annotate src/fftw-3.3.3/rdft/scalar/r2cb/r2cb_128.c @ 95:89f5e221ed7b

Add FFTW3
author Chris Cannam <cannam@all-day-breakfast.com>
date Wed, 20 Mar 2013 15:35:50 +0000
parents
children
rev   line source
cannam@95 1 /*
cannam@95 2 * Copyright (c) 2003, 2007-11 Matteo Frigo
cannam@95 3 * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
cannam@95 4 *
cannam@95 5 * This program is free software; you can redistribute it and/or modify
cannam@95 6 * it under the terms of the GNU General Public License as published by
cannam@95 7 * the Free Software Foundation; either version 2 of the License, or
cannam@95 8 * (at your option) any later version.
cannam@95 9 *
cannam@95 10 * This program is distributed in the hope that it will be useful,
cannam@95 11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
cannam@95 12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
cannam@95 13 * GNU General Public License for more details.
cannam@95 14 *
cannam@95 15 * You should have received a copy of the GNU General Public License
cannam@95 16 * along with this program; if not, write to the Free Software
cannam@95 17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
cannam@95 18 *
cannam@95 19 */
cannam@95 20
cannam@95 21 /* This file was automatically generated --- DO NOT EDIT */
cannam@95 22 /* Generated on Sun Nov 25 07:41:09 EST 2012 */
cannam@95 23
cannam@95 24 #include "codelet-rdft.h"
cannam@95 25
cannam@95 26 #ifdef HAVE_FMA
cannam@95 27
cannam@95 28 /* Generated by: ../../../genfft/gen_r2cb.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -sign 1 -n 128 -name r2cb_128 -include r2cb.h */
cannam@95 29
cannam@95 30 /*
cannam@95 31 * This function contains 956 FP additions, 540 FP multiplications,
cannam@95 32 * (or, 416 additions, 0 multiplications, 540 fused multiply/add),
cannam@95 33 * 242 stack variables, 36 constants, and 256 memory accesses
cannam@95 34 */
cannam@95 35 #include "r2cb.h"
cannam@95 36
cannam@95 37 static void r2cb_128(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
cannam@95 38 {
cannam@95 39 DK(KP1_715457220, +1.715457220000544139804539968569540274084981599);
cannam@95 40 DK(KP1_606415062, +1.606415062961289819613353025926283847759138854);
cannam@95 41 DK(KP599376933, +0.599376933681923766271389869014404232837890546);
cannam@95 42 DK(KP741650546, +0.741650546272035369581266691172079863842265220);
cannam@95 43 DK(KP1_978353019, +1.978353019929561946903347476032486127967379067);
cannam@95 44 DK(KP1_940062506, +1.940062506389087985207968414572200502913731924);
cannam@95 45 DK(KP148335987, +0.148335987538347428753676511486911367000625355);
cannam@95 46 DK(KP250486960, +0.250486960191305461595702160124721208578685568);
cannam@95 47 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
cannam@95 48 DK(KP1_807978586, +1.807978586246886663172400594461074097420264050);
cannam@95 49 DK(KP1_481902250, +1.481902250709918182351233794990325459457910619);
cannam@95 50 DK(KP472964775, +0.472964775891319928124438237972992463904131113);
cannam@95 51 DK(KP906347169, +0.906347169019147157946142717268914412664134293);
cannam@95 52 DK(KP1_997590912, +1.997590912410344785429543209518201388886407229);
cannam@95 53 DK(KP1_883088130, +1.883088130366041556825018805199004714371179592);
cannam@95 54 DK(KP049126849, +0.049126849769467254105343321271313617079695752);
cannam@95 55 DK(KP357805721, +0.357805721314524104672487743774474392487532769);
cannam@95 56 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
cannam@95 57 DK(KP1_763842528, +1.763842528696710059425513727320776699016885241);
cannam@95 58 DK(KP1_913880671, +1.913880671464417729871595773960539938965698411);
cannam@95 59 DK(KP534511135, +0.534511135950791641089685961295362908582039528);
cannam@95 60 DK(KP303346683, +0.303346683607342391675883946941299872384187453);
cannam@95 61 DK(KP1_662939224, +1.662939224605090474157576755235811513477121624);
cannam@95 62 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
cannam@95 63 DK(KP1_990369453, +1.990369453344393772489673906218959843150949737);
cannam@95 64 DK(KP1_546020906, +1.546020906725473921621813219516939601942082586);
cannam@95 65 DK(KP098491403, +0.098491403357164253077197521291327432293052451);
cannam@95 66 DK(KP820678790, +0.820678790828660330972281985331011598767386482);
cannam@95 67 DK(KP1_961570560, +1.961570560806460898252364472268478073947867462);
cannam@95 68 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
cannam@95 69 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
cannam@95 70 DK(KP1_847759065, +1.847759065022573512256366378793576573644833252);
cannam@95 71 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
cannam@95 72 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
cannam@95 73 DK(KP1_414213562, +1.414213562373095048801688724209698078569671875);
cannam@95 74 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
cannam@95 75 {
cannam@95 76 INT i;
cannam@95 77 for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(512, rs), MAKE_VOLATILE_STRIDE(512, csr), MAKE_VOLATILE_STRIDE(512, csi)) {
cannam@95 78 E T9H, T9I, T9X, T9Y;
cannam@95 79 {
cannam@95 80 E Tdr, T9, Tcl, Ta9, T6b, T2d, T91, T7j, Tg, Tds, Tcm, Tae, T92, T7m, T6c;
cannam@95 81 E T2o, Tdu, Tw, Tco, Tap, TeM, Tdx, T6f, T2G, T6e, T2P, T94, T7t, Tcp, Tak;
cannam@95 82 E T95, T7q, TdM, T1i, TcL, TbD, Tf0, Te6, T6q, T42, T6B, T5t, T9r, T8j, TcA;
cannam@95 83 E TaY, T9g, T7S, TdA, TM, Tcv, TaN, TeP, TdI, T6i, T38, T6l, T3F, T9b, T7J;
cannam@95 84 E Tcs, Taw, T98, T7y, T1N, TeW, T6x, T4H, Te8, TdV, T6w, T4Q, T9j, T86, TcO;
cannam@95 85 E TcI, T9k, T83, TbI, Tbl, T22, TeV, Te0, Te9, T58, T6u, T6t, T5h, T9m, T8d;
cannam@95 86 E TcP, TcF, T9n, T8a, TbJ, Tbw, Te3, T1x, TcB, TbG, Tf1, TdP, T6C, T4p, T6r;
cannam@95 87 E T5w, T9h, T8m, TcM, Tb9, T9s, T7Z, TaB, TaG, TdF, T11, Tct, TaQ, TeQ, TdD;
cannam@95 88 E T6m, T3v, T7B, T7E, T6j, T3I, T99, T7M;
cannam@95 89 {
cannam@95 90 E TaU, TaX, T7Q, T7R, Tbk, Tbf;
cannam@95 91 {
cannam@95 92 E Td, T2e, Tc, Tab, T2m, Te, T2f, T2g;
cannam@95 93 {
cannam@95 94 E T7h, T27, T2c, T7i;
cannam@95 95 {
cannam@95 96 E T4, T26, T29, T25, T3, T28, T8, T2a;
cannam@95 97 T4 = Cr[WS(csr, 32)];
cannam@95 98 T26 = Ci[WS(csi, 32)];
cannam@95 99 {
cannam@95 100 E T1, T2, T6, T7;
cannam@95 101 T1 = Cr[0];
cannam@95 102 T2 = Cr[WS(csr, 64)];
cannam@95 103 T6 = Cr[WS(csr, 16)];
cannam@95 104 T7 = Cr[WS(csr, 48)];
cannam@95 105 T29 = Ci[WS(csi, 16)];
cannam@95 106 T25 = T1 - T2;
cannam@95 107 T3 = T1 + T2;
cannam@95 108 T28 = T6 - T7;
cannam@95 109 T8 = T6 + T7;
cannam@95 110 T2a = Ci[WS(csi, 48)];
cannam@95 111 }
cannam@95 112 {
cannam@95 113 E Ta7, T5, Ta8, T2b;
cannam@95 114 Ta7 = FNMS(KP2_000000000, T4, T3);
cannam@95 115 T5 = FMA(KP2_000000000, T4, T3);
cannam@95 116 T7h = FMA(KP2_000000000, T26, T25);
cannam@95 117 T27 = FNMS(KP2_000000000, T26, T25);
cannam@95 118 Ta8 = T29 - T2a;
cannam@95 119 T2b = T29 + T2a;
cannam@95 120 Tdr = FNMS(KP2_000000000, T8, T5);
cannam@95 121 T9 = FMA(KP2_000000000, T8, T5);
cannam@95 122 Tcl = FMA(KP2_000000000, Ta8, Ta7);
cannam@95 123 Ta9 = FNMS(KP2_000000000, Ta8, Ta7);
cannam@95 124 T2c = T28 - T2b;
cannam@95 125 T7i = T28 + T2b;
cannam@95 126 }
cannam@95 127 }
cannam@95 128 {
cannam@95 129 E Ta, Tb, T2k, T2l;
cannam@95 130 Ta = Cr[WS(csr, 8)];
cannam@95 131 T6b = FNMS(KP1_414213562, T2c, T27);
cannam@95 132 T2d = FMA(KP1_414213562, T2c, T27);
cannam@95 133 T91 = FMA(KP1_414213562, T7i, T7h);
cannam@95 134 T7j = FNMS(KP1_414213562, T7i, T7h);
cannam@95 135 Tb = Cr[WS(csr, 56)];
cannam@95 136 T2k = Ci[WS(csi, 8)];
cannam@95 137 T2l = Ci[WS(csi, 56)];
cannam@95 138 Td = Cr[WS(csr, 40)];
cannam@95 139 T2e = Ta - Tb;
cannam@95 140 Tc = Ta + Tb;
cannam@95 141 Tab = T2k - T2l;
cannam@95 142 T2m = T2k + T2l;
cannam@95 143 Te = Cr[WS(csr, 24)];
cannam@95 144 T2f = Ci[WS(csi, 40)];
cannam@95 145 T2g = Ci[WS(csi, 24)];
cannam@95 146 }
cannam@95 147 }
cannam@95 148 {
cannam@95 149 E Tag, Taj, T7o, T7p;
cannam@95 150 {
cannam@95 151 E T2q, Tk, Tam, T2K, T2H, Tn, Tan, T2t, Tu, Tah, T2E, T2N, Tr, T2v, T2y;
cannam@95 152 E Tai;
cannam@95 153 {
cannam@95 154 E Tl, Tm, T2r, T2s;
cannam@95 155 {
cannam@95 156 E Ti, Tj, T2j, Tf, T2I, T2J;
cannam@95 157 Ti = Cr[WS(csr, 4)];
cannam@95 158 T2j = Td - Te;
cannam@95 159 Tf = Td + Te;
cannam@95 160 {
cannam@95 161 E Tac, T2h, T7k, T2n;
cannam@95 162 Tac = T2f - T2g;
cannam@95 163 T2h = T2f + T2g;
cannam@95 164 T7k = T2m - T2j;
cannam@95 165 T2n = T2j + T2m;
cannam@95 166 {
cannam@95 167 E Taa, Tad, T7l, T2i;
cannam@95 168 Taa = Tc - Tf;
cannam@95 169 Tg = Tc + Tf;
cannam@95 170 Tad = Tab - Tac;
cannam@95 171 Tds = Tac + Tab;
cannam@95 172 T7l = T2e + T2h;
cannam@95 173 T2i = T2e - T2h;
cannam@95 174 Tcm = Taa + Tad;
cannam@95 175 Tae = Taa - Tad;
cannam@95 176 T92 = FMA(KP414213562, T7k, T7l);
cannam@95 177 T7m = FNMS(KP414213562, T7l, T7k);
cannam@95 178 T6c = FMA(KP414213562, T2i, T2n);
cannam@95 179 T2o = FNMS(KP414213562, T2n, T2i);
cannam@95 180 Tj = Cr[WS(csr, 60)];
cannam@95 181 }
cannam@95 182 }
cannam@95 183 T2I = Ci[WS(csi, 4)];
cannam@95 184 T2J = Ci[WS(csi, 60)];
cannam@95 185 Tl = Cr[WS(csr, 36)];
cannam@95 186 T2q = Ti - Tj;
cannam@95 187 Tk = Ti + Tj;
cannam@95 188 Tam = T2I - T2J;
cannam@95 189 T2K = T2I + T2J;
cannam@95 190 Tm = Cr[WS(csr, 28)];
cannam@95 191 }
cannam@95 192 T2r = Ci[WS(csi, 36)];
cannam@95 193 T2s = Ci[WS(csi, 28)];
cannam@95 194 {
cannam@95 195 E Ts, Tt, T2B, T2C;
cannam@95 196 Ts = Cr[WS(csr, 12)];
cannam@95 197 T2H = Tl - Tm;
cannam@95 198 Tn = Tl + Tm;
cannam@95 199 Tan = T2r - T2s;
cannam@95 200 T2t = T2r + T2s;
cannam@95 201 Tt = Cr[WS(csr, 52)];
cannam@95 202 T2B = Ci[WS(csi, 12)];
cannam@95 203 T2C = Ci[WS(csi, 52)];
cannam@95 204 {
cannam@95 205 E Tp, T2A, T2D, Tq, T2w, T2x;
cannam@95 206 Tp = Cr[WS(csr, 20)];
cannam@95 207 Tu = Ts + Tt;
cannam@95 208 T2A = Ts - Tt;
cannam@95 209 Tah = T2C - T2B;
cannam@95 210 T2D = T2B + T2C;
cannam@95 211 Tq = Cr[WS(csr, 44)];
cannam@95 212 T2w = Ci[WS(csi, 20)];
cannam@95 213 T2x = Ci[WS(csi, 44)];
cannam@95 214 T2E = T2A - T2D;
cannam@95 215 T2N = T2A + T2D;
cannam@95 216 Tr = Tp + Tq;
cannam@95 217 T2v = Tp - Tq;
cannam@95 218 T2y = T2w + T2x;
cannam@95 219 Tai = T2w - T2x;
cannam@95 220 }
cannam@95 221 }
cannam@95 222 }
cannam@95 223 {
cannam@95 224 E T2M, Tdv, Tdw, T2u, T2F, T7s, T7r, T2L, T2O;
cannam@95 225 {
cannam@95 226 E To, T2z, Tv, Tal, Tao;
cannam@95 227 Tag = Tk - Tn;
cannam@95 228 To = Tk + Tn;
cannam@95 229 T2M = T2v + T2y;
cannam@95 230 T2z = T2v - T2y;
cannam@95 231 Tv = Tr + Tu;
cannam@95 232 Tal = Tr - Tu;
cannam@95 233 Tao = Tam - Tan;
cannam@95 234 Tdv = Tan + Tam;
cannam@95 235 Tdu = To - Tv;
cannam@95 236 Tw = To + Tv;
cannam@95 237 Tco = Tao - Tal;
cannam@95 238 Tap = Tal + Tao;
cannam@95 239 Tdw = Tai + Tah;
cannam@95 240 Taj = Tah - Tai;
cannam@95 241 T7o = T2q + T2t;
cannam@95 242 T2u = T2q - T2t;
cannam@95 243 T2F = T2z + T2E;
cannam@95 244 T7s = T2E - T2z;
cannam@95 245 }
cannam@95 246 T7r = T2K - T2H;
cannam@95 247 T2L = T2H + T2K;
cannam@95 248 TeM = Tdw + Tdv;
cannam@95 249 Tdx = Tdv - Tdw;
cannam@95 250 T6f = FNMS(KP707106781, T2F, T2u);
cannam@95 251 T2G = FMA(KP707106781, T2F, T2u);
cannam@95 252 T2O = T2M - T2N;
cannam@95 253 T7p = T2M + T2N;
cannam@95 254 T6e = FNMS(KP707106781, T2O, T2L);
cannam@95 255 T2P = FMA(KP707106781, T2O, T2L);
cannam@95 256 T94 = FMA(KP707106781, T7s, T7r);
cannam@95 257 T7t = FNMS(KP707106781, T7s, T7r);
cannam@95 258 }
cannam@95 259 }
cannam@95 260 {
cannam@95 261 E T3M, T16, TbA, T5o, T5l, T19, TbB, T3P, T1g, TaV, T40, T5r, T1d, T3R, T3U;
cannam@95 262 E TaW;
cannam@95 263 {
cannam@95 264 E T17, T18, T3N, T3O;
cannam@95 265 {
cannam@95 266 E T14, T15, T5m, T5n;
cannam@95 267 T14 = Cr[WS(csr, 1)];
cannam@95 268 Tcp = Tag - Taj;
cannam@95 269 Tak = Tag + Taj;
cannam@95 270 T95 = FMA(KP707106781, T7p, T7o);
cannam@95 271 T7q = FNMS(KP707106781, T7p, T7o);
cannam@95 272 T15 = Cr[WS(csr, 63)];
cannam@95 273 T5m = Ci[WS(csi, 1)];
cannam@95 274 T5n = Ci[WS(csi, 63)];
cannam@95 275 T17 = Cr[WS(csr, 33)];
cannam@95 276 T3M = T14 - T15;
cannam@95 277 T16 = T14 + T15;
cannam@95 278 TbA = T5m - T5n;
cannam@95 279 T5o = T5m + T5n;
cannam@95 280 T18 = Cr[WS(csr, 31)];
cannam@95 281 }
cannam@95 282 T3N = Ci[WS(csi, 33)];
cannam@95 283 T3O = Ci[WS(csi, 31)];
cannam@95 284 {
cannam@95 285 E T1e, T1f, T3X, T3Y;
cannam@95 286 T1e = Cr[WS(csr, 15)];
cannam@95 287 T5l = T17 - T18;
cannam@95 288 T19 = T17 + T18;
cannam@95 289 TbB = T3N - T3O;
cannam@95 290 T3P = T3N + T3O;
cannam@95 291 T1f = Cr[WS(csr, 49)];
cannam@95 292 T3X = Ci[WS(csi, 15)];
cannam@95 293 T3Y = Ci[WS(csi, 49)];
cannam@95 294 {
cannam@95 295 E T1b, T3W, T3Z, T1c, T3S, T3T;
cannam@95 296 T1b = Cr[WS(csr, 17)];
cannam@95 297 T1g = T1e + T1f;
cannam@95 298 T3W = T1e - T1f;
cannam@95 299 TaV = T3Y - T3X;
cannam@95 300 T3Z = T3X + T3Y;
cannam@95 301 T1c = Cr[WS(csr, 47)];
cannam@95 302 T3S = Ci[WS(csi, 17)];
cannam@95 303 T3T = Ci[WS(csi, 47)];
cannam@95 304 T40 = T3W - T3Z;
cannam@95 305 T5r = T3W + T3Z;
cannam@95 306 T1d = T1b + T1c;
cannam@95 307 T3R = T1b - T1c;
cannam@95 308 T3U = T3S + T3T;
cannam@95 309 TaW = T3S - T3T;
cannam@95 310 }
cannam@95 311 }
cannam@95 312 }
cannam@95 313 {
cannam@95 314 E T5q, Te4, Te5, T3Q, T41, T8i, T8h, T5p, T5s;
cannam@95 315 {
cannam@95 316 E T1a, T3V, T1h, Tbz, TbC;
cannam@95 317 TaU = T16 - T19;
cannam@95 318 T1a = T16 + T19;
cannam@95 319 T5q = T3R + T3U;
cannam@95 320 T3V = T3R - T3U;
cannam@95 321 T1h = T1d + T1g;
cannam@95 322 Tbz = T1d - T1g;
cannam@95 323 TbC = TbA - TbB;
cannam@95 324 Te4 = TbB + TbA;
cannam@95 325 TdM = T1a - T1h;
cannam@95 326 T1i = T1a + T1h;
cannam@95 327 TcL = TbC - Tbz;
cannam@95 328 TbD = Tbz + TbC;
cannam@95 329 Te5 = TaW + TaV;
cannam@95 330 TaX = TaV - TaW;
cannam@95 331 T7Q = T3M + T3P;
cannam@95 332 T3Q = T3M - T3P;
cannam@95 333 T41 = T3V + T40;
cannam@95 334 T8i = T40 - T3V;
cannam@95 335 }
cannam@95 336 T8h = T5o - T5l;
cannam@95 337 T5p = T5l + T5o;
cannam@95 338 Tf0 = Te5 + Te4;
cannam@95 339 Te6 = Te4 - Te5;
cannam@95 340 T6q = FNMS(KP707106781, T41, T3Q);
cannam@95 341 T42 = FMA(KP707106781, T41, T3Q);
cannam@95 342 T5s = T5q - T5r;
cannam@95 343 T7R = T5q + T5r;
cannam@95 344 T6B = FNMS(KP707106781, T5s, T5p);
cannam@95 345 T5t = FMA(KP707106781, T5s, T5p);
cannam@95 346 T9r = FMA(KP707106781, T8i, T8h);
cannam@95 347 T8j = FNMS(KP707106781, T8i, T8h);
cannam@95 348 }
cannam@95 349 }
cannam@95 350 }
cannam@95 351 }
cannam@95 352 {
cannam@95 353 E Tas, Tav, T7w, T7x;
cannam@95 354 {
cannam@95 355 E T2S, TA, TaK, T3A, T3x, TD, TaL, T2V, TK, Tat, T36, T3D, TH, T2X, T30;
cannam@95 356 E Tau;
cannam@95 357 {
cannam@95 358 E TB, TC, T2T, T2U;
cannam@95 359 {
cannam@95 360 E Ty, Tz, T3y, T3z;
cannam@95 361 Ty = Cr[WS(csr, 2)];
cannam@95 362 TcA = TaU - TaX;
cannam@95 363 TaY = TaU + TaX;
cannam@95 364 T9g = FMA(KP707106781, T7R, T7Q);
cannam@95 365 T7S = FNMS(KP707106781, T7R, T7Q);
cannam@95 366 Tz = Cr[WS(csr, 62)];
cannam@95 367 T3y = Ci[WS(csi, 2)];
cannam@95 368 T3z = Ci[WS(csi, 62)];
cannam@95 369 TB = Cr[WS(csr, 34)];
cannam@95 370 T2S = Ty - Tz;
cannam@95 371 TA = Ty + Tz;
cannam@95 372 TaK = T3y - T3z;
cannam@95 373 T3A = T3y + T3z;
cannam@95 374 TC = Cr[WS(csr, 30)];
cannam@95 375 }
cannam@95 376 T2T = Ci[WS(csi, 34)];
cannam@95 377 T2U = Ci[WS(csi, 30)];
cannam@95 378 {
cannam@95 379 E TI, TJ, T33, T34;
cannam@95 380 TI = Cr[WS(csr, 14)];
cannam@95 381 T3x = TB - TC;
cannam@95 382 TD = TB + TC;
cannam@95 383 TaL = T2T - T2U;
cannam@95 384 T2V = T2T + T2U;
cannam@95 385 TJ = Cr[WS(csr, 50)];
cannam@95 386 T33 = Ci[WS(csi, 14)];
cannam@95 387 T34 = Ci[WS(csi, 50)];
cannam@95 388 {
cannam@95 389 E TF, T32, T35, TG, T2Y, T2Z;
cannam@95 390 TF = Cr[WS(csr, 18)];
cannam@95 391 TK = TI + TJ;
cannam@95 392 T32 = TI - TJ;
cannam@95 393 Tat = T34 - T33;
cannam@95 394 T35 = T33 + T34;
cannam@95 395 TG = Cr[WS(csr, 46)];
cannam@95 396 T2Y = Ci[WS(csi, 18)];
cannam@95 397 T2Z = Ci[WS(csi, 46)];
cannam@95 398 T36 = T32 - T35;
cannam@95 399 T3D = T32 + T35;
cannam@95 400 TH = TF + TG;
cannam@95 401 T2X = TF - TG;
cannam@95 402 T30 = T2Y + T2Z;
cannam@95 403 Tau = T2Y - T2Z;
cannam@95 404 }
cannam@95 405 }
cannam@95 406 }
cannam@95 407 {
cannam@95 408 E T3C, TdG, TdH, T2W, T37, T7I, T7H, T3B, T3E;
cannam@95 409 {
cannam@95 410 E TE, T31, TL, TaJ, TaM;
cannam@95 411 Tas = TA - TD;
cannam@95 412 TE = TA + TD;
cannam@95 413 T3C = T2X + T30;
cannam@95 414 T31 = T2X - T30;
cannam@95 415 TL = TH + TK;
cannam@95 416 TaJ = TH - TK;
cannam@95 417 TaM = TaK - TaL;
cannam@95 418 TdG = TaL + TaK;
cannam@95 419 TdA = TE - TL;
cannam@95 420 TM = TE + TL;
cannam@95 421 Tcv = TaM - TaJ;
cannam@95 422 TaN = TaJ + TaM;
cannam@95 423 TdH = Tau + Tat;
cannam@95 424 Tav = Tat - Tau;
cannam@95 425 T7w = T2S + T2V;
cannam@95 426 T2W = T2S - T2V;
cannam@95 427 T37 = T31 + T36;
cannam@95 428 T7I = T36 - T31;
cannam@95 429 }
cannam@95 430 T7H = T3A - T3x;
cannam@95 431 T3B = T3x + T3A;
cannam@95 432 TeP = TdH + TdG;
cannam@95 433 TdI = TdG - TdH;
cannam@95 434 T6i = FNMS(KP707106781, T37, T2W);
cannam@95 435 T38 = FMA(KP707106781, T37, T2W);
cannam@95 436 T3E = T3C - T3D;
cannam@95 437 T7x = T3C + T3D;
cannam@95 438 T6l = FNMS(KP707106781, T3E, T3B);
cannam@95 439 T3F = FMA(KP707106781, T3E, T3B);
cannam@95 440 T9b = FMA(KP707106781, T7I, T7H);
cannam@95 441 T7J = FNMS(KP707106781, T7I, T7H);
cannam@95 442 }
cannam@95 443 }
cannam@95 444 {
cannam@95 445 E T4r, T4I, T1F, Tbb, T4u, T4L, Tbj, TdS, T1I, Tbd, T4N, T4A, T4B, T1L, Tbc;
cannam@95 446 E T4E, T1M, Tbg;
cannam@95 447 {
cannam@95 448 E T1z, T1A, T1C, T1D, Tbi, Tbh;
cannam@95 449 T1z = Cr[WS(csr, 5)];
cannam@95 450 Tcs = Tas - Tav;
cannam@95 451 Taw = Tas + Tav;
cannam@95 452 T98 = FMA(KP707106781, T7x, T7w);
cannam@95 453 T7y = FNMS(KP707106781, T7x, T7w);
cannam@95 454 T1A = Cr[WS(csr, 59)];
cannam@95 455 T1C = Cr[WS(csr, 37)];
cannam@95 456 T1D = Cr[WS(csr, 27)];
cannam@95 457 {
cannam@95 458 E T4s, T1B, T1E, T4t, T4J, T4K;
cannam@95 459 T4s = Ci[WS(csi, 37)];
cannam@95 460 T4r = T1z - T1A;
cannam@95 461 T1B = T1z + T1A;
cannam@95 462 T4I = T1C - T1D;
cannam@95 463 T1E = T1C + T1D;
cannam@95 464 T4t = Ci[WS(csi, 27)];
cannam@95 465 T4J = Ci[WS(csi, 5)];
cannam@95 466 T4K = Ci[WS(csi, 59)];
cannam@95 467 T1F = T1B + T1E;
cannam@95 468 Tbb = T1B - T1E;
cannam@95 469 T4u = T4s + T4t;
cannam@95 470 Tbi = T4s - T4t;
cannam@95 471 Tbh = T4J - T4K;
cannam@95 472 T4L = T4J + T4K;
cannam@95 473 }
cannam@95 474 {
cannam@95 475 E T1J, T4w, T4z, T1K, T4C, T4D;
cannam@95 476 {
cannam@95 477 E T1G, T1H, T4x, T4y;
cannam@95 478 T1G = Cr[WS(csr, 21)];
cannam@95 479 Tbj = Tbh - Tbi;
cannam@95 480 TdS = Tbi + Tbh;
cannam@95 481 T1H = Cr[WS(csr, 43)];
cannam@95 482 T4x = Ci[WS(csi, 21)];
cannam@95 483 T4y = Ci[WS(csi, 43)];
cannam@95 484 T1J = Cr[WS(csr, 11)];
cannam@95 485 T4w = T1G - T1H;
cannam@95 486 T1I = T1G + T1H;
cannam@95 487 Tbd = T4x - T4y;
cannam@95 488 T4z = T4x + T4y;
cannam@95 489 T1K = Cr[WS(csr, 53)];
cannam@95 490 T4C = Ci[WS(csi, 11)];
cannam@95 491 T4D = Ci[WS(csi, 53)];
cannam@95 492 }
cannam@95 493 T4N = T4w + T4z;
cannam@95 494 T4A = T4w - T4z;
cannam@95 495 T4B = T1J - T1K;
cannam@95 496 T1L = T1J + T1K;
cannam@95 497 Tbc = T4D - T4C;
cannam@95 498 T4E = T4C + T4D;
cannam@95 499 }
cannam@95 500 }
cannam@95 501 T1M = T1I + T1L;
cannam@95 502 Tbg = T1I - T1L;
cannam@95 503 {
cannam@95 504 E TdT, Tbe, T4F, T4O;
cannam@95 505 TdT = Tbd + Tbc;
cannam@95 506 Tbe = Tbc - Tbd;
cannam@95 507 T4F = T4B - T4E;
cannam@95 508 T4O = T4B + T4E;
cannam@95 509 {
cannam@95 510 E TdR, TdU, T81, T4v, T4G, T85;
cannam@95 511 TdR = T1F - T1M;
cannam@95 512 T1N = T1F + T1M;
cannam@95 513 TeW = TdT + TdS;
cannam@95 514 TdU = TdS - TdT;
cannam@95 515 T81 = T4r + T4u;
cannam@95 516 T4v = T4r - T4u;
cannam@95 517 T4G = T4A + T4F;
cannam@95 518 T85 = T4F - T4A;
cannam@95 519 {
cannam@95 520 E T84, T4M, T4P, T82, TcG, TcH;
cannam@95 521 T84 = T4L - T4I;
cannam@95 522 T4M = T4I + T4L;
cannam@95 523 T6x = FNMS(KP707106781, T4G, T4v);
cannam@95 524 T4H = FMA(KP707106781, T4G, T4v);
cannam@95 525 Te8 = TdR + TdU;
cannam@95 526 TdV = TdR - TdU;
cannam@95 527 T4P = T4N - T4O;
cannam@95 528 T82 = T4N + T4O;
cannam@95 529 Tbk = Tbg + Tbj;
cannam@95 530 TcG = Tbj - Tbg;
cannam@95 531 T6w = FNMS(KP707106781, T4P, T4M);
cannam@95 532 T4Q = FMA(KP707106781, T4P, T4M);
cannam@95 533 T9j = FMA(KP707106781, T85, T84);
cannam@95 534 T86 = FNMS(KP707106781, T85, T84);
cannam@95 535 TcH = Tbb - Tbe;
cannam@95 536 Tbf = Tbb + Tbe;
cannam@95 537 TcO = FMA(KP414213562, TcG, TcH);
cannam@95 538 TcI = FNMS(KP414213562, TcH, TcG);
cannam@95 539 T9k = FMA(KP707106781, T82, T81);
cannam@95 540 T83 = FNMS(KP707106781, T82, T81);
cannam@95 541 }
cannam@95 542 }
cannam@95 543 }
cannam@95 544 }
cannam@95 545 }
cannam@95 546 {
cannam@95 547 E T88, T89, Tbv, Tbq;
cannam@95 548 {
cannam@95 549 E T4S, T59, T4V, Tbm, T1U, T5c, TdX, Tbu, T1X, T53, Tbo, T52, T20, T54, T5e;
cannam@95 550 E T51;
cannam@95 551 {
cannam@95 552 E T1R, T1Q, T1S, T1O, T1P;
cannam@95 553 T1O = Cr[WS(csr, 3)];
cannam@95 554 T1P = Cr[WS(csr, 61)];
cannam@95 555 T1R = Cr[WS(csr, 29)];
cannam@95 556 TbI = FMA(KP414213562, Tbf, Tbk);
cannam@95 557 Tbl = FNMS(KP414213562, Tbk, Tbf);
cannam@95 558 T1Q = T1O + T1P;
cannam@95 559 T4S = T1O - T1P;
cannam@95 560 T1S = Cr[WS(csr, 35)];
cannam@95 561 {
cannam@95 562 E Tbt, Tbs, T4X, T50;
cannam@95 563 {
cannam@95 564 E T5a, T5b, T4T, T4U, T1T;
cannam@95 565 T4T = Ci[WS(csi, 29)];
cannam@95 566 T4U = Ci[WS(csi, 35)];
cannam@95 567 T1T = T1R + T1S;
cannam@95 568 T59 = T1R - T1S;
cannam@95 569 T5a = Ci[WS(csi, 3)];
cannam@95 570 Tbt = T4T - T4U;
cannam@95 571 T4V = T4T + T4U;
cannam@95 572 T5b = Ci[WS(csi, 61)];
cannam@95 573 Tbm = T1Q - T1T;
cannam@95 574 T1U = T1Q + T1T;
cannam@95 575 T5c = T5a + T5b;
cannam@95 576 Tbs = T5b - T5a;
cannam@95 577 }
cannam@95 578 {
cannam@95 579 E T4Y, T4Z, T1V, T1W, T1Y, T1Z;
cannam@95 580 T1V = Cr[WS(csr, 13)];
cannam@95 581 T1W = Cr[WS(csr, 51)];
cannam@95 582 TdX = Tbt + Tbs;
cannam@95 583 Tbu = Tbs - Tbt;
cannam@95 584 T4Y = Ci[WS(csi, 13)];
cannam@95 585 T4X = T1V - T1W;
cannam@95 586 T1X = T1V + T1W;
cannam@95 587 T4Z = Ci[WS(csi, 51)];
cannam@95 588 T1Y = Cr[WS(csr, 19)];
cannam@95 589 T1Z = Cr[WS(csr, 45)];
cannam@95 590 T53 = Ci[WS(csi, 19)];
cannam@95 591 Tbo = T4Y - T4Z;
cannam@95 592 T50 = T4Y + T4Z;
cannam@95 593 T52 = T1Y - T1Z;
cannam@95 594 T20 = T1Y + T1Z;
cannam@95 595 T54 = Ci[WS(csi, 45)];
cannam@95 596 }
cannam@95 597 T5e = T4X + T50;
cannam@95 598 T51 = T4X - T50;
cannam@95 599 }
cannam@95 600 }
cannam@95 601 {
cannam@95 602 E T21, Tbr, T55, Tbn;
cannam@95 603 T21 = T1X + T20;
cannam@95 604 Tbr = T1X - T20;
cannam@95 605 T55 = T53 + T54;
cannam@95 606 Tbn = T54 - T53;
cannam@95 607 {
cannam@95 608 E T4W, TdW, Tbp, T5f, TdZ, T57, T8c, TdY, T56;
cannam@95 609 T88 = T4S + T4V;
cannam@95 610 T4W = T4S - T4V;
cannam@95 611 T22 = T1U + T21;
cannam@95 612 TdW = T1U - T21;
cannam@95 613 TdY = Tbo + Tbn;
cannam@95 614 Tbp = Tbn - Tbo;
cannam@95 615 T56 = T52 - T55;
cannam@95 616 T5f = T52 + T55;
cannam@95 617 TeV = TdY + TdX;
cannam@95 618 TdZ = TdX - TdY;
cannam@95 619 T57 = T51 + T56;
cannam@95 620 T8c = T56 - T51;
cannam@95 621 {
cannam@95 622 E T8b, T5d, T5g, TcD, TcE;
cannam@95 623 T8b = T59 + T5c;
cannam@95 624 T5d = T59 - T5c;
cannam@95 625 T5g = T5e - T5f;
cannam@95 626 T89 = T5e + T5f;
cannam@95 627 Te0 = TdW + TdZ;
cannam@95 628 Te9 = TdZ - TdW;
cannam@95 629 T58 = FMA(KP707106781, T57, T4W);
cannam@95 630 T6u = FNMS(KP707106781, T57, T4W);
cannam@95 631 T6t = FNMS(KP707106781, T5g, T5d);
cannam@95 632 T5h = FMA(KP707106781, T5g, T5d);
cannam@95 633 Tbv = Tbr + Tbu;
cannam@95 634 TcD = Tbu - Tbr;
cannam@95 635 TcE = Tbm - Tbp;
cannam@95 636 Tbq = Tbm + Tbp;
cannam@95 637 T9m = FNMS(KP707106781, T8c, T8b);
cannam@95 638 T8d = FMA(KP707106781, T8c, T8b);
cannam@95 639 TcP = FNMS(KP414213562, TcD, TcE);
cannam@95 640 TcF = FMA(KP414213562, TcE, TcD);
cannam@95 641 }
cannam@95 642 }
cannam@95 643 }
cannam@95 644 }
cannam@95 645 {
cannam@95 646 E Tb3, Tb8, T7V, T7Y;
cannam@95 647 {
cannam@95 648 E T7T, T4c, TaZ, T1p, TdO, Tb2, T7U, T47, T1t, T4e, T1s, Tb5, T4m, T1u, T4f;
cannam@95 649 E T4g;
cannam@95 650 {
cannam@95 651 E T1m, T43, T1l, Tb0, T4b, T1n, T44, T45;
cannam@95 652 {
cannam@95 653 E T1j, T1k, T49, T4a;
cannam@95 654 T1j = Cr[WS(csr, 9)];
cannam@95 655 T9n = FMA(KP707106781, T89, T88);
cannam@95 656 T8a = FNMS(KP707106781, T89, T88);
cannam@95 657 TbJ = FNMS(KP414213562, Tbq, Tbv);
cannam@95 658 Tbw = FMA(KP414213562, Tbv, Tbq);
cannam@95 659 T1k = Cr[WS(csr, 55)];
cannam@95 660 T49 = Ci[WS(csi, 9)];
cannam@95 661 T4a = Ci[WS(csi, 55)];
cannam@95 662 T1m = Cr[WS(csr, 41)];
cannam@95 663 T43 = T1j - T1k;
cannam@95 664 T1l = T1j + T1k;
cannam@95 665 Tb0 = T49 - T4a;
cannam@95 666 T4b = T49 + T4a;
cannam@95 667 T1n = Cr[WS(csr, 23)];
cannam@95 668 T44 = Ci[WS(csi, 41)];
cannam@95 669 T45 = Ci[WS(csi, 23)];
cannam@95 670 }
cannam@95 671 {
cannam@95 672 E T1q, T1r, T4k, T4l;
cannam@95 673 T1q = Cr[WS(csr, 7)];
cannam@95 674 {
cannam@95 675 E T48, T1o, Tb1, T46;
cannam@95 676 T48 = T1m - T1n;
cannam@95 677 T1o = T1m + T1n;
cannam@95 678 Tb1 = T44 - T45;
cannam@95 679 T46 = T44 + T45;
cannam@95 680 T7T = T4b - T48;
cannam@95 681 T4c = T48 + T4b;
cannam@95 682 TaZ = T1l - T1o;
cannam@95 683 T1p = T1l + T1o;
cannam@95 684 TdO = Tb1 + Tb0;
cannam@95 685 Tb2 = Tb0 - Tb1;
cannam@95 686 T7U = T43 + T46;
cannam@95 687 T47 = T43 - T46;
cannam@95 688 T1r = Cr[WS(csr, 57)];
cannam@95 689 }
cannam@95 690 T4k = Ci[WS(csi, 7)];
cannam@95 691 T4l = Ci[WS(csi, 57)];
cannam@95 692 T1t = Cr[WS(csr, 25)];
cannam@95 693 T4e = T1q - T1r;
cannam@95 694 T1s = T1q + T1r;
cannam@95 695 Tb5 = T4l - T4k;
cannam@95 696 T4m = T4k + T4l;
cannam@95 697 T1u = Cr[WS(csr, 39)];
cannam@95 698 T4f = Ci[WS(csi, 25)];
cannam@95 699 T4g = Ci[WS(csi, 39)];
cannam@95 700 }
cannam@95 701 }
cannam@95 702 {
cannam@95 703 E T7W, TdN, T7X, T5u, T4d, T4o, T5v, T8k, T8l;
cannam@95 704 {
cannam@95 705 E T4n, T1w, T4i, TbE, TbF, Tb4, Tb7;
cannam@95 706 {
cannam@95 707 E T4j, T1v, Tb6, T4h;
cannam@95 708 T4j = T1t - T1u;
cannam@95 709 T1v = T1t + T1u;
cannam@95 710 Tb6 = T4f - T4g;
cannam@95 711 T4h = T4f + T4g;
cannam@95 712 T7W = T4j + T4m;
cannam@95 713 T4n = T4j - T4m;
cannam@95 714 Tb4 = T1s - T1v;
cannam@95 715 T1w = T1s + T1v;
cannam@95 716 TdN = Tb6 + Tb5;
cannam@95 717 Tb7 = Tb5 - Tb6;
cannam@95 718 T7X = T4e + T4h;
cannam@95 719 T4i = T4e - T4h;
cannam@95 720 }
cannam@95 721 Tb3 = TaZ - Tb2;
cannam@95 722 TbE = TaZ + Tb2;
cannam@95 723 TbF = Tb7 - Tb4;
cannam@95 724 Tb8 = Tb4 + Tb7;
cannam@95 725 Te3 = T1p - T1w;
cannam@95 726 T1x = T1p + T1w;
cannam@95 727 TcB = TbE - TbF;
cannam@95 728 TbG = TbE + TbF;
cannam@95 729 T5u = FMA(KP414213562, T47, T4c);
cannam@95 730 T4d = FNMS(KP414213562, T4c, T47);
cannam@95 731 T4o = FMA(KP414213562, T4n, T4i);
cannam@95 732 T5v = FNMS(KP414213562, T4i, T4n);
cannam@95 733 }
cannam@95 734 Tf1 = TdO + TdN;
cannam@95 735 TdP = TdN - TdO;
cannam@95 736 T6C = T4o - T4d;
cannam@95 737 T4p = T4d + T4o;
cannam@95 738 T7V = FNMS(KP414213562, T7U, T7T);
cannam@95 739 T8k = FMA(KP414213562, T7T, T7U);
cannam@95 740 T8l = FMA(KP414213562, T7W, T7X);
cannam@95 741 T7Y = FNMS(KP414213562, T7X, T7W);
cannam@95 742 T6r = T5u - T5v;
cannam@95 743 T5w = T5u + T5v;
cannam@95 744 T9h = T8k + T8l;
cannam@95 745 T8m = T8k - T8l;
cannam@95 746 }
cannam@95 747 }
cannam@95 748 {
cannam@95 749 E T7z, T3i, Tax, TT, TdC, TaA, T7A, T3d, TX, T3k, TW, TaD, T3s, TY, T3l;
cannam@95 750 E T3m;
cannam@95 751 {
cannam@95 752 E TQ, T39, TP, Tay, T3h, TR, T3a, T3b;
cannam@95 753 {
cannam@95 754 E TN, TO, T3f, T3g;
cannam@95 755 TN = Cr[WS(csr, 10)];
cannam@95 756 TcM = Tb8 - Tb3;
cannam@95 757 Tb9 = Tb3 + Tb8;
cannam@95 758 T9s = T7V - T7Y;
cannam@95 759 T7Z = T7V + T7Y;
cannam@95 760 TO = Cr[WS(csr, 54)];
cannam@95 761 T3f = Ci[WS(csi, 10)];
cannam@95 762 T3g = Ci[WS(csi, 54)];
cannam@95 763 TQ = Cr[WS(csr, 42)];
cannam@95 764 T39 = TN - TO;
cannam@95 765 TP = TN + TO;
cannam@95 766 Tay = T3f - T3g;
cannam@95 767 T3h = T3f + T3g;
cannam@95 768 TR = Cr[WS(csr, 22)];
cannam@95 769 T3a = Ci[WS(csi, 42)];
cannam@95 770 T3b = Ci[WS(csi, 22)];
cannam@95 771 }
cannam@95 772 {
cannam@95 773 E TU, TV, T3q, T3r;
cannam@95 774 TU = Cr[WS(csr, 6)];
cannam@95 775 {
cannam@95 776 E T3e, TS, Taz, T3c;
cannam@95 777 T3e = TQ - TR;
cannam@95 778 TS = TQ + TR;
cannam@95 779 Taz = T3a - T3b;
cannam@95 780 T3c = T3a + T3b;
cannam@95 781 T7z = T3h - T3e;
cannam@95 782 T3i = T3e + T3h;
cannam@95 783 Tax = TP - TS;
cannam@95 784 TT = TP + TS;
cannam@95 785 TdC = Taz + Tay;
cannam@95 786 TaA = Tay - Taz;
cannam@95 787 T7A = T39 + T3c;
cannam@95 788 T3d = T39 - T3c;
cannam@95 789 TV = Cr[WS(csr, 58)];
cannam@95 790 }
cannam@95 791 T3q = Ci[WS(csi, 6)];
cannam@95 792 T3r = Ci[WS(csi, 58)];
cannam@95 793 TX = Cr[WS(csr, 26)];
cannam@95 794 T3k = TU - TV;
cannam@95 795 TW = TU + TV;
cannam@95 796 TaD = T3r - T3q;
cannam@95 797 T3s = T3q + T3r;
cannam@95 798 TY = Cr[WS(csr, 38)];
cannam@95 799 T3l = Ci[WS(csi, 26)];
cannam@95 800 T3m = Ci[WS(csi, 38)];
cannam@95 801 }
cannam@95 802 }
cannam@95 803 {
cannam@95 804 E T7C, TdB, T7D, T3G, T3j, T3u, T3H, T7K, T7L;
cannam@95 805 {
cannam@95 806 E T3t, T10, T3o, TaO, TaP, TaC, TaF;
cannam@95 807 {
cannam@95 808 E T3p, TZ, TaE, T3n;
cannam@95 809 T3p = TX - TY;
cannam@95 810 TZ = TX + TY;
cannam@95 811 TaE = T3l - T3m;
cannam@95 812 T3n = T3l + T3m;
cannam@95 813 T7C = T3p + T3s;
cannam@95 814 T3t = T3p - T3s;
cannam@95 815 TaC = TW - TZ;
cannam@95 816 T10 = TW + TZ;
cannam@95 817 TdB = TaE + TaD;
cannam@95 818 TaF = TaD - TaE;
cannam@95 819 T7D = T3k + T3n;
cannam@95 820 T3o = T3k - T3n;
cannam@95 821 }
cannam@95 822 TaB = Tax - TaA;
cannam@95 823 TaO = Tax + TaA;
cannam@95 824 TaP = TaF - TaC;
cannam@95 825 TaG = TaC + TaF;
cannam@95 826 TdF = TT - T10;
cannam@95 827 T11 = TT + T10;
cannam@95 828 Tct = TaO - TaP;
cannam@95 829 TaQ = TaO + TaP;
cannam@95 830 T3G = FMA(KP414213562, T3d, T3i);
cannam@95 831 T3j = FNMS(KP414213562, T3i, T3d);
cannam@95 832 T3u = FMA(KP414213562, T3t, T3o);
cannam@95 833 T3H = FNMS(KP414213562, T3o, T3t);
cannam@95 834 }
cannam@95 835 TeQ = TdC + TdB;
cannam@95 836 TdD = TdB - TdC;
cannam@95 837 T6m = T3u - T3j;
cannam@95 838 T3v = T3j + T3u;
cannam@95 839 T7B = FNMS(KP414213562, T7A, T7z);
cannam@95 840 T7K = FMA(KP414213562, T7z, T7A);
cannam@95 841 T7L = FMA(KP414213562, T7C, T7D);
cannam@95 842 T7E = FNMS(KP414213562, T7D, T7C);
cannam@95 843 T6j = T3G - T3H;
cannam@95 844 T3I = T3G + T3H;
cannam@95 845 T99 = T7K + T7L;
cannam@95 846 T7M = T7K - T7L;
cannam@95 847 }
cannam@95 848 }
cannam@95 849 }
cannam@95 850 }
cannam@95 851 }
cannam@95 852 {
cannam@95 853 E Tcw, T9c, T7F, Tev, Teu, TeD, Tep, TeG, Tez, TeE, Tes;
cannam@95 854 {
cannam@95 855 E TbX, TbY, Tc7, TbP, Tar, Tc5, Tc1, Tc0, Tc4, Tba, TbS, TbL, TbQ, TaS, Tbx;
cannam@95 856 E Tc8;
cannam@95 857 {
cannam@95 858 E TeO, TaH, TeR, TeL, TeU, TeZ, Tf2, TeX, Tfh, Tfn, Tfo, Tfm;
cannam@95 859 {
cannam@95 860 E T12, Tfg, Tfj, Tx, Tff, T24, Tfi, Tfk, Th, T1y, T23;
cannam@95 861 TeO = TM - T11;
cannam@95 862 T12 = TM + T11;
cannam@95 863 Tcw = TaG - TaB;
cannam@95 864 TaH = TaB + TaG;
cannam@95 865 T9c = T7B - T7E;
cannam@95 866 T7F = T7B + T7E;
cannam@95 867 Tfg = TeQ + TeP;
cannam@95 868 TeR = TeP - TeQ;
cannam@95 869 TeL = FNMS(KP2_000000000, Tg, T9);
cannam@95 870 Th = FMA(KP2_000000000, Tg, T9);
cannam@95 871 T1y = T1i + T1x;
cannam@95 872 TeU = T1i - T1x;
cannam@95 873 TeZ = T1N - T22;
cannam@95 874 T23 = T1N + T22;
cannam@95 875 Tfj = Tf1 + Tf0;
cannam@95 876 Tf2 = Tf0 - Tf1;
cannam@95 877 Tx = FMA(KP2_000000000, Tw, Th);
cannam@95 878 Tff = FNMS(KP2_000000000, Tw, Th);
cannam@95 879 T24 = T1y + T23;
cannam@95 880 Tfi = T1y - T23;
cannam@95 881 TeX = TeV - TeW;
cannam@95 882 Tfk = TeW + TeV;
cannam@95 883 {
cannam@95 884 E T13, Tfp, Tfl, Tfq;
cannam@95 885 T13 = FMA(KP2_000000000, T12, Tx);
cannam@95 886 Tfp = FNMS(KP2_000000000, T12, Tx);
cannam@95 887 Tfh = FNMS(KP2_000000000, Tfg, Tff);
cannam@95 888 Tfn = FMA(KP2_000000000, Tfg, Tff);
cannam@95 889 Tfl = Tfj - Tfk;
cannam@95 890 Tfq = Tfk + Tfj;
cannam@95 891 R0[0] = FMA(KP2_000000000, T24, T13);
cannam@95 892 R0[WS(rs, 32)] = FNMS(KP2_000000000, T24, T13);
cannam@95 893 R0[WS(rs, 48)] = FMA(KP2_000000000, Tfq, Tfp);
cannam@95 894 R0[WS(rs, 16)] = FNMS(KP2_000000000, Tfq, Tfp);
cannam@95 895 Tfo = Tfi + Tfl;
cannam@95 896 Tfm = Tfi - Tfl;
cannam@95 897 }
cannam@95 898 }
cannam@95 899 {
cannam@95 900 E Tf7, TeN, Tfa, Tf3, Tf8, TeS;
cannam@95 901 R0[WS(rs, 8)] = FMA(KP1_414213562, Tfm, Tfh);
cannam@95 902 R0[WS(rs, 40)] = FNMS(KP1_414213562, Tfm, Tfh);
cannam@95 903 R0[WS(rs, 56)] = FMA(KP1_414213562, Tfo, Tfn);
cannam@95 904 R0[WS(rs, 24)] = FNMS(KP1_414213562, Tfo, Tfn);
cannam@95 905 Tf7 = FMA(KP2_000000000, TeM, TeL);
cannam@95 906 TeN = FNMS(KP2_000000000, TeM, TeL);
cannam@95 907 Tfa = Tf2 - TeZ;
cannam@95 908 Tf3 = TeZ + Tf2;
cannam@95 909 Tf8 = TeO + TeR;
cannam@95 910 TeS = TeO - TeR;
cannam@95 911 {
cannam@95 912 E TbH, TbK, TaI, TaR;
cannam@95 913 {
cannam@95 914 E Taf, Tf9, Tfd, Tf5, TeT, Tfb, TeY, Taq;
cannam@95 915 TbX = FNMS(KP1_414213562, Tae, Ta9);
cannam@95 916 Taf = FMA(KP1_414213562, Tae, Ta9);
cannam@95 917 Tf9 = FNMS(KP1_414213562, Tf8, Tf7);
cannam@95 918 Tfd = FMA(KP1_414213562, Tf8, Tf7);
cannam@95 919 Tf5 = FNMS(KP1_414213562, TeS, TeN);
cannam@95 920 TeT = FMA(KP1_414213562, TeS, TeN);
cannam@95 921 Tfb = TeU - TeX;
cannam@95 922 TeY = TeU + TeX;
cannam@95 923 Taq = FNMS(KP414213562, Tap, Tak);
cannam@95 924 TbY = FMA(KP414213562, Tak, Tap);
cannam@95 925 Tc7 = FNMS(KP707106781, TbG, TbD);
cannam@95 926 TbH = FMA(KP707106781, TbG, TbD);
cannam@95 927 {
cannam@95 928 E Tfc, Tfe, Tf6, Tf4;
cannam@95 929 Tfc = FNMS(KP414213562, Tfb, Tfa);
cannam@95 930 Tfe = FMA(KP414213562, Tfa, Tfb);
cannam@95 931 Tf6 = FMA(KP414213562, TeY, Tf3);
cannam@95 932 Tf4 = FNMS(KP414213562, Tf3, TeY);
cannam@95 933 TbP = FNMS(KP1_847759065, Taq, Taf);
cannam@95 934 Tar = FMA(KP1_847759065, Taq, Taf);
cannam@95 935 R0[WS(rs, 44)] = FMA(KP1_847759065, Tfc, Tf9);
cannam@95 936 R0[WS(rs, 12)] = FNMS(KP1_847759065, Tfc, Tf9);
cannam@95 937 R0[WS(rs, 60)] = FMA(KP1_847759065, Tfe, Tfd);
cannam@95 938 R0[WS(rs, 28)] = FNMS(KP1_847759065, Tfe, Tfd);
cannam@95 939 R0[WS(rs, 52)] = FMA(KP1_847759065, Tf6, Tf5);
cannam@95 940 R0[WS(rs, 20)] = FNMS(KP1_847759065, Tf6, Tf5);
cannam@95 941 R0[WS(rs, 4)] = FMA(KP1_847759065, Tf4, TeT);
cannam@95 942 R0[WS(rs, 36)] = FNMS(KP1_847759065, Tf4, TeT);
cannam@95 943 TbK = TbI + TbJ;
cannam@95 944 Tc5 = TbI - TbJ;
cannam@95 945 }
cannam@95 946 }
cannam@95 947 Tc1 = FNMS(KP707106781, TaH, Taw);
cannam@95 948 TaI = FMA(KP707106781, TaH, Taw);
cannam@95 949 TaR = FMA(KP707106781, TaQ, TaN);
cannam@95 950 Tc0 = FNMS(KP707106781, TaQ, TaN);
cannam@95 951 Tc4 = FNMS(KP707106781, Tb9, TaY);
cannam@95 952 Tba = FMA(KP707106781, Tb9, TaY);
cannam@95 953 TbS = FNMS(KP923879532, TbK, TbH);
cannam@95 954 TbL = FMA(KP923879532, TbK, TbH);
cannam@95 955 TbQ = FMA(KP198912367, TaI, TaR);
cannam@95 956 TaS = FNMS(KP198912367, TaR, TaI);
cannam@95 957 Tbx = Tbl + Tbw;
cannam@95 958 Tc8 = Tbw - Tbl;
cannam@95 959 }
cannam@95 960 }
cannam@95 961 }
cannam@95 962 {
cannam@95 963 E Ten, Teo, Tex, Tef, Tdz, Ter, Teq, TdQ, Tei, Teb, Teg, TdK, Te1, Tey;
cannam@95 964 {
cannam@95 965 E Te7, Tea, TdE, TdJ;
cannam@95 966 {
cannam@95 967 E Tdt, TbR, TbV, TbN, TaT, TbT, Tby, Tdy;
cannam@95 968 Ten = FMA(KP2_000000000, Tds, Tdr);
cannam@95 969 Tdt = FNMS(KP2_000000000, Tds, Tdr);
cannam@95 970 TbR = FNMS(KP1_961570560, TbQ, TbP);
cannam@95 971 TbV = FMA(KP1_961570560, TbQ, TbP);
cannam@95 972 TbN = FNMS(KP1_961570560, TaS, Tar);
cannam@95 973 TaT = FMA(KP1_961570560, TaS, Tar);
cannam@95 974 TbT = FNMS(KP923879532, Tbx, Tba);
cannam@95 975 Tby = FMA(KP923879532, Tbx, Tba);
cannam@95 976 Tdy = Tdu - Tdx;
cannam@95 977 Teo = Tdu + Tdx;
cannam@95 978 Tex = Te6 - Te3;
cannam@95 979 Te7 = Te3 + Te6;
cannam@95 980 {
cannam@95 981 E TbU, TbW, TbO, TbM;
cannam@95 982 TbU = FNMS(KP820678790, TbT, TbS);
cannam@95 983 TbW = FMA(KP820678790, TbS, TbT);
cannam@95 984 TbO = FMA(KP098491403, Tby, TbL);
cannam@95 985 TbM = FNMS(KP098491403, TbL, Tby);
cannam@95 986 Tef = FNMS(KP1_414213562, Tdy, Tdt);
cannam@95 987 Tdz = FMA(KP1_414213562, Tdy, Tdt);
cannam@95 988 R0[WS(rs, 41)] = FMA(KP1_546020906, TbU, TbR);
cannam@95 989 R0[WS(rs, 9)] = FNMS(KP1_546020906, TbU, TbR);
cannam@95 990 R0[WS(rs, 57)] = FMA(KP1_546020906, TbW, TbV);
cannam@95 991 R0[WS(rs, 25)] = FNMS(KP1_546020906, TbW, TbV);
cannam@95 992 R0[WS(rs, 49)] = FMA(KP1_990369453, TbO, TbN);
cannam@95 993 R0[WS(rs, 17)] = FNMS(KP1_990369453, TbO, TbN);
cannam@95 994 R0[WS(rs, 1)] = FMA(KP1_990369453, TbM, TaT);
cannam@95 995 R0[WS(rs, 33)] = FNMS(KP1_990369453, TbM, TaT);
cannam@95 996 Tea = Te8 + Te9;
cannam@95 997 Tev = Te8 - Te9;
cannam@95 998 }
cannam@95 999 }
cannam@95 1000 Ter = TdA - TdD;
cannam@95 1001 TdE = TdA + TdD;
cannam@95 1002 TdJ = TdF + TdI;
cannam@95 1003 Teq = TdI - TdF;
cannam@95 1004 Teu = TdM - TdP;
cannam@95 1005 TdQ = TdM + TdP;
cannam@95 1006 Tei = FNMS(KP707106781, Tea, Te7);
cannam@95 1007 Teb = FMA(KP707106781, Tea, Te7);
cannam@95 1008 Teg = FMA(KP414213562, TdE, TdJ);
cannam@95 1009 TdK = FNMS(KP414213562, TdJ, TdE);
cannam@95 1010 Te1 = TdV + Te0;
cannam@95 1011 Tey = Te0 - TdV;
cannam@95 1012 }
cannam@95 1013 {
cannam@95 1014 E Tcd, TbZ, Tcg, Tc9, Tce, Tc2;
cannam@95 1015 {
cannam@95 1016 E Teh, Tel, Ted, TdL, Tej, Te2;
cannam@95 1017 Teh = FNMS(KP1_847759065, Teg, Tef);
cannam@95 1018 Tel = FMA(KP1_847759065, Teg, Tef);
cannam@95 1019 Ted = FNMS(KP1_847759065, TdK, Tdz);
cannam@95 1020 TdL = FMA(KP1_847759065, TdK, Tdz);
cannam@95 1021 Tej = FNMS(KP707106781, Te1, TdQ);
cannam@95 1022 Te2 = FMA(KP707106781, Te1, TdQ);
cannam@95 1023 {
cannam@95 1024 E Tek, Tem, Tee, Tec;
cannam@95 1025 Tek = FNMS(KP668178637, Tej, Tei);
cannam@95 1026 Tem = FMA(KP668178637, Tei, Tej);
cannam@95 1027 Tee = FMA(KP198912367, Te2, Teb);
cannam@95 1028 Tec = FNMS(KP198912367, Teb, Te2);
cannam@95 1029 Tcd = FMA(KP1_847759065, TbY, TbX);
cannam@95 1030 TbZ = FNMS(KP1_847759065, TbY, TbX);
cannam@95 1031 R0[WS(rs, 42)] = FMA(KP1_662939224, Tek, Teh);
cannam@95 1032 R0[WS(rs, 10)] = FNMS(KP1_662939224, Tek, Teh);
cannam@95 1033 R0[WS(rs, 58)] = FMA(KP1_662939224, Tem, Tel);
cannam@95 1034 R0[WS(rs, 26)] = FNMS(KP1_662939224, Tem, Tel);
cannam@95 1035 R0[WS(rs, 50)] = FMA(KP1_961570560, Tee, Ted);
cannam@95 1036 R0[WS(rs, 18)] = FNMS(KP1_961570560, Tee, Ted);
cannam@95 1037 R0[WS(rs, 2)] = FMA(KP1_961570560, Tec, TdL);
cannam@95 1038 R0[WS(rs, 34)] = FNMS(KP1_961570560, Tec, TdL);
cannam@95 1039 }
cannam@95 1040 }
cannam@95 1041 Tcg = FMA(KP923879532, Tc8, Tc7);
cannam@95 1042 Tc9 = FNMS(KP923879532, Tc8, Tc7);
cannam@95 1043 Tce = FMA(KP668178637, Tc0, Tc1);
cannam@95 1044 Tc2 = FNMS(KP668178637, Tc1, Tc0);
cannam@95 1045 {
cannam@95 1046 E Tcf, Tcj, Tcb, Tc3, Tch, Tc6;
cannam@95 1047 Tcf = FNMS(KP1_662939224, Tce, Tcd);
cannam@95 1048 Tcj = FMA(KP1_662939224, Tce, Tcd);
cannam@95 1049 Tcb = FMA(KP1_662939224, Tc2, TbZ);
cannam@95 1050 Tc3 = FNMS(KP1_662939224, Tc2, TbZ);
cannam@95 1051 Tch = FMA(KP923879532, Tc5, Tc4);
cannam@95 1052 Tc6 = FNMS(KP923879532, Tc5, Tc4);
cannam@95 1053 {
cannam@95 1054 E Tci, Tck, Tcc, Tca;
cannam@95 1055 Tci = FNMS(KP303346683, Tch, Tcg);
cannam@95 1056 Tck = FMA(KP303346683, Tcg, Tch);
cannam@95 1057 Tcc = FMA(KP534511135, Tc6, Tc9);
cannam@95 1058 Tca = FNMS(KP534511135, Tc9, Tc6);
cannam@95 1059 TeD = FMA(KP1_414213562, Teo, Ten);
cannam@95 1060 Tep = FNMS(KP1_414213562, Teo, Ten);
cannam@95 1061 R0[WS(rs, 45)] = FMA(KP1_913880671, Tci, Tcf);
cannam@95 1062 R0[WS(rs, 13)] = FNMS(KP1_913880671, Tci, Tcf);
cannam@95 1063 R0[WS(rs, 61)] = FMA(KP1_913880671, Tck, Tcj);
cannam@95 1064 R0[WS(rs, 29)] = FNMS(KP1_913880671, Tck, Tcj);
cannam@95 1065 R0[WS(rs, 53)] = FMA(KP1_763842528, Tcc, Tcb);
cannam@95 1066 R0[WS(rs, 21)] = FNMS(KP1_763842528, Tcc, Tcb);
cannam@95 1067 R0[WS(rs, 5)] = FMA(KP1_763842528, Tca, Tc3);
cannam@95 1068 R0[WS(rs, 37)] = FNMS(KP1_763842528, Tca, Tc3);
cannam@95 1069 }
cannam@95 1070 }
cannam@95 1071 TeG = FMA(KP707106781, Tey, Tex);
cannam@95 1072 Tez = FNMS(KP707106781, Tey, Tex);
cannam@95 1073 TeE = FMA(KP414213562, Teq, Ter);
cannam@95 1074 Tes = FNMS(KP414213562, Ter, Teq);
cannam@95 1075 }
cannam@95 1076 }
cannam@95 1077 }
cannam@95 1078 {
cannam@95 1079 E T5L, T5M, T61, T62;
cannam@95 1080 {
cannam@95 1081 E Td3, Td4, Tdd, TcV, Tcr, Tdb, Td7, Td6, Tda, TcC, TcY, TcR, TcW, Tcy, TcJ;
cannam@95 1082 E Tde;
cannam@95 1083 {
cannam@95 1084 E TcN, TcQ, Tcu, Tcx;
cannam@95 1085 {
cannam@95 1086 E Tcn, TeF, TeJ, TeB, Tet, TeH, Tew, Tcq;
cannam@95 1087 Td3 = FMA(KP1_414213562, Tcm, Tcl);
cannam@95 1088 Tcn = FNMS(KP1_414213562, Tcm, Tcl);
cannam@95 1089 TeF = FNMS(KP1_847759065, TeE, TeD);
cannam@95 1090 TeJ = FMA(KP1_847759065, TeE, TeD);
cannam@95 1091 TeB = FMA(KP1_847759065, Tes, Tep);
cannam@95 1092 Tet = FNMS(KP1_847759065, Tes, Tep);
cannam@95 1093 TeH = FMA(KP707106781, Tev, Teu);
cannam@95 1094 Tew = FNMS(KP707106781, Tev, Teu);
cannam@95 1095 Tcq = FNMS(KP414213562, Tcp, Tco);
cannam@95 1096 Td4 = FMA(KP414213562, Tco, Tcp);
cannam@95 1097 Tdd = FMA(KP707106781, TcM, TcL);
cannam@95 1098 TcN = FNMS(KP707106781, TcM, TcL);
cannam@95 1099 {
cannam@95 1100 E TeI, TeK, TeC, TeA;
cannam@95 1101 TeI = FNMS(KP198912367, TeH, TeG);
cannam@95 1102 TeK = FMA(KP198912367, TeG, TeH);
cannam@95 1103 TeC = FMA(KP668178637, Tew, Tez);
cannam@95 1104 TeA = FNMS(KP668178637, Tez, Tew);
cannam@95 1105 TcV = FMA(KP1_847759065, Tcq, Tcn);
cannam@95 1106 Tcr = FNMS(KP1_847759065, Tcq, Tcn);
cannam@95 1107 R0[WS(rs, 46)] = FMA(KP1_961570560, TeI, TeF);
cannam@95 1108 R0[WS(rs, 14)] = FNMS(KP1_961570560, TeI, TeF);
cannam@95 1109 R0[WS(rs, 62)] = FMA(KP1_961570560, TeK, TeJ);
cannam@95 1110 R0[WS(rs, 30)] = FNMS(KP1_961570560, TeK, TeJ);
cannam@95 1111 R0[WS(rs, 54)] = FMA(KP1_662939224, TeC, TeB);
cannam@95 1112 R0[WS(rs, 22)] = FNMS(KP1_662939224, TeC, TeB);
cannam@95 1113 R0[WS(rs, 6)] = FMA(KP1_662939224, TeA, Tet);
cannam@95 1114 R0[WS(rs, 38)] = FNMS(KP1_662939224, TeA, Tet);
cannam@95 1115 TcQ = TcO - TcP;
cannam@95 1116 Tdb = TcO + TcP;
cannam@95 1117 }
cannam@95 1118 }
cannam@95 1119 Td7 = FMA(KP707106781, Tct, Tcs);
cannam@95 1120 Tcu = FNMS(KP707106781, Tct, Tcs);
cannam@95 1121 Tcx = FNMS(KP707106781, Tcw, Tcv);
cannam@95 1122 Td6 = FMA(KP707106781, Tcw, Tcv);
cannam@95 1123 Tda = FMA(KP707106781, TcB, TcA);
cannam@95 1124 TcC = FNMS(KP707106781, TcB, TcA);
cannam@95 1125 TcY = FNMS(KP923879532, TcQ, TcN);
cannam@95 1126 TcR = FMA(KP923879532, TcQ, TcN);
cannam@95 1127 TcW = FMA(KP668178637, Tcu, Tcx);
cannam@95 1128 Tcy = FNMS(KP668178637, Tcx, Tcu);
cannam@95 1129 TcJ = TcF - TcI;
cannam@95 1130 Tde = TcI + TcF;
cannam@95 1131 }
cannam@95 1132 {
cannam@95 1133 E Tdj, Td5, Tdm, Tdf, Tdk, Td8;
cannam@95 1134 {
cannam@95 1135 E TcX, Td1, TcT, Tcz, TcZ, TcK;
cannam@95 1136 TcX = FNMS(KP1_662939224, TcW, TcV);
cannam@95 1137 Td1 = FMA(KP1_662939224, TcW, TcV);
cannam@95 1138 TcT = FNMS(KP1_662939224, Tcy, Tcr);
cannam@95 1139 Tcz = FMA(KP1_662939224, Tcy, Tcr);
cannam@95 1140 TcZ = FNMS(KP923879532, TcJ, TcC);
cannam@95 1141 TcK = FMA(KP923879532, TcJ, TcC);
cannam@95 1142 {
cannam@95 1143 E Td0, Td2, TcU, TcS;
cannam@95 1144 Td0 = FNMS(KP534511135, TcZ, TcY);
cannam@95 1145 Td2 = FMA(KP534511135, TcY, TcZ);
cannam@95 1146 TcU = FMA(KP303346683, TcK, TcR);
cannam@95 1147 TcS = FNMS(KP303346683, TcR, TcK);
cannam@95 1148 Tdj = FMA(KP1_847759065, Td4, Td3);
cannam@95 1149 Td5 = FNMS(KP1_847759065, Td4, Td3);
cannam@95 1150 R0[WS(rs, 43)] = FMA(KP1_763842528, Td0, TcX);
cannam@95 1151 R0[WS(rs, 11)] = FNMS(KP1_763842528, Td0, TcX);
cannam@95 1152 R0[WS(rs, 59)] = FMA(KP1_763842528, Td2, Td1);
cannam@95 1153 R0[WS(rs, 27)] = FNMS(KP1_763842528, Td2, Td1);
cannam@95 1154 R0[WS(rs, 51)] = FMA(KP1_913880671, TcU, TcT);
cannam@95 1155 R0[WS(rs, 19)] = FNMS(KP1_913880671, TcU, TcT);
cannam@95 1156 R0[WS(rs, 3)] = FMA(KP1_913880671, TcS, Tcz);
cannam@95 1157 R0[WS(rs, 35)] = FNMS(KP1_913880671, TcS, Tcz);
cannam@95 1158 }
cannam@95 1159 }
cannam@95 1160 Tdm = FMA(KP923879532, Tde, Tdd);
cannam@95 1161 Tdf = FNMS(KP923879532, Tde, Tdd);
cannam@95 1162 Tdk = FMA(KP198912367, Td6, Td7);
cannam@95 1163 Td8 = FNMS(KP198912367, Td7, Td6);
cannam@95 1164 {
cannam@95 1165 E T5F, T2R, T5G, T3K, T64, T5S, T5X, T5x, T5U, T4q, T4R, T63, T5P, T5i, T5V;
cannam@95 1166 E T5A;
cannam@95 1167 {
cannam@95 1168 E T5N, T5O, T5R, T3w, T3J, T5Q, T5y, T5z;
cannam@95 1169 {
cannam@95 1170 E T2p, Tdl, Tdp, Tdh, Td9, Tdn, Tdc, T2Q;
cannam@95 1171 T5N = FNMS(KP1_847759065, T2o, T2d);
cannam@95 1172 T2p = FMA(KP1_847759065, T2o, T2d);
cannam@95 1173 Tdl = FNMS(KP1_961570560, Tdk, Tdj);
cannam@95 1174 Tdp = FMA(KP1_961570560, Tdk, Tdj);
cannam@95 1175 Tdh = FMA(KP1_961570560, Td8, Td5);
cannam@95 1176 Td9 = FNMS(KP1_961570560, Td8, Td5);
cannam@95 1177 Tdn = FMA(KP923879532, Tdb, Tda);
cannam@95 1178 Tdc = FNMS(KP923879532, Tdb, Tda);
cannam@95 1179 T2Q = FNMS(KP198912367, T2P, T2G);
cannam@95 1180 T5O = FMA(KP198912367, T2G, T2P);
cannam@95 1181 T5R = FNMS(KP923879532, T3v, T38);
cannam@95 1182 T3w = FMA(KP923879532, T3v, T38);
cannam@95 1183 {
cannam@95 1184 E Tdo, Tdq, Tdi, Tdg;
cannam@95 1185 Tdo = FNMS(KP098491403, Tdn, Tdm);
cannam@95 1186 Tdq = FMA(KP098491403, Tdm, Tdn);
cannam@95 1187 Tdi = FMA(KP820678790, Tdc, Tdf);
cannam@95 1188 Tdg = FNMS(KP820678790, Tdf, Tdc);
cannam@95 1189 T5F = FNMS(KP1_961570560, T2Q, T2p);
cannam@95 1190 T2R = FMA(KP1_961570560, T2Q, T2p);
cannam@95 1191 R0[WS(rs, 47)] = FMA(KP1_990369453, Tdo, Tdl);
cannam@95 1192 R0[WS(rs, 15)] = FNMS(KP1_990369453, Tdo, Tdl);
cannam@95 1193 R0[WS(rs, 63)] = FMA(KP1_990369453, Tdq, Tdp);
cannam@95 1194 R0[WS(rs, 31)] = FNMS(KP1_990369453, Tdq, Tdp);
cannam@95 1195 R0[WS(rs, 55)] = FMA(KP1_546020906, Tdi, Tdh);
cannam@95 1196 R0[WS(rs, 23)] = FNMS(KP1_546020906, Tdi, Tdh);
cannam@95 1197 R0[WS(rs, 7)] = FMA(KP1_546020906, Tdg, Td9);
cannam@95 1198 R0[WS(rs, 39)] = FNMS(KP1_546020906, Tdg, Td9);
cannam@95 1199 T3J = FMA(KP923879532, T3I, T3F);
cannam@95 1200 T5Q = FNMS(KP923879532, T3I, T3F);
cannam@95 1201 }
cannam@95 1202 }
cannam@95 1203 T5G = FMA(KP098491403, T3w, T3J);
cannam@95 1204 T3K = FNMS(KP098491403, T3J, T3w);
cannam@95 1205 T64 = FMA(KP820678790, T5Q, T5R);
cannam@95 1206 T5S = FNMS(KP820678790, T5R, T5Q);
cannam@95 1207 T5X = FNMS(KP923879532, T5w, T5t);
cannam@95 1208 T5x = FMA(KP923879532, T5w, T5t);
cannam@95 1209 T5U = FNMS(KP923879532, T4p, T42);
cannam@95 1210 T4q = FMA(KP923879532, T4p, T42);
cannam@95 1211 T4R = FNMS(KP198912367, T4Q, T4H);
cannam@95 1212 T5y = FMA(KP198912367, T4H, T4Q);
cannam@95 1213 T63 = FMA(KP1_961570560, T5O, T5N);
cannam@95 1214 T5P = FNMS(KP1_961570560, T5O, T5N);
cannam@95 1215 T5z = FNMS(KP198912367, T58, T5h);
cannam@95 1216 T5i = FMA(KP198912367, T5h, T58);
cannam@95 1217 T5V = T5y - T5z;
cannam@95 1218 T5A = T5y + T5z;
cannam@95 1219 }
cannam@95 1220 {
cannam@95 1221 E T5W, T5I, T5Z, T5J;
cannam@95 1222 {
cannam@95 1223 E T5D, T3L, T67, T5B, T5Y, T5j, T65, T69, T66, T5k;
cannam@95 1224 T5D = FNMS(KP1_990369453, T3K, T2R);
cannam@95 1225 T3L = FMA(KP1_990369453, T3K, T2R);
cannam@95 1226 T5W = FNMS(KP980785280, T5V, T5U);
cannam@95 1227 T67 = FMA(KP980785280, T5V, T5U);
cannam@95 1228 T5I = FNMS(KP980785280, T5A, T5x);
cannam@95 1229 T5B = FMA(KP980785280, T5A, T5x);
cannam@95 1230 T5Y = T5i - T4R;
cannam@95 1231 T5j = T4R + T5i;
cannam@95 1232 T65 = FNMS(KP1_546020906, T64, T63);
cannam@95 1233 T69 = FMA(KP1_546020906, T64, T63);
cannam@95 1234 T5Z = FNMS(KP980785280, T5Y, T5X);
cannam@95 1235 T66 = FMA(KP980785280, T5Y, T5X);
cannam@95 1236 T5J = FNMS(KP980785280, T5j, T4q);
cannam@95 1237 T5k = FMA(KP980785280, T5j, T4q);
cannam@95 1238 {
cannam@95 1239 E T68, T6a, T5E, T5C;
cannam@95 1240 T68 = FNMS(KP357805721, T67, T66);
cannam@95 1241 T6a = FMA(KP357805721, T66, T67);
cannam@95 1242 T5E = FMA(KP049126849, T5k, T5B);
cannam@95 1243 T5C = FNMS(KP049126849, T5B, T5k);
cannam@95 1244 R1[WS(rs, 60)] = FMA(KP1_883088130, T6a, T69);
cannam@95 1245 R1[WS(rs, 28)] = FNMS(KP1_883088130, T6a, T69);
cannam@95 1246 R1[WS(rs, 44)] = FMA(KP1_883088130, T68, T65);
cannam@95 1247 R1[WS(rs, 12)] = FNMS(KP1_883088130, T68, T65);
cannam@95 1248 R1[0] = FMA(KP1_997590912, T5C, T3L);
cannam@95 1249 R1[WS(rs, 32)] = FNMS(KP1_997590912, T5C, T3L);
cannam@95 1250 R1[WS(rs, 16)] = FNMS(KP1_997590912, T5E, T5D);
cannam@95 1251 R1[WS(rs, 48)] = FMA(KP1_997590912, T5E, T5D);
cannam@95 1252 }
cannam@95 1253 }
cannam@95 1254 {
cannam@95 1255 E T5H, T5K, T5T, T60;
cannam@95 1256 T5L = FMA(KP1_990369453, T5G, T5F);
cannam@95 1257 T5H = FNMS(KP1_990369453, T5G, T5F);
cannam@95 1258 T5K = FNMS(KP906347169, T5J, T5I);
cannam@95 1259 T5M = FMA(KP906347169, T5I, T5J);
cannam@95 1260 T61 = FMA(KP1_546020906, T5S, T5P);
cannam@95 1261 T5T = FNMS(KP1_546020906, T5S, T5P);
cannam@95 1262 T60 = FNMS(KP472964775, T5Z, T5W);
cannam@95 1263 T62 = FMA(KP472964775, T5W, T5Z);
cannam@95 1264 R1[WS(rs, 40)] = FMA(KP1_481902250, T5K, T5H);
cannam@95 1265 R1[WS(rs, 8)] = FNMS(KP1_481902250, T5K, T5H);
cannam@95 1266 R1[WS(rs, 4)] = FMA(KP1_807978586, T60, T5T);
cannam@95 1267 R1[WS(rs, 36)] = FNMS(KP1_807978586, T60, T5T);
cannam@95 1268 }
cannam@95 1269 }
cannam@95 1270 }
cannam@95 1271 }
cannam@95 1272 }
cannam@95 1273 {
cannam@95 1274 E T8B, T8C, T8R, T8S;
cannam@95 1275 {
cannam@95 1276 E T8v, T7v, T8w, T7O, T8N, T8n, T8U, T8I, T8T, T8F, T8K, T80, T87, T8e, T8L;
cannam@95 1277 E T8q;
cannam@95 1278 {
cannam@95 1279 E T8D, T8E, T8H, T8G, T8o, T8p;
cannam@95 1280 {
cannam@95 1281 E T7n, T7u, T7G, T7N;
cannam@95 1282 T8D = FMA(KP1_847759065, T7m, T7j);
cannam@95 1283 T7n = FNMS(KP1_847759065, T7m, T7j);
cannam@95 1284 R1[WS(rs, 52)] = FMA(KP1_807978586, T62, T61);
cannam@95 1285 R1[WS(rs, 20)] = FNMS(KP1_807978586, T62, T61);
cannam@95 1286 R1[WS(rs, 56)] = FMA(KP1_481902250, T5M, T5L);
cannam@95 1287 R1[WS(rs, 24)] = FNMS(KP1_481902250, T5M, T5L);
cannam@95 1288 T7u = FNMS(KP668178637, T7t, T7q);
cannam@95 1289 T8E = FMA(KP668178637, T7q, T7t);
cannam@95 1290 T8H = FMA(KP923879532, T7F, T7y);
cannam@95 1291 T7G = FNMS(KP923879532, T7F, T7y);
cannam@95 1292 T7N = FMA(KP923879532, T7M, T7J);
cannam@95 1293 T8G = FNMS(KP923879532, T7M, T7J);
cannam@95 1294 T8v = FNMS(KP1_662939224, T7u, T7n);
cannam@95 1295 T7v = FMA(KP1_662939224, T7u, T7n);
cannam@95 1296 T8w = FMA(KP303346683, T7G, T7N);
cannam@95 1297 T7O = FNMS(KP303346683, T7N, T7G);
cannam@95 1298 }
cannam@95 1299 T8N = FNMS(KP923879532, T8m, T8j);
cannam@95 1300 T8n = FMA(KP923879532, T8m, T8j);
cannam@95 1301 T8U = FMA(KP534511135, T8G, T8H);
cannam@95 1302 T8I = FNMS(KP534511135, T8H, T8G);
cannam@95 1303 T8T = FMA(KP1_662939224, T8E, T8D);
cannam@95 1304 T8F = FNMS(KP1_662939224, T8E, T8D);
cannam@95 1305 T8K = FMA(KP923879532, T7Z, T7S);
cannam@95 1306 T80 = FNMS(KP923879532, T7Z, T7S);
cannam@95 1307 T87 = FNMS(KP668178637, T86, T83);
cannam@95 1308 T8o = FMA(KP668178637, T83, T86);
cannam@95 1309 T8p = FMA(KP668178637, T8a, T8d);
cannam@95 1310 T8e = FNMS(KP668178637, T8d, T8a);
cannam@95 1311 T8L = T8o + T8p;
cannam@95 1312 T8q = T8o - T8p;
cannam@95 1313 }
cannam@95 1314 {
cannam@95 1315 E T8M, T8y, T8P, T8z;
cannam@95 1316 {
cannam@95 1317 E T8t, T7P, T8X, T8r, T8O, T8f, T8V, T8Z, T8W, T8g;
cannam@95 1318 T8t = FNMS(KP1_913880671, T7O, T7v);
cannam@95 1319 T7P = FMA(KP1_913880671, T7O, T7v);
cannam@95 1320 T8M = FNMS(KP831469612, T8L, T8K);
cannam@95 1321 T8X = FMA(KP831469612, T8L, T8K);
cannam@95 1322 T8y = FNMS(KP831469612, T8q, T8n);
cannam@95 1323 T8r = FMA(KP831469612, T8q, T8n);
cannam@95 1324 T8O = T8e - T87;
cannam@95 1325 T8f = T87 + T8e;
cannam@95 1326 T8V = FNMS(KP1_763842528, T8U, T8T);
cannam@95 1327 T8Z = FMA(KP1_763842528, T8U, T8T);
cannam@95 1328 T8P = FNMS(KP831469612, T8O, T8N);
cannam@95 1329 T8W = FMA(KP831469612, T8O, T8N);
cannam@95 1330 T8z = FNMS(KP831469612, T8f, T80);
cannam@95 1331 T8g = FMA(KP831469612, T8f, T80);
cannam@95 1332 {
cannam@95 1333 E T8Y, T90, T8u, T8s;
cannam@95 1334 T8Y = FNMS(KP250486960, T8X, T8W);
cannam@95 1335 T90 = FMA(KP250486960, T8W, T8X);
cannam@95 1336 T8u = FMA(KP148335987, T8g, T8r);
cannam@95 1337 T8s = FNMS(KP148335987, T8r, T8g);
cannam@95 1338 R1[WS(rs, 61)] = FMA(KP1_940062506, T90, T8Z);
cannam@95 1339 R1[WS(rs, 29)] = FNMS(KP1_940062506, T90, T8Z);
cannam@95 1340 R1[WS(rs, 45)] = FMA(KP1_940062506, T8Y, T8V);
cannam@95 1341 R1[WS(rs, 13)] = FNMS(KP1_940062506, T8Y, T8V);
cannam@95 1342 R1[WS(rs, 1)] = FMA(KP1_978353019, T8s, T7P);
cannam@95 1343 R1[WS(rs, 33)] = FNMS(KP1_978353019, T8s, T7P);
cannam@95 1344 R1[WS(rs, 17)] = FNMS(KP1_978353019, T8u, T8t);
cannam@95 1345 R1[WS(rs, 49)] = FMA(KP1_978353019, T8u, T8t);
cannam@95 1346 }
cannam@95 1347 }
cannam@95 1348 {
cannam@95 1349 E T8x, T8A, T8J, T8Q;
cannam@95 1350 T8B = FMA(KP1_913880671, T8w, T8v);
cannam@95 1351 T8x = FNMS(KP1_913880671, T8w, T8v);
cannam@95 1352 T8A = FNMS(KP741650546, T8z, T8y);
cannam@95 1353 T8C = FMA(KP741650546, T8y, T8z);
cannam@95 1354 T8R = FMA(KP1_763842528, T8I, T8F);
cannam@95 1355 T8J = FNMS(KP1_763842528, T8I, T8F);
cannam@95 1356 T8Q = FNMS(KP599376933, T8P, T8M);
cannam@95 1357 T8S = FMA(KP599376933, T8M, T8P);
cannam@95 1358 R1[WS(rs, 41)] = FMA(KP1_606415062, T8A, T8x);
cannam@95 1359 R1[WS(rs, 9)] = FNMS(KP1_606415062, T8A, T8x);
cannam@95 1360 R1[WS(rs, 5)] = FMA(KP1_715457220, T8Q, T8J);
cannam@95 1361 R1[WS(rs, 37)] = FNMS(KP1_715457220, T8Q, T8J);
cannam@95 1362 }
cannam@95 1363 }
cannam@95 1364 }
cannam@95 1365 {
cannam@95 1366 E T6R, T6S, T77, T78;
cannam@95 1367 {
cannam@95 1368 E T6L, T6h, T6M, T6o, T73, T6D, T7a, T6Y, T79, T6V, T70, T6s, T6y, T6v, T71;
cannam@95 1369 E T6G;
cannam@95 1370 {
cannam@95 1371 E T6T, T6U, T6X, T6W, T6E, T6F;
cannam@95 1372 {
cannam@95 1373 E T6d, T6g, T6k, T6n;
cannam@95 1374 T6T = FMA(KP1_847759065, T6c, T6b);
cannam@95 1375 T6d = FNMS(KP1_847759065, T6c, T6b);
cannam@95 1376 R1[WS(rs, 53)] = FMA(KP1_715457220, T8S, T8R);
cannam@95 1377 R1[WS(rs, 21)] = FNMS(KP1_715457220, T8S, T8R);
cannam@95 1378 R1[WS(rs, 57)] = FMA(KP1_606415062, T8C, T8B);
cannam@95 1379 R1[WS(rs, 25)] = FNMS(KP1_606415062, T8C, T8B);
cannam@95 1380 T6g = FNMS(KP668178637, T6f, T6e);
cannam@95 1381 T6U = FMA(KP668178637, T6e, T6f);
cannam@95 1382 T6X = FMA(KP923879532, T6j, T6i);
cannam@95 1383 T6k = FNMS(KP923879532, T6j, T6i);
cannam@95 1384 T6n = FNMS(KP923879532, T6m, T6l);
cannam@95 1385 T6W = FMA(KP923879532, T6m, T6l);
cannam@95 1386 T6L = FMA(KP1_662939224, T6g, T6d);
cannam@95 1387 T6h = FNMS(KP1_662939224, T6g, T6d);
cannam@95 1388 T6M = FMA(KP534511135, T6k, T6n);
cannam@95 1389 T6o = FNMS(KP534511135, T6n, T6k);
cannam@95 1390 }
cannam@95 1391 T73 = FMA(KP923879532, T6C, T6B);
cannam@95 1392 T6D = FNMS(KP923879532, T6C, T6B);
cannam@95 1393 T7a = FMA(KP303346683, T6W, T6X);
cannam@95 1394 T6Y = FNMS(KP303346683, T6X, T6W);
cannam@95 1395 T79 = FMA(KP1_662939224, T6U, T6T);
cannam@95 1396 T6V = FNMS(KP1_662939224, T6U, T6T);
cannam@95 1397 T70 = FMA(KP923879532, T6r, T6q);
cannam@95 1398 T6s = FNMS(KP923879532, T6r, T6q);
cannam@95 1399 T6y = FNMS(KP668178637, T6x, T6w);
cannam@95 1400 T6E = FMA(KP668178637, T6w, T6x);
cannam@95 1401 T6F = FNMS(KP668178637, T6t, T6u);
cannam@95 1402 T6v = FMA(KP668178637, T6u, T6t);
cannam@95 1403 T71 = T6E + T6F;
cannam@95 1404 T6G = T6E - T6F;
cannam@95 1405 }
cannam@95 1406 {
cannam@95 1407 E T72, T6O, T75, T6P;
cannam@95 1408 {
cannam@95 1409 E T6J, T6p, T7d, T6H, T74, T6z, T7b, T7f, T7c, T6A;
cannam@95 1410 T6J = FNMS(KP1_763842528, T6o, T6h);
cannam@95 1411 T6p = FMA(KP1_763842528, T6o, T6h);
cannam@95 1412 T72 = FNMS(KP831469612, T71, T70);
cannam@95 1413 T7d = FMA(KP831469612, T71, T70);
cannam@95 1414 T6O = FNMS(KP831469612, T6G, T6D);
cannam@95 1415 T6H = FMA(KP831469612, T6G, T6D);
cannam@95 1416 T74 = T6y + T6v;
cannam@95 1417 T6z = T6v - T6y;
cannam@95 1418 T7b = FNMS(KP1_913880671, T7a, T79);
cannam@95 1419 T7f = FMA(KP1_913880671, T7a, T79);
cannam@95 1420 T75 = FNMS(KP831469612, T74, T73);
cannam@95 1421 T7c = FMA(KP831469612, T74, T73);
cannam@95 1422 T6P = FNMS(KP831469612, T6z, T6s);
cannam@95 1423 T6A = FMA(KP831469612, T6z, T6s);
cannam@95 1424 {
cannam@95 1425 E T7e, T7g, T6K, T6I;
cannam@95 1426 T7e = FNMS(KP148335987, T7d, T7c);
cannam@95 1427 T7g = FMA(KP148335987, T7c, T7d);
cannam@95 1428 T6K = FMA(KP250486960, T6A, T6H);
cannam@95 1429 T6I = FNMS(KP250486960, T6H, T6A);
cannam@95 1430 R1[WS(rs, 62)] = FMA(KP1_978353019, T7g, T7f);
cannam@95 1431 R1[WS(rs, 30)] = FNMS(KP1_978353019, T7g, T7f);
cannam@95 1432 R1[WS(rs, 46)] = FMA(KP1_978353019, T7e, T7b);
cannam@95 1433 R1[WS(rs, 14)] = FNMS(KP1_978353019, T7e, T7b);
cannam@95 1434 R1[WS(rs, 2)] = FMA(KP1_940062506, T6I, T6p);
cannam@95 1435 R1[WS(rs, 34)] = FNMS(KP1_940062506, T6I, T6p);
cannam@95 1436 R1[WS(rs, 18)] = FNMS(KP1_940062506, T6K, T6J);
cannam@95 1437 R1[WS(rs, 50)] = FMA(KP1_940062506, T6K, T6J);
cannam@95 1438 }
cannam@95 1439 }
cannam@95 1440 {
cannam@95 1441 E T6N, T6Q, T6Z, T76;
cannam@95 1442 T6R = FMA(KP1_763842528, T6M, T6L);
cannam@95 1443 T6N = FNMS(KP1_763842528, T6M, T6L);
cannam@95 1444 T6Q = FNMS(KP599376933, T6P, T6O);
cannam@95 1445 T6S = FMA(KP599376933, T6O, T6P);
cannam@95 1446 T77 = FMA(KP1_913880671, T6Y, T6V);
cannam@95 1447 T6Z = FNMS(KP1_913880671, T6Y, T6V);
cannam@95 1448 T76 = FNMS(KP741650546, T75, T72);
cannam@95 1449 T78 = FMA(KP741650546, T72, T75);
cannam@95 1450 R1[WS(rs, 42)] = FMA(KP1_715457220, T6Q, T6N);
cannam@95 1451 R1[WS(rs, 10)] = FNMS(KP1_715457220, T6Q, T6N);
cannam@95 1452 R1[WS(rs, 6)] = FMA(KP1_606415062, T76, T6Z);
cannam@95 1453 R1[WS(rs, 38)] = FNMS(KP1_606415062, T76, T6Z);
cannam@95 1454 }
cannam@95 1455 }
cannam@95 1456 }
cannam@95 1457 {
cannam@95 1458 E T9B, T97, T9C, T9e, T9T, T9t, Ta0, T9O, T9Z, T9L, T9Q, T9i, T9l, T9o, T9R;
cannam@95 1459 E T9w;
cannam@95 1460 {
cannam@95 1461 E T9J, T9K, T9N, T9M, T9u, T9v;
cannam@95 1462 {
cannam@95 1463 E T93, T96, T9a, T9d;
cannam@95 1464 T9J = FMA(KP1_847759065, T92, T91);
cannam@95 1465 T93 = FNMS(KP1_847759065, T92, T91);
cannam@95 1466 R1[WS(rs, 54)] = FMA(KP1_606415062, T78, T77);
cannam@95 1467 R1[WS(rs, 22)] = FNMS(KP1_606415062, T78, T77);
cannam@95 1468 R1[WS(rs, 58)] = FMA(KP1_715457220, T6S, T6R);
cannam@95 1469 R1[WS(rs, 26)] = FNMS(KP1_715457220, T6S, T6R);
cannam@95 1470 T96 = FNMS(KP198912367, T95, T94);
cannam@95 1471 T9K = FMA(KP198912367, T94, T95);
cannam@95 1472 T9N = FMA(KP923879532, T99, T98);
cannam@95 1473 T9a = FNMS(KP923879532, T99, T98);
cannam@95 1474 T9d = FNMS(KP923879532, T9c, T9b);
cannam@95 1475 T9M = FMA(KP923879532, T9c, T9b);
cannam@95 1476 T9B = FMA(KP1_961570560, T96, T93);
cannam@95 1477 T97 = FNMS(KP1_961570560, T96, T93);
cannam@95 1478 T9C = FMA(KP820678790, T9a, T9d);
cannam@95 1479 T9e = FNMS(KP820678790, T9d, T9a);
cannam@95 1480 }
cannam@95 1481 T9T = FMA(KP923879532, T9s, T9r);
cannam@95 1482 T9t = FNMS(KP923879532, T9s, T9r);
cannam@95 1483 Ta0 = FMA(KP098491403, T9M, T9N);
cannam@95 1484 T9O = FNMS(KP098491403, T9N, T9M);
cannam@95 1485 T9Z = FMA(KP1_961570560, T9K, T9J);
cannam@95 1486 T9L = FNMS(KP1_961570560, T9K, T9J);
cannam@95 1487 T9Q = FMA(KP923879532, T9h, T9g);
cannam@95 1488 T9i = FNMS(KP923879532, T9h, T9g);
cannam@95 1489 T9l = FNMS(KP198912367, T9k, T9j);
cannam@95 1490 T9u = FMA(KP198912367, T9j, T9k);
cannam@95 1491 T9v = FMA(KP198912367, T9m, T9n);
cannam@95 1492 T9o = FNMS(KP198912367, T9n, T9m);
cannam@95 1493 T9R = T9u + T9v;
cannam@95 1494 T9w = T9u - T9v;
cannam@95 1495 }
cannam@95 1496 {
cannam@95 1497 E T9S, T9E, T9V, T9F;
cannam@95 1498 {
cannam@95 1499 E T9z, T9f, Ta3, T9x, T9U, T9p, Ta1, Ta5, Ta2, T9q;
cannam@95 1500 T9z = FNMS(KP1_546020906, T9e, T97);
cannam@95 1501 T9f = FMA(KP1_546020906, T9e, T97);
cannam@95 1502 T9S = FNMS(KP980785280, T9R, T9Q);
cannam@95 1503 Ta3 = FMA(KP980785280, T9R, T9Q);
cannam@95 1504 T9E = FNMS(KP980785280, T9w, T9t);
cannam@95 1505 T9x = FMA(KP980785280, T9w, T9t);
cannam@95 1506 T9U = T9l - T9o;
cannam@95 1507 T9p = T9l + T9o;
cannam@95 1508 Ta1 = FNMS(KP1_990369453, Ta0, T9Z);
cannam@95 1509 Ta5 = FMA(KP1_990369453, Ta0, T9Z);
cannam@95 1510 T9V = FNMS(KP980785280, T9U, T9T);
cannam@95 1511 Ta2 = FMA(KP980785280, T9U, T9T);
cannam@95 1512 T9F = FMA(KP980785280, T9p, T9i);
cannam@95 1513 T9q = FNMS(KP980785280, T9p, T9i);
cannam@95 1514 {
cannam@95 1515 E Ta4, Ta6, T9A, T9y;
cannam@95 1516 Ta4 = FNMS(KP049126849, Ta3, Ta2);
cannam@95 1517 Ta6 = FMA(KP049126849, Ta2, Ta3);
cannam@95 1518 T9A = FMA(KP357805721, T9q, T9x);
cannam@95 1519 T9y = FNMS(KP357805721, T9x, T9q);
cannam@95 1520 R1[WS(rs, 63)] = FMA(KP1_997590912, Ta6, Ta5);
cannam@95 1521 R1[WS(rs, 31)] = FNMS(KP1_997590912, Ta6, Ta5);
cannam@95 1522 R1[WS(rs, 47)] = FMA(KP1_997590912, Ta4, Ta1);
cannam@95 1523 R1[WS(rs, 15)] = FNMS(KP1_997590912, Ta4, Ta1);
cannam@95 1524 R1[WS(rs, 3)] = FMA(KP1_883088130, T9y, T9f);
cannam@95 1525 R1[WS(rs, 35)] = FNMS(KP1_883088130, T9y, T9f);
cannam@95 1526 R1[WS(rs, 19)] = FNMS(KP1_883088130, T9A, T9z);
cannam@95 1527 R1[WS(rs, 51)] = FMA(KP1_883088130, T9A, T9z);
cannam@95 1528 }
cannam@95 1529 }
cannam@95 1530 {
cannam@95 1531 E T9D, T9G, T9P, T9W;
cannam@95 1532 T9H = FMA(KP1_546020906, T9C, T9B);
cannam@95 1533 T9D = FNMS(KP1_546020906, T9C, T9B);
cannam@95 1534 T9G = FNMS(KP472964775, T9F, T9E);
cannam@95 1535 T9I = FMA(KP472964775, T9E, T9F);
cannam@95 1536 T9X = FMA(KP1_990369453, T9O, T9L);
cannam@95 1537 T9P = FNMS(KP1_990369453, T9O, T9L);
cannam@95 1538 T9W = FNMS(KP906347169, T9V, T9S);
cannam@95 1539 T9Y = FMA(KP906347169, T9S, T9V);
cannam@95 1540 R1[WS(rs, 43)] = FMA(KP1_807978586, T9G, T9D);
cannam@95 1541 R1[WS(rs, 11)] = FNMS(KP1_807978586, T9G, T9D);
cannam@95 1542 R1[WS(rs, 7)] = FMA(KP1_481902250, T9W, T9P);
cannam@95 1543 R1[WS(rs, 39)] = FNMS(KP1_481902250, T9W, T9P);
cannam@95 1544 }
cannam@95 1545 }
cannam@95 1546 }
cannam@95 1547 }
cannam@95 1548 }
cannam@95 1549 }
cannam@95 1550 }
cannam@95 1551 }
cannam@95 1552 R1[WS(rs, 55)] = FMA(KP1_481902250, T9Y, T9X);
cannam@95 1553 R1[WS(rs, 23)] = FNMS(KP1_481902250, T9Y, T9X);
cannam@95 1554 R1[WS(rs, 59)] = FMA(KP1_807978586, T9I, T9H);
cannam@95 1555 R1[WS(rs, 27)] = FNMS(KP1_807978586, T9I, T9H);
cannam@95 1556 }
cannam@95 1557 }
cannam@95 1558 }
cannam@95 1559
cannam@95 1560 static const kr2c_desc desc = { 128, "r2cb_128", {416, 0, 540, 0}, &GENUS };
cannam@95 1561
cannam@95 1562 void X(codelet_r2cb_128) (planner *p) {
cannam@95 1563 X(kr2c_register) (p, r2cb_128, &desc);
cannam@95 1564 }
cannam@95 1565
cannam@95 1566 #else /* HAVE_FMA */
cannam@95 1567
cannam@95 1568 /* Generated by: ../../../genfft/gen_r2cb.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 128 -name r2cb_128 -include r2cb.h */
cannam@95 1569
cannam@95 1570 /*
cannam@95 1571 * This function contains 956 FP additions, 342 FP multiplications,
cannam@95 1572 * (or, 812 additions, 198 multiplications, 144 fused multiply/add),
cannam@95 1573 * 198 stack variables, 39 constants, and 256 memory accesses
cannam@95 1574 */
cannam@95 1575 #include "r2cb.h"
cannam@95 1576
cannam@95 1577 static void r2cb_128(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
cannam@95 1578 {
cannam@95 1579 DK(KP1_028205488, +1.028205488386443453187387677937631545216098241);
cannam@95 1580 DK(KP1_715457220, +1.715457220000544139804539968569540274084981599);
cannam@95 1581 DK(KP1_606415062, +1.606415062961289819613353025926283847759138854);
cannam@95 1582 DK(KP1_191398608, +1.191398608984866686934073057659939779023852677);
cannam@95 1583 DK(KP1_940062506, +1.940062506389087985207968414572200502913731924);
cannam@95 1584 DK(KP485960359, +0.485960359806527779896548324154942236641981567);
cannam@95 1585 DK(KP293460948, +0.293460948910723503317700259293435639412430633);
cannam@95 1586 DK(KP1_978353019, +1.978353019929561946903347476032486127967379067);
cannam@95 1587 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
cannam@95 1588 DK(KP555570233, +0.555570233019602224742830813948532874374937191);
cannam@95 1589 DK(KP855110186, +0.855110186860564188641933713777597068609157259);
cannam@95 1590 DK(KP1_807978586, +1.807978586246886663172400594461074097420264050);
cannam@95 1591 DK(KP1_481902250, +1.481902250709918182351233794990325459457910619);
cannam@95 1592 DK(KP1_343117909, +1.343117909694036801250753700854843606457501264);
cannam@95 1593 DK(KP1_883088130, +1.883088130366041556825018805199004714371179592);
cannam@95 1594 DK(KP673779706, +0.673779706784440101378506425238295140955533559);
cannam@95 1595 DK(KP098135348, +0.098135348654836028509909953885365316629490726);
cannam@95 1596 DK(KP1_997590912, +1.997590912410344785429543209518201388886407229);
cannam@95 1597 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
cannam@95 1598 DK(KP195090322, +0.195090322016128267848284868477022240927691618);
cannam@95 1599 DK(KP580569354, +0.580569354508924735272384751634790549382952557);
cannam@95 1600 DK(KP1_913880671, +1.913880671464417729871595773960539938965698411);
cannam@95 1601 DK(KP942793473, +0.942793473651995297112775251810508755314920638);
cannam@95 1602 DK(KP1_763842528, +1.763842528696710059425513727320776699016885241);
cannam@95 1603 DK(KP1_111140466, +1.111140466039204449485661627897065748749874382);
cannam@95 1604 DK(KP1_662939224, +1.662939224605090474157576755235811513477121624);
cannam@95 1605 DK(KP1_268786568, +1.268786568327290996430343226450986741351374190);
cannam@95 1606 DK(KP1_546020906, +1.546020906725473921621813219516939601942082586);
cannam@95 1607 DK(KP196034280, +0.196034280659121203988391127777283691722273346);
cannam@95 1608 DK(KP1_990369453, +1.990369453344393772489673906218959843150949737);
cannam@95 1609 DK(KP390180644, +0.390180644032256535696569736954044481855383236);
cannam@95 1610 DK(KP1_961570560, +1.961570560806460898252364472268478073947867462);
cannam@95 1611 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
cannam@95 1612 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
cannam@95 1613 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
cannam@95 1614 DK(KP765366864, +0.765366864730179543456919968060797733522689125);
cannam@95 1615 DK(KP1_847759065, +1.847759065022573512256366378793576573644833252);
cannam@95 1616 DK(KP1_414213562, +1.414213562373095048801688724209698078569671875);
cannam@95 1617 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
cannam@95 1618 {
cannam@95 1619 INT i;
cannam@95 1620 for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(512, rs), MAKE_VOLATILE_STRIDE(512, csr), MAKE_VOLATILE_STRIDE(512, csi)) {
cannam@95 1621 E Ta, T6q, T2a, T5k, T8x, Tbx, TcF, Ten, Th, T6r, T2j, T5l, T8E, Tby, TcI;
cannam@95 1622 E Teo, Tx, T6t, TcM, Teq, TcP, Ter, T2t, T5n, T2C, T5o, T8Q, TbA, T8X, TbB;
cannam@95 1623 E T6w, T7L, T1j, T6L, Tde, TeC, TdL, TeR, T3v, T5z, T4I, T5O, T9O, TbM, TaV;
cannam@95 1624 E Tc1, T78, T7Z, TN, T6z, TcU, Teu, Td8, Tey, T2N, T5r, T3j, T5v, T9a, TbE;
cannam@95 1625 E T9A, TbI, T6H, T7O, T1O, T7V, T48, T4u, Tds, TeG, T5E, T5K, Taf, TbP, Tdp;
cannam@95 1626 E TeF, T6U, T72, Tam, TbQ, T23, T7U, T4r, T4v, Tdz, TeJ, T5H, T5L, Tay, TbS;
cannam@95 1627 E Tdw, TeI, T6Z, T73, TaF, TbT, T1y, T75, Tdl, TeQ, TdI, TeD, T3O, T5N, T4z;
cannam@95 1628 E T5A, Ta3, Tc0, TaO, TbN, T6O, T80, T12, T6E, Td1, Tex, Td5, Tev, T36, T5u;
cannam@95 1629 E T3a, T5s, T9p, TbH, T9t, TbF, T6C, T7P;
cannam@95 1630 {
cannam@95 1631 E T5, T8s, T3, T8q, T9, T8u, T29, T8v, T6, T26;
cannam@95 1632 {
cannam@95 1633 E T4, T8r, T1, T2;
cannam@95 1634 T4 = Cr[WS(csr, 32)];
cannam@95 1635 T5 = KP2_000000000 * T4;
cannam@95 1636 T8r = Ci[WS(csi, 32)];
cannam@95 1637 T8s = KP2_000000000 * T8r;
cannam@95 1638 T1 = Cr[0];
cannam@95 1639 T2 = Cr[WS(csr, 64)];
cannam@95 1640 T3 = T1 + T2;
cannam@95 1641 T8q = T1 - T2;
cannam@95 1642 {
cannam@95 1643 E T7, T8, T27, T28;
cannam@95 1644 T7 = Cr[WS(csr, 16)];
cannam@95 1645 T8 = Cr[WS(csr, 48)];
cannam@95 1646 T9 = KP2_000000000 * (T7 + T8);
cannam@95 1647 T8u = T7 - T8;
cannam@95 1648 T27 = Ci[WS(csi, 16)];
cannam@95 1649 T28 = Ci[WS(csi, 48)];
cannam@95 1650 T29 = KP2_000000000 * (T27 - T28);
cannam@95 1651 T8v = T27 + T28;
cannam@95 1652 }
cannam@95 1653 }
cannam@95 1654 T6 = T3 + T5;
cannam@95 1655 Ta = T6 + T9;
cannam@95 1656 T6q = T6 - T9;
cannam@95 1657 T26 = T3 - T5;
cannam@95 1658 T2a = T26 - T29;
cannam@95 1659 T5k = T26 + T29;
cannam@95 1660 {
cannam@95 1661 E T8t, T8w, TcD, TcE;
cannam@95 1662 T8t = T8q - T8s;
cannam@95 1663 T8w = KP1_414213562 * (T8u - T8v);
cannam@95 1664 T8x = T8t + T8w;
cannam@95 1665 Tbx = T8t - T8w;
cannam@95 1666 TcD = T8q + T8s;
cannam@95 1667 TcE = KP1_414213562 * (T8u + T8v);
cannam@95 1668 TcF = TcD - TcE;
cannam@95 1669 Ten = TcD + TcE;
cannam@95 1670 }
cannam@95 1671 }
cannam@95 1672 {
cannam@95 1673 E Td, T8y, T2e, T8C, Tg, T8B, T2h, T8z, T2b, T2i;
cannam@95 1674 {
cannam@95 1675 E Tb, Tc, T2c, T2d;
cannam@95 1676 Tb = Cr[WS(csr, 8)];
cannam@95 1677 Tc = Cr[WS(csr, 56)];
cannam@95 1678 Td = Tb + Tc;
cannam@95 1679 T8y = Tb - Tc;
cannam@95 1680 T2c = Ci[WS(csi, 8)];
cannam@95 1681 T2d = Ci[WS(csi, 56)];
cannam@95 1682 T2e = T2c - T2d;
cannam@95 1683 T8C = T2c + T2d;
cannam@95 1684 }
cannam@95 1685 {
cannam@95 1686 E Te, Tf, T2f, T2g;
cannam@95 1687 Te = Cr[WS(csr, 40)];
cannam@95 1688 Tf = Cr[WS(csr, 24)];
cannam@95 1689 Tg = Te + Tf;
cannam@95 1690 T8B = Te - Tf;
cannam@95 1691 T2f = Ci[WS(csi, 40)];
cannam@95 1692 T2g = Ci[WS(csi, 24)];
cannam@95 1693 T2h = T2f - T2g;
cannam@95 1694 T8z = T2f + T2g;
cannam@95 1695 }
cannam@95 1696 Th = KP2_000000000 * (Td + Tg);
cannam@95 1697 T6r = KP2_000000000 * (T2h + T2e);
cannam@95 1698 T2b = Td - Tg;
cannam@95 1699 T2i = T2e - T2h;
cannam@95 1700 T2j = KP1_414213562 * (T2b - T2i);
cannam@95 1701 T5l = KP1_414213562 * (T2b + T2i);
cannam@95 1702 {
cannam@95 1703 E T8A, T8D, TcG, TcH;
cannam@95 1704 T8A = T8y - T8z;
cannam@95 1705 T8D = T8B + T8C;
cannam@95 1706 T8E = FNMS(KP765366864, T8D, KP1_847759065 * T8A);
cannam@95 1707 Tby = FMA(KP765366864, T8A, KP1_847759065 * T8D);
cannam@95 1708 TcG = T8y + T8z;
cannam@95 1709 TcH = T8C - T8B;
cannam@95 1710 TcI = FNMS(KP1_847759065, TcH, KP765366864 * TcG);
cannam@95 1711 Teo = FMA(KP1_847759065, TcG, KP765366864 * TcH);
cannam@95 1712 }
cannam@95 1713 }
cannam@95 1714 {
cannam@95 1715 E Tl, T8G, T2x, T8V, To, T8U, T2A, T8H, Tv, T8S, T2o, T8O, Ts, T8R, T2r;
cannam@95 1716 E T8L;
cannam@95 1717 {
cannam@95 1718 E Tj, Tk, T2y, T2z;
cannam@95 1719 Tj = Cr[WS(csr, 4)];
cannam@95 1720 Tk = Cr[WS(csr, 60)];
cannam@95 1721 Tl = Tj + Tk;
cannam@95 1722 T8G = Tj - Tk;
cannam@95 1723 {
cannam@95 1724 E T2v, T2w, Tm, Tn;
cannam@95 1725 T2v = Ci[WS(csi, 4)];
cannam@95 1726 T2w = Ci[WS(csi, 60)];
cannam@95 1727 T2x = T2v - T2w;
cannam@95 1728 T8V = T2v + T2w;
cannam@95 1729 Tm = Cr[WS(csr, 36)];
cannam@95 1730 Tn = Cr[WS(csr, 28)];
cannam@95 1731 To = Tm + Tn;
cannam@95 1732 T8U = Tm - Tn;
cannam@95 1733 }
cannam@95 1734 T2y = Ci[WS(csi, 36)];
cannam@95 1735 T2z = Ci[WS(csi, 28)];
cannam@95 1736 T2A = T2y - T2z;
cannam@95 1737 T8H = T2y + T2z;
cannam@95 1738 {
cannam@95 1739 E Tt, Tu, T8M, T2m, T2n, T8N;
cannam@95 1740 Tt = Cr[WS(csr, 12)];
cannam@95 1741 Tu = Cr[WS(csr, 52)];
cannam@95 1742 T8M = Tt - Tu;
cannam@95 1743 T2m = Ci[WS(csi, 52)];
cannam@95 1744 T2n = Ci[WS(csi, 12)];
cannam@95 1745 T8N = T2n + T2m;
cannam@95 1746 Tv = Tt + Tu;
cannam@95 1747 T8S = T8M + T8N;
cannam@95 1748 T2o = T2m - T2n;
cannam@95 1749 T8O = T8M - T8N;
cannam@95 1750 }
cannam@95 1751 {
cannam@95 1752 E Tq, Tr, T8J, T2p, T2q, T8K;
cannam@95 1753 Tq = Cr[WS(csr, 20)];
cannam@95 1754 Tr = Cr[WS(csr, 44)];
cannam@95 1755 T8J = Tq - Tr;
cannam@95 1756 T2p = Ci[WS(csi, 20)];
cannam@95 1757 T2q = Ci[WS(csi, 44)];
cannam@95 1758 T8K = T2p + T2q;
cannam@95 1759 Ts = Tq + Tr;
cannam@95 1760 T8R = T8J + T8K;
cannam@95 1761 T2r = T2p - T2q;
cannam@95 1762 T8L = T8J - T8K;
cannam@95 1763 }
cannam@95 1764 }
cannam@95 1765 {
cannam@95 1766 E Tp, Tw, TcK, TcL;
cannam@95 1767 Tp = Tl + To;
cannam@95 1768 Tw = Ts + Tv;
cannam@95 1769 Tx = KP2_000000000 * (Tp + Tw);
cannam@95 1770 T6t = Tp - Tw;
cannam@95 1771 TcK = T8G + T8H;
cannam@95 1772 TcL = KP707106781 * (T8R + T8S);
cannam@95 1773 TcM = TcK - TcL;
cannam@95 1774 Teq = TcK + TcL;
cannam@95 1775 }
cannam@95 1776 {
cannam@95 1777 E TcN, TcO, T2l, T2s;
cannam@95 1778 TcN = KP707106781 * (T8L - T8O);
cannam@95 1779 TcO = T8V - T8U;
cannam@95 1780 TcP = TcN + TcO;
cannam@95 1781 Ter = TcO - TcN;
cannam@95 1782 T2l = Tl - To;
cannam@95 1783 T2s = T2o - T2r;
cannam@95 1784 T2t = T2l + T2s;
cannam@95 1785 T5n = T2l - T2s;
cannam@95 1786 }
cannam@95 1787 {
cannam@95 1788 E T2u, T2B, T8I, T8P;
cannam@95 1789 T2u = Ts - Tv;
cannam@95 1790 T2B = T2x - T2A;
cannam@95 1791 T2C = T2u + T2B;
cannam@95 1792 T5o = T2B - T2u;
cannam@95 1793 T8I = T8G - T8H;
cannam@95 1794 T8P = KP707106781 * (T8L + T8O);
cannam@95 1795 T8Q = T8I + T8P;
cannam@95 1796 TbA = T8I - T8P;
cannam@95 1797 }
cannam@95 1798 {
cannam@95 1799 E T8T, T8W, T6u, T6v;
cannam@95 1800 T8T = KP707106781 * (T8R - T8S);
cannam@95 1801 T8W = T8U + T8V;
cannam@95 1802 T8X = T8T + T8W;
cannam@95 1803 TbB = T8W - T8T;
cannam@95 1804 T6u = T2A + T2x;
cannam@95 1805 T6v = T2r + T2o;
cannam@95 1806 T6w = T6u - T6v;
cannam@95 1807 T7L = KP2_000000000 * (T6v + T6u);
cannam@95 1808 }
cannam@95 1809 }
cannam@95 1810 {
cannam@95 1811 E T17, T9E, T4D, TaT, T1a, TaS, T4G, T9F, T1h, TaQ, T3q, T9M, T1e, TaP, T3t;
cannam@95 1812 E T9J;
cannam@95 1813 {
cannam@95 1814 E T15, T16, T4E, T4F;
cannam@95 1815 T15 = Cr[WS(csr, 1)];
cannam@95 1816 T16 = Cr[WS(csr, 63)];
cannam@95 1817 T17 = T15 + T16;
cannam@95 1818 T9E = T15 - T16;
cannam@95 1819 {
cannam@95 1820 E T4B, T4C, T18, T19;
cannam@95 1821 T4B = Ci[WS(csi, 1)];
cannam@95 1822 T4C = Ci[WS(csi, 63)];
cannam@95 1823 T4D = T4B - T4C;
cannam@95 1824 TaT = T4B + T4C;
cannam@95 1825 T18 = Cr[WS(csr, 33)];
cannam@95 1826 T19 = Cr[WS(csr, 31)];
cannam@95 1827 T1a = T18 + T19;
cannam@95 1828 TaS = T18 - T19;
cannam@95 1829 }
cannam@95 1830 T4E = Ci[WS(csi, 33)];
cannam@95 1831 T4F = Ci[WS(csi, 31)];
cannam@95 1832 T4G = T4E - T4F;
cannam@95 1833 T9F = T4E + T4F;
cannam@95 1834 {
cannam@95 1835 E T1f, T1g, T9K, T3o, T3p, T9L;
cannam@95 1836 T1f = Cr[WS(csr, 15)];
cannam@95 1837 T1g = Cr[WS(csr, 49)];
cannam@95 1838 T9K = T1f - T1g;
cannam@95 1839 T3o = Ci[WS(csi, 49)];
cannam@95 1840 T3p = Ci[WS(csi, 15)];
cannam@95 1841 T9L = T3p + T3o;
cannam@95 1842 T1h = T1f + T1g;
cannam@95 1843 TaQ = T9K + T9L;
cannam@95 1844 T3q = T3o - T3p;
cannam@95 1845 T9M = T9K - T9L;
cannam@95 1846 }
cannam@95 1847 {
cannam@95 1848 E T1c, T1d, T9H, T3r, T3s, T9I;
cannam@95 1849 T1c = Cr[WS(csr, 17)];
cannam@95 1850 T1d = Cr[WS(csr, 47)];
cannam@95 1851 T9H = T1c - T1d;
cannam@95 1852 T3r = Ci[WS(csi, 17)];
cannam@95 1853 T3s = Ci[WS(csi, 47)];
cannam@95 1854 T9I = T3r + T3s;
cannam@95 1855 T1e = T1c + T1d;
cannam@95 1856 TaP = T9H + T9I;
cannam@95 1857 T3t = T3r - T3s;
cannam@95 1858 T9J = T9H - T9I;
cannam@95 1859 }
cannam@95 1860 }
cannam@95 1861 {
cannam@95 1862 E T1b, T1i, Tdc, Tdd;
cannam@95 1863 T1b = T17 + T1a;
cannam@95 1864 T1i = T1e + T1h;
cannam@95 1865 T1j = T1b + T1i;
cannam@95 1866 T6L = T1b - T1i;
cannam@95 1867 Tdc = T9E + T9F;
cannam@95 1868 Tdd = KP707106781 * (TaP + TaQ);
cannam@95 1869 Tde = Tdc - Tdd;
cannam@95 1870 TeC = Tdc + Tdd;
cannam@95 1871 }
cannam@95 1872 {
cannam@95 1873 E TdJ, TdK, T3n, T3u;
cannam@95 1874 TdJ = KP707106781 * (T9J - T9M);
cannam@95 1875 TdK = TaT - TaS;
cannam@95 1876 TdL = TdJ + TdK;
cannam@95 1877 TeR = TdK - TdJ;
cannam@95 1878 T3n = T17 - T1a;
cannam@95 1879 T3u = T3q - T3t;
cannam@95 1880 T3v = T3n + T3u;
cannam@95 1881 T5z = T3n - T3u;
cannam@95 1882 }
cannam@95 1883 {
cannam@95 1884 E T4A, T4H, T9G, T9N;
cannam@95 1885 T4A = T1e - T1h;
cannam@95 1886 T4H = T4D - T4G;
cannam@95 1887 T4I = T4A + T4H;
cannam@95 1888 T5O = T4H - T4A;
cannam@95 1889 T9G = T9E - T9F;
cannam@95 1890 T9N = KP707106781 * (T9J + T9M);
cannam@95 1891 T9O = T9G + T9N;
cannam@95 1892 TbM = T9G - T9N;
cannam@95 1893 }
cannam@95 1894 {
cannam@95 1895 E TaR, TaU, T76, T77;
cannam@95 1896 TaR = KP707106781 * (TaP - TaQ);
cannam@95 1897 TaU = TaS + TaT;
cannam@95 1898 TaV = TaR + TaU;
cannam@95 1899 Tc1 = TaU - TaR;
cannam@95 1900 T76 = T4G + T4D;
cannam@95 1901 T77 = T3t + T3q;
cannam@95 1902 T78 = T76 - T77;
cannam@95 1903 T7Z = T77 + T76;
cannam@95 1904 }
cannam@95 1905 }
cannam@95 1906 {
cannam@95 1907 E TB, T90, T3e, T9y, TE, T9x, T3h, T91, TL, T9v, T2I, T98, TI, T9u, T2L;
cannam@95 1908 E T95;
cannam@95 1909 {
cannam@95 1910 E Tz, TA, T3f, T3g;
cannam@95 1911 Tz = Cr[WS(csr, 2)];
cannam@95 1912 TA = Cr[WS(csr, 62)];
cannam@95 1913 TB = Tz + TA;
cannam@95 1914 T90 = Tz - TA;
cannam@95 1915 {
cannam@95 1916 E T3c, T3d, TC, TD;
cannam@95 1917 T3c = Ci[WS(csi, 2)];
cannam@95 1918 T3d = Ci[WS(csi, 62)];
cannam@95 1919 T3e = T3c - T3d;
cannam@95 1920 T9y = T3c + T3d;
cannam@95 1921 TC = Cr[WS(csr, 34)];
cannam@95 1922 TD = Cr[WS(csr, 30)];
cannam@95 1923 TE = TC + TD;
cannam@95 1924 T9x = TC - TD;
cannam@95 1925 }
cannam@95 1926 T3f = Ci[WS(csi, 34)];
cannam@95 1927 T3g = Ci[WS(csi, 30)];
cannam@95 1928 T3h = T3f - T3g;
cannam@95 1929 T91 = T3f + T3g;
cannam@95 1930 {
cannam@95 1931 E TJ, TK, T96, T2G, T2H, T97;
cannam@95 1932 TJ = Cr[WS(csr, 14)];
cannam@95 1933 TK = Cr[WS(csr, 50)];
cannam@95 1934 T96 = TJ - TK;
cannam@95 1935 T2G = Ci[WS(csi, 50)];
cannam@95 1936 T2H = Ci[WS(csi, 14)];
cannam@95 1937 T97 = T2H + T2G;
cannam@95 1938 TL = TJ + TK;
cannam@95 1939 T9v = T96 + T97;
cannam@95 1940 T2I = T2G - T2H;
cannam@95 1941 T98 = T96 - T97;
cannam@95 1942 }
cannam@95 1943 {
cannam@95 1944 E TG, TH, T93, T2J, T2K, T94;
cannam@95 1945 TG = Cr[WS(csr, 18)];
cannam@95 1946 TH = Cr[WS(csr, 46)];
cannam@95 1947 T93 = TG - TH;
cannam@95 1948 T2J = Ci[WS(csi, 18)];
cannam@95 1949 T2K = Ci[WS(csi, 46)];
cannam@95 1950 T94 = T2J + T2K;
cannam@95 1951 TI = TG + TH;
cannam@95 1952 T9u = T93 + T94;
cannam@95 1953 T2L = T2J - T2K;
cannam@95 1954 T95 = T93 - T94;
cannam@95 1955 }
cannam@95 1956 }
cannam@95 1957 {
cannam@95 1958 E TF, TM, TcS, TcT;
cannam@95 1959 TF = TB + TE;
cannam@95 1960 TM = TI + TL;
cannam@95 1961 TN = TF + TM;
cannam@95 1962 T6z = TF - TM;
cannam@95 1963 TcS = T90 + T91;
cannam@95 1964 TcT = KP707106781 * (T9u + T9v);
cannam@95 1965 TcU = TcS - TcT;
cannam@95 1966 Teu = TcS + TcT;
cannam@95 1967 }
cannam@95 1968 {
cannam@95 1969 E Td6, Td7, T2F, T2M;
cannam@95 1970 Td6 = KP707106781 * (T95 - T98);
cannam@95 1971 Td7 = T9y - T9x;
cannam@95 1972 Td8 = Td6 + Td7;
cannam@95 1973 Tey = Td7 - Td6;
cannam@95 1974 T2F = TB - TE;
cannam@95 1975 T2M = T2I - T2L;
cannam@95 1976 T2N = T2F + T2M;
cannam@95 1977 T5r = T2F - T2M;
cannam@95 1978 }
cannam@95 1979 {
cannam@95 1980 E T3b, T3i, T92, T99;
cannam@95 1981 T3b = TI - TL;
cannam@95 1982 T3i = T3e - T3h;
cannam@95 1983 T3j = T3b + T3i;
cannam@95 1984 T5v = T3i - T3b;
cannam@95 1985 T92 = T90 - T91;
cannam@95 1986 T99 = KP707106781 * (T95 + T98);
cannam@95 1987 T9a = T92 + T99;
cannam@95 1988 TbE = T92 - T99;
cannam@95 1989 }
cannam@95 1990 {
cannam@95 1991 E T9w, T9z, T6F, T6G;
cannam@95 1992 T9w = KP707106781 * (T9u - T9v);
cannam@95 1993 T9z = T9x + T9y;
cannam@95 1994 T9A = T9w + T9z;
cannam@95 1995 TbI = T9z - T9w;
cannam@95 1996 T6F = T3h + T3e;
cannam@95 1997 T6G = T2L + T2I;
cannam@95 1998 T6H = T6F - T6G;
cannam@95 1999 T7O = T6G + T6F;
cannam@95 2000 }
cannam@95 2001 }
cannam@95 2002 {
cannam@95 2003 E T1G, Taj, T3Q, Ta5, T46, Tak, T6R, Ta6, T1N, Tag, Tah, T3X, T3Z, Taa, Tad;
cannam@95 2004 E T6S, Tdn, Tdo;
cannam@95 2005 {
cannam@95 2006 E T1A, T1B, T1C, T1D, T1E, T1F;
cannam@95 2007 T1A = Cr[WS(csr, 5)];
cannam@95 2008 T1B = Cr[WS(csr, 59)];
cannam@95 2009 T1C = T1A + T1B;
cannam@95 2010 T1D = Cr[WS(csr, 37)];
cannam@95 2011 T1E = Cr[WS(csr, 27)];
cannam@95 2012 T1F = T1D + T1E;
cannam@95 2013 T1G = T1C + T1F;
cannam@95 2014 Taj = T1D - T1E;
cannam@95 2015 T3Q = T1C - T1F;
cannam@95 2016 Ta5 = T1A - T1B;
cannam@95 2017 }
cannam@95 2018 {
cannam@95 2019 E T40, T41, T42, T43, T44, T45;
cannam@95 2020 T40 = Ci[WS(csi, 5)];
cannam@95 2021 T41 = Ci[WS(csi, 59)];
cannam@95 2022 T42 = T40 - T41;
cannam@95 2023 T43 = Ci[WS(csi, 37)];
cannam@95 2024 T44 = Ci[WS(csi, 27)];
cannam@95 2025 T45 = T43 - T44;
cannam@95 2026 T46 = T42 - T45;
cannam@95 2027 Tak = T40 + T41;
cannam@95 2028 T6R = T45 + T42;
cannam@95 2029 Ta6 = T43 + T44;
cannam@95 2030 }
cannam@95 2031 {
cannam@95 2032 E T1J, Ta8, T3W, Ta9, T1M, Tab, T3T, Tac;
cannam@95 2033 {
cannam@95 2034 E T1H, T1I, T3U, T3V;
cannam@95 2035 T1H = Cr[WS(csr, 21)];
cannam@95 2036 T1I = Cr[WS(csr, 43)];
cannam@95 2037 T1J = T1H + T1I;
cannam@95 2038 Ta8 = T1H - T1I;
cannam@95 2039 T3U = Ci[WS(csi, 21)];
cannam@95 2040 T3V = Ci[WS(csi, 43)];
cannam@95 2041 T3W = T3U - T3V;
cannam@95 2042 Ta9 = T3U + T3V;
cannam@95 2043 }
cannam@95 2044 {
cannam@95 2045 E T1K, T1L, T3R, T3S;
cannam@95 2046 T1K = Cr[WS(csr, 11)];
cannam@95 2047 T1L = Cr[WS(csr, 53)];
cannam@95 2048 T1M = T1K + T1L;
cannam@95 2049 Tab = T1K - T1L;
cannam@95 2050 T3R = Ci[WS(csi, 53)];
cannam@95 2051 T3S = Ci[WS(csi, 11)];
cannam@95 2052 T3T = T3R - T3S;
cannam@95 2053 Tac = T3S + T3R;
cannam@95 2054 }
cannam@95 2055 T1N = T1J + T1M;
cannam@95 2056 Tag = Ta8 + Ta9;
cannam@95 2057 Tah = Tab + Tac;
cannam@95 2058 T3X = T3T - T3W;
cannam@95 2059 T3Z = T1J - T1M;
cannam@95 2060 Taa = Ta8 - Ta9;
cannam@95 2061 Tad = Tab - Tac;
cannam@95 2062 T6S = T3W + T3T;
cannam@95 2063 }
cannam@95 2064 T1O = T1G + T1N;
cannam@95 2065 T7V = T6S + T6R;
cannam@95 2066 {
cannam@95 2067 E T3Y, T47, Tdq, Tdr;
cannam@95 2068 T3Y = T3Q + T3X;
cannam@95 2069 T47 = T3Z + T46;
cannam@95 2070 T48 = FNMS(KP382683432, T47, KP923879532 * T3Y);
cannam@95 2071 T4u = FMA(KP382683432, T3Y, KP923879532 * T47);
cannam@95 2072 Tdq = KP707106781 * (Taa - Tad);
cannam@95 2073 Tdr = Tak - Taj;
cannam@95 2074 Tds = Tdq + Tdr;
cannam@95 2075 TeG = Tdr - Tdq;
cannam@95 2076 }
cannam@95 2077 {
cannam@95 2078 E T5C, T5D, Ta7, Tae;
cannam@95 2079 T5C = T3Q - T3X;
cannam@95 2080 T5D = T46 - T3Z;
cannam@95 2081 T5E = FNMS(KP923879532, T5D, KP382683432 * T5C);
cannam@95 2082 T5K = FMA(KP923879532, T5C, KP382683432 * T5D);
cannam@95 2083 Ta7 = Ta5 - Ta6;
cannam@95 2084 Tae = KP707106781 * (Taa + Tad);
cannam@95 2085 Taf = Ta7 + Tae;
cannam@95 2086 TbP = Ta7 - Tae;
cannam@95 2087 }
cannam@95 2088 Tdn = Ta5 + Ta6;
cannam@95 2089 Tdo = KP707106781 * (Tag + Tah);
cannam@95 2090 Tdp = Tdn - Tdo;
cannam@95 2091 TeF = Tdn + Tdo;
cannam@95 2092 {
cannam@95 2093 E T6Q, T6T, Tai, Tal;
cannam@95 2094 T6Q = T1G - T1N;
cannam@95 2095 T6T = T6R - T6S;
cannam@95 2096 T6U = T6Q - T6T;
cannam@95 2097 T72 = T6Q + T6T;
cannam@95 2098 Tai = KP707106781 * (Tag - Tah);
cannam@95 2099 Tal = Taj + Tak;
cannam@95 2100 Tam = Tai + Tal;
cannam@95 2101 TbQ = Tal - Tai;
cannam@95 2102 }
cannam@95 2103 }
cannam@95 2104 {
cannam@95 2105 E T1V, TaC, T49, Tao, T4p, TaD, T6W, Tap, T22, Taz, TaA, T4g, T4i, Tat, Taw;
cannam@95 2106 E T6X, Tdu, Tdv;
cannam@95 2107 {
cannam@95 2108 E T1P, T1Q, T1R, T1S, T1T, T1U;
cannam@95 2109 T1P = Cr[WS(csr, 3)];
cannam@95 2110 T1Q = Cr[WS(csr, 61)];
cannam@95 2111 T1R = T1P + T1Q;
cannam@95 2112 T1S = Cr[WS(csr, 29)];
cannam@95 2113 T1T = Cr[WS(csr, 35)];
cannam@95 2114 T1U = T1S + T1T;
cannam@95 2115 T1V = T1R + T1U;
cannam@95 2116 TaC = T1S - T1T;
cannam@95 2117 T49 = T1R - T1U;
cannam@95 2118 Tao = T1P - T1Q;
cannam@95 2119 }
cannam@95 2120 {
cannam@95 2121 E T4j, T4k, T4l, T4m, T4n, T4o;
cannam@95 2122 T4j = Ci[WS(csi, 61)];
cannam@95 2123 T4k = Ci[WS(csi, 3)];
cannam@95 2124 T4l = T4j - T4k;
cannam@95 2125 T4m = Ci[WS(csi, 29)];
cannam@95 2126 T4n = Ci[WS(csi, 35)];
cannam@95 2127 T4o = T4m - T4n;
cannam@95 2128 T4p = T4l - T4o;
cannam@95 2129 TaD = T4k + T4j;
cannam@95 2130 T6W = T4o + T4l;
cannam@95 2131 Tap = T4m + T4n;
cannam@95 2132 }
cannam@95 2133 {
cannam@95 2134 E T1Y, Tar, T4f, Tas, T21, Tau, T4c, Tav;
cannam@95 2135 {
cannam@95 2136 E T1W, T1X, T4d, T4e;
cannam@95 2137 T1W = Cr[WS(csr, 13)];
cannam@95 2138 T1X = Cr[WS(csr, 51)];
cannam@95 2139 T1Y = T1W + T1X;
cannam@95 2140 Tar = T1W - T1X;
cannam@95 2141 T4d = Ci[WS(csi, 13)];
cannam@95 2142 T4e = Ci[WS(csi, 51)];
cannam@95 2143 T4f = T4d - T4e;
cannam@95 2144 Tas = T4d + T4e;
cannam@95 2145 }
cannam@95 2146 {
cannam@95 2147 E T1Z, T20, T4a, T4b;
cannam@95 2148 T1Z = Cr[WS(csr, 19)];
cannam@95 2149 T20 = Cr[WS(csr, 45)];
cannam@95 2150 T21 = T1Z + T20;
cannam@95 2151 Tau = T1Z - T20;
cannam@95 2152 T4a = Ci[WS(csi, 45)];
cannam@95 2153 T4b = Ci[WS(csi, 19)];
cannam@95 2154 T4c = T4a - T4b;
cannam@95 2155 Tav = T4b + T4a;
cannam@95 2156 }
cannam@95 2157 T22 = T1Y + T21;
cannam@95 2158 Taz = Tar + Tas;
cannam@95 2159 TaA = Tau + Tav;
cannam@95 2160 T4g = T4c - T4f;
cannam@95 2161 T4i = T1Y - T21;
cannam@95 2162 Tat = Tar - Tas;
cannam@95 2163 Taw = Tau - Tav;
cannam@95 2164 T6X = T4f + T4c;
cannam@95 2165 }
cannam@95 2166 T23 = T1V + T22;
cannam@95 2167 T7U = T6X + T6W;
cannam@95 2168 {
cannam@95 2169 E T4h, T4q, Tdx, Tdy;
cannam@95 2170 T4h = T49 + T4g;
cannam@95 2171 T4q = T4i + T4p;
cannam@95 2172 T4r = FMA(KP923879532, T4h, KP382683432 * T4q);
cannam@95 2173 T4v = FNMS(KP382683432, T4h, KP923879532 * T4q);
cannam@95 2174 Tdx = KP707106781 * (Tat - Taw);
cannam@95 2175 Tdy = TaC + TaD;
cannam@95 2176 Tdz = Tdx - Tdy;
cannam@95 2177 TeJ = Tdx + Tdy;
cannam@95 2178 }
cannam@95 2179 {
cannam@95 2180 E T5F, T5G, Taq, Tax;
cannam@95 2181 T5F = T49 - T4g;
cannam@95 2182 T5G = T4p - T4i;
cannam@95 2183 T5H = FMA(KP382683432, T5F, KP923879532 * T5G);
cannam@95 2184 T5L = FNMS(KP923879532, T5F, KP382683432 * T5G);
cannam@95 2185 Taq = Tao - Tap;
cannam@95 2186 Tax = KP707106781 * (Tat + Taw);
cannam@95 2187 Tay = Taq + Tax;
cannam@95 2188 TbS = Taq - Tax;
cannam@95 2189 }
cannam@95 2190 Tdu = Tao + Tap;
cannam@95 2191 Tdv = KP707106781 * (Taz + TaA);
cannam@95 2192 Tdw = Tdu - Tdv;
cannam@95 2193 TeI = Tdu + Tdv;
cannam@95 2194 {
cannam@95 2195 E T6V, T6Y, TaB, TaE;
cannam@95 2196 T6V = T1V - T22;
cannam@95 2197 T6Y = T6W - T6X;
cannam@95 2198 T6Z = T6V + T6Y;
cannam@95 2199 T73 = T6Y - T6V;
cannam@95 2200 TaB = KP707106781 * (Taz - TaA);
cannam@95 2201 TaE = TaC - TaD;
cannam@95 2202 TaF = TaB + TaE;
cannam@95 2203 TbT = TaE - TaB;
cannam@95 2204 }
cannam@95 2205 }
cannam@95 2206 {
cannam@95 2207 E T1m, T3z, T1p, T3C, T3w, T3D, Tdg, Tdf, T9U, T9R, T1t, T3I, T1w, T3L, T3F;
cannam@95 2208 E T3M, Tdj, Tdi, Ta1, T9Y;
cannam@95 2209 {
cannam@95 2210 E T9P, T9T, T9S, T9Q;
cannam@95 2211 {
cannam@95 2212 E T1k, T1l, T3x, T3y;
cannam@95 2213 T1k = Cr[WS(csr, 9)];
cannam@95 2214 T1l = Cr[WS(csr, 55)];
cannam@95 2215 T1m = T1k + T1l;
cannam@95 2216 T9P = T1k - T1l;
cannam@95 2217 T3x = Ci[WS(csi, 9)];
cannam@95 2218 T3y = Ci[WS(csi, 55)];
cannam@95 2219 T3z = T3x - T3y;
cannam@95 2220 T9T = T3x + T3y;
cannam@95 2221 }
cannam@95 2222 {
cannam@95 2223 E T1n, T1o, T3A, T3B;
cannam@95 2224 T1n = Cr[WS(csr, 41)];
cannam@95 2225 T1o = Cr[WS(csr, 23)];
cannam@95 2226 T1p = T1n + T1o;
cannam@95 2227 T9S = T1n - T1o;
cannam@95 2228 T3A = Ci[WS(csi, 41)];
cannam@95 2229 T3B = Ci[WS(csi, 23)];
cannam@95 2230 T3C = T3A - T3B;
cannam@95 2231 T9Q = T3A + T3B;
cannam@95 2232 }
cannam@95 2233 T3w = T1m - T1p;
cannam@95 2234 T3D = T3z - T3C;
cannam@95 2235 Tdg = T9T - T9S;
cannam@95 2236 Tdf = T9P + T9Q;
cannam@95 2237 T9U = T9S + T9T;
cannam@95 2238 T9R = T9P - T9Q;
cannam@95 2239 }
cannam@95 2240 {
cannam@95 2241 E T9W, Ta0, T9Z, T9X;
cannam@95 2242 {
cannam@95 2243 E T1r, T1s, T3G, T3H;
cannam@95 2244 T1r = Cr[WS(csr, 7)];
cannam@95 2245 T1s = Cr[WS(csr, 57)];
cannam@95 2246 T1t = T1r + T1s;
cannam@95 2247 T9W = T1r - T1s;
cannam@95 2248 T3G = Ci[WS(csi, 57)];
cannam@95 2249 T3H = Ci[WS(csi, 7)];
cannam@95 2250 T3I = T3G - T3H;
cannam@95 2251 Ta0 = T3H + T3G;
cannam@95 2252 }
cannam@95 2253 {
cannam@95 2254 E T1u, T1v, T3J, T3K;
cannam@95 2255 T1u = Cr[WS(csr, 25)];
cannam@95 2256 T1v = Cr[WS(csr, 39)];
cannam@95 2257 T1w = T1u + T1v;
cannam@95 2258 T9Z = T1u - T1v;
cannam@95 2259 T3J = Ci[WS(csi, 25)];
cannam@95 2260 T3K = Ci[WS(csi, 39)];
cannam@95 2261 T3L = T3J - T3K;
cannam@95 2262 T9X = T3J + T3K;
cannam@95 2263 }
cannam@95 2264 T3F = T1t - T1w;
cannam@95 2265 T3M = T3I - T3L;
cannam@95 2266 Tdj = T9Z + Ta0;
cannam@95 2267 Tdi = T9W + T9X;
cannam@95 2268 Ta1 = T9Z - Ta0;
cannam@95 2269 T9Y = T9W - T9X;
cannam@95 2270 }
cannam@95 2271 {
cannam@95 2272 E T1q, T1x, Tdh, Tdk;
cannam@95 2273 T1q = T1m + T1p;
cannam@95 2274 T1x = T1t + T1w;
cannam@95 2275 T1y = T1q + T1x;
cannam@95 2276 T75 = T1q - T1x;
cannam@95 2277 Tdh = FNMS(KP923879532, Tdg, KP382683432 * Tdf);
cannam@95 2278 Tdk = FNMS(KP923879532, Tdj, KP382683432 * Tdi);
cannam@95 2279 Tdl = Tdh + Tdk;
cannam@95 2280 TeQ = Tdh - Tdk;
cannam@95 2281 }
cannam@95 2282 {
cannam@95 2283 E TdG, TdH, T3E, T3N;
cannam@95 2284 TdG = FMA(KP923879532, Tdf, KP382683432 * Tdg);
cannam@95 2285 TdH = FMA(KP923879532, Tdi, KP382683432 * Tdj);
cannam@95 2286 TdI = TdG - TdH;
cannam@95 2287 TeD = TdG + TdH;
cannam@95 2288 T3E = T3w - T3D;
cannam@95 2289 T3N = T3F + T3M;
cannam@95 2290 T3O = KP707106781 * (T3E + T3N);
cannam@95 2291 T5N = KP707106781 * (T3E - T3N);
cannam@95 2292 }
cannam@95 2293 {
cannam@95 2294 E T4x, T4y, T9V, Ta2;
cannam@95 2295 T4x = T3w + T3D;
cannam@95 2296 T4y = T3M - T3F;
cannam@95 2297 T4z = KP707106781 * (T4x + T4y);
cannam@95 2298 T5A = KP707106781 * (T4y - T4x);
cannam@95 2299 T9V = FNMS(KP382683432, T9U, KP923879532 * T9R);
cannam@95 2300 Ta2 = FMA(KP923879532, T9Y, KP382683432 * Ta1);
cannam@95 2301 Ta3 = T9V + Ta2;
cannam@95 2302 Tc0 = T9V - Ta2;
cannam@95 2303 }
cannam@95 2304 {
cannam@95 2305 E TaM, TaN, T6M, T6N;
cannam@95 2306 TaM = FMA(KP382683432, T9R, KP923879532 * T9U);
cannam@95 2307 TaN = FNMS(KP382683432, T9Y, KP923879532 * Ta1);
cannam@95 2308 TaO = TaM + TaN;
cannam@95 2309 TbN = TaN - TaM;
cannam@95 2310 T6M = T3L + T3I;
cannam@95 2311 T6N = T3C + T3z;
cannam@95 2312 T6O = T6M - T6N;
cannam@95 2313 T80 = T6N + T6M;
cannam@95 2314 }
cannam@95 2315 }
cannam@95 2316 {
cannam@95 2317 E TQ, T2R, TT, T2U, T2O, T2V, TcW, TcV, T9g, T9d, TX, T30, T10, T33, T2X;
cannam@95 2318 E T34, TcZ, TcY, T9n, T9k;
cannam@95 2319 {
cannam@95 2320 E T9b, T9f, T9e, T9c;
cannam@95 2321 {
cannam@95 2322 E TO, TP, T2P, T2Q;
cannam@95 2323 TO = Cr[WS(csr, 10)];
cannam@95 2324 TP = Cr[WS(csr, 54)];
cannam@95 2325 TQ = TO + TP;
cannam@95 2326 T9b = TO - TP;
cannam@95 2327 T2P = Ci[WS(csi, 10)];
cannam@95 2328 T2Q = Ci[WS(csi, 54)];
cannam@95 2329 T2R = T2P - T2Q;
cannam@95 2330 T9f = T2P + T2Q;
cannam@95 2331 }
cannam@95 2332 {
cannam@95 2333 E TR, TS, T2S, T2T;
cannam@95 2334 TR = Cr[WS(csr, 42)];
cannam@95 2335 TS = Cr[WS(csr, 22)];
cannam@95 2336 TT = TR + TS;
cannam@95 2337 T9e = TR - TS;
cannam@95 2338 T2S = Ci[WS(csi, 42)];
cannam@95 2339 T2T = Ci[WS(csi, 22)];
cannam@95 2340 T2U = T2S - T2T;
cannam@95 2341 T9c = T2S + T2T;
cannam@95 2342 }
cannam@95 2343 T2O = TQ - TT;
cannam@95 2344 T2V = T2R - T2U;
cannam@95 2345 TcW = T9f - T9e;
cannam@95 2346 TcV = T9b + T9c;
cannam@95 2347 T9g = T9e + T9f;
cannam@95 2348 T9d = T9b - T9c;
cannam@95 2349 }
cannam@95 2350 {
cannam@95 2351 E T9i, T9m, T9l, T9j;
cannam@95 2352 {
cannam@95 2353 E TV, TW, T2Y, T2Z;
cannam@95 2354 TV = Cr[WS(csr, 6)];
cannam@95 2355 TW = Cr[WS(csr, 58)];
cannam@95 2356 TX = TV + TW;
cannam@95 2357 T9i = TV - TW;
cannam@95 2358 T2Y = Ci[WS(csi, 58)];
cannam@95 2359 T2Z = Ci[WS(csi, 6)];
cannam@95 2360 T30 = T2Y - T2Z;
cannam@95 2361 T9m = T2Z + T2Y;
cannam@95 2362 }
cannam@95 2363 {
cannam@95 2364 E TY, TZ, T31, T32;
cannam@95 2365 TY = Cr[WS(csr, 26)];
cannam@95 2366 TZ = Cr[WS(csr, 38)];
cannam@95 2367 T10 = TY + TZ;
cannam@95 2368 T9l = TY - TZ;
cannam@95 2369 T31 = Ci[WS(csi, 26)];
cannam@95 2370 T32 = Ci[WS(csi, 38)];
cannam@95 2371 T33 = T31 - T32;
cannam@95 2372 T9j = T31 + T32;
cannam@95 2373 }
cannam@95 2374 T2X = TX - T10;
cannam@95 2375 T34 = T30 - T33;
cannam@95 2376 TcZ = T9l + T9m;
cannam@95 2377 TcY = T9i + T9j;
cannam@95 2378 T9n = T9l - T9m;
cannam@95 2379 T9k = T9i - T9j;
cannam@95 2380 }
cannam@95 2381 {
cannam@95 2382 E TU, T11, TcX, Td0;
cannam@95 2383 TU = TQ + TT;
cannam@95 2384 T11 = TX + T10;
cannam@95 2385 T12 = TU + T11;
cannam@95 2386 T6E = TU - T11;
cannam@95 2387 TcX = FNMS(KP923879532, TcW, KP382683432 * TcV);
cannam@95 2388 Td0 = FNMS(KP923879532, TcZ, KP382683432 * TcY);
cannam@95 2389 Td1 = TcX + Td0;
cannam@95 2390 Tex = TcX - Td0;
cannam@95 2391 }
cannam@95 2392 {
cannam@95 2393 E Td3, Td4, T2W, T35;
cannam@95 2394 Td3 = FMA(KP923879532, TcV, KP382683432 * TcW);
cannam@95 2395 Td4 = FMA(KP923879532, TcY, KP382683432 * TcZ);
cannam@95 2396 Td5 = Td3 - Td4;
cannam@95 2397 Tev = Td3 + Td4;
cannam@95 2398 T2W = T2O - T2V;
cannam@95 2399 T35 = T2X + T34;
cannam@95 2400 T36 = KP707106781 * (T2W + T35);
cannam@95 2401 T5u = KP707106781 * (T2W - T35);
cannam@95 2402 }
cannam@95 2403 {
cannam@95 2404 E T38, T39, T9h, T9o;
cannam@95 2405 T38 = T2O + T2V;
cannam@95 2406 T39 = T34 - T2X;
cannam@95 2407 T3a = KP707106781 * (T38 + T39);
cannam@95 2408 T5s = KP707106781 * (T39 - T38);
cannam@95 2409 T9h = FNMS(KP382683432, T9g, KP923879532 * T9d);
cannam@95 2410 T9o = FMA(KP923879532, T9k, KP382683432 * T9n);
cannam@95 2411 T9p = T9h + T9o;
cannam@95 2412 TbH = T9h - T9o;
cannam@95 2413 }
cannam@95 2414 {
cannam@95 2415 E T9r, T9s, T6A, T6B;
cannam@95 2416 T9r = FMA(KP382683432, T9d, KP923879532 * T9g);
cannam@95 2417 T9s = FNMS(KP382683432, T9k, KP923879532 * T9n);
cannam@95 2418 T9t = T9r + T9s;
cannam@95 2419 TbF = T9s - T9r;
cannam@95 2420 T6A = T33 + T30;
cannam@95 2421 T6B = T2U + T2R;
cannam@95 2422 T6C = T6A - T6B;
cannam@95 2423 T7P = T6B + T6A;
cannam@95 2424 }
cannam@95 2425 }
cannam@95 2426 {
cannam@95 2427 E T13, T8f, Ty, T8e, T25, T8h, T8k, T8p, Ti, T14, T8o;
cannam@95 2428 T13 = KP2_000000000 * (TN + T12);
cannam@95 2429 T8f = KP2_000000000 * (T7P + T7O);
cannam@95 2430 Ti = Ta + Th;
cannam@95 2431 Ty = Ti + Tx;
cannam@95 2432 T8e = Ti - Tx;
cannam@95 2433 {
cannam@95 2434 E T1z, T24, T8i, T8j;
cannam@95 2435 T1z = T1j + T1y;
cannam@95 2436 T24 = T1O + T23;
cannam@95 2437 T25 = KP2_000000000 * (T1z + T24);
cannam@95 2438 T8h = T1z - T24;
cannam@95 2439 T8i = T80 + T7Z;
cannam@95 2440 T8j = T7V + T7U;
cannam@95 2441 T8k = T8i - T8j;
cannam@95 2442 T8p = KP2_000000000 * (T8j + T8i);
cannam@95 2443 }
cannam@95 2444 T14 = Ty + T13;
cannam@95 2445 R0[WS(rs, 32)] = T14 - T25;
cannam@95 2446 R0[0] = T14 + T25;
cannam@95 2447 T8o = Ty - T13;
cannam@95 2448 R0[WS(rs, 16)] = T8o - T8p;
cannam@95 2449 R0[WS(rs, 48)] = T8o + T8p;
cannam@95 2450 {
cannam@95 2451 E T8g, T8l, T8m, T8n;
cannam@95 2452 T8g = T8e - T8f;
cannam@95 2453 T8l = KP1_414213562 * (T8h - T8k);
cannam@95 2454 R0[WS(rs, 40)] = T8g - T8l;
cannam@95 2455 R0[WS(rs, 8)] = T8g + T8l;
cannam@95 2456 T8m = T8e + T8f;
cannam@95 2457 T8n = KP1_414213562 * (T8h + T8k);
cannam@95 2458 R0[WS(rs, 24)] = T8m - T8n;
cannam@95 2459 R0[WS(rs, 56)] = T8m + T8n;
cannam@95 2460 }
cannam@95 2461 }
cannam@95 2462 {
cannam@95 2463 E T7M, T86, T82, T8a, T7R, T87, T7X, T89, T7K, T7Y, T81;
cannam@95 2464 T7K = Ta - Th;
cannam@95 2465 T7M = T7K - T7L;
cannam@95 2466 T86 = T7K + T7L;
cannam@95 2467 T7Y = T1O - T23;
cannam@95 2468 T81 = T7Z - T80;
cannam@95 2469 T82 = T7Y + T81;
cannam@95 2470 T8a = T81 - T7Y;
cannam@95 2471 {
cannam@95 2472 E T7N, T7Q, T7T, T7W;
cannam@95 2473 T7N = TN - T12;
cannam@95 2474 T7Q = T7O - T7P;
cannam@95 2475 T7R = KP1_414213562 * (T7N - T7Q);
cannam@95 2476 T87 = KP1_414213562 * (T7N + T7Q);
cannam@95 2477 T7T = T1j - T1y;
cannam@95 2478 T7W = T7U - T7V;
cannam@95 2479 T7X = T7T + T7W;
cannam@95 2480 T89 = T7T - T7W;
cannam@95 2481 }
cannam@95 2482 {
cannam@95 2483 E T7S, T83, T8c, T8d;
cannam@95 2484 T7S = T7M + T7R;
cannam@95 2485 T83 = FNMS(KP765366864, T82, KP1_847759065 * T7X);
cannam@95 2486 R0[WS(rs, 36)] = T7S - T83;
cannam@95 2487 R0[WS(rs, 4)] = T7S + T83;
cannam@95 2488 T8c = T86 + T87;
cannam@95 2489 T8d = FMA(KP1_847759065, T89, KP765366864 * T8a);
cannam@95 2490 R0[WS(rs, 28)] = T8c - T8d;
cannam@95 2491 R0[WS(rs, 60)] = T8c + T8d;
cannam@95 2492 }
cannam@95 2493 {
cannam@95 2494 E T84, T85, T88, T8b;
cannam@95 2495 T84 = T7M - T7R;
cannam@95 2496 T85 = FMA(KP765366864, T7X, KP1_847759065 * T82);
cannam@95 2497 R0[WS(rs, 20)] = T84 - T85;
cannam@95 2498 R0[WS(rs, 52)] = T84 + T85;
cannam@95 2499 T88 = T86 - T87;
cannam@95 2500 T8b = FNMS(KP1_847759065, T8a, KP765366864 * T89);
cannam@95 2501 R0[WS(rs, 44)] = T88 - T8b;
cannam@95 2502 R0[WS(rs, 12)] = T88 + T8b;
cannam@95 2503 }
cannam@95 2504 }
cannam@95 2505 {
cannam@95 2506 E T2E, T4O, T4K, T4S, T3l, T4P, T4t, T4R;
cannam@95 2507 {
cannam@95 2508 E T2k, T2D, T4w, T4J;
cannam@95 2509 T2k = T2a + T2j;
cannam@95 2510 T2D = FNMS(KP765366864, T2C, KP1_847759065 * T2t);
cannam@95 2511 T2E = T2k + T2D;
cannam@95 2512 T4O = T2k - T2D;
cannam@95 2513 T4w = T4u + T4v;
cannam@95 2514 T4J = T4z + T4I;
cannam@95 2515 T4K = T4w + T4J;
cannam@95 2516 T4S = T4J - T4w;
cannam@95 2517 }
cannam@95 2518 {
cannam@95 2519 E T37, T3k, T3P, T4s;
cannam@95 2520 T37 = T2N + T36;
cannam@95 2521 T3k = T3a + T3j;
cannam@95 2522 T3l = FNMS(KP390180644, T3k, KP1_961570560 * T37);
cannam@95 2523 T4P = FMA(KP390180644, T37, KP1_961570560 * T3k);
cannam@95 2524 T3P = T3v + T3O;
cannam@95 2525 T4s = T48 + T4r;
cannam@95 2526 T4t = T3P + T4s;
cannam@95 2527 T4R = T3P - T4s;
cannam@95 2528 }
cannam@95 2529 {
cannam@95 2530 E T3m, T4L, T4U, T4V;
cannam@95 2531 T3m = T2E + T3l;
cannam@95 2532 T4L = FNMS(KP196034280, T4K, KP1_990369453 * T4t);
cannam@95 2533 R0[WS(rs, 33)] = T3m - T4L;
cannam@95 2534 R0[WS(rs, 1)] = T3m + T4L;
cannam@95 2535 T4U = T4O + T4P;
cannam@95 2536 T4V = FMA(KP1_546020906, T4R, KP1_268786568 * T4S);
cannam@95 2537 R0[WS(rs, 25)] = T4U - T4V;
cannam@95 2538 R0[WS(rs, 57)] = T4U + T4V;
cannam@95 2539 }
cannam@95 2540 {
cannam@95 2541 E T4M, T4N, T4Q, T4T;
cannam@95 2542 T4M = T2E - T3l;
cannam@95 2543 T4N = FMA(KP196034280, T4t, KP1_990369453 * T4K);
cannam@95 2544 R0[WS(rs, 17)] = T4M - T4N;
cannam@95 2545 R0[WS(rs, 49)] = T4M + T4N;
cannam@95 2546 T4Q = T4O - T4P;
cannam@95 2547 T4T = FNMS(KP1_546020906, T4S, KP1_268786568 * T4R);
cannam@95 2548 R0[WS(rs, 41)] = T4Q - T4T;
cannam@95 2549 R0[WS(rs, 9)] = T4Q + T4T;
cannam@95 2550 }
cannam@95 2551 }
cannam@95 2552 {
cannam@95 2553 E T6y, T7e, T7a, T7i, T6J, T7f, T71, T7h;
cannam@95 2554 {
cannam@95 2555 E T6s, T6x, T74, T79;
cannam@95 2556 T6s = T6q - T6r;
cannam@95 2557 T6x = KP1_414213562 * (T6t - T6w);
cannam@95 2558 T6y = T6s + T6x;
cannam@95 2559 T7e = T6s - T6x;
cannam@95 2560 T74 = KP707106781 * (T72 + T73);
cannam@95 2561 T79 = T75 + T78;
cannam@95 2562 T7a = T74 + T79;
cannam@95 2563 T7i = T79 - T74;
cannam@95 2564 }
cannam@95 2565 {
cannam@95 2566 E T6D, T6I, T6P, T70;
cannam@95 2567 T6D = T6z + T6C;
cannam@95 2568 T6I = T6E + T6H;
cannam@95 2569 T6J = FNMS(KP765366864, T6I, KP1_847759065 * T6D);
cannam@95 2570 T7f = FMA(KP765366864, T6D, KP1_847759065 * T6I);
cannam@95 2571 T6P = T6L + T6O;
cannam@95 2572 T70 = KP707106781 * (T6U + T6Z);
cannam@95 2573 T71 = T6P + T70;
cannam@95 2574 T7h = T6P - T70;
cannam@95 2575 }
cannam@95 2576 {
cannam@95 2577 E T6K, T7b, T7k, T7l;
cannam@95 2578 T6K = T6y + T6J;
cannam@95 2579 T7b = FNMS(KP390180644, T7a, KP1_961570560 * T71);
cannam@95 2580 R0[WS(rs, 34)] = T6K - T7b;
cannam@95 2581 R0[WS(rs, 2)] = T6K + T7b;
cannam@95 2582 T7k = T7e + T7f;
cannam@95 2583 T7l = FMA(KP1_662939224, T7h, KP1_111140466 * T7i);
cannam@95 2584 R0[WS(rs, 26)] = T7k - T7l;
cannam@95 2585 R0[WS(rs, 58)] = T7k + T7l;
cannam@95 2586 }
cannam@95 2587 {
cannam@95 2588 E T7c, T7d, T7g, T7j;
cannam@95 2589 T7c = T6y - T6J;
cannam@95 2590 T7d = FMA(KP390180644, T71, KP1_961570560 * T7a);
cannam@95 2591 R0[WS(rs, 18)] = T7c - T7d;
cannam@95 2592 R0[WS(rs, 50)] = T7c + T7d;
cannam@95 2593 T7g = T7e - T7f;
cannam@95 2594 T7j = FNMS(KP1_662939224, T7i, KP1_111140466 * T7h);
cannam@95 2595 R0[WS(rs, 42)] = T7g - T7j;
cannam@95 2596 R0[WS(rs, 10)] = T7g + T7j;
cannam@95 2597 }
cannam@95 2598 }
cannam@95 2599 {
cannam@95 2600 E T4Y, T5c, T58, T5g, T51, T5d, T55, T5f;
cannam@95 2601 {
cannam@95 2602 E T4W, T4X, T56, T57;
cannam@95 2603 T4W = T2a - T2j;
cannam@95 2604 T4X = FMA(KP765366864, T2t, KP1_847759065 * T2C);
cannam@95 2605 T4Y = T4W - T4X;
cannam@95 2606 T5c = T4W + T4X;
cannam@95 2607 T56 = T48 - T4r;
cannam@95 2608 T57 = T4I - T4z;
cannam@95 2609 T58 = T56 + T57;
cannam@95 2610 T5g = T57 - T56;
cannam@95 2611 }
cannam@95 2612 {
cannam@95 2613 E T4Z, T50, T53, T54;
cannam@95 2614 T4Z = T2N - T36;
cannam@95 2615 T50 = T3j - T3a;
cannam@95 2616 T51 = FNMS(KP1_662939224, T50, KP1_111140466 * T4Z);
cannam@95 2617 T5d = FMA(KP1_662939224, T4Z, KP1_111140466 * T50);
cannam@95 2618 T53 = T3v - T3O;
cannam@95 2619 T54 = T4v - T4u;
cannam@95 2620 T55 = T53 + T54;
cannam@95 2621 T5f = T53 - T54;
cannam@95 2622 }
cannam@95 2623 {
cannam@95 2624 E T52, T59, T5i, T5j;
cannam@95 2625 T52 = T4Y + T51;
cannam@95 2626 T59 = FNMS(KP942793473, T58, KP1_763842528 * T55);
cannam@95 2627 R0[WS(rs, 37)] = T52 - T59;
cannam@95 2628 R0[WS(rs, 5)] = T52 + T59;
cannam@95 2629 T5i = T5c + T5d;
cannam@95 2630 T5j = FMA(KP1_913880671, T5f, KP580569354 * T5g);
cannam@95 2631 R0[WS(rs, 29)] = T5i - T5j;
cannam@95 2632 R0[WS(rs, 61)] = T5i + T5j;
cannam@95 2633 }
cannam@95 2634 {
cannam@95 2635 E T5a, T5b, T5e, T5h;
cannam@95 2636 T5a = T4Y - T51;
cannam@95 2637 T5b = FMA(KP942793473, T55, KP1_763842528 * T58);
cannam@95 2638 R0[WS(rs, 21)] = T5a - T5b;
cannam@95 2639 R0[WS(rs, 53)] = T5a + T5b;
cannam@95 2640 T5e = T5c - T5d;
cannam@95 2641 T5h = FNMS(KP1_913880671, T5g, KP580569354 * T5f);
cannam@95 2642 R0[WS(rs, 45)] = T5e - T5h;
cannam@95 2643 R0[WS(rs, 13)] = T5e + T5h;
cannam@95 2644 }
cannam@95 2645 }
cannam@95 2646 {
cannam@95 2647 E T7o, T7C, T7y, T7G, T7r, T7D, T7v, T7F;
cannam@95 2648 {
cannam@95 2649 E T7m, T7n, T7w, T7x;
cannam@95 2650 T7m = T6q + T6r;
cannam@95 2651 T7n = KP1_414213562 * (T6t + T6w);
cannam@95 2652 T7o = T7m - T7n;
cannam@95 2653 T7C = T7m + T7n;
cannam@95 2654 T7w = KP707106781 * (T6U - T6Z);
cannam@95 2655 T7x = T78 - T75;
cannam@95 2656 T7y = T7w + T7x;
cannam@95 2657 T7G = T7x - T7w;
cannam@95 2658 }
cannam@95 2659 {
cannam@95 2660 E T7p, T7q, T7t, T7u;
cannam@95 2661 T7p = T6z - T6C;
cannam@95 2662 T7q = T6H - T6E;
cannam@95 2663 T7r = FNMS(KP1_847759065, T7q, KP765366864 * T7p);
cannam@95 2664 T7D = FMA(KP1_847759065, T7p, KP765366864 * T7q);
cannam@95 2665 T7t = T6L - T6O;
cannam@95 2666 T7u = KP707106781 * (T73 - T72);
cannam@95 2667 T7v = T7t + T7u;
cannam@95 2668 T7F = T7t - T7u;
cannam@95 2669 }
cannam@95 2670 {
cannam@95 2671 E T7s, T7z, T7I, T7J;
cannam@95 2672 T7s = T7o + T7r;
cannam@95 2673 T7z = FNMS(KP1_111140466, T7y, KP1_662939224 * T7v);
cannam@95 2674 R0[WS(rs, 38)] = T7s - T7z;
cannam@95 2675 R0[WS(rs, 6)] = T7s + T7z;
cannam@95 2676 T7I = T7C + T7D;
cannam@95 2677 T7J = FMA(KP1_961570560, T7F, KP390180644 * T7G);
cannam@95 2678 R0[WS(rs, 30)] = T7I - T7J;
cannam@95 2679 R0[WS(rs, 62)] = T7I + T7J;
cannam@95 2680 }
cannam@95 2681 {
cannam@95 2682 E T7A, T7B, T7E, T7H;
cannam@95 2683 T7A = T7o - T7r;
cannam@95 2684 T7B = FMA(KP1_111140466, T7v, KP1_662939224 * T7y);
cannam@95 2685 R0[WS(rs, 22)] = T7A - T7B;
cannam@95 2686 R0[WS(rs, 54)] = T7A + T7B;
cannam@95 2687 T7E = T7C - T7D;
cannam@95 2688 T7H = FNMS(KP1_961570560, T7G, KP390180644 * T7F);
cannam@95 2689 R0[WS(rs, 46)] = T7E - T7H;
cannam@95 2690 R0[WS(rs, 14)] = T7E + T7H;
cannam@95 2691 }
cannam@95 2692 }
cannam@95 2693 {
cannam@95 2694 E T5q, T5U, T5Q, T5Y, T5x, T5V, T5J, T5X;
cannam@95 2695 {
cannam@95 2696 E T5m, T5p, T5M, T5P;
cannam@95 2697 T5m = T5k - T5l;
cannam@95 2698 T5p = FNMS(KP1_847759065, T5o, KP765366864 * T5n);
cannam@95 2699 T5q = T5m + T5p;
cannam@95 2700 T5U = T5m - T5p;
cannam@95 2701 T5M = T5K + T5L;
cannam@95 2702 T5P = T5N + T5O;
cannam@95 2703 T5Q = T5M + T5P;
cannam@95 2704 T5Y = T5P - T5M;
cannam@95 2705 }
cannam@95 2706 {
cannam@95 2707 E T5t, T5w, T5B, T5I;
cannam@95 2708 T5t = T5r + T5s;
cannam@95 2709 T5w = T5u + T5v;
cannam@95 2710 T5x = FNMS(KP1_111140466, T5w, KP1_662939224 * T5t);
cannam@95 2711 T5V = FMA(KP1_111140466, T5t, KP1_662939224 * T5w);
cannam@95 2712 T5B = T5z + T5A;
cannam@95 2713 T5I = T5E + T5H;
cannam@95 2714 T5J = T5B + T5I;
cannam@95 2715 T5X = T5B - T5I;
cannam@95 2716 }
cannam@95 2717 {
cannam@95 2718 E T5y, T5R, T60, T61;
cannam@95 2719 T5y = T5q + T5x;
cannam@95 2720 T5R = FNMS(KP580569354, T5Q, KP1_913880671 * T5J);
cannam@95 2721 R0[WS(rs, 35)] = T5y - T5R;
cannam@95 2722 R0[WS(rs, 3)] = T5y + T5R;
cannam@95 2723 T60 = T5U + T5V;
cannam@95 2724 T61 = FMA(KP1_763842528, T5X, KP942793473 * T5Y);
cannam@95 2725 R0[WS(rs, 27)] = T60 - T61;
cannam@95 2726 R0[WS(rs, 59)] = T60 + T61;
cannam@95 2727 }
cannam@95 2728 {
cannam@95 2729 E T5S, T5T, T5W, T5Z;
cannam@95 2730 T5S = T5q - T5x;
cannam@95 2731 T5T = FMA(KP580569354, T5J, KP1_913880671 * T5Q);
cannam@95 2732 R0[WS(rs, 19)] = T5S - T5T;
cannam@95 2733 R0[WS(rs, 51)] = T5S + T5T;
cannam@95 2734 T5W = T5U - T5V;
cannam@95 2735 T5Z = FNMS(KP1_763842528, T5Y, KP942793473 * T5X);
cannam@95 2736 R0[WS(rs, 43)] = T5W - T5Z;
cannam@95 2737 R0[WS(rs, 11)] = T5W + T5Z;
cannam@95 2738 }
cannam@95 2739 }
cannam@95 2740 {
cannam@95 2741 E T64, T6i, T6e, T6m, T67, T6j, T6b, T6l;
cannam@95 2742 {
cannam@95 2743 E T62, T63, T6c, T6d;
cannam@95 2744 T62 = T5k + T5l;
cannam@95 2745 T63 = FMA(KP1_847759065, T5n, KP765366864 * T5o);
cannam@95 2746 T64 = T62 - T63;
cannam@95 2747 T6i = T62 + T63;
cannam@95 2748 T6c = T5E - T5H;
cannam@95 2749 T6d = T5O - T5N;
cannam@95 2750 T6e = T6c + T6d;
cannam@95 2751 T6m = T6d - T6c;
cannam@95 2752 }
cannam@95 2753 {
cannam@95 2754 E T65, T66, T69, T6a;
cannam@95 2755 T65 = T5r - T5s;
cannam@95 2756 T66 = T5v - T5u;
cannam@95 2757 T67 = FNMS(KP1_961570560, T66, KP390180644 * T65);
cannam@95 2758 T6j = FMA(KP1_961570560, T65, KP390180644 * T66);
cannam@95 2759 T69 = T5z - T5A;
cannam@95 2760 T6a = T5L - T5K;
cannam@95 2761 T6b = T69 + T6a;
cannam@95 2762 T6l = T69 - T6a;
cannam@95 2763 }
cannam@95 2764 {
cannam@95 2765 E T68, T6f, T6o, T6p;
cannam@95 2766 T68 = T64 + T67;
cannam@95 2767 T6f = FNMS(KP1_268786568, T6e, KP1_546020906 * T6b);
cannam@95 2768 R0[WS(rs, 39)] = T68 - T6f;
cannam@95 2769 R0[WS(rs, 7)] = T68 + T6f;
cannam@95 2770 T6o = T6i + T6j;
cannam@95 2771 T6p = FMA(KP1_990369453, T6l, KP196034280 * T6m);
cannam@95 2772 R0[WS(rs, 31)] = T6o - T6p;
cannam@95 2773 R0[WS(rs, 63)] = T6o + T6p;
cannam@95 2774 }
cannam@95 2775 {
cannam@95 2776 E T6g, T6h, T6k, T6n;
cannam@95 2777 T6g = T64 - T67;
cannam@95 2778 T6h = FMA(KP1_268786568, T6b, KP1_546020906 * T6e);
cannam@95 2779 R0[WS(rs, 23)] = T6g - T6h;
cannam@95 2780 R0[WS(rs, 55)] = T6g + T6h;
cannam@95 2781 T6k = T6i - T6j;
cannam@95 2782 T6n = FNMS(KP1_990369453, T6m, KP196034280 * T6l);
cannam@95 2783 R0[WS(rs, 47)] = T6k - T6n;
cannam@95 2784 R0[WS(rs, 15)] = T6k + T6n;
cannam@95 2785 }
cannam@95 2786 }
cannam@95 2787 {
cannam@95 2788 E T8Z, Tb1, T9C, Tb2, Tbe, Tbq, Tbb, Tbp, TaX, Tbs, Tb5, Tbi, TaI, Tbt, Tb4;
cannam@95 2789 E Tbl;
cannam@95 2790 {
cannam@95 2791 E T8F, T8Y, Tb9, Tba;
cannam@95 2792 T8F = T8x + T8E;
cannam@95 2793 T8Y = FNMS(KP390180644, T8X, KP1_961570560 * T8Q);
cannam@95 2794 T8Z = T8F + T8Y;
cannam@95 2795 Tb1 = T8F - T8Y;
cannam@95 2796 {
cannam@95 2797 E T9q, T9B, Tbc, Tbd;
cannam@95 2798 T9q = T9a + T9p;
cannam@95 2799 T9B = T9t + T9A;
cannam@95 2800 T9C = FNMS(KP196034280, T9B, KP1_990369453 * T9q);
cannam@95 2801 Tb2 = FMA(KP196034280, T9q, KP1_990369453 * T9B);
cannam@95 2802 Tbc = T9a - T9p;
cannam@95 2803 Tbd = T9A - T9t;
cannam@95 2804 Tbe = FNMS(KP1_546020906, Tbd, KP1_268786568 * Tbc);
cannam@95 2805 Tbq = FMA(KP1_546020906, Tbc, KP1_268786568 * Tbd);
cannam@95 2806 }
cannam@95 2807 Tb9 = T8x - T8E;
cannam@95 2808 Tba = FMA(KP390180644, T8Q, KP1_961570560 * T8X);
cannam@95 2809 Tbb = Tb9 - Tba;
cannam@95 2810 Tbp = Tb9 + Tba;
cannam@95 2811 {
cannam@95 2812 E TaW, Tbg, TaL, Tbh, TaJ, TaK;
cannam@95 2813 TaW = TaO + TaV;
cannam@95 2814 Tbg = T9O - Ta3;
cannam@95 2815 TaJ = FMA(KP195090322, Taf, KP980785280 * Tam);
cannam@95 2816 TaK = FNMS(KP195090322, Tay, KP980785280 * TaF);
cannam@95 2817 TaL = TaJ + TaK;
cannam@95 2818 Tbh = TaK - TaJ;
cannam@95 2819 TaX = TaL + TaW;
cannam@95 2820 Tbs = Tbg - Tbh;
cannam@95 2821 Tb5 = TaW - TaL;
cannam@95 2822 Tbi = Tbg + Tbh;
cannam@95 2823 }
cannam@95 2824 {
cannam@95 2825 E Ta4, Tbk, TaH, Tbj, Tan, TaG;
cannam@95 2826 Ta4 = T9O + Ta3;
cannam@95 2827 Tbk = TaV - TaO;
cannam@95 2828 Tan = FNMS(KP195090322, Tam, KP980785280 * Taf);
cannam@95 2829 TaG = FMA(KP980785280, Tay, KP195090322 * TaF);
cannam@95 2830 TaH = Tan + TaG;
cannam@95 2831 Tbj = Tan - TaG;
cannam@95 2832 TaI = Ta4 + TaH;
cannam@95 2833 Tbt = Tbk - Tbj;
cannam@95 2834 Tb4 = Ta4 - TaH;
cannam@95 2835 Tbl = Tbj + Tbk;
cannam@95 2836 }
cannam@95 2837 }
cannam@95 2838 {
cannam@95 2839 E T9D, TaY, Tbr, Tbu;
cannam@95 2840 T9D = T8Z + T9C;
cannam@95 2841 TaY = FNMS(KP098135348, TaX, KP1_997590912 * TaI);
cannam@95 2842 R1[WS(rs, 32)] = T9D - TaY;
cannam@95 2843 R1[0] = T9D + TaY;
cannam@95 2844 Tbr = Tbp - Tbq;
cannam@95 2845 Tbu = FNMS(KP1_883088130, Tbt, KP673779706 * Tbs);
cannam@95 2846 R1[WS(rs, 44)] = Tbr - Tbu;
cannam@95 2847 R1[WS(rs, 12)] = Tbr + Tbu;
cannam@95 2848 }
cannam@95 2849 {
cannam@95 2850 E Tbv, Tbw, TaZ, Tb0;
cannam@95 2851 Tbv = Tbp + Tbq;
cannam@95 2852 Tbw = FMA(KP1_883088130, Tbs, KP673779706 * Tbt);
cannam@95 2853 R1[WS(rs, 28)] = Tbv - Tbw;
cannam@95 2854 R1[WS(rs, 60)] = Tbv + Tbw;
cannam@95 2855 TaZ = T8Z - T9C;
cannam@95 2856 Tb0 = FMA(KP098135348, TaI, KP1_997590912 * TaX);
cannam@95 2857 R1[WS(rs, 16)] = TaZ - Tb0;
cannam@95 2858 R1[WS(rs, 48)] = TaZ + Tb0;
cannam@95 2859 }
cannam@95 2860 {
cannam@95 2861 E Tb3, Tb6, Tbf, Tbm;
cannam@95 2862 Tb3 = Tb1 - Tb2;
cannam@95 2863 Tb6 = FNMS(KP1_481902250, Tb5, KP1_343117909 * Tb4);
cannam@95 2864 R1[WS(rs, 40)] = Tb3 - Tb6;
cannam@95 2865 R1[WS(rs, 8)] = Tb3 + Tb6;
cannam@95 2866 Tbf = Tbb + Tbe;
cannam@95 2867 Tbm = FNMS(KP855110186, Tbl, KP1_807978586 * Tbi);
cannam@95 2868 R1[WS(rs, 36)] = Tbf - Tbm;
cannam@95 2869 R1[WS(rs, 4)] = Tbf + Tbm;
cannam@95 2870 }
cannam@95 2871 {
cannam@95 2872 E Tbn, Tbo, Tb7, Tb8;
cannam@95 2873 Tbn = Tbb - Tbe;
cannam@95 2874 Tbo = FMA(KP855110186, Tbi, KP1_807978586 * Tbl);
cannam@95 2875 R1[WS(rs, 20)] = Tbn - Tbo;
cannam@95 2876 R1[WS(rs, 52)] = Tbn + Tbo;
cannam@95 2877 Tb7 = Tb1 + Tb2;
cannam@95 2878 Tb8 = FMA(KP1_481902250, Tb4, KP1_343117909 * Tb5);
cannam@95 2879 R1[WS(rs, 24)] = Tb7 - Tb8;
cannam@95 2880 R1[WS(rs, 56)] = Tb7 + Tb8;
cannam@95 2881 }
cannam@95 2882 }
cannam@95 2883 {
cannam@95 2884 E TcR, TdR, Tda, TdS, Te4, Teg, Te1, Tef, TdN, Tei, TdV, Te8, TdC, Tej, TdU;
cannam@95 2885 E Teb;
cannam@95 2886 {
cannam@95 2887 E TcJ, TcQ, TdZ, Te0;
cannam@95 2888 TcJ = TcF + TcI;
cannam@95 2889 TcQ = FNMS(KP1_111140466, TcP, KP1_662939224 * TcM);
cannam@95 2890 TcR = TcJ + TcQ;
cannam@95 2891 TdR = TcJ - TcQ;
cannam@95 2892 {
cannam@95 2893 E Td2, Td9, Te2, Te3;
cannam@95 2894 Td2 = TcU + Td1;
cannam@95 2895 Td9 = Td5 + Td8;
cannam@95 2896 Tda = FNMS(KP580569354, Td9, KP1_913880671 * Td2);
cannam@95 2897 TdS = FMA(KP580569354, Td2, KP1_913880671 * Td9);
cannam@95 2898 Te2 = TcU - Td1;
cannam@95 2899 Te3 = Td8 - Td5;
cannam@95 2900 Te4 = FNMS(KP1_763842528, Te3, KP942793473 * Te2);
cannam@95 2901 Teg = FMA(KP1_763842528, Te2, KP942793473 * Te3);
cannam@95 2902 }
cannam@95 2903 TdZ = TcF - TcI;
cannam@95 2904 Te0 = FMA(KP1_111140466, TcM, KP1_662939224 * TcP);
cannam@95 2905 Te1 = TdZ - Te0;
cannam@95 2906 Tef = TdZ + Te0;
cannam@95 2907 {
cannam@95 2908 E TdM, Te6, TdF, Te7, TdD, TdE;
cannam@95 2909 TdM = TdI + TdL;
cannam@95 2910 Te6 = Tde - Tdl;
cannam@95 2911 TdD = FMA(KP555570233, Tdp, KP831469612 * Tds);
cannam@95 2912 TdE = FNMS(KP555570233, Tdw, KP831469612 * Tdz);
cannam@95 2913 TdF = TdD + TdE;
cannam@95 2914 Te7 = TdE - TdD;
cannam@95 2915 TdN = TdF + TdM;
cannam@95 2916 Tei = Te6 - Te7;
cannam@95 2917 TdV = TdM - TdF;
cannam@95 2918 Te8 = Te6 + Te7;
cannam@95 2919 }
cannam@95 2920 {
cannam@95 2921 E Tdm, Tea, TdB, Te9, Tdt, TdA;
cannam@95 2922 Tdm = Tde + Tdl;
cannam@95 2923 Tea = TdL - TdI;
cannam@95 2924 Tdt = FNMS(KP555570233, Tds, KP831469612 * Tdp);
cannam@95 2925 TdA = FMA(KP831469612, Tdw, KP555570233 * Tdz);
cannam@95 2926 TdB = Tdt + TdA;
cannam@95 2927 Te9 = Tdt - TdA;
cannam@95 2928 TdC = Tdm + TdB;
cannam@95 2929 Tej = Tea - Te9;
cannam@95 2930 TdU = Tdm - TdB;
cannam@95 2931 Teb = Te9 + Tea;
cannam@95 2932 }
cannam@95 2933 }
cannam@95 2934 {
cannam@95 2935 E Tdb, TdO, Teh, Tek;
cannam@95 2936 Tdb = TcR + Tda;
cannam@95 2937 TdO = FNMS(KP293460948, TdN, KP1_978353019 * TdC);
cannam@95 2938 R1[WS(rs, 33)] = Tdb - TdO;
cannam@95 2939 R1[WS(rs, 1)] = Tdb + TdO;
cannam@95 2940 Teh = Tef - Teg;
cannam@95 2941 Tek = FNMS(KP1_940062506, Tej, KP485960359 * Tei);
cannam@95 2942 R1[WS(rs, 45)] = Teh - Tek;
cannam@95 2943 R1[WS(rs, 13)] = Teh + Tek;
cannam@95 2944 }
cannam@95 2945 {
cannam@95 2946 E Tel, Tem, TdP, TdQ;
cannam@95 2947 Tel = Tef + Teg;
cannam@95 2948 Tem = FMA(KP1_940062506, Tei, KP485960359 * Tej);
cannam@95 2949 R1[WS(rs, 29)] = Tel - Tem;
cannam@95 2950 R1[WS(rs, 61)] = Tel + Tem;
cannam@95 2951 TdP = TcR - Tda;
cannam@95 2952 TdQ = FMA(KP293460948, TdC, KP1_978353019 * TdN);
cannam@95 2953 R1[WS(rs, 17)] = TdP - TdQ;
cannam@95 2954 R1[WS(rs, 49)] = TdP + TdQ;
cannam@95 2955 }
cannam@95 2956 {
cannam@95 2957 E TdT, TdW, Te5, Tec;
cannam@95 2958 TdT = TdR - TdS;
cannam@95 2959 TdW = FNMS(KP1_606415062, TdV, KP1_191398608 * TdU);
cannam@95 2960 R1[WS(rs, 41)] = TdT - TdW;
cannam@95 2961 R1[WS(rs, 9)] = TdT + TdW;
cannam@95 2962 Te5 = Te1 + Te4;
cannam@95 2963 Tec = FNMS(KP1_028205488, Teb, KP1_715457220 * Te8);
cannam@95 2964 R1[WS(rs, 37)] = Te5 - Tec;
cannam@95 2965 R1[WS(rs, 5)] = Te5 + Tec;
cannam@95 2966 }
cannam@95 2967 {
cannam@95 2968 E Ted, Tee, TdX, TdY;
cannam@95 2969 Ted = Te1 - Te4;
cannam@95 2970 Tee = FMA(KP1_028205488, Te8, KP1_715457220 * Teb);
cannam@95 2971 R1[WS(rs, 21)] = Ted - Tee;
cannam@95 2972 R1[WS(rs, 53)] = Ted + Tee;
cannam@95 2973 TdX = TdR + TdS;
cannam@95 2974 TdY = FMA(KP1_606415062, TdU, KP1_191398608 * TdV);
cannam@95 2975 R1[WS(rs, 25)] = TdX - TdY;
cannam@95 2976 R1[WS(rs, 57)] = TdX + TdY;
cannam@95 2977 }
cannam@95 2978 }
cannam@95 2979 {
cannam@95 2980 E TbD, Tc7, TbK, Tc8, Tck, Tcw, Tch, Tcv, Tc3, Tcy, Tcb, Tco, TbW, Tcz, Tca;
cannam@95 2981 E Tcr;
cannam@95 2982 {
cannam@95 2983 E Tbz, TbC, Tcf, Tcg;
cannam@95 2984 Tbz = Tbx - Tby;
cannam@95 2985 TbC = FNMS(KP1_662939224, TbB, KP1_111140466 * TbA);
cannam@95 2986 TbD = Tbz + TbC;
cannam@95 2987 Tc7 = Tbz - TbC;
cannam@95 2988 {
cannam@95 2989 E TbG, TbJ, Tci, Tcj;
cannam@95 2990 TbG = TbE + TbF;
cannam@95 2991 TbJ = TbH + TbI;
cannam@95 2992 TbK = FNMS(KP942793473, TbJ, KP1_763842528 * TbG);
cannam@95 2993 Tc8 = FMA(KP942793473, TbG, KP1_763842528 * TbJ);
cannam@95 2994 Tci = TbE - TbF;
cannam@95 2995 Tcj = TbI - TbH;
cannam@95 2996 Tck = FNMS(KP1_913880671, Tcj, KP580569354 * Tci);
cannam@95 2997 Tcw = FMA(KP1_913880671, Tci, KP580569354 * Tcj);
cannam@95 2998 }
cannam@95 2999 Tcf = Tbx + Tby;
cannam@95 3000 Tcg = FMA(KP1_662939224, TbA, KP1_111140466 * TbB);
cannam@95 3001 Tch = Tcf - Tcg;
cannam@95 3002 Tcv = Tcf + Tcg;
cannam@95 3003 {
cannam@95 3004 E Tc2, Tcm, TbZ, Tcn, TbX, TbY;
cannam@95 3005 Tc2 = Tc0 + Tc1;
cannam@95 3006 Tcm = TbM - TbN;
cannam@95 3007 TbX = FMA(KP831469612, TbP, KP555570233 * TbQ);
cannam@95 3008 TbY = FNMS(KP831469612, TbS, KP555570233 * TbT);
cannam@95 3009 TbZ = TbX + TbY;
cannam@95 3010 Tcn = TbY - TbX;
cannam@95 3011 Tc3 = TbZ + Tc2;
cannam@95 3012 Tcy = Tcm - Tcn;
cannam@95 3013 Tcb = Tc2 - TbZ;
cannam@95 3014 Tco = Tcm + Tcn;
cannam@95 3015 }
cannam@95 3016 {
cannam@95 3017 E TbO, Tcq, TbV, Tcp, TbR, TbU;
cannam@95 3018 TbO = TbM + TbN;
cannam@95 3019 Tcq = Tc1 - Tc0;
cannam@95 3020 TbR = FNMS(KP831469612, TbQ, KP555570233 * TbP);
cannam@95 3021 TbU = FMA(KP555570233, TbS, KP831469612 * TbT);
cannam@95 3022 TbV = TbR + TbU;
cannam@95 3023 Tcp = TbR - TbU;
cannam@95 3024 TbW = TbO + TbV;
cannam@95 3025 Tcz = Tcq - Tcp;
cannam@95 3026 Tca = TbO - TbV;
cannam@95 3027 Tcr = Tcp + Tcq;
cannam@95 3028 }
cannam@95 3029 }
cannam@95 3030 {
cannam@95 3031 E TbL, Tc4, Tcx, TcA;
cannam@95 3032 TbL = TbD + TbK;
cannam@95 3033 Tc4 = FNMS(KP485960359, Tc3, KP1_940062506 * TbW);
cannam@95 3034 R1[WS(rs, 34)] = TbL - Tc4;
cannam@95 3035 R1[WS(rs, 2)] = TbL + Tc4;
cannam@95 3036 Tcx = Tcv - Tcw;
cannam@95 3037 TcA = FNMS(KP1_978353019, Tcz, KP293460948 * Tcy);
cannam@95 3038 R1[WS(rs, 46)] = Tcx - TcA;
cannam@95 3039 R1[WS(rs, 14)] = Tcx + TcA;
cannam@95 3040 }
cannam@95 3041 {
cannam@95 3042 E TcB, TcC, Tc5, Tc6;
cannam@95 3043 TcB = Tcv + Tcw;
cannam@95 3044 TcC = FMA(KP1_978353019, Tcy, KP293460948 * Tcz);
cannam@95 3045 R1[WS(rs, 30)] = TcB - TcC;
cannam@95 3046 R1[WS(rs, 62)] = TcB + TcC;
cannam@95 3047 Tc5 = TbD - TbK;
cannam@95 3048 Tc6 = FMA(KP485960359, TbW, KP1_940062506 * Tc3);
cannam@95 3049 R1[WS(rs, 18)] = Tc5 - Tc6;
cannam@95 3050 R1[WS(rs, 50)] = Tc5 + Tc6;
cannam@95 3051 }
cannam@95 3052 {
cannam@95 3053 E Tc9, Tcc, Tcl, Tcs;
cannam@95 3054 Tc9 = Tc7 - Tc8;
cannam@95 3055 Tcc = FNMS(KP1_715457220, Tcb, KP1_028205488 * Tca);
cannam@95 3056 R1[WS(rs, 42)] = Tc9 - Tcc;
cannam@95 3057 R1[WS(rs, 10)] = Tc9 + Tcc;
cannam@95 3058 Tcl = Tch + Tck;
cannam@95 3059 Tcs = FNMS(KP1_191398608, Tcr, KP1_606415062 * Tco);
cannam@95 3060 R1[WS(rs, 38)] = Tcl - Tcs;
cannam@95 3061 R1[WS(rs, 6)] = Tcl + Tcs;
cannam@95 3062 }
cannam@95 3063 {
cannam@95 3064 E Tct, Tcu, Tcd, Tce;
cannam@95 3065 Tct = Tch - Tck;
cannam@95 3066 Tcu = FMA(KP1_191398608, Tco, KP1_606415062 * Tcr);
cannam@95 3067 R1[WS(rs, 22)] = Tct - Tcu;
cannam@95 3068 R1[WS(rs, 54)] = Tct + Tcu;
cannam@95 3069 Tcd = Tc7 + Tc8;
cannam@95 3070 Tce = FMA(KP1_715457220, Tca, KP1_028205488 * Tcb);
cannam@95 3071 R1[WS(rs, 26)] = Tcd - Tce;
cannam@95 3072 R1[WS(rs, 58)] = Tcd + Tce;
cannam@95 3073 }
cannam@95 3074 }
cannam@95 3075 {
cannam@95 3076 E Tet, TeX, TeA, TeY, Tfa, Tfm, Tf7, Tfl, TeT, Tfo, Tf1, Tfe, TeM, Tfp, Tf0;
cannam@95 3077 E Tfh;
cannam@95 3078 {
cannam@95 3079 E Tep, Tes, Tf5, Tf6;
cannam@95 3080 Tep = Ten - Teo;
cannam@95 3081 Tes = FNMS(KP1_961570560, Ter, KP390180644 * Teq);
cannam@95 3082 Tet = Tep + Tes;
cannam@95 3083 TeX = Tep - Tes;
cannam@95 3084 {
cannam@95 3085 E Tew, Tez, Tf8, Tf9;
cannam@95 3086 Tew = Teu - Tev;
cannam@95 3087 Tez = Tex + Tey;
cannam@95 3088 TeA = FNMS(KP1_268786568, Tez, KP1_546020906 * Tew);
cannam@95 3089 TeY = FMA(KP1_268786568, Tew, KP1_546020906 * Tez);
cannam@95 3090 Tf8 = Teu + Tev;
cannam@95 3091 Tf9 = Tey - Tex;
cannam@95 3092 Tfa = FNMS(KP1_990369453, Tf9, KP196034280 * Tf8);
cannam@95 3093 Tfm = FMA(KP1_990369453, Tf8, KP196034280 * Tf9);
cannam@95 3094 }
cannam@95 3095 Tf5 = Ten + Teo;
cannam@95 3096 Tf6 = FMA(KP1_961570560, Teq, KP390180644 * Ter);
cannam@95 3097 Tf7 = Tf5 - Tf6;
cannam@95 3098 Tfl = Tf5 + Tf6;
cannam@95 3099 {
cannam@95 3100 E TeS, Tfc, TeP, Tfd, TeN, TeO;
cannam@95 3101 TeS = TeQ + TeR;
cannam@95 3102 Tfc = TeC + TeD;
cannam@95 3103 TeN = FMA(KP980785280, TeF, KP195090322 * TeG);
cannam@95 3104 TeO = FMA(KP980785280, TeI, KP195090322 * TeJ);
cannam@95 3105 TeP = TeN - TeO;
cannam@95 3106 Tfd = TeN + TeO;
cannam@95 3107 TeT = TeP + TeS;
cannam@95 3108 Tfo = Tfc + Tfd;
cannam@95 3109 Tf1 = TeS - TeP;
cannam@95 3110 Tfe = Tfc - Tfd;
cannam@95 3111 }
cannam@95 3112 {
cannam@95 3113 E TeE, Tfg, TeL, Tff, TeH, TeK;
cannam@95 3114 TeE = TeC - TeD;
cannam@95 3115 Tfg = TeR - TeQ;
cannam@95 3116 TeH = FNMS(KP980785280, TeG, KP195090322 * TeF);
cannam@95 3117 TeK = FNMS(KP980785280, TeJ, KP195090322 * TeI);
cannam@95 3118 TeL = TeH + TeK;
cannam@95 3119 Tff = TeH - TeK;
cannam@95 3120 TeM = TeE + TeL;
cannam@95 3121 Tfp = Tfg - Tff;
cannam@95 3122 Tf0 = TeE - TeL;
cannam@95 3123 Tfh = Tff + Tfg;
cannam@95 3124 }
cannam@95 3125 }
cannam@95 3126 {
cannam@95 3127 E TeB, TeU, Tfn, Tfq;
cannam@95 3128 TeB = Tet + TeA;
cannam@95 3129 TeU = FNMS(KP673779706, TeT, KP1_883088130 * TeM);
cannam@95 3130 R1[WS(rs, 35)] = TeB - TeU;
cannam@95 3131 R1[WS(rs, 3)] = TeB + TeU;
cannam@95 3132 Tfn = Tfl - Tfm;
cannam@95 3133 Tfq = FNMS(KP1_997590912, Tfp, KP098135348 * Tfo);
cannam@95 3134 R1[WS(rs, 47)] = Tfn - Tfq;
cannam@95 3135 R1[WS(rs, 15)] = Tfn + Tfq;
cannam@95 3136 }
cannam@95 3137 {
cannam@95 3138 E Tfr, Tfs, TeV, TeW;
cannam@95 3139 Tfr = Tfl + Tfm;
cannam@95 3140 Tfs = FMA(KP1_997590912, Tfo, KP098135348 * Tfp);
cannam@95 3141 R1[WS(rs, 31)] = Tfr - Tfs;
cannam@95 3142 R1[WS(rs, 63)] = Tfr + Tfs;
cannam@95 3143 TeV = Tet - TeA;
cannam@95 3144 TeW = FMA(KP673779706, TeM, KP1_883088130 * TeT);
cannam@95 3145 R1[WS(rs, 19)] = TeV - TeW;
cannam@95 3146 R1[WS(rs, 51)] = TeV + TeW;
cannam@95 3147 }
cannam@95 3148 {
cannam@95 3149 E TeZ, Tf2, Tfb, Tfi;
cannam@95 3150 TeZ = TeX - TeY;
cannam@95 3151 Tf2 = FNMS(KP1_807978586, Tf1, KP855110186 * Tf0);
cannam@95 3152 R1[WS(rs, 43)] = TeZ - Tf2;
cannam@95 3153 R1[WS(rs, 11)] = TeZ + Tf2;
cannam@95 3154 Tfb = Tf7 + Tfa;
cannam@95 3155 Tfi = FNMS(KP1_343117909, Tfh, KP1_481902250 * Tfe);
cannam@95 3156 R1[WS(rs, 39)] = Tfb - Tfi;
cannam@95 3157 R1[WS(rs, 7)] = Tfb + Tfi;
cannam@95 3158 }
cannam@95 3159 {
cannam@95 3160 E Tfj, Tfk, Tf3, Tf4;
cannam@95 3161 Tfj = Tf7 - Tfa;
cannam@95 3162 Tfk = FMA(KP1_343117909, Tfe, KP1_481902250 * Tfh);
cannam@95 3163 R1[WS(rs, 23)] = Tfj - Tfk;
cannam@95 3164 R1[WS(rs, 55)] = Tfj + Tfk;
cannam@95 3165 Tf3 = TeX + TeY;
cannam@95 3166 Tf4 = FMA(KP1_807978586, Tf0, KP855110186 * Tf1);
cannam@95 3167 R1[WS(rs, 27)] = Tf3 - Tf4;
cannam@95 3168 R1[WS(rs, 59)] = Tf3 + Tf4;
cannam@95 3169 }
cannam@95 3170 }
cannam@95 3171 }
cannam@95 3172 }
cannam@95 3173 }
cannam@95 3174
cannam@95 3175 static const kr2c_desc desc = { 128, "r2cb_128", {812, 198, 144, 0}, &GENUS };
cannam@95 3176
cannam@95 3177 void X(codelet_r2cb_128) (planner *p) {
cannam@95 3178 X(kr2c_register) (p, r2cb_128, &desc);
cannam@95 3179 }
cannam@95 3180
cannam@95 3181 #endif /* HAVE_FMA */