t3fv_25.c

快速fft变换

		    T21 = VSUB(Tf, Ti);
		    T2R = VMUL(LDK(KP951056516), T21);
		    T22 = VFMA(LDK(KP951056516), T20, VMUL(LDK(KP587785252), T21));
	       }
	       {
		    V T1P, T1I, T1L, T1M, T1B, T1E, T1F, T1O;
		    T1O = LD(&(x[WS(rs, 3)]), ms, &(x[WS(rs, 1)]));
		    T1P = VZMULJ(T2, T1O);
		    {
			 V T1H, T1K, T1A, T1D;
			 T1H = LD(&(x[WS(rs, 13)]), ms, &(x[WS(rs, 1)]));
			 T1I = VZMULJ(T1G, T1H);
			 T1K = LD(&(x[WS(rs, 18)]), ms, &(x[0]));
			 T1L = VZMULJ(T1J, T1K);
			 T1M = VADD(T1I, T1L);
			 T1A = LD(&(x[WS(rs, 8)]), ms, &(x[0]));
			 T1B = VZMULJ(TA, T1A);
			 T1D = LD(&(x[WS(rs, 23)]), ms, &(x[WS(rs, 1)]));
			 T1E = VZMULJ(T1C, T1D);
			 T1F = VADD(T1B, T1E);
		    }
		    {
			 V T1T, T1U, T1N, T1Q, T1R;
			 T1T = VSUB(T1B, T1E);
			 T1U = VSUB(T1I, T1L);
			 T1V = VFMA(LDK(KP475528258), T1T, VMUL(LDK(KP293892626), T1U));
			 T2K = VFNMS(LDK(KP293892626), T1T, VMUL(LDK(KP475528258), T1U));
			 T1N = VMUL(LDK(KP559016994), VSUB(T1F, T1M));
			 T1Q = VADD(T1F, T1M);
			 T1R = VFNMS(LDK(KP250000000), T1Q, T1P);
			 T1S = VADD(T1N, T1R);
			 T3A = VADD(T1P, T1Q);
			 T2L = VSUB(T1R, T1N);
		    }
	       }
	       {
		    V TH, Tz, TD, TE, Tr, Tu, Tv, TG;
		    TG = LD(&(x[WS(rs, 1)]), ms, &(x[WS(rs, 1)]));
		    TH = VZMULJ(T1, TG);
		    {
			 V Ty, TC, Tq, Tt;
			 Ty = LD(&(x[WS(rs, 11)]), ms, &(x[WS(rs, 1)]));
			 Tz = VZMULJ(Tx, Ty);
			 TC = LD(&(x[WS(rs, 16)]), ms, &(x[0]));
			 TD = VZMULJ(TB, TC);
			 TE = VADD(Tz, TD);
			 Tq = LD(&(x[WS(rs, 6)]), ms, &(x[0]));
			 Tr = VZMULJ(Tp, Tq);
			 Tt = LD(&(x[WS(rs, 21)]), ms, &(x[WS(rs, 1)]));
			 Tu = VZMULJ(Ts, Tt);
			 Tv = VADD(Tr, Tu);
		    }
		    {
			 V TL, TM, TF, TI, TJ;
			 TL = VSUB(Tr, Tu);
			 TM = VSUB(Tz, TD);
			 TN = VFMA(LDK(KP475528258), TL, VMUL(LDK(KP293892626), TM));
			 T2G = VFNMS(LDK(KP293892626), TL, VMUL(LDK(KP475528258), TM));
			 TF = VMUL(LDK(KP559016994), VSUB(Tv, TE));
			 TI = VADD(Tv, TE);
			 TJ = VFNMS(LDK(KP250000000), TI, TH);
			 TK = VADD(TF, TJ);
			 T3w = VADD(TH, TI);
			 T2H = VSUB(TJ, TF);
		    }
	       }
	       {
		    V T13, TW, TZ, T10, TQ, TS, TT, T12;
		    T12 = LD(&(x[WS(rs, 4)]), ms, &(x[0]));
		    T13 = VZMULJ(T3, T12);
		    {
			 V TV, TY, TP, TR;
			 TV = LD(&(x[WS(rs, 14)]), ms, &(x[0]));
			 TW = VZMULJ(TU, TV);
			 TY = LD(&(x[WS(rs, 19)]), ms, &(x[WS(rs, 1)]));
			 TZ = VZMULJ(TX, TY);
			 T10 = VADD(TW, TZ);
			 TP = LD(&(x[WS(rs, 9)]), ms, &(x[WS(rs, 1)]));
			 TQ = VZMULJ(T4, TP);
			 TR = LD(&(x[WS(rs, 24)]), ms, &(x[0]));
			 TS = VZMULJ(T8, TR);
			 TT = VADD(TQ, TS);
		    }
		    {
			 V T17, T18, T11, T14, T15;
			 T17 = VSUB(TQ, TS);
			 T18 = VSUB(TW, TZ);
			 T19 = VFMA(LDK(KP475528258), T17, VMUL(LDK(KP293892626), T18));
			 T2D = VFNMS(LDK(KP293892626), T17, VMUL(LDK(KP475528258), T18));
			 T11 = VMUL(LDK(KP559016994), VSUB(TT, T10));
			 T14 = VADD(TT, T10);
			 T15 = VFNMS(LDK(KP250000000), T14, T13);
			 T16 = VADD(T11, T15);
			 T3x = VADD(T13, T14);
			 T2E = VSUB(T15, T11);
		    }
	       }
	       {
		    V T1s, T1l, T1o, T1p, T1e, T1h, T1i, T1r;
		    T1r = LD(&(x[WS(rs, 2)]), ms, &(x[0]));
		    T1s = VZMULJ(Tw, T1r);
		    {
			 V T1k, T1n, T1d, T1g;
			 T1k = LD(&(x[WS(rs, 12)]), ms, &(x[0]));
			 T1l = VZMULJ(T1j, T1k);
			 T1n = LD(&(x[WS(rs, 17)]), ms, &(x[WS(rs, 1)]));
			 T1o = VZMULJ(T1m, T1n);
			 T1p = VADD(T1l, T1o);
			 T1d = LD(&(x[WS(rs, 7)]), ms, &(x[WS(rs, 1)]));
			 T1e = VZMULJ(T1c, T1d);
			 T1g = LD(&(x[WS(rs, 22)]), ms, &(x[0]));
			 T1h = VZMULJ(T1f, T1g);
			 T1i = VADD(T1e, T1h);
		    }
		    {
			 V T1w, T1x, T1q, T1t, T1u;
			 T1w = VSUB(T1e, T1h);
			 T1x = VSUB(T1l, T1o);
			 T1y = VFMA(LDK(KP475528258), T1w, VMUL(LDK(KP293892626), T1x));
			 T2N = VFNMS(LDK(KP293892626), T1w, VMUL(LDK(KP475528258), T1x));
			 T1q = VMUL(LDK(KP559016994), VSUB(T1i, T1p));
			 T1t = VADD(T1i, T1p);
			 T1u = VFNMS(LDK(KP250000000), T1t, T1s);
			 T1v = VADD(T1q, T1u);
			 T3z = VADD(T1s, T1t);
			 T2O = VSUB(T1u, T1q);
		    }
	       }
	       {
		    V T3J, T3K, T3D, T3E, T3C, T3F, T3L, T3G;
		    {
			 V T3H, T3I, T3y, T3B;
			 T3H = VSUB(T3w, T3x);
			 T3I = VSUB(T3z, T3A);
			 T3J = VBYI(VFMA(LDK(KP951056516), T3H, VMUL(LDK(KP587785252), T3I)));
			 T3K = VBYI(VFNMS(LDK(KP587785252), T3H, VMUL(LDK(KP951056516), T3I)));
			 T3D = VADD(Tl, Tm);
			 T3y = VADD(T3w, T3x);
			 T3B = VADD(T3z, T3A);
			 T3E = VADD(T3y, T3B);
			 T3C = VMUL(LDK(KP559016994), VSUB(T3y, T3B));
			 T3F = VFNMS(LDK(KP250000000), T3E, T3D);
		    }
		    ST(&(x[0]), VADD(T3D, T3E), ms, &(x[0]));
		    T3L = VSUB(T3F, T3C);
		    ST(&(x[WS(rs, 10)]), VADD(T3K, T3L), ms, &(x[0]));
		    ST(&(x[WS(rs, 15)]), VSUB(T3L, T3K), ms, &(x[WS(rs, 1)]));
		    T3G = VADD(T3C, T3F);
		    ST(&(x[WS(rs, 5)]), VSUB(T3G, T3J), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 20)]), VADD(T3J, T3G), ms, &(x[0]));
	       }
	       {
		    V To, T2n, T2o, T2p, T2x, T2y, T2z, T2u, T2v, T2w, T2q, T2r, T2s, T29, T2i;
		    V T2e, T2g, T1Y, T2j, T2b, T2c, T2B, T2C;
		    To = VADD(Tk, Tn);
		    T2n = VFMA(LDK(KP1_688655851), TN, VMUL(LDK(KP535826794), TK));
		    T2o = VFMA(LDK(KP1_541026485), T19, VMUL(LDK(KP637423989), T16));
		    T2p = VSUB(T2n, T2o);
		    T2x = VFMA(LDK(KP851558583), T1y, VMUL(LDK(KP904827052), T1v));
		    T2y = VFMA(LDK(KP1_984229402), T1V, VMUL(LDK(KP125333233), T1S));
		    T2z = VADD(T2x, T2y);
		    T2u = VFNMS(LDK(KP844327925), TK, VMUL(LDK(KP1_071653589), TN));
		    T2v = VFNMS(LDK(KP1_274847979), T19, VMUL(LDK(KP770513242), T16));
		    T2w = VADD(T2u, T2v);
		    T2q = VFNMS(LDK(KP425779291), T1v, VMUL(LDK(KP1_809654104), T1y));
		    T2r = VFNMS(LDK(KP992114701), T1S, VMUL(LDK(KP250666467), T1V));
		    T2s = VADD(T2q, T2r);
		    {
			 V T23, T24, T25, T26, T27, T28;
			 T23 = VFMA(LDK(KP1_937166322), TN, VMUL(LDK(KP248689887), TK));
			 T24 = VFMA(LDK(KP1_071653589), T19, VMUL(LDK(KP844327925), T16));
			 T25 = VADD(T23, T24);
			 T26 = VFMA(LDK(KP1_752613360), T1y, VMUL(LDK(KP481753674), T1v));
			 T27 = VFMA(LDK(KP1_457937254), T1V, VMUL(LDK(KP684547105), T1S));
			 T28 = VADD(T26, T27);
			 T29 = VADD(T25, T28);
			 T2i = VSUB(T27, T26);
			 T2e = VMUL(LDK(KP559016994), VSUB(T28, T25));
			 T2g = VSUB(T24, T23);
		    }
		    {
			 V TO, T1a, T1b, T1z, T1W, T1X;
			 TO = VFNMS(LDK(KP497379774), TN, VMUL(LDK(KP968583161), TK));
			 T1a = VFNMS(LDK(KP1_688655851), T19, VMUL(LDK(KP535826794), T16));
			 T1b = VADD(TO, T1a);
			 T1z = VFNMS(LDK(KP963507348), T1y, VMUL(LDK(KP876306680), T1v));
			 T1W = VFNMS(LDK(KP1_369094211), T1V, VMUL(LDK(KP728968627), T1S));
			 T1X = VADD(T1z, T1W);
			 T1Y = VADD(T1b, T1X);
			 T2j = VMUL(LDK(KP559016994), VSUB(T1b, T1X));
			 T2b = VSUB(T1a, TO);
			 T2c = VSUB(T1z, T1W);
		    }
		    {
			 V T1Z, T2a, T2t, T2A;
			 T1Z = VADD(To, T1Y);
			 T2a = VBYI(VADD(T22, T29));
			 ST(&(x[WS(rs, 1)]), VSUB(T1Z, T2a), ms, &(x[WS(rs, 1)]));
			 ST(&(x[WS(rs, 24)]), VADD(T1Z, T2a), ms, &(x[0]));
			 T2t = VADD(To, VADD(T2p, T2s));
			 T2A = VBYI(VADD(T22, VSUB(T2w, T2z)));
			 ST(&(x[WS(rs, 21)]), VSUB(T2t, T2A), ms, &(x[WS(rs, 1)]));
			 ST(&(x[WS(rs, 4)]), VADD(T2t, T2A), ms, &(x[0]));
		    }
		    T2B = VBYI(VADD(T22, VFMA(LDK(KP309016994), T2w, VFMA(LDK(KP587785252), VSUB(T2r, T2q), VFNMS(LDK(KP951056516), VADD(T2n, T2o), VMUL(LDK(KP809016994), T2z))))));
		    T2C = VFMA(LDK(KP309016994), T2p, VFMA(LDK(KP951056516), VSUB(T2u, T2v), VFMA(LDK(KP587785252), VSUB(T2y, T2x), VFNMS(LDK(KP809016994), T2s, To))));
		    ST(&(x[WS(rs, 9)]), VADD(T2B, T2C), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 16)]), VSUB(T2C, T2B), ms, &(x[0]));
		    {
			 V T2f, T2l, T2k, T2m, T2d, T2h;
			 T2d = VFMS(LDK(KP250000000), T29, T22);
			 T2f = VBYI(VADD(VFMA(LDK(KP587785252), T2b, VMUL(LDK(KP951056516), T2c)), VSUB(T2d, T2e)));
			 T2l = VBYI(VADD(VFNMS(LDK(KP587785252), T2c, VMUL(LDK(KP951056516), T2b)), VADD(T2d, T2e)));
			 T2h = VFNMS(LDK(KP250000000), T1Y, To);
			 T2k = VFMA(LDK(KP587785252), T2g, VFNMS(LDK(KP951056516), T2i, VSUB(T2h, T2j)));
			 T2m = VFMA(LDK(KP951056516), T2g, VADD(T2j, VFMA(LDK(KP587785252), T2i, T2h)));
			 ST(&(x[WS(rs, 11)]), VADD(T2f, T2k), ms, &(x[WS(rs, 1)]));
			 ST(&(x[WS(rs, 19)]), VSUB(T2m, T2l), ms, &(x[WS(rs, 1)]));
			 ST(&(x[WS(rs, 14)]), VSUB(T2k, T2f), ms, &(x[0]));
			 ST(&(x[WS(rs, 6)]), VADD(T2l, T2m), ms, &(x[0]));
		    }
	       }
	       {
		    V T2S, T2U, T2F, T2I, T2J, T2Y, T2Z, T30, T2M, T2P, T2Q, T2V, T2W, T2X, T3a;
		    V T3l, T3b, T3k, T3f, T3p, T3i, T3o, T32, T33;
		    T2S = VFNMS(LDK(KP587785252), T20, T2R);
		    T2U = VSUB(Tn, Tk);
		    T2F = VFNMS(LDK(KP125333233), T2E, VMUL(LDK(KP1_984229402), T2D));
		    T2I = VFMA(LDK(KP1_457937254), T2G, VMUL(LDK(KP684547105), T2H));
		    T2J = VSUB(T2F, T2I);
		    T2Y = VFNMS(LDK(KP1_996053456), T2N, VMUL(LDK(KP062790519), T2O));
		    T2Z = VFMA(LDK(KP1_541026485), T2K, VMUL(LDK(KP637423989), T2L));
		    T30 = VSUB(T2Y, T2Z);
		    T2M = VFNMS(LDK(KP770513242), T2L, VMUL(LDK(KP1_274847979), T2K));
		    T2P = VFMA(LDK(KP125581039), T2N, VMUL(LDK(KP998026728), T2O));
		    T2Q = VSUB(T2M, T2P);
		    T2V = VFNMS(LDK(KP1_369094211), T2G, VMUL(LDK(KP728968627), T2H));
		    T2W = VFMA(LDK(KP250666467), T2D, VMUL(LDK(KP992114701), T2E));
		    T2X = VSUB(T2V, T2W);
		    {
			 V T34, T35, T36, T37, T38, T39;
			 T34 = VFNMS(LDK(KP481753674), T2H, VMUL(LDK(KP1_752613360), T2G));
			 T35 = VFMA(LDK(KP851558583), T2D, VMUL(LDK(KP904827052), T2E));
			 T36 = VSUB(T34, T35);
			 T37 = VFNMS(LDK(KP844327925), T2O, VMUL(LDK(KP1_071653589), T2N));
			 T38 = VFNMS(LDK(KP998026728), T2L, VMUL(LDK(KP125581039), T2K));
			 T39 = VADD(T37, T38);
			 T3a = VMUL(LDK(KP559016994), VSUB(T36, T39));
			 T3l = VSUB(T37, T38);
			 T3b = VADD(T36, T39);
			 T3k = VADD(T34, T35);
		    }
		    {
			 V T3d, T3e, T3m, T3g, T3h, T3n;
			 T3d = VFNMS(LDK(KP425779291), T2E, VMUL(LDK(KP1_809654104), T2D));
			 T3e = VFMA(LDK(KP963507348), T2G, VMUL(LDK(KP876306680), T2H));
			 T3m = VADD(T3e, T3d);
			 T3g = VFMA(LDK(KP1_688655851), T2N, VMUL(LDK(KP535826794), T2O));
			 T3h = VFMA(LDK(KP1_996053456), T2K, VMUL(LDK(KP062790519), T2L));
			 T3n = VADD(T3g, T3h);
			 T3f = VSUB(T3d, T3e);
			 T3p = VADD(T3m, T3n);
			 T3i = VSUB(T3g, T3h);
			 T3o = VMUL(LDK(KP559016994), VSUB(T3m, T3n));
		    }
		    {
			 V T3u, T3v, T2T, T31;
			 T3u = VBYI(VADD(T2S, T3b));
			 T3v = VADD(T2U, T3p);
			 ST(&(x[WS(rs, 2)]), VADD(T3u, T3v), ms, &(x[0]));
			 ST(&(x[WS(rs, 23)]), VSUB(T3v, T3u), ms, &(x[WS(rs, 1)]));
			 T2T = VBYI(VSUB(VADD(T2J, T2Q), T2S));
			 T31 = VADD(T2U, VADD(T2X, T30));
			 ST(&(x[WS(rs, 3)]), VADD(T2T, T31), ms, &(x[WS(rs, 1)]));
			 ST(&(x[WS(rs, 22)]), VSUB(T31, T2T), ms, &(x[0]));
		    }
		    T32 = VFMA(LDK(KP309016994), T2X, VFNMS(LDK(KP809016994), T30, VFNMS(LDK(KP587785252), VADD(T2P, T2M), VFNMS(LDK(KP951056516), VADD(T2I, T2F), T2U))));
		    T33 = VBYI(VSUB(VFNMS(LDK(KP587785252), VADD(T2Y, T2Z), VFNMS(LDK(KP809016994), T2Q, VFNMS(LDK(KP951056516), VADD(T2V, T2W), VMUL(LDK(KP309016994), T2J)))), T2S));
		    ST(&(x[WS(rs, 17)]), VSUB(T32, T33), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 8)]), VADD(T32, T33), ms, &(x[0]));
		    {
			 V T3j, T3s, T3r, T3t, T3c, T3q;
			 T3c = VFNMS(LDK(KP250000000), T3b, T2S);
			 T3j = VBYI(VADD(T3a, VADD(T3c, VFNMS(LDK(KP587785252), T3i, VMUL(LDK(KP951056516), T3f)))));
			 T3s = VBYI(VADD(T3c, VSUB(VFMA(LDK(KP587785252), T3f, VMUL(LDK(KP951056516), T3i)), T3a)));
			 T3q = VFNMS(LDK(KP250000000), T3p, T2U);
			 T3r = VFMA(LDK(KP951056516), T3k, VFMA(LDK(KP587785252), T3l, VADD(T3o, T3q)));
			 T3t = VFMA(LDK(KP587785252), T3k, VSUB(VFNMS(LDK(KP951056516), T3l, T3q), T3o));
			 ST(&(x[WS(rs, 7)]), VADD(T3j, T3r), ms, &(x[WS(rs, 1)]));
			 ST(&(x[WS(rs, 13)]), VSUB(T3t, T3s), ms, &(x[WS(rs, 1)]));
			 ST(&(x[WS(rs, 18)]), VSUB(T3r, T3j), ms, &(x[0]));
			 ST(&(x[WS(rs, 12)]), VADD(T3s, T3t), 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, "t3fv_25", twinstr, &GENUS, {190, 150, 78, 0}, 0, 0, 0 };

void X(codelet_t3fv_25) (planner *p) {
     X(kdft_dit_register) (p, t3fv_25, &desc);
}
#endif				/* HAVE_FMA */