t2sv_32.c

快速fft变换

			 TA = VFMA(Tu, Tv, VMUL(Ty, Tz));
			 T3C = VFNMS(Ty, Tv, VMUL(Tu, Tz));
			 TE = LD(&(ri[WS(rs, 24)]), ms, &(ri[0]));
			 TI = LD(&(ii[WS(rs, 24)]), ms, &(ii[0]));
			 TJ = VFMA(TD, TE, VMUL(TH, TI));
			 T3D = VFNMS(TH, TE, VMUL(TD, TI));
		    }
		    {
			 V Tr, TK, T8a, T8b;
			 Tr = VADD(T1, Tq);
			 TK = VADD(TA, TJ);
			 TL = VADD(Tr, TK);
			 T6f = VSUB(Tr, TK);
			 T8a = VSUB(T7G, T7F);
			 T8b = VSUB(TA, TJ);
			 T8c = VSUB(T8a, T8b);
			 T8q = VADD(T8b, T8a);
		    }
		    {
			 V T3B, T3E, T7E, T7H;
			 T3B = VSUB(T1, Tq);
			 T3E = VSUB(T3C, T3D);
			 T3F = VSUB(T3B, T3E);
			 T5t = VADD(T3B, T3E);
			 T7E = VADD(T3C, T3D);
			 T7H = VADD(T7F, T7G);
			 T7I = VADD(T7E, T7H);
			 T7W = VSUB(T7H, T7E);
		    }
	       }
	       {
		    V T2e, T4g, T2w, T4z, T2j, T4h, T2n, T4y;
		    {
			 V T2c, T2d, T2r, T2v;
			 T2c = LD(&(ri[WS(rs, 1)]), ms, &(ri[WS(rs, 1)]));
			 T2d = LD(&(ii[WS(rs, 1)]), ms, &(ii[WS(rs, 1)]));
			 T2e = VFMA(T2, T2c, VMUL(T5, T2d));
			 T4g = VFNMS(T5, T2c, VMUL(T2, T2d));
			 T2r = LD(&(ri[WS(rs, 25)]), ms, &(ri[WS(rs, 1)]));
			 T2v = LD(&(ii[WS(rs, 25)]), ms, &(ii[WS(rs, 1)]));
			 T2w = VFMA(T2q, T2r, VMUL(T2u, T2v));
			 T4z = VFNMS(T2u, T2r, VMUL(T2q, T2v));
		    }
		    {
			 V T2g, T2i, T2l, T2m;
			 T2g = LD(&(ri[WS(rs, 17)]), ms, &(ri[WS(rs, 1)]));
			 T2i = LD(&(ii[WS(rs, 17)]), ms, &(ii[WS(rs, 1)]));
			 T2j = VFMA(T2f, T2g, VMUL(T2h, T2i));
			 T4h = VFNMS(T2h, T2g, VMUL(T2f, T2i));
			 T2l = LD(&(ri[WS(rs, 9)]), ms, &(ri[WS(rs, 1)]));
			 T2m = LD(&(ii[WS(rs, 9)]), ms, &(ii[WS(rs, 1)]));
			 T2n = VFMA(T9, T2l, VMUL(Te, T2m));
			 T4y = VFNMS(Te, T2l, VMUL(T9, T2m));
		    }
		    {
			 V T2k, T2x, T6w, T6x;
			 T2k = VADD(T2e, T2j);
			 T2x = VADD(T2n, T2w);
			 T2y = VADD(T2k, T2x);
			 T6B = VSUB(T2k, T2x);
			 T6w = VADD(T4g, T4h);
			 T6x = VADD(T4y, T4z);
			 T6y = VSUB(T6w, T6x);
			 T7j = VADD(T6w, T6x);
		    }
		    {
			 V T4i, T4j, T4x, T4A;
			 T4i = VSUB(T4g, T4h);
			 T4j = VSUB(T2n, T2w);
			 T4k = VADD(T4i, T4j);
			 T5J = VSUB(T4i, T4j);
			 T4x = VSUB(T2e, T2j);
			 T4A = VSUB(T4y, T4z);
			 T4B = VSUB(T4x, T4A);
			 T5G = VADD(T4x, T4A);
		    }
	       }
	       {
		    V T31, T4Y, T3f, T4J, T36, T4Z, T3a, T4I;
		    {
			 V T2W, T30, T3c, T3e;
			 T2W = LD(&(ri[WS(rs, 31)]), ms, &(ri[WS(rs, 1)]));
			 T30 = LD(&(ii[WS(rs, 31)]), ms, &(ii[WS(rs, 1)]));
			 T31 = VFMA(T2V, T2W, VMUL(T2Z, T30));
			 T4Y = VFNMS(T2Z, T2W, VMUL(T2V, T30));
			 T3c = LD(&(ri[WS(rs, 23)]), ms, &(ri[WS(rs, 1)]));
			 T3e = LD(&(ii[WS(rs, 23)]), ms, &(ii[WS(rs, 1)]));
			 T3f = VFMA(T3b, T3c, VMUL(T3d, T3e));
			 T4J = VFNMS(T3d, T3c, VMUL(T3b, T3e));
		    }
		    {
			 V T33, T35, T38, T39;
			 T33 = LD(&(ri[WS(rs, 15)]), ms, &(ri[WS(rs, 1)]));
			 T35 = LD(&(ii[WS(rs, 15)]), ms, &(ii[WS(rs, 1)]));
			 T36 = VFMA(T32, T33, VMUL(T34, T35));
			 T4Z = VFNMS(T34, T33, VMUL(T32, T35));
			 T38 = LD(&(ri[WS(rs, 7)]), ms, &(ri[WS(rs, 1)]));
			 T39 = LD(&(ii[WS(rs, 7)]), ms, &(ii[WS(rs, 1)]));
			 T3a = VFMA(TR, T38, VMUL(TS, T39));
			 T4I = VFNMS(TS, T38, VMUL(TR, T39));
		    }
		    {
			 V T37, T3g, T6M, T6N;
			 T37 = VADD(T31, T36);
			 T3g = VADD(T3a, T3f);
			 T3h = VADD(T37, T3g);
			 T6H = VSUB(T37, T3g);
			 T6M = VADD(T4Y, T4Z);
			 T6N = VADD(T4I, T4J);
			 T6O = VSUB(T6M, T6N);
			 T7o = VADD(T6M, T6N);
		    }
		    {
			 V T4H, T4K, T50, T51;
			 T4H = VSUB(T31, T36);
			 T4K = VSUB(T4I, T4J);
			 T4L = VSUB(T4H, T4K);
			 T5N = VADD(T4H, T4K);
			 T50 = VSUB(T4Y, T4Z);
			 T51 = VSUB(T3a, T3f);
			 T52 = VADD(T50, T51);
			 T5Q = VSUB(T50, T51);
		    }
	       }
	       {
		    V TQ, T3G, T1g, T3N, TX, T3H, T17, T3M;
		    {
			 V TN, TP, T1b, T1f;
			 TN = LD(&(ri[WS(rs, 4)]), ms, &(ri[0]));
			 TP = LD(&(ii[WS(rs, 4)]), ms, &(ii[0]));
			 TQ = VFMA(TM, TN, VMUL(TO, TP));
			 T3G = VFNMS(TO, TN, VMUL(TM, TP));
			 T1b = LD(&(ri[WS(rs, 12)]), ms, &(ri[0]));
			 T1f = LD(&(ii[WS(rs, 12)]), ms, &(ii[0]));
			 T1g = VFMA(T1a, T1b, VMUL(T1e, T1f));
			 T3N = VFNMS(T1e, T1b, VMUL(T1a, T1f));
		    }
		    {
			 V TU, TW, T12, T16;
			 TU = LD(&(ri[WS(rs, 20)]), ms, &(ri[0]));
			 TW = LD(&(ii[WS(rs, 20)]), ms, &(ii[0]));
			 TX = VFMA(TT, TU, VMUL(TV, TW));
			 T3H = VFNMS(TV, TU, VMUL(TT, TW));
			 T12 = LD(&(ri[WS(rs, 28)]), ms, &(ri[0]));
			 T16 = LD(&(ii[WS(rs, 28)]), ms, &(ii[0]));
			 T17 = VFMA(T11, T12, VMUL(T15, T16));
			 T3M = VFNMS(T15, T12, VMUL(T11, T16));
		    }
		    {
			 V TY, T1h, T6g, T6h;
			 TY = VADD(TQ, TX);
			 T1h = VADD(T17, T1g);
			 T1i = VADD(TY, T1h);
			 T7V = VSUB(T1h, TY);
			 T6g = VADD(T3G, T3H);
			 T6h = VADD(T3M, T3N);
			 T6i = VSUB(T6g, T6h);
			 T7D = VADD(T6g, T6h);
		    }
		    {
			 V T3I, T3J, T3L, T3O;
			 T3I = VSUB(T3G, T3H);
			 T3J = VSUB(TQ, TX);
			 T3K = VSUB(T3I, T3J);
			 T5u = VADD(T3J, T3I);
			 T3L = VSUB(T17, T1g);
			 T3O = VSUB(T3M, T3N);
			 T3P = VADD(T3L, T3O);
			 T5v = VSUB(T3L, T3O);
		    }
	       }
	       {
		    V T1m, T3S, T1C, T3Z, T1r, T3T, T1x, T3Y;
		    {
			 V T1k, T1l, T1z, T1B;
			 T1k = LD(&(ri[WS(rs, 2)]), ms, &(ri[0]));
			 T1l = LD(&(ii[WS(rs, 2)]), ms, &(ii[0]));
			 T1m = VFMA(T8, T1k, VMUL(Td, T1l));
			 T3S = VFNMS(Td, T1k, VMUL(T8, T1l));
			 T1z = LD(&(ri[WS(rs, 26)]), ms, &(ri[0]));
			 T1B = LD(&(ii[WS(rs, 26)]), ms, &(ii[0]));
			 T1C = VFMA(T1y, T1z, VMUL(T1A, T1B));
			 T3Z = VFNMS(T1A, T1z, VMUL(T1y, T1B));
		    }
		    {
			 V T1o, T1q, T1u, T1w;
			 T1o = LD(&(ri[WS(rs, 18)]), ms, &(ri[0]));
			 T1q = LD(&(ii[WS(rs, 18)]), ms, &(ii[0]));
			 T1r = VFMA(T1n, T1o, VMUL(T1p, T1q));
			 T3T = VFNMS(T1p, T1o, VMUL(T1n, T1q));
			 T1u = LD(&(ri[WS(rs, 10)]), ms, &(ri[0]));
			 T1w = LD(&(ii[WS(rs, 10)]), ms, &(ii[0]));
			 T1x = VFMA(T1t, T1u, VMUL(T1v, T1w));
			 T3Y = VFNMS(T1v, T1u, VMUL(T1t, T1w));
		    }
		    {
			 V T1s, T1D, T6k, T6l;
			 T1s = VADD(T1m, T1r);
			 T1D = VADD(T1x, T1C);
			 T1E = VADD(T1s, T1D);
			 T6n = VSUB(T1s, T1D);
			 T6k = VADD(T3S, T3T);
			 T6l = VADD(T3Y, T3Z);
			 T6m = VSUB(T6k, T6l);
			 T7e = VADD(T6k, T6l);
		    }
		    {
			 V T3U, T3V, T3X, T40;
			 T3U = VSUB(T3S, T3T);
			 T3V = VSUB(T1x, T1C);
			 T3W = VADD(T3U, T3V);
			 T5y = VSUB(T3U, T3V);
			 T3X = VSUB(T1m, T1r);
			 T40 = VSUB(T3Y, T3Z);
			 T41 = VSUB(T3X, T40);
			 T5z = VADD(T3X, T40);
		    }
	       }
	       {
		    V T1J, T43, T27, T4a, T1U, T44, T20, T49;
		    {
			 V T1G, T1I, T24, T26;
			 T1G = LD(&(ri[WS(rs, 30)]), ms, &(ri[0]));
			 T1I = LD(&(ii[WS(rs, 30)]), ms, &(ii[0]));
			 T1J = VFMA(T1F, T1G, VMUL(T1H, T1I));
			 T43 = VFNMS(T1H, T1G, VMUL(T1F, T1I));
			 T24 = LD(&(ri[WS(rs, 22)]), ms, &(ri[0]));
			 T26 = LD(&(ii[WS(rs, 22)]), ms, &(ii[0]));
			 T27 = VFMA(T23, T24, VMUL(T25, T26));
			 T4a = VFNMS(T25, T24, VMUL(T23, T26));
		    }
		    {
			 V T1R, T1T, T1X, T1Z;
			 T1R = LD(&(ri[WS(rs, 14)]), ms, &(ri[0]));
			 T1T = LD(&(ii[WS(rs, 14)]), ms, &(ii[0]));
			 T1U = VFMA(T1Q, T1R, VMUL(T1S, T1T));
			 T44 = VFNMS(T1S, T1R, VMUL(T1Q, T1T));
			 T1X = LD(&(ri[WS(rs, 6)]), ms, &(ri[0]));
			 T1Z = LD(&(ii[WS(rs, 6)]), ms, &(ii[0]));
			 T20 = VFMA(T1W, T1X, VMUL(T1Y, T1Z));
			 T49 = VFNMS(T1Y, T1X, VMUL(T1W, T1Z));
		    }
		    {
			 V T1V, T28, T6q, T6r;
			 T1V = VADD(T1J, T1U);
			 T28 = VADD(T20, T27);
			 T29 = VADD(T1V, T28);
			 T6p = VSUB(T1V, T28);
			 T6q = VADD(T43, T44);
			 T6r = VADD(T49, T4a);
			 T6s = VSUB(T6q, T6r);
			 T7f = VADD(T6q, T6r);
		    }
		    {
			 V T45, T46, T48, T4b;
			 T45 = VSUB(T43, T44);
			 T46 = VSUB(T20, T27);
			 T47 = VADD(T45, T46);
			 T5B = VSUB(T45, T46);
			 T48 = VSUB(T1J, T1U);
			 T4b = VSUB(T49, T4a);
			 T4c = VSUB(T48, T4b);
			 T5C = VADD(T48, T4b);
		    }
	       }
	       {
		    V T2B, T4r, T2G, T4s, T4q, T4t, T2M, T4m, T2P, T4n, T4l, T4o;
		    {
			 V T2z, T2A, T2D, T2F;
			 T2z = LD(&(ri[WS(rs, 5)]), ms, &(ri[WS(rs, 1)]));
			 T2A = LD(&(ii[WS(rs, 5)]), ms, &(ii[WS(rs, 1)]));
			 T2B = VFMA(T21, T2z, VMUL(T22, T2A));
			 T4r = VFNMS(T22, T2z, VMUL(T21, T2A));
			 T2D = LD(&(ri[WS(rs, 21)]), ms, &(ri[WS(rs, 1)]));
			 T2F = LD(&(ii[WS(rs, 21)]), ms, &(ii[WS(rs, 1)]));
			 T2G = VFMA(T2C, T2D, VMUL(T2E, T2F));
			 T4s = VFNMS(T2E, T2D, VMUL(T2C, T2F));
		    }
		    T4q = VSUB(T2B, T2G);
		    T4t = VSUB(T4r, T4s);
		    {
			 V T2J, T2L, T2N, T2O;
			 T2J = LD(&(ri[WS(rs, 29)]), ms, &(ri[WS(rs, 1)]));
			 T2L = LD(&(ii[WS(rs, 29)]), ms, &(ii[WS(rs, 1)]));
			 T2M = VFMA(T2I, T2J, VMUL(T2K, T2L));
			 T4m = VFNMS(T2K, T2J, VMUL(T2I, T2L));
			 T2N = LD(&(ri[WS(rs, 13)]), ms, &(ri[WS(rs, 1)]));
			 T2O = LD(&(ii[WS(rs, 13)]), ms, &(ii[WS(rs, 1)]));
			 T2P = VFMA(T1M, T2N, VMUL(T1P, T2O));
			 T4n = VFNMS(T1P, T2N, VMUL(T1M, T2O));
		    }
		    T4l = VSUB(T2M, T2P);
		    T4o = VSUB(T4m, T4n);
		    {
			 V T2H, T2Q, T6C, T6D;
			 T2H = VADD(T2B, T2G);
			 T2Q = VADD(T2M, T2P);
			 T2R = VADD(T2H, T2Q);
			 T6z = VSUB(T2Q, T2H);
			 T6C = VADD(T4r, T4s);
			 T6D = VADD(T4m, T4n);
			 T6E = VSUB(T6C, T6D);
			 T7k = VADD(T6C, T6D);
		    }
		    {
			 V T4p, T4u, T4C, T4D;
			 T4p = VSUB(T4l, T4o);
			 T4u = VADD(T4q, T4t);
			 T4v = VMUL(LDK(KP707106781), VSUB(T4p, T4u));
			 T5H = VMUL(LDK(KP707106781), VADD(T4u, T4p));
			 T4C = VSUB(T4t, T4q);
			 T4D = VADD(T4l, T4o);
			 T4E = VMUL(LDK(KP707106781), VSUB(T4C, T4D));
			 T5K = VMUL(LDK(KP707106781), VADD(T4C, T4D));
		    }
	       }
	       {
		    V T3k, T4M, T3p, T4N, T4O, T4P, T3t, T4S, T3w, T4T, T4R, T4U;
		    {
			 V T3i, T3j, T3m, T3o;
			 T3i = LD(&(ri[WS(rs, 3)]), ms, &(ri[WS(rs, 1)]));
			 T3j = LD(&(ii[WS(rs, 3)]), ms, &(ii[WS(rs, 1)]));
			 T3k = VFMA(T3, T3i, VMUL(T6, T3j));
			 T4M = VFNMS(T6, T3i, VMUL(T3, T3j));
			 T3m = LD(&(ri[WS(rs, 19)]), ms, &(ri[WS(rs, 1)]));
			 T3o = LD(&(ii[WS(rs, 19)]), ms, &(ii[WS(rs, 1)]));
			 T3p = VFMA(T3l, T3m, VMUL(T3n, T3o));
			 T4N = VFNMS(T3n, T3m, VMUL(T3l, T3o));
		    }
		    T4O = VSUB(T4M, T4N);
		    T4P = VSUB(T3k, T3p);
		    {
			 V T3r, T3s, T3u, T3v;
			 T3r = LD(&(ri[WS(rs, 27)]), ms, &(ri[WS(rs, 1)]));
			 T3s = LD(&(ii[WS(rs, 27)]), ms, &(ii[WS(rs, 1)]));
			 T3t = VFMA(Th, T3r, VMUL(Tl, T3s));
			 T4S = VFNMS(Tl, T3r, VMUL(Th, T3s));
			 T3u = LD(&(ri[WS(rs, 11)]), ms, &(ri[WS(rs, 1)]));
			 T3v = LD(&(ii[WS(rs, 11)]), ms, &(ii[WS(rs, 1)]));
			 T3w = VFMA(Tg, T3u, VMUL(Tk, T3v));
			 T4T = VFNMS(Tk, T3u, VMUL(Tg, T3v));
		    }
		    T4R = VSUB(T3t, T3w);
		    T4U = VSUB(T4S, T4T);
		    {
			 V T3q, T3x, T6I, T6J;
			 T3q = VADD(T3k, T3p);
			 T3x = VADD(T3t, T3w);
			 T3y = VADD(T3q, T3x);
			 T6P = VSUB(T3x, T3q);
			 T6I = VADD(T4M, T4N);
			 T6J = VADD(T4S, T4T);
			 T6K = VSUB(T6I, T6J);
			 T7p = VADD(T6I, T6J);
		    }
		    {
			 V T4Q, T4V, T53, T54;
			 T4Q = VSUB(T4O, T4P);
			 T4V = VADD(T4R, T4U);