t1bv_25.c

用于DFT计算的c语言的库的最新版本,包含丰富的函数库.

		    T1k = VSUB(T19, T1b);
		    T1l = VFMA(LDK(KP475528258), T1j, VMUL(LDK(KP293892626), T1k));
		    T2v = VFNMS(LDK(KP475528258), T1k, VMUL(LDK(KP293892626), T1j));
		    T1d = VMUL(LDK(KP559016994), VSUB(T17, T1c));
		    T1g = VADD(T17, T1c);
		    T1h = VFNMS(LDK(KP250000000), T1g, T1f);
		    T1i = VADD(T1d, T1h);
		    T3e = VADD(T1f, T1g);
		    T2u = VSUB(T1h, T1d);
	       }
	  }
	  {
	       V Tg, T7, T9, Td, T2, T4, Tc, Tf;
	       Tf = LD(&(x[WS(rs, 4)]), ms, &(x[0]));
	       Tg = BYTW(&(W[TWVL * 6]), Tf);
	       {
		    V T6, T8, T1, T3;
		    T6 = LD(&(x[WS(rs, 14)]), ms, &(x[0]));
		    T7 = BYTW(&(W[TWVL * 26]), T6);
		    T8 = LD(&(x[WS(rs, 19)]), ms, &(x[WS(rs, 1)]));
		    T9 = BYTW(&(W[TWVL * 36]), T8);
		    Td = VADD(T7, T9);
		    T1 = LD(&(x[WS(rs, 9)]), ms, &(x[WS(rs, 1)]));
		    T2 = BYTW(&(W[TWVL * 16]), T1);
		    T3 = LD(&(x[WS(rs, 24)]), ms, &(x[0]));
		    T4 = BYTW(&(W[TWVL * 46]), T3);
		    Tc = VADD(T2, T4);
	       }
	       {
		    V T5, Ta, Te, Th, Ti;
		    T5 = VSUB(T2, T4);
		    Ta = VSUB(T7, T9);
		    Tb = VFMA(LDK(KP475528258), T5, VMUL(LDK(KP293892626), Ta));
		    T2i = VFNMS(LDK(KP475528258), Ta, VMUL(LDK(KP293892626), T5));
		    Te = VMUL(LDK(KP559016994), VSUB(Tc, Td));
		    Th = VADD(Tc, Td);
		    Ti = VFNMS(LDK(KP250000000), Th, Tg);
		    Tj = VADD(Te, Ti);
		    T3b = VADD(Tg, Th);
		    T2h = VSUB(Ti, Te);
	       }
	  }
	  {
	       V TA, Tr, Tt, Tx, Tm, To, Tw, Tz;
	       Tz = LD(&(x[WS(rs, 1)]), ms, &(x[WS(rs, 1)]));
	       TA = BYTW(&(W[0]), Tz);
	       {
		    V Tq, Ts, Tl, Tn;
		    Tq = LD(&(x[WS(rs, 11)]), ms, &(x[WS(rs, 1)]));
		    Tr = BYTW(&(W[TWVL * 20]), Tq);
		    Ts = LD(&(x[WS(rs, 16)]), ms, &(x[0]));
		    Tt = BYTW(&(W[TWVL * 30]), Ts);
		    Tx = VADD(Tr, Tt);
		    Tl = LD(&(x[WS(rs, 6)]), ms, &(x[0]));
		    Tm = BYTW(&(W[TWVL * 10]), Tl);
		    Tn = LD(&(x[WS(rs, 21)]), ms, &(x[WS(rs, 1)]));
		    To = BYTW(&(W[TWVL * 40]), Tn);
		    Tw = VADD(Tm, To);
	       }
	       {
		    V Tp, Tu, Ty, TB, TC;
		    Tp = VSUB(Tm, To);
		    Tu = VSUB(Tr, Tt);
		    Tv = VFMA(LDK(KP475528258), Tp, VMUL(LDK(KP293892626), Tu));
		    T2k = VFNMS(LDK(KP475528258), Tu, VMUL(LDK(KP293892626), Tp));
		    Ty = VMUL(LDK(KP559016994), VSUB(Tw, Tx));
		    TB = VADD(Tw, Tx);
		    TC = VFNMS(LDK(KP250000000), TB, TA);
		    TD = VADD(Ty, TC);
		    T3a = VADD(TA, TB);
		    T2l = VSUB(TC, Ty);
	       }
	  }
	  {
	       V TV, TP, TR, TS, TK, TM, TN, TU;
	       TU = LD(&(x[WS(rs, 2)]), ms, &(x[0]));
	       TV = BYTW(&(W[TWVL * 2]), TU);
	       {
		    V TO, TQ, TJ, TL;
		    TO = LD(&(x[WS(rs, 12)]), ms, &(x[0]));
		    TP = BYTW(&(W[TWVL * 22]), TO);
		    TQ = LD(&(x[WS(rs, 17)]), ms, &(x[WS(rs, 1)]));
		    TR = BYTW(&(W[TWVL * 32]), TQ);
		    TS = VADD(TP, TR);
		    TJ = LD(&(x[WS(rs, 7)]), ms, &(x[WS(rs, 1)]));
		    TK = BYTW(&(W[TWVL * 12]), TJ);
		    TL = LD(&(x[WS(rs, 22)]), ms, &(x[0]));
		    TM = BYTW(&(W[TWVL * 42]), TL);
		    TN = VADD(TK, TM);
	       }
	       {
		    V TZ, T10, TT, TW, TX;
		    TZ = VSUB(TK, TM);
		    T10 = VSUB(TP, TR);
		    T11 = VFMA(LDK(KP475528258), TZ, VMUL(LDK(KP293892626), T10));
		    T2s = VFNMS(LDK(KP475528258), T10, VMUL(LDK(KP293892626), TZ));
		    TT = VMUL(LDK(KP559016994), VSUB(TN, TS));
		    TW = VADD(TN, TS);
		    TX = VFNMS(LDK(KP250000000), TW, TV);
		    TY = VADD(TT, TX);
		    T3d = VADD(TV, TW);
		    T2r = VSUB(TX, TT);
	       }
	  }
	  {
	       V T3g, T3o, T3k, T3l, T3j, T3m, T3p, T3n;
	       {
		    V T3c, T3f, T3h, T3i;
		    T3c = VSUB(T3a, T3b);
		    T3f = VSUB(T3d, T3e);
		    T3g = VBYI(VFMA(LDK(KP951056516), T3c, VMUL(LDK(KP587785252), T3f)));
		    T3o = VBYI(VFNMS(LDK(KP951056516), T3f, VMUL(LDK(KP587785252), T3c)));
		    T3k = VADD(T1A, T1B);
		    T3h = VADD(T3a, T3b);
		    T3i = VADD(T3d, T3e);
		    T3l = VADD(T3h, T3i);
		    T3j = VMUL(LDK(KP559016994), VSUB(T3h, T3i));
		    T3m = VFNMS(LDK(KP250000000), T3l, T3k);
	       }
	       ST(&(x[0]), VADD(T3k, T3l), ms, &(x[0]));
	       T3p = VSUB(T3m, T3j);
	       ST(&(x[WS(rs, 10)]), VADD(T3o, T3p), ms, &(x[0]));
	       ST(&(x[WS(rs, 15)]), VSUB(T3p, T3o), ms, &(x[WS(rs, 1)]));
	       T3n = VADD(T3j, T3m);
	       ST(&(x[WS(rs, 5)]), VADD(T3g, T3n), ms, &(x[WS(rs, 1)]));
	       ST(&(x[WS(rs, 20)]), VSUB(T3n, T3g), ms, &(x[0]));
	  }
	  {
	       V T2z, T2M, T2U, T2V, T2W, T34, T35, T36, T2X, T2Y, T2Z, T31, T32, T33, T2n;
	       V T2N, T2E, T2K, T2y, T2H, T2A, T2G, T38, T39;
	       T2z = VSUB(T1C, T1z);
	       T2M = VFNMS(LDK(KP951056516), T1R, T2L);
	       T2U = VFMA(LDK(KP1_369094211), T2k, VMUL(LDK(KP728968627), T2l));
	       T2V = VFNMS(LDK(KP992114701), T2h, VMUL(LDK(KP250666467), T2i));
	       T2W = VADD(T2U, T2V);
	       T34 = VFNMS(LDK(KP125581039), T2s, VMUL(LDK(KP998026728), T2r));
	       T35 = VFMA(LDK(KP1_274847979), T2v, VMUL(LDK(KP770513242), T2u));
	       T36 = VADD(T34, T35);
	       T2X = VFMA(LDK(KP1_996053456), T2s, VMUL(LDK(KP062790519), T2r));
	       T2Y = VFNMS(LDK(KP637423989), T2u, VMUL(LDK(KP1_541026485), T2v));
	       T2Z = VADD(T2X, T2Y);
	       T31 = VFNMS(LDK(KP1_457937254), T2k, VMUL(LDK(KP684547105), T2l));
	       T32 = VFMA(LDK(KP1_984229402), T2i, VMUL(LDK(KP125333233), T2h));
	       T33 = VADD(T31, T32);
	       {
		    V T2j, T2m, T2I, T2C, T2D, T2J;
		    T2j = VFNMS(LDK(KP851558583), T2i, VMUL(LDK(KP904827052), T2h));
		    T2m = VFMA(LDK(KP1_752613360), T2k, VMUL(LDK(KP481753674), T2l));
		    T2I = VADD(T2m, T2j);
		    T2C = VFMA(LDK(KP1_071653589), T2s, VMUL(LDK(KP844327925), T2r));
		    T2D = VFMA(LDK(KP125581039), T2v, VMUL(LDK(KP998026728), T2u));
		    T2J = VADD(T2C, T2D);
		    T2n = VSUB(T2j, T2m);
		    T2N = VADD(T2I, T2J);
		    T2E = VSUB(T2C, T2D);
		    T2K = VMUL(LDK(KP559016994), VSUB(T2I, T2J));
	       }
	       {
		    V T2o, T2p, T2q, T2t, T2w, T2x;
		    T2o = VFNMS(LDK(KP963507348), T2k, VMUL(LDK(KP876306680), T2l));
		    T2p = VFMA(LDK(KP1_809654104), T2i, VMUL(LDK(KP425779291), T2h));
		    T2q = VSUB(T2o, T2p);
		    T2t = VFNMS(LDK(KP1_688655851), T2s, VMUL(LDK(KP535826794), T2r));
		    T2w = VFNMS(LDK(KP1_996053456), T2v, VMUL(LDK(KP062790519), T2u));
		    T2x = VADD(T2t, T2w);
		    T2y = VMUL(LDK(KP559016994), VSUB(T2q, T2x));
		    T2H = VSUB(T2t, T2w);
		    T2A = VADD(T2q, T2x);
		    T2G = VADD(T2o, T2p);
	       }
	       {
		    V T2S, T2T, T30, T37;
		    T2S = VADD(T2z, T2A);
		    T2T = VBYI(VADD(T2M, T2N));
		    ST(&(x[WS(rs, 23)]), VSUB(T2S, T2T), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 2)]), VADD(T2S, T2T), ms, &(x[0]));
		    T30 = VADD(T2z, VADD(T2W, T2Z));
		    T37 = VBYI(VSUB(VADD(T33, T36), T2M));
		    ST(&(x[WS(rs, 22)]), VSUB(T30, T37), ms, &(x[0]));
		    ST(&(x[WS(rs, 3)]), VADD(T30, T37), ms, &(x[WS(rs, 1)]));
	       }
	       T38 = VBYI(VSUB(VFMA(LDK(KP951056516), VSUB(T2U, T2V), VFMA(LDK(KP309016994), T33, VFNMS(LDK(KP809016994), T36, VMUL(LDK(KP587785252), VSUB(T2X, T2Y))))), T2M));
	       T39 = VFMA(LDK(KP309016994), T2W, VFMA(LDK(KP951056516), VSUB(T32, T31), VFMA(LDK(KP587785252), VSUB(T35, T34), VFNMS(LDK(KP809016994), T2Z, T2z))));
	       ST(&(x[WS(rs, 8)]), VADD(T38, T39), ms, &(x[0]));
	       ST(&(x[WS(rs, 17)]), VSUB(T39, T38), ms, &(x[WS(rs, 1)]));
	       {
		    V T2F, T2Q, T2P, T2R, T2B, T2O;
		    T2B = VFNMS(LDK(KP250000000), T2A, T2z);
		    T2F = VFMA(LDK(KP951056516), T2n, VADD(T2y, VFNMS(LDK(KP587785252), T2E, T2B)));
		    T2Q = VFMA(LDK(KP587785252), T2n, VFMA(LDK(KP951056516), T2E, VSUB(T2B, T2y)));
		    T2O = VFNMS(LDK(KP250000000), T2N, T2M);
		    T2P = VBYI(VADD(VFMA(LDK(KP951056516), T2G, VMUL(LDK(KP587785252), T2H)), VADD(T2K, T2O)));
		    T2R = VBYI(VADD(VFNMS(LDK(KP951056516), T2H, VMUL(LDK(KP587785252), T2G)), VSUB(T2O, T2K)));
		    ST(&(x[WS(rs, 18)]), VSUB(T2F, T2P), ms, &(x[0]));
		    ST(&(x[WS(rs, 12)]), VADD(T2Q, T2R), ms, &(x[0]));
		    ST(&(x[WS(rs, 7)]), VADD(T2F, T2P), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 13)]), VSUB(T2Q, T2R), ms, &(x[WS(rs, 1)]));
	       }
	  }
	  {
	       V T1D, T1T, T21, T22, T23, T2b, T2c, T2d, T24, T25, T26, T28, T29, T2a, TF;
	       V T1U, T1I, T1O, T1o, T1L, T1E, T1K, T2f, T2g;
	       T1D = VADD(T1z, T1C);
	       T1T = VADD(T1Q, T1S);
	       T21 = VFMA(LDK(KP1_688655851), Tv, VMUL(LDK(KP535826794), TD));
	       T22 = VFMA(LDK(KP1_541026485), Tb, VMUL(LDK(KP637423989), Tj));
	       T23 = VSUB(T21, T22);
	       T2b = VFMA(LDK(KP851558583), T11, VMUL(LDK(KP904827052), TY));
	       T2c = VFMA(LDK(KP1_984229402), T1l, VMUL(LDK(KP125333233), T1i));
	       T2d = VADD(T2b, T2c);
	       T24 = VFNMS(LDK(KP425779291), TY, VMUL(LDK(KP1_809654104), T11));
	       T25 = VFNMS(LDK(KP992114701), T1i, VMUL(LDK(KP250666467), T1l));
	       T26 = VADD(T24, T25);
	       T28 = VFNMS(LDK(KP1_071653589), Tv, VMUL(LDK(KP844327925), TD));
	       T29 = VFNMS(LDK(KP770513242), Tj, VMUL(LDK(KP1_274847979), Tb));
	       T2a = VADD(T28, T29);
	       {
		    V Tk, TE, T1M, T1G, T1H, T1N;
		    Tk = VFMA(LDK(KP1_071653589), Tb, VMUL(LDK(KP844327925), Tj));
		    TE = VFMA(LDK(KP1_937166322), Tv, VMUL(LDK(KP248689887), TD));
		    T1M = VADD(TE, Tk);
		    T1G = VFMA(LDK(KP1_752613360), T11, VMUL(LDK(KP481753674), TY));
		    T1H = VFMA(LDK(KP1_457937254), T1l, VMUL(LDK(KP684547105), T1i));
		    T1N = VADD(T1G, T1H);
		    TF = VSUB(Tk, TE);
		    T1U = VADD(T1M, T1N);
		    T1I = VSUB(T1G, T1H);
		    T1O = VMUL(LDK(KP559016994), VSUB(T1M, T1N));
	       }
	       {
		    V TG, TH, TI, T12, T1m, T1n;
		    TG = VFNMS(LDK(KP497379774), Tv, VMUL(LDK(KP968583161), TD));
		    TH = VFNMS(LDK(KP1_688655851), Tb, VMUL(LDK(KP535826794), Tj));
		    TI = VADD(TG, TH);
		    T12 = VFNMS(LDK(KP963507348), T11, VMUL(LDK(KP876306680), TY));
		    T1m = VFNMS(LDK(KP1_369094211), T1l, VMUL(LDK(KP728968627), T1i));
		    T1n = VADD(T12, T1m);
		    T1o = VMUL(LDK(KP559016994), VSUB(TI, T1n));
		    T1L = VSUB(T12, T1m);
		    T1E = VADD(TI, T1n);
		    T1K = VSUB(TG, TH);
	       }
	       {
		    V T1Z, T20, T27, T2e;
		    T1Z = VADD(T1D, T1E);
		    T20 = VBYI(VADD(T1T, T1U));
		    ST(&(x[WS(rs, 24)]), VSUB(T1Z, T20), ms, &(x[0]));
		    ST(&(x[WS(rs, 1)]), VADD(T1Z, T20), ms, &(x[WS(rs, 1)]));
		    T27 = VADD(T1D, VADD(T23, T26));
		    T2e = VBYI(VSUB(VADD(T2a, T2d), T1T));
		    ST(&(x[WS(rs, 21)]), VSUB(T27, T2e), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 4)]), VADD(T27, T2e), ms, &(x[0]));
	       }
	       T2f = VBYI(VSUB(VFMA(LDK(KP309016994), T2a, VFMA(LDK(KP951056516), VADD(T21, T22), VFNMS(LDK(KP809016994), T2d, VMUL(LDK(KP587785252), VSUB(T24, T25))))), T1T));
	       T2g = VFMA(LDK(KP951056516), VSUB(T29, T28), VFMA(LDK(KP309016994), T23, VFMA(LDK(KP587785252), VSUB(T2c, T2b), VFNMS(LDK(KP809016994), T26, T1D))));
	       ST(&(x[WS(rs, 9)]), VADD(T2f, T2g), ms, &(x[WS(rs, 1)]));
	       ST(&(x[WS(rs, 16)]), VSUB(T2g, T2f), ms, &(x[0]));
	       {
		    V T1J, T1X, T1W, T1Y, T1F, T1V;
		    T1F = VFNMS(LDK(KP250000000), T1E, T1D);
		    T1J = VFMA(LDK(KP951056516), TF, VADD(T1o, VFNMS(LDK(KP587785252), T1I, T1F)));
		    T1X = VFMA(LDK(KP587785252), TF, VFMA(LDK(KP951056516), T1I, VSUB(T1F, T1o)));
		    T1V = VFNMS(LDK(KP250000000), T1U, T1T);
		    T1W = VBYI(VADD(VFMA(LDK(KP951056516), T1K, VMUL(LDK(KP587785252), T1L)), VADD(T1O, T1V)));
		    T1Y = VBYI(VADD(VFNMS(LDK(KP951056516), T1L, VMUL(LDK(KP587785252), T1K)), VSUB(T1V, T1O)));
		    ST(&(x[WS(rs, 19)]), VSUB(T1J, T1W), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 11)]), VADD(T1X, T1Y), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 6)]), VADD(T1J, T1W), ms, &(x[0]));
		    ST(&(x[WS(rs, 14)]), VSUB(T1X, T1Y), 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, "t1bv_25", twinstr, &GENUS, {171, 111, 77, 0}, 0, 0, 0 };

void X(codelet_t1bv_25) (planner *p) {
     X(kdft_dit_register) (p, t1bv_25, &desc);
}
#endif				/* HAVE_FMA */