Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.8/rdft/scalar/r2cf/r2cf_32.c @ 82:d0c2a83c1364

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Add FFTW 3.3.8 source, and a Linux build
author Chris Cannam
date Tue, 19 Nov 2019 14:52:55 +0000
parents
children
rev   line source
Chris@82 1 /*
Chris@82 2 * Copyright (c) 2003, 2007-14 Matteo Frigo
Chris@82 3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
Chris@82 4 *
Chris@82 5 * This program is free software; you can redistribute it and/or modify
Chris@82 6 * it under the terms of the GNU General Public License as published by
Chris@82 7 * the Free Software Foundation; either version 2 of the License, or
Chris@82 8 * (at your option) any later version.
Chris@82 9 *
Chris@82 10 * This program is distributed in the hope that it will be useful,
Chris@82 11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
Chris@82 12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
Chris@82 13 * GNU General Public License for more details.
Chris@82 14 *
Chris@82 15 * You should have received a copy of the GNU General Public License
Chris@82 16 * along with this program; if not, write to the Free Software
Chris@82 17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
Chris@82 18 *
Chris@82 19 */
Chris@82 20
Chris@82 21 /* This file was automatically generated --- DO NOT EDIT */
Chris@82 22 /* Generated on Thu May 24 08:06:27 EDT 2018 */
Chris@82 23
Chris@82 24 #include "rdft/codelet-rdft.h"
Chris@82 25
Chris@82 26 #if defined(ARCH_PREFERS_FMA) || defined(ISA_EXTENSION_PREFERS_FMA)
Chris@82 27
Chris@82 28 /* Generated by: ../../../genfft/gen_r2cf.native -fma -compact -variables 4 -pipeline-latency 4 -n 32 -name r2cf_32 -include rdft/scalar/r2cf.h */
Chris@82 29
Chris@82 30 /*
Chris@82 31 * This function contains 156 FP additions, 68 FP multiplications,
Chris@82 32 * (or, 88 additions, 0 multiplications, 68 fused multiply/add),
Chris@82 33 * 54 stack variables, 7 constants, and 64 memory accesses
Chris@82 34 */
Chris@82 35 #include "rdft/scalar/r2cf.h"
Chris@82 36
Chris@82 37 static void r2cf_32(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
Chris@82 38 {
Chris@82 39 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
Chris@82 40 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
Chris@82 41 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
Chris@82 42 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
Chris@82 43 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
Chris@82 44 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
Chris@82 45 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
Chris@82 46 {
Chris@82 47 INT i;
Chris@82 48 for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(128, rs), MAKE_VOLATILE_STRIDE(128, csr), MAKE_VOLATILE_STRIDE(128, csi)) {
Chris@82 49 E T7, T2b, Tv, T1h, Te, T2n, Ty, T1i, Tt, T2d, TF, T1l, Tm, T2c, TC;
Chris@82 50 E T1k, T1Z, T22, T2k, T2j, T1e, T1C, T19, T1B, T1S, T1V, T2h, T2g, TX, T1z;
Chris@82 51 E TS, T1y;
Chris@82 52 {
Chris@82 53 E T1, T2, T3, T4, T5, T6;
Chris@82 54 T1 = R0[0];
Chris@82 55 T2 = R0[WS(rs, 8)];
Chris@82 56 T3 = T1 + T2;
Chris@82 57 T4 = R0[WS(rs, 4)];
Chris@82 58 T5 = R0[WS(rs, 12)];
Chris@82 59 T6 = T4 + T5;
Chris@82 60 T7 = T3 + T6;
Chris@82 61 T2b = T3 - T6;
Chris@82 62 Tv = T1 - T2;
Chris@82 63 T1h = T4 - T5;
Chris@82 64 }
Chris@82 65 {
Chris@82 66 E Ta, Tw, Td, Tx;
Chris@82 67 {
Chris@82 68 E T8, T9, Tb, Tc;
Chris@82 69 T8 = R0[WS(rs, 2)];
Chris@82 70 T9 = R0[WS(rs, 10)];
Chris@82 71 Ta = T8 + T9;
Chris@82 72 Tw = T8 - T9;
Chris@82 73 Tb = R0[WS(rs, 14)];
Chris@82 74 Tc = R0[WS(rs, 6)];
Chris@82 75 Td = Tb + Tc;
Chris@82 76 Tx = Tb - Tc;
Chris@82 77 }
Chris@82 78 Te = Ta + Td;
Chris@82 79 T2n = Td - Ta;
Chris@82 80 Ty = Tw + Tx;
Chris@82 81 T1i = Tx - Tw;
Chris@82 82 }
Chris@82 83 {
Chris@82 84 E Tp, TD, Ts, TE;
Chris@82 85 {
Chris@82 86 E Tn, To, Tq, Tr;
Chris@82 87 Tn = R0[WS(rs, 15)];
Chris@82 88 To = R0[WS(rs, 7)];
Chris@82 89 Tp = Tn + To;
Chris@82 90 TD = Tn - To;
Chris@82 91 Tq = R0[WS(rs, 3)];
Chris@82 92 Tr = R0[WS(rs, 11)];
Chris@82 93 Ts = Tq + Tr;
Chris@82 94 TE = Tq - Tr;
Chris@82 95 }
Chris@82 96 Tt = Tp + Ts;
Chris@82 97 T2d = Tp - Ts;
Chris@82 98 TF = FMA(KP414213562, TE, TD);
Chris@82 99 T1l = FNMS(KP414213562, TD, TE);
Chris@82 100 }
Chris@82 101 {
Chris@82 102 E Ti, TA, Tl, TB;
Chris@82 103 {
Chris@82 104 E Tg, Th, Tj, Tk;
Chris@82 105 Tg = R0[WS(rs, 1)];
Chris@82 106 Th = R0[WS(rs, 9)];
Chris@82 107 Ti = Tg + Th;
Chris@82 108 TA = Tg - Th;
Chris@82 109 Tj = R0[WS(rs, 5)];
Chris@82 110 Tk = R0[WS(rs, 13)];
Chris@82 111 Tl = Tj + Tk;
Chris@82 112 TB = Tj - Tk;
Chris@82 113 }
Chris@82 114 Tm = Ti + Tl;
Chris@82 115 T2c = Ti - Tl;
Chris@82 116 TC = FNMS(KP414213562, TB, TA);
Chris@82 117 T1k = FMA(KP414213562, TA, TB);
Chris@82 118 }
Chris@82 119 {
Chris@82 120 E T11, T1X, T1c, T1Y, T14, T20, T17, T21, T1d, T18;
Chris@82 121 {
Chris@82 122 E TZ, T10, T1a, T1b;
Chris@82 123 TZ = R1[WS(rs, 15)];
Chris@82 124 T10 = R1[WS(rs, 7)];
Chris@82 125 T11 = TZ - T10;
Chris@82 126 T1X = TZ + T10;
Chris@82 127 T1a = R1[WS(rs, 11)];
Chris@82 128 T1b = R1[WS(rs, 3)];
Chris@82 129 T1c = T1a - T1b;
Chris@82 130 T1Y = T1b + T1a;
Chris@82 131 }
Chris@82 132 {
Chris@82 133 E T12, T13, T15, T16;
Chris@82 134 T12 = R1[WS(rs, 1)];
Chris@82 135 T13 = R1[WS(rs, 9)];
Chris@82 136 T14 = T12 - T13;
Chris@82 137 T20 = T12 + T13;
Chris@82 138 T15 = R1[WS(rs, 13)];
Chris@82 139 T16 = R1[WS(rs, 5)];
Chris@82 140 T17 = T15 - T16;
Chris@82 141 T21 = T15 + T16;
Chris@82 142 }
Chris@82 143 T1Z = T1X + T1Y;
Chris@82 144 T22 = T20 + T21;
Chris@82 145 T2k = T21 - T20;
Chris@82 146 T2j = T1X - T1Y;
Chris@82 147 T1d = T17 - T14;
Chris@82 148 T1e = FMA(KP707106781, T1d, T1c);
Chris@82 149 T1C = FNMS(KP707106781, T1d, T1c);
Chris@82 150 T18 = T14 + T17;
Chris@82 151 T19 = FMA(KP707106781, T18, T11);
Chris@82 152 T1B = FNMS(KP707106781, T18, T11);
Chris@82 153 }
Chris@82 154 {
Chris@82 155 E TK, T1Q, TV, T1R, TN, T1T, TQ, T1U, TW, TR;
Chris@82 156 {
Chris@82 157 E TI, TJ, TT, TU;
Chris@82 158 TI = R1[0];
Chris@82 159 TJ = R1[WS(rs, 8)];
Chris@82 160 TK = TI - TJ;
Chris@82 161 T1Q = TI + TJ;
Chris@82 162 TT = R1[WS(rs, 4)];
Chris@82 163 TU = R1[WS(rs, 12)];
Chris@82 164 TV = TT - TU;
Chris@82 165 T1R = TT + TU;
Chris@82 166 }
Chris@82 167 {
Chris@82 168 E TL, TM, TO, TP;
Chris@82 169 TL = R1[WS(rs, 2)];
Chris@82 170 TM = R1[WS(rs, 10)];
Chris@82 171 TN = TL - TM;
Chris@82 172 T1T = TL + TM;
Chris@82 173 TO = R1[WS(rs, 14)];
Chris@82 174 TP = R1[WS(rs, 6)];
Chris@82 175 TQ = TO - TP;
Chris@82 176 T1U = TO + TP;
Chris@82 177 }
Chris@82 178 T1S = T1Q + T1R;
Chris@82 179 T1V = T1T + T1U;
Chris@82 180 T2h = T1U - T1T;
Chris@82 181 T2g = T1Q - T1R;
Chris@82 182 TW = TN - TQ;
Chris@82 183 TX = FMA(KP707106781, TW, TV);
Chris@82 184 T1z = FNMS(KP707106781, TW, TV);
Chris@82 185 TR = TN + TQ;
Chris@82 186 TS = FMA(KP707106781, TR, TK);
Chris@82 187 T1y = FNMS(KP707106781, TR, TK);
Chris@82 188 }
Chris@82 189 {
Chris@82 190 E Tf, Tu, T27, T28, T29, T2a;
Chris@82 191 Tf = T7 + Te;
Chris@82 192 Tu = Tm + Tt;
Chris@82 193 T27 = Tf + Tu;
Chris@82 194 T28 = T1S + T1V;
Chris@82 195 T29 = T1Z + T22;
Chris@82 196 T2a = T28 + T29;
Chris@82 197 Cr[WS(csr, 8)] = Tf - Tu;
Chris@82 198 Ci[WS(csi, 8)] = T29 - T28;
Chris@82 199 Cr[WS(csr, 16)] = T27 - T2a;
Chris@82 200 Cr[0] = T27 + T2a;
Chris@82 201 }
Chris@82 202 {
Chris@82 203 E T1P, T25, T24, T26, T1W, T23;
Chris@82 204 T1P = T7 - Te;
Chris@82 205 T25 = Tt - Tm;
Chris@82 206 T1W = T1S - T1V;
Chris@82 207 T23 = T1Z - T22;
Chris@82 208 T24 = T1W + T23;
Chris@82 209 T26 = T23 - T1W;
Chris@82 210 Cr[WS(csr, 12)] = FNMS(KP707106781, T24, T1P);
Chris@82 211 Ci[WS(csi, 12)] = FMS(KP707106781, T26, T25);
Chris@82 212 Cr[WS(csr, 4)] = FMA(KP707106781, T24, T1P);
Chris@82 213 Ci[WS(csi, 4)] = FMA(KP707106781, T26, T25);
Chris@82 214 }
Chris@82 215 {
Chris@82 216 E T2f, T2v, T2p, T2r, T2m, T2q, T2u, T2w, T2e, T2o;
Chris@82 217 T2e = T2c + T2d;
Chris@82 218 T2f = FMA(KP707106781, T2e, T2b);
Chris@82 219 T2v = FNMS(KP707106781, T2e, T2b);
Chris@82 220 T2o = T2d - T2c;
Chris@82 221 T2p = FNMS(KP707106781, T2o, T2n);
Chris@82 222 T2r = FMA(KP707106781, T2o, T2n);
Chris@82 223 {
Chris@82 224 E T2i, T2l, T2s, T2t;
Chris@82 225 T2i = FMA(KP414213562, T2h, T2g);
Chris@82 226 T2l = FNMS(KP414213562, T2k, T2j);
Chris@82 227 T2m = T2i + T2l;
Chris@82 228 T2q = T2l - T2i;
Chris@82 229 T2s = FNMS(KP414213562, T2g, T2h);
Chris@82 230 T2t = FMA(KP414213562, T2j, T2k);
Chris@82 231 T2u = T2s + T2t;
Chris@82 232 T2w = T2t - T2s;
Chris@82 233 }
Chris@82 234 Cr[WS(csr, 14)] = FNMS(KP923879532, T2m, T2f);
Chris@82 235 Ci[WS(csi, 14)] = FMS(KP923879532, T2u, T2r);
Chris@82 236 Cr[WS(csr, 2)] = FMA(KP923879532, T2m, T2f);
Chris@82 237 Ci[WS(csi, 2)] = FMA(KP923879532, T2u, T2r);
Chris@82 238 Ci[WS(csi, 6)] = FMS(KP923879532, T2q, T2p);
Chris@82 239 Cr[WS(csr, 6)] = FMA(KP923879532, T2w, T2v);
Chris@82 240 Ci[WS(csi, 10)] = FMA(KP923879532, T2q, T2p);
Chris@82 241 Cr[WS(csr, 10)] = FNMS(KP923879532, T2w, T2v);
Chris@82 242 }
Chris@82 243 {
Chris@82 244 E TH, T1t, T1s, T1u, T1g, T1o, T1n, T1p;
Chris@82 245 {
Chris@82 246 E Tz, TG, T1q, T1r;
Chris@82 247 Tz = FMA(KP707106781, Ty, Tv);
Chris@82 248 TG = TC + TF;
Chris@82 249 TH = FMA(KP923879532, TG, Tz);
Chris@82 250 T1t = FNMS(KP923879532, TG, Tz);
Chris@82 251 T1q = FMA(KP198912367, T19, T1e);
Chris@82 252 T1r = FMA(KP198912367, TS, TX);
Chris@82 253 T1s = T1q - T1r;
Chris@82 254 T1u = T1r + T1q;
Chris@82 255 }
Chris@82 256 {
Chris@82 257 E TY, T1f, T1j, T1m;
Chris@82 258 TY = FNMS(KP198912367, TX, TS);
Chris@82 259 T1f = FNMS(KP198912367, T1e, T19);
Chris@82 260 T1g = TY + T1f;
Chris@82 261 T1o = T1f - TY;
Chris@82 262 T1j = FNMS(KP707106781, T1i, T1h);
Chris@82 263 T1m = T1k + T1l;
Chris@82 264 T1n = FNMS(KP923879532, T1m, T1j);
Chris@82 265 T1p = FMA(KP923879532, T1m, T1j);
Chris@82 266 }
Chris@82 267 Cr[WS(csr, 15)] = FNMS(KP980785280, T1g, TH);
Chris@82 268 Ci[WS(csi, 15)] = FMA(KP980785280, T1s, T1p);
Chris@82 269 Cr[WS(csr, 1)] = FMA(KP980785280, T1g, TH);
Chris@82 270 Ci[WS(csi, 1)] = FMS(KP980785280, T1s, T1p);
Chris@82 271 Ci[WS(csi, 7)] = FMA(KP980785280, T1o, T1n);
Chris@82 272 Cr[WS(csr, 7)] = FMA(KP980785280, T1u, T1t);
Chris@82 273 Ci[WS(csi, 9)] = FMS(KP980785280, T1o, T1n);
Chris@82 274 Cr[WS(csr, 9)] = FNMS(KP980785280, T1u, T1t);
Chris@82 275 }
Chris@82 276 {
Chris@82 277 E T1x, T1N, T1M, T1O, T1E, T1I, T1H, T1J;
Chris@82 278 {
Chris@82 279 E T1v, T1w, T1K, T1L;
Chris@82 280 T1v = FNMS(KP707106781, Ty, Tv);
Chris@82 281 T1w = T1k - T1l;
Chris@82 282 T1x = FMA(KP923879532, T1w, T1v);
Chris@82 283 T1N = FNMS(KP923879532, T1w, T1v);
Chris@82 284 T1K = FNMS(KP668178637, T1y, T1z);
Chris@82 285 T1L = FNMS(KP668178637, T1B, T1C);
Chris@82 286 T1M = T1K - T1L;
Chris@82 287 T1O = T1K + T1L;
Chris@82 288 }
Chris@82 289 {
Chris@82 290 E T1A, T1D, T1F, T1G;
Chris@82 291 T1A = FMA(KP668178637, T1z, T1y);
Chris@82 292 T1D = FMA(KP668178637, T1C, T1B);
Chris@82 293 T1E = T1A + T1D;
Chris@82 294 T1I = T1D - T1A;
Chris@82 295 T1F = FMA(KP707106781, T1i, T1h);
Chris@82 296 T1G = TF - TC;
Chris@82 297 T1H = FNMS(KP923879532, T1G, T1F);
Chris@82 298 T1J = FMA(KP923879532, T1G, T1F);
Chris@82 299 }
Chris@82 300 Cr[WS(csr, 13)] = FNMS(KP831469612, T1E, T1x);
Chris@82 301 Ci[WS(csi, 13)] = FMS(KP831469612, T1M, T1J);
Chris@82 302 Cr[WS(csr, 3)] = FMA(KP831469612, T1E, T1x);
Chris@82 303 Ci[WS(csi, 3)] = FMA(KP831469612, T1M, T1J);
Chris@82 304 Ci[WS(csi, 5)] = FMS(KP831469612, T1I, T1H);
Chris@82 305 Cr[WS(csr, 5)] = FNMS(KP831469612, T1O, T1N);
Chris@82 306 Ci[WS(csi, 11)] = FMA(KP831469612, T1I, T1H);
Chris@82 307 Cr[WS(csr, 11)] = FMA(KP831469612, T1O, T1N);
Chris@82 308 }
Chris@82 309 }
Chris@82 310 }
Chris@82 311 }
Chris@82 312
Chris@82 313 static const kr2c_desc desc = { 32, "r2cf_32", {88, 0, 68, 0}, &GENUS };
Chris@82 314
Chris@82 315 void X(codelet_r2cf_32) (planner *p) {
Chris@82 316 X(kr2c_register) (p, r2cf_32, &desc);
Chris@82 317 }
Chris@82 318
Chris@82 319 #else
Chris@82 320
Chris@82 321 /* Generated by: ../../../genfft/gen_r2cf.native -compact -variables 4 -pipeline-latency 4 -n 32 -name r2cf_32 -include rdft/scalar/r2cf.h */
Chris@82 322
Chris@82 323 /*
Chris@82 324 * This function contains 156 FP additions, 42 FP multiplications,
Chris@82 325 * (or, 140 additions, 26 multiplications, 16 fused multiply/add),
Chris@82 326 * 54 stack variables, 7 constants, and 64 memory accesses
Chris@82 327 */
Chris@82 328 #include "rdft/scalar/r2cf.h"
Chris@82 329
Chris@82 330 static void r2cf_32(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
Chris@82 331 {
Chris@82 332 DK(KP555570233, +0.555570233019602224742830813948532874374937191);
Chris@82 333 DK(KP831469612, +0.831469612302545237078788377617905756738560812);
Chris@82 334 DK(KP195090322, +0.195090322016128267848284868477022240927691618);
Chris@82 335 DK(KP980785280, +0.980785280403230449126182236134239036973933731);
Chris@82 336 DK(KP382683432, +0.382683432365089771728459984030398866761344562);
Chris@82 337 DK(KP923879532, +0.923879532511286756128183189396788286822416626);
Chris@82 338 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
Chris@82 339 {
Chris@82 340 INT i;
Chris@82 341 for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(128, rs), MAKE_VOLATILE_STRIDE(128, csr), MAKE_VOLATILE_STRIDE(128, csi)) {
Chris@82 342 E T7, T2b, Tv, T1l, Te, T2o, Ty, T1k, Tt, T2d, TF, T1h, Tm, T2c, TC;
Chris@82 343 E T1i, T1Z, T22, T2k, T2j, T1e, T1C, T19, T1B, T1S, T1V, T2h, T2g, TX, T1z;
Chris@82 344 E TS, T1y;
Chris@82 345 {
Chris@82 346 E T1, T2, T3, T4, T5, T6;
Chris@82 347 T1 = R0[0];
Chris@82 348 T2 = R0[WS(rs, 8)];
Chris@82 349 T3 = T1 + T2;
Chris@82 350 T4 = R0[WS(rs, 4)];
Chris@82 351 T5 = R0[WS(rs, 12)];
Chris@82 352 T6 = T4 + T5;
Chris@82 353 T7 = T3 + T6;
Chris@82 354 T2b = T3 - T6;
Chris@82 355 Tv = T1 - T2;
Chris@82 356 T1l = T4 - T5;
Chris@82 357 }
Chris@82 358 {
Chris@82 359 E Ta, Tw, Td, Tx;
Chris@82 360 {
Chris@82 361 E T8, T9, Tb, Tc;
Chris@82 362 T8 = R0[WS(rs, 2)];
Chris@82 363 T9 = R0[WS(rs, 10)];
Chris@82 364 Ta = T8 + T9;
Chris@82 365 Tw = T8 - T9;
Chris@82 366 Tb = R0[WS(rs, 14)];
Chris@82 367 Tc = R0[WS(rs, 6)];
Chris@82 368 Td = Tb + Tc;
Chris@82 369 Tx = Tb - Tc;
Chris@82 370 }
Chris@82 371 Te = Ta + Td;
Chris@82 372 T2o = Td - Ta;
Chris@82 373 Ty = KP707106781 * (Tw + Tx);
Chris@82 374 T1k = KP707106781 * (Tx - Tw);
Chris@82 375 }
Chris@82 376 {
Chris@82 377 E Tp, TD, Ts, TE;
Chris@82 378 {
Chris@82 379 E Tn, To, Tq, Tr;
Chris@82 380 Tn = R0[WS(rs, 15)];
Chris@82 381 To = R0[WS(rs, 7)];
Chris@82 382 Tp = Tn + To;
Chris@82 383 TD = Tn - To;
Chris@82 384 Tq = R0[WS(rs, 3)];
Chris@82 385 Tr = R0[WS(rs, 11)];
Chris@82 386 Ts = Tq + Tr;
Chris@82 387 TE = Tq - Tr;
Chris@82 388 }
Chris@82 389 Tt = Tp + Ts;
Chris@82 390 T2d = Tp - Ts;
Chris@82 391 TF = FMA(KP923879532, TD, KP382683432 * TE);
Chris@82 392 T1h = FNMS(KP923879532, TE, KP382683432 * TD);
Chris@82 393 }
Chris@82 394 {
Chris@82 395 E Ti, TA, Tl, TB;
Chris@82 396 {
Chris@82 397 E Tg, Th, Tj, Tk;
Chris@82 398 Tg = R0[WS(rs, 1)];
Chris@82 399 Th = R0[WS(rs, 9)];
Chris@82 400 Ti = Tg + Th;
Chris@82 401 TA = Tg - Th;
Chris@82 402 Tj = R0[WS(rs, 5)];
Chris@82 403 Tk = R0[WS(rs, 13)];
Chris@82 404 Tl = Tj + Tk;
Chris@82 405 TB = Tj - Tk;
Chris@82 406 }
Chris@82 407 Tm = Ti + Tl;
Chris@82 408 T2c = Ti - Tl;
Chris@82 409 TC = FNMS(KP382683432, TB, KP923879532 * TA);
Chris@82 410 T1i = FMA(KP382683432, TA, KP923879532 * TB);
Chris@82 411 }
Chris@82 412 {
Chris@82 413 E T11, T1X, T1d, T1Y, T14, T20, T17, T21, T1a, T18;
Chris@82 414 {
Chris@82 415 E TZ, T10, T1b, T1c;
Chris@82 416 TZ = R1[WS(rs, 15)];
Chris@82 417 T10 = R1[WS(rs, 7)];
Chris@82 418 T11 = TZ - T10;
Chris@82 419 T1X = TZ + T10;
Chris@82 420 T1b = R1[WS(rs, 3)];
Chris@82 421 T1c = R1[WS(rs, 11)];
Chris@82 422 T1d = T1b - T1c;
Chris@82 423 T1Y = T1b + T1c;
Chris@82 424 }
Chris@82 425 {
Chris@82 426 E T12, T13, T15, T16;
Chris@82 427 T12 = R1[WS(rs, 1)];
Chris@82 428 T13 = R1[WS(rs, 9)];
Chris@82 429 T14 = T12 - T13;
Chris@82 430 T20 = T12 + T13;
Chris@82 431 T15 = R1[WS(rs, 13)];
Chris@82 432 T16 = R1[WS(rs, 5)];
Chris@82 433 T17 = T15 - T16;
Chris@82 434 T21 = T15 + T16;
Chris@82 435 }
Chris@82 436 T1Z = T1X + T1Y;
Chris@82 437 T22 = T20 + T21;
Chris@82 438 T2k = T21 - T20;
Chris@82 439 T2j = T1X - T1Y;
Chris@82 440 T1a = KP707106781 * (T17 - T14);
Chris@82 441 T1e = T1a - T1d;
Chris@82 442 T1C = T1d + T1a;
Chris@82 443 T18 = KP707106781 * (T14 + T17);
Chris@82 444 T19 = T11 + T18;
Chris@82 445 T1B = T11 - T18;
Chris@82 446 }
Chris@82 447 {
Chris@82 448 E TK, T1Q, TW, T1R, TN, T1T, TQ, T1U, TT, TR;
Chris@82 449 {
Chris@82 450 E TI, TJ, TU, TV;
Chris@82 451 TI = R1[0];
Chris@82 452 TJ = R1[WS(rs, 8)];
Chris@82 453 TK = TI - TJ;
Chris@82 454 T1Q = TI + TJ;
Chris@82 455 TU = R1[WS(rs, 4)];
Chris@82 456 TV = R1[WS(rs, 12)];
Chris@82 457 TW = TU - TV;
Chris@82 458 T1R = TU + TV;
Chris@82 459 }
Chris@82 460 {
Chris@82 461 E TL, TM, TO, TP;
Chris@82 462 TL = R1[WS(rs, 2)];
Chris@82 463 TM = R1[WS(rs, 10)];
Chris@82 464 TN = TL - TM;
Chris@82 465 T1T = TL + TM;
Chris@82 466 TO = R1[WS(rs, 14)];
Chris@82 467 TP = R1[WS(rs, 6)];
Chris@82 468 TQ = TO - TP;
Chris@82 469 T1U = TO + TP;
Chris@82 470 }
Chris@82 471 T1S = T1Q + T1R;
Chris@82 472 T1V = T1T + T1U;
Chris@82 473 T2h = T1U - T1T;
Chris@82 474 T2g = T1Q - T1R;
Chris@82 475 TT = KP707106781 * (TQ - TN);
Chris@82 476 TX = TT - TW;
Chris@82 477 T1z = TW + TT;
Chris@82 478 TR = KP707106781 * (TN + TQ);
Chris@82 479 TS = TK + TR;
Chris@82 480 T1y = TK - TR;
Chris@82 481 }
Chris@82 482 {
Chris@82 483 E Tf, Tu, T27, T28, T29, T2a;
Chris@82 484 Tf = T7 + Te;
Chris@82 485 Tu = Tm + Tt;
Chris@82 486 T27 = Tf + Tu;
Chris@82 487 T28 = T1S + T1V;
Chris@82 488 T29 = T1Z + T22;
Chris@82 489 T2a = T28 + T29;
Chris@82 490 Cr[WS(csr, 8)] = Tf - Tu;
Chris@82 491 Ci[WS(csi, 8)] = T29 - T28;
Chris@82 492 Cr[WS(csr, 16)] = T27 - T2a;
Chris@82 493 Cr[0] = T27 + T2a;
Chris@82 494 }
Chris@82 495 {
Chris@82 496 E T1P, T25, T24, T26, T1W, T23;
Chris@82 497 T1P = T7 - Te;
Chris@82 498 T25 = Tt - Tm;
Chris@82 499 T1W = T1S - T1V;
Chris@82 500 T23 = T1Z - T22;
Chris@82 501 T24 = KP707106781 * (T1W + T23);
Chris@82 502 T26 = KP707106781 * (T23 - T1W);
Chris@82 503 Cr[WS(csr, 12)] = T1P - T24;
Chris@82 504 Ci[WS(csi, 12)] = T26 - T25;
Chris@82 505 Cr[WS(csr, 4)] = T1P + T24;
Chris@82 506 Ci[WS(csi, 4)] = T25 + T26;
Chris@82 507 }
Chris@82 508 {
Chris@82 509 E T2f, T2v, T2p, T2r, T2m, T2q, T2u, T2w, T2e, T2n;
Chris@82 510 T2e = KP707106781 * (T2c + T2d);
Chris@82 511 T2f = T2b + T2e;
Chris@82 512 T2v = T2b - T2e;
Chris@82 513 T2n = KP707106781 * (T2d - T2c);
Chris@82 514 T2p = T2n - T2o;
Chris@82 515 T2r = T2o + T2n;
Chris@82 516 {
Chris@82 517 E T2i, T2l, T2s, T2t;
Chris@82 518 T2i = FMA(KP923879532, T2g, KP382683432 * T2h);
Chris@82 519 T2l = FNMS(KP382683432, T2k, KP923879532 * T2j);
Chris@82 520 T2m = T2i + T2l;
Chris@82 521 T2q = T2l - T2i;
Chris@82 522 T2s = FNMS(KP382683432, T2g, KP923879532 * T2h);
Chris@82 523 T2t = FMA(KP382683432, T2j, KP923879532 * T2k);
Chris@82 524 T2u = T2s + T2t;
Chris@82 525 T2w = T2t - T2s;
Chris@82 526 }
Chris@82 527 Cr[WS(csr, 14)] = T2f - T2m;
Chris@82 528 Ci[WS(csi, 14)] = T2u - T2r;
Chris@82 529 Cr[WS(csr, 2)] = T2f + T2m;
Chris@82 530 Ci[WS(csi, 2)] = T2r + T2u;
Chris@82 531 Ci[WS(csi, 6)] = T2p + T2q;
Chris@82 532 Cr[WS(csr, 6)] = T2v + T2w;
Chris@82 533 Ci[WS(csi, 10)] = T2q - T2p;
Chris@82 534 Cr[WS(csr, 10)] = T2v - T2w;
Chris@82 535 }
Chris@82 536 {
Chris@82 537 E TH, T1t, T1s, T1u, T1g, T1o, T1n, T1p;
Chris@82 538 {
Chris@82 539 E Tz, TG, T1q, T1r;
Chris@82 540 Tz = Tv + Ty;
Chris@82 541 TG = TC + TF;
Chris@82 542 TH = Tz + TG;
Chris@82 543 T1t = Tz - TG;
Chris@82 544 T1q = FNMS(KP195090322, TS, KP980785280 * TX);
Chris@82 545 T1r = FMA(KP195090322, T19, KP980785280 * T1e);
Chris@82 546 T1s = T1q + T1r;
Chris@82 547 T1u = T1r - T1q;
Chris@82 548 }
Chris@82 549 {
Chris@82 550 E TY, T1f, T1j, T1m;
Chris@82 551 TY = FMA(KP980785280, TS, KP195090322 * TX);
Chris@82 552 T1f = FNMS(KP195090322, T1e, KP980785280 * T19);
Chris@82 553 T1g = TY + T1f;
Chris@82 554 T1o = T1f - TY;
Chris@82 555 T1j = T1h - T1i;
Chris@82 556 T1m = T1k - T1l;
Chris@82 557 T1n = T1j - T1m;
Chris@82 558 T1p = T1m + T1j;
Chris@82 559 }
Chris@82 560 Cr[WS(csr, 15)] = TH - T1g;
Chris@82 561 Ci[WS(csi, 15)] = T1s - T1p;
Chris@82 562 Cr[WS(csr, 1)] = TH + T1g;
Chris@82 563 Ci[WS(csi, 1)] = T1p + T1s;
Chris@82 564 Ci[WS(csi, 7)] = T1n + T1o;
Chris@82 565 Cr[WS(csr, 7)] = T1t + T1u;
Chris@82 566 Ci[WS(csi, 9)] = T1o - T1n;
Chris@82 567 Cr[WS(csr, 9)] = T1t - T1u;
Chris@82 568 }
Chris@82 569 {
Chris@82 570 E T1x, T1N, T1M, T1O, T1E, T1I, T1H, T1J;
Chris@82 571 {
Chris@82 572 E T1v, T1w, T1K, T1L;
Chris@82 573 T1v = Tv - Ty;
Chris@82 574 T1w = T1i + T1h;
Chris@82 575 T1x = T1v + T1w;
Chris@82 576 T1N = T1v - T1w;
Chris@82 577 T1K = FNMS(KP555570233, T1y, KP831469612 * T1z);
Chris@82 578 T1L = FMA(KP555570233, T1B, KP831469612 * T1C);
Chris@82 579 T1M = T1K + T1L;
Chris@82 580 T1O = T1L - T1K;
Chris@82 581 }
Chris@82 582 {
Chris@82 583 E T1A, T1D, T1F, T1G;
Chris@82 584 T1A = FMA(KP831469612, T1y, KP555570233 * T1z);
Chris@82 585 T1D = FNMS(KP555570233, T1C, KP831469612 * T1B);
Chris@82 586 T1E = T1A + T1D;
Chris@82 587 T1I = T1D - T1A;
Chris@82 588 T1F = TF - TC;
Chris@82 589 T1G = T1l + T1k;
Chris@82 590 T1H = T1F - T1G;
Chris@82 591 T1J = T1G + T1F;
Chris@82 592 }
Chris@82 593 Cr[WS(csr, 13)] = T1x - T1E;
Chris@82 594 Ci[WS(csi, 13)] = T1M - T1J;
Chris@82 595 Cr[WS(csr, 3)] = T1x + T1E;
Chris@82 596 Ci[WS(csi, 3)] = T1J + T1M;
Chris@82 597 Ci[WS(csi, 5)] = T1H + T1I;
Chris@82 598 Cr[WS(csr, 5)] = T1N + T1O;
Chris@82 599 Ci[WS(csi, 11)] = T1I - T1H;
Chris@82 600 Cr[WS(csr, 11)] = T1N - T1O;
Chris@82 601 }
Chris@82 602 }
Chris@82 603 }
Chris@82 604 }
Chris@82 605
Chris@82 606 static const kr2c_desc desc = { 32, "r2cf_32", {140, 26, 16, 0}, &GENUS };
Chris@82 607
Chris@82 608 void X(codelet_r2cf_32) (planner *p) {
Chris@82 609 X(kr2c_register) (p, r2cf_32, &desc);
Chris@82 610 }
Chris@82 611
Chris@82 612 #endif