t2fv_25.c

快速fft变换

		    T1v = VSUB(T1g, T1i);
		    T1w = VSUB(T1l, T1n);
		    T1x = VFMA(LDK(KP475528258), T1v, VMUL(LDK(KP293892626), T1w));
		    T2m = VFNMS(LDK(KP293892626), T1v, VMUL(LDK(KP475528258), T1w));
		    T1p = VMUL(LDK(KP559016994), VSUB(T1j, T1o));
		    T1s = VADD(T1j, T1o);
		    T1t = VFNMS(LDK(KP250000000), T1s, T1r);
		    T1u = VADD(T1p, T1t);
		    T3c = VADD(T1r, T1s);
		    T2n = VSUB(T1t, T1p);
	       }
	  }
	  {
	       V Ts, Tm, To, Tp, Th, Tj, Tk, Tr;
	       Tr = LD(&(x[WS(rs, 1)]), ms, &(x[WS(rs, 1)]));
	       Ts = BYTWJ(&(W[0]), Tr);
	       {
		    V Tl, Tn, Tg, Ti;
		    Tl = LD(&(x[WS(rs, 11)]), ms, &(x[WS(rs, 1)]));
		    Tm = BYTWJ(&(W[TWVL * 20]), Tl);
		    Tn = LD(&(x[WS(rs, 16)]), ms, &(x[0]));
		    To = BYTWJ(&(W[TWVL * 30]), Tn);
		    Tp = VADD(Tm, To);
		    Tg = LD(&(x[WS(rs, 6)]), ms, &(x[0]));
		    Th = BYTWJ(&(W[TWVL * 10]), Tg);
		    Ti = LD(&(x[WS(rs, 21)]), ms, &(x[WS(rs, 1)]));
		    Tj = BYTWJ(&(W[TWVL * 40]), Ti);
		    Tk = VADD(Th, Tj);
	       }
	       {
		    V Tw, Tx, Tq, Tt, Tu;
		    Tw = VSUB(Th, Tj);
		    Tx = VSUB(Tm, To);
		    Ty = VFMA(LDK(KP475528258), Tw, VMUL(LDK(KP293892626), Tx));
		    T2i = VFNMS(LDK(KP293892626), Tw, VMUL(LDK(KP475528258), Tx));
		    Tq = VMUL(LDK(KP559016994), VSUB(Tk, Tp));
		    Tt = VADD(Tk, Tp);
		    Tu = VFNMS(LDK(KP250000000), Tt, Ts);
		    Tv = VADD(Tq, Tu);
		    T38 = VADD(Ts, Tt);
		    T2j = VSUB(Tu, Tq);
	       }
	  }
	  {
	       V TM, TG, TI, TJ, TB, TD, TE, TL;
	       TL = LD(&(x[WS(rs, 4)]), ms, &(x[0]));
	       TM = BYTWJ(&(W[TWVL * 6]), TL);
	       {
		    V TF, TH, TA, TC;
		    TF = LD(&(x[WS(rs, 14)]), ms, &(x[0]));
		    TG = BYTWJ(&(W[TWVL * 26]), TF);
		    TH = LD(&(x[WS(rs, 19)]), ms, &(x[WS(rs, 1)]));
		    TI = BYTWJ(&(W[TWVL * 36]), TH);
		    TJ = VADD(TG, TI);
		    TA = LD(&(x[WS(rs, 9)]), ms, &(x[WS(rs, 1)]));
		    TB = BYTWJ(&(W[TWVL * 16]), TA);
		    TC = LD(&(x[WS(rs, 24)]), ms, &(x[0]));
		    TD = BYTWJ(&(W[TWVL * 46]), TC);
		    TE = VADD(TB, TD);
	       }
	       {
		    V TQ, TR, TK, TN, TO;
		    TQ = VSUB(TB, TD);
		    TR = VSUB(TG, TI);
		    TS = VFMA(LDK(KP475528258), TQ, VMUL(LDK(KP293892626), TR));
		    T2f = VFNMS(LDK(KP293892626), TQ, VMUL(LDK(KP475528258), TR));
		    TK = VMUL(LDK(KP559016994), VSUB(TE, TJ));
		    TN = VADD(TE, TJ);
		    TO = VFNMS(LDK(KP250000000), TN, TM);
		    TP = VADD(TK, TO);
		    T39 = VADD(TM, TN);
		    T2g = VSUB(TO, TK);
	       }
	  }
	  {
	       V T17, T11, T13, T14, TW, TY, TZ, T16;
	       T16 = LD(&(x[WS(rs, 2)]), ms, &(x[0]));
	       T17 = BYTWJ(&(W[TWVL * 2]), T16);
	       {
		    V T10, T12, TV, TX;
		    T10 = LD(&(x[WS(rs, 12)]), ms, &(x[0]));
		    T11 = BYTWJ(&(W[TWVL * 22]), T10);
		    T12 = LD(&(x[WS(rs, 17)]), ms, &(x[WS(rs, 1)]));
		    T13 = BYTWJ(&(W[TWVL * 32]), T12);
		    T14 = VADD(T11, T13);
		    TV = LD(&(x[WS(rs, 7)]), ms, &(x[WS(rs, 1)]));
		    TW = BYTWJ(&(W[TWVL * 12]), TV);
		    TX = LD(&(x[WS(rs, 22)]), ms, &(x[0]));
		    TY = BYTWJ(&(W[TWVL * 42]), TX);
		    TZ = VADD(TW, TY);
	       }
	       {
		    V T1b, T1c, T15, T18, T19;
		    T1b = VSUB(TW, TY);
		    T1c = VSUB(T11, T13);
		    T1d = VFMA(LDK(KP475528258), T1b, VMUL(LDK(KP293892626), T1c));
		    T2p = VFNMS(LDK(KP293892626), T1b, VMUL(LDK(KP475528258), T1c));
		    T15 = VMUL(LDK(KP559016994), VSUB(TZ, T14));
		    T18 = VADD(TZ, T14);
		    T19 = VFNMS(LDK(KP250000000), T18, T17);
		    T1a = VADD(T15, T19);
		    T3b = VADD(T17, T18);
		    T2q = VSUB(T19, T15);
	       }
	  }
	  {
	       V T3l, T3m, T3f, T3g, T3e, T3h, T3n, T3i;
	       {
		    V T3j, T3k, T3a, T3d;
		    T3j = VSUB(T38, T39);
		    T3k = VSUB(T3b, T3c);
		    T3l = VBYI(VFMA(LDK(KP951056516), T3j, VMUL(LDK(KP587785252), T3k)));
		    T3m = VBYI(VFNMS(LDK(KP587785252), T3j, VMUL(LDK(KP951056516), T3k)));
		    T3f = VADD(Tc, Td);
		    T3a = VADD(T38, T39);
		    T3d = VADD(T3b, T3c);
		    T3g = VADD(T3a, T3d);
		    T3e = VMUL(LDK(KP559016994), VSUB(T3a, T3d));
		    T3h = VFNMS(LDK(KP250000000), T3g, T3f);
	       }
	       ST(&(x[0]), VADD(T3f, T3g), ms, &(x[0]));
	       T3n = VSUB(T3h, T3e);
	       ST(&(x[WS(rs, 10)]), VADD(T3m, T3n), ms, &(x[0]));
	       ST(&(x[WS(rs, 15)]), VSUB(T3n, T3m), ms, &(x[WS(rs, 1)]));
	       T3i = VADD(T3e, T3h);
	       ST(&(x[WS(rs, 5)]), VSUB(T3i, T3l), ms, &(x[WS(rs, 1)]));
	       ST(&(x[WS(rs, 20)]), VADD(T3l, T3i), ms, &(x[0]));
	  }
	  {
	       V Tf, T1Z, T20, T21, T29, T2a, T2b, T26, T27, T28, T22, T23, T24, T1L, T1U;
	       V T1Q, T1S, T1A, T1V, T1N, T1O, T2d, T2e;
	       Tf = VADD(Tb, Te);
	       T1Z = VFMA(LDK(KP1_688655851), Ty, VMUL(LDK(KP535826794), Tv));
	       T20 = VFMA(LDK(KP1_541026485), TS, VMUL(LDK(KP637423989), TP));
	       T21 = VSUB(T1Z, T20);
	       T29 = VFMA(LDK(KP851558583), T1d, VMUL(LDK(KP904827052), T1a));
	       T2a = VFMA(LDK(KP1_984229402), T1x, VMUL(LDK(KP125333233), T1u));
	       T2b = VADD(T29, T2a);
	       T26 = VFNMS(LDK(KP844327925), Tv, VMUL(LDK(KP1_071653589), Ty));
	       T27 = VFNMS(LDK(KP1_274847979), TS, VMUL(LDK(KP770513242), TP));
	       T28 = VADD(T26, T27);
	       T22 = VFNMS(LDK(KP425779291), T1a, VMUL(LDK(KP1_809654104), T1d));
	       T23 = VFNMS(LDK(KP992114701), T1u, VMUL(LDK(KP250666467), T1x));
	       T24 = VADD(T22, T23);
	       {
		    V T1F, T1G, T1H, T1I, T1J, T1K;
		    T1F = VFMA(LDK(KP1_937166322), Ty, VMUL(LDK(KP248689887), Tv));
		    T1G = VFMA(LDK(KP1_071653589), TS, VMUL(LDK(KP844327925), TP));
		    T1H = VADD(T1F, T1G);
		    T1I = VFMA(LDK(KP1_752613360), T1d, VMUL(LDK(KP481753674), T1a));
		    T1J = VFMA(LDK(KP1_457937254), T1x, VMUL(LDK(KP684547105), T1u));
		    T1K = VADD(T1I, T1J);
		    T1L = VADD(T1H, T1K);
		    T1U = VSUB(T1J, T1I);
		    T1Q = VMUL(LDK(KP559016994), VSUB(T1K, T1H));
		    T1S = VSUB(T1G, T1F);
	       }
	       {
		    V Tz, TT, TU, T1e, T1y, T1z;
		    Tz = VFNMS(LDK(KP497379774), Ty, VMUL(LDK(KP968583161), Tv));
		    TT = VFNMS(LDK(KP1_688655851), TS, VMUL(LDK(KP535826794), TP));
		    TU = VADD(Tz, TT);
		    T1e = VFNMS(LDK(KP963507348), T1d, VMUL(LDK(KP876306680), T1a));
		    T1y = VFNMS(LDK(KP1_369094211), T1x, VMUL(LDK(KP728968627), T1u));
		    T1z = VADD(T1e, T1y);
		    T1A = VADD(TU, T1z);
		    T1V = VMUL(LDK(KP559016994), VSUB(TU, T1z));
		    T1N = VSUB(TT, Tz);
		    T1O = VSUB(T1e, T1y);
	       }
	       {
		    V T1B, T1M, T25, T2c;
		    T1B = VADD(Tf, T1A);
		    T1M = VBYI(VADD(T1E, T1L));
		    ST(&(x[WS(rs, 1)]), VSUB(T1B, T1M), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 24)]), VADD(T1B, T1M), ms, &(x[0]));
		    T25 = VADD(Tf, VADD(T21, T24));
		    T2c = VBYI(VADD(T1E, VSUB(T28, T2b)));
		    ST(&(x[WS(rs, 21)]), VSUB(T25, T2c), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 4)]), VADD(T25, T2c), ms, &(x[0]));
	       }
	       T2d = VBYI(VADD(T1E, VFMA(LDK(KP309016994), T28, VFMA(LDK(KP587785252), VSUB(T23, T22), VFNMS(LDK(KP951056516), VADD(T1Z, T20), VMUL(LDK(KP809016994), T2b))))));
	       T2e = VFMA(LDK(KP309016994), T21, VFMA(LDK(KP951056516), VSUB(T26, T27), VFMA(LDK(KP587785252), VSUB(T2a, T29), VFNMS(LDK(KP809016994), T24, Tf))));
	       ST(&(x[WS(rs, 9)]), VADD(T2d, T2e), ms, &(x[WS(rs, 1)]));
	       ST(&(x[WS(rs, 16)]), VSUB(T2e, T2d), ms, &(x[0]));
	       {
		    V T1R, T1X, T1W, T1Y, T1P, T1T;
		    T1P = VFMS(LDK(KP250000000), T1L, T1E);
		    T1R = VBYI(VADD(VFMA(LDK(KP587785252), T1N, VMUL(LDK(KP951056516), T1O)), VSUB(T1P, T1Q)));
		    T1X = VBYI(VADD(VFNMS(LDK(KP587785252), T1O, VMUL(LDK(KP951056516), T1N)), VADD(T1P, T1Q)));
		    T1T = VFNMS(LDK(KP250000000), T1A, Tf);
		    T1W = VFMA(LDK(KP587785252), T1S, VFNMS(LDK(KP951056516), T1U, VSUB(T1T, T1V)));
		    T1Y = VFMA(LDK(KP951056516), T1S, VADD(T1V, VFMA(LDK(KP587785252), T1U, T1T)));
		    ST(&(x[WS(rs, 11)]), VADD(T1R, T1W), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 19)]), VSUB(T1Y, T1X), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 14)]), VSUB(T1W, T1R), ms, &(x[0]));
		    ST(&(x[WS(rs, 6)]), VADD(T1X, T1Y), ms, &(x[0]));
	       }
	  }
	  {
	       V T2u, T2w, T2h, T2k, T2l, T2A, T2B, T2C, T2o, T2r, T2s, T2x, T2y, T2z, T2M;
	       V T2X, T2N, T2W, T2R, T31, T2U, T30, T2E, T2F;
	       T2u = VFNMS(LDK(KP587785252), T1C, T2t);
	       T2w = VSUB(Te, Tb);
	       T2h = VFNMS(LDK(KP125333233), T2g, VMUL(LDK(KP1_984229402), T2f));
	       T2k = VFMA(LDK(KP1_457937254), T2i, VMUL(LDK(KP684547105), T2j));
	       T2l = VSUB(T2h, T2k);
	       T2A = VFNMS(LDK(KP1_996053456), T2p, VMUL(LDK(KP062790519), T2q));
	       T2B = VFMA(LDK(KP1_541026485), T2m, VMUL(LDK(KP637423989), T2n));
	       T2C = VSUB(T2A, T2B);
	       T2o = VFNMS(LDK(KP770513242), T2n, VMUL(LDK(KP1_274847979), T2m));
	       T2r = VFMA(LDK(KP125581039), T2p, VMUL(LDK(KP998026728), T2q));
	       T2s = VSUB(T2o, T2r);
	       T2x = VFNMS(LDK(KP1_369094211), T2i, VMUL(LDK(KP728968627), T2j));
	       T2y = VFMA(LDK(KP250666467), T2f, VMUL(LDK(KP992114701), T2g));
	       T2z = VSUB(T2x, T2y);
	       {
		    V T2G, T2H, T2I, T2J, T2K, T2L;
		    T2G = VFNMS(LDK(KP481753674), T2j, VMUL(LDK(KP1_752613360), T2i));
		    T2H = VFMA(LDK(KP851558583), T2f, VMUL(LDK(KP904827052), T2g));
		    T2I = VSUB(T2G, T2H);
		    T2J = VFNMS(LDK(KP844327925), T2q, VMUL(LDK(KP1_071653589), T2p));
		    T2K = VFNMS(LDK(KP998026728), T2n, VMUL(LDK(KP125581039), T2m));
		    T2L = VADD(T2J, T2K);
		    T2M = VMUL(LDK(KP559016994), VSUB(T2I, T2L));
		    T2X = VSUB(T2J, T2K);
		    T2N = VADD(T2I, T2L);
		    T2W = VADD(T2G, T2H);
	       }
	       {
		    V T2P, T2Q, T2Y, T2S, T2T, T2Z;
		    T2P = VFNMS(LDK(KP425779291), T2g, VMUL(LDK(KP1_809654104), T2f));
		    T2Q = VFMA(LDK(KP963507348), T2i, VMUL(LDK(KP876306680), T2j));
		    T2Y = VADD(T2Q, T2P);
		    T2S = VFMA(LDK(KP1_688655851), T2p, VMUL(LDK(KP535826794), T2q));
		    T2T = VFMA(LDK(KP1_996053456), T2m, VMUL(LDK(KP062790519), T2n));
		    T2Z = VADD(T2S, T2T);
		    T2R = VSUB(T2P, T2Q);
		    T31 = VADD(T2Y, T2Z);
		    T2U = VSUB(T2S, T2T);
		    T30 = VMUL(LDK(KP559016994), VSUB(T2Y, T2Z));
	       }
	       {
		    V T36, T37, T2v, T2D;
		    T36 = VBYI(VADD(T2u, T2N));
		    T37 = VADD(T2w, T31);
		    ST(&(x[WS(rs, 2)]), VADD(T36, T37), ms, &(x[0]));
		    ST(&(x[WS(rs, 23)]), VSUB(T37, T36), ms, &(x[WS(rs, 1)]));
		    T2v = VBYI(VSUB(VADD(T2l, T2s), T2u));
		    T2D = VADD(T2w, VADD(T2z, T2C));
		    ST(&(x[WS(rs, 3)]), VADD(T2v, T2D), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 22)]), VSUB(T2D, T2v), ms, &(x[0]));
	       }
	       T2E = VFMA(LDK(KP309016994), T2z, VFNMS(LDK(KP809016994), T2C, VFNMS(LDK(KP587785252), VADD(T2r, T2o), VFNMS(LDK(KP951056516), VADD(T2k, T2h), T2w))));
	       T2F = VBYI(VSUB(VFNMS(LDK(KP587785252), VADD(T2A, T2B), VFNMS(LDK(KP809016994), T2s, VFNMS(LDK(KP951056516), VADD(T2x, T2y), VMUL(LDK(KP309016994), T2l)))), T2u));
	       ST(&(x[WS(rs, 17)]), VSUB(T2E, T2F), ms, &(x[WS(rs, 1)]));
	       ST(&(x[WS(rs, 8)]), VADD(T2E, T2F), ms, &(x[0]));
	       {
		    V T2V, T34, T33, T35, T2O, T32;
		    T2O = VFNMS(LDK(KP250000000), T2N, T2u);
		    T2V = VBYI(VADD(T2M, VADD(T2O, VFNMS(LDK(KP587785252), T2U, VMUL(LDK(KP951056516), T2R)))));
		    T34 = VBYI(VADD(T2O, VSUB(VFMA(LDK(KP587785252), T2R, VMUL(LDK(KP951056516), T2U)), T2M)));
		    T32 = VFNMS(LDK(KP250000000), T31, T2w);
		    T33 = VFMA(LDK(KP951056516), T2W, VFMA(LDK(KP587785252), T2X, VADD(T30, T32)));
		    T35 = VFMA(LDK(KP587785252), T2W, VSUB(VFNMS(LDK(KP951056516), T2X, T32), T30));
		    ST(&(x[WS(rs, 7)]), VADD(T2V, T33), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 13)]), VSUB(T35, T34), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 18)]), VSUB(T33, T2V), ms, &(x[0]));
		    ST(&(x[WS(rs, 12)]), VADD(T34, T35), ms, &(x[0]));
	       }
	  }
     }
}

static const tw_instr twinstr[] = {
     VTW(0, 1),
     VTW(0, 2),
     VTW(0, 3),
     VTW(0, 4),
     VTW(0, 5),
     VTW(0, 6),
     VTW(0, 7),
     VTW(0, 8),
     VTW(0, 9),
     VTW(0, 10),
     VTW(0, 11),
     VTW(0, 12),
     VTW(0, 13),
     VTW(0, 14),
     VTW(0, 15),
     VTW(0, 16),
     VTW(0, 17),
     VTW(0, 18),
     VTW(0, 19),
     VTW(0, 20),
     VTW(0, 21),
     VTW(0, 22),
     VTW(0, 23),
     VTW(0, 24),
     {TW_NEXT, VL, 0}
};

static const ct_desc desc = { 25, "t2fv_25", twinstr, &GENUS, {170, 110, 78, 0}, 0, 0, 0 };

void X(codelet_t2fv_25) (planner *p) {
     X(kdft_dit_register) (p, t2fv_25, &desc);
}
#endif				/* HAVE_FMA */