t1sv_32.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.

		    T2p = VADD(T2j, T2o);
		    T2A = VADD(T2u, T2z);
		    T2B = VADD(T2p, T2A);
		    T67 = VSUB(T2p, T2A);
		    T6c = VADD(T4o, T4p);
		    T6d = VADD(T48, T49);
		    T6e = VSUB(T6c, T6d);
		    T6O = VADD(T6c, T6d);
	       }
	       {
		    V T47, T4a, T4q, T4r;
		    T47 = VSUB(T2j, T2o);
		    T4a = VSUB(T48, T49);
		    T4b = VSUB(T47, T4a);
		    T5d = VADD(T47, T4a);
		    T4q = VSUB(T4o, T4p);
		    T4r = VSUB(T2u, T2z);
		    T4s = VADD(T4q, T4r);
		    T5g = VSUB(T4q, T4r);
	       }
	  }
	  {
	       V To, T36, TE, T3d, Tt, T37, Tz, T3c;
	       {
		    V Tl, Tn, Tk, Tm;
		    Tl = LD(&(ri[WS(rs, 4)]), ms, &(ri[0]));
		    Tn = LD(&(ii[WS(rs, 4)]), ms, &(ii[0]));
		    Tk = LDW(&(W[TWVL * 6]));
		    Tm = LDW(&(W[TWVL * 7]));
		    To = VFMA(Tk, Tl, VMUL(Tm, Tn));
		    T36 = VFNMS(Tm, Tl, VMUL(Tk, Tn));
	       }
	       {
		    V TB, TD, TA, TC;
		    TB = LD(&(ri[WS(rs, 12)]), ms, &(ri[0]));
		    TD = LD(&(ii[WS(rs, 12)]), ms, &(ii[0]));
		    TA = LDW(&(W[TWVL * 22]));
		    TC = LDW(&(W[TWVL * 23]));
		    TE = VFMA(TA, TB, VMUL(TC, TD));
		    T3d = VFNMS(TC, TB, VMUL(TA, TD));
	       }
	       {
		    V Tq, Ts, Tp, Tr;
		    Tq = LD(&(ri[WS(rs, 20)]), ms, &(ri[0]));
		    Ts = LD(&(ii[WS(rs, 20)]), ms, &(ii[0]));
		    Tp = LDW(&(W[TWVL * 38]));
		    Tr = LDW(&(W[TWVL * 39]));
		    Tt = VFMA(Tp, Tq, VMUL(Tr, Ts));
		    T37 = VFNMS(Tr, Tq, VMUL(Tp, Ts));
	       }
	       {
		    V Tw, Ty, Tv, Tx;
		    Tw = LD(&(ri[WS(rs, 28)]), ms, &(ri[0]));
		    Ty = LD(&(ii[WS(rs, 28)]), ms, &(ii[0]));
		    Tv = LDW(&(W[TWVL * 54]));
		    Tx = LDW(&(W[TWVL * 55]));
		    Tz = VFMA(Tv, Tw, VMUL(Tx, Ty));
		    T3c = VFNMS(Tx, Tw, VMUL(Tv, Ty));
	       }
	       {
		    V Tu, TF, T5G, T5H;
		    Tu = VADD(To, Tt);
		    TF = VADD(Tz, TE);
		    TG = VADD(Tu, TF);
		    T7l = VSUB(TF, Tu);
		    T5G = VADD(T36, T37);
		    T5H = VADD(T3c, T3d);
		    T5I = VSUB(T5G, T5H);
		    T73 = VADD(T5G, T5H);
	       }
	       {
		    V T38, T39, T3b, T3e;
		    T38 = VSUB(T36, T37);
		    T39 = VSUB(To, Tt);
		    T3a = VSUB(T38, T39);
		    T4U = VADD(T39, T38);
		    T3b = VSUB(Tz, TE);
		    T3e = VSUB(T3c, T3d);
		    T3f = VADD(T3b, T3e);
		    T4V = VSUB(T3b, T3e);
	       }
	  }
	  {
	       V TM, T3i, T12, T3p, TR, T3j, TX, T3o;
	       {
		    V TJ, TL, TI, TK;
		    TJ = LD(&(ri[WS(rs, 2)]), ms, &(ri[0]));
		    TL = LD(&(ii[WS(rs, 2)]), ms, &(ii[0]));
		    TI = LDW(&(W[TWVL * 2]));
		    TK = LDW(&(W[TWVL * 3]));
		    TM = VFMA(TI, TJ, VMUL(TK, TL));
		    T3i = VFNMS(TK, TJ, VMUL(TI, TL));
	       }
	       {
		    V TZ, T11, TY, T10;
		    TZ = LD(&(ri[WS(rs, 26)]), ms, &(ri[0]));
		    T11 = LD(&(ii[WS(rs, 26)]), ms, &(ii[0]));
		    TY = LDW(&(W[TWVL * 50]));
		    T10 = LDW(&(W[TWVL * 51]));
		    T12 = VFMA(TY, TZ, VMUL(T10, T11));
		    T3p = VFNMS(T10, TZ, VMUL(TY, T11));
	       }
	       {
		    V TO, TQ, TN, TP;
		    TO = LD(&(ri[WS(rs, 18)]), ms, &(ri[0]));
		    TQ = LD(&(ii[WS(rs, 18)]), ms, &(ii[0]));
		    TN = LDW(&(W[TWVL * 34]));
		    TP = LDW(&(W[TWVL * 35]));
		    TR = VFMA(TN, TO, VMUL(TP, TQ));
		    T3j = VFNMS(TP, TO, VMUL(TN, TQ));
	       }
	       {
		    V TU, TW, TT, TV;
		    TU = LD(&(ri[WS(rs, 10)]), ms, &(ri[0]));
		    TW = LD(&(ii[WS(rs, 10)]), ms, &(ii[0]));
		    TT = LDW(&(W[TWVL * 18]));
		    TV = LDW(&(W[TWVL * 19]));
		    TX = VFMA(TT, TU, VMUL(TV, TW));
		    T3o = VFNMS(TV, TU, VMUL(TT, TW));
	       }
	       {
		    V TS, T13, T5K, T5L;
		    TS = VADD(TM, TR);
		    T13 = VADD(TX, T12);
		    T14 = VADD(TS, T13);
		    T5N = VSUB(TS, T13);
		    T5K = VADD(T3i, T3j);
		    T5L = VADD(T3o, T3p);
		    T5M = VSUB(T5K, T5L);
		    T6E = VADD(T5K, T5L);
	       }
	       {
		    V T3k, T3l, T3n, T3q;
		    T3k = VSUB(T3i, T3j);
		    T3l = VSUB(TX, T12);
		    T3m = VADD(T3k, T3l);
		    T4Y = VSUB(T3k, T3l);
		    T3n = VSUB(TM, TR);
		    T3q = VSUB(T3o, T3p);
		    T3r = VSUB(T3n, T3q);
		    T4Z = VADD(T3n, T3q);
	       }
	  }
	  {
	       V T19, T3t, T1p, T3A, T1e, T3u, T1k, T3z;
	       {
		    V T16, T18, T15, T17;
		    T16 = LD(&(ri[WS(rs, 30)]), ms, &(ri[0]));
		    T18 = LD(&(ii[WS(rs, 30)]), ms, &(ii[0]));
		    T15 = LDW(&(W[TWVL * 58]));
		    T17 = LDW(&(W[TWVL * 59]));
		    T19 = VFMA(T15, T16, VMUL(T17, T18));
		    T3t = VFNMS(T17, T16, VMUL(T15, T18));
	       }
	       {
		    V T1m, T1o, T1l, T1n;
		    T1m = LD(&(ri[WS(rs, 22)]), ms, &(ri[0]));
		    T1o = LD(&(ii[WS(rs, 22)]), ms, &(ii[0]));
		    T1l = LDW(&(W[TWVL * 42]));
		    T1n = LDW(&(W[TWVL * 43]));
		    T1p = VFMA(T1l, T1m, VMUL(T1n, T1o));
		    T3A = VFNMS(T1n, T1m, VMUL(T1l, T1o));
	       }
	       {
		    V T1b, T1d, T1a, T1c;
		    T1b = LD(&(ri[WS(rs, 14)]), ms, &(ri[0]));
		    T1d = LD(&(ii[WS(rs, 14)]), ms, &(ii[0]));
		    T1a = LDW(&(W[TWVL * 26]));
		    T1c = LDW(&(W[TWVL * 27]));
		    T1e = VFMA(T1a, T1b, VMUL(T1c, T1d));
		    T3u = VFNMS(T1c, T1b, VMUL(T1a, T1d));
	       }
	       {
		    V T1h, T1j, T1g, T1i;
		    T1h = LD(&(ri[WS(rs, 6)]), ms, &(ri[0]));
		    T1j = LD(&(ii[WS(rs, 6)]), ms, &(ii[0]));
		    T1g = LDW(&(W[TWVL * 10]));
		    T1i = LDW(&(W[TWVL * 11]));
		    T1k = VFMA(T1g, T1h, VMUL(T1i, T1j));
		    T3z = VFNMS(T1i, T1h, VMUL(T1g, T1j));
	       }
	       {
		    V T1f, T1q, T5Q, T5R;
		    T1f = VADD(T19, T1e);
		    T1q = VADD(T1k, T1p);
		    T1r = VADD(T1f, T1q);
		    T5P = VSUB(T1f, T1q);
		    T5Q = VADD(T3t, T3u);
		    T5R = VADD(T3z, T3A);
		    T5S = VSUB(T5Q, T5R);
		    T6F = VADD(T5Q, T5R);
	       }
	       {
		    V T3v, T3w, T3y, T3B;
		    T3v = VSUB(T3t, T3u);
		    T3w = VSUB(T1k, T1p);
		    T3x = VADD(T3v, T3w);
		    T51 = VSUB(T3v, T3w);
		    T3y = VSUB(T19, T1e);
		    T3B = VSUB(T3z, T3A);
		    T3C = VSUB(T3y, T3B);
		    T52 = VADD(T3y, T3B);
	       }
	  }
	  {
	       V T1V, T3R, T20, T3S, T3Q, T3T, T26, T3M, T2b, T3N, T3L, T3O;
	       {
		    V T1S, T1U, T1R, T1T;
		    T1S = LD(&(ri[WS(rs, 5)]), ms, &(ri[WS(rs, 1)]));
		    T1U = LD(&(ii[WS(rs, 5)]), ms, &(ii[WS(rs, 1)]));
		    T1R = LDW(&(W[TWVL * 8]));
		    T1T = LDW(&(W[TWVL * 9]));
		    T1V = VFMA(T1R, T1S, VMUL(T1T, T1U));
		    T3R = VFNMS(T1T, T1S, VMUL(T1R, T1U));
	       }
	       {
		    V T1X, T1Z, T1W, T1Y;
		    T1X = LD(&(ri[WS(rs, 21)]), ms, &(ri[WS(rs, 1)]));
		    T1Z = LD(&(ii[WS(rs, 21)]), ms, &(ii[WS(rs, 1)]));
		    T1W = LDW(&(W[TWVL * 40]));
		    T1Y = LDW(&(W[TWVL * 41]));
		    T20 = VFMA(T1W, T1X, VMUL(T1Y, T1Z));
		    T3S = VFNMS(T1Y, T1X, VMUL(T1W, T1Z));
	       }
	       T3Q = VSUB(T1V, T20);
	       T3T = VSUB(T3R, T3S);
	       {
		    V T23, T25, T22, T24;
		    T23 = LD(&(ri[WS(rs, 29)]), ms, &(ri[WS(rs, 1)]));
		    T25 = LD(&(ii[WS(rs, 29)]), ms, &(ii[WS(rs, 1)]));
		    T22 = LDW(&(W[TWVL * 56]));
		    T24 = LDW(&(W[TWVL * 57]));
		    T26 = VFMA(T22, T23, VMUL(T24, T25));
		    T3M = VFNMS(T24, T23, VMUL(T22, T25));
	       }
	       {
		    V T28, T2a, T27, T29;
		    T28 = LD(&(ri[WS(rs, 13)]), ms, &(ri[WS(rs, 1)]));
		    T2a = LD(&(ii[WS(rs, 13)]), ms, &(ii[WS(rs, 1)]));
		    T27 = LDW(&(W[TWVL * 24]));
		    T29 = LDW(&(W[TWVL * 25]));
		    T2b = VFMA(T27, T28, VMUL(T29, T2a));
		    T3N = VFNMS(T29, T28, VMUL(T27, T2a));
	       }
	       T3L = VSUB(T26, T2b);
	       T3O = VSUB(T3M, T3N);
	       {
		    V T21, T2c, T62, T63;
		    T21 = VADD(T1V, T20);
		    T2c = VADD(T26, T2b);
		    T2d = VADD(T21, T2c);
		    T5Z = VSUB(T2c, T21);
		    T62 = VADD(T3R, T3S);
		    T63 = VADD(T3M, T3N);
		    T64 = VSUB(T62, T63);
		    T6K = VADD(T62, T63);
	       }
	       {
		    V T3P, T3U, T42, T43;
		    T3P = VSUB(T3L, T3O);
		    T3U = VADD(T3Q, T3T);
		    T3V = VMUL(LDK(KP707106781), VSUB(T3P, T3U));
		    T57 = VMUL(LDK(KP707106781), VADD(T3U, T3P));
		    T42 = VSUB(T3T, T3Q);
		    T43 = VADD(T3L, T3O);
		    T44 = VMUL(LDK(KP707106781), VSUB(T42, T43));
		    T5a = VMUL(LDK(KP707106781), VADD(T42, T43));
	       }
	  }
	  {
	       V T2G, T4c, T2L, T4d, T4e, T4f, T2R, T4i, T2W, T4j, T4h, T4k;
	       {
		    V T2D, T2F, T2C, T2E;
		    T2D = LD(&(ri[WS(rs, 3)]), ms, &(ri[WS(rs, 1)]));
		    T2F = LD(&(ii[WS(rs, 3)]), ms, &(ii[WS(rs, 1)]));
		    T2C = LDW(&(W[TWVL * 4]));
		    T2E = LDW(&(W[TWVL * 5]));
		    T2G = VFMA(T2C, T2D, VMUL(T2E, T2F));
		    T4c = VFNMS(T2E, T2D, VMUL(T2C, T2F));
	       }
	       {
		    V T2I, T2K, T2H, T2J;
		    T2I = LD(&(ri[WS(rs, 19)]), ms, &(ri[WS(rs, 1)]));
		    T2K = LD(&(ii[WS(rs, 19)]), ms, &(ii[WS(rs, 1)]));
		    T2H = LDW(&(W[TWVL * 36]));
		    T2J = LDW(&(W[TWVL * 37]));
		    T2L = VFMA(T2H, T2I, VMUL(T2J, T2K));
		    T4d = VFNMS(T2J, T2I, VMUL(T2H, T2K));
	       }
	       T4e = VSUB(T4c, T4d);
	       T4f = VSUB(T2G, T2L);
	       {
		    V T2O, T2Q, T2N, T2P;
		    T2O = LD(&(ri[WS(rs, 27)]), ms, &(ri[WS(rs, 1)]));
		    T2Q = LD(&(ii[WS(rs, 27)]), ms, &(ii[WS(rs, 1)]));
		    T2N = LDW(&(W[TWVL * 52]));
		    T2P = LDW(&(W[TWVL * 53]));
		    T2R = VFMA(T2N, T2O, VMUL(T2P, T2Q));
		    T4i = VFNMS(T2P, T2O, VMUL(T2N, T2Q));
	       }
	       {
		    V T2T, T2V, T2S, T2U;
		    T2T = LD(&(ri[WS(rs, 11)]), ms, &(ri[WS(rs, 1)]));
		    T2V = LD(&(ii[WS(rs, 11)]), ms, &(ii[WS(rs, 1)]));
		    T2S = LDW(&(W[TWVL * 20]));
		    T2U = LDW(&(W[TWVL * 21]));
		    T2W = VFMA(T2S, T2T, VMUL(T2U, T2V));
		    T4j = VFNMS(T2U, T2T, VMUL(T2S, T2V));
	       }
	       T4h = VSUB(T2R, T2W);
	       T4k = VSUB(T4i, T4j);
	       {
		    V T2M, T2X, T68, T69;
		    T2M = VADD(T2G, T2L);
		    T2X = VADD(T2R, T2W);
		    T2Y = VADD(T2M, T2X);
		    T6f = VSUB(T2X, T2M);
		    T68 = VADD(T4c, T4d);
		    T69 = VADD(T4i, T4j);
		    T6a = VSUB(T68, T69);
		    T6P = VADD(T68, T69);
	       }
	       {
		    V T4g, T4l, T4t, T4u;
		    T4g = VSUB(T4e, T4f);
		    T4l = VADD(T4h, T4k);
		    T4m = VMUL(LDK(KP707106781), VSUB(T4g, T4l));
		    T5h = VMUL(LDK(KP707106781), VADD(T4g, T4l));
		    T4t = VSUB(T4h, T4k);
		    T4u = VADD(T4f, T4e);
		    T4v = VMUL(LDK(KP707106781), VSUB(T4t, T4u));
		    T5e = VMUL(LDK(KP707106781), VADD(T4u, T4t));
	       }
	  }
	  {
	       V T1t, T6X, T7a, T7c, T30, T7b, T70, T71;
	       {
		    V TH, T1s, T72, T79;
		    TH = VADD(Tj, TG);
		    T1s = VADD(T14, T1r);
		    T1t = VADD(TH, T1s);
		    T6X = VSUB(TH, T1s);
		    T72 = VADD(T6E, T6F);
		    T79 = VADD(T73, T78);
		    T7a = VADD(T72, T79);
		    T7c = VSUB(T79, T72);
	       }
	       {
		    V T2e, T2Z, T6Y, T6Z;
		    T2e = VADD(T1Q, T2d);
		    T2Z = VADD(T2B, T2Y);
		    T30 = VADD(T2e, T2Z);
		    T7b = VSUB(T2Z, T2e);
		    T6Y = VADD(T6J, T6K);
		    T6Z = VADD(T6O, T6P);
		    T70 = VSUB(T6Y, T6Z);