Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.8/rdft/scalar/r2cb/r2cb_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 08:07:29 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_r2cb.native -fma -compact -variables 4 -pipeline-latency 4 -sign 1 -n 32 -name r2cb_32 -include rdft/scalar/r2cb.h */
Chris@82 29
Chris@82 30 /*
Chris@82 31 * This function contains 156 FP additions, 84 FP multiplications,
Chris@82 32 * (or, 72 additions, 0 multiplications, 84 fused multiply/add),
Chris@82 33 * 54 stack variables, 9 constants, and 64 memory accesses
Chris@82 34 */
Chris@82 35 #include "rdft/scalar/r2cb.h"
Chris@82 36
Chris@82 37 static void r2cb_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(KP1_662939224, +1.662939224605090474157576755235811513477121624);
Chris@82 40 DK(KP668178637, +0.668178637919298919997757686523080761552472251);
Chris@82 41 DK(KP1_961570560, +1.961570560806460898252364472268478073947867462);
Chris@82 42 DK(KP198912367, +0.198912367379658006911597622644676228597850501);
Chris@82 43 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
Chris@82 44 DK(KP1_847759065, +1.847759065022573512256366378793576573644833252);
Chris@82 45 DK(KP414213562, +0.414213562373095048801688724209698078569671875);
Chris@82 46 DK(KP1_414213562, +1.414213562373095048801688724209698078569671875);
Chris@82 47 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
Chris@82 48 {
Chris@82 49 INT i;
Chris@82 50 for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(128, rs), MAKE_VOLATILE_STRIDE(128, csr), MAKE_VOLATILE_STRIDE(128, csi)) {
Chris@82 51 E T5, T1R, Tz, T1t, T8, T1S, TE, T1u, Tg, T1X, T2m, TK, TP, T1x, T1U;
Chris@82 52 E T1w, To, T28, T2p, TW, T1d, T1D, T20, T1A, Tv, T23, T2q, T25, T1g, T1B;
Chris@82 53 E T17, T1E;
Chris@82 54 {
Chris@82 55 E T4, Ty, T3, Tx, T1, T2;
Chris@82 56 T4 = Cr[WS(csr, 8)];
Chris@82 57 Ty = Ci[WS(csi, 8)];
Chris@82 58 T1 = Cr[0];
Chris@82 59 T2 = Cr[WS(csr, 16)];
Chris@82 60 T3 = T1 + T2;
Chris@82 61 Tx = T1 - T2;
Chris@82 62 T5 = FMA(KP2_000000000, T4, T3);
Chris@82 63 T1R = FNMS(KP2_000000000, T4, T3);
Chris@82 64 Tz = FNMS(KP2_000000000, Ty, Tx);
Chris@82 65 T1t = FMA(KP2_000000000, Ty, Tx);
Chris@82 66 }
Chris@82 67 {
Chris@82 68 E T6, T7, TA, TB, TC, TD;
Chris@82 69 T6 = Cr[WS(csr, 4)];
Chris@82 70 T7 = Cr[WS(csr, 12)];
Chris@82 71 TA = T6 - T7;
Chris@82 72 TB = Ci[WS(csi, 4)];
Chris@82 73 TC = Ci[WS(csi, 12)];
Chris@82 74 TD = TB + TC;
Chris@82 75 T8 = T6 + T7;
Chris@82 76 T1S = TB - TC;
Chris@82 77 TE = TA - TD;
Chris@82 78 T1u = TA + TD;
Chris@82 79 }
Chris@82 80 {
Chris@82 81 E Tc, TG, TO, T1V, Tf, TL, TJ, T1W;
Chris@82 82 {
Chris@82 83 E Ta, Tb, TM, TN;
Chris@82 84 Ta = Cr[WS(csr, 2)];
Chris@82 85 Tb = Cr[WS(csr, 14)];
Chris@82 86 Tc = Ta + Tb;
Chris@82 87 TG = Ta - Tb;
Chris@82 88 TM = Ci[WS(csi, 2)];
Chris@82 89 TN = Ci[WS(csi, 14)];
Chris@82 90 TO = TM + TN;
Chris@82 91 T1V = TM - TN;
Chris@82 92 }
Chris@82 93 {
Chris@82 94 E Td, Te, TH, TI;
Chris@82 95 Td = Cr[WS(csr, 10)];
Chris@82 96 Te = Cr[WS(csr, 6)];
Chris@82 97 Tf = Td + Te;
Chris@82 98 TL = Td - Te;
Chris@82 99 TH = Ci[WS(csi, 10)];
Chris@82 100 TI = Ci[WS(csi, 6)];
Chris@82 101 TJ = TH + TI;
Chris@82 102 T1W = TH - TI;
Chris@82 103 }
Chris@82 104 Tg = Tc + Tf;
Chris@82 105 T1X = T1V - T1W;
Chris@82 106 T2m = T1W + T1V;
Chris@82 107 TK = TG - TJ;
Chris@82 108 TP = TL + TO;
Chris@82 109 T1x = TG + TJ;
Chris@82 110 T1U = Tc - Tf;
Chris@82 111 T1w = TO - TL;
Chris@82 112 }
Chris@82 113 {
Chris@82 114 E Tk, TS, T1c, T26, Tn, T19, TV, T27;
Chris@82 115 {
Chris@82 116 E Ti, Tj, T1a, T1b;
Chris@82 117 Ti = Cr[WS(csr, 1)];
Chris@82 118 Tj = Cr[WS(csr, 15)];
Chris@82 119 Tk = Ti + Tj;
Chris@82 120 TS = Ti - Tj;
Chris@82 121 T1a = Ci[WS(csi, 1)];
Chris@82 122 T1b = Ci[WS(csi, 15)];
Chris@82 123 T1c = T1a + T1b;
Chris@82 124 T26 = T1a - T1b;
Chris@82 125 }
Chris@82 126 {
Chris@82 127 E Tl, Tm, TT, TU;
Chris@82 128 Tl = Cr[WS(csr, 9)];
Chris@82 129 Tm = Cr[WS(csr, 7)];
Chris@82 130 Tn = Tl + Tm;
Chris@82 131 T19 = Tl - Tm;
Chris@82 132 TT = Ci[WS(csi, 9)];
Chris@82 133 TU = Ci[WS(csi, 7)];
Chris@82 134 TV = TT + TU;
Chris@82 135 T27 = TT - TU;
Chris@82 136 }
Chris@82 137 To = Tk + Tn;
Chris@82 138 T28 = T26 - T27;
Chris@82 139 T2p = T27 + T26;
Chris@82 140 TW = TS - TV;
Chris@82 141 T1d = T19 + T1c;
Chris@82 142 T1D = T1c - T19;
Chris@82 143 T20 = Tk - Tn;
Chris@82 144 T1A = TS + TV;
Chris@82 145 }
Chris@82 146 {
Chris@82 147 E Tr, TX, T10, T22, Tu, T12, T15, T21;
Chris@82 148 {
Chris@82 149 E Tp, Tq, TY, TZ;
Chris@82 150 Tp = Cr[WS(csr, 5)];
Chris@82 151 Tq = Cr[WS(csr, 11)];
Chris@82 152 Tr = Tp + Tq;
Chris@82 153 TX = Tp - Tq;
Chris@82 154 TY = Ci[WS(csi, 5)];
Chris@82 155 TZ = Ci[WS(csi, 11)];
Chris@82 156 T10 = TY + TZ;
Chris@82 157 T22 = TY - TZ;
Chris@82 158 }
Chris@82 159 {
Chris@82 160 E Ts, Tt, T13, T14;
Chris@82 161 Ts = Cr[WS(csr, 3)];
Chris@82 162 Tt = Cr[WS(csr, 13)];
Chris@82 163 Tu = Ts + Tt;
Chris@82 164 T12 = Ts - Tt;
Chris@82 165 T13 = Ci[WS(csi, 3)];
Chris@82 166 T14 = Ci[WS(csi, 13)];
Chris@82 167 T15 = T13 + T14;
Chris@82 168 T21 = T14 - T13;
Chris@82 169 }
Chris@82 170 Tv = Tr + Tu;
Chris@82 171 T23 = T21 - T22;
Chris@82 172 T2q = T22 + T21;
Chris@82 173 T25 = Tr - Tu;
Chris@82 174 {
Chris@82 175 E T1e, T1f, T11, T16;
Chris@82 176 T1e = TX + T10;
Chris@82 177 T1f = T12 + T15;
Chris@82 178 T1g = T1e - T1f;
Chris@82 179 T1B = T1e + T1f;
Chris@82 180 T11 = TX - T10;
Chris@82 181 T16 = T12 - T15;
Chris@82 182 T17 = T11 + T16;
Chris@82 183 T1E = T16 - T11;
Chris@82 184 }
Chris@82 185 }
Chris@82 186 {
Chris@82 187 E Tw, T2w, Th, T2v, T9;
Chris@82 188 Tw = To + Tv;
Chris@82 189 T2w = T2q + T2p;
Chris@82 190 T9 = FMA(KP2_000000000, T8, T5);
Chris@82 191 Th = FMA(KP2_000000000, Tg, T9);
Chris@82 192 T2v = FNMS(KP2_000000000, Tg, T9);
Chris@82 193 R0[WS(rs, 8)] = FNMS(KP2_000000000, Tw, Th);
Chris@82 194 R0[WS(rs, 12)] = FMA(KP2_000000000, T2w, T2v);
Chris@82 195 R0[0] = FMA(KP2_000000000, Tw, Th);
Chris@82 196 R0[WS(rs, 4)] = FNMS(KP2_000000000, T2w, T2v);
Chris@82 197 }
Chris@82 198 {
Chris@82 199 E T2n, T2t, T2s, T2u, T2l, T2o, T2r;
Chris@82 200 T2l = FNMS(KP2_000000000, T8, T5);
Chris@82 201 T2n = FNMS(KP2_000000000, T2m, T2l);
Chris@82 202 T2t = FMA(KP2_000000000, T2m, T2l);
Chris@82 203 T2o = To - Tv;
Chris@82 204 T2r = T2p - T2q;
Chris@82 205 T2s = T2o - T2r;
Chris@82 206 T2u = T2o + T2r;
Chris@82 207 R0[WS(rs, 10)] = FNMS(KP1_414213562, T2s, T2n);
Chris@82 208 R0[WS(rs, 14)] = FMA(KP1_414213562, T2u, T2t);
Chris@82 209 R0[WS(rs, 2)] = FMA(KP1_414213562, T2s, T2n);
Chris@82 210 R0[WS(rs, 6)] = FNMS(KP1_414213562, T2u, T2t);
Chris@82 211 }
Chris@82 212 {
Chris@82 213 E TR, T1j, T1i, T1k;
Chris@82 214 {
Chris@82 215 E TF, TQ, T18, T1h;
Chris@82 216 TF = FMA(KP1_414213562, TE, Tz);
Chris@82 217 TQ = FNMS(KP414213562, TP, TK);
Chris@82 218 TR = FMA(KP1_847759065, TQ, TF);
Chris@82 219 T1j = FNMS(KP1_847759065, TQ, TF);
Chris@82 220 T18 = FMA(KP707106781, T17, TW);
Chris@82 221 T1h = FMA(KP707106781, T1g, T1d);
Chris@82 222 T1i = FNMS(KP198912367, T1h, T18);
Chris@82 223 T1k = FMA(KP198912367, T18, T1h);
Chris@82 224 }
Chris@82 225 R1[WS(rs, 8)] = FNMS(KP1_961570560, T1i, TR);
Chris@82 226 R1[WS(rs, 12)] = FMA(KP1_961570560, T1k, T1j);
Chris@82 227 R1[0] = FMA(KP1_961570560, T1i, TR);
Chris@82 228 R1[WS(rs, 4)] = FNMS(KP1_961570560, T1k, T1j);
Chris@82 229 }
Chris@82 230 {
Chris@82 231 E T2f, T2j, T2i, T2k;
Chris@82 232 {
Chris@82 233 E T2d, T2e, T2g, T2h;
Chris@82 234 T2d = FMA(KP2_000000000, T1S, T1R);
Chris@82 235 T2e = T1U + T1X;
Chris@82 236 T2f = FNMS(KP1_414213562, T2e, T2d);
Chris@82 237 T2j = FMA(KP1_414213562, T2e, T2d);
Chris@82 238 T2g = T28 - T25;
Chris@82 239 T2h = T20 - T23;
Chris@82 240 T2i = FNMS(KP414213562, T2h, T2g);
Chris@82 241 T2k = FMA(KP414213562, T2g, T2h);
Chris@82 242 }
Chris@82 243 R0[WS(rs, 3)] = FNMS(KP1_847759065, T2i, T2f);
Chris@82 244 R0[WS(rs, 15)] = FMA(KP1_847759065, T2k, T2j);
Chris@82 245 R0[WS(rs, 11)] = FMA(KP1_847759065, T2i, T2f);
Chris@82 246 R0[WS(rs, 7)] = FNMS(KP1_847759065, T2k, T2j);
Chris@82 247 }
Chris@82 248 {
Chris@82 249 E T1n, T1r, T1q, T1s;
Chris@82 250 {
Chris@82 251 E T1l, T1m, T1o, T1p;
Chris@82 252 T1l = FNMS(KP1_414213562, TE, Tz);
Chris@82 253 T1m = FMA(KP414213562, TK, TP);
Chris@82 254 T1n = FNMS(KP1_847759065, T1m, T1l);
Chris@82 255 T1r = FMA(KP1_847759065, T1m, T1l);
Chris@82 256 T1o = FNMS(KP707106781, T1g, T1d);
Chris@82 257 T1p = FNMS(KP707106781, T17, TW);
Chris@82 258 T1q = FNMS(KP668178637, T1p, T1o);
Chris@82 259 T1s = FMA(KP668178637, T1o, T1p);
Chris@82 260 }
Chris@82 261 R1[WS(rs, 2)] = FNMS(KP1_662939224, T1q, T1n);
Chris@82 262 R1[WS(rs, 14)] = FMA(KP1_662939224, T1s, T1r);
Chris@82 263 R1[WS(rs, 10)] = FMA(KP1_662939224, T1q, T1n);
Chris@82 264 R1[WS(rs, 6)] = FNMS(KP1_662939224, T1s, T1r);
Chris@82 265 }
Chris@82 266 {
Chris@82 267 E T1L, T1P, T1O, T1Q;
Chris@82 268 {
Chris@82 269 E T1J, T1K, T1M, T1N;
Chris@82 270 T1J = FMA(KP1_414213562, T1u, T1t);
Chris@82 271 T1K = FMA(KP414213562, T1w, T1x);
Chris@82 272 T1L = FNMS(KP1_847759065, T1K, T1J);
Chris@82 273 T1P = FMA(KP1_847759065, T1K, T1J);
Chris@82 274 T1M = FMA(KP707106781, T1E, T1D);
Chris@82 275 T1N = FMA(KP707106781, T1B, T1A);
Chris@82 276 T1O = FNMS(KP198912367, T1N, T1M);
Chris@82 277 T1Q = FMA(KP198912367, T1M, T1N);
Chris@82 278 }
Chris@82 279 R1[WS(rs, 3)] = FNMS(KP1_961570560, T1O, T1L);
Chris@82 280 R1[WS(rs, 15)] = FMA(KP1_961570560, T1Q, T1P);
Chris@82 281 R1[WS(rs, 11)] = FMA(KP1_961570560, T1O, T1L);
Chris@82 282 R1[WS(rs, 7)] = FNMS(KP1_961570560, T1Q, T1P);
Chris@82 283 }
Chris@82 284 {
Chris@82 285 E T1Z, T2b, T2a, T2c;
Chris@82 286 {
Chris@82 287 E T1T, T1Y, T24, T29;
Chris@82 288 T1T = FNMS(KP2_000000000, T1S, T1R);
Chris@82 289 T1Y = T1U - T1X;
Chris@82 290 T1Z = FMA(KP1_414213562, T1Y, T1T);
Chris@82 291 T2b = FNMS(KP1_414213562, T1Y, T1T);
Chris@82 292 T24 = T20 + T23;
Chris@82 293 T29 = T25 + T28;
Chris@82 294 T2a = FNMS(KP414213562, T29, T24);
Chris@82 295 T2c = FMA(KP414213562, T24, T29);
Chris@82 296 }
Chris@82 297 R0[WS(rs, 9)] = FNMS(KP1_847759065, T2a, T1Z);
Chris@82 298 R0[WS(rs, 13)] = FMA(KP1_847759065, T2c, T2b);
Chris@82 299 R0[WS(rs, 1)] = FMA(KP1_847759065, T2a, T1Z);
Chris@82 300 R0[WS(rs, 5)] = FNMS(KP1_847759065, T2c, T2b);
Chris@82 301 }
Chris@82 302 {
Chris@82 303 E T1z, T1H, T1G, T1I;
Chris@82 304 {
Chris@82 305 E T1v, T1y, T1C, T1F;
Chris@82 306 T1v = FNMS(KP1_414213562, T1u, T1t);
Chris@82 307 T1y = FNMS(KP414213562, T1x, T1w);
Chris@82 308 T1z = FNMS(KP1_847759065, T1y, T1v);
Chris@82 309 T1H = FMA(KP1_847759065, T1y, T1v);
Chris@82 310 T1C = FNMS(KP707106781, T1B, T1A);
Chris@82 311 T1F = FNMS(KP707106781, T1E, T1D);
Chris@82 312 T1G = FNMS(KP668178637, T1F, T1C);
Chris@82 313 T1I = FMA(KP668178637, T1C, T1F);
Chris@82 314 }
Chris@82 315 R1[WS(rs, 9)] = FNMS(KP1_662939224, T1G, T1z);
Chris@82 316 R1[WS(rs, 13)] = FMA(KP1_662939224, T1I, T1H);
Chris@82 317 R1[WS(rs, 1)] = FMA(KP1_662939224, T1G, T1z);
Chris@82 318 R1[WS(rs, 5)] = FNMS(KP1_662939224, T1I, T1H);
Chris@82 319 }
Chris@82 320 }
Chris@82 321 }
Chris@82 322 }
Chris@82 323
Chris@82 324 static const kr2c_desc desc = { 32, "r2cb_32", {72, 0, 84, 0}, &GENUS };
Chris@82 325
Chris@82 326 void X(codelet_r2cb_32) (planner *p) {
Chris@82 327 X(kr2c_register) (p, r2cb_32, &desc);
Chris@82 328 }
Chris@82 329
Chris@82 330 #else
Chris@82 331
Chris@82 332 /* Generated by: ../../../genfft/gen_r2cb.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 32 -name r2cb_32 -include rdft/scalar/r2cb.h */
Chris@82 333
Chris@82 334 /*
Chris@82 335 * This function contains 156 FP additions, 50 FP multiplications,
Chris@82 336 * (or, 140 additions, 34 multiplications, 16 fused multiply/add),
Chris@82 337 * 54 stack variables, 9 constants, and 64 memory accesses
Chris@82 338 */
Chris@82 339 #include "rdft/scalar/r2cb.h"
Chris@82 340
Chris@82 341 static void r2cb_32(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
Chris@82 342 {
Chris@82 343 DK(KP1_662939224, +1.662939224605090474157576755235811513477121624);
Chris@82 344 DK(KP1_111140466, +1.111140466039204449485661627897065748749874382);
Chris@82 345 DK(KP1_961570560, +1.961570560806460898252364472268478073947867462);
Chris@82 346 DK(KP390180644, +0.390180644032256535696569736954044481855383236);
Chris@82 347 DK(KP765366864, +0.765366864730179543456919968060797733522689125);
Chris@82 348 DK(KP1_847759065, +1.847759065022573512256366378793576573644833252);
Chris@82 349 DK(KP707106781, +0.707106781186547524400844362104849039284835938);
Chris@82 350 DK(KP1_414213562, +1.414213562373095048801688724209698078569671875);
Chris@82 351 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
Chris@82 352 {
Chris@82 353 INT i;
Chris@82 354 for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(128, rs), MAKE_VOLATILE_STRIDE(128, csr), MAKE_VOLATILE_STRIDE(128, csi)) {
Chris@82 355 E T9, T2c, TB, T1y, T6, T2b, Ty, T1v, Th, T2e, T2f, TD, TK, T1C, T1F;
Chris@82 356 E T1h, Tp, T2i, T2m, TN, T13, T1K, T1Y, T1k, Tw, TU, T1l, TW, T1V, T2j;
Chris@82 357 E T1R, T2l;
Chris@82 358 {
Chris@82 359 E T7, T8, T1w, Tz, TA, T1x;
Chris@82 360 T7 = Cr[WS(csr, 4)];
Chris@82 361 T8 = Cr[WS(csr, 12)];
Chris@82 362 T1w = T7 - T8;
Chris@82 363 Tz = Ci[WS(csi, 4)];
Chris@82 364 TA = Ci[WS(csi, 12)];
Chris@82 365 T1x = Tz + TA;
Chris@82 366 T9 = KP2_000000000 * (T7 + T8);
Chris@82 367 T2c = KP1_414213562 * (T1w + T1x);
Chris@82 368 TB = KP2_000000000 * (Tz - TA);
Chris@82 369 T1y = KP1_414213562 * (T1w - T1x);
Chris@82 370 }
Chris@82 371 {
Chris@82 372 E T5, T1u, T3, T1s;
Chris@82 373 {
Chris@82 374 E T4, T1t, T1, T2;
Chris@82 375 T4 = Cr[WS(csr, 8)];
Chris@82 376 T5 = KP2_000000000 * T4;
Chris@82 377 T1t = Ci[WS(csi, 8)];
Chris@82 378 T1u = KP2_000000000 * T1t;
Chris@82 379 T1 = Cr[0];
Chris@82 380 T2 = Cr[WS(csr, 16)];
Chris@82 381 T3 = T1 + T2;
Chris@82 382 T1s = T1 - T2;
Chris@82 383 }
Chris@82 384 T6 = T3 + T5;
Chris@82 385 T2b = T1s + T1u;
Chris@82 386 Ty = T3 - T5;
Chris@82 387 T1v = T1s - T1u;
Chris@82 388 }
Chris@82 389 {
Chris@82 390 E Td, T1A, TG, T1E, Tg, T1D, TJ, T1B;
Chris@82 391 {
Chris@82 392 E Tb, Tc, TE, TF;
Chris@82 393 Tb = Cr[WS(csr, 2)];
Chris@82 394 Tc = Cr[WS(csr, 14)];
Chris@82 395 Td = Tb + Tc;
Chris@82 396 T1A = Tb - Tc;
Chris@82 397 TE = Ci[WS(csi, 2)];
Chris@82 398 TF = Ci[WS(csi, 14)];
Chris@82 399 TG = TE - TF;
Chris@82 400 T1E = TE + TF;
Chris@82 401 }
Chris@82 402 {
Chris@82 403 E Te, Tf, TH, TI;
Chris@82 404 Te = Cr[WS(csr, 10)];
Chris@82 405 Tf = Cr[WS(csr, 6)];
Chris@82 406 Tg = Te + Tf;
Chris@82 407 T1D = Te - Tf;
Chris@82 408 TH = Ci[WS(csi, 10)];
Chris@82 409 TI = Ci[WS(csi, 6)];
Chris@82 410 TJ = TH - TI;
Chris@82 411 T1B = TH + TI;
Chris@82 412 }
Chris@82 413 Th = KP2_000000000 * (Td + Tg);
Chris@82 414 T2e = T1A + T1B;
Chris@82 415 T2f = T1E - T1D;
Chris@82 416 TD = Td - Tg;
Chris@82 417 TK = TG - TJ;
Chris@82 418 T1C = T1A - T1B;
Chris@82 419 T1F = T1D + T1E;
Chris@82 420 T1h = KP2_000000000 * (TJ + TG);
Chris@82 421 }
Chris@82 422 {
Chris@82 423 E Tl, T1I, TZ, T1X, To, T1W, T12, T1J;
Chris@82 424 {
Chris@82 425 E Tj, Tk, TX, TY;
Chris@82 426 Tj = Cr[WS(csr, 1)];
Chris@82 427 Tk = Cr[WS(csr, 15)];
Chris@82 428 Tl = Tj + Tk;
Chris@82 429 T1I = Tj - Tk;
Chris@82 430 TX = Ci[WS(csi, 1)];
Chris@82 431 TY = Ci[WS(csi, 15)];
Chris@82 432 TZ = TX - TY;
Chris@82 433 T1X = TX + TY;
Chris@82 434 }
Chris@82 435 {
Chris@82 436 E Tm, Tn, T10, T11;
Chris@82 437 Tm = Cr[WS(csr, 9)];
Chris@82 438 Tn = Cr[WS(csr, 7)];
Chris@82 439 To = Tm + Tn;
Chris@82 440 T1W = Tm - Tn;
Chris@82 441 T10 = Ci[WS(csi, 9)];
Chris@82 442 T11 = Ci[WS(csi, 7)];
Chris@82 443 T12 = T10 - T11;
Chris@82 444 T1J = T10 + T11;
Chris@82 445 }
Chris@82 446 Tp = Tl + To;
Chris@82 447 T2i = T1I + T1J;
Chris@82 448 T2m = T1X - T1W;
Chris@82 449 TN = Tl - To;
Chris@82 450 T13 = TZ - T12;
Chris@82 451 T1K = T1I - T1J;
Chris@82 452 T1Y = T1W + T1X;
Chris@82 453 T1k = T12 + TZ;
Chris@82 454 }
Chris@82 455 {
Chris@82 456 E Ts, T1L, TT, T1M, Tv, T1O, TQ, T1P;
Chris@82 457 {
Chris@82 458 E Tq, Tr, TR, TS;
Chris@82 459 Tq = Cr[WS(csr, 5)];
Chris@82 460 Tr = Cr[WS(csr, 11)];
Chris@82 461 Ts = Tq + Tr;
Chris@82 462 T1L = Tq - Tr;
Chris@82 463 TR = Ci[WS(csi, 5)];
Chris@82 464 TS = Ci[WS(csi, 11)];
Chris@82 465 TT = TR - TS;
Chris@82 466 T1M = TR + TS;
Chris@82 467 }
Chris@82 468 {
Chris@82 469 E Tt, Tu, TO, TP;
Chris@82 470 Tt = Cr[WS(csr, 3)];
Chris@82 471 Tu = Cr[WS(csr, 13)];
Chris@82 472 Tv = Tt + Tu;
Chris@82 473 T1O = Tt - Tu;
Chris@82 474 TO = Ci[WS(csi, 13)];
Chris@82 475 TP = Ci[WS(csi, 3)];
Chris@82 476 TQ = TO - TP;
Chris@82 477 T1P = TP + TO;
Chris@82 478 }
Chris@82 479 Tw = Ts + Tv;
Chris@82 480 TU = TQ - TT;
Chris@82 481 T1l = TT + TQ;
Chris@82 482 TW = Ts - Tv;
Chris@82 483 {
Chris@82 484 E T1T, T1U, T1N, T1Q;
Chris@82 485 T1T = T1L + T1M;
Chris@82 486 T1U = T1O + T1P;
Chris@82 487 T1V = KP707106781 * (T1T - T1U);
Chris@82 488 T2j = KP707106781 * (T1T + T1U);
Chris@82 489 T1N = T1L - T1M;
Chris@82 490 T1Q = T1O - T1P;
Chris@82 491 T1R = KP707106781 * (T1N + T1Q);
Chris@82 492 T2l = KP707106781 * (T1N - T1Q);
Chris@82 493 }
Chris@82 494 }
Chris@82 495 {
Chris@82 496 E Tx, T1r, Ti, T1q, Ta;
Chris@82 497 Tx = KP2_000000000 * (Tp + Tw);
Chris@82 498 T1r = KP2_000000000 * (T1l + T1k);
Chris@82 499 Ta = T6 + T9;
Chris@82 500 Ti = Ta + Th;
Chris@82 501 T1q = Ta - Th;
Chris@82 502 R0[WS(rs, 8)] = Ti - Tx;
Chris@82 503 R0[WS(rs, 12)] = T1q + T1r;
Chris@82 504 R0[0] = Ti + Tx;
Chris@82 505 R0[WS(rs, 4)] = T1q - T1r;
Chris@82 506 }
Chris@82 507 {
Chris@82 508 E T1i, T1o, T1n, T1p, T1g, T1j, T1m;
Chris@82 509 T1g = T6 - T9;
Chris@82 510 T1i = T1g - T1h;
Chris@82 511 T1o = T1g + T1h;
Chris@82 512 T1j = Tp - Tw;
Chris@82 513 T1m = T1k - T1l;
Chris@82 514 T1n = KP1_414213562 * (T1j - T1m);
Chris@82 515 T1p = KP1_414213562 * (T1j + T1m);
Chris@82 516 R0[WS(rs, 10)] = T1i - T1n;
Chris@82 517 R0[WS(rs, 14)] = T1o + T1p;
Chris@82 518 R0[WS(rs, 2)] = T1i + T1n;
Chris@82 519 R0[WS(rs, 6)] = T1o - T1p;
Chris@82 520 }
Chris@82 521 {
Chris@82 522 E TM, T16, T15, T17;
Chris@82 523 {
Chris@82 524 E TC, TL, TV, T14;
Chris@82 525 TC = Ty - TB;
Chris@82 526 TL = KP1_414213562 * (TD - TK);
Chris@82 527 TM = TC + TL;
Chris@82 528 T16 = TC - TL;
Chris@82 529 TV = TN + TU;
Chris@82 530 T14 = TW + T13;
Chris@82 531 T15 = FNMS(KP765366864, T14, KP1_847759065 * TV);
Chris@82 532 T17 = FMA(KP765366864, TV, KP1_847759065 * T14);
Chris@82 533 }
Chris@82 534 R0[WS(rs, 9)] = TM - T15;
Chris@82 535 R0[WS(rs, 13)] = T16 + T17;
Chris@82 536 R0[WS(rs, 1)] = TM + T15;
Chris@82 537 R0[WS(rs, 5)] = T16 - T17;
Chris@82 538 }
Chris@82 539 {
Chris@82 540 E T2t, T2x, T2w, T2y;
Chris@82 541 {
Chris@82 542 E T2r, T2s, T2u, T2v;
Chris@82 543 T2r = T2b + T2c;
Chris@82 544 T2s = FMA(KP1_847759065, T2e, KP765366864 * T2f);
Chris@82 545 T2t = T2r - T2s;
Chris@82 546 T2x = T2r + T2s;
Chris@82 547 T2u = T2i + T2j;
Chris@82 548 T2v = T2m - T2l;
Chris@82 549 T2w = FNMS(KP1_961570560, T2v, KP390180644 * T2u);
Chris@82 550 T2y = FMA(KP1_961570560, T2u, KP390180644 * T2v);
Chris@82 551 }
Chris@82 552 R1[WS(rs, 11)] = T2t - T2w;
Chris@82 553 R1[WS(rs, 15)] = T2x + T2y;
Chris@82 554 R1[WS(rs, 3)] = T2t + T2w;
Chris@82 555 R1[WS(rs, 7)] = T2x - T2y;
Chris@82 556 }
Chris@82 557 {
Chris@82 558 E T1a, T1e, T1d, T1f;
Chris@82 559 {
Chris@82 560 E T18, T19, T1b, T1c;
Chris@82 561 T18 = Ty + TB;
Chris@82 562 T19 = KP1_414213562 * (TD + TK);
Chris@82 563 T1a = T18 - T19;
Chris@82 564 T1e = T18 + T19;
Chris@82 565 T1b = TN - TU;
Chris@82 566 T1c = T13 - TW;
Chris@82 567 T1d = FNMS(KP1_847759065, T1c, KP765366864 * T1b);
Chris@82 568 T1f = FMA(KP1_847759065, T1b, KP765366864 * T1c);
Chris@82 569 }
Chris@82 570 R0[WS(rs, 11)] = T1a - T1d;
Chris@82 571 R0[WS(rs, 15)] = T1e + T1f;
Chris@82 572 R0[WS(rs, 3)] = T1a + T1d;
Chris@82 573 R0[WS(rs, 7)] = T1e - T1f;
Chris@82 574 }
Chris@82 575 {
Chris@82 576 E T25, T29, T28, T2a;
Chris@82 577 {
Chris@82 578 E T23, T24, T26, T27;
Chris@82 579 T23 = T1v - T1y;
Chris@82 580 T24 = FMA(KP765366864, T1C, KP1_847759065 * T1F);
Chris@82 581 T25 = T23 - T24;
Chris@82 582 T29 = T23 + T24;
Chris@82 583 T26 = T1K - T1R;
Chris@82 584 T27 = T1Y - T1V;
Chris@82 585 T28 = FNMS(KP1_662939224, T27, KP1_111140466 * T26);
Chris@82 586 T2a = FMA(KP1_662939224, T26, KP1_111140466 * T27);
Chris@82 587 }
Chris@82 588 R1[WS(rs, 10)] = T25 - T28;
Chris@82 589 R1[WS(rs, 14)] = T29 + T2a;
Chris@82 590 R1[WS(rs, 2)] = T25 + T28;
Chris@82 591 R1[WS(rs, 6)] = T29 - T2a;
Chris@82 592 }
Chris@82 593 {
Chris@82 594 E T2h, T2p, T2o, T2q;
Chris@82 595 {
Chris@82 596 E T2d, T2g, T2k, T2n;
Chris@82 597 T2d = T2b - T2c;
Chris@82 598 T2g = FNMS(KP1_847759065, T2f, KP765366864 * T2e);
Chris@82 599 T2h = T2d + T2g;
Chris@82 600 T2p = T2d - T2g;
Chris@82 601 T2k = T2i - T2j;
Chris@82 602 T2n = T2l + T2m;
Chris@82 603 T2o = FNMS(KP1_111140466, T2n, KP1_662939224 * T2k);
Chris@82 604 T2q = FMA(KP1_111140466, T2k, KP1_662939224 * T2n);
Chris@82 605 }
Chris@82 606 R1[WS(rs, 9)] = T2h - T2o;
Chris@82 607 R1[WS(rs, 13)] = T2p + T2q;
Chris@82 608 R1[WS(rs, 1)] = T2h + T2o;
Chris@82 609 R1[WS(rs, 5)] = T2p - T2q;
Chris@82 610 }
Chris@82 611 {
Chris@82 612 E T1H, T21, T20, T22;
Chris@82 613 {
Chris@82 614 E T1z, T1G, T1S, T1Z;
Chris@82 615 T1z = T1v + T1y;
Chris@82 616 T1G = FNMS(KP765366864, T1F, KP1_847759065 * T1C);
Chris@82 617 T1H = T1z + T1G;
Chris@82 618 T21 = T1z - T1G;
Chris@82 619 T1S = T1K + T1R;
Chris@82 620 T1Z = T1V + T1Y;
Chris@82 621 T20 = FNMS(KP390180644, T1Z, KP1_961570560 * T1S);
Chris@82 622 T22 = FMA(KP390180644, T1S, KP1_961570560 * T1Z);
Chris@82 623 }
Chris@82 624 R1[WS(rs, 8)] = T1H - T20;
Chris@82 625 R1[WS(rs, 12)] = T21 + T22;
Chris@82 626 R1[0] = T1H + T20;
Chris@82 627 R1[WS(rs, 4)] = T21 - T22;
Chris@82 628 }
Chris@82 629 }
Chris@82 630 }
Chris@82 631 }
Chris@82 632
Chris@82 633 static const kr2c_desc desc = { 32, "r2cb_32", {140, 34, 16, 0}, &GENUS };
Chris@82 634
Chris@82 635 void X(codelet_r2cb_32) (planner *p) {
Chris@82 636 X(kr2c_register) (p, r2cb_32, &desc);
Chris@82 637 }
Chris@82 638
Chris@82 639 #endif