n2bv_64.c

快速fft变换

		    TY = VSUB(TW, TX);
		    TZ = VFNMS(LDK(KP555570233), TY, VMUL(LDK(KP831469612), TT));
		    T2h = VFMA(LDK(KP555570233), TT, VMUL(LDK(KP831469612), TY));
		    T6v = VADD(T4I, T4J);
		    T6w = VADD(T4F, T4G);
		    T6x = VSUB(T6v, T6w);
		    T75 = VADD(T6v, T6w);
	       }
	       {
		    V T3r, T3s, T4H, T4K;
		    T3r = VADD(TS, TP);
		    T3s = VADD(TW, TX);
		    T3t = VFNMS(LDK(KP195090322), T3s, VMUL(LDK(KP980785280), T3r));
		    T3E = VFMA(LDK(KP195090322), T3r, VMUL(LDK(KP980785280), T3s));
		    T4H = VSUB(T4F, T4G);
		    T4K = VSUB(T4I, T4J);
		    T4L = VFNMS(LDK(KP382683432), T4K, VMUL(LDK(KP923879532), T4H));
		    T5p = VFMA(LDK(KP923879532), T4K, VMUL(LDK(KP382683432), T4H));
	       }
	  }
	  {
	       V T21, T5h, T26, T5g, T1Y, T5d, T27, T5c, T55, T56, T1J, T57, T29, T58, T59;
	       V T1Q, T5a, T2a;
	       {
		    V T1Z, T20, T24, T25;
		    T1Z = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
		    T20 = LD(&(xi[WS(is, 47)]), ivs, &(xi[WS(is, 1)]));
		    T21 = VSUB(T1Z, T20);
		    T5h = VADD(T1Z, T20);
		    T24 = LD(&(xi[WS(is, 63)]), ivs, &(xi[WS(is, 1)]));
		    T25 = LD(&(xi[WS(is, 31)]), ivs, &(xi[WS(is, 1)]));
		    T26 = VSUB(T24, T25);
		    T5g = VADD(T24, T25);
	       }
	       {
		    V T1S, T1T, T1U, T1V, T1W, T1X;
		    T1S = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
		    T1T = LD(&(xi[WS(is, 39)]), ivs, &(xi[WS(is, 1)]));
		    T1U = VSUB(T1S, T1T);
		    T1V = LD(&(xi[WS(is, 55)]), ivs, &(xi[WS(is, 1)]));
		    T1W = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
		    T1X = VSUB(T1V, T1W);
		    T1Y = VMUL(LDK(KP707106781), VSUB(T1U, T1X));
		    T5d = VADD(T1V, T1W);
		    T27 = VMUL(LDK(KP707106781), VADD(T1U, T1X));
		    T5c = VADD(T1S, T1T);
	       }
	       {
		    V T1F, T1I, T1M, T1P;
		    {
			 V T1D, T1E, T1G, T1H;
			 T1D = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
			 T1E = LD(&(xi[WS(is, 35)]), ivs, &(xi[WS(is, 1)]));
			 T1F = VSUB(T1D, T1E);
			 T55 = VADD(T1D, T1E);
			 T1G = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
			 T1H = LD(&(xi[WS(is, 51)]), ivs, &(xi[WS(is, 1)]));
			 T1I = VSUB(T1G, T1H);
			 T56 = VADD(T1G, T1H);
		    }
		    T1J = VFNMS(LDK(KP382683432), T1I, VMUL(LDK(KP923879532), T1F));
		    T57 = VSUB(T55, T56);
		    T29 = VFMA(LDK(KP382683432), T1F, VMUL(LDK(KP923879532), T1I));
		    {
			 V T1K, T1L, T1N, T1O;
			 T1K = LD(&(xi[WS(is, 59)]), ivs, &(xi[WS(is, 1)]));
			 T1L = LD(&(xi[WS(is, 27)]), ivs, &(xi[WS(is, 1)]));
			 T1M = VSUB(T1K, T1L);
			 T58 = VADD(T1K, T1L);
			 T1N = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
			 T1O = LD(&(xi[WS(is, 43)]), ivs, &(xi[WS(is, 1)]));
			 T1P = VSUB(T1N, T1O);
			 T59 = VADD(T1N, T1O);
		    }
		    T1Q = VFMA(LDK(KP923879532), T1M, VMUL(LDK(KP382683432), T1P));
		    T5a = VSUB(T58, T59);
		    T2a = VFNMS(LDK(KP382683432), T1M, VMUL(LDK(KP923879532), T1P));
	       }
	       {
		    V T1R, T22, T6k, T6l;
		    T1R = VSUB(T1J, T1Q);
		    T22 = VSUB(T1Y, T21);
		    T23 = VSUB(T1R, T22);
		    T2N = VADD(T22, T1R);
		    T6k = VADD(T5g, T5h);
		    T6l = VADD(T5c, T5d);
		    T6m = VSUB(T6k, T6l);
		    T70 = VADD(T6k, T6l);
	       }
	       {
		    V T6n, T6o, T28, T2b;
		    T6n = VADD(T55, T56);
		    T6o = VADD(T58, T59);
		    T6p = VSUB(T6n, T6o);
		    T71 = VADD(T6n, T6o);
		    T28 = VSUB(T26, T27);
		    T2b = VSUB(T29, T2a);
		    T2c = VSUB(T28, T2b);
		    T2O = VADD(T28, T2b);
	       }
	       {
		    V T3g, T3h, T5b, T5e;
		    T3g = VADD(T26, T27);
		    T3h = VADD(T1J, T1Q);
		    T3i = VADD(T3g, T3h);
		    T3Y = VSUB(T3g, T3h);
		    T5b = VMUL(LDK(KP707106781), VSUB(T57, T5a));
		    T5e = VSUB(T5c, T5d);
		    T5f = VSUB(T5b, T5e);
		    T5R = VADD(T5e, T5b);
	       }
	       {
		    V T5i, T5j, T3j, T3k;
		    T5i = VSUB(T5g, T5h);
		    T5j = VMUL(LDK(KP707106781), VADD(T57, T5a));
		    T5k = VSUB(T5i, T5j);
		    T5S = VADD(T5i, T5j);
		    T3j = VADD(T21, T1Y);
		    T3k = VADD(T29, T2a);
		    T3l = VADD(T3j, T3k);
		    T3Z = VSUB(T3k, T3j);
	       }
	  }
	  {
	       V T1q, T50, T1v, T4Z, T1n, T4W, T1w, T4V, T4O, T4P, T18, T4Q, T1y, T4R, T4S;
	       V T1f, T4T, T1z;
	       {
		    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);
		    T50 = 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);
		    T4Z = VADD(T1t, T1u);
	       }
	       {
		    V T1h, T1i, T1j, T1k, T1l, T1m;
		    T1h = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
		    T1i = LD(&(xi[WS(is, 41)]), ivs, &(xi[WS(is, 1)]));
		    T1j = VSUB(T1h, T1i);
		    T1k = LD(&(xi[WS(is, 57)]), ivs, &(xi[WS(is, 1)]));
		    T1l = LD(&(xi[WS(is, 25)]), ivs, &(xi[WS(is, 1)]));
		    T1m = VSUB(T1k, T1l);
		    T1n = VMUL(LDK(KP707106781), VSUB(T1j, T1m));
		    T4W = VADD(T1k, T1l);
		    T1w = VMUL(LDK(KP707106781), VADD(T1j, T1m));
		    T4V = VADD(T1h, T1i);
	       }
	       {
		    V T14, T17, T1b, T1e;
		    {
			 V T12, T13, T15, T16;
			 T12 = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
			 T13 = LD(&(xi[WS(is, 37)]), ivs, &(xi[WS(is, 1)]));
			 T14 = VSUB(T12, T13);
			 T4O = VADD(T12, T13);
			 T15 = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
			 T16 = LD(&(xi[WS(is, 53)]), ivs, &(xi[WS(is, 1)]));
			 T17 = VSUB(T15, T16);
			 T4P = VADD(T15, T16);
		    }
		    T18 = VFNMS(LDK(KP382683432), T17, VMUL(LDK(KP923879532), T14));
		    T4Q = VSUB(T4O, T4P);
		    T1y = VFMA(LDK(KP382683432), T14, VMUL(LDK(KP923879532), T17));
		    {
			 V T19, T1a, T1c, T1d;
			 T19 = LD(&(xi[WS(is, 61)]), ivs, &(xi[WS(is, 1)]));
			 T1a = LD(&(xi[WS(is, 29)]), ivs, &(xi[WS(is, 1)]));
			 T1b = VSUB(T19, T1a);
			 T4R = VADD(T19, T1a);
			 T1c = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
			 T1d = LD(&(xi[WS(is, 45)]), ivs, &(xi[WS(is, 1)]));
			 T1e = VSUB(T1c, T1d);
			 T4S = VADD(T1c, T1d);
		    }
		    T1f = VFMA(LDK(KP923879532), T1b, VMUL(LDK(KP382683432), T1e));
		    T4T = VSUB(T4R, T4S);
		    T1z = VFNMS(LDK(KP382683432), T1b, VMUL(LDK(KP923879532), T1e));
	       }
	       {
		    V T1g, T1r, T6d, T6e;
		    T1g = VSUB(T18, T1f);
		    T1r = VSUB(T1n, T1q);
		    T1s = VSUB(T1g, T1r);
		    T2K = VADD(T1r, T1g);
		    T6d = VADD(T4Z, T50);
		    T6e = VADD(T4V, T4W);
		    T6f = VSUB(T6d, T6e);
		    T6X = VADD(T6d, T6e);
	       }
	       {
		    V T6g, T6h, T1x, T1A;
		    T6g = VADD(T4O, T4P);
		    T6h = VADD(T4R, T4S);
		    T6i = VSUB(T6g, T6h);
		    T6Y = VADD(T6g, T6h);
		    T1x = VSUB(T1v, T1w);
		    T1A = VSUB(T1y, T1z);
		    T1B = VSUB(T1x, T1A);
		    T2L = VADD(T1x, T1A);
	       }
	       {
		    V T39, T3a, T4U, T4X;
		    T39 = VADD(T1v, T1w);
		    T3a = VADD(T18, T1f);
		    T3b = VADD(T39, T3a);
		    T3V = VSUB(T39, T3a);
		    T4U = VMUL(LDK(KP707106781), VSUB(T4Q, T4T));
		    T4X = VSUB(T4V, T4W);
		    T4Y = VSUB(T4U, T4X);
		    T5O = VADD(T4X, T4U);
	       }
	       {
		    V T51, T52, T3c, T3d;
		    T51 = VSUB(T4Z, T50);
		    T52 = VMUL(LDK(KP707106781), VADD(T4Q, T4T));
		    T53 = VSUB(T51, T52);
		    T5P = VADD(T51, T52);
		    T3c = VADD(T1q, T1n);
		    T3d = VADD(T1y, T1z);
		    T3e = VADD(T3c, T3d);
		    T3W = VSUB(T3d, T3c);
	       }
	  }
	  {
	       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 T7h, T7l, T7k, T7m;
		    {
			 V T7f, T7g, T7i, T7j;
			 T7f = VADD(T78, T79);
			 T7g = VADD(T74, T75);
			 T7h = VSUB(T7f, T7g);
			 T7l = VADD(T7f, T7g);
			 T7i = VADD(T6X, T6Y);
			 T7j = VADD(T70, T71);
			 T7k = VBYI(VSUB(T7i, T7j));
			 T7m = VADD(T7i, T7j);
		    }
		    T7n = VSUB(T7h, T7k);
		    STM2(&(xo[96]), T7n, ovs, &(xo[0]));
		    T7o = VADD(T7l, T7m);
		    STM2(&(xo[0]), T7o, ovs, &(xo[0]));
		    T7p = VADD(T7h, T7k);
		    STM2(&(xo[32]), T7p, ovs, &(xo[0]));
		    T7q = VSUB(T7l, T7m);
		    STM2(&(xo[64]), T7q, ovs, &(xo[0]));
	       }
	       {
		    V T76, T7a, T73, T7b, T6Z, T72;
		    T76 = VSUB(T74, T75);
		    T7a = VSUB(T78, T79);
		    T6Z = VSUB(T6X, T6Y);
		    T72 = VSUB(T70, T71);
		    T73 = VMUL(LDK(KP707106781), VSUB(T6Z, T72));
		    T7b = VMUL(LDK(KP707106781), VADD(T6Z, T72));
		    {
			 V T77, T7c, T7d, T7e;
			 T77 = VBYI(VSUB(T73, T76));
			 T7c = VSUB(T7a, T7b);
			 T7r = VADD(T77, T7c);
			 STM2(&(xo[48]), T7r, ovs, &(xo[0]));
			 T7s = VSUB(T7c, T77);
			 STM2(&(xo[80]), T7s, ovs, &(xo[0]));
			 T7d = VBYI(VADD(T76, T73));
			 T7e = VADD(T7a, T7b);
			 T7t = VADD(T7d, T7e);
			 STM2(&(xo[16]), T7t, ovs, &(xo[0]));
			 T7u = VSUB(T7e, T7d);
			 STM2(&(xo[112]), T7u, ovs, &(xo[0]));
		    }
	       }
	       {
		    V T6C, T6S, T6I, T6P, T6r, T6Q, T6L, T6T, T6y, T6H;
		    T6y = VMUL(LDK(KP707106781), VSUB(T6u, T6x));
		    T6C = VSUB(T6y, T6B);
		    T6S = VADD(T6B, T6y);
		    T6H = VMUL(LDK(KP707106781), VADD(T6u, T6x));
		    T6I = VSUB(T6G, T6H);
		    T6P = VADD(T6G, T6H);
		    {
			 V T6j, T6q, T6J, T6K;
			 T6j = VFNMS(LDK(KP382683432), T6i, VMUL(LDK(KP923879532), T6f));
			 T6q = VFMA(LDK(KP923879532), T6m, VMUL(LDK(KP382683432), T6p));
			 T6r = VSUB(T6j, T6q);
			 T6Q = VADD(T6j, T6q);
			 T6J = VFMA(LDK(KP382683432), T6f, VMUL(LDK(KP923879532), T6i));
			 T6K = VFNMS(LDK(KP382683432), T6m, VMUL(LDK(KP923879532), T6p));
			 T6L = VSUB(T6J, T6K);
			 T6T = VADD(T6J, T6K);
		    }
		    {
			 V T6D, T6M, T6V, T6W;
			 T6D = VBYI(VSUB(T6r, T6C));
			 T6M = VSUB(T6I, T6L);
			 T7v = VADD(T6D, T6M);
			 STM2(&(xo[40]), T7v, ovs, &(xo[0]));
			 T7w = VSUB(T6M, T6D);
			 STM2(&(xo[88]), T7w, ovs, &(xo[0]));
			 T6V = VSUB(T6P, T6Q);
			 T6W = VBYI(VSUB(T6T, T6S));
			 T7x = VSUB(T6V, T6W);
			 STM2(&(xo[72]), T7x, ovs, &(xo[0]));
			 T7y = VADD(T6V, T6W);
			 STM2(&(xo[56]), T7y, ovs, &(xo[0]));
		    }
		    {
			 V T6N, T6O, T6R, T6U;
			 T6N = VBYI(VADD(T6C, T6r));
			 T6O = VADD(T6I, T6L);
			 T7z = VADD(T6N, T6O);
			 STM2(&(xo[24]), T7z, ovs, &(xo[0]));
			 T7A = VSUB(T6O, T6N);
			 STM2(&(xo[104]), T7A, ovs, &(xo[0]));
			 T6R = VADD(T6P, T6Q);
			 T6U = VBYI(VADD(T6S, T6T));
			 T7B = VSUB(T6R, T6U);
			 STM2(&(xo[120]), T7B, ovs, &(xo[0]));
			 T7C = VADD(T6R, T6U);
			 STM2(&(xo[8]), T7C, ovs, &(xo[0]));
		    }
	       }
	       {
		    V T5N, T68, T61, T69, T5U, T65, T5Y, T66;
		    {
			 V T5L, T5M, T5Z, T60;
			 T5L = VADD(T4p, T4w);
			 T5M = VADD(T5o, T5p);
			 T5N = VSUB(T5L, T5M);
			 T68 = VADD(T5L, T5M);
			 T5Z = VFNMS(LDK(KP195090322), T5O, VMUL(LDK(KP980785280), T5P));
			 T60 = VFMA(LDK(KP195090322), T5R, VMUL(LDK(KP980785280), T5S));
			 T61 = VSUB(T5Z, T60);
			 T69 = VADD(T5Z, T60);
		    }
		    {
			 V T5Q, T5T, T5W, T5X;
			 T5Q = VFMA(LDK(KP980785280), T5O, VMUL(LDK(KP195090322), T5P));
			 T5T = VFNMS(LDK(KP195090322), T5S, VMUL(LDK(KP980785280), T5R));
			 T5U = VSUB(T5Q, T5T);
			 T65 = VADD(T5Q, T5T);
			 T5W = VADD(T4E, T4L);
			 T5X = VADD(T5u, T5r);
			 T5Y = VSUB(T5W, T5X);
			 T66 = VADD(T5X, T5W);
		    }
		    {
			 V T5V, T62, T6b, T6c;
			 T5V = VADD(T5N, T5U);
			 T62 = VBYI(VADD(T5Y, T61));
			 T7D = VSUB(T5V, T62);
			 STM2(&(xo[100]), T7D, ovs, &(xo[0]));
			 T7E = VADD(T5V, T62);
			 STM2(&(xo[28]), T7E, ovs, &(xo[0]));
			 T6b = VBYI(VADD(T66, T65));
			 T6c = VADD(T68, T69);
			 T7F = VADD(T6b, T6c);
			 STM2(&(xo[4]), T7F, ovs, &(xo[0]));
			 T7G = VSUB(T6c, T6b);