t3bv_25.c

快速fft变换

		    T20 = VFNMS(LDK(KP250000000), T1Z, T1Y);
		    T2d = VSUB(T1M, T1P);
		    T2e = VMUL(LDK(KP951056516), T2d);
		    T39 = VMUL(LDK(KP587785252), T2d);
	       }
	       {
		    V T1B, T1u, T1x, T1y, T1n, T1q, T1r, T1A;
		    T1A = LD(&(x[WS(rs, 3)]), ms, &(x[WS(rs, 1)]));
		    T1B = VZMUL(Td, T1A);
		    {
			 V T1t, T1w, T1m, T1p;
			 T1t = LD(&(x[WS(rs, 13)]), ms, &(x[WS(rs, 1)]));
			 T1u = VZMUL(T1s, T1t);
			 T1w = LD(&(x[WS(rs, 18)]), ms, &(x[0]));
			 T1x = VZMUL(T1v, T1w);
			 T1y = VADD(T1u, T1x);
			 T1m = LD(&(x[WS(rs, 8)]), ms, &(x[0]));
			 T1n = VZMUL(TF, T1m);
			 T1p = LD(&(x[WS(rs, 23)]), ms, &(x[WS(rs, 1)]));
			 T1q = VZMUL(T1o, T1p);
			 T1r = VADD(T1n, T1q);
		    }
		    {
			 V T1F, T1G, T1z, T1C, T1D;
			 T1F = VSUB(T1n, T1q);
			 T1G = VSUB(T1u, T1x);
			 T1H = VFMA(LDK(KP475528258), T1F, VMUL(LDK(KP293892626), T1G));
			 T2T = VFNMS(LDK(KP475528258), T1G, VMUL(LDK(KP293892626), T1F));
			 T1z = VMUL(LDK(KP559016994), VSUB(T1r, T1y));
			 T1C = VADD(T1r, T1y);
			 T1D = VFNMS(LDK(KP250000000), T1C, T1B);
			 T1E = VADD(T1z, T1D);
			 T3C = VADD(T1B, T1C);
			 T2S = VSUB(T1D, T1z);
		    }
	       }
	       {
		    V Tp, Tc, Ti, Tm, T3, T6, Tl, To;
		    To = LD(&(x[WS(rs, 4)]), ms, &(x[0]));
		    Tp = VZMUL(Te, To);
		    {
			 V Tb, Th, T2, T5;
			 Tb = LD(&(x[WS(rs, 14)]), ms, &(x[0]));
			 Tc = VZMUL(Ta, Tb);
			 Th = LD(&(x[WS(rs, 19)]), ms, &(x[WS(rs, 1)]));
			 Ti = VZMUL(Tg, Th);
			 Tm = VADD(Tc, Ti);
			 T2 = LD(&(x[WS(rs, 9)]), ms, &(x[WS(rs, 1)]));
			 T3 = VZMUL(T1, T2);
			 T5 = LD(&(x[WS(rs, 24)]), ms, &(x[0]));
			 T6 = VZMUL(T4, T5);
			 Tl = VADD(T3, T6);
		    }
		    {
			 V T7, Tj, Tn, Tq, Tr;
			 T7 = VSUB(T3, T6);
			 Tj = VSUB(Tc, Ti);
			 Tk = VFMA(LDK(KP475528258), T7, VMUL(LDK(KP293892626), Tj));
			 T2G = VFNMS(LDK(KP475528258), Tj, VMUL(LDK(KP293892626), T7));
			 Tn = VMUL(LDK(KP559016994), VSUB(Tl, Tm));
			 Tq = VADD(Tl, Tm);
			 Tr = VFNMS(LDK(KP250000000), Tq, Tp);
			 Ts = VADD(Tn, Tr);
			 T3z = VADD(Tp, Tq);
			 T2F = VSUB(Tr, Tn);
		    }
	       }
	       {
		    V TP, TE, TI, TM, Tw, Tz, TL, TO;
		    TO = LD(&(x[WS(rs, 1)]), ms, &(x[WS(rs, 1)]));
		    TP = VZMUL(T8, TO);
		    {
			 V TD, TH, Tv, Ty;
			 TD = LD(&(x[WS(rs, 11)]), ms, &(x[WS(rs, 1)]));
			 TE = VZMUL(TC, TD);
			 TH = LD(&(x[WS(rs, 16)]), ms, &(x[0]));
			 TI = VZMUL(TG, TH);
			 TM = VADD(TE, TI);
			 Tv = LD(&(x[WS(rs, 6)]), ms, &(x[0]));
			 Tw = VZMUL(Tu, Tv);
			 Ty = LD(&(x[WS(rs, 21)]), ms, &(x[WS(rs, 1)]));
			 Tz = VZMUL(Tx, Ty);
			 TL = VADD(Tw, Tz);
		    }
		    {
			 V TA, TJ, TN, TQ, TR;
			 TA = VSUB(Tw, Tz);
			 TJ = VSUB(TE, TI);
			 TK = VFMA(LDK(KP475528258), TA, VMUL(LDK(KP293892626), TJ));
			 T2I = VFNMS(LDK(KP475528258), TJ, VMUL(LDK(KP293892626), TA));
			 TN = VMUL(LDK(KP559016994), VSUB(TL, TM));
			 TQ = VADD(TL, TM);
			 TR = VFNMS(LDK(KP250000000), TQ, TP);
			 TS = VADD(TN, TR);
			 T3y = VADD(TP, TQ);
			 T2J = VSUB(TR, TN);
		    }
	       }
	       {
		    V T1e, T17, T1a, T1b, T10, T13, T14, T1d;
		    T1d = LD(&(x[WS(rs, 2)]), ms, &(x[0]));
		    T1e = VZMUL(TB, T1d);
		    {
			 V T16, T19, TZ, T12;
			 T16 = LD(&(x[WS(rs, 12)]), ms, &(x[0]));
			 T17 = VZMUL(T15, T16);
			 T19 = LD(&(x[WS(rs, 17)]), ms, &(x[WS(rs, 1)]));
			 T1a = VZMUL(T18, T19);
			 T1b = VADD(T17, T1a);
			 TZ = LD(&(x[WS(rs, 7)]), ms, &(x[WS(rs, 1)]));
			 T10 = VZMUL(TY, TZ);
			 T12 = LD(&(x[WS(rs, 22)]), ms, &(x[0]));
			 T13 = VZMUL(T11, T12);
			 T14 = VADD(T10, T13);
		    }
		    {
			 V T1i, T1j, T1c, T1f, T1g;
			 T1i = VSUB(T10, T13);
			 T1j = VSUB(T17, T1a);
			 T1k = VFMA(LDK(KP475528258), T1i, VMUL(LDK(KP293892626), T1j));
			 T2Q = VFNMS(LDK(KP475528258), T1j, VMUL(LDK(KP293892626), T1i));
			 T1c = VMUL(LDK(KP559016994), VSUB(T14, T1b));
			 T1f = VADD(T14, T1b);
			 T1g = VFNMS(LDK(KP250000000), T1f, T1e);
			 T1h = VADD(T1c, T1g);
			 T3B = VADD(T1e, T1f);
			 T2P = VSUB(T1g, T1c);
		    }
	       }
	       {
		    V T3E, T3M, T3I, T3J, T3H, T3K, T3N, T3L;
		    {
			 V T3A, T3D, T3F, T3G;
			 T3A = VSUB(T3y, T3z);
			 T3D = VSUB(T3B, T3C);
			 T3E = VBYI(VFMA(LDK(KP951056516), T3A, VMUL(LDK(KP587785252), T3D)));
			 T3M = VBYI(VFNMS(LDK(KP951056516), T3D, VMUL(LDK(KP587785252), T3A)));
			 T3I = VADD(T1Y, T1Z);
			 T3F = VADD(T3y, T3z);
			 T3G = VADD(T3B, T3C);
			 T3J = VADD(T3F, T3G);
			 T3H = VMUL(LDK(KP559016994), VSUB(T3F, T3G));
			 T3K = VFNMS(LDK(KP250000000), T3J, T3I);
		    }
		    ST(&(x[0]), VADD(T3I, T3J), ms, &(x[0]));
		    T3N = VSUB(T3K, T3H);
		    ST(&(x[WS(rs, 10)]), VADD(T3M, T3N), ms, &(x[0]));
		    ST(&(x[WS(rs, 15)]), VSUB(T3N, T3M), ms, &(x[WS(rs, 1)]));
		    T3L = VADD(T3H, T3K);
		    ST(&(x[WS(rs, 5)]), VADD(T3E, T3L), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 20)]), VSUB(T3L, T3E), ms, &(x[0]));
	       }
	       {
		    V T2X, T3a, T3i, T3j, T3k, T3s, T3t, T3u, T3l, T3m, T3n, T3p, T3q, T3r, T2L;
		    V T3b, T32, T38, T2W, T35, T2Y, T34, T3w, T3x;
		    T2X = VSUB(T20, T1X);
		    T3a = VFNMS(LDK(KP951056516), T2f, T39);
		    T3i = VFMA(LDK(KP1_369094211), T2I, VMUL(LDK(KP728968627), T2J));
		    T3j = VFNMS(LDK(KP992114701), T2F, VMUL(LDK(KP250666467), T2G));
		    T3k = VADD(T3i, T3j);
		    T3s = VFNMS(LDK(KP125581039), T2Q, VMUL(LDK(KP998026728), T2P));
		    T3t = VFMA(LDK(KP1_274847979), T2T, VMUL(LDK(KP770513242), T2S));
		    T3u = VADD(T3s, T3t);
		    T3l = VFMA(LDK(KP1_996053456), T2Q, VMUL(LDK(KP062790519), T2P));
		    T3m = VFNMS(LDK(KP637423989), T2S, VMUL(LDK(KP1_541026485), T2T));
		    T3n = VADD(T3l, T3m);
		    T3p = VFNMS(LDK(KP1_457937254), T2I, VMUL(LDK(KP684547105), T2J));
		    T3q = VFMA(LDK(KP1_984229402), T2G, VMUL(LDK(KP125333233), T2F));
		    T3r = VADD(T3p, T3q);
		    {
			 V T2H, T2K, T36, T30, T31, T37;
			 T2H = VFNMS(LDK(KP851558583), T2G, VMUL(LDK(KP904827052), T2F));
			 T2K = VFMA(LDK(KP1_752613360), T2I, VMUL(LDK(KP481753674), T2J));
			 T36 = VADD(T2K, T2H);
			 T30 = VFMA(LDK(KP1_071653589), T2Q, VMUL(LDK(KP844327925), T2P));
			 T31 = VFMA(LDK(KP125581039), T2T, VMUL(LDK(KP998026728), T2S));
			 T37 = VADD(T30, T31);
			 T2L = VSUB(T2H, T2K);
			 T3b = VADD(T36, T37);
			 T32 = VSUB(T30, T31);
			 T38 = VMUL(LDK(KP559016994), VSUB(T36, T37));
		    }
		    {
			 V T2M, T2N, T2O, T2R, T2U, T2V;
			 T2M = VFNMS(LDK(KP963507348), T2I, VMUL(LDK(KP876306680), T2J));
			 T2N = VFMA(LDK(KP1_809654104), T2G, VMUL(LDK(KP425779291), T2F));
			 T2O = VSUB(T2M, T2N);
			 T2R = VFNMS(LDK(KP1_688655851), T2Q, VMUL(LDK(KP535826794), T2P));
			 T2U = VFNMS(LDK(KP1_996053456), T2T, VMUL(LDK(KP062790519), T2S));
			 T2V = VADD(T2R, T2U);
			 T2W = VMUL(LDK(KP559016994), VSUB(T2O, T2V));
			 T35 = VSUB(T2R, T2U);
			 T2Y = VADD(T2O, T2V);
			 T34 = VADD(T2M, T2N);
		    }
		    {
			 V T3g, T3h, T3o, T3v;
			 T3g = VADD(T2X, T2Y);
			 T3h = VBYI(VADD(T3a, T3b));
			 ST(&(x[WS(rs, 23)]), VSUB(T3g, T3h), ms, &(x[WS(rs, 1)]));
			 ST(&(x[WS(rs, 2)]), VADD(T3g, T3h), ms, &(x[0]));
			 T3o = VADD(T2X, VADD(T3k, T3n));
			 T3v = VBYI(VSUB(VADD(T3r, T3u), T3a));
			 ST(&(x[WS(rs, 22)]), VSUB(T3o, T3v), ms, &(x[0]));
			 ST(&(x[WS(rs, 3)]), VADD(T3o, T3v), ms, &(x[WS(rs, 1)]));
		    }
		    T3w = VBYI(VSUB(VFMA(LDK(KP951056516), VSUB(T3i, T3j), VFMA(LDK(KP309016994), T3r, VFNMS(LDK(KP809016994), T3u, VMUL(LDK(KP587785252), VSUB(T3l, T3m))))), T3a));
		    T3x = VFMA(LDK(KP309016994), T3k, VFMA(LDK(KP951056516), VSUB(T3q, T3p), VFMA(LDK(KP587785252), VSUB(T3t, T3s), VFNMS(LDK(KP809016994), T3n, T2X))));
		    ST(&(x[WS(rs, 8)]), VADD(T3w, T3x), ms, &(x[0]));
		    ST(&(x[WS(rs, 17)]), VSUB(T3x, T3w), ms, &(x[WS(rs, 1)]));
		    {
			 V T33, T3e, T3d, T3f, T2Z, T3c;
			 T2Z = VFNMS(LDK(KP250000000), T2Y, T2X);
			 T33 = VFMA(LDK(KP951056516), T2L, VADD(T2W, VFNMS(LDK(KP587785252), T32, T2Z)));
			 T3e = VFMA(LDK(KP587785252), T2L, VFMA(LDK(KP951056516), T32, VSUB(T2Z, T2W)));
			 T3c = VFNMS(LDK(KP250000000), T3b, T3a);
			 T3d = VBYI(VADD(VFMA(LDK(KP951056516), T34, VMUL(LDK(KP587785252), T35)), VADD(T38, T3c)));
			 T3f = VBYI(VADD(VFNMS(LDK(KP951056516), T35, VMUL(LDK(KP587785252), T34)), VSUB(T3c, T38)));
			 ST(&(x[WS(rs, 18)]), VSUB(T33, T3d), ms, &(x[0]));
			 ST(&(x[WS(rs, 12)]), VADD(T3e, T3f), ms, &(x[0]));
			 ST(&(x[WS(rs, 7)]), VADD(T33, T3d), ms, &(x[WS(rs, 1)]));
			 ST(&(x[WS(rs, 13)]), VSUB(T3e, T3f), ms, &(x[WS(rs, 1)]));
		    }
	       }
	       {
		    V T21, T2h, T2p, T2q, T2r, T2z, T2A, T2B, T2s, T2t, T2u, T2w, T2x, T2y, TU;
		    V T2i, T26, T2c, T1K, T29, T22, T28, T2D, T2E;
		    T21 = VADD(T1X, T20);
		    T2h = VADD(T2e, T2g);
		    T2p = VFMA(LDK(KP1_688655851), TK, VMUL(LDK(KP535826794), TS));
		    T2q = VFMA(LDK(KP1_541026485), Tk, VMUL(LDK(KP637423989), Ts));
		    T2r = VSUB(T2p, T2q);
		    T2z = VFMA(LDK(KP851558583), T1k, VMUL(LDK(KP904827052), T1h));
		    T2A = VFMA(LDK(KP1_984229402), T1H, VMUL(LDK(KP125333233), T1E));
		    T2B = VADD(T2z, T2A);
		    T2s = VFNMS(LDK(KP425779291), T1h, VMUL(LDK(KP1_809654104), T1k));
		    T2t = VFNMS(LDK(KP992114701), T1E, VMUL(LDK(KP250666467), T1H));
		    T2u = VADD(T2s, T2t);
		    T2w = VFNMS(LDK(KP1_071653589), TK, VMUL(LDK(KP844327925), TS));
		    T2x = VFNMS(LDK(KP770513242), Ts, VMUL(LDK(KP1_274847979), Tk));
		    T2y = VADD(T2w, T2x);
		    {
			 V Tt, TT, T2a, T24, T25, T2b;
			 Tt = VFMA(LDK(KP1_071653589), Tk, VMUL(LDK(KP844327925), Ts));
			 TT = VFMA(LDK(KP1_937166322), TK, VMUL(LDK(KP248689887), TS));
			 T2a = VADD(TT, Tt);
			 T24 = VFMA(LDK(KP1_752613360), T1k, VMUL(LDK(KP481753674), T1h));
			 T25 = VFMA(LDK(KP1_457937254), T1H, VMUL(LDK(KP684547105), T1E));
			 T2b = VADD(T24, T25);
			 TU = VSUB(Tt, TT);
			 T2i = VADD(T2a, T2b);
			 T26 = VSUB(T24, T25);
			 T2c = VMUL(LDK(KP559016994), VSUB(T2a, T2b));
		    }
		    {
			 V TV, TW, TX, T1l, T1I, T1J;
			 TV = VFNMS(LDK(KP497379774), TK, VMUL(LDK(KP968583161), TS));
			 TW = VFNMS(LDK(KP1_688655851), Tk, VMUL(LDK(KP535826794), Ts));
			 TX = VADD(TV, TW);
			 T1l = VFNMS(LDK(KP963507348), T1k, VMUL(LDK(KP876306680), T1h));
			 T1I = VFNMS(LDK(KP1_369094211), T1H, VMUL(LDK(KP728968627), T1E));
			 T1J = VADD(T1l, T1I);
			 T1K = VMUL(LDK(KP559016994), VSUB(TX, T1J));
			 T29 = VSUB(T1l, T1I);
			 T22 = VADD(TX, T1J);
			 T28 = VSUB(TV, TW);
		    }
		    {
			 V T2n, T2o, T2v, T2C;
			 T2n = VADD(T21, T22);
			 T2o = VBYI(VADD(T2h, T2i));
			 ST(&(x[WS(rs, 24)]), VSUB(T2n, T2o), ms, &(x[0]));
			 ST(&(x[WS(rs, 1)]), VADD(T2n, T2o), ms, &(x[WS(rs, 1)]));
			 T2v = VADD(T21, VADD(T2r, T2u));
			 T2C = VBYI(VSUB(VADD(T2y, T2B), T2h));
			 ST(&(x[WS(rs, 21)]), VSUB(T2v, T2C), ms, &(x[WS(rs, 1)]));
			 ST(&(x[WS(rs, 4)]), VADD(T2v, T2C), ms, &(x[0]));
		    }
		    T2D = VBYI(VSUB(VFMA(LDK(KP309016994), T2y, VFMA(LDK(KP951056516), VADD(T2p, T2q), VFNMS(LDK(KP809016994), T2B, VMUL(LDK(KP587785252), VSUB(T2s, T2t))))), T2h));
		    T2E = VFMA(LDK(KP951056516), VSUB(T2x, T2w), VFMA(LDK(KP309016994), T2r, VFMA(LDK(KP587785252), VSUB(T2A, T2z), VFNMS(LDK(KP809016994), T2u, T21))));
		    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 T27, T2l, T2k, T2m, T23, T2j;
			 T23 = VFNMS(LDK(KP250000000), T22, T21);
			 T27 = VFMA(LDK(KP951056516), TU, VADD(T1K, VFNMS(LDK(KP587785252), T26, T23)));
			 T2l = VFMA(LDK(KP587785252), TU, VFMA(LDK(KP951056516), T26, VSUB(T23, T1K)));
			 T2j = VFNMS(LDK(KP250000000), T2i, T2h);
			 T2k = VBYI(VADD(VFMA(LDK(KP951056516), T28, VMUL(LDK(KP587785252), T29)), VADD(T2c, T2j)));
			 T2m = VBYI(VADD(VFNMS(LDK(KP951056516), T29, VMUL(LDK(KP587785252), T28)), VSUB(T2j, T2c)));
			 ST(&(x[WS(rs, 19)]), VSUB(T27, T2k), ms, &(x[WS(rs, 1)]));
			 ST(&(x[WS(rs, 11)]), VADD(T2l, T2m), ms, &(x[WS(rs, 1)]));
			 ST(&(x[WS(rs, 6)]), VADD(T27, T2k), ms, &(x[0]));
			 ST(&(x[WS(rs, 14)]), VSUB(T2l, T2m), ms, &(x[0]));
		    }
	       }
	  }
     }
}

static const tw_instr twinstr[] = {
     VTW(0, 1),
     VTW(0, 3),
     VTW(0, 9),
     VTW(0, 24),
     {TW_NEXT, VL, 0}
};

static const ct_desc desc = { 25, "t3bv_25", twinstr, &GENUS, {191, 151, 77, 0}, 0, 0, 0 };

void X(codelet_t3bv_25) (planner *p) {
     X(kdft_dit_register) (p, t3bv_25, &desc);
}
#endif				/* HAVE_FMA */