Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.8/dft/scalar/codelets/n1_9.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 08:04:10 EDT 2018 */
Chris@82 23
Chris@82 24 #include "dft/codelet-dft.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_notw.native -fma -compact -variables 4 -pipeline-latency 4 -n 9 -name n1_9 -include dft/scalar/n.h */
Chris@82 29
Chris@82 30 /*
Chris@82 31 * This function contains 80 FP additions, 56 FP multiplications,
Chris@82 32 * (or, 24 additions, 0 multiplications, 56 fused multiply/add),
Chris@82 33 * 41 stack variables, 10 constants, and 36 memory accesses
Chris@82 34 */
Chris@82 35 #include "dft/scalar/n.h"
Chris@82 36
Chris@82 37 static void n1_9(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
Chris@82 38 {
Chris@82 39 DK(KP954188894, +0.954188894138671133499268364187245676532219158);
Chris@82 40 DK(KP363970234, +0.363970234266202361351047882776834043890471784);
Chris@82 41 DK(KP852868531, +0.852868531952443209628250963940074071936020296);
Chris@82 42 DK(KP492403876, +0.492403876506104029683371512294761506835321626);
Chris@82 43 DK(KP984807753, +0.984807753012208059366743024589523013670643252);
Chris@82 44 DK(KP777861913, +0.777861913430206160028177977318626690410586096);
Chris@82 45 DK(KP839099631, +0.839099631177280011763127298123181364687434283);
Chris@82 46 DK(KP176326980, +0.176326980708464973471090386868618986121633062);
Chris@82 47 DK(KP866025403, +0.866025403784438646763723170752936183471402627);
Chris@82 48 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
Chris@82 49 {
Chris@82 50 INT i;
Chris@82 51 for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(36, is), MAKE_VOLATILE_STRIDE(36, os)) {
Chris@82 52 E T5, TL, Tm, Tl, T1f, TM, Ta, T1c, TF, TW, TI, TX, Tf, T1d, Ts;
Chris@82 53 E TZ, Tx, T10;
Chris@82 54 {
Chris@82 55 E T1, T2, T3, T4;
Chris@82 56 T1 = ri[0];
Chris@82 57 T2 = ri[WS(is, 3)];
Chris@82 58 T3 = ri[WS(is, 6)];
Chris@82 59 T4 = T2 + T3;
Chris@82 60 T5 = T1 + T4;
Chris@82 61 TL = FNMS(KP500000000, T4, T1);
Chris@82 62 Tm = T3 - T2;
Chris@82 63 }
Chris@82 64 {
Chris@82 65 E Th, Ti, Tj, Tk;
Chris@82 66 Th = ii[0];
Chris@82 67 Ti = ii[WS(is, 3)];
Chris@82 68 Tj = ii[WS(is, 6)];
Chris@82 69 Tk = Ti + Tj;
Chris@82 70 Tl = FNMS(KP500000000, Tk, Th);
Chris@82 71 T1f = Th + Tk;
Chris@82 72 TM = Ti - Tj;
Chris@82 73 }
Chris@82 74 {
Chris@82 75 E T6, Tz, T9, TE, TC, TH, TD, TG;
Chris@82 76 T6 = ri[WS(is, 1)];
Chris@82 77 Tz = ii[WS(is, 1)];
Chris@82 78 {
Chris@82 79 E T7, T8, TA, TB;
Chris@82 80 T7 = ri[WS(is, 4)];
Chris@82 81 T8 = ri[WS(is, 7)];
Chris@82 82 T9 = T7 + T8;
Chris@82 83 TE = T7 - T8;
Chris@82 84 TA = ii[WS(is, 4)];
Chris@82 85 TB = ii[WS(is, 7)];
Chris@82 86 TC = TA + TB;
Chris@82 87 TH = TB - TA;
Chris@82 88 }
Chris@82 89 Ta = T6 + T9;
Chris@82 90 T1c = Tz + TC;
Chris@82 91 TD = FNMS(KP500000000, TC, Tz);
Chris@82 92 TF = FNMS(KP866025403, TE, TD);
Chris@82 93 TW = FMA(KP866025403, TE, TD);
Chris@82 94 TG = FNMS(KP500000000, T9, T6);
Chris@82 95 TI = FNMS(KP866025403, TH, TG);
Chris@82 96 TX = FMA(KP866025403, TH, TG);
Chris@82 97 }
Chris@82 98 {
Chris@82 99 E Tb, Tt, Te, Tw, Tr, Tu, To, Tv;
Chris@82 100 Tb = ri[WS(is, 2)];
Chris@82 101 Tt = ii[WS(is, 2)];
Chris@82 102 {
Chris@82 103 E Tc, Td, Tp, Tq;
Chris@82 104 Tc = ri[WS(is, 5)];
Chris@82 105 Td = ri[WS(is, 8)];
Chris@82 106 Te = Tc + Td;
Chris@82 107 Tw = Td - Tc;
Chris@82 108 Tp = ii[WS(is, 5)];
Chris@82 109 Tq = ii[WS(is, 8)];
Chris@82 110 Tr = Tp - Tq;
Chris@82 111 Tu = Tp + Tq;
Chris@82 112 }
Chris@82 113 Tf = Tb + Te;
Chris@82 114 T1d = Tt + Tu;
Chris@82 115 To = FNMS(KP500000000, Te, Tb);
Chris@82 116 Ts = FMA(KP866025403, Tr, To);
Chris@82 117 TZ = FNMS(KP866025403, Tr, To);
Chris@82 118 Tv = FNMS(KP500000000, Tu, Tt);
Chris@82 119 Tx = FMA(KP866025403, Tw, Tv);
Chris@82 120 T10 = FNMS(KP866025403, Tw, Tv);
Chris@82 121 }
Chris@82 122 {
Chris@82 123 E T1e, Tg, T1b, T1i, T1g, T1h;
Chris@82 124 T1e = T1c - T1d;
Chris@82 125 Tg = Ta + Tf;
Chris@82 126 T1b = FNMS(KP500000000, Tg, T5);
Chris@82 127 ro[0] = T5 + Tg;
Chris@82 128 ro[WS(os, 3)] = FMA(KP866025403, T1e, T1b);
Chris@82 129 ro[WS(os, 6)] = FNMS(KP866025403, T1e, T1b);
Chris@82 130 T1i = Tf - Ta;
Chris@82 131 T1g = T1c + T1d;
Chris@82 132 T1h = FNMS(KP500000000, T1g, T1f);
Chris@82 133 io[WS(os, 3)] = FMA(KP866025403, T1i, T1h);
Chris@82 134 io[0] = T1f + T1g;
Chris@82 135 io[WS(os, 6)] = FNMS(KP866025403, T1i, T1h);
Chris@82 136 }
Chris@82 137 {
Chris@82 138 E Tn, TN, TK, TS, TQ, TU, TR, TT;
Chris@82 139 Tn = FMA(KP866025403, Tm, Tl);
Chris@82 140 TN = FMA(KP866025403, TM, TL);
Chris@82 141 {
Chris@82 142 E Ty, TJ, TO, TP;
Chris@82 143 Ty = FNMS(KP176326980, Tx, Ts);
Chris@82 144 TJ = FNMS(KP839099631, TI, TF);
Chris@82 145 TK = FNMS(KP777861913, TJ, Ty);
Chris@82 146 TS = FMA(KP777861913, TJ, Ty);
Chris@82 147 TO = FMA(KP176326980, Ts, Tx);
Chris@82 148 TP = FMA(KP839099631, TF, TI);
Chris@82 149 TQ = FMA(KP777861913, TP, TO);
Chris@82 150 TU = FNMS(KP777861913, TP, TO);
Chris@82 151 }
Chris@82 152 io[WS(os, 1)] = FNMS(KP984807753, TK, Tn);
Chris@82 153 ro[WS(os, 1)] = FMA(KP984807753, TQ, TN);
Chris@82 154 TR = FNMS(KP492403876, TQ, TN);
Chris@82 155 ro[WS(os, 4)] = FMA(KP852868531, TS, TR);
Chris@82 156 ro[WS(os, 7)] = FNMS(KP852868531, TS, TR);
Chris@82 157 TT = FMA(KP492403876, TK, Tn);
Chris@82 158 io[WS(os, 7)] = FNMS(KP852868531, TU, TT);
Chris@82 159 io[WS(os, 4)] = FMA(KP852868531, TU, TT);
Chris@82 160 }
Chris@82 161 {
Chris@82 162 E TV, T17, T12, T1a, T16, T18, T13, T19;
Chris@82 163 TV = FNMS(KP866025403, TM, TL);
Chris@82 164 T17 = FNMS(KP866025403, Tm, Tl);
Chris@82 165 {
Chris@82 166 E TY, T11, T14, T15;
Chris@82 167 TY = FMA(KP176326980, TX, TW);
Chris@82 168 T11 = FNMS(KP363970234, T10, TZ);
Chris@82 169 T12 = FNMS(KP954188894, T11, TY);
Chris@82 170 T1a = FMA(KP954188894, T11, TY);
Chris@82 171 T14 = FNMS(KP176326980, TW, TX);
Chris@82 172 T15 = FMA(KP363970234, TZ, T10);
Chris@82 173 T16 = FNMS(KP954188894, T15, T14);
Chris@82 174 T18 = FMA(KP954188894, T15, T14);
Chris@82 175 }
Chris@82 176 ro[WS(os, 2)] = FMA(KP984807753, T12, TV);
Chris@82 177 io[WS(os, 2)] = FNMS(KP984807753, T18, T17);
Chris@82 178 T13 = FNMS(KP492403876, T12, TV);
Chris@82 179 ro[WS(os, 5)] = FNMS(KP852868531, T16, T13);
Chris@82 180 ro[WS(os, 8)] = FMA(KP852868531, T16, T13);
Chris@82 181 T19 = FMA(KP492403876, T18, T17);
Chris@82 182 io[WS(os, 5)] = FNMS(KP852868531, T1a, T19);
Chris@82 183 io[WS(os, 8)] = FMA(KP852868531, T1a, T19);
Chris@82 184 }
Chris@82 185 }
Chris@82 186 }
Chris@82 187 }
Chris@82 188
Chris@82 189 static const kdft_desc desc = { 9, "n1_9", {24, 0, 56, 0}, &GENUS, 0, 0, 0, 0 };
Chris@82 190
Chris@82 191 void X(codelet_n1_9) (planner *p) {
Chris@82 192 X(kdft_register) (p, n1_9, &desc);
Chris@82 193 }
Chris@82 194
Chris@82 195 #else
Chris@82 196
Chris@82 197 /* Generated by: ../../../genfft/gen_notw.native -compact -variables 4 -pipeline-latency 4 -n 9 -name n1_9 -include dft/scalar/n.h */
Chris@82 198
Chris@82 199 /*
Chris@82 200 * This function contains 80 FP additions, 40 FP multiplications,
Chris@82 201 * (or, 60 additions, 20 multiplications, 20 fused multiply/add),
Chris@82 202 * 39 stack variables, 8 constants, and 36 memory accesses
Chris@82 203 */
Chris@82 204 #include "dft/scalar/n.h"
Chris@82 205
Chris@82 206 static void n1_9(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
Chris@82 207 {
Chris@82 208 DK(KP939692620, +0.939692620785908384054109277324731469936208134);
Chris@82 209 DK(KP342020143, +0.342020143325668733044099614682259580763083368);
Chris@82 210 DK(KP984807753, +0.984807753012208059366743024589523013670643252);
Chris@82 211 DK(KP173648177, +0.173648177666930348851716626769314796000375677);
Chris@82 212 DK(KP642787609, +0.642787609686539326322643409907263432907559884);
Chris@82 213 DK(KP766044443, +0.766044443118978035202392650555416673935832457);
Chris@82 214 DK(KP500000000, +0.500000000000000000000000000000000000000000000);
Chris@82 215 DK(KP866025403, +0.866025403784438646763723170752936183471402627);
Chris@82 216 {
Chris@82 217 INT i;
Chris@82 218 for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(36, is), MAKE_VOLATILE_STRIDE(36, os)) {
Chris@82 219 E T5, TO, Th, Tk, T1g, TR, Ta, T1c, Tq, TW, Tv, TX, Tf, T1d, TB;
Chris@82 220 E T10, TG, TZ;
Chris@82 221 {
Chris@82 222 E T1, T2, T3, T4;
Chris@82 223 T1 = ri[0];
Chris@82 224 T2 = ri[WS(is, 3)];
Chris@82 225 T3 = ri[WS(is, 6)];
Chris@82 226 T4 = T2 + T3;
Chris@82 227 T5 = T1 + T4;
Chris@82 228 TO = KP866025403 * (T3 - T2);
Chris@82 229 Th = FNMS(KP500000000, T4, T1);
Chris@82 230 }
Chris@82 231 {
Chris@82 232 E TP, Ti, Tj, TQ;
Chris@82 233 TP = ii[0];
Chris@82 234 Ti = ii[WS(is, 3)];
Chris@82 235 Tj = ii[WS(is, 6)];
Chris@82 236 TQ = Ti + Tj;
Chris@82 237 Tk = KP866025403 * (Ti - Tj);
Chris@82 238 T1g = TP + TQ;
Chris@82 239 TR = FNMS(KP500000000, TQ, TP);
Chris@82 240 }
Chris@82 241 {
Chris@82 242 E T6, Ts, T9, Tr, Tp, Tt, Tm, Tu;
Chris@82 243 T6 = ri[WS(is, 1)];
Chris@82 244 Ts = ii[WS(is, 1)];
Chris@82 245 {
Chris@82 246 E T7, T8, Tn, To;
Chris@82 247 T7 = ri[WS(is, 4)];
Chris@82 248 T8 = ri[WS(is, 7)];
Chris@82 249 T9 = T7 + T8;
Chris@82 250 Tr = KP866025403 * (T8 - T7);
Chris@82 251 Tn = ii[WS(is, 4)];
Chris@82 252 To = ii[WS(is, 7)];
Chris@82 253 Tp = KP866025403 * (Tn - To);
Chris@82 254 Tt = Tn + To;
Chris@82 255 }
Chris@82 256 Ta = T6 + T9;
Chris@82 257 T1c = Ts + Tt;
Chris@82 258 Tm = FNMS(KP500000000, T9, T6);
Chris@82 259 Tq = Tm + Tp;
Chris@82 260 TW = Tm - Tp;
Chris@82 261 Tu = FNMS(KP500000000, Tt, Ts);
Chris@82 262 Tv = Tr + Tu;
Chris@82 263 TX = Tu - Tr;
Chris@82 264 }
Chris@82 265 {
Chris@82 266 E Tb, TD, Te, TC, TA, TE, Tx, TF;
Chris@82 267 Tb = ri[WS(is, 2)];
Chris@82 268 TD = ii[WS(is, 2)];
Chris@82 269 {
Chris@82 270 E Tc, Td, Ty, Tz;
Chris@82 271 Tc = ri[WS(is, 5)];
Chris@82 272 Td = ri[WS(is, 8)];
Chris@82 273 Te = Tc + Td;
Chris@82 274 TC = KP866025403 * (Td - Tc);
Chris@82 275 Ty = ii[WS(is, 5)];
Chris@82 276 Tz = ii[WS(is, 8)];
Chris@82 277 TA = KP866025403 * (Ty - Tz);
Chris@82 278 TE = Ty + Tz;
Chris@82 279 }
Chris@82 280 Tf = Tb + Te;
Chris@82 281 T1d = TD + TE;
Chris@82 282 Tx = FNMS(KP500000000, Te, Tb);
Chris@82 283 TB = Tx + TA;
Chris@82 284 T10 = Tx - TA;
Chris@82 285 TF = FNMS(KP500000000, TE, TD);
Chris@82 286 TG = TC + TF;
Chris@82 287 TZ = TF - TC;
Chris@82 288 }
Chris@82 289 {
Chris@82 290 E T1e, Tg, T1b, T1f, T1h, T1i;
Chris@82 291 T1e = KP866025403 * (T1c - T1d);
Chris@82 292 Tg = Ta + Tf;
Chris@82 293 T1b = FNMS(KP500000000, Tg, T5);
Chris@82 294 ro[0] = T5 + Tg;
Chris@82 295 ro[WS(os, 3)] = T1b + T1e;
Chris@82 296 ro[WS(os, 6)] = T1b - T1e;
Chris@82 297 T1f = KP866025403 * (Tf - Ta);
Chris@82 298 T1h = T1c + T1d;
Chris@82 299 T1i = FNMS(KP500000000, T1h, T1g);
Chris@82 300 io[WS(os, 3)] = T1f + T1i;
Chris@82 301 io[0] = T1g + T1h;
Chris@82 302 io[WS(os, 6)] = T1i - T1f;
Chris@82 303 }
Chris@82 304 {
Chris@82 305 E Tl, TS, TI, TN, TM, TT, TJ, TU;
Chris@82 306 Tl = Th + Tk;
Chris@82 307 TS = TO + TR;
Chris@82 308 {
Chris@82 309 E Tw, TH, TK, TL;
Chris@82 310 Tw = FMA(KP766044443, Tq, KP642787609 * Tv);
Chris@82 311 TH = FMA(KP173648177, TB, KP984807753 * TG);
Chris@82 312 TI = Tw + TH;
Chris@82 313 TN = KP866025403 * (TH - Tw);
Chris@82 314 TK = FNMS(KP642787609, Tq, KP766044443 * Tv);
Chris@82 315 TL = FNMS(KP984807753, TB, KP173648177 * TG);
Chris@82 316 TM = KP866025403 * (TK - TL);
Chris@82 317 TT = TK + TL;
Chris@82 318 }
Chris@82 319 ro[WS(os, 1)] = Tl + TI;
Chris@82 320 io[WS(os, 1)] = TS + TT;
Chris@82 321 TJ = FNMS(KP500000000, TI, Tl);
Chris@82 322 ro[WS(os, 7)] = TJ - TM;
Chris@82 323 ro[WS(os, 4)] = TJ + TM;
Chris@82 324 TU = FNMS(KP500000000, TT, TS);
Chris@82 325 io[WS(os, 4)] = TN + TU;
Chris@82 326 io[WS(os, 7)] = TU - TN;
Chris@82 327 }
Chris@82 328 {
Chris@82 329 E TV, T14, T12, T13, T17, T1a, T18, T19;
Chris@82 330 TV = Th - Tk;
Chris@82 331 T14 = TR - TO;
Chris@82 332 {
Chris@82 333 E TY, T11, T15, T16;
Chris@82 334 TY = FMA(KP173648177, TW, KP984807753 * TX);
Chris@82 335 T11 = FNMS(KP939692620, T10, KP342020143 * TZ);
Chris@82 336 T12 = TY + T11;
Chris@82 337 T13 = KP866025403 * (T11 - TY);
Chris@82 338 T15 = FNMS(KP984807753, TW, KP173648177 * TX);
Chris@82 339 T16 = FMA(KP342020143, T10, KP939692620 * TZ);
Chris@82 340 T17 = T15 - T16;
Chris@82 341 T1a = KP866025403 * (T15 + T16);
Chris@82 342 }
Chris@82 343 ro[WS(os, 2)] = TV + T12;
Chris@82 344 io[WS(os, 2)] = T14 + T17;
Chris@82 345 T18 = FNMS(KP500000000, T17, T14);
Chris@82 346 io[WS(os, 5)] = T13 + T18;
Chris@82 347 io[WS(os, 8)] = T18 - T13;
Chris@82 348 T19 = FNMS(KP500000000, T12, TV);
Chris@82 349 ro[WS(os, 8)] = T19 - T1a;
Chris@82 350 ro[WS(os, 5)] = T19 + T1a;
Chris@82 351 }
Chris@82 352 }
Chris@82 353 }
Chris@82 354 }
Chris@82 355
Chris@82 356 static const kdft_desc desc = { 9, "n1_9", {60, 20, 20, 0}, &GENUS, 0, 0, 0, 0 };
Chris@82 357
Chris@82 358 void X(codelet_n1_9) (planner *p) {
Chris@82 359 X(kdft_register) (p, n1_9, &desc);
Chris@82 360 }
Chris@82 361
Chris@82 362 #endif