t2fv_64.c

快速fft变换

		    T1c = BYTWJ(&(W[TWVL * 82]), T1b);
		    T1d = VSUB(T1a, T1c);
		    T5D = VADD(T1a, T1c);
	       }
	       T7E = VADD(T5z, T5A);
	       T7F = VADD(T5D, T5C);
	       T7V = VSUB(T7E, T7F);
	       T5B = VSUB(T5z, T5A);
	       T5E = VSUB(T5C, T5D);
	       T5F = VFMA(LDK(KP923879532), T5B, VMUL(LDK(KP382683432), T5E));
	       T6u = VFNMS(LDK(KP382683432), T5B, VMUL(LDK(KP923879532), T5E));
	       {
		    V T1e, T1j, T1p, T1q;
		    T1e = VMUL(LDK(KP707106781), VSUB(T18, T1d));
		    T1j = VSUB(T1g, T1i);
		    T1k = VSUB(T1e, T1j);
		    T4e = VADD(T1j, T1e);
		    T1p = VSUB(T1m, T1o);
		    T1q = VMUL(LDK(KP707106781), VADD(T1d, T18));
		    T1r = VADD(T1p, T1q);
		    T4d = VSUB(T1p, T1q);
	       }
	  }
	  {
	       V TG, TI, T5G, TY, T10, T5H, TO, T5K, TT, T5J, T5I, T5L;
	       {
		    V TF, TH, TX, TZ;
		    TF = LD(&(x[WS(rs, 62)]), ms, &(x[0]));
		    TG = BYTWJ(&(W[TWVL * 122]), TF);
		    TH = LD(&(x[WS(rs, 30)]), ms, &(x[0]));
		    TI = BYTWJ(&(W[TWVL * 58]), TH);
		    T5G = VADD(TG, TI);
		    TX = LD(&(x[WS(rs, 14)]), ms, &(x[0]));
		    TY = BYTWJ(&(W[TWVL * 26]), TX);
		    TZ = LD(&(x[WS(rs, 46)]), ms, &(x[0]));
		    T10 = BYTWJ(&(W[TWVL * 90]), TZ);
		    T5H = VADD(TY, T10);
	       }
	       {
		    V TL, TN, TK, TM;
		    TK = LD(&(x[WS(rs, 6)]), ms, &(x[0]));
		    TL = BYTWJ(&(W[TWVL * 10]), TK);
		    TM = LD(&(x[WS(rs, 38)]), ms, &(x[0]));
		    TN = BYTWJ(&(W[TWVL * 74]), TM);
		    TO = VSUB(TL, TN);
		    T5K = VADD(TL, TN);
	       }
	       {
		    V TQ, TS, TP, TR;
		    TP = LD(&(x[WS(rs, 54)]), ms, &(x[0]));
		    TQ = BYTWJ(&(W[TWVL * 106]), TP);
		    TR = LD(&(x[WS(rs, 22)]), ms, &(x[0]));
		    TS = BYTWJ(&(W[TWVL * 42]), TR);
		    TT = VSUB(TQ, TS);
		    T5J = VADD(TQ, TS);
	       }
	       T7B = VADD(T5G, T5H);
	       T7C = VADD(T5K, T5J);
	       T7W = VSUB(T7B, T7C);
	       T5I = VSUB(T5G, T5H);
	       T5L = VSUB(T5J, T5K);
	       T5M = VFNMS(LDK(KP382683432), T5L, VMUL(LDK(KP923879532), T5I));
	       T6v = VFMA(LDK(KP382683432), T5I, VMUL(LDK(KP923879532), T5L));
	       {
		    V TJ, TU, TW, T11;
		    TJ = VSUB(TG, TI);
		    TU = VMUL(LDK(KP707106781), VADD(TO, TT));
		    TV = VADD(TJ, TU);
		    T4g = VSUB(TJ, TU);
		    TW = VMUL(LDK(KP707106781), VSUB(TT, TO));
		    T11 = VSUB(TY, T10);
		    T12 = VSUB(TW, T11);
		    T4h = VADD(T11, TW);
	       }
	  }
	  {
	       V Tl, T5r, TB, T5v, Tq, T5s, Tw, T5u, Tr, TC;
	       {
		    V Ti, Tk, Th, Tj;
		    Th = LD(&(x[WS(rs, 4)]), ms, &(x[0]));
		    Ti = BYTWJ(&(W[TWVL * 6]), Th);
		    Tj = LD(&(x[WS(rs, 36)]), ms, &(x[0]));
		    Tk = BYTWJ(&(W[TWVL * 70]), Tj);
		    Tl = VSUB(Ti, Tk);
		    T5r = VADD(Ti, Tk);
	       }
	       {
		    V Ty, TA, Tx, Tz;
		    Tx = LD(&(x[WS(rs, 12)]), ms, &(x[0]));
		    Ty = BYTWJ(&(W[TWVL * 22]), Tx);
		    Tz = LD(&(x[WS(rs, 44)]), ms, &(x[0]));
		    TA = BYTWJ(&(W[TWVL * 86]), Tz);
		    TB = VSUB(Ty, TA);
		    T5v = VADD(Ty, TA);
	       }
	       {
		    V Tn, Tp, Tm, To;
		    Tm = LD(&(x[WS(rs, 20)]), ms, &(x[0]));
		    Tn = BYTWJ(&(W[TWVL * 38]), Tm);
		    To = LD(&(x[WS(rs, 52)]), ms, &(x[0]));
		    Tp = BYTWJ(&(W[TWVL * 102]), To);
		    Tq = VSUB(Tn, Tp);
		    T5s = VADD(Tn, Tp);
	       }
	       {
		    V Tt, Tv, Ts, Tu;
		    Ts = LD(&(x[WS(rs, 60)]), ms, &(x[0]));
		    Tt = BYTWJ(&(W[TWVL * 118]), Ts);
		    Tu = LD(&(x[WS(rs, 28)]), ms, &(x[0]));
		    Tv = BYTWJ(&(W[TWVL * 54]), Tu);
		    Tw = VSUB(Tt, Tv);
		    T5u = VADD(Tt, Tv);
	       }
	       T7h = VADD(T5r, T5s);
	       T7i = VADD(T5u, T5v);
	       Tr = VFNMS(LDK(KP382683432), Tq, VMUL(LDK(KP923879532), Tl));
	       TC = VFMA(LDK(KP923879532), Tw, VMUL(LDK(KP382683432), TB));
	       TD = VADD(Tr, TC);
	       T4C = VSUB(TC, Tr);
	       {
		    V T3f, T3g, T5t, T5w;
		    T3f = VFNMS(LDK(KP923879532), TB, VMUL(LDK(KP382683432), Tw));
		    T3g = VFMA(LDK(KP382683432), Tl, VMUL(LDK(KP923879532), Tq));
		    T3h = VSUB(T3f, T3g);
		    T4b = VADD(T3g, T3f);
		    T5t = VSUB(T5r, T5s);
		    T5w = VSUB(T5u, T5v);
		    T5x = VMUL(LDK(KP707106781), VADD(T5t, T5w));
		    T6s = VMUL(LDK(KP707106781), VSUB(T5w, T5t));
	       }
	  }
	  {
	       V T1z, T5V, T1P, T5T, T1E, T5W, T1K, T5S;
	       {
		    V T1w, T1y, T1v, T1x;
		    T1v = LD(&(x[WS(rs, 61)]), ms, &(x[WS(rs, 1)]));
		    T1w = BYTWJ(&(W[TWVL * 120]), T1v);
		    T1x = LD(&(x[WS(rs, 29)]), ms, &(x[WS(rs, 1)]));
		    T1y = BYTWJ(&(W[TWVL * 56]), T1x);
		    T1z = VSUB(T1w, T1y);
		    T5V = VADD(T1w, T1y);
	       }
	       {
		    V T1M, T1O, T1L, T1N;
		    T1L = LD(&(x[WS(rs, 21)]), ms, &(x[WS(rs, 1)]));
		    T1M = BYTWJ(&(W[TWVL * 40]), T1L);
		    T1N = LD(&(x[WS(rs, 53)]), ms, &(x[WS(rs, 1)]));
		    T1O = BYTWJ(&(W[TWVL * 104]), T1N);
		    T1P = VSUB(T1M, T1O);
		    T5T = VADD(T1M, T1O);
	       }
	       {
		    V T1B, T1D, T1A, T1C;
		    T1A = LD(&(x[WS(rs, 13)]), ms, &(x[WS(rs, 1)]));
		    T1B = BYTWJ(&(W[TWVL * 24]), T1A);
		    T1C = LD(&(x[WS(rs, 45)]), ms, &(x[WS(rs, 1)]));
		    T1D = BYTWJ(&(W[TWVL * 88]), T1C);
		    T1E = VSUB(T1B, T1D);
		    T5W = VADD(T1B, T1D);
	       }
	       {
		    V T1H, T1J, T1G, T1I;
		    T1G = LD(&(x[WS(rs, 5)]), ms, &(x[WS(rs, 1)]));
		    T1H = BYTWJ(&(W[TWVL * 8]), T1G);
		    T1I = LD(&(x[WS(rs, 37)]), ms, &(x[WS(rs, 1)]));
		    T1J = BYTWJ(&(W[TWVL * 72]), T1I);
		    T1K = VSUB(T1H, T1J);
		    T5S = VADD(T1H, T1J);
	       }
	       {
		    V T1F, T1Q, T7o, T7p;
		    T1F = VFNMS(LDK(KP923879532), T1E, VMUL(LDK(KP382683432), T1z));
		    T1Q = VFMA(LDK(KP382683432), T1K, VMUL(LDK(KP923879532), T1P));
		    T1R = VSUB(T1F, T1Q);
		    T4m = VADD(T1Q, T1F);
		    T7o = VADD(T5S, T5T);
		    T7p = VADD(T5V, T5W);
		    T7q = VADD(T7o, T7p);
		    T80 = VSUB(T7p, T7o);
	       }
	       {
		    V T2h, T2i, T5U, T5X;
		    T2h = VFNMS(LDK(KP382683432), T1P, VMUL(LDK(KP923879532), T1K));
		    T2i = VFMA(LDK(KP923879532), T1z, VMUL(LDK(KP382683432), T1E));
		    T2j = VADD(T2h, T2i);
		    T4p = VSUB(T2i, T2h);
		    T5U = VSUB(T5S, T5T);
		    T5X = VSUB(T5V, T5W);
		    T5Y = VMUL(LDK(KP707106781), VADD(T5U, T5X));
		    T63 = VMUL(LDK(KP707106781), VSUB(T5X, T5U));
	       }
	  }
	  {
	       V T2H, T69, T2X, T6d, T2M, T6a, T2S, T6c;
	       {
		    V T2E, T2G, T2D, T2F;
		    T2D = LD(&(x[WS(rs, 3)]), ms, &(x[WS(rs, 1)]));
		    T2E = BYTWJ(&(W[TWVL * 4]), T2D);
		    T2F = LD(&(x[WS(rs, 35)]), ms, &(x[WS(rs, 1)]));
		    T2G = BYTWJ(&(W[TWVL * 68]), T2F);
		    T2H = VSUB(T2E, T2G);
		    T69 = VADD(T2E, T2G);
	       }
	       {
		    V T2U, T2W, T2T, T2V;
		    T2T = LD(&(x[WS(rs, 11)]), ms, &(x[WS(rs, 1)]));
		    T2U = BYTWJ(&(W[TWVL * 20]), T2T);
		    T2V = LD(&(x[WS(rs, 43)]), ms, &(x[WS(rs, 1)]));
		    T2W = BYTWJ(&(W[TWVL * 84]), T2V);
		    T2X = VSUB(T2U, T2W);
		    T6d = VADD(T2U, T2W);
	       }
	       {
		    V T2J, T2L, T2I, T2K;
		    T2I = LD(&(x[WS(rs, 19)]), ms, &(x[WS(rs, 1)]));
		    T2J = BYTWJ(&(W[TWVL * 36]), T2I);
		    T2K = LD(&(x[WS(rs, 51)]), ms, &(x[WS(rs, 1)]));
		    T2L = BYTWJ(&(W[TWVL * 100]), T2K);
		    T2M = VSUB(T2J, T2L);
		    T6a = VADD(T2J, T2L);
	       }
	       {
		    V T2P, T2R, T2O, T2Q;
		    T2O = LD(&(x[WS(rs, 59)]), ms, &(x[WS(rs, 1)]));
		    T2P = BYTWJ(&(W[TWVL * 116]), T2O);
		    T2Q = LD(&(x[WS(rs, 27)]), ms, &(x[WS(rs, 1)]));
		    T2R = BYTWJ(&(W[TWVL * 52]), T2Q);
		    T2S = VSUB(T2P, T2R);
		    T6c = VADD(T2P, T2R);
	       }
	       {
		    V T2N, T2Y, T7v, T7w;
		    T2N = VFNMS(LDK(KP382683432), T2M, VMUL(LDK(KP923879532), T2H));
		    T2Y = VFMA(LDK(KP923879532), T2S, VMUL(LDK(KP382683432), T2X));
		    T2Z = VADD(T2N, T2Y);
		    T4w = VSUB(T2Y, T2N);
		    T7v = VADD(T69, T6a);
		    T7w = VADD(T6c, T6d);
		    T7x = VADD(T7v, T7w);
		    T83 = VSUB(T7w, T7v);
	       }
	       {
		    V T31, T32, T6b, T6e;
		    T31 = VFNMS(LDK(KP923879532), T2X, VMUL(LDK(KP382683432), T2S));
		    T32 = VFMA(LDK(KP382683432), T2H, VMUL(LDK(KP923879532), T2M));
		    T33 = VSUB(T31, T32);
		    T4t = VADD(T32, T31);
		    T6b = VSUB(T69, T6a);
		    T6e = VSUB(T6c, T6d);
		    T6f = VMUL(LDK(KP707106781), VADD(T6b, T6e));
		    T6k = VMUL(LDK(KP707106781), VSUB(T6e, T6b));
	       }
	  }
	  {
	       V T7k, T7M, T7R, T7T, T7z, T7I, T7H, T7N, T7O, T7S;
	       {
		    V T7g, T7j, T7P, T7Q;
		    T7g = VADD(T7e, T7f);
		    T7j = VADD(T7h, T7i);
		    T7k = VSUB(T7g, T7j);
		    T7M = VADD(T7g, T7j);
		    T7P = VADD(T7n, T7q);
		    T7Q = VADD(T7u, T7x);
		    T7R = VADD(T7P, T7Q);
		    T7T = VBYI(VSUB(T7Q, T7P));
	       }
	       {
		    V T7r, T7y, T7D, T7G;
		    T7r = VSUB(T7n, T7q);
		    T7y = VSUB(T7u, T7x);
		    T7z = VMUL(LDK(KP707106781), VADD(T7r, T7y));
		    T7I = VMUL(LDK(KP707106781), VSUB(T7y, T7r));
		    T7D = VADD(T7B, T7C);
		    T7G = VADD(T7E, T7F);
		    T7H = VSUB(T7D, T7G);
		    T7N = VADD(T7G, T7D);
	       }
	       T7O = VADD(T7M, T7N);
	       ST(&(x[WS(rs, 32)]), VSUB(T7O, T7R), ms, &(x[0]));
	       ST(&(x[0]), VADD(T7O, T7R), ms, &(x[0]));
	       T7S = VSUB(T7M, T7N);
	       ST(&(x[WS(rs, 48)]), VSUB(T7S, T7T), ms, &(x[0]));
	       ST(&(x[WS(rs, 16)]), VADD(T7S, T7T), ms, &(x[0]));
	       {
		    V T7A, T7J, T7K, T7L;
		    T7A = VADD(T7k, T7z);
		    T7J = VBYI(VADD(T7H, T7I));
		    ST(&(x[WS(rs, 56)]), VSUB(T7A, T7J), ms, &(x[0]));
		    ST(&(x[WS(rs, 8)]), VADD(T7A, T7J), ms, &(x[0]));
		    T7K = VSUB(T7k, T7z);
		    T7L = VBYI(VSUB(T7I, T7H));
		    ST(&(x[WS(rs, 40)]), VSUB(T7K, T7L), ms, &(x[0]));
		    ST(&(x[WS(rs, 24)]), VADD(T7K, T7L), ms, &(x[0]));
	       }
	  }
	  {
	       V T7Y, T8j, T8c, T8k, T85, T8g, T89, T8h;
	       {
		    V T7U, T7X, T8a, T8b;
		    T7U = VSUB(T7e, T7f);
		    T7X = VMUL(LDK(KP707106781), VADD(T7V, T7W));
		    T7Y = VADD(T7U, T7X);
		    T8j = VSUB(T7U, T7X);
		    T8a = VFNMS(LDK(KP382683432), T7Z, VMUL(LDK(KP923879532), T80));
		    T8b = VFMA(LDK(KP382683432), T82, VMUL(LDK(KP923879532), T83));
		    T8c = VADD(T8a, T8b);
		    T8k = VSUB(T8b, T8a);
	       }
	       {
		    V T81, T84, T87, T88;
		    T81 = VFMA(LDK(KP923879532), T7Z, VMUL(LDK(KP382683432), T80));
		    T84 = VFNMS(LDK(KP382683432), T83, VMUL(LDK(KP923879532), T82));
		    T85 = VADD(T81, T84);
		    T8g = VSUB(T84, T81);
		    T87 = VSUB(T7i, T7h);
		    T88 = VMUL(LDK(KP707106781), VSUB(T7W, T7V));
		    T89 = VADD(T87, T88);
		    T8h = VSUB(T88, T87);
	       }
	       {
		    V T86, T8d, T8m, T8n;
		    T86 = VADD(T7Y, T85);
		    T8d = VBYI(VADD(T89, T8c));
		    ST(&(x[WS(rs, 60)]), VSUB(T86, T8d), ms, &(x[0]));
		    ST(&(x[WS(rs, 4)]), VADD(T86, T8d), ms, &(x[0]));
		    T8m = VBYI(VADD(T8h, T8g));
		    T8n = VADD(T8j, T8k);
		    ST(&(x[WS(rs, 12)]), VADD(T8m, T8n), ms, &(x[0]));
		    ST(&(x[WS(rs, 52)]), VSUB(T8n, T8m), ms, &(x[0]));
	       }
	       {
		    V T8e, T8f, T8i, T8l;
		    T8e = VSUB(T7Y, T85);
		    T8f = VBYI(VSUB(T8c, T89));
		    ST(&(x[WS(rs, 36)]), VSUB(T8e, T8f), ms, &(x[0]));
		    ST(&(x[WS(rs, 28)]), VADD(T8e, T8f), ms, &(x[0]));
		    T8i = VBYI(VSUB(T8g, T8h));
		    T8l = VSUB(T8j, T8k);
		    ST(&(x[WS(rs, 20)]), VADD(T8i, T8l), ms, &(x[0]));
		    ST(&(x[WS(rs, 44)]), VSUB(T8l, T8i), ms, &(x[0]));
	       }
	  }
	  {
	       V T5O, T6H, T6x, T6F, T6n, T6I, T6A, T6E;
	       {
		    V T5y, T5N, T6t, T6w;
		    T5y = VADD(T5q, T5x);
		    T5N = VADD(T5F, T5M);
		    T5O = VADD(T5y, T5N);
		    T6H = VSUB(T5y, T5N);
		    T6t = VADD(T6r, T6s);
		    T6w = VADD(T6u, T6v);
		    T6x = VADD(T6t, T6w);
		    T6F = VSUB(T6w, T6t);
		    {
			 V T65, T6y, T6m, T6z;
			 {
			      V T5Z, T64, T6g, T6l;
			      T5Z = VADD(T5R, T5Y);
			      T64 = VADD(T62, T63);
			      T65 = VFMA(LDK(KP980785280), T5Z, VMUL(LDK(KP195090322), T64));
			      T6y = VFNMS(LDK(KP195090322), T5Z, VMUL(LDK(KP980785280), T64));
			      T6g = VADD(T68, T6f);
			      T6l = VADD(T6j, T6k);
			      T6m = VFNMS(LDK(KP195090322), T6l, VMUL(LDK(KP980785280), T6g));
			      T6z = VFMA(LDK(KP195090322), T6g, VMUL(LDK(KP980785280), T6l));
			 }
			 T6n = VADD(T65, T6m);
			 T6I = VSUB(T6z, T6y);
			 T6A = VADD(T6y, T6z);
			 T6E = VSUB(T6m, T65);
		    }
	       }
	       {
		    V T6o, T6B, T6K, T6L;