Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.5/rdft/scalar/r2cb/r2cb_20.c @ 56:af97cad61ff0

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Add updated build of PortAudio for OSX
author Chris Cannam <cannam@all-day-breakfast.com>
date Tue, 03 Jan 2017 15:10:52 +0000
parents 2cd0e3b3e1fd
children
rev   line source
Chris@42 1 /*
Chris@42 2 * Copyright (c) 2003, 2007-14 Matteo Frigo
Chris@42 3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
Chris@42 4 *
Chris@42 5 * This program is free software; you can redistribute it and/or modify
Chris@42 6 * it under the terms of the GNU General Public License as published by
Chris@42 7 * the Free Software Foundation; either version 2 of the License, or
Chris@42 8 * (at your option) any later version.
Chris@42 9 *
Chris@42 10 * This program is distributed in the hope that it will be useful,
Chris@42 11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
Chris@42 12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
Chris@42 13 * GNU General Public License for more details.
Chris@42 14 *
Chris@42 15 * You should have received a copy of the GNU General Public License
Chris@42 16 * along with this program; if not, write to the Free Software
Chris@42 17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
Chris@42 18 *
Chris@42 19 */
Chris@42 20
Chris@42 21 /* This file was automatically generated --- DO NOT EDIT */
Chris@42 22 /* Generated on Sat Jul 30 16:49:33 EDT 2016 */
Chris@42 23
Chris@42 24 #include "codelet-rdft.h"
Chris@42 25
Chris@42 26 #ifdef HAVE_FMA
Chris@42 27
Chris@42 28 /* Generated by: ../../../genfft/gen_r2cb.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -sign 1 -n 20 -name r2cb_20 -include r2cb.h */
Chris@42 29
Chris@42 30 /*
Chris@42 31 * This function contains 86 FP additions, 44 FP multiplications,
Chris@42 32 * (or, 42 additions, 0 multiplications, 44 fused multiply/add),
Chris@42 33 * 69 stack variables, 5 constants, and 40 memory accesses
Chris@42 34 */
Chris@42 35 #include "r2cb.h"
Chris@42 36
Chris@42 37 static void r2cb_20(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
Chris@42 38 {
Chris@42 39 DK(KP1_902113032, +1.902113032590307144232878666758764286811397268);
Chris@42 40 DK(KP1_118033988, +1.118033988749894848204586834365638117720309180);
Chris@42 41 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
Chris@42 42 DK(KP618033988, +0.618033988749894848204586834365638117720309180);
Chris@42 43 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
Chris@42 44 {
Chris@42 45 INT i;
Chris@42 46 for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(80, rs), MAKE_VOLATILE_STRIDE(80, csr), MAKE_VOLATILE_STRIDE(80, csi)) {
Chris@42 47 E TY, T1o, T1m, T14, T12, TX, T1n, T1j, TZ, T13;
Chris@42 48 {
Chris@42 49 E Tr, TD, Tl, T5, T1a, T1l, T1d, T1k, TT, T10, TO, T11, TE, TF, Tk;
Chris@42 50 E TI, TC, T1i, To, TG, T16;
Chris@42 51 {
Chris@42 52 E T4, Tq, T1, T2;
Chris@42 53 T4 = Cr[WS(csr, 5)];
Chris@42 54 Tq = Ci[WS(csi, 5)];
Chris@42 55 T1 = Cr[0];
Chris@42 56 T2 = Cr[WS(csr, 10)];
Chris@42 57 {
Chris@42 58 E Ts, T8, T19, TR, T18, Tb, TS, Tv, Tx, Tf, Ty, T1c, TM, T1b, Ti;
Chris@42 59 E Tz, Tt, Tu, TN, TA;
Chris@42 60 {
Chris@42 61 E TP, TQ, T9, Ta;
Chris@42 62 {
Chris@42 63 E T6, T7, Tp, T3;
Chris@42 64 T6 = Cr[WS(csr, 4)];
Chris@42 65 T7 = Cr[WS(csr, 6)];
Chris@42 66 TP = Ci[WS(csi, 4)];
Chris@42 67 Tp = T1 - T2;
Chris@42 68 T3 = T1 + T2;
Chris@42 69 Ts = T6 - T7;
Chris@42 70 T8 = T6 + T7;
Chris@42 71 Tr = FMA(KP2_000000000, Tq, Tp);
Chris@42 72 TD = FNMS(KP2_000000000, Tq, Tp);
Chris@42 73 Tl = FMA(KP2_000000000, T4, T3);
Chris@42 74 T5 = FNMS(KP2_000000000, T4, T3);
Chris@42 75 TQ = Ci[WS(csi, 6)];
Chris@42 76 }
Chris@42 77 T9 = Cr[WS(csr, 9)];
Chris@42 78 Ta = Cr[WS(csr, 1)];
Chris@42 79 Tt = Ci[WS(csi, 9)];
Chris@42 80 T19 = TP + TQ;
Chris@42 81 TR = TP - TQ;
Chris@42 82 T18 = T9 - Ta;
Chris@42 83 Tb = T9 + Ta;
Chris@42 84 Tu = Ci[WS(csi, 1)];
Chris@42 85 }
Chris@42 86 {
Chris@42 87 E TK, TL, Td, Te, Tg, Th;
Chris@42 88 Td = Cr[WS(csr, 8)];
Chris@42 89 Te = Cr[WS(csr, 2)];
Chris@42 90 TK = Ci[WS(csi, 8)];
Chris@42 91 TS = Tt - Tu;
Chris@42 92 Tv = Tt + Tu;
Chris@42 93 Tx = Td - Te;
Chris@42 94 Tf = Td + Te;
Chris@42 95 TL = Ci[WS(csi, 2)];
Chris@42 96 Tg = Cr[WS(csr, 7)];
Chris@42 97 Th = Cr[WS(csr, 3)];
Chris@42 98 Ty = Ci[WS(csi, 7)];
Chris@42 99 T1c = TK + TL;
Chris@42 100 TM = TK - TL;
Chris@42 101 T1b = Tg - Th;
Chris@42 102 Ti = Tg + Th;
Chris@42 103 Tz = Ci[WS(csi, 3)];
Chris@42 104 }
Chris@42 105 T1a = T18 + T19;
Chris@42 106 T1l = T19 - T18;
Chris@42 107 T1d = T1b + T1c;
Chris@42 108 T1k = T1c - T1b;
Chris@42 109 TT = TR - TS;
Chris@42 110 T10 = TS + TR;
Chris@42 111 TN = Tz - Ty;
Chris@42 112 TA = Ty + Tz;
Chris@42 113 TO = TM - TN;
Chris@42 114 T11 = TN + TM;
Chris@42 115 {
Chris@42 116 E Tm, Tc, Tj, Tn, Tw, TB;
Chris@42 117 Tm = T8 + Tb;
Chris@42 118 Tc = T8 - Tb;
Chris@42 119 Tj = Tf - Ti;
Chris@42 120 Tn = Tf + Ti;
Chris@42 121 TE = Ts - Tv;
Chris@42 122 Tw = Ts + Tv;
Chris@42 123 TB = Tx - TA;
Chris@42 124 TF = Tx + TA;
Chris@42 125 Tk = Tc + Tj;
Chris@42 126 TI = Tc - Tj;
Chris@42 127 TC = Tw + TB;
Chris@42 128 T1i = Tw - TB;
Chris@42 129 TY = Tm - Tn;
Chris@42 130 To = Tm + Tn;
Chris@42 131 }
Chris@42 132 }
Chris@42 133 }
Chris@42 134 R0[WS(rs, 5)] = FMA(KP2_000000000, Tk, T5);
Chris@42 135 R1[WS(rs, 7)] = FMA(KP2_000000000, TC, Tr);
Chris@42 136 TG = TE + TF;
Chris@42 137 T16 = TE - TF;
Chris@42 138 R0[0] = FMA(KP2_000000000, To, Tl);
Chris@42 139 {
Chris@42 140 E TU, TW, T1g, T1e, T15, TV, TJ, TH, T1h, T1f, T17;
Chris@42 141 TU = FNMS(KP618033988, TT, TO);
Chris@42 142 TW = FMA(KP618033988, TO, TT);
Chris@42 143 R1[WS(rs, 2)] = FMA(KP2_000000000, TG, TD);
Chris@42 144 TH = FNMS(KP500000000, Tk, T5);
Chris@42 145 T1g = FNMS(KP618033988, T1a, T1d);
Chris@42 146 T1e = FMA(KP618033988, T1d, T1a);
Chris@42 147 T15 = FNMS(KP500000000, TG, TD);
Chris@42 148 TV = FMA(KP1_118033988, TI, TH);
Chris@42 149 TJ = FNMS(KP1_118033988, TI, TH);
Chris@42 150 T1o = FMA(KP618033988, T1k, T1l);
Chris@42 151 T1m = FNMS(KP618033988, T1l, T1k);
Chris@42 152 R0[WS(rs, 3)] = FNMS(KP1_902113032, TW, TV);
Chris@42 153 R0[WS(rs, 7)] = FMA(KP1_902113032, TW, TV);
Chris@42 154 R0[WS(rs, 1)] = FMA(KP1_902113032, TU, TJ);
Chris@42 155 R0[WS(rs, 9)] = FNMS(KP1_902113032, TU, TJ);
Chris@42 156 T1f = FNMS(KP1_118033988, T16, T15);
Chris@42 157 T17 = FMA(KP1_118033988, T16, T15);
Chris@42 158 T1h = FNMS(KP500000000, TC, Tr);
Chris@42 159 R1[WS(rs, 6)] = FNMS(KP1_902113032, T1g, T1f);
Chris@42 160 R1[WS(rs, 8)] = FMA(KP1_902113032, T1g, T1f);
Chris@42 161 R1[WS(rs, 4)] = FMA(KP1_902113032, T1e, T17);
Chris@42 162 R1[0] = FNMS(KP1_902113032, T1e, T17);
Chris@42 163 T14 = FNMS(KP618033988, T10, T11);
Chris@42 164 T12 = FMA(KP618033988, T11, T10);
Chris@42 165 TX = FNMS(KP500000000, To, Tl);
Chris@42 166 T1n = FMA(KP1_118033988, T1i, T1h);
Chris@42 167 T1j = FNMS(KP1_118033988, T1i, T1h);
Chris@42 168 }
Chris@42 169 }
Chris@42 170 R1[WS(rs, 5)] = FNMS(KP1_902113032, T1o, T1n);
Chris@42 171 R1[WS(rs, 9)] = FMA(KP1_902113032, T1o, T1n);
Chris@42 172 R1[WS(rs, 3)] = FMA(KP1_902113032, T1m, T1j);
Chris@42 173 R1[WS(rs, 1)] = FNMS(KP1_902113032, T1m, T1j);
Chris@42 174 TZ = FMA(KP1_118033988, TY, TX);
Chris@42 175 T13 = FNMS(KP1_118033988, TY, TX);
Chris@42 176 R0[WS(rs, 4)] = FNMS(KP1_902113032, T14, T13);
Chris@42 177 R0[WS(rs, 6)] = FMA(KP1_902113032, T14, T13);
Chris@42 178 R0[WS(rs, 2)] = FMA(KP1_902113032, T12, TZ);
Chris@42 179 R0[WS(rs, 8)] = FNMS(KP1_902113032, T12, TZ);
Chris@42 180 }
Chris@42 181 }
Chris@42 182 }
Chris@42 183
Chris@42 184 static const kr2c_desc desc = { 20, "r2cb_20", {42, 0, 44, 0}, &GENUS };
Chris@42 185
Chris@42 186 void X(codelet_r2cb_20) (planner *p) {
Chris@42 187 X(kr2c_register) (p, r2cb_20, &desc);
Chris@42 188 }
Chris@42 189
Chris@42 190 #else /* HAVE_FMA */
Chris@42 191
Chris@42 192 /* Generated by: ../../../genfft/gen_r2cb.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 20 -name r2cb_20 -include r2cb.h */
Chris@42 193
Chris@42 194 /*
Chris@42 195 * This function contains 86 FP additions, 30 FP multiplications,
Chris@42 196 * (or, 70 additions, 14 multiplications, 16 fused multiply/add),
Chris@42 197 * 50 stack variables, 5 constants, and 40 memory accesses
Chris@42 198 */
Chris@42 199 #include "r2cb.h"
Chris@42 200
Chris@42 201 static void r2cb_20(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
Chris@42 202 {
Chris@42 203 DK(KP1_118033988, +1.118033988749894848204586834365638117720309180);
Chris@42 204 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
Chris@42 205 DK(KP1_902113032, +1.902113032590307144232878666758764286811397268);
Chris@42 206 DK(KP1_175570504, +1.175570504584946258337411909278145537195304875);
Chris@42 207 DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
Chris@42 208 {
Chris@42 209 INT i;
Chris@42 210 for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(80, rs), MAKE_VOLATILE_STRIDE(80, csr), MAKE_VOLATILE_STRIDE(80, csi)) {
Chris@42 211 E T6, TF, Tm, Tt, TQ, T1n, T1f, T12, T1m, TV, T13, T1c, Td, Tk, Tl;
Chris@42 212 E Ty, TD, TE, Tn, To, Tp, TG, TH, TI;
Chris@42 213 {
Chris@42 214 E T5, Ts, T3, Tq;
Chris@42 215 {
Chris@42 216 E T4, Tr, T1, T2;
Chris@42 217 T4 = Cr[WS(csr, 5)];
Chris@42 218 T5 = KP2_000000000 * T4;
Chris@42 219 Tr = Ci[WS(csi, 5)];
Chris@42 220 Ts = KP2_000000000 * Tr;
Chris@42 221 T1 = Cr[0];
Chris@42 222 T2 = Cr[WS(csr, 10)];
Chris@42 223 T3 = T1 + T2;
Chris@42 224 Tq = T1 - T2;
Chris@42 225 }
Chris@42 226 T6 = T3 - T5;
Chris@42 227 TF = Tq - Ts;
Chris@42 228 Tm = T3 + T5;
Chris@42 229 Tt = Tq + Ts;
Chris@42 230 }
Chris@42 231 {
Chris@42 232 E T9, Tu, TO, T1b, Tc, T1a, Tx, TP, Tg, Tz, TT, T1e, Tj, T1d, TC;
Chris@42 233 E TU;
Chris@42 234 {
Chris@42 235 E T7, T8, TM, TN;
Chris@42 236 T7 = Cr[WS(csr, 4)];
Chris@42 237 T8 = Cr[WS(csr, 6)];
Chris@42 238 T9 = T7 + T8;
Chris@42 239 Tu = T7 - T8;
Chris@42 240 TM = Ci[WS(csi, 4)];
Chris@42 241 TN = Ci[WS(csi, 6)];
Chris@42 242 TO = TM - TN;
Chris@42 243 T1b = TM + TN;
Chris@42 244 }
Chris@42 245 {
Chris@42 246 E Ta, Tb, Tv, Tw;
Chris@42 247 Ta = Cr[WS(csr, 9)];
Chris@42 248 Tb = Cr[WS(csr, 1)];
Chris@42 249 Tc = Ta + Tb;
Chris@42 250 T1a = Ta - Tb;
Chris@42 251 Tv = Ci[WS(csi, 9)];
Chris@42 252 Tw = Ci[WS(csi, 1)];
Chris@42 253 Tx = Tv + Tw;
Chris@42 254 TP = Tv - Tw;
Chris@42 255 }
Chris@42 256 {
Chris@42 257 E Te, Tf, TR, TS;
Chris@42 258 Te = Cr[WS(csr, 8)];
Chris@42 259 Tf = Cr[WS(csr, 2)];
Chris@42 260 Tg = Te + Tf;
Chris@42 261 Tz = Te - Tf;
Chris@42 262 TR = Ci[WS(csi, 8)];
Chris@42 263 TS = Ci[WS(csi, 2)];
Chris@42 264 TT = TR - TS;
Chris@42 265 T1e = TR + TS;
Chris@42 266 }
Chris@42 267 {
Chris@42 268 E Th, Ti, TA, TB;
Chris@42 269 Th = Cr[WS(csr, 7)];
Chris@42 270 Ti = Cr[WS(csr, 3)];
Chris@42 271 Tj = Th + Ti;
Chris@42 272 T1d = Th - Ti;
Chris@42 273 TA = Ci[WS(csi, 7)];
Chris@42 274 TB = Ci[WS(csi, 3)];
Chris@42 275 TC = TA + TB;
Chris@42 276 TU = TB - TA;
Chris@42 277 }
Chris@42 278 TQ = TO - TP;
Chris@42 279 T1n = T1e - T1d;
Chris@42 280 T1f = T1d + T1e;
Chris@42 281 T12 = TP + TO;
Chris@42 282 T1m = T1b - T1a;
Chris@42 283 TV = TT - TU;
Chris@42 284 T13 = TU + TT;
Chris@42 285 T1c = T1a + T1b;
Chris@42 286 Td = T9 - Tc;
Chris@42 287 Tk = Tg - Tj;
Chris@42 288 Tl = Td + Tk;
Chris@42 289 Ty = Tu + Tx;
Chris@42 290 TD = Tz - TC;
Chris@42 291 TE = Ty + TD;
Chris@42 292 Tn = T9 + Tc;
Chris@42 293 To = Tg + Tj;
Chris@42 294 Tp = Tn + To;
Chris@42 295 TG = Tu - Tx;
Chris@42 296 TH = Tz + TC;
Chris@42 297 TI = TG + TH;
Chris@42 298 }
Chris@42 299 R0[WS(rs, 5)] = FMA(KP2_000000000, Tl, T6);
Chris@42 300 R1[WS(rs, 7)] = FMA(KP2_000000000, TE, Tt);
Chris@42 301 R1[WS(rs, 2)] = FMA(KP2_000000000, TI, TF);
Chris@42 302 R0[0] = FMA(KP2_000000000, Tp, Tm);
Chris@42 303 {
Chris@42 304 E TW, TY, TL, TX, TJ, TK;
Chris@42 305 TW = FNMS(KP1_902113032, TV, KP1_175570504 * TQ);
Chris@42 306 TY = FMA(KP1_902113032, TQ, KP1_175570504 * TV);
Chris@42 307 TJ = FNMS(KP500000000, Tl, T6);
Chris@42 308 TK = KP1_118033988 * (Td - Tk);
Chris@42 309 TL = TJ - TK;
Chris@42 310 TX = TK + TJ;
Chris@42 311 R0[WS(rs, 1)] = TL - TW;
Chris@42 312 R0[WS(rs, 7)] = TX + TY;
Chris@42 313 R0[WS(rs, 9)] = TL + TW;
Chris@42 314 R0[WS(rs, 3)] = TX - TY;
Chris@42 315 }
Chris@42 316 {
Chris@42 317 E T1g, T1i, T19, T1h, T17, T18;
Chris@42 318 T1g = FNMS(KP1_902113032, T1f, KP1_175570504 * T1c);
Chris@42 319 T1i = FMA(KP1_902113032, T1c, KP1_175570504 * T1f);
Chris@42 320 T17 = FNMS(KP500000000, TI, TF);
Chris@42 321 T18 = KP1_118033988 * (TG - TH);
Chris@42 322 T19 = T17 - T18;
Chris@42 323 T1h = T18 + T17;
Chris@42 324 R1[WS(rs, 8)] = T19 - T1g;
Chris@42 325 R1[WS(rs, 4)] = T1h + T1i;
Chris@42 326 R1[WS(rs, 6)] = T19 + T1g;
Chris@42 327 R1[0] = T1h - T1i;
Chris@42 328 }
Chris@42 329 {
Chris@42 330 E T1o, T1q, T1l, T1p, T1j, T1k;
Chris@42 331 T1o = FNMS(KP1_902113032, T1n, KP1_175570504 * T1m);
Chris@42 332 T1q = FMA(KP1_902113032, T1m, KP1_175570504 * T1n);
Chris@42 333 T1j = FNMS(KP500000000, TE, Tt);
Chris@42 334 T1k = KP1_118033988 * (Ty - TD);
Chris@42 335 T1l = T1j - T1k;
Chris@42 336 T1p = T1k + T1j;
Chris@42 337 R1[WS(rs, 3)] = T1l - T1o;
Chris@42 338 R1[WS(rs, 9)] = T1p + T1q;
Chris@42 339 R1[WS(rs, 1)] = T1l + T1o;
Chris@42 340 R1[WS(rs, 5)] = T1p - T1q;
Chris@42 341 }
Chris@42 342 {
Chris@42 343 E T14, T16, T11, T15, TZ, T10;
Chris@42 344 T14 = FNMS(KP1_902113032, T13, KP1_175570504 * T12);
Chris@42 345 T16 = FMA(KP1_902113032, T12, KP1_175570504 * T13);
Chris@42 346 TZ = FNMS(KP500000000, Tp, Tm);
Chris@42 347 T10 = KP1_118033988 * (Tn - To);
Chris@42 348 T11 = TZ - T10;
Chris@42 349 T15 = T10 + TZ;
Chris@42 350 R0[WS(rs, 6)] = T11 - T14;
Chris@42 351 R0[WS(rs, 2)] = T15 + T16;
Chris@42 352 R0[WS(rs, 4)] = T11 + T14;
Chris@42 353 R0[WS(rs, 8)] = T15 - T16;
Chris@42 354 }
Chris@42 355 }
Chris@42 356 }
Chris@42 357 }
Chris@42 358
Chris@42 359 static const kr2c_desc desc = { 20, "r2cb_20", {70, 14, 16, 0}, &GENUS };
Chris@42 360
Chris@42 361 void X(codelet_r2cb_20) (planner *p) {
Chris@42 362 X(kr2c_register) (p, r2cb_20, &desc);
Chris@42 363 }
Chris@42 364
Chris@42 365 #endif /* HAVE_FMA */