n2fv_64.c

快速fft变换

		    TY = VADD(TW, TX);
		    TZ = VFNMS(LDK(KP195090322), TY, VMUL(LDK(KP980785280), TT));
		    T2q = VFMA(LDK(KP980785280), TY, VMUL(LDK(KP195090322), TT));
		    T6D = VADD(T4y, T4z);
		    T6E = VADD(T4C, T4B);
		    T6F = VADD(T6D, T6E);
		    T6U = VSUB(T6D, T6E);
	       }
	       {
		    V T3c, T3d, T4A, T4D;
		    T3c = VSUB(TW, TX);
		    T3d = VADD(TS, TP);
		    T3e = VFMA(LDK(KP831469612), T3c, VMUL(LDK(KP555570233), T3d));
		    T3D = VFNMS(LDK(KP555570233), T3c, VMUL(LDK(KP831469612), T3d));
		    T4A = VSUB(T4y, T4z);
		    T4D = VSUB(T4B, T4C);
		    T4E = VFMA(LDK(KP923879532), T4A, VMUL(LDK(KP382683432), T4D));
		    T5t = VFNMS(LDK(KP382683432), T4A, VMUL(LDK(KP923879532), T4D));
	       }
	  }
	  {
	       V T1F, T55, T2a, T56, T1M, T5h, T27, T5g, T58, T59, T1U, T5a, T25, T5b, T5c;
	       V T21, T5d, T24;
	       {
		    V T1D, T1E, T28, T29;
		    T1D = LD(&(xi[WS(is, 63)]), ivs, &(xi[WS(is, 1)]));
		    T1E = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
		    T1F = VSUB(T1D, T1E);
		    T55 = VADD(T1D, T1E);
		    T28 = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
		    T29 = LD(&(xi[WS(is, 47)]), ivs, &(xi[WS(is, 1)]));
		    T2a = VSUB(T28, T29);
		    T56 = VADD(T28, T29);
	       }
	       {
		    V T1G, T1H, T1I, T1J, T1K, T1L;
		    T1G = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
		    T1H = LD(&(xi[WS(is, 39)]), ivs, &(xi[WS(is, 1)]));
		    T1I = VSUB(T1G, T1H);
		    T1J = LD(&(xi[WS(is, 55)]), ivs, &(xi[WS(is, 1)]));
		    T1K = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
		    T1L = VSUB(T1J, T1K);
		    T1M = VMUL(LDK(KP707106781), VADD(T1I, T1L));
		    T5h = VADD(T1G, T1H);
		    T27 = VMUL(LDK(KP707106781), VSUB(T1L, T1I));
		    T5g = VADD(T1J, T1K);
	       }
	       {
		    V T1Q, T1T, T1X, T20;
		    {
			 V T1O, T1P, T1R, T1S;
			 T1O = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
			 T1P = LD(&(xi[WS(is, 35)]), ivs, &(xi[WS(is, 1)]));
			 T1Q = VSUB(T1O, T1P);
			 T58 = VADD(T1O, T1P);
			 T1R = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
			 T1S = LD(&(xi[WS(is, 51)]), ivs, &(xi[WS(is, 1)]));
			 T1T = VSUB(T1R, T1S);
			 T59 = VADD(T1R, T1S);
		    }
		    T1U = VFNMS(LDK(KP382683432), T1T, VMUL(LDK(KP923879532), T1Q));
		    T5a = VSUB(T58, T59);
		    T25 = VFMA(LDK(KP382683432), T1Q, VMUL(LDK(KP923879532), T1T));
		    {
			 V T1V, T1W, T1Y, T1Z;
			 T1V = LD(&(xi[WS(is, 59)]), ivs, &(xi[WS(is, 1)]));
			 T1W = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
			 T1X = VSUB(T1V, T1W);
			 T5b = VADD(T1V, T1W);
			 T1Y = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
			 T1Z = LD(&(xi[WS(is, 43)]), ivs, &(xi[WS(is, 1)]));
			 T20 = VSUB(T1Y, T1Z);
			 T5c = VADD(T1Y, T1Z);
		    }
		    T21 = VFMA(LDK(KP923879532), T1X, VMUL(LDK(KP382683432), T20));
		    T5d = VSUB(T5b, T5c);
		    T24 = VFNMS(LDK(KP923879532), T20, VMUL(LDK(KP382683432), T1X));
	       }
	       {
		    V T1N, T22, T6r, T6s;
		    T1N = VADD(T1F, T1M);
		    T22 = VADD(T1U, T21);
		    T23 = VSUB(T1N, T22);
		    T2N = VADD(T1N, T22);
		    T6r = VADD(T55, T56);
		    T6s = VADD(T5h, T5g);
		    T6t = VADD(T6r, T6s);
		    T71 = VSUB(T6r, T6s);
	       }
	       {
		    V T6u, T6v, T26, T2b;
		    T6u = VADD(T58, T59);
		    T6v = VADD(T5b, T5c);
		    T6w = VADD(T6u, T6v);
		    T72 = VSUB(T6v, T6u);
		    T26 = VSUB(T24, T25);
		    T2b = VSUB(T27, T2a);
		    T2c = VSUB(T26, T2b);
		    T2O = VADD(T2b, T26);
	       }
	       {
		    V T3r, T3s, T57, T5e;
		    T3r = VSUB(T1F, T1M);
		    T3s = VADD(T25, T24);
		    T3t = VADD(T3r, T3s);
		    T41 = VSUB(T3r, T3s);
		    T57 = VSUB(T55, T56);
		    T5e = VMUL(LDK(KP707106781), VADD(T5a, T5d));
		    T5f = VADD(T57, T5e);
		    T5R = VSUB(T57, T5e);
	       }
	       {
		    V T5i, T5j, T3u, T3v;
		    T5i = VSUB(T5g, T5h);
		    T5j = VMUL(LDK(KP707106781), VSUB(T5d, T5a));
		    T5k = VADD(T5i, T5j);
		    T5S = VSUB(T5j, T5i);
		    T3u = VADD(T2a, T27);
		    T3v = VSUB(T21, T1U);
		    T3w = VADD(T3u, T3v);
		    T42 = VSUB(T3v, T3u);
	       }
	  }
	  {
	       V T1q, T4P, T1v, T4O, T1n, T50, T1w, T4Z, T4U, T4V, T18, T4W, T1z, T4R, T4S;
	       V T1f, T4T, T1y;
	       {
		    V T1o, T1p, T1t, T1u;
		    T1o = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
		    T1p = LD(&(xi[WS(is, 49)]), ivs, &(xi[WS(is, 1)]));
		    T1q = VSUB(T1o, T1p);
		    T4P = VADD(T1o, T1p);
		    T1t = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
		    T1u = LD(&(xi[WS(is, 33)]), ivs, &(xi[WS(is, 1)]));
		    T1v = VSUB(T1t, T1u);
		    T4O = VADD(T1t, T1u);
	       }
	       {
		    V T1h, T1i, T1j, T1k, T1l, T1m;
		    T1h = LD(&(xi[WS(is, 57)]), ivs, &(xi[WS(is, 1)]));
		    T1i = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
		    T1j = VSUB(T1h, T1i);
		    T1k = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
		    T1l = LD(&(xi[WS(is, 41)]), ivs, &(xi[WS(is, 1)]));
		    T1m = VSUB(T1k, T1l);
		    T1n = VMUL(LDK(KP707106781), VSUB(T1j, T1m));
		    T50 = VADD(T1k, T1l);
		    T1w = VMUL(LDK(KP707106781), VADD(T1m, T1j));
		    T4Z = VADD(T1h, T1i);
	       }
	       {
		    V T14, T17, T1b, T1e;
		    {
			 V T12, T13, T15, T16;
			 T12 = LD(&(xi[WS(is, 61)]), ivs, &(xi[WS(is, 1)]));
			 T13 = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
			 T14 = VSUB(T12, T13);
			 T4U = VADD(T12, T13);
			 T15 = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
			 T16 = LD(&(xi[WS(is, 45)]), ivs, &(xi[WS(is, 1)]));
			 T17 = VSUB(T15, T16);
			 T4V = VADD(T15, T16);
		    }
		    T18 = VFNMS(LDK(KP923879532), T17, VMUL(LDK(KP382683432), T14));
		    T4W = VSUB(T4U, T4V);
		    T1z = VFMA(LDK(KP923879532), T14, VMUL(LDK(KP382683432), T17));
		    {
			 V T19, T1a, T1c, T1d;
			 T19 = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
			 T1a = LD(&(xi[WS(is, 37)]), ivs, &(xi[WS(is, 1)]));
			 T1b = VSUB(T19, T1a);
			 T4R = VADD(T19, T1a);
			 T1c = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
			 T1d = LD(&(xi[WS(is, 53)]), ivs, &(xi[WS(is, 1)]));
			 T1e = VSUB(T1c, T1d);
			 T4S = VADD(T1c, T1d);
		    }
		    T1f = VFMA(LDK(KP382683432), T1b, VMUL(LDK(KP923879532), T1e));
		    T4T = VSUB(T4R, T4S);
		    T1y = VFNMS(LDK(KP382683432), T1e, VMUL(LDK(KP923879532), T1b));
	       }
	       {
		    V T1g, T1r, T6k, T6l;
		    T1g = VSUB(T18, T1f);
		    T1r = VSUB(T1n, T1q);
		    T1s = VSUB(T1g, T1r);
		    T2K = VADD(T1r, T1g);
		    T6k = VADD(T4O, T4P);
		    T6l = VADD(T50, T4Z);
		    T6m = VADD(T6k, T6l);
		    T6Y = VSUB(T6k, T6l);
	       }
	       {
		    V T6n, T6o, T1x, T1A;
		    T6n = VADD(T4R, T4S);
		    T6o = VADD(T4U, T4V);
		    T6p = VADD(T6n, T6o);
		    T6Z = VSUB(T6o, T6n);
		    T1x = VADD(T1v, T1w);
		    T1A = VADD(T1y, T1z);
		    T1B = VSUB(T1x, T1A);
		    T2L = VADD(T1x, T1A);
	       }
	       {
		    V T3k, T3l, T4Q, T4X;
		    T3k = VSUB(T1v, T1w);
		    T3l = VADD(T1f, T18);
		    T3m = VADD(T3k, T3l);
		    T3Y = VSUB(T3k, T3l);
		    T4Q = VSUB(T4O, T4P);
		    T4X = VMUL(LDK(KP707106781), VADD(T4T, T4W));
		    T4Y = VADD(T4Q, T4X);
		    T5O = VSUB(T4Q, T4X);
	       }
	       {
		    V T51, T52, T3n, T3o;
		    T51 = VSUB(T4Z, T50);
		    T52 = VMUL(LDK(KP707106781), VSUB(T4W, T4T));
		    T53 = VADD(T51, T52);
		    T5P = VSUB(T52, T51);
		    T3n = VADD(T1q, T1n);
		    T3o = VSUB(T1z, T1y);
		    T3p = VADD(T3n, T3o);
		    T3Z = VSUB(T3o, T3n);
	       }
	  }
	  {
	       V T7n, T7o, T7p, T7q, T7r, T7s, T7t, T7u, T7v, T7w, T7x, T7y, T7z, T7A, T7B;
	       V T7C, T7D, T7E, T7F, T7G, T7H, T7I, T7J, T7K;
	       {
		    V T6N, T6R, T6Q, T6S;
		    {
			 V T6L, T6M, T6O, T6P;
			 T6L = VADD(T6f, T6i);
			 T6M = VADD(T6F, T6C);
			 T6N = VADD(T6L, T6M);
			 T6R = VSUB(T6L, T6M);
			 T6O = VADD(T6m, T6p);
			 T6P = VADD(T6t, T6w);
			 T6Q = VADD(T6O, T6P);
			 T6S = VBYI(VSUB(T6P, T6O));
		    }
		    T7n = VSUB(T6N, T6Q);
		    STM2(&(xo[64]), T7n, ovs, &(xo[0]));
		    T7o = VADD(T6R, T6S);
		    STM2(&(xo[32]), T7o, ovs, &(xo[0]));
		    T7p = VADD(T6N, T6Q);
		    STM2(&(xo[0]), T7p, ovs, &(xo[0]));
		    T7q = VSUB(T6R, T6S);
		    STM2(&(xo[96]), T7q, ovs, &(xo[0]));
	       }
	       {
		    V T6j, T6G, T6y, T6H, T6q, T6x;
		    T6j = VSUB(T6f, T6i);
		    T6G = VSUB(T6C, T6F);
		    T6q = VSUB(T6m, T6p);
		    T6x = VSUB(T6t, T6w);
		    T6y = VMUL(LDK(KP707106781), VADD(T6q, T6x));
		    T6H = VMUL(LDK(KP707106781), VSUB(T6x, T6q));
		    {
			 V T6z, T6I, T6J, T6K;
			 T6z = VADD(T6j, T6y);
			 T6I = VBYI(VADD(T6G, T6H));
			 T7r = VSUB(T6z, T6I);
			 STM2(&(xo[112]), T7r, ovs, &(xo[0]));
			 T7s = VADD(T6z, T6I);
			 STM2(&(xo[16]), T7s, ovs, &(xo[0]));
			 T6J = VSUB(T6j, T6y);
			 T6K = VBYI(VSUB(T6H, T6G));
			 T7t = VSUB(T6J, T6K);
			 STM2(&(xo[80]), T7t, ovs, &(xo[0]));
			 T7u = VADD(T6J, T6K);
			 STM2(&(xo[48]), T7u, ovs, &(xo[0]));
		    }
	       }
	       {
		    V T6X, T7i, T78, T7g, T74, T7f, T7b, T7j, T6W, T77;
		    T6W = VMUL(LDK(KP707106781), VADD(T6U, T6V));
		    T6X = VADD(T6T, T6W);
		    T7i = VSUB(T6T, T6W);
		    T77 = VMUL(LDK(KP707106781), VSUB(T6V, T6U));
		    T78 = VADD(T76, T77);
		    T7g = VSUB(T77, T76);
		    {
			 V T70, T73, T79, T7a;
			 T70 = VFMA(LDK(KP923879532), T6Y, VMUL(LDK(KP382683432), T6Z));
			 T73 = VFNMS(LDK(KP382683432), T72, VMUL(LDK(KP923879532), T71));
			 T74 = VADD(T70, T73);
			 T7f = VSUB(T73, T70);
			 T79 = VFNMS(LDK(KP382683432), T6Y, VMUL(LDK(KP923879532), T6Z));
			 T7a = VFMA(LDK(KP382683432), T71, VMUL(LDK(KP923879532), T72));
			 T7b = VADD(T79, T7a);
			 T7j = VSUB(T7a, T79);
		    }
		    {
			 V T75, T7c, T7l, T7m;
			 T75 = VADD(T6X, T74);
			 T7c = VBYI(VADD(T78, T7b));
			 T7v = VSUB(T75, T7c);
			 STM2(&(xo[120]), T7v, ovs, &(xo[0]));
			 T7w = VADD(T75, T7c);
			 STM2(&(xo[8]), T7w, ovs, &(xo[0]));
			 T7l = VBYI(VADD(T7g, T7f));
			 T7m = VADD(T7i, T7j);
			 T7x = VADD(T7l, T7m);
			 STM2(&(xo[24]), T7x, ovs, &(xo[0]));
			 T7y = VSUB(T7m, T7l);
			 STM2(&(xo[104]), T7y, ovs, &(xo[0]));
		    }
		    {
			 V T7d, T7e, T7h, T7k;
			 T7d = VSUB(T6X, T74);
			 T7e = VBYI(VSUB(T7b, T78));
			 T7z = VSUB(T7d, T7e);
			 STM2(&(xo[72]), T7z, ovs, &(xo[0]));
			 T7A = VADD(T7d, T7e);
			 STM2(&(xo[56]), T7A, ovs, &(xo[0]));
			 T7h = VBYI(VSUB(T7f, T7g));
			 T7k = VSUB(T7i, T7j);
			 T7B = VADD(T7h, T7k);
			 STM2(&(xo[40]), T7B, ovs, &(xo[0]));
			 T7C = VSUB(T7k, T7h);
			 STM2(&(xo[88]), T7C, ovs, &(xo[0]));
		    }
	       }
	       {
		    V T5N, T68, T61, T69, T5U, T65, T5Y, T66;
		    {
			 V T5L, T5M, T5Z, T60;
			 T5L = VSUB(T4p, T4w);
			 T5M = VSUB(T5u, T5t);
			 T5N = VADD(T5L, T5M);
			 T68 = VSUB(T5L, T5M);
			 T5Z = VFNMS(LDK(KP555570233), T5O, VMUL(LDK(KP831469612), T5P));
			 T60 = VFMA(LDK(KP555570233), T5R, VMUL(LDK(KP831469612), T5S));
			 T61 = VADD(T5Z, T60);
			 T69 = VSUB(T60, T5Z);
		    }
		    {
			 V T5Q, T5T, T5W, T5X;
			 T5Q = VFMA(LDK(KP831469612), T5O, VMUL(LDK(KP555570233), T5P));
			 T5T = VFNMS(LDK(KP555570233), T5S, VMUL(LDK(KP831469612), T5R));
			 T5U = VADD(T5Q, T5T);
			 T65 = VSUB(T5T, T5Q);
			 T5W = VSUB(T5r, T5q);
			 T5X = VSUB(T4L, T4E);
			 T5Y = VADD(T5W, T5X);
			 T66 = VSUB(T5X, T5W);
		    }
		    {
			 V T5V, T62, T6b, T6c;
			 T5V = VADD(T5N, T5U);
			 T62 = VBYI(VADD(T5Y, T61));
			 T7D = VSUB(T5V, T62);
			 STM2(&(xo[116]), T7D, ovs, &(xo[0]));
			 T7E = VADD(T5V, T62);
			 STM2(&(xo[12]), T7E, ovs, &(xo[0]));
			 T6b = VBYI(VADD(T66, T65));
			 T6c = VADD(T68, T69);
			 T7F = VADD(T6b, T6c);
			 STM2(&(xo[20]), T7F, ovs, &(xo[0]));
			 T7G = VSUB(T6c, T6b);