Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.8/rdft/simd/common/hc2cbdftv_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 08:08:12 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 -dif -sign 1 -name hc2cbdftv_32 -include rdft/simd/hc2cbv.h */
Chris@82 29
Chris@82 30 /*
Chris@82 31 * This function contains 249 FP additions, 192 FP multiplications,
Chris@82 32 * (or, 119 additions, 62 multiplications, 130 fused multiply/add),
Chris@82 33 * 143 stack variables, 7 constants, and 64 memory accesses
Chris@82 34 */
Chris@82 35 #include "rdft/simd/hc2cbv.h"
Chris@82 36
Chris@82 37 static void hc2cbdftv_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(KP980785280, +0.980785280403230449126182236134239036973933731);
Chris@82 40 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
Chris@82 41 DVK(KP198912367, +0.198912367379658006911597622644676228597850501);
Chris@82 42 DVK(KP668178637, +0.668178637919298919997757686523080761552472251);
Chris@82 43 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
Chris@82 44 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
Chris@82 45 DVK(KP414213562, +0.414213562373095048801688724209698078569671875);
Chris@82 46 {
Chris@82 47 INT m;
Chris@82 48 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 49 V Ts, T1S, T3p, T45, T3A, T48, T1b, T1V, T1o, T2G, T2o, T2Y, T2z, T31, T1L;
Chris@82 50 V T2H, T2J, T2K, TJ, T1c, T3D, T46, T10, T1d, T2r, T2A, T3w, T49, T1D, T1M;
Chris@82 51 V T2u, T2B;
Chris@82 52 {
Chris@82 53 V T4, T1i, T15, T1j, Tb, T1m, T16, T1l, T1G, T1F, Tj, T3m, T18, T1J, T1I;
Chris@82 54 V Tq, T3n, T19, T2, T3, T13, T14, T5, T6, T7, T8, T9, Ta, Tf, Ti;
Chris@82 55 V Td, Te, Tg, Th, Tm, Tp, Tk, Tl, Tn, To, Tc, Tr, T3l, T3o, T3y;
Chris@82 56 V T3z, T17, T1a, T1k, T1n, T2m, T2n, T2x, T2y, T1H, T1K;
Chris@82 57 T2 = LD(&(Rp[0]), ms, &(Rp[0]));
Chris@82 58 T3 = LD(&(Rm[WS(rs, 15)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 59 T4 = VFNMSCONJ(T3, T2);
Chris@82 60 T1i = VFMACONJ(T3, T2);
Chris@82 61 T13 = LD(&(Rp[WS(rs, 8)]), ms, &(Rp[0]));
Chris@82 62 T14 = LD(&(Rm[WS(rs, 7)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 63 T15 = VFNMSCONJ(T14, T13);
Chris@82 64 T1j = VFMACONJ(T14, T13);
Chris@82 65 T5 = LD(&(Rp[WS(rs, 4)]), ms, &(Rp[0]));
Chris@82 66 T6 = LD(&(Rm[WS(rs, 11)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 67 T7 = VFNMSCONJ(T6, T5);
Chris@82 68 T8 = LD(&(Rp[WS(rs, 12)]), ms, &(Rp[0]));
Chris@82 69 T9 = LD(&(Rm[WS(rs, 3)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 70 Ta = VFMSCONJ(T9, T8);
Chris@82 71 Tb = VADD(T7, Ta);
Chris@82 72 T1m = VFMACONJ(T9, T8);
Chris@82 73 T16 = VSUB(T7, Ta);
Chris@82 74 T1l = VFMACONJ(T6, T5);
Chris@82 75 Td = LD(&(Rp[WS(rs, 10)]), ms, &(Rp[0]));
Chris@82 76 Te = LD(&(Rm[WS(rs, 5)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 77 Tf = VFNMSCONJ(Te, Td);
Chris@82 78 T1G = VFMACONJ(Te, Td);
Chris@82 79 Tg = LD(&(Rp[WS(rs, 2)]), ms, &(Rp[0]));
Chris@82 80 Th = LD(&(Rm[WS(rs, 13)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 81 Ti = VFNMSCONJ(Th, Tg);
Chris@82 82 T1F = VFMACONJ(Th, Tg);
Chris@82 83 Tj = VFMA(LDK(KP414213562), Ti, Tf);
Chris@82 84 T3m = VSUB(T1F, T1G);
Chris@82 85 T18 = VFNMS(LDK(KP414213562), Tf, Ti);
Chris@82 86 Tk = LD(&(Rp[WS(rs, 6)]), ms, &(Rp[0]));
Chris@82 87 Tl = LD(&(Rm[WS(rs, 9)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 88 Tm = VFNMSCONJ(Tl, Tk);
Chris@82 89 T1J = VFMACONJ(Tl, Tk);
Chris@82 90 Tn = LD(&(Rp[WS(rs, 14)]), ms, &(Rp[0]));
Chris@82 91 To = LD(&(Rm[WS(rs, 1)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 92 Tp = VFMSCONJ(To, Tn);
Chris@82 93 T1I = VFMACONJ(To, Tn);
Chris@82 94 Tq = VFNMS(LDK(KP414213562), Tp, Tm);
Chris@82 95 T3n = VSUB(T1I, T1J);
Chris@82 96 T19 = VFMA(LDK(KP414213562), Tm, Tp);
Chris@82 97 Tc = VFNMS(LDK(KP707106781), Tb, T4);
Chris@82 98 Tr = VSUB(Tj, Tq);
Chris@82 99 Ts = VFMA(LDK(KP923879532), Tr, Tc);
Chris@82 100 T1S = VFNMS(LDK(KP923879532), Tr, Tc);
Chris@82 101 T3l = VSUB(T1i, T1j);
Chris@82 102 T3o = VADD(T3m, T3n);
Chris@82 103 T3p = VFMA(LDK(KP707106781), T3o, T3l);
Chris@82 104 T45 = VFNMS(LDK(KP707106781), T3o, T3l);
Chris@82 105 T3y = VSUB(T1l, T1m);
Chris@82 106 T3z = VSUB(T3m, T3n);
Chris@82 107 T3A = VFMA(LDK(KP707106781), T3z, T3y);
Chris@82 108 T48 = VFNMS(LDK(KP707106781), T3z, T3y);
Chris@82 109 T17 = VFNMS(LDK(KP707106781), T16, T15);
Chris@82 110 T1a = VSUB(T18, T19);
Chris@82 111 T1b = VFNMS(LDK(KP923879532), T1a, T17);
Chris@82 112 T1V = VFMA(LDK(KP923879532), T1a, T17);
Chris@82 113 T1k = VADD(T1i, T1j);
Chris@82 114 T1n = VADD(T1l, T1m);
Chris@82 115 T1o = VSUB(T1k, T1n);
Chris@82 116 T2G = VADD(T1k, T1n);
Chris@82 117 T2m = VFMA(LDK(KP707106781), Tb, T4);
Chris@82 118 T2n = VADD(T18, T19);
Chris@82 119 T2o = VFNMS(LDK(KP923879532), T2n, T2m);
Chris@82 120 T2Y = VFMA(LDK(KP923879532), T2n, T2m);
Chris@82 121 T2x = VFMA(LDK(KP707106781), T16, T15);
Chris@82 122 T2y = VADD(Tj, Tq);
Chris@82 123 T2z = VFNMS(LDK(KP923879532), T2y, T2x);
Chris@82 124 T31 = VFMA(LDK(KP923879532), T2y, T2x);
Chris@82 125 T1H = VADD(T1F, T1G);
Chris@82 126 T1K = VADD(T1I, T1J);
Chris@82 127 T1L = VSUB(T1H, T1K);
Chris@82 128 T2H = VADD(T1H, T1K);
Chris@82 129 }
Chris@82 130 {
Chris@82 131 V Tv, T3q, TG, T1r, TM, T3t, TX, T1y, TC, T3r, TH, T1u, TT, T3u, TY;
Chris@82 132 V T1B, Tt, Tu, T1p, TE, TF, T1q, TK, TL, T1w, TV, TW, T1x, Ty, T1s;
Chris@82 133 V TB, T1t, Tw, Tx, Tz, TA, TP, T1z, TS, T1A, TN, TO, TQ, TR, TD;
Chris@82 134 V TI, T3B, T3C, TU, TZ, T2p, T2q, T3s, T3v, T1v, T1C, T2s, T2t;
Chris@82 135 Tt = LD(&(Rp[WS(rs, 1)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 136 Tu = LD(&(Rm[WS(rs, 14)]), -ms, &(Rm[0]));
Chris@82 137 T1p = VFMACONJ(Tu, Tt);
Chris@82 138 TE = LD(&(Rp[WS(rs, 9)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 139 TF = LD(&(Rm[WS(rs, 6)]), -ms, &(Rm[0]));
Chris@82 140 T1q = VFMACONJ(TF, TE);
Chris@82 141 Tv = VFNMSCONJ(Tu, Tt);
Chris@82 142 T3q = VSUB(T1p, T1q);
Chris@82 143 TG = VFNMSCONJ(TF, TE);
Chris@82 144 T1r = VADD(T1p, T1q);
Chris@82 145 TK = LD(&(Rp[WS(rs, 15)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 146 TL = LD(&(Rm[0]), -ms, &(Rm[0]));
Chris@82 147 T1w = VFMACONJ(TL, TK);
Chris@82 148 TV = LD(&(Rp[WS(rs, 7)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 149 TW = LD(&(Rm[WS(rs, 8)]), -ms, &(Rm[0]));
Chris@82 150 T1x = VFMACONJ(TW, TV);
Chris@82 151 TM = VFMSCONJ(TL, TK);
Chris@82 152 T3t = VSUB(T1w, T1x);
Chris@82 153 TX = VFNMSCONJ(TW, TV);
Chris@82 154 T1y = VADD(T1w, T1x);
Chris@82 155 Tw = LD(&(Rp[WS(rs, 5)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 156 Tx = LD(&(Rm[WS(rs, 10)]), -ms, &(Rm[0]));
Chris@82 157 Ty = VFNMSCONJ(Tx, Tw);
Chris@82 158 T1s = VFMACONJ(Tx, Tw);
Chris@82 159 Tz = LD(&(Rp[WS(rs, 13)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 160 TA = LD(&(Rm[WS(rs, 2)]), -ms, &(Rm[0]));
Chris@82 161 TB = VFMSCONJ(TA, Tz);
Chris@82 162 T1t = VFMACONJ(TA, Tz);
Chris@82 163 TC = VADD(Ty, TB);
Chris@82 164 T3r = VSUB(T1s, T1t);
Chris@82 165 TH = VSUB(Ty, TB);
Chris@82 166 T1u = VADD(T1s, T1t);
Chris@82 167 TN = LD(&(Rp[WS(rs, 3)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 168 TO = LD(&(Rm[WS(rs, 12)]), -ms, &(Rm[0]));
Chris@82 169 TP = VFNMSCONJ(TO, TN);
Chris@82 170 T1z = VFMACONJ(TO, TN);
Chris@82 171 TQ = LD(&(Rp[WS(rs, 11)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 172 TR = LD(&(Rm[WS(rs, 4)]), -ms, &(Rm[0]));
Chris@82 173 TS = VFMSCONJ(TR, TQ);
Chris@82 174 T1A = VFMACONJ(TR, TQ);
Chris@82 175 TT = VADD(TP, TS);
Chris@82 176 T3u = VSUB(T1A, T1z);
Chris@82 177 TY = VSUB(TS, TP);
Chris@82 178 T1B = VADD(T1z, T1A);
Chris@82 179 T2J = VADD(T1r, T1u);
Chris@82 180 T2K = VADD(T1y, T1B);
Chris@82 181 TD = VFNMS(LDK(KP707106781), TC, Tv);
Chris@82 182 TI = VFNMS(LDK(KP707106781), TH, TG);
Chris@82 183 TJ = VFMA(LDK(KP668178637), TI, TD);
Chris@82 184 T1c = VFNMS(LDK(KP668178637), TD, TI);
Chris@82 185 T3B = VFMA(LDK(KP414213562), T3q, T3r);
Chris@82 186 T3C = VFMA(LDK(KP414213562), T3t, T3u);
Chris@82 187 T3D = VSUB(T3B, T3C);
Chris@82 188 T46 = VADD(T3B, T3C);
Chris@82 189 TU = VFNMS(LDK(KP707106781), TT, TM);
Chris@82 190 TZ = VFMA(LDK(KP707106781), TY, TX);
Chris@82 191 T10 = VFNMS(LDK(KP668178637), TZ, TU);
Chris@82 192 T1d = VFMA(LDK(KP668178637), TU, TZ);
Chris@82 193 T2p = VFMA(LDK(KP707106781), TH, TG);
Chris@82 194 T2q = VFMA(LDK(KP707106781), TC, Tv);
Chris@82 195 T2r = VFMA(LDK(KP198912367), T2q, T2p);
Chris@82 196 T2A = VFNMS(LDK(KP198912367), T2p, T2q);
Chris@82 197 T3s = VFNMS(LDK(KP414213562), T3r, T3q);
Chris@82 198 T3v = VFNMS(LDK(KP414213562), T3u, T3t);
Chris@82 199 T3w = VADD(T3s, T3v);
Chris@82 200 T49 = VSUB(T3s, T3v);
Chris@82 201 T1v = VSUB(T1r, T1u);
Chris@82 202 T1C = VSUB(T1y, T1B);
Chris@82 203 T1D = VADD(T1v, T1C);
Chris@82 204 T1M = VSUB(T1v, T1C);
Chris@82 205 T2s = VFNMS(LDK(KP707106781), TY, TX);
Chris@82 206 T2t = VFMA(LDK(KP707106781), TT, TM);
Chris@82 207 T2u = VFNMS(LDK(KP198912367), T2t, T2s);
Chris@82 208 T2B = VFMA(LDK(KP198912367), T2s, T2t);
Chris@82 209 }
Chris@82 210 {
Chris@82 211 V T3f, T38, T4p, T4v, T3T, T3Z, T2a, T2i, T4b, T4h, T1O, T20, T2M, T2U, T3F;
Chris@82 212 V T3L, T1g, T3X, T2g, T3J, T2E, T4l, T2S, T4f, T1Y, T4t, T26, T43, T34, T3P;
Chris@82 213 V T3e, T3j, T36, T37, T35, T4n, T4o, T4m, T4u, T3R, T3S, T3Q, T3Y, T28, T29;
Chris@82 214 V T27, T2h, T47, T4a, T44, T4g, T1E, T1N, T1h, T1Z;
Chris@82 215 T36 = VADD(T2G, T2H);
Chris@82 216 T37 = VADD(T2J, T2K);
Chris@82 217 T3f = VADD(T36, T37);
Chris@82 218 T35 = LDW(&(W[TWVL * 30]));
Chris@82 219 T38 = VZMUL(T35, VSUB(T36, T37));
Chris@82 220 T4n = VFMA(LDK(KP923879532), T46, T45);
Chris@82 221 T4o = VFNMS(LDK(KP923879532), T49, T48);
Chris@82 222 T4m = LDW(&(W[TWVL * 10]));
Chris@82 223 T4p = VZMUL(T4m, VFNMSI(T4o, T4n));
Chris@82 224 T4u = LDW(&(W[TWVL * 50]));
Chris@82 225 T4v = VZMUL(T4u, VFMAI(T4o, T4n));
Chris@82 226 T3R = VFMA(LDK(KP923879532), T3w, T3p);
Chris@82 227 T3S = VFMA(LDK(KP923879532), T3D, T3A);
Chris@82 228 T3Q = LDW(&(W[TWVL * 58]));
Chris@82 229 T3T = VZMUL(T3Q, VFNMSI(T3S, T3R));
Chris@82 230 T3Y = LDW(&(W[TWVL * 2]));
Chris@82 231 T3Z = VZMUL(T3Y, VFMAI(T3S, T3R));
Chris@82 232 T28 = VFMA(LDK(KP707106781), T1D, T1o);
Chris@82 233 T29 = VFMA(LDK(KP707106781), T1M, T1L);
Chris@82 234 T27 = LDW(&(W[TWVL * 6]));
Chris@82 235 T2a = VZMUL(T27, VFMAI(T29, T28));
Chris@82 236 T2h = LDW(&(W[TWVL * 54]));
Chris@82 237 T2i = VZMUL(T2h, VFNMSI(T29, T28));
Chris@82 238 T47 = VFNMS(LDK(KP923879532), T46, T45);
Chris@82 239 T4a = VFMA(LDK(KP923879532), T49, T48);
Chris@82 240 T44 = LDW(&(W[TWVL * 18]));
Chris@82 241 T4b = VZMUL(T44, VFMAI(T4a, T47));
Chris@82 242 T4g = LDW(&(W[TWVL * 42]));
Chris@82 243 T4h = VZMUL(T4g, VFNMSI(T4a, T47));
Chris@82 244 T1E = VFNMS(LDK(KP707106781), T1D, T1o);
Chris@82 245 T1N = VFNMS(LDK(KP707106781), T1M, T1L);
Chris@82 246 T1h = LDW(&(W[TWVL * 22]));
Chris@82 247 T1O = VZMUL(T1h, VFNMSI(T1N, T1E));
Chris@82 248 T1Z = LDW(&(W[TWVL * 38]));
Chris@82 249 T20 = VZMUL(T1Z, VFMAI(T1N, T1E));
Chris@82 250 {
Chris@82 251 V T2I, T2L, T2F, T2T, T3x, T3E, T3k, T3K, T12, T2e, T1f, T2f, T11, T1e, T1;
Chris@82 252 V T3W, T2d, T3I, T2w, T2Q, T2D, T2R, T2v, T2C, T2l, T4k, T2P, T4e, T1U, T24;
Chris@82 253 V T1X, T25, T1T, T1W, T1R, T4s, T23, T42, T30, T3c, T33, T3d, T2Z, T32, T2X;
Chris@82 254 V T3O, T3b, T3i;
Chris@82 255 T2I = VSUB(T2G, T2H);
Chris@82 256 T2L = VSUB(T2J, T2K);
Chris@82 257 T2F = LDW(&(W[TWVL * 46]));
Chris@82 258 T2M = VZMUL(T2F, VFNMSI(T2L, T2I));
Chris@82 259 T2T = LDW(&(W[TWVL * 14]));
Chris@82 260 T2U = VZMUL(T2T, VFMAI(T2L, T2I));
Chris@82 261 T3x = VFNMS(LDK(KP923879532), T3w, T3p);
Chris@82 262 T3E = VFNMS(LDK(KP923879532), T3D, T3A);
Chris@82 263 T3k = LDW(&(W[TWVL * 26]));
Chris@82 264 T3F = VZMUL(T3k, VFNMSI(T3E, T3x));
Chris@82 265 T3K = LDW(&(W[TWVL * 34]));
Chris@82 266 T3L = VZMUL(T3K, VFMAI(T3E, T3x));
Chris@82 267 T11 = VADD(TJ, T10);
Chris@82 268 T12 = VFNMS(LDK(KP831469612), T11, Ts);
Chris@82 269 T2e = VFMA(LDK(KP831469612), T11, Ts);
Chris@82 270 T1e = VADD(T1c, T1d);
Chris@82 271 T1f = VFNMS(LDK(KP831469612), T1e, T1b);
Chris@82 272 T2f = VFMA(LDK(KP831469612), T1e, T1b);
Chris@82 273 T1 = LDW(&(W[TWVL * 24]));
Chris@82 274 T1g = VZMULI(T1, VFMAI(T1f, T12));
Chris@82 275 T3W = LDW(&(W[TWVL * 4]));
Chris@82 276 T3X = VZMULI(T3W, VFNMSI(T2f, T2e));
Chris@82 277 T2d = LDW(&(W[TWVL * 56]));
Chris@82 278 T2g = VZMULI(T2d, VFMAI(T2f, T2e));
Chris@82 279 T3I = LDW(&(W[TWVL * 36]));
Chris@82 280 T3J = VZMULI(T3I, VFNMSI(T1f, T12));
Chris@82 281 T2v = VSUB(T2r, T2u);
Chris@82 282 T2w = VFMA(LDK(KP980785280), T2v, T2o);
Chris@82 283 T2Q = VFNMS(LDK(KP980785280), T2v, T2o);
Chris@82 284 T2C = VSUB(T2A, T2B);
Chris@82 285 T2D = VFNMS(LDK(KP980785280), T2C, T2z);
Chris@82 286 T2R = VFMA(LDK(KP980785280), T2C, T2z);
Chris@82 287 T2l = LDW(&(W[TWVL * 48]));
Chris@82 288 T2E = VZMULI(T2l, VFMAI(T2D, T2w));
Chris@82 289 T4k = LDW(&(W[TWVL * 12]));
Chris@82 290 T4l = VZMULI(T4k, VFNMSI(T2D, T2w));
Chris@82 291 T2P = LDW(&(W[TWVL * 16]));
Chris@82 292 T2S = VZMULI(T2P, VFMAI(T2R, T2Q));
Chris@82 293 T4e = LDW(&(W[TWVL * 44]));
Chris@82 294 T4f = VZMULI(T4e, VFNMSI(T2R, T2Q));
Chris@82 295 T1T = VSUB(T1d, T1c);
Chris@82 296 T1U = VFNMS(LDK(KP831469612), T1T, T1S);
Chris@82 297 T24 = VFMA(LDK(KP831469612), T1T, T1S);
Chris@82 298 T1W = VSUB(TJ, T10);
Chris@82 299 T1X = VFNMS(LDK(KP831469612), T1W, T1V);
Chris@82 300 T25 = VFMA(LDK(KP831469612), T1W, T1V);
Chris@82 301 T1R = LDW(&(W[TWVL * 40]));
Chris@82 302 T1Y = VZMULI(T1R, VFMAI(T1X, T1U));
Chris@82 303 T4s = LDW(&(W[TWVL * 52]));
Chris@82 304 T4t = VZMULI(T4s, VFNMSI(T25, T24));
Chris@82 305 T23 = LDW(&(W[TWVL * 8]));
Chris@82 306 T26 = VZMULI(T23, VFMAI(T25, T24));
Chris@82 307 T42 = LDW(&(W[TWVL * 20]));
Chris@82 308 T43 = VZMULI(T42, VFNMSI(T1X, T1U));
Chris@82 309 T2Z = VADD(T2A, T2B);
Chris@82 310 T30 = VFNMS(LDK(KP980785280), T2Z, T2Y);
Chris@82 311 T3c = VFMA(LDK(KP980785280), T2Z, T2Y);
Chris@82 312 T32 = VADD(T2r, T2u);
Chris@82 313 T33 = VFNMS(LDK(KP980785280), T32, T31);
Chris@82 314 T3d = VFMA(LDK(KP980785280), T32, T31);
Chris@82 315 T2X = LDW(&(W[TWVL * 32]));
Chris@82 316 T34 = VZMULI(T2X, VFMAI(T33, T30));
Chris@82 317 T3O = LDW(&(W[TWVL * 60]));
Chris@82 318 T3P = VZMULI(T3O, VFNMSI(T3d, T3c));
Chris@82 319 T3b = LDW(&(W[0]));
Chris@82 320 T3e = VZMULI(T3b, VFMAI(T3d, T3c));
Chris@82 321 T3i = LDW(&(W[TWVL * 28]));
Chris@82 322 T3j = VZMULI(T3i, VFNMSI(T33, T30));
Chris@82 323 }
Chris@82 324 {
Chris@82 325 V T1P, T4w, T2j, T4c, T4x, T1Q, T4d, T2k, T21, T4q, T2b, T4i, T4r, T22, T4j;
Chris@82 326 V T2c, T2N, T40, T3g, T3G, T41, T2O, T3H, T3h, T2V, T3U, T39, T3M, T3V, T2W;
Chris@82 327 V T3N, T3a;
Chris@82 328 T1P = VADD(T1g, T1O);
Chris@82 329 ST(&(Rp[WS(rs, 6)]), T1P, ms, &(Rp[0]));
Chris@82 330 T4w = VADD(T4t, T4v);
Chris@82 331 ST(&(Rp[WS(rs, 13)]), T4w, ms, &(Rp[WS(rs, 1)]));
Chris@82 332 T2j = VADD(T2g, T2i);
Chris@82 333 ST(&(Rp[WS(rs, 14)]), T2j, ms, &(Rp[0]));
Chris@82 334 T4c = VADD(T43, T4b);
Chris@82 335 ST(&(Rp[WS(rs, 5)]), T4c, ms, &(Rp[WS(rs, 1)]));
Chris@82 336 T4x = VCONJ(VSUB(T4v, T4t));
Chris@82 337 ST(&(Rm[WS(rs, 13)]), T4x, -ms, &(Rm[WS(rs, 1)]));
Chris@82 338 T1Q = VCONJ(VSUB(T1O, T1g));
Chris@82 339 ST(&(Rm[WS(rs, 6)]), T1Q, -ms, &(Rm[0]));
Chris@82 340 T4d = VCONJ(VSUB(T4b, T43));
Chris@82 341 ST(&(Rm[WS(rs, 5)]), T4d, -ms, &(Rm[WS(rs, 1)]));
Chris@82 342 T2k = VCONJ(VSUB(T2i, T2g));
Chris@82 343 ST(&(Rm[WS(rs, 14)]), T2k, -ms, &(Rm[0]));
Chris@82 344 T21 = VADD(T1Y, T20);
Chris@82 345 ST(&(Rp[WS(rs, 10)]), T21, ms, &(Rp[0]));
Chris@82 346 T4q = VADD(T4l, T4p);
Chris@82 347 ST(&(Rp[WS(rs, 3)]), T4q, ms, &(Rp[WS(rs, 1)]));
Chris@82 348 T2b = VADD(T26, T2a);
Chris@82 349 ST(&(Rp[WS(rs, 2)]), T2b, ms, &(Rp[0]));
Chris@82 350 T4i = VADD(T4f, T4h);
Chris@82 351 ST(&(Rp[WS(rs, 11)]), T4i, ms, &(Rp[WS(rs, 1)]));
Chris@82 352 T4r = VCONJ(VSUB(T4p, T4l));
Chris@82 353 ST(&(Rm[WS(rs, 3)]), T4r, -ms, &(Rm[WS(rs, 1)]));
Chris@82 354 T22 = VCONJ(VSUB(T20, T1Y));
Chris@82 355 ST(&(Rm[WS(rs, 10)]), T22, -ms, &(Rm[0]));
Chris@82 356 T4j = VCONJ(VSUB(T4h, T4f));
Chris@82 357 ST(&(Rm[WS(rs, 11)]), T4j, -ms, &(Rm[WS(rs, 1)]));
Chris@82 358 T2c = VCONJ(VSUB(T2a, T26));
Chris@82 359 ST(&(Rm[WS(rs, 2)]), T2c, -ms, &(Rm[0]));
Chris@82 360 T2N = VADD(T2E, T2M);
Chris@82 361 ST(&(Rp[WS(rs, 12)]), T2N, ms, &(Rp[0]));
Chris@82 362 T40 = VADD(T3X, T3Z);
Chris@82 363 ST(&(Rp[WS(rs, 1)]), T40, ms, &(Rp[WS(rs, 1)]));
Chris@82 364 T3g = VADD(T3e, T3f);
Chris@82 365 ST(&(Rp[0]), T3g, ms, &(Rp[0]));
Chris@82 366 T3G = VADD(T3j, T3F);
Chris@82 367 ST(&(Rp[WS(rs, 7)]), T3G, ms, &(Rp[WS(rs, 1)]));
Chris@82 368 T41 = VCONJ(VSUB(T3Z, T3X));
Chris@82 369 ST(&(Rm[WS(rs, 1)]), T41, -ms, &(Rm[WS(rs, 1)]));
Chris@82 370 T2O = VCONJ(VSUB(T2M, T2E));
Chris@82 371 ST(&(Rm[WS(rs, 12)]), T2O, -ms, &(Rm[0]));
Chris@82 372 T3H = VCONJ(VSUB(T3F, T3j));
Chris@82 373 ST(&(Rm[WS(rs, 7)]), T3H, -ms, &(Rm[WS(rs, 1)]));
Chris@82 374 T3h = VCONJ(VSUB(T3f, T3e));
Chris@82 375 ST(&(Rm[0]), T3h, -ms, &(Rm[0]));
Chris@82 376 T2V = VADD(T2S, T2U);
Chris@82 377 ST(&(Rp[WS(rs, 4)]), T2V, ms, &(Rp[0]));
Chris@82 378 T3U = VADD(T3P, T3T);
Chris@82 379 ST(&(Rp[WS(rs, 15)]), T3U, ms, &(Rp[WS(rs, 1)]));
Chris@82 380 T39 = VADD(T34, T38);
Chris@82 381 ST(&(Rp[WS(rs, 8)]), T39, ms, &(Rp[0]));
Chris@82 382 T3M = VADD(T3J, T3L);
Chris@82 383 ST(&(Rp[WS(rs, 9)]), T3M, ms, &(Rp[WS(rs, 1)]));
Chris@82 384 T3V = VCONJ(VSUB(T3T, T3P));
Chris@82 385 ST(&(Rm[WS(rs, 15)]), T3V, -ms, &(Rm[WS(rs, 1)]));
Chris@82 386 T2W = VCONJ(VSUB(T2U, T2S));
Chris@82 387 ST(&(Rm[WS(rs, 4)]), T2W, -ms, &(Rm[0]));
Chris@82 388 T3N = VCONJ(VSUB(T3L, T3J));
Chris@82 389 ST(&(Rm[WS(rs, 9)]), T3N, -ms, &(Rm[WS(rs, 1)]));
Chris@82 390 T3a = VCONJ(VSUB(T38, T34));
Chris@82 391 ST(&(Rm[WS(rs, 8)]), T3a, -ms, &(Rm[0]));
Chris@82 392 }
Chris@82 393 }
Chris@82 394 }
Chris@82 395 }
Chris@82 396 VLEAVE();
Chris@82 397 }
Chris@82 398
Chris@82 399 static const tw_instr twinstr[] = {
Chris@82 400 VTW(1, 1),
Chris@82 401 VTW(1, 2),
Chris@82 402 VTW(1, 3),
Chris@82 403 VTW(1, 4),
Chris@82 404 VTW(1, 5),
Chris@82 405 VTW(1, 6),
Chris@82 406 VTW(1, 7),
Chris@82 407 VTW(1, 8),
Chris@82 408 VTW(1, 9),
Chris@82 409 VTW(1, 10),
Chris@82 410 VTW(1, 11),
Chris@82 411 VTW(1, 12),
Chris@82 412 VTW(1, 13),
Chris@82 413 VTW(1, 14),
Chris@82 414 VTW(1, 15),
Chris@82 415 VTW(1, 16),
Chris@82 416 VTW(1, 17),
Chris@82 417 VTW(1, 18),
Chris@82 418 VTW(1, 19),
Chris@82 419 VTW(1, 20),
Chris@82 420 VTW(1, 21),
Chris@82 421 VTW(1, 22),
Chris@82 422 VTW(1, 23),
Chris@82 423 VTW(1, 24),
Chris@82 424 VTW(1, 25),
Chris@82 425 VTW(1, 26),
Chris@82 426 VTW(1, 27),
Chris@82 427 VTW(1, 28),
Chris@82 428 VTW(1, 29),
Chris@82 429 VTW(1, 30),
Chris@82 430 VTW(1, 31),
Chris@82 431 {TW_NEXT, VL, 0}
Chris@82 432 };
Chris@82 433
Chris@82 434 static const hc2c_desc desc = { 32, XSIMD_STRING("hc2cbdftv_32"), twinstr, &GENUS, {119, 62, 130, 0} };
Chris@82 435
Chris@82 436 void XSIMD(codelet_hc2cbdftv_32) (planner *p) {
Chris@82 437 X(khc2c_register) (p, hc2cbdftv_32, &desc, HC2C_VIA_DFT);
Chris@82 438 }
Chris@82 439 #else
Chris@82 440
Chris@82 441 /* Generated by: ../../../genfft/gen_hc2cdft_c.native -simd -compact -variables 4 -pipeline-latency 8 -trivial-stores -variables 32 -no-generate-bytw -n 32 -dif -sign 1 -name hc2cbdftv_32 -include rdft/simd/hc2cbv.h */
Chris@82 442
Chris@82 443 /*
Chris@82 444 * This function contains 249 FP additions, 104 FP multiplications,
Chris@82 445 * (or, 233 additions, 88 multiplications, 16 fused multiply/add),
Chris@82 446 * 161 stack variables, 7 constants, and 64 memory accesses
Chris@82 447 */
Chris@82 448 #include "rdft/simd/hc2cbv.h"
Chris@82 449
Chris@82 450 static void hc2cbdftv_32(R *Rp, R *Ip, R *Rm, R *Im, const R *W, stride rs, INT mb, INT me, INT ms)
Chris@82 451 {
Chris@82 452 DVK(KP195090322, +0.195090322016128267848284868477022240927691618);
Chris@82 453 DVK(KP980785280, +0.980785280403230449126182236134239036973933731);
Chris@82 454 DVK(KP555570233, +0.555570233019602224742830813948532874374937191);
Chris@82 455 DVK(KP831469612, +0.831469612302545237078788377617905756738560812);
Chris@82 456 DVK(KP923879532, +0.923879532511286756128183189396788286822416626);
Chris@82 457 DVK(KP382683432, +0.382683432365089771728459984030398866761344562);
Chris@82 458 DVK(KP707106781, +0.707106781186547524400844362104849039284835938);
Chris@82 459 {
Chris@82 460 INT m;
Chris@82 461 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@82 462 V T1W, T21, Tf, T2c, T1t, T2r, T3T, T4m, Ty, T2q, T3P, T4n, T1n, T2d, T1T;
Chris@82 463 V T22, T1E, T24, T3I, T4p, TU, T2n, T1i, T2h, T1L, T25, T3L, T4q, T1f, T2o;
Chris@82 464 V T1j, T2k;
Chris@82 465 {
Chris@82 466 V T2, T4, T1Z, T1p, T1r, T20, T9, T1U, Td, T1V, T3, T1q, T6, T8, T7;
Chris@82 467 V Tc, Tb, Ta, T5, Te, T1o, T1s, T3R, T3S, Tj, T1N, Tw, T1Q, Tn, T1O;
Chris@82 468 V Ts, T1R, Tg, Ti, Th, Tv, Tu, Tt, Tk, Tm, Tl, Tp, Tr, Tq, To;
Chris@82 469 V Tx, T3N, T3O, T1l, T1m, T1P, T1S;
Chris@82 470 T2 = LD(&(Rp[0]), ms, &(Rp[0]));
Chris@82 471 T3 = LD(&(Rm[WS(rs, 15)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 472 T4 = VCONJ(T3);
Chris@82 473 T1Z = VADD(T2, T4);
Chris@82 474 T1p = LD(&(Rp[WS(rs, 8)]), ms, &(Rp[0]));
Chris@82 475 T1q = LD(&(Rm[WS(rs, 7)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 476 T1r = VCONJ(T1q);
Chris@82 477 T20 = VADD(T1p, T1r);
Chris@82 478 T6 = LD(&(Rp[WS(rs, 4)]), ms, &(Rp[0]));
Chris@82 479 T7 = LD(&(Rm[WS(rs, 11)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 480 T8 = VCONJ(T7);
Chris@82 481 T9 = VSUB(T6, T8);
Chris@82 482 T1U = VADD(T6, T8);
Chris@82 483 Tc = LD(&(Rp[WS(rs, 12)]), ms, &(Rp[0]));
Chris@82 484 Ta = LD(&(Rm[WS(rs, 3)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 485 Tb = VCONJ(Ta);
Chris@82 486 Td = VSUB(Tb, Tc);
Chris@82 487 T1V = VADD(Tb, Tc);
Chris@82 488 T1W = VSUB(T1U, T1V);
Chris@82 489 T21 = VSUB(T1Z, T20);
Chris@82 490 T5 = VSUB(T2, T4);
Chris@82 491 Te = VMUL(LDK(KP707106781), VADD(T9, Td));
Chris@82 492 Tf = VSUB(T5, Te);
Chris@82 493 T2c = VADD(T5, Te);
Chris@82 494 T1o = VMUL(LDK(KP707106781), VSUB(T9, Td));
Chris@82 495 T1s = VSUB(T1p, T1r);
Chris@82 496 T1t = VSUB(T1o, T1s);
Chris@82 497 T2r = VADD(T1s, T1o);
Chris@82 498 T3R = VADD(T1Z, T20);
Chris@82 499 T3S = VADD(T1U, T1V);
Chris@82 500 T3T = VSUB(T3R, T3S);
Chris@82 501 T4m = VADD(T3R, T3S);
Chris@82 502 Tg = LD(&(Rp[WS(rs, 2)]), ms, &(Rp[0]));
Chris@82 503 Th = LD(&(Rm[WS(rs, 13)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 504 Ti = VCONJ(Th);
Chris@82 505 Tj = VSUB(Tg, Ti);
Chris@82 506 T1N = VADD(Tg, Ti);
Chris@82 507 Tv = LD(&(Rp[WS(rs, 14)]), ms, &(Rp[0]));
Chris@82 508 Tt = LD(&(Rm[WS(rs, 1)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 509 Tu = VCONJ(Tt);
Chris@82 510 Tw = VSUB(Tu, Tv);
Chris@82 511 T1Q = VADD(Tu, Tv);
Chris@82 512 Tk = LD(&(Rp[WS(rs, 10)]), ms, &(Rp[0]));
Chris@82 513 Tl = LD(&(Rm[WS(rs, 5)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 514 Tm = VCONJ(Tl);
Chris@82 515 Tn = VSUB(Tk, Tm);
Chris@82 516 T1O = VADD(Tk, Tm);
Chris@82 517 Tp = LD(&(Rp[WS(rs, 6)]), ms, &(Rp[0]));
Chris@82 518 Tq = LD(&(Rm[WS(rs, 9)]), -ms, &(Rm[WS(rs, 1)]));
Chris@82 519 Tr = VCONJ(Tq);
Chris@82 520 Ts = VSUB(Tp, Tr);
Chris@82 521 T1R = VADD(Tp, Tr);
Chris@82 522 To = VFMA(LDK(KP382683432), Tj, VMUL(LDK(KP923879532), Tn));
Chris@82 523 Tx = VFNMS(LDK(KP382683432), Tw, VMUL(LDK(KP923879532), Ts));
Chris@82 524 Ty = VSUB(To, Tx);
Chris@82 525 T2q = VADD(To, Tx);
Chris@82 526 T3N = VADD(T1N, T1O);
Chris@82 527 T3O = VADD(T1Q, T1R);
Chris@82 528 T3P = VSUB(T3N, T3O);
Chris@82 529 T4n = VADD(T3N, T3O);
Chris@82 530 T1l = VFNMS(LDK(KP382683432), Tn, VMUL(LDK(KP923879532), Tj));
Chris@82 531 T1m = VFMA(LDK(KP923879532), Tw, VMUL(LDK(KP382683432), Ts));
Chris@82 532 T1n = VSUB(T1l, T1m);
Chris@82 533 T2d = VADD(T1l, T1m);
Chris@82 534 T1P = VSUB(T1N, T1O);
Chris@82 535 T1S = VSUB(T1Q, T1R);
Chris@82 536 T1T = VMUL(LDK(KP707106781), VSUB(T1P, T1S));
Chris@82 537 T22 = VMUL(LDK(KP707106781), VADD(T1P, T1S));
Chris@82 538 }
Chris@82 539 {
Chris@82 540 V TD, T1B, TR, T1y, TH, T1C, TM, T1z, TA, TC, TB, TO, TQ, TP, TG;
Chris@82 541 V TF, TE, TJ, TL, TK, T1A, T1D, T3G, T3H, TN, T2f, TT, T2g, TI, TS;
Chris@82 542 V TY, T1I, T1c, T1F, T12, T1J, T17, T1G, TV, TX, TW, T1b, T1a, T19, T11;
Chris@82 543 V T10, TZ, T14, T16, T15, T1H, T1K, T3J, T3K, T18, T2i, T1e, T2j, T13, T1d;
Chris@82 544 TA = LD(&(Rp[WS(rs, 5)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 545 TB = LD(&(Rm[WS(rs, 10)]), -ms, &(Rm[0]));
Chris@82 546 TC = VCONJ(TB);
Chris@82 547 TD = VSUB(TA, TC);
Chris@82 548 T1B = VADD(TA, TC);
Chris@82 549 TO = LD(&(Rp[WS(rs, 1)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 550 TP = LD(&(Rm[WS(rs, 14)]), -ms, &(Rm[0]));
Chris@82 551 TQ = VCONJ(TP);
Chris@82 552 TR = VSUB(TO, TQ);
Chris@82 553 T1y = VADD(TO, TQ);
Chris@82 554 TG = LD(&(Rp[WS(rs, 13)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 555 TE = LD(&(Rm[WS(rs, 2)]), -ms, &(Rm[0]));
Chris@82 556 TF = VCONJ(TE);
Chris@82 557 TH = VSUB(TF, TG);
Chris@82 558 T1C = VADD(TF, TG);
Chris@82 559 TJ = LD(&(Rp[WS(rs, 9)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 560 TK = LD(&(Rm[WS(rs, 6)]), -ms, &(Rm[0]));
Chris@82 561 TL = VCONJ(TK);
Chris@82 562 TM = VSUB(TJ, TL);
Chris@82 563 T1z = VADD(TJ, TL);
Chris@82 564 T1A = VSUB(T1y, T1z);
Chris@82 565 T1D = VSUB(T1B, T1C);
Chris@82 566 T1E = VFNMS(LDK(KP382683432), T1D, VMUL(LDK(KP923879532), T1A));
Chris@82 567 T24 = VFMA(LDK(KP382683432), T1A, VMUL(LDK(KP923879532), T1D));
Chris@82 568 T3G = VADD(T1y, T1z);
Chris@82 569 T3H = VADD(T1B, T1C);
Chris@82 570 T3I = VSUB(T3G, T3H);
Chris@82 571 T4p = VADD(T3G, T3H);
Chris@82 572 TI = VMUL(LDK(KP707106781), VSUB(TD, TH));
Chris@82 573 TN = VSUB(TI, TM);
Chris@82 574 T2f = VADD(TM, TI);
Chris@82 575 TS = VMUL(LDK(KP707106781), VADD(TD, TH));
Chris@82 576 TT = VSUB(TR, TS);
Chris@82 577 T2g = VADD(TR, TS);
Chris@82 578 TU = VFMA(LDK(KP831469612), TN, VMUL(LDK(KP555570233), TT));
Chris@82 579 T2n = VFNMS(LDK(KP195090322), T2f, VMUL(LDK(KP980785280), T2g));
Chris@82 580 T1i = VFNMS(LDK(KP555570233), TN, VMUL(LDK(KP831469612), TT));
Chris@82 581 T2h = VFMA(LDK(KP980785280), T2f, VMUL(LDK(KP195090322), T2g));
Chris@82 582 TV = LD(&(Rp[WS(rs, 3)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 583 TW = LD(&(Rm[WS(rs, 12)]), -ms, &(Rm[0]));
Chris@82 584 TX = VCONJ(TW);
Chris@82 585 TY = VSUB(TV, TX);
Chris@82 586 T1I = VADD(TV, TX);
Chris@82 587 T1b = LD(&(Rp[WS(rs, 15)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 588 T19 = LD(&(Rm[0]), -ms, &(Rm[0]));
Chris@82 589 T1a = VCONJ(T19);
Chris@82 590 T1c = VSUB(T1a, T1b);
Chris@82 591 T1F = VADD(T1a, T1b);
Chris@82 592 T11 = LD(&(Rp[WS(rs, 11)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 593 TZ = LD(&(Rm[WS(rs, 4)]), -ms, &(Rm[0]));
Chris@82 594 T10 = VCONJ(TZ);
Chris@82 595 T12 = VSUB(T10, T11);
Chris@82 596 T1J = VADD(T10, T11);
Chris@82 597 T14 = LD(&(Rp[WS(rs, 7)]), ms, &(Rp[WS(rs, 1)]));
Chris@82 598 T15 = LD(&(Rm[WS(rs, 8)]), -ms, &(Rm[0]));
Chris@82 599 T16 = VCONJ(T15);
Chris@82 600 T17 = VSUB(T14, T16);
Chris@82 601 T1G = VADD(T14, T16);
Chris@82 602 T1H = VSUB(T1F, T1G);
Chris@82 603 T1K = VSUB(T1I, T1J);
Chris@82 604 T1L = VFMA(LDK(KP923879532), T1H, VMUL(LDK(KP382683432), T1K));
Chris@82 605 T25 = VFNMS(LDK(KP382683432), T1H, VMUL(LDK(KP923879532), T1K));
Chris@82 606 T3J = VADD(T1F, T1G);
Chris@82 607 T3K = VADD(T1I, T1J);
Chris@82 608 T3L = VSUB(T3J, T3K);
Chris@82 609 T4q = VADD(T3J, T3K);
Chris@82 610 T13 = VMUL(LDK(KP707106781), VSUB(TY, T12));
Chris@82 611 T18 = VSUB(T13, T17);
Chris@82 612 T2i = VADD(T17, T13);
Chris@82 613 T1d = VMUL(LDK(KP707106781), VADD(TY, T12));
Chris@82 614 T1e = VSUB(T1c, T1d);
Chris@82 615 T2j = VADD(T1c, T1d);
Chris@82 616 T1f = VFNMS(LDK(KP555570233), T1e, VMUL(LDK(KP831469612), T18));
Chris@82 617 T2o = VFMA(LDK(KP195090322), T2i, VMUL(LDK(KP980785280), T2j));
Chris@82 618 T1j = VFMA(LDK(KP555570233), T18, VMUL(LDK(KP831469612), T1e));
Chris@82 619 T2k = VFNMS(LDK(KP195090322), T2j, VMUL(LDK(KP980785280), T2i));
Chris@82 620 }
Chris@82 621 {
Chris@82 622 V T4L, T4G, T4s, T4y, T3W, T4g, T42, T4a, T3g, T4e, T3o, T3E, T1w, T46, T2M;
Chris@82 623 V T40, T2u, T4w, T2C, T4k, T36, T3A, T3i, T3s, T28, T2O, T2w, T2G, T2Y, T4K;
Chris@82 624 V T3y, T4C;
Chris@82 625 {
Chris@82 626 V T4E, T4F, T4D, T4o, T4r, T4l, T4x, T3Q, T48, T3V, T49, T3M, T3U, T3F, T4f;
Chris@82 627 V T41, T47, T3c, T3n, T3f, T3m, T3a, T3b, T3d, T3e, T39, T4d, T3l, T3D, T1h;
Chris@82 628 V T2K, T1v, T2L, Tz, T1g, T1k, T1u, T1, T45, T2J, T3Z, T2m, T2A, T2t, T2B;
Chris@82 629 V T2e, T2l, T2p, T2s, T2b, T4v, T2z, T4j;
Chris@82 630 T4E = VADD(T4m, T4n);
Chris@82 631 T4F = VADD(T4p, T4q);
Chris@82 632 T4L = VADD(T4E, T4F);
Chris@82 633 T4D = LDW(&(W[TWVL * 30]));
Chris@82 634 T4G = VZMUL(T4D, VSUB(T4E, T4F));
Chris@82 635 T4o = VSUB(T4m, T4n);
Chris@82 636 T4r = VBYI(VSUB(T4p, T4q));
Chris@82 637 T4l = LDW(&(W[TWVL * 46]));
Chris@82 638 T4s = VZMUL(T4l, VSUB(T4o, T4r));
Chris@82 639 T4x = LDW(&(W[TWVL * 14]));
Chris@82 640 T4y = VZMUL(T4x, VADD(T4o, T4r));
Chris@82 641 T3M = VMUL(LDK(KP707106781), VSUB(T3I, T3L));
Chris@82 642 T3Q = VBYI(VSUB(T3M, T3P));
Chris@82 643 T48 = VBYI(VADD(T3P, T3M));
Chris@82 644 T3U = VMUL(LDK(KP707106781), VADD(T3I, T3L));
Chris@82 645 T3V = VSUB(T3T, T3U);
Chris@82 646 T49 = VADD(T3T, T3U);
Chris@82 647 T3F = LDW(&(W[TWVL * 22]));
Chris@82 648 T3W = VZMUL(T3F, VADD(T3Q, T3V));
Chris@82 649 T4f = LDW(&(W[TWVL * 54]));
Chris@82 650 T4g = VZMUL(T4f, VSUB(T49, T48));
Chris@82 651 T41 = LDW(&(W[TWVL * 38]));
Chris@82 652 T42 = VZMUL(T41, VSUB(T3V, T3Q));
Chris@82 653 T47 = LDW(&(W[TWVL * 6]));
Chris@82 654 T4a = VZMUL(T47, VADD(T48, T49));
Chris@82 655 T3a = VADD(T1t, T1n);
Chris@82 656 T3b = VADD(TU, T1f);
Chris@82 657 T3c = VBYI(VADD(T3a, T3b));
Chris@82 658 T3n = VBYI(VSUB(T3b, T3a));
Chris@82 659 T3d = VADD(Tf, Ty);
Chris@82 660 T3e = VADD(T1i, T1j);
Chris@82 661 T3f = VADD(T3d, T3e);
Chris@82 662 T3m = VSUB(T3d, T3e);
Chris@82 663 T39 = LDW(&(W[TWVL * 4]));
Chris@82 664 T3g = VZMULI(T39, VADD(T3c, T3f));
Chris@82 665 T4d = LDW(&(W[TWVL * 56]));
Chris@82 666 T4e = VZMULI(T4d, VSUB(T3f, T3c));
Chris@82 667 T3l = LDW(&(W[TWVL * 36]));
Chris@82 668 T3o = VZMULI(T3l, VSUB(T3m, T3n));
Chris@82 669 T3D = LDW(&(W[TWVL * 24]));
Chris@82 670 T3E = VZMULI(T3D, VADD(T3n, T3m));
Chris@82 671 Tz = VSUB(Tf, Ty);
Chris@82 672 T1g = VSUB(TU, T1f);
Chris@82 673 T1h = VSUB(Tz, T1g);
Chris@82 674 T2K = VADD(Tz, T1g);
Chris@82 675 T1k = VSUB(T1i, T1j);
Chris@82 676 T1u = VSUB(T1n, T1t);
Chris@82 677 T1v = VBYI(VSUB(T1k, T1u));
Chris@82 678 T2L = VBYI(VADD(T1u, T1k));
Chris@82 679 T1 = LDW(&(W[TWVL * 20]));
Chris@82 680 T1w = VZMULI(T1, VADD(T1h, T1v));
Chris@82 681 T45 = LDW(&(W[TWVL * 8]));
Chris@82 682 T46 = VZMULI(T45, VADD(T2K, T2L));
Chris@82 683 T2J = LDW(&(W[TWVL * 52]));
Chris@82 684 T2M = VZMULI(T2J, VSUB(T2K, T2L));
Chris@82 685 T3Z = LDW(&(W[TWVL * 40]));
Chris@82 686 T40 = VZMULI(T3Z, VSUB(T1h, T1v));
Chris@82 687 T2e = VSUB(T2c, T2d);
Chris@82 688 T2l = VSUB(T2h, T2k);
Chris@82 689 T2m = VSUB(T2e, T2l);
Chris@82 690 T2A = VADD(T2e, T2l);
Chris@82 691 T2p = VSUB(T2n, T2o);
Chris@82 692 T2s = VSUB(T2q, T2r);
Chris@82 693 T2t = VBYI(VSUB(T2p, T2s));
Chris@82 694 T2B = VBYI(VADD(T2s, T2p));
Chris@82 695 T2b = LDW(&(W[TWVL * 44]));
Chris@82 696 T2u = VZMULI(T2b, VSUB(T2m, T2t));
Chris@82 697 T4v = LDW(&(W[TWVL * 16]));
Chris@82 698 T4w = VZMULI(T4v, VADD(T2m, T2t));
Chris@82 699 T2z = LDW(&(W[TWVL * 12]));
Chris@82 700 T2C = VZMULI(T2z, VADD(T2A, T2B));
Chris@82 701 T4j = LDW(&(W[TWVL * 48]));
Chris@82 702 T4k = VZMULI(T4j, VSUB(T2A, T2B));
Chris@82 703 {
Chris@82 704 V T32, T3q, T35, T3r, T30, T31, T33, T34, T2Z, T3z, T3h, T3p, T1Y, T2E, T27;
Chris@82 705 V T2F, T1M, T1X, T23, T26, T1x, T2N, T2v, T2D, T2U, T3x, T2X, T3w, T2S, T2T;
Chris@82 706 V T2V, T2W, T2R, T4J, T3v, T4B;
Chris@82 707 T30 = VADD(T21, T22);
Chris@82 708 T31 = VADD(T1E, T1L);
Chris@82 709 T32 = VADD(T30, T31);
Chris@82 710 T3q = VSUB(T30, T31);
Chris@82 711 T33 = VADD(T1W, T1T);
Chris@82 712 T34 = VADD(T24, T25);
Chris@82 713 T35 = VBYI(VADD(T33, T34));
Chris@82 714 T3r = VBYI(VSUB(T34, T33));
Chris@82 715 T2Z = LDW(&(W[TWVL * 58]));
Chris@82 716 T36 = VZMUL(T2Z, VSUB(T32, T35));
Chris@82 717 T3z = LDW(&(W[TWVL * 26]));
Chris@82 718 T3A = VZMUL(T3z, VADD(T3q, T3r));
Chris@82 719 T3h = LDW(&(W[TWVL * 2]));
Chris@82 720 T3i = VZMUL(T3h, VADD(T32, T35));
Chris@82 721 T3p = LDW(&(W[TWVL * 34]));
Chris@82 722 T3s = VZMUL(T3p, VSUB(T3q, T3r));
Chris@82 723 T1M = VSUB(T1E, T1L);
Chris@82 724 T1X = VSUB(T1T, T1W);
Chris@82 725 T1Y = VBYI(VSUB(T1M, T1X));
Chris@82 726 T2E = VBYI(VADD(T1X, T1M));
Chris@82 727 T23 = VSUB(T21, T22);
Chris@82 728 T26 = VSUB(T24, T25);
Chris@82 729 T27 = VSUB(T23, T26);
Chris@82 730 T2F = VADD(T23, T26);
Chris@82 731 T1x = LDW(&(W[TWVL * 18]));
Chris@82 732 T28 = VZMUL(T1x, VADD(T1Y, T27));
Chris@82 733 T2N = LDW(&(W[TWVL * 50]));
Chris@82 734 T2O = VZMUL(T2N, VSUB(T2F, T2E));
Chris@82 735 T2v = LDW(&(W[TWVL * 42]));
Chris@82 736 T2w = VZMUL(T2v, VSUB(T27, T1Y));
Chris@82 737 T2D = LDW(&(W[TWVL * 10]));
Chris@82 738 T2G = VZMUL(T2D, VADD(T2E, T2F));
Chris@82 739 T2S = VADD(T2c, T2d);
Chris@82 740 T2T = VADD(T2n, T2o);
Chris@82 741 T2U = VADD(T2S, T2T);
Chris@82 742 T3x = VSUB(T2S, T2T);
Chris@82 743 T2V = VADD(T2r, T2q);
Chris@82 744 T2W = VADD(T2h, T2k);
Chris@82 745 T2X = VBYI(VADD(T2V, T2W));
Chris@82 746 T3w = VBYI(VSUB(T2W, T2V));
Chris@82 747 T2R = LDW(&(W[TWVL * 60]));
Chris@82 748 T2Y = VZMULI(T2R, VSUB(T2U, T2X));
Chris@82 749 T4J = LDW(&(W[0]));
Chris@82 750 T4K = VZMULI(T4J, VADD(T2X, T2U));
Chris@82 751 T3v = LDW(&(W[TWVL * 28]));
Chris@82 752 T3y = VZMULI(T3v, VADD(T3w, T3x));
Chris@82 753 T4B = LDW(&(W[TWVL * 32]));
Chris@82 754 T4C = VZMULI(T4B, VSUB(T3x, T3w));
Chris@82 755 }
Chris@82 756 }
Chris@82 757 {
Chris@82 758 V T29, T4M, T2P, T4t, T4N, T2a, T4u, T2Q, T2x, T4H, T2H, T4z, T4I, T2y, T4A;
Chris@82 759 V T2I, T37, T4h, T3B, T3X, T4i, T38, T3Y, T3C, T3j, T4b, T3t, T43, T4c, T3k;
Chris@82 760 V T44, T3u;
Chris@82 761 T29 = VADD(T1w, T28);
Chris@82 762 ST(&(Rp[WS(rs, 5)]), T29, ms, &(Rp[WS(rs, 1)]));
Chris@82 763 T4M = VADD(T4K, T4L);
Chris@82 764 ST(&(Rp[0]), T4M, ms, &(Rp[0]));
Chris@82 765 T2P = VADD(T2M, T2O);
Chris@82 766 ST(&(Rp[WS(rs, 13)]), T2P, ms, &(Rp[WS(rs, 1)]));
Chris@82 767 T4t = VADD(T4k, T4s);
Chris@82 768 ST(&(Rp[WS(rs, 12)]), T4t, ms, &(Rp[0]));
Chris@82 769 T4N = VCONJ(VSUB(T4L, T4K));
Chris@82 770 ST(&(Rm[0]), T4N, -ms, &(Rm[0]));
Chris@82 771 T2a = VCONJ(VSUB(T28, T1w));
Chris@82 772 ST(&(Rm[WS(rs, 5)]), T2a, -ms, &(Rm[WS(rs, 1)]));
Chris@82 773 T4u = VCONJ(VSUB(T4s, T4k));
Chris@82 774 ST(&(Rm[WS(rs, 12)]), T4u, -ms, &(Rm[0]));
Chris@82 775 T2Q = VCONJ(VSUB(T2O, T2M));
Chris@82 776 ST(&(Rm[WS(rs, 13)]), T2Q, -ms, &(Rm[WS(rs, 1)]));
Chris@82 777 T2x = VADD(T2u, T2w);
Chris@82 778 ST(&(Rp[WS(rs, 11)]), T2x, ms, &(Rp[WS(rs, 1)]));
Chris@82 779 T4H = VADD(T4C, T4G);
Chris@82 780 ST(&(Rp[WS(rs, 8)]), T4H, ms, &(Rp[0]));
Chris@82 781 T2H = VADD(T2C, T2G);
Chris@82 782 ST(&(Rp[WS(rs, 3)]), T2H, ms, &(Rp[WS(rs, 1)]));
Chris@82 783 T4z = VADD(T4w, T4y);
Chris@82 784 ST(&(Rp[WS(rs, 4)]), T4z, ms, &(Rp[0]));
Chris@82 785 T4I = VCONJ(VSUB(T4G, T4C));
Chris@82 786 ST(&(Rm[WS(rs, 8)]), T4I, -ms, &(Rm[0]));
Chris@82 787 T2y = VCONJ(VSUB(T2w, T2u));
Chris@82 788 ST(&(Rm[WS(rs, 11)]), T2y, -ms, &(Rm[WS(rs, 1)]));
Chris@82 789 T4A = VCONJ(VSUB(T4y, T4w));
Chris@82 790 ST(&(Rm[WS(rs, 4)]), T4A, -ms, &(Rm[0]));
Chris@82 791 T2I = VCONJ(VSUB(T2G, T2C));
Chris@82 792 ST(&(Rm[WS(rs, 3)]), T2I, -ms, &(Rm[WS(rs, 1)]));
Chris@82 793 T37 = VADD(T2Y, T36);
Chris@82 794 ST(&(Rp[WS(rs, 15)]), T37, ms, &(Rp[WS(rs, 1)]));
Chris@82 795 T4h = VADD(T4e, T4g);
Chris@82 796 ST(&(Rp[WS(rs, 14)]), T4h, ms, &(Rp[0]));
Chris@82 797 T3B = VADD(T3y, T3A);
Chris@82 798 ST(&(Rp[WS(rs, 7)]), T3B, ms, &(Rp[WS(rs, 1)]));
Chris@82 799 T3X = VADD(T3E, T3W);
Chris@82 800 ST(&(Rp[WS(rs, 6)]), T3X, ms, &(Rp[0]));
Chris@82 801 T4i = VCONJ(VSUB(T4g, T4e));
Chris@82 802 ST(&(Rm[WS(rs, 14)]), T4i, -ms, &(Rm[0]));
Chris@82 803 T38 = VCONJ(VSUB(T36, T2Y));
Chris@82 804 ST(&(Rm[WS(rs, 15)]), T38, -ms, &(Rm[WS(rs, 1)]));
Chris@82 805 T3Y = VCONJ(VSUB(T3W, T3E));
Chris@82 806 ST(&(Rm[WS(rs, 6)]), T3Y, -ms, &(Rm[0]));
Chris@82 807 T3C = VCONJ(VSUB(T3A, T3y));
Chris@82 808 ST(&(Rm[WS(rs, 7)]), T3C, -ms, &(Rm[WS(rs, 1)]));
Chris@82 809 T3j = VADD(T3g, T3i);
Chris@82 810 ST(&(Rp[WS(rs, 1)]), T3j, ms, &(Rp[WS(rs, 1)]));
Chris@82 811 T4b = VADD(T46, T4a);
Chris@82 812 ST(&(Rp[WS(rs, 2)]), T4b, ms, &(Rp[0]));
Chris@82 813 T3t = VADD(T3o, T3s);
Chris@82 814 ST(&(Rp[WS(rs, 9)]), T3t, ms, &(Rp[WS(rs, 1)]));
Chris@82 815 T43 = VADD(T40, T42);
Chris@82 816 ST(&(Rp[WS(rs, 10)]), T43, ms, &(Rp[0]));
Chris@82 817 T4c = VCONJ(VSUB(T4a, T46));
Chris@82 818 ST(&(Rm[WS(rs, 2)]), T4c, -ms, &(Rm[0]));
Chris@82 819 T3k = VCONJ(VSUB(T3i, T3g));
Chris@82 820 ST(&(Rm[WS(rs, 1)]), T3k, -ms, &(Rm[WS(rs, 1)]));
Chris@82 821 T44 = VCONJ(VSUB(T42, T40));
Chris@82 822 ST(&(Rm[WS(rs, 10)]), T44, -ms, &(Rm[0]));
Chris@82 823 T3u = VCONJ(VSUB(T3s, T3o));
Chris@82 824 ST(&(Rm[WS(rs, 9)]), T3u, -ms, &(Rm[WS(rs, 1)]));
Chris@82 825 }
Chris@82 826 }
Chris@82 827 }
Chris@82 828 }
Chris@82 829 VLEAVE();
Chris@82 830 }
Chris@82 831
Chris@82 832 static const tw_instr twinstr[] = {
Chris@82 833 VTW(1, 1),
Chris@82 834 VTW(1, 2),
Chris@82 835 VTW(1, 3),
Chris@82 836 VTW(1, 4),
Chris@82 837 VTW(1, 5),
Chris@82 838 VTW(1, 6),
Chris@82 839 VTW(1, 7),
Chris@82 840 VTW(1, 8),
Chris@82 841 VTW(1, 9),
Chris@82 842 VTW(1, 10),
Chris@82 843 VTW(1, 11),
Chris@82 844 VTW(1, 12),
Chris@82 845 VTW(1, 13),
Chris@82 846 VTW(1, 14),
Chris@82 847 VTW(1, 15),
Chris@82 848 VTW(1, 16),
Chris@82 849 VTW(1, 17),
Chris@82 850 VTW(1, 18),
Chris@82 851 VTW(1, 19),
Chris@82 852 VTW(1, 20),
Chris@82 853 VTW(1, 21),
Chris@82 854 VTW(1, 22),
Chris@82 855 VTW(1, 23),
Chris@82 856 VTW(1, 24),
Chris@82 857 VTW(1, 25),
Chris@82 858 VTW(1, 26),
Chris@82 859 VTW(1, 27),
Chris@82 860 VTW(1, 28),
Chris@82 861 VTW(1, 29),
Chris@82 862 VTW(1, 30),
Chris@82 863 VTW(1, 31),
Chris@82 864 {TW_NEXT, VL, 0}
Chris@82 865 };
Chris@82 866
Chris@82 867 static const hc2c_desc desc = { 32, XSIMD_STRING("hc2cbdftv_32"), twinstr, &GENUS, {233, 88, 16, 0} };
Chris@82 868
Chris@82 869 void XSIMD(codelet_hc2cbdftv_32) (planner *p) {
Chris@82 870 X(khc2c_register) (p, hc2cbdftv_32, &desc, HC2C_VIA_DFT);
Chris@82 871 }
Chris@82 872 #endif