Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.8/rdft/simd/common/hc2cfdftv_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 08:08:11 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_hc2cdft_c.native -fma -simd -compact -variables 4 -pipeline-latency 8 -trivial-stores -variables 32 -no-generate-bytw -n 32 -dit -name hc2cfdftv_32 -include rdft/simd/hc2cfv.h */
Chris@82 29
Chris@82 30 /*
Chris@82 31 * This function contains 249 FP additions, 224 FP multiplications,
Chris@82 32 * (or, 119 additions, 94 multiplications, 130 fused multiply/add),
Chris@82 33 * 154 stack variables, 8 constants, and 64 memory accesses
Chris@82 34 */
Chris@82 35 #include "rdft/simd/hc2cfv.h"
Chris@82 36
Chris@82 37 static void hc2cfdftv_32(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
Chris@82 38 {
Chris@82 39 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
Chris@82 40 DVK(KP668178637, +0.668178637919298919997757686523080761552472251);
Chris@82 41 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
Chris@82 42 DVK(KP198912367, +0.198912367379658006911597622644676228597850501);
Chris@82 43 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
Chris@82 44 DVK(KP500000000, +0.500000000000000000000000000000000000000000000);
Chris@82 45 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
Chris@82 46 DVK(KP414213562, +0.414213562373095048801688724209698078569671875);
Chris@82 47 {
Chris@82 48 INT m;
Chris@82 49 for (m = mb, W = W + ((mb - 1) * ((TWVL / VL) * 62)); m < me; m = m + VL, Rp = Rp + (VL * ms), Ip = Ip + (VL * ms), Rm = Rm - (VL * ms), Im = Im - (VL * ms), W = W + (TWVL * 62), MAKE_VOLATILE_STRIDE(128, rs)) {
Chris@82 50 V T47, T48, T4l, T3w, T3F, T3B, T41, Ts, T2y, T1Q, T2B, T27, T2J, T3a, T40;
Chris@82 51 V T1X, T2C, T43, T44, T4a, T4b, T4m, T3p, T3E, T15, T2K, T1u, T2F, T3h, T3C;
Chris@82 52 V T1n, T2E, T2a, T2z, T1a, T18, TU, T3m, T3f, T1r, T1p, T13, T3n, T3e, TB;
Chris@82 53 V T3k, T1l, T3c, TK, T3j, T1g, T3b, T3l, T3o, TL, T14, T1s, T1t, T3d, T3g;
Chris@82 54 V T1b, T1m, T28, T29, T3Q, T3W, T3T, T3X, T3O, T3P, T3R, T3S, T3U, T3Z, T3V;
Chris@82 55 V T3Y;
Chris@82 56 {
Chris@82 57 V T1U, T1S, T3, T3u, T7, T1z, T1D, T3t, T24, T22, Tc, Tg, Th, T3q, T1J;
Chris@82 58 V Tl, Tp, Tq, T3r, T1O, T3s, T3v, T3z, T3A, T8, Tr, T1E, T1P, T25, T26;
Chris@82 59 V T38, T39, T1V, T1W;
Chris@82 60 {
Chris@82 61 V T1, T2, T5, T6, T1T, T1R, T4, T1x, T1y, T1B, T1C, T1w, T1A, T23, T21;
Chris@82 62 V T1I, T1G, Ta, Tb, T9, T1H, Te, Tf, Td, T1F, T1N, T1L, Tj, Tk, Ti;
Chris@82 63 V T1M, Tn, To, Tm, T1K;
Chris@82 64 T1 = LD(&(Rp[0]), ms, &(Rp[0]));
Chris@82 65 T2 = LD(&(Rm[0]), -ms, &(Rm[0]));
Chris@82 66 T1T = LDW(&(W[0]));
Chris@82 67 T1U = VZMULIJ(T1T, VFNMSCONJ(T2, T1));
Chris@82 68 T5 = LD(&(Rp[WS(rs, 8)]), ms, &(Rp[0]));
Chris@82 69 T6 = LD(&(Rm[WS(rs, 8)]), -ms, &(Rm[0]));
Chris@82 70 T1R = LDW(&(W[TWVL * 32]));
Chris@82 71 T1S = VZMULIJ(T1R, VFNMSCONJ(T6, T5));
Chris@82 72 T3 = VFMACONJ(T2, T1);
Chris@82 73 T3u = VADD(T1U, T1S);
Chris@82 74 T4 = LDW(&(W[TWVL * 30]));
Chris@82 75 T7 = VZMULJ(T4, VFMACONJ(T6, T5));
Chris@82 76 T1x = LD(&(Rp[WS(rs, 12)]), ms, &(Rp[0]));
Chris@82 77 T1y = LD(&(Rm[WS(rs, 12)]), -ms, &(Rm[0]));
Chris@82 78 T1w = LDW(&(W[TWVL * 48]));
Chris@82 79 T1z = VZMULIJ(T1w, VFNMSCONJ(T1y, T1x));
Chris@82 80 T1B = LD(&(Rp[WS(rs, 4)]), ms, &(Rp[0]));
Chris@82 81 T1C = LD(&(Rm[WS(rs, 4)]), -ms, &(Rm[0]));
Chris@82 82 T1A = LDW(&(W[TWVL * 16]));
Chris@82 83 T1D = VZMULIJ(T1A, VFNMSCONJ(T1C, T1B));
Chris@82 84 T3t = VADD(T1D, T1z);
Chris@82 85 T23 = LDW(&(W[TWVL * 46]));
Chris@82 86 T24 = VZMULJ(T23, VFMACONJ(T1y, T1x));
Chris@82 87 T21 = LDW(&(W[TWVL * 14]));
Chris@82 88 T22 = VZMULJ(T21, VFMACONJ(T1C, T1B));
Chris@82 89 Ta = LD(&(Rp[WS(rs, 2)]), ms, &(Rp[0]));
Chris@82 90 Tb = LD(&(Rm[WS(rs, 2)]), -ms, &(Rm[0]));
Chris@82 91 T9 = LDW(&(W[TWVL * 6]));
Chris@82 92 Tc = VZMULJ(T9, VFMACONJ(Tb, Ta));
Chris@82 93 T1H = LDW(&(W[TWVL * 8]));
Chris@82 94 T1I = VZMULIJ(T1H, VFNMSCONJ(Tb, Ta));
Chris@82 95 Te = LD(&(Rp[WS(rs, 10)]), ms, &(Rp[0]));
Chris@82 96 Tf = LD(&(Rm[WS(rs, 10)]), -ms, &(Rm[0]));
Chris@82 97 Td = LDW(&(W[TWVL * 38]));
Chris@82 98 Tg = VZMULJ(Td, VFMACONJ(Tf, Te));
Chris@82 99 T1F = LDW(&(W[TWVL * 40]));
Chris@82 100 T1G = VZMULIJ(T1F, VFNMSCONJ(Tf, Te));
Chris@82 101 Th = VSUB(Tc, Tg);
Chris@82 102 T3q = VADD(T1I, T1G);
Chris@82 103 T1J = VSUB(T1G, T1I);
Chris@82 104 Tj = LD(&(Rp[WS(rs, 14)]), ms, &(Rp[0]));
Chris@82 105 Tk = LD(&(Rm[WS(rs, 14)]), -ms, &(Rm[0]));
Chris@82 106 Ti = LDW(&(W[TWVL * 54]));
Chris@82 107 Tl = VZMULJ(Ti, VFMACONJ(Tk, Tj));
Chris@82 108 T1M = LDW(&(W[TWVL * 56]));
Chris@82 109 T1N = VZMULIJ(T1M, VFNMSCONJ(Tk, Tj));
Chris@82 110 Tn = LD(&(Rp[WS(rs, 6)]), ms, &(Rp[0]));
Chris@82 111 To = LD(&(Rm[WS(rs, 6)]), -ms, &(Rm[0]));
Chris@82 112 Tm = LDW(&(W[TWVL * 22]));
Chris@82 113 Tp = VZMULJ(Tm, VFMACONJ(To, Tn));
Chris@82 114 T1K = LDW(&(W[TWVL * 24]));
Chris@82 115 T1L = VZMULIJ(T1K, VFNMSCONJ(To, Tn));
Chris@82 116 Tq = VSUB(Tl, Tp);
Chris@82 117 T3r = VADD(T1N, T1L);
Chris@82 118 T1O = VSUB(T1L, T1N);
Chris@82 119 }
Chris@82 120 T47 = VADD(T3u, T3t);
Chris@82 121 T48 = VADD(T3q, T3r);
Chris@82 122 T4l = VSUB(T48, T47);
Chris@82 123 T3s = VSUB(T3q, T3r);
Chris@82 124 T3v = VSUB(T3t, T3u);
Chris@82 125 T3w = VFNMS(LDK(KP414213562), T3v, T3s);
Chris@82 126 T3F = VFMA(LDK(KP414213562), T3s, T3v);
Chris@82 127 T3z = VADD(Tl, Tp);
Chris@82 128 T3A = VADD(Tc, Tg);
Chris@82 129 T3B = VSUB(T3z, T3A);
Chris@82 130 T41 = VADD(T3A, T3z);
Chris@82 131 T8 = VSUB(T3, T7);
Chris@82 132 Tr = VADD(Th, Tq);
Chris@82 133 Ts = VFNMS(LDK(KP707106781), Tr, T8);
Chris@82 134 T2y = VFMA(LDK(KP707106781), Tr, T8);
Chris@82 135 T1E = VSUB(T1z, T1D);
Chris@82 136 T1P = VSUB(T1J, T1O);
Chris@82 137 T1Q = VFNMS(LDK(KP707106781), T1P, T1E);
Chris@82 138 T2B = VFMA(LDK(KP707106781), T1P, T1E);
Chris@82 139 T25 = VSUB(T22, T24);
Chris@82 140 T26 = VSUB(Tq, Th);
Chris@82 141 T27 = VFMA(LDK(KP707106781), T26, T25);
Chris@82 142 T2J = VFNMS(LDK(KP707106781), T26, T25);
Chris@82 143 T38 = VADD(T3, T7);
Chris@82 144 T39 = VADD(T22, T24);
Chris@82 145 T3a = VSUB(T38, T39);
Chris@82 146 T40 = VADD(T38, T39);
Chris@82 147 T1V = VSUB(T1S, T1U);
Chris@82 148 T1W = VADD(T1J, T1O);
Chris@82 149 T1X = VFNMS(LDK(KP707106781), T1W, T1V);
Chris@82 150 T2C = VFMA(LDK(KP707106781), T1W, T1V);
Chris@82 151 }
Chris@82 152 {
Chris@82 153 V TP, TT, TN, TO, TM, T19, TR, TS, TQ, T17, TY, T12, TW, TX, TV;
Chris@82 154 V T1q, T10, T11, TZ, T1o, Tw, T1i, TA, T1k, Tu, Tv, Tt, T1h, Ty, Tz;
Chris@82 155 V Tx, T1j, TF, T1f, TJ, T1d, TD, TE, TC, T1e, TH, TI, TG, T1c;
Chris@82 156 TN = LD(&(Rp[WS(rs, 3)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 157 TO = LD(&(Rm[WS(rs, 3)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 158 TM = LDW(&(W[TWVL * 10]));
Chris@82 159 TP = VZMULJ(TM, VFMACONJ(TO, TN));
Chris@82 160 T19 = LDW(&(W[TWVL * 12]));
Chris@82 161 T1a = VZMULIJ(T19, VFNMSCONJ(TO, TN));
Chris@82 162 TR = LD(&(Rp[WS(rs, 11)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 163 TS = LD(&(Rm[WS(rs, 11)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 164 TQ = LDW(&(W[TWVL * 42]));
Chris@82 165 TT = VZMULJ(TQ, VFMACONJ(TS, TR));
Chris@82 166 T17 = LDW(&(W[TWVL * 44]));
Chris@82 167 T18 = VZMULIJ(T17, VFNMSCONJ(TS, TR));
Chris@82 168 TU = VSUB(TP, TT);
Chris@82 169 T3m = VADD(T1a, T18);
Chris@82 170 T3f = VADD(TP, TT);
Chris@82 171 TW = LD(&(Rp[WS(rs, 15)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 172 TX = LD(&(Rm[WS(rs, 15)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 173 TV = LDW(&(W[TWVL * 58]));
Chris@82 174 TY = VZMULJ(TV, VFMACONJ(TX, TW));
Chris@82 175 T1q = LDW(&(W[TWVL * 60]));
Chris@82 176 T1r = VZMULIJ(T1q, VFNMSCONJ(TX, TW));
Chris@82 177 T10 = LD(&(Rp[WS(rs, 7)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 178 T11 = LD(&(Rm[WS(rs, 7)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 179 TZ = LDW(&(W[TWVL * 26]));
Chris@82 180 T12 = VZMULJ(TZ, VFMACONJ(T11, T10));
Chris@82 181 T1o = LDW(&(W[TWVL * 28]));
Chris@82 182 T1p = VZMULIJ(T1o, VFNMSCONJ(T11, T10));
Chris@82 183 T13 = VSUB(TY, T12);
Chris@82 184 T3n = VADD(T1r, T1p);
Chris@82 185 T3e = VADD(TY, T12);
Chris@82 186 Tu = LD(&(Rp[WS(rs, 5)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 187 Tv = LD(&(Rm[WS(rs, 5)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 188 Tt = LDW(&(W[TWVL * 18]));
Chris@82 189 Tw = VZMULJ(Tt, VFMACONJ(Tv, Tu));
Chris@82 190 T1h = LDW(&(W[TWVL * 20]));
Chris@82 191 T1i = VZMULIJ(T1h, VFNMSCONJ(Tv, Tu));
Chris@82 192 Ty = LD(&(Rp[WS(rs, 13)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 193 Tz = LD(&(Rm[WS(rs, 13)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 194 Tx = LDW(&(W[TWVL * 50]));
Chris@82 195 TA = VZMULJ(Tx, VFMACONJ(Tz, Ty));
Chris@82 196 T1j = LDW(&(W[TWVL * 52]));
Chris@82 197 T1k = VZMULIJ(T1j, VFNMSCONJ(Tz, Ty));
Chris@82 198 TB = VSUB(Tw, TA);
Chris@82 199 T3k = VADD(T1i, T1k);
Chris@82 200 T1l = VSUB(T1i, T1k);
Chris@82 201 T3c = VADD(Tw, TA);
Chris@82 202 TD = LD(&(Rp[WS(rs, 1)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 203 TE = LD(&(Rm[WS(rs, 1)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 204 TC = LDW(&(W[TWVL * 2]));
Chris@82 205 TF = VZMULJ(TC, VFMACONJ(TE, TD));
Chris@82 206 T1e = LDW(&(W[TWVL * 4]));
Chris@82 207 T1f = VZMULIJ(T1e, VFNMSCONJ(TE, TD));
Chris@82 208 TH = LD(&(Rp[WS(rs, 9)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 209 TI = LD(&(Rm[WS(rs, 9)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 210 TG = LDW(&(W[TWVL * 34]));
Chris@82 211 TJ = VZMULJ(TG, VFMACONJ(TI, TH));
Chris@82 212 T1c = LDW(&(W[TWVL * 36]));
Chris@82 213 T1d = VZMULIJ(T1c, VFNMSCONJ(TI, TH));
Chris@82 214 TK = VSUB(TF, TJ);
Chris@82 215 T3j = VADD(T1f, T1d);
Chris@82 216 T1g = VSUB(T1d, T1f);
Chris@82 217 T3b = VADD(TF, TJ);
Chris@82 218 }
Chris@82 219 T43 = VADD(T3b, T3c);
Chris@82 220 T44 = VADD(T3e, T3f);
Chris@82 221 T4a = VADD(T3j, T3k);
Chris@82 222 T4b = VADD(T3n, T3m);
Chris@82 223 T4m = VSUB(T4a, T4b);
Chris@82 224 T3l = VSUB(T3j, T3k);
Chris@82 225 T3o = VSUB(T3m, T3n);
Chris@82 226 T3p = VFMA(LDK(KP414213562), T3o, T3l);
Chris@82 227 T3E = VFNMS(LDK(KP414213562), T3l, T3o);
Chris@82 228 TL = VFMA(LDK(KP414213562), TK, TB);
Chris@82 229 T14 = VFNMS(LDK(KP414213562), T13, TU);
Chris@82 230 T15 = VSUB(TL, T14);
Chris@82 231 T2K = VADD(TL, T14);
Chris@82 232 T1s = VSUB(T1p, T1r);
Chris@82 233 T1t = VADD(T1g, T1l);
Chris@82 234 T1u = VFNMS(LDK(KP707106781), T1t, T1s);
Chris@82 235 T2F = VFMA(LDK(KP707106781), T1t, T1s);
Chris@82 236 T3d = VSUB(T3b, T3c);
Chris@82 237 T3g = VSUB(T3e, T3f);
Chris@82 238 T3h = VADD(T3d, T3g);
Chris@82 239 T3C = VSUB(T3g, T3d);
Chris@82 240 T1b = VSUB(T18, T1a);
Chris@82 241 T1m = VSUB(T1g, T1l);
Chris@82 242 T1n = VFNMS(LDK(KP707106781), T1m, T1b);
Chris@82 243 T2E = VFMA(LDK(KP707106781), T1m, T1b);
Chris@82 244 T28 = VFMA(LDK(KP414213562), TU, T13);
Chris@82 245 T29 = VFNMS(LDK(KP414213562), TB, TK);
Chris@82 246 T2a = VSUB(T28, T29);
Chris@82 247 T2z = VADD(T29, T28);
Chris@82 248 {
Chris@82 249 V T4o, T4u, T4r, T4v, T4k, T4n, T4p, T4q, T4s, T4x, T4t, T4w, T3y, T3K, T3H;
Chris@82 250 V T3L, T3i, T3x, T3D, T3G, T3I, T3N, T3J, T3M, T46, T4g, T4d, T4h, T42, T45;
Chris@82 251 V T49, T4c, T4e, T4j, T4f, T4i;
Chris@82 252 T4k = VSUB(T40, T41);
Chris@82 253 T4n = VADD(T4l, T4m);
Chris@82 254 T4o = VFMA(LDK(KP707106781), T4n, T4k);
Chris@82 255 T4u = VFNMS(LDK(KP707106781), T4n, T4k);
Chris@82 256 T4p = VSUB(T44, T43);
Chris@82 257 T4q = VSUB(T4m, T4l);
Chris@82 258 T4r = VFMA(LDK(KP707106781), T4q, T4p);
Chris@82 259 T4v = VFNMS(LDK(KP707106781), T4q, T4p);
Chris@82 260 T4s = VCONJ(VMUL(LDK(KP500000000), VFNMSI(T4r, T4o)));
Chris@82 261 ST(&(Rm[WS(rs, 3)]), T4s, -ms, &(Rm[WS(rs, 1)]));
Chris@82 262 T4x = VCONJ(VMUL(LDK(KP500000000), VFMAI(T4v, T4u)));
Chris@82 263 ST(&(Rm[WS(rs, 11)]), T4x, -ms, &(Rm[WS(rs, 1)]));
Chris@82 264 T4t = VMUL(LDK(KP500000000), VFMAI(T4r, T4o));
Chris@82 265 ST(&(Rp[WS(rs, 4)]), T4t, ms, &(Rp[0]));
Chris@82 266 T4w = VMUL(LDK(KP500000000), VFNMSI(T4v, T4u));
Chris@82 267 ST(&(Rp[WS(rs, 12)]), T4w, ms, &(Rp[0]));
Chris@82 268 T3i = VFNMS(LDK(KP707106781), T3h, T3a);
Chris@82 269 T3x = VSUB(T3p, T3w);
Chris@82 270 T3y = VFMA(LDK(KP923879532), T3x, T3i);
Chris@82 271 T3K = VFNMS(LDK(KP923879532), T3x, T3i);
Chris@82 272 T3D = VFNMS(LDK(KP707106781), T3C, T3B);
Chris@82 273 T3G = VSUB(T3E, T3F);
Chris@82 274 T3H = VFNMS(LDK(KP923879532), T3G, T3D);
Chris@82 275 T3L = VFMA(LDK(KP923879532), T3G, T3D);
Chris@82 276 T3I = VMUL(LDK(KP500000000), VFNMSI(T3H, T3y));
Chris@82 277 ST(&(Rp[WS(rs, 6)]), T3I, ms, &(Rp[0]));
Chris@82 278 T3N = VMUL(LDK(KP500000000), VFMAI(T3L, T3K));
Chris@82 279 ST(&(Rp[WS(rs, 10)]), T3N, ms, &(Rp[0]));
Chris@82 280 T3J = VCONJ(VMUL(LDK(KP500000000), VFMAI(T3H, T3y)));
Chris@82 281 ST(&(Rm[WS(rs, 5)]), T3J, -ms, &(Rm[WS(rs, 1)]));
Chris@82 282 T3M = VCONJ(VMUL(LDK(KP500000000), VFNMSI(T3L, T3K)));
Chris@82 283 ST(&(Rm[WS(rs, 9)]), T3M, -ms, &(Rm[WS(rs, 1)]));
Chris@82 284 T42 = VADD(T40, T41);
Chris@82 285 T45 = VADD(T43, T44);
Chris@82 286 T46 = VSUB(T42, T45);
Chris@82 287 T4g = VADD(T42, T45);
Chris@82 288 T49 = VADD(T47, T48);
Chris@82 289 T4c = VADD(T4a, T4b);
Chris@82 290 T4d = VSUB(T49, T4c);
Chris@82 291 T4h = VADD(T49, T4c);
Chris@82 292 T4e = VMUL(LDK(KP500000000), VFMAI(T4d, T46));
Chris@82 293 ST(&(Rp[WS(rs, 8)]), T4e, ms, &(Rp[0]));
Chris@82 294 T4j = VCONJ(VMUL(LDK(KP500000000), VADD(T4h, T4g)));
Chris@82 295 ST(&(Rm[WS(rs, 15)]), T4j, -ms, &(Rm[WS(rs, 1)]));
Chris@82 296 T4f = VCONJ(VMUL(LDK(KP500000000), VFNMSI(T4d, T46)));
Chris@82 297 ST(&(Rm[WS(rs, 7)]), T4f, -ms, &(Rm[WS(rs, 1)]));
Chris@82 298 T4i = VMUL(LDK(KP500000000), VSUB(T4g, T4h));
Chris@82 299 ST(&(Rp[0]), T4i, ms, &(Rp[0]));
Chris@82 300 }
Chris@82 301 T3O = VFMA(LDK(KP707106781), T3h, T3a);
Chris@82 302 T3P = VADD(T3F, T3E);
Chris@82 303 T3Q = VFMA(LDK(KP923879532), T3P, T3O);
Chris@82 304 T3W = VFNMS(LDK(KP923879532), T3P, T3O);
Chris@82 305 T3R = VFMA(LDK(KP707106781), T3C, T3B);
Chris@82 306 T3S = VADD(T3w, T3p);
Chris@82 307 T3T = VFMA(LDK(KP923879532), T3S, T3R);
Chris@82 308 T3X = VFNMS(LDK(KP923879532), T3S, T3R);
Chris@82 309 T3U = VCONJ(VMUL(LDK(KP500000000), VFNMSI(T3T, T3Q)));
Chris@82 310 ST(&(Rm[WS(rs, 1)]), T3U, -ms, &(Rm[WS(rs, 1)]));
Chris@82 311 T3Z = VCONJ(VMUL(LDK(KP500000000), VFMAI(T3X, T3W)));
Chris@82 312 ST(&(Rm[WS(rs, 13)]), T3Z, -ms, &(Rm[WS(rs, 1)]));
Chris@82 313 T3V = VMUL(LDK(KP500000000), VFMAI(T3T, T3Q));
Chris@82 314 ST(&(Rp[WS(rs, 2)]), T3V, ms, &(Rp[0]));
Chris@82 315 T3Y = VMUL(LDK(KP500000000), VFNMSI(T3X, T3W));
Chris@82 316 ST(&(Rp[WS(rs, 14)]), T3Y, ms, &(Rp[0]));
Chris@82 317 {
Chris@82 318 V T2I, T35, T2S, T31, T2P, T34, T2T, T2Y, T2A, T2Z, T2H, T30, T2D, T2G, T2L;
Chris@82 319 V T2W, T2O, T2X, T2M, T2N, T2Q, T36, T37, T2R, T2U, T32, T33, T2V, T20, T2v;
Chris@82 320 V T2i, T2r, T2f, T2u, T2j, T2o, T16, T2p, T1Z, T2q, T1v, T1Y, T2b, T2m, T2e;
Chris@82 321 V T2n, T2c, T2d, T2g, T2w, T2x, T2h, T2k, T2s, T2t, T2l;
Chris@82 322 T2A = VFNMS(LDK(KP923879532), T2z, T2y);
Chris@82 323 T2Z = VFMA(LDK(KP923879532), T2K, T2J);
Chris@82 324 T2D = VFMA(LDK(KP198912367), T2C, T2B);
Chris@82 325 T2G = VFNMS(LDK(KP198912367), T2F, T2E);
Chris@82 326 T2H = VSUB(T2D, T2G);
Chris@82 327 T30 = VADD(T2D, T2G);
Chris@82 328 T2I = VFMA(LDK(KP980785280), T2H, T2A);
Chris@82 329 T35 = VFNMS(LDK(KP980785280), T30, T2Z);
Chris@82 330 T2S = VFNMS(LDK(KP980785280), T2H, T2A);
Chris@82 331 T31 = VFMA(LDK(KP980785280), T30, T2Z);
Chris@82 332 T2L = VFNMS(LDK(KP923879532), T2K, T2J);
Chris@82 333 T2W = VFMA(LDK(KP923879532), T2z, T2y);
Chris@82 334 T2M = VFMA(LDK(KP198912367), T2E, T2F);
Chris@82 335 T2N = VFNMS(LDK(KP198912367), T2B, T2C);
Chris@82 336 T2O = VSUB(T2M, T2N);
Chris@82 337 T2X = VADD(T2N, T2M);
Chris@82 338 T2P = VFMA(LDK(KP980785280), T2O, T2L);
Chris@82 339 T34 = VFNMS(LDK(KP980785280), T2X, T2W);
Chris@82 340 T2T = VFNMS(LDK(KP980785280), T2O, T2L);
Chris@82 341 T2Y = VFMA(LDK(KP980785280), T2X, T2W);
Chris@82 342 T2Q = VCONJ(VMUL(LDK(KP500000000), VFNMSI(T2P, T2I)));
Chris@82 343 ST(&(Rm[WS(rs, 6)]), T2Q, -ms, &(Rm[0]));
Chris@82 344 T36 = VCONJ(VMUL(LDK(KP500000000), VFNMSI(T35, T34)));
Chris@82 345 ST(&(Rm[WS(rs, 14)]), T36, -ms, &(Rm[0]));
Chris@82 346 T37 = VMUL(LDK(KP500000000), VFMAI(T35, T34));
Chris@82 347 ST(&(Rp[WS(rs, 15)]), T37, ms, &(Rp[WS(rs, 1)]));
Chris@82 348 T2R = VMUL(LDK(KP500000000), VFMAI(T2P, T2I));
Chris@82 349 ST(&(Rp[WS(rs, 7)]), T2R, ms, &(Rp[WS(rs, 1)]));
Chris@82 350 T2U = VMUL(LDK(KP500000000), VFNMSI(T2T, T2S));
Chris@82 351 ST(&(Rp[WS(rs, 9)]), T2U, ms, &(Rp[WS(rs, 1)]));
Chris@82 352 T32 = VMUL(LDK(KP500000000), VFNMSI(T31, T2Y));
Chris@82 353 ST(&(Rp[WS(rs, 1)]), T32, ms, &(Rp[WS(rs, 1)]));
Chris@82 354 T33 = VCONJ(VMUL(LDK(KP500000000), VFMAI(T31, T2Y)));
Chris@82 355 ST(&(Rm[0]), T33, -ms, &(Rm[0]));
Chris@82 356 T2V = VCONJ(VMUL(LDK(KP500000000), VFMAI(T2T, T2S)));
Chris@82 357 ST(&(Rm[WS(rs, 8)]), T2V, -ms, &(Rm[0]));
Chris@82 358 T16 = VFNMS(LDK(KP923879532), T15, Ts);
Chris@82 359 T2p = VFMA(LDK(KP923879532), T2a, T27);
Chris@82 360 T1v = VFMA(LDK(KP668178637), T1u, T1n);
Chris@82 361 T1Y = VFNMS(LDK(KP668178637), T1X, T1Q);
Chris@82 362 T1Z = VSUB(T1v, T1Y);
Chris@82 363 T2q = VADD(T1Y, T1v);
Chris@82 364 T20 = VFMA(LDK(KP831469612), T1Z, T16);
Chris@82 365 T2v = VFNMS(LDK(KP831469612), T2q, T2p);
Chris@82 366 T2i = VFNMS(LDK(KP831469612), T1Z, T16);
Chris@82 367 T2r = VFMA(LDK(KP831469612), T2q, T2p);
Chris@82 368 T2b = VFNMS(LDK(KP923879532), T2a, T27);
Chris@82 369 T2m = VFMA(LDK(KP923879532), T15, Ts);
Chris@82 370 T2c = VFNMS(LDK(KP668178637), T1n, T1u);
Chris@82 371 T2d = VFMA(LDK(KP668178637), T1Q, T1X);
Chris@82 372 T2e = VSUB(T2c, T2d);
Chris@82 373 T2n = VADD(T2d, T2c);
Chris@82 374 T2f = VFNMS(LDK(KP831469612), T2e, T2b);
Chris@82 375 T2u = VFNMS(LDK(KP831469612), T2n, T2m);
Chris@82 376 T2j = VFMA(LDK(KP831469612), T2e, T2b);
Chris@82 377 T2o = VFMA(LDK(KP831469612), T2n, T2m);
Chris@82 378 T2g = VMUL(LDK(KP500000000), VFNMSI(T2f, T20));
Chris@82 379 ST(&(Rp[WS(rs, 5)]), T2g, ms, &(Rp[WS(rs, 1)]));
Chris@82 380 T2w = VMUL(LDK(KP500000000), VFNMSI(T2v, T2u));
Chris@82 381 ST(&(Rp[WS(rs, 13)]), T2w, ms, &(Rp[WS(rs, 1)]));
Chris@82 382 T2x = VCONJ(VMUL(LDK(KP500000000), VFMAI(T2v, T2u)));
Chris@82 383 ST(&(Rm[WS(rs, 12)]), T2x, -ms, &(Rm[0]));
Chris@82 384 T2h = VCONJ(VMUL(LDK(KP500000000), VFMAI(T2f, T20)));
Chris@82 385 ST(&(Rm[WS(rs, 4)]), T2h, -ms, &(Rm[0]));
Chris@82 386 T2k = VMUL(LDK(KP500000000), VFMAI(T2j, T2i));
Chris@82 387 ST(&(Rp[WS(rs, 11)]), T2k, ms, &(Rp[WS(rs, 1)]));
Chris@82 388 T2s = VMUL(LDK(KP500000000), VFMAI(T2r, T2o));
Chris@82 389 ST(&(Rp[WS(rs, 3)]), T2s, ms, &(Rp[WS(rs, 1)]));
Chris@82 390 T2t = VCONJ(VMUL(LDK(KP500000000), VFNMSI(T2r, T2o)));
Chris@82 391 ST(&(Rm[WS(rs, 2)]), T2t, -ms, &(Rm[0]));
Chris@82 392 T2l = VCONJ(VMUL(LDK(KP500000000), VFNMSI(T2j, T2i)));
Chris@82 393 ST(&(Rm[WS(rs, 10)]), T2l, -ms, &(Rm[0]));
Chris@82 394 }
Chris@82 395 }
Chris@82 396 }
Chris@82 397 VLEAVE();
Chris@82 398 }
Chris@82 399
Chris@82 400 static const tw_instr twinstr[] = {
Chris@82 401 VTW(1, 1),
Chris@82 402 VTW(1, 2),
Chris@82 403 VTW(1, 3),
Chris@82 404 VTW(1, 4),
Chris@82 405 VTW(1, 5),
Chris@82 406 VTW(1, 6),
Chris@82 407 VTW(1, 7),
Chris@82 408 VTW(1, 8),
Chris@82 409 VTW(1, 9),
Chris@82 410 VTW(1, 10),
Chris@82 411 VTW(1, 11),
Chris@82 412 VTW(1, 12),
Chris@82 413 VTW(1, 13),
Chris@82 414 VTW(1, 14),
Chris@82 415 VTW(1, 15),
Chris@82 416 VTW(1, 16),
Chris@82 417 VTW(1, 17),
Chris@82 418 VTW(1, 18),
Chris@82 419 VTW(1, 19),
Chris@82 420 VTW(1, 20),
Chris@82 421 VTW(1, 21),
Chris@82 422 VTW(1, 22),
Chris@82 423 VTW(1, 23),
Chris@82 424 VTW(1, 24),
Chris@82 425 VTW(1, 25),
Chris@82 426 VTW(1, 26),
Chris@82 427 VTW(1, 27),
Chris@82 428 VTW(1, 28),
Chris@82 429 VTW(1, 29),
Chris@82 430 VTW(1, 30),
Chris@82 431 VTW(1, 31),
Chris@82 432 {TW_NEXT, VL, 0}
Chris@82 433 };
Chris@82 434
Chris@82 435 static const hc2c_desc desc = { 32, XSIMD_STRING("hc2cfdftv_32"), twinstr, &GENUS, {119, 94, 130, 0} };
Chris@82 436
Chris@82 437 void XSIMD(codelet_hc2cfdftv_32) (planner *p) {
Chris@82 438 X(khc2c_register) (p, hc2cfdftv_32, &desc, HC2C_VIA_DFT);
Chris@82 439 }
Chris@82 440 #else
Chris@82 441
Chris@82 442 /* Generated by: ../../../genfft/gen_hc2cdft_c.native -simd -compact -variables 4 -pipeline-latency 8 -trivial-stores -variables 32 -no-generate-bytw -n 32 -dit -name hc2cfdftv_32 -include rdft/simd/hc2cfv.h */
Chris@82 443
Chris@82 444 /*
Chris@82 445 * This function contains 249 FP additions, 133 FP multiplications,
Chris@82 446 * (or, 233 additions, 117 multiplications, 16 fused multiply/add),
Chris@82 447 * 130 stack variables, 9 constants, and 64 memory accesses
Chris@82 448 */
Chris@82 449 #include "rdft/simd/hc2cfv.h"
Chris@82 450
Chris@82 451 static void hc2cfdftv_32(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
Chris@82 452 {
Chris@82 453 DVK(KP555570233, +0.555570233019602224742830813948532874374937191);
Chris@82 454 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
Chris@82 455 DVK(KP195090322, +0.195090322016128267848284868477022240927691618);
Chris@82 456 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
Chris@82 457 DVK(KP382683432, +0.382683432365089771728459984030398866761344562);
Chris@82 458 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
Chris@82 459 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
Chris@82 460 DVK(KP353553390, +0.353553390593273762200422181052424519642417969);
Chris@82 461 DVK(KP500000000, +0.500000000000000000000000000000000000000000000);
Chris@82 462 {
Chris@82 463 INT m;
Chris@82 464 for (m = mb, W = W + ((mb - 1) * ((TWVL / VL) * 62)); m < me; m = m + VL, Rp = Rp + (VL * ms), Ip = Ip + (VL * ms), Rm = Rm - (VL * ms), Im = Im - (VL * ms), W = W + (TWVL * 62), MAKE_VOLATILE_STRIDE(128, rs)) {
Chris@82 465 V Ta, T2m, Tx, T2h, T3R, T4h, T3q, T4g, T3B, T4n, T3E, T4o, T1B, T2S, T1O;
Chris@82 466 V T2R, TV, T2p, T1i, T2o, T3L, T4q, T3I, T4r, T3w, T4k, T3t, T4j, T26, T2V;
Chris@82 467 V T2d, T2U;
Chris@82 468 {
Chris@82 469 V T4, T1m, T1H, T2j, T1M, T2l, T9, T1o, Tf, T1r, Tq, T1w, Tv, T1y, Tk;
Chris@82 470 V T1t, Tl, Tw, T3P, T3Q, T3o, T3p, T3z, T3A, T3C, T3D, T1p, T1N, T1A, T1C;
Chris@82 471 V T1u, T1z;
Chris@82 472 {
Chris@82 473 V T1, T3, T2, T1l, T1G, T1F, T1E, T1D, T2i, T1L, T1K, T1J, T1I, T2k, T6;
Chris@82 474 V T8, T7, T5, T1n, Tc, Te, Td, Tb, T1q, Tn, Tp, To, Tm, T1v, Ts;
Chris@82 475 V Tu, Tt, Tr, T1x, Th, Tj, Ti, Tg, T1s;
Chris@82 476 T1 = LD(&(Rp[0]), ms, &(Rp[0]));
Chris@82 477 T2 = LD(&(Rm[0]), -ms, &(Rm[0]));
Chris@82 478 T3 = VCONJ(T2);
Chris@82 479 T4 = VADD(T1, T3);
Chris@82 480 T1l = LDW(&(W[0]));
Chris@82 481 T1m = VZMULIJ(T1l, VSUB(T3, T1));
Chris@82 482 T1G = LD(&(Rp[WS(rs, 4)]), ms, &(Rp[0]));
Chris@82 483 T1E = LD(&(Rm[WS(rs, 4)]), -ms, &(Rm[0]));
Chris@82 484 T1F = VCONJ(T1E);
Chris@82 485 T1D = LDW(&(W[TWVL * 16]));
Chris@82 486 T1H = VZMULIJ(T1D, VSUB(T1F, T1G));
Chris@82 487 T2i = LDW(&(W[TWVL * 14]));
Chris@82 488 T2j = VZMULJ(T2i, VADD(T1G, T1F));
Chris@82 489 T1L = LD(&(Rp[WS(rs, 12)]), ms, &(Rp[0]));
Chris@82 490 T1J = LD(&(Rm[WS(rs, 12)]), -ms, &(Rm[0]));
Chris@82 491 T1K = VCONJ(T1J);
Chris@82 492 T1I = LDW(&(W[TWVL * 48]));
Chris@82 493 T1M = VZMULIJ(T1I, VSUB(T1K, T1L));
Chris@82 494 T2k = LDW(&(W[TWVL * 46]));
Chris@82 495 T2l = VZMULJ(T2k, VADD(T1L, T1K));
Chris@82 496 T6 = LD(&(Rp[WS(rs, 8)]), ms, &(Rp[0]));
Chris@82 497 T7 = LD(&(Rm[WS(rs, 8)]), -ms, &(Rm[0]));
Chris@82 498 T8 = VCONJ(T7);
Chris@82 499 T5 = LDW(&(W[TWVL * 30]));
Chris@82 500 T9 = VZMULJ(T5, VADD(T6, T8));
Chris@82 501 T1n = LDW(&(W[TWVL * 32]));
Chris@82 502 T1o = VZMULIJ(T1n, VSUB(T8, T6));
Chris@82 503 Tc = LD(&(Rp[WS(rs, 2)]), ms, &(Rp[0]));
Chris@82 504 Td = LD(&(Rm[WS(rs, 2)]), -ms, &(Rm[0]));
Chris@82 505 Te = VCONJ(Td);
Chris@82 506 Tb = LDW(&(W[TWVL * 6]));
Chris@82 507 Tf = VZMULJ(Tb, VADD(Tc, Te));
Chris@82 508 T1q = LDW(&(W[TWVL * 8]));
Chris@82 509 T1r = VZMULIJ(T1q, VSUB(Te, Tc));
Chris@82 510 Tn = LD(&(Rp[WS(rs, 14)]), ms, &(Rp[0]));
Chris@82 511 To = LD(&(Rm[WS(rs, 14)]), -ms, &(Rm[0]));
Chris@82 512 Tp = VCONJ(To);
Chris@82 513 Tm = LDW(&(W[TWVL * 54]));
Chris@82 514 Tq = VZMULJ(Tm, VADD(Tn, Tp));
Chris@82 515 T1v = LDW(&(W[TWVL * 56]));
Chris@82 516 T1w = VZMULIJ(T1v, VSUB(Tp, Tn));
Chris@82 517 Ts = LD(&(Rp[WS(rs, 6)]), ms, &(Rp[0]));
Chris@82 518 Tt = LD(&(Rm[WS(rs, 6)]), -ms, &(Rm[0]));
Chris@82 519 Tu = VCONJ(Tt);
Chris@82 520 Tr = LDW(&(W[TWVL * 22]));
Chris@82 521 Tv = VZMULJ(Tr, VADD(Ts, Tu));
Chris@82 522 T1x = LDW(&(W[TWVL * 24]));
Chris@82 523 T1y = VZMULIJ(T1x, VSUB(Tu, Ts));
Chris@82 524 Th = LD(&(Rp[WS(rs, 10)]), ms, &(Rp[0]));
Chris@82 525 Ti = LD(&(Rm[WS(rs, 10)]), -ms, &(Rm[0]));
Chris@82 526 Tj = VCONJ(Ti);
Chris@82 527 Tg = LDW(&(W[TWVL * 38]));
Chris@82 528 Tk = VZMULJ(Tg, VADD(Th, Tj));
Chris@82 529 T1s = LDW(&(W[TWVL * 40]));
Chris@82 530 T1t = VZMULIJ(T1s, VSUB(Tj, Th));
Chris@82 531 }
Chris@82 532 Ta = VMUL(LDK(KP500000000), VSUB(T4, T9));
Chris@82 533 T2m = VSUB(T2j, T2l);
Chris@82 534 Tl = VSUB(Tf, Tk);
Chris@82 535 Tw = VSUB(Tq, Tv);
Chris@82 536 Tx = VMUL(LDK(KP353553390), VADD(Tl, Tw));
Chris@82 537 T2h = VMUL(LDK(KP707106781), VSUB(Tw, Tl));
Chris@82 538 T3P = VADD(Tq, Tv);
Chris@82 539 T3Q = VADD(Tf, Tk);
Chris@82 540 T3R = VSUB(T3P, T3Q);
Chris@82 541 T4h = VADD(T3Q, T3P);
Chris@82 542 T3o = VADD(T4, T9);
Chris@82 543 T3p = VADD(T2j, T2l);
Chris@82 544 T3q = VMUL(LDK(KP500000000), VSUB(T3o, T3p));
Chris@82 545 T4g = VADD(T3o, T3p);
Chris@82 546 T3z = VADD(T1m, T1o);
Chris@82 547 T3A = VADD(T1H, T1M);
Chris@82 548 T3B = VSUB(T3z, T3A);
Chris@82 549 T4n = VADD(T3z, T3A);
Chris@82 550 T3C = VADD(T1w, T1y);
Chris@82 551 T3D = VADD(T1r, T1t);
Chris@82 552 T3E = VSUB(T3C, T3D);
Chris@82 553 T4o = VADD(T3D, T3C);
Chris@82 554 T1p = VSUB(T1m, T1o);
Chris@82 555 T1N = VSUB(T1H, T1M);
Chris@82 556 T1u = VSUB(T1r, T1t);
Chris@82 557 T1z = VSUB(T1w, T1y);
Chris@82 558 T1A = VMUL(LDK(KP707106781), VADD(T1u, T1z));
Chris@82 559 T1C = VMUL(LDK(KP707106781), VSUB(T1z, T1u));
Chris@82 560 T1B = VADD(T1p, T1A);
Chris@82 561 T2S = VADD(T1N, T1C);
Chris@82 562 T1O = VSUB(T1C, T1N);
Chris@82 563 T2R = VSUB(T1p, T1A);
Chris@82 564 }
Chris@82 565 {
Chris@82 566 V TD, T1R, T1b, T29, T1g, T2b, TI, T1T, TO, T1Y, T10, T22, T15, T24, TT;
Chris@82 567 V T1W, TJ, TU, T16, T1h, T3J, T3K, T3G, T3H, T3u, T3v, T3r, T3s, T25, T2c;
Chris@82 568 V T20, T27, T1U, T1Z;
Chris@82 569 {
Chris@82 570 V TA, TC, TB, Tz, T1Q, T18, T1a, T19, T17, T28, T1d, T1f, T1e, T1c, T2a;
Chris@82 571 V TF, TH, TG, TE, T1S, TL, TN, TM, TK, T1X, TX, TZ, TY, TW, T21;
Chris@82 572 V T12, T14, T13, T11, T23, TQ, TS, TR, TP, T1V;
Chris@82 573 TA = LD(&(Rp[WS(rs, 1)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 574 TB = LD(&(Rm[WS(rs, 1)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 575 TC = VCONJ(TB);
Chris@82 576 Tz = LDW(&(W[TWVL * 2]));
Chris@82 577 TD = VZMULJ(Tz, VADD(TA, TC));
Chris@82 578 T1Q = LDW(&(W[TWVL * 4]));
Chris@82 579 T1R = VZMULIJ(T1Q, VSUB(TC, TA));
Chris@82 580 T18 = LD(&(Rp[WS(rs, 3)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 581 T19 = LD(&(Rm[WS(rs, 3)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 582 T1a = VCONJ(T19);
Chris@82 583 T17 = LDW(&(W[TWVL * 10]));
Chris@82 584 T1b = VZMULJ(T17, VADD(T18, T1a));
Chris@82 585 T28 = LDW(&(W[TWVL * 12]));
Chris@82 586 T29 = VZMULIJ(T28, VSUB(T1a, T18));
Chris@82 587 T1d = LD(&(Rp[WS(rs, 11)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 588 T1e = LD(&(Rm[WS(rs, 11)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 589 T1f = VCONJ(T1e);
Chris@82 590 T1c = LDW(&(W[TWVL * 42]));
Chris@82 591 T1g = VZMULJ(T1c, VADD(T1d, T1f));
Chris@82 592 T2a = LDW(&(W[TWVL * 44]));
Chris@82 593 T2b = VZMULIJ(T2a, VSUB(T1f, T1d));
Chris@82 594 TF = LD(&(Rp[WS(rs, 9)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 595 TG = LD(&(Rm[WS(rs, 9)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 596 TH = VCONJ(TG);
Chris@82 597 TE = LDW(&(W[TWVL * 34]));
Chris@82 598 TI = VZMULJ(TE, VADD(TF, TH));
Chris@82 599 T1S = LDW(&(W[TWVL * 36]));
Chris@82 600 T1T = VZMULIJ(T1S, VSUB(TH, TF));
Chris@82 601 TL = LD(&(Rp[WS(rs, 5)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 602 TM = LD(&(Rm[WS(rs, 5)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 603 TN = VCONJ(TM);
Chris@82 604 TK = LDW(&(W[TWVL * 18]));
Chris@82 605 TO = VZMULJ(TK, VADD(TL, TN));
Chris@82 606 T1X = LDW(&(W[TWVL * 20]));
Chris@82 607 T1Y = VZMULIJ(T1X, VSUB(TN, TL));
Chris@82 608 TX = LD(&(Rp[WS(rs, 15)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 609 TY = LD(&(Rm[WS(rs, 15)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 610 TZ = VCONJ(TY);
Chris@82 611 TW = LDW(&(W[TWVL * 58]));
Chris@82 612 T10 = VZMULJ(TW, VADD(TX, TZ));
Chris@82 613 T21 = LDW(&(W[TWVL * 60]));
Chris@82 614 T22 = VZMULIJ(T21, VSUB(TZ, TX));
Chris@82 615 T12 = LD(&(Rp[WS(rs, 7)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 616 T13 = LD(&(Rm[WS(rs, 7)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 617 T14 = VCONJ(T13);
Chris@82 618 T11 = LDW(&(W[TWVL * 26]));
Chris@82 619 T15 = VZMULJ(T11, VADD(T12, T14));
Chris@82 620 T23 = LDW(&(W[TWVL * 28]));
Chris@82 621 T24 = VZMULIJ(T23, VSUB(T14, T12));
Chris@82 622 TQ = LD(&(Rp[WS(rs, 13)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 623 TR = LD(&(Rm[WS(rs, 13)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 624 TS = VCONJ(TR);
Chris@82 625 TP = LDW(&(W[TWVL * 50]));
Chris@82 626 TT = VZMULJ(TP, VADD(TQ, TS));
Chris@82 627 T1V = LDW(&(W[TWVL * 52]));
Chris@82 628 T1W = VZMULIJ(T1V, VSUB(TS, TQ));
Chris@82 629 }
Chris@82 630 TJ = VSUB(TD, TI);
Chris@82 631 TU = VSUB(TO, TT);
Chris@82 632 TV = VFNMS(LDK(KP382683432), TU, VMUL(LDK(KP923879532), TJ));
Chris@82 633 T2p = VFMA(LDK(KP382683432), TJ, VMUL(LDK(KP923879532), TU));
Chris@82 634 T16 = VSUB(T10, T15);
Chris@82 635 T1h = VSUB(T1b, T1g);
Chris@82 636 T1i = VFMA(LDK(KP923879532), T16, VMUL(LDK(KP382683432), T1h));
Chris@82 637 T2o = VFNMS(LDK(KP923879532), T1h, VMUL(LDK(KP382683432), T16));
Chris@82 638 T3J = VADD(T1Y, T1W);
Chris@82 639 T3K = VADD(T1R, T1T);
Chris@82 640 T3L = VSUB(T3J, T3K);
Chris@82 641 T4q = VADD(T3K, T3J);
Chris@82 642 T3G = VADD(T22, T24);
Chris@82 643 T3H = VADD(T29, T2b);
Chris@82 644 T3I = VSUB(T3G, T3H);
Chris@82 645 T4r = VADD(T3G, T3H);
Chris@82 646 T3u = VADD(T10, T15);
Chris@82 647 T3v = VADD(T1b, T1g);
Chris@82 648 T3w = VSUB(T3u, T3v);
Chris@82 649 T4k = VADD(T3u, T3v);
Chris@82 650 T3r = VADD(TD, TI);
Chris@82 651 T3s = VADD(TO, TT);
Chris@82 652 T3t = VSUB(T3r, T3s);
Chris@82 653 T4j = VADD(T3r, T3s);
Chris@82 654 T25 = VSUB(T22, T24);
Chris@82 655 T2c = VSUB(T29, T2b);
Chris@82 656 T1U = VSUB(T1R, T1T);
Chris@82 657 T1Z = VSUB(T1W, T1Y);
Chris@82 658 T20 = VMUL(LDK(KP707106781), VADD(T1U, T1Z));
Chris@82 659 T27 = VMUL(LDK(KP707106781), VSUB(T1Z, T1U));
Chris@82 660 T26 = VADD(T20, T25);
Chris@82 661 T2V = VADD(T27, T2c);
Chris@82 662 T2d = VSUB(T27, T2c);
Chris@82 663 T2U = VSUB(T25, T20);
Chris@82 664 }
Chris@82 665 {
Chris@82 666 V T4m, T4w, T4t, T4x, T4i, T4l, T4p, T4s, T4u, T4z, T4v, T4y, T4E, T4L, T4H;
Chris@82 667 V T4K, T4A, T4F, T4D, T4G, T4B, T4C, T4I, T4N, T4J, T4M, T3O, T4c, T4d, T3X;
Chris@82 668 V T40, T46, T49, T41, T3y, T47, T3T, T45, T3N, T44, T3W, T48, T3x, T3S, T3F;
Chris@82 669 V T3M, T3U, T3V, T3Y, T4e, T4f, T3Z, T42, T4a, T4b, T43;
Chris@82 670 T4i = VADD(T4g, T4h);
Chris@82 671 T4l = VADD(T4j, T4k);
Chris@82 672 T4m = VADD(T4i, T4l);
Chris@82 673 T4w = VSUB(T4i, T4l);
Chris@82 674 T4p = VADD(T4n, T4o);
Chris@82 675 T4s = VADD(T4q, T4r);
Chris@82 676 T4t = VADD(T4p, T4s);
Chris@82 677 T4x = VBYI(VSUB(T4s, T4p));
Chris@82 678 T4u = VCONJ(VMUL(LDK(KP500000000), VSUB(T4m, T4t)));
Chris@82 679 ST(&(Rm[WS(rs, 15)]), T4u, -ms, &(Rm[WS(rs, 1)]));
Chris@82 680 T4z = VMUL(LDK(KP500000000), VADD(T4w, T4x));
Chris@82 681 ST(&(Rp[WS(rs, 8)]), T4z, ms, &(Rp[0]));
Chris@82 682 T4v = VMUL(LDK(KP500000000), VADD(T4m, T4t));
Chris@82 683 ST(&(Rp[0]), T4v, ms, &(Rp[0]));
Chris@82 684 T4y = VCONJ(VMUL(LDK(KP500000000), VSUB(T4w, T4x)));
Chris@82 685 ST(&(Rm[WS(rs, 7)]), T4y, -ms, &(Rm[WS(rs, 1)]));
Chris@82 686 T4A = VMUL(LDK(KP500000000), VSUB(T4g, T4h));
Chris@82 687 T4F = VSUB(T4k, T4j);
Chris@82 688 T4B = VSUB(T4n, T4o);
Chris@82 689 T4C = VSUB(T4r, T4q);
Chris@82 690 T4D = VMUL(LDK(KP353553390), VADD(T4B, T4C));
Chris@82 691 T4G = VMUL(LDK(KP707106781), VSUB(T4C, T4B));
Chris@82 692 T4E = VADD(T4A, T4D);
Chris@82 693 T4L = VMUL(LDK(KP500000000), VBYI(VSUB(T4G, T4F)));
Chris@82 694 T4H = VMUL(LDK(KP500000000), VBYI(VADD(T4F, T4G)));
Chris@82 695 T4K = VSUB(T4A, T4D);
Chris@82 696 T4I = VCONJ(VSUB(T4E, T4H));
Chris@82 697 ST(&(Rm[WS(rs, 3)]), T4I, -ms, &(Rm[WS(rs, 1)]));
Chris@82 698 T4N = VADD(T4K, T4L);
Chris@82 699 ST(&(Rp[WS(rs, 12)]), T4N, ms, &(Rp[0]));
Chris@82 700 T4J = VADD(T4E, T4H);
Chris@82 701 ST(&(Rp[WS(rs, 4)]), T4J, ms, &(Rp[0]));
Chris@82 702 T4M = VCONJ(VSUB(T4K, T4L));
Chris@82 703 ST(&(Rm[WS(rs, 11)]), T4M, -ms, &(Rm[WS(rs, 1)]));
Chris@82 704 T3x = VMUL(LDK(KP353553390), VADD(T3t, T3w));
Chris@82 705 T3y = VADD(T3q, T3x);
Chris@82 706 T47 = VSUB(T3q, T3x);
Chris@82 707 T3S = VMUL(LDK(KP707106781), VSUB(T3w, T3t));
Chris@82 708 T3T = VADD(T3R, T3S);
Chris@82 709 T45 = VSUB(T3S, T3R);
Chris@82 710 T3F = VFMA(LDK(KP923879532), T3B, VMUL(LDK(KP382683432), T3E));
Chris@82 711 T3M = VFNMS(LDK(KP382683432), T3L, VMUL(LDK(KP923879532), T3I));
Chris@82 712 T3N = VMUL(LDK(KP500000000), VADD(T3F, T3M));
Chris@82 713 T44 = VSUB(T3M, T3F);
Chris@82 714 T3U = VFNMS(LDK(KP382683432), T3B, VMUL(LDK(KP923879532), T3E));
Chris@82 715 T3V = VFMA(LDK(KP923879532), T3L, VMUL(LDK(KP382683432), T3I));
Chris@82 716 T3W = VADD(T3U, T3V);
Chris@82 717 T48 = VMUL(LDK(KP500000000), VSUB(T3V, T3U));
Chris@82 718 T3O = VADD(T3y, T3N);
Chris@82 719 T4c = VMUL(LDK(KP500000000), VBYI(VADD(T45, T44)));
Chris@82 720 T4d = VADD(T47, T48);
Chris@82 721 T3X = VMUL(LDK(KP500000000), VBYI(VADD(T3T, T3W)));
Chris@82 722 T40 = VSUB(T3y, T3N);
Chris@82 723 T46 = VMUL(LDK(KP500000000), VBYI(VSUB(T44, T45)));
Chris@82 724 T49 = VSUB(T47, T48);
Chris@82 725 T41 = VMUL(LDK(KP500000000), VBYI(VSUB(T3W, T3T)));
Chris@82 726 T3Y = VCONJ(VSUB(T3O, T3X));
Chris@82 727 ST(&(Rm[WS(rs, 1)]), T3Y, -ms, &(Rm[WS(rs, 1)]));
Chris@82 728 T4e = VADD(T4c, T4d);
Chris@82 729 ST(&(Rp[WS(rs, 6)]), T4e, ms, &(Rp[0]));
Chris@82 730 T4f = VCONJ(VSUB(T4d, T4c));
Chris@82 731 ST(&(Rm[WS(rs, 5)]), T4f, -ms, &(Rm[WS(rs, 1)]));
Chris@82 732 T3Z = VADD(T3O, T3X);
Chris@82 733 ST(&(Rp[WS(rs, 2)]), T3Z, ms, &(Rp[0]));
Chris@82 734 T42 = VCONJ(VSUB(T40, T41));
Chris@82 735 ST(&(Rm[WS(rs, 13)]), T42, -ms, &(Rm[WS(rs, 1)]));
Chris@82 736 T4a = VADD(T46, T49);
Chris@82 737 ST(&(Rp[WS(rs, 10)]), T4a, ms, &(Rp[0]));
Chris@82 738 T4b = VCONJ(VSUB(T49, T46));
Chris@82 739 ST(&(Rm[WS(rs, 9)]), T4b, -ms, &(Rm[WS(rs, 1)]));
Chris@82 740 T43 = VADD(T40, T41);
Chris@82 741 ST(&(Rp[WS(rs, 14)]), T43, ms, &(Rp[0]));
Chris@82 742 {
Chris@82 743 V T2g, T2K, T2L, T2v, T2y, T2E, T2H, T2z, T1k, T2F, T2u, T2G, T2f, T2C, T2r;
Chris@82 744 V T2D, Ty, T1j, T2s, T2t, T1P, T2e, T2n, T2q, T2w, T2M, T2N, T2x, T2A, T2I;
Chris@82 745 V T2J, T2B;
Chris@82 746 Ty = VADD(Ta, Tx);
Chris@82 747 T1j = VMUL(LDK(KP500000000), VADD(TV, T1i));
Chris@82 748 T1k = VADD(Ty, T1j);
Chris@82 749 T2F = VSUB(Ty, T1j);
Chris@82 750 T2s = VFNMS(LDK(KP195090322), T1B, VMUL(LDK(KP980785280), T1O));
Chris@82 751 T2t = VFMA(LDK(KP195090322), T26, VMUL(LDK(KP980785280), T2d));
Chris@82 752 T2u = VADD(T2s, T2t);
Chris@82 753 T2G = VMUL(LDK(KP500000000), VSUB(T2t, T2s));
Chris@82 754 T1P = VFMA(LDK(KP980785280), T1B, VMUL(LDK(KP195090322), T1O));
Chris@82 755 T2e = VFNMS(LDK(KP195090322), T2d, VMUL(LDK(KP980785280), T26));
Chris@82 756 T2f = VMUL(LDK(KP500000000), VADD(T1P, T2e));
Chris@82 757 T2C = VSUB(T2e, T1P);
Chris@82 758 T2n = VSUB(T2h, T2m);
Chris@82 759 T2q = VSUB(T2o, T2p);
Chris@82 760 T2r = VADD(T2n, T2q);
Chris@82 761 T2D = VSUB(T2q, T2n);
Chris@82 762 T2g = VADD(T1k, T2f);
Chris@82 763 T2K = VMUL(LDK(KP500000000), VBYI(VADD(T2D, T2C)));
Chris@82 764 T2L = VADD(T2F, T2G);
Chris@82 765 T2v = VMUL(LDK(KP500000000), VBYI(VADD(T2r, T2u)));
Chris@82 766 T2y = VSUB(T1k, T2f);
Chris@82 767 T2E = VMUL(LDK(KP500000000), VBYI(VSUB(T2C, T2D)));
Chris@82 768 T2H = VSUB(T2F, T2G);
Chris@82 769 T2z = VMUL(LDK(KP500000000), VBYI(VSUB(T2u, T2r)));
Chris@82 770 T2w = VCONJ(VSUB(T2g, T2v));
Chris@82 771 ST(&(Rm[0]), T2w, -ms, &(Rm[0]));
Chris@82 772 T2M = VADD(T2K, T2L);
Chris@82 773 ST(&(Rp[WS(rs, 7)]), T2M, ms, &(Rp[WS(rs, 1)]));
Chris@82 774 T2N = VCONJ(VSUB(T2L, T2K));
Chris@82 775 ST(&(Rm[WS(rs, 6)]), T2N, -ms, &(Rm[0]));
Chris@82 776 T2x = VADD(T2g, T2v);
Chris@82 777 ST(&(Rp[WS(rs, 1)]), T2x, ms, &(Rp[WS(rs, 1)]));
Chris@82 778 T2A = VCONJ(VSUB(T2y, T2z));
Chris@82 779 ST(&(Rm[WS(rs, 14)]), T2A, -ms, &(Rm[0]));
Chris@82 780 T2I = VADD(T2E, T2H);
Chris@82 781 ST(&(Rp[WS(rs, 9)]), T2I, ms, &(Rp[WS(rs, 1)]));
Chris@82 782 T2J = VCONJ(VSUB(T2H, T2E));
Chris@82 783 ST(&(Rm[WS(rs, 8)]), T2J, -ms, &(Rm[0]));
Chris@82 784 T2B = VADD(T2y, T2z);
Chris@82 785 ST(&(Rp[WS(rs, 15)]), T2B, ms, &(Rp[WS(rs, 1)]));
Chris@82 786 }
Chris@82 787 {
Chris@82 788 V T2Y, T3k, T3l, T35, T38, T3e, T3h, T39, T2Q, T3f, T34, T3g, T2X, T3c, T31;
Chris@82 789 V T3d, T2O, T2P, T32, T33, T2T, T2W, T2Z, T30, T36, T3m, T3n, T37, T3a, T3i;
Chris@82 790 V T3j, T3b;
Chris@82 791 T2O = VSUB(Ta, Tx);
Chris@82 792 T2P = VMUL(LDK(KP500000000), VADD(T2p, T2o));
Chris@82 793 T2Q = VADD(T2O, T2P);
Chris@82 794 T3f = VSUB(T2O, T2P);
Chris@82 795 T32 = VFNMS(LDK(KP555570233), T2R, VMUL(LDK(KP831469612), T2S));
Chris@82 796 T33 = VFMA(LDK(KP555570233), T2U, VMUL(LDK(KP831469612), T2V));
Chris@82 797 T34 = VADD(T32, T33);
Chris@82 798 T3g = VMUL(LDK(KP500000000), VSUB(T33, T32));
Chris@82 799 T2T = VFMA(LDK(KP831469612), T2R, VMUL(LDK(KP555570233), T2S));
Chris@82 800 T2W = VFNMS(LDK(KP555570233), T2V, VMUL(LDK(KP831469612), T2U));
Chris@82 801 T2X = VMUL(LDK(KP500000000), VADD(T2T, T2W));
Chris@82 802 T3c = VSUB(T2W, T2T);
Chris@82 803 T2Z = VADD(T2m, T2h);
Chris@82 804 T30 = VSUB(T1i, TV);
Chris@82 805 T31 = VADD(T2Z, T30);
Chris@82 806 T3d = VSUB(T30, T2Z);
Chris@82 807 T2Y = VADD(T2Q, T2X);
Chris@82 808 T3k = VMUL(LDK(KP500000000), VBYI(VADD(T3d, T3c)));
Chris@82 809 T3l = VADD(T3f, T3g);
Chris@82 810 T35 = VMUL(LDK(KP500000000), VBYI(VADD(T31, T34)));
Chris@82 811 T38 = VSUB(T2Q, T2X);
Chris@82 812 T3e = VMUL(LDK(KP500000000), VBYI(VSUB(T3c, T3d)));
Chris@82 813 T3h = VSUB(T3f, T3g);
Chris@82 814 T39 = VMUL(LDK(KP500000000), VBYI(VSUB(T34, T31)));
Chris@82 815 T36 = VCONJ(VSUB(T2Y, T35));
Chris@82 816 ST(&(Rm[WS(rs, 2)]), T36, -ms, &(Rm[0]));
Chris@82 817 T3m = VADD(T3k, T3l);
Chris@82 818 ST(&(Rp[WS(rs, 5)]), T3m, ms, &(Rp[WS(rs, 1)]));
Chris@82 819 T3n = VCONJ(VSUB(T3l, T3k));
Chris@82 820 ST(&(Rm[WS(rs, 4)]), T3n, -ms, &(Rm[0]));
Chris@82 821 T37 = VADD(T2Y, T35);
Chris@82 822 ST(&(Rp[WS(rs, 3)]), T37, ms, &(Rp[WS(rs, 1)]));
Chris@82 823 T3a = VCONJ(VSUB(T38, T39));
Chris@82 824 ST(&(Rm[WS(rs, 12)]), T3a, -ms, &(Rm[0]));
Chris@82 825 T3i = VADD(T3e, T3h);
Chris@82 826 ST(&(Rp[WS(rs, 11)]), T3i, ms, &(Rp[WS(rs, 1)]));
Chris@82 827 T3j = VCONJ(VSUB(T3h, T3e));
Chris@82 828 ST(&(Rm[WS(rs, 10)]), T3j, -ms, &(Rm[0]));
Chris@82 829 T3b = VADD(T38, T39);
Chris@82 830 ST(&(Rp[WS(rs, 13)]), T3b, ms, &(Rp[WS(rs, 1)]));
Chris@82 831 }
Chris@82 832 }
Chris@82 833 }
Chris@82 834 }
Chris@82 835 VLEAVE();
Chris@82 836 }
Chris@82 837
Chris@82 838 static const tw_instr twinstr[] = {
Chris@82 839 VTW(1, 1),
Chris@82 840 VTW(1, 2),
Chris@82 841 VTW(1, 3),
Chris@82 842 VTW(1, 4),
Chris@82 843 VTW(1, 5),
Chris@82 844 VTW(1, 6),
Chris@82 845 VTW(1, 7),
Chris@82 846 VTW(1, 8),
Chris@82 847 VTW(1, 9),
Chris@82 848 VTW(1, 10),
Chris@82 849 VTW(1, 11),
Chris@82 850 VTW(1, 12),
Chris@82 851 VTW(1, 13),
Chris@82 852 VTW(1, 14),
Chris@82 853 VTW(1, 15),
Chris@82 854 VTW(1, 16),
Chris@82 855 VTW(1, 17),
Chris@82 856 VTW(1, 18),
Chris@82 857 VTW(1, 19),
Chris@82 858 VTW(1, 20),
Chris@82 859 VTW(1, 21),
Chris@82 860 VTW(1, 22),
Chris@82 861 VTW(1, 23),
Chris@82 862 VTW(1, 24),
Chris@82 863 VTW(1, 25),
Chris@82 864 VTW(1, 26),
Chris@82 865 VTW(1, 27),
Chris@82 866 VTW(1, 28),
Chris@82 867 VTW(1, 29),
Chris@82 868 VTW(1, 30),
Chris@82 869 VTW(1, 31),
Chris@82 870 {TW_NEXT, VL, 0}
Chris@82 871 };
Chris@82 872
Chris@82 873 static const hc2c_desc desc = { 32, XSIMD_STRING("hc2cfdftv_32"), twinstr, &GENUS, {233, 117, 16, 0} };
Chris@82 874
Chris@82 875 void XSIMD(codelet_hc2cfdftv_32) (planner *p) {
Chris@82 876 X(khc2c_register) (p, hc2cfdftv_32, &desc, HC2C_VIA_DFT);
Chris@82 877 }
Chris@82 878 #endif