cannam@127 1141 T3o = FNMS(KP500000000, T2S, T2L);
cannam@127 1142 T3s = T3o - T3r;
cannam@127 1143 T3A = T3o + T3r;
cannam@127 1144 T3w = FNMS(KP500000000, T3v, T3u);
cannam@127 1145 T3y = T3w - T3x;
cannam@127 1146 T3C = T3x + T3w;
cannam@127 1147 {
cannam@127 1148 E T3n, T3t, T3z, T3B;
cannam@127 1149 T3n = W[2];
cannam@127 1150 T3t = W[3];
cannam@127 1151 rio[WS(vs, 2) + WS(rs, 3)] = FMA(T3n, T3s, T3t * T3y);
cannam@127 1152 iio[WS(vs, 2) + WS(rs, 3)] = FNMS(T3t, T3s, T3n * T3y);
cannam@127 1153 T3z = W[6];
cannam@127 1154 T3B = W[7];
cannam@127 1155 rio[WS(vs, 4) + WS(rs, 3)] = FMA(T3z, T3A, T3B * T3C);
cannam@127 1156 iio[WS(vs, 4) + WS(rs, 3)] = FNMS(T3B, T3A, T3z * T3C);
cannam@127 1157 }
cannam@127 1158 }
cannam@127 1159 {
cannam@127 1160 E T1E, T1M, T1K, T1O, T1A, T1I;
cannam@127 1161 T1A = FNMS(KP500000000, T14, TX);
cannam@127 1162 T1E = T1A - T1D;
cannam@127 1163 T1M = T1A + T1D;
cannam@127 1164 T1I = FNMS(KP500000000, T1H, T1G);
cannam@127 1165 T1K = T1I - T1J;
cannam@127 1166 T1O = T1J + T1I;
cannam@127 1167 {
cannam@127 1168 E T1z, T1F, T1L, T1N;
cannam@127 1169 T1z = W[2];
cannam@127 1170 T1F = W[3];
cannam@127 1171 rio[WS(vs, 2) + WS(rs, 1)] = FMA(T1z, T1E, T1F * T1K);
cannam@127 1172 iio[WS(vs, 2) + WS(rs, 1)] = FNMS(T1F, T1E, T1z * T1K);
cannam@127 1173 T1L = W[6];
cannam@127 1174 T1N = W[7];
cannam@127 1175 rio[WS(vs, 4) + WS(rs, 1)] = FMA(T1L, T1M, T1N * T1O);
cannam@127 1176 iio[WS(vs, 4) + WS(rs, 1)] = FNMS(T1N, T1M, T1L * T1O);
cannam@127 1177 }
cannam@127 1178 }
cannam@127 1179 {
cannam@127 1180 E T4m, T4u, T4s, T4w, T4i, T4q;
cannam@127 1181 T4i = FNMS(KP500000000, T3M, T3F);
cannam@127 1182 T4m = T4i - T4l;
cannam@127 1183 T4u = T4i + T4l;
cannam@127 1184 T4q = FNMS(KP500000000, T4p, T4o);
cannam@127 1185 T4s = T4q - T4r;
cannam@127 1186 T4w = T4r + T4q;
cannam@127 1187 {
cannam@127 1188 E T4h, T4n, T4t, T4v;
cannam@127 1189 T4h = W[2];
cannam@127 1190 T4n = W[3];
cannam@127 1191 rio[WS(vs, 2) + WS(rs, 4)] = FMA(T4h, T4m, T4n * T4s);
cannam@127 1192 iio[WS(vs, 2) + WS(rs, 4)] = FNMS(T4n, T4m, T4h * T4s);
cannam@127 1193 T4t = W[6];
cannam@127 1194 T4v = W[7];
cannam@127 1195 rio[WS(vs, 4) + WS(rs, 4)] = FMA(T4t, T4u, T4v * T4w);
cannam@127 1196 iio[WS(vs, 4) + WS(rs, 4)] = FNMS(T4v, T4u, T4t * T4w);
cannam@127 1197 }
cannam@127 1198 }
cannam@127 1199 {
cannam@127 1200 E TK, TS, TQ, TU, TG, TO;
cannam@127 1201 TG = FNMS(KP500000000, Ta, T3);
cannam@127 1202 TK = TG - TJ;
cannam@127 1203 TS = TG + TJ;
cannam@127 1204 TO = FNMS(KP500000000, TN, TM);
cannam@127 1205 TQ = TO - TP;
cannam@127 1206 TU = TP + TO;
cannam@127 1207 {
cannam@127 1208 E TF, TL, TR, TT;
cannam@127 1209 TF = W[2];
cannam@127 1210 TL = W[3];
cannam@127 1211 rio[WS(vs, 2)] = FMA(TF, TK, TL * TQ);
cannam@127 1212 iio[WS(vs, 2)] = FNMS(TL, TK, TF * TQ);
cannam@127 1213 TR = W[6];
cannam@127 1214 TT = W[7];
cannam@127 1215 rio[WS(vs, 4)] = FMA(TR, TS, TT * TU);
cannam@127 1216 iio[WS(vs, 4)] = FNMS(TT, TS, TR * TU);
cannam@127 1217 }
cannam@127 1218 }
cannam@127 1219 {
cannam@127 1220 E T2c, T2m, T2k, T2o, T24, T2j;
cannam@127 1221 T24 = FNMS(KP500000000, T23, T20);
cannam@127 1222 T2c = T24 + T2b;
cannam@127 1223 T2m = T24 - T2b;
cannam@127 1224 T2j = FNMS(KP500000000, T2i, T2h);
cannam@127 1225 T2k = T2e + T2j;
cannam@127 1226 T2o = T2j - T2e;
cannam@127 1227 {
cannam@127 1228 E T1Z, T2d, T2l, T2n;
cannam@127 1229 T1Z = W[0];
cannam@127 1230 T2d = W[1];
cannam@127 1231 rio[WS(vs, 1) + WS(rs, 2)] = FMA(T1Z, T2c, T2d * T2k);
cannam@127 1232 iio[WS(vs, 1) + WS(rs, 2)] = FNMS(T2d, T2c, T1Z * T2k);
cannam@127 1233 T2l = W[8];
cannam@127 1234 T2n = W[9];
cannam@127 1235 rio[WS(vs, 5) + WS(rs, 2)] = FMA(T2l, T2m, T2n * T2o);
cannam@127 1236 iio[WS(vs, 5) + WS(rs, 2)] = FNMS(T2n, T2m, T2l * T2o);
cannam@127 1237 }
cannam@127 1238 }
cannam@127 1239 {
cannam@127 1240 E T40, T4a, T48, T4c, T3S, T47;
cannam@127 1241 T3S = FNMS(KP500000000, T3R, T3O);
cannam@127 1242 T40 = T3S + T3Z;
cannam@127 1243 T4a = T3S - T3Z;
cannam@127 1244 T47 = FNMS(KP500000000, T46, T45);
cannam@127 1245 T48 = T42 + T47;
cannam@127 1246 T4c = T47 - T42;
cannam@127 1247 {
cannam@127 1248 E T3N, T41, T49, T4b;
cannam@127 1249 T3N = W[0];
cannam@127 1250 T41 = W[1];
cannam@127 1251 rio[WS(vs, 1) + WS(rs, 4)] = FMA(T3N, T40, T41 * T48);
cannam@127 1252 iio[WS(vs, 1) + WS(rs, 4)] = FNMS(T41, T40, T3N * T48);
cannam@127 1253 T49 = W[8];
cannam@127 1254 T4b = W[9];
cannam@127 1255 rio[WS(vs, 5) + WS(rs, 4)] = FMA(T49, T4a, T4b * T4c);
cannam@127 1256 iio[WS(vs, 5) + WS(rs, 4)] = FNMS(T4b, T4a, T49 * T4c);
cannam@127 1257 }
cannam@127 1258 }
cannam@127 1259 {
cannam@127 1260 E T1i, T1s, T1q, T1u, T1a, T1p;
cannam@127 1261 T1a = FNMS(KP500000000, T19, T16);
cannam@127 1262 T1i = T1a + T1h;
cannam@127 1263 T1s = T1a - T1h;
cannam@127 1264 T1p = FNMS(KP500000000, T1o, T1n);
cannam@127 1265 T1q = T1k + T1p;
cannam@127 1266 T1u = T1p - T1k;
cannam@127 1267 {
cannam@127 1268 E T15, T1j, T1r, T1t;
cannam@127 1269 T15 = W[0];
cannam@127 1270 T1j = W[1];
cannam@127 1271 rio[WS(vs, 1) + WS(rs, 1)] = FMA(T15, T1i, T1j * T1q);
cannam@127 1272 iio[WS(vs, 1) + WS(rs, 1)] = FNMS(T1j, T1i, T15 * T1q);
cannam@127 1273 T1r = W[8];
cannam@127 1274 T1t = W[9];
cannam@127 1275 rio[WS(vs, 5) + WS(rs, 1)] = FMA(T1r, T1s, T1t * T1u);
cannam@127 1276 iio[WS(vs, 5) + WS(rs, 1)] = FNMS(T1t, T1s, T1r * T1u);
cannam@127 1277 }
cannam@127 1278 }
cannam@127 1279 {
cannam@127 1280 E T4U, T54, T52, T56, T4M, T51;
cannam@127 1281 T4M = FNMS(KP500000000, T4L, T4I);
cannam@127 1282 T4U = T4M + T4T;
cannam@127 1283 T54 = T4M - T4T;
cannam@127 1284 T51 = FNMS(KP500000000, T50, T4Z);
cannam@127 1285 T52 = T4W + T51;
cannam@127 1286 T56 = T51 - T4W;
cannam@127 1287 {
cannam@127 1288 E T4H, T4V, T53, T55;
cannam@127 1289 T4H = W[0];
cannam@127 1290 T4V = W[1];
cannam@127 1291 rio[WS(vs, 1) + WS(rs, 5)] = FMA(T4H, T4U, T4V * T52);
cannam@127 1292 iio[WS(vs, 1) + WS(rs, 5)] = FNMS(T4V, T4U, T4H * T52);
cannam@127 1293 T53 = W[8];
cannam@127 1294 T55 = W[9];
cannam@127 1295 rio[WS(vs, 5) + WS(rs, 5)] = FMA(T53, T54, T55 * T56);
cannam@127 1296 iio[WS(vs, 5) + WS(rs, 5)] = FNMS(T55, T54, T53 * T56);
cannam@127 1297 }
cannam@127 1298 }
cannam@127 1299 }
cannam@127 1300 }
cannam@127 1301 }
cannam@127 1302
cannam@127 1303 static const tw_instr twinstr[] = {
cannam@127 1304 {TW_FULL, 0, 6},
cannam@127 1305 {TW_NEXT, 1, 0}
cannam@127 1306 };
cannam@127 1307
cannam@127 1308 static const ct_desc desc = { 6, "q1_6", twinstr, &GENUS, {192, 84, 84, 0}, 0, 0, 0 };
cannam@127 1309
cannam@127 1310 void X(codelet_q1_6) (planner *p) {
cannam@127 1311 X(kdft_difsq_register) (p, q1_6, &desc);
cannam@127 1312 }
cannam@127 1313 #endif /* HAVE_FMA */