自学教程：C++ FMS函数代码示例

static void r2cbIII_5(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs){     DK(KP1_902113032, +1.902113032590307144232878666758764286811397268);     DK(KP1_118033988, +1.118033988749894848204586834365638117720309180);     DK(KP500000000, +0.500000000000000000000000000000000000000000000);     DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);     DK(KP618033988, +0.618033988749894848204586834365638117720309180);     {	  INT i;	  for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(rs), MAKE_VOLATILE_STRIDE(csr), MAKE_VOLATILE_STRIDE(csi)) {	       E T1, T2, T3, Tc, Ta, T8, T9;	       T8 = Ci[WS(csi, 1)];	       T9 = Ci[0];	       T1 = Cr[WS(csr, 2)];	       T2 = Cr[WS(csr, 1)];	       T3 = Cr[0];	       Tc = FMS(KP618033988, T8, T9);	       Ta = FMA(KP618033988, T9, T8);	       {		    E T6, T4, T5, T7, Tb;		    T6 = T3 - T2;		    T4 = T2 + T3;		    R0[0] = FMA(KP2_000000000, T4, T1);		    T5 = FNMS(KP500000000, T4, T1);		    T7 = FNMS(KP1_118033988, T6, T5);		    Tb = FMA(KP1_118033988, T6, T5);		    R0[WS(rs, 2)] = FNMS(KP1_902113032, Ta, T7);		    R1[0] = -(FMA(KP1_902113032, Ta, T7));		    R1[WS(rs, 1)] = FMS(KP1_902113032, Tc, Tb);		    R0[WS(rs, 1)] = FMA(KP1_902113032, Tc, Tb);	       }	  }     }}

static void hc2rIII_5(const R *ri, const R *ii, R *O, stride ris, stride iis, stride os, INT v, INT ivs, INT ovs){     DK(KP1_902113032, +1.902113032590307144232878666758764286811397268);     DK(KP1_118033988, +1.118033988749894848204586834365638117720309180);     DK(KP500000000, +0.500000000000000000000000000000000000000000000);     DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);     DK(KP618033988, +0.618033988749894848204586834365638117720309180);     INT i;     for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, O = O + ovs, MAKE_VOLATILE_STRIDE(ris), MAKE_VOLATILE_STRIDE(iis), MAKE_VOLATILE_STRIDE(os)) {	  E T1, T2, T3, Tc, Ta, T8, T9;	  T8 = ii[WS(iis, 1)];	  T9 = ii[0];	  T1 = ri[WS(ris, 2)];	  T2 = ri[WS(ris, 1)];	  T3 = ri[0];	  Tc = FMS(KP618033988, T8, T9);	  Ta = FMA(KP618033988, T9, T8);	  {	       E T6, T4, T5, T7, Tb;	       T6 = T3 - T2;	       T4 = T2 + T3;	       O[0] = FMA(KP2_000000000, T4, T1);	       T5 = FNMS(KP500000000, T4, T1);	       T7 = FNMS(KP1_118033988, T6, T5);	       Tb = FMA(KP1_118033988, T6, T5);	       O[WS(os, 4)] = FNMS(KP1_902113032, Ta, T7);	       O[WS(os, 1)] = -(FMA(KP1_902113032, Ta, T7));	       O[WS(os, 3)] = FMS(KP1_902113032, Tc, Tb);	       O[WS(os, 2)] = FMA(KP1_902113032, Tc, Tb);	  }     }}

    double findMedianSortedArrays(int A[], int m, int B[], int n) {        int total = m + n ;        if(total%2 == 1)        	return FMS(A,m,B,n,total/2+1);        else        	return (FMS(A,m,B,n,total/2+1) + FMS(A,m,B,n,total/2))/2;    }

static void r2cfII_8(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs){     DK(KP923879532, +0.923879532511286756128183189396788286822416626);     DK(KP707106781, +0.707106781186547524400844362104849039284835938);     DK(KP414213562, +0.414213562373095048801688724209698078569671875);     {	  INT i;	  for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(rs), MAKE_VOLATILE_STRIDE(csr), MAKE_VOLATILE_STRIDE(csi)) {	       E Te, T8, Td, T5, Tj, Tl, Tf, Tb;	       {		    E T1, Th, T9, Ti, T4, Ta;		    T1 = R0[0];		    Th = R0[WS(rs, 2)];		    {			 E T2, T3, T6, T7;			 T2 = R0[WS(rs, 1)];			 T3 = R0[WS(rs, 3)];			 T6 = R1[0];			 T7 = R1[WS(rs, 2)];			 T9 = R1[WS(rs, 3)];			 Ti = T2 + T3;			 T4 = T2 - T3;			 Te = FMA(KP414213562, T6, T7);			 T8 = FNMS(KP414213562, T7, T6);			 Ta = R1[WS(rs, 1)];		    }		    Td = FNMS(KP707106781, T4, T1);		    T5 = FMA(KP707106781, T4, T1);		    Tj = FMA(KP707106781, Ti, Th);		    Tl = FNMS(KP707106781, Ti, Th);		    Tf = FMA(KP414213562, T9, Ta);		    Tb = FMS(KP414213562, Ta, T9);	       }	       {		    E Tk, Tg, Tc, Tm;		    Tk = Te + Tf;		    Tg = Te - Tf;		    Tc = T8 + Tb;		    Tm = Tb - T8;		    Cr[WS(csr, 1)] = FMA(KP923879532, Tg, Td);		    Cr[WS(csr, 2)] = FNMS(KP923879532, Tg, Td);		    Ci[WS(csi, 3)] = FNMS(KP923879532, Tk, Tj);		    Ci[0] = -(FMA(KP923879532, Tk, Tj));		    Ci[WS(csi, 1)] = FMA(KP923879532, Tm, Tl);		    Ci[WS(csi, 2)] = FMS(KP923879532, Tm, Tl);		    Cr[0] = FMA(KP923879532, Tc, T5);		    Cr[WS(csr, 3)] = FNMS(KP923879532, Tc, T5);	       }	  }     }}

 double FMS(int a[], int m , int b[], int n, const int k) { 	if(m > n ) 		return FMS(b,n,a,m,k); 	if(m == 0) 		return b[k-1]; 	if(k == 1) 		return min(a[0],b[0]); 	int pa = min(m,k/2); 	int pb = k - pa; 	if(a[pa-1] < b[pb-1]) 		return FMS(a+pa,m-pa,b,n,k-pa); 	else 		return FMS(a,m,b+pb,n-pb,k-pb); }

static void hc2rIII_5(const R *ri, const R *ii, R *O, stride ris, stride iis, stride os, int v, int ivs, int ovs){     DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);     DK(KP1_118033988, +1.118033988749894848204586834365638117720309180);     DK(KP500000000, +0.500000000000000000000000000000000000000000000);     DK(KP1_175570504, +1.175570504584946258337411909278145537195304875);     DK(KP1_902113032, +1.902113032590307144232878666758764286811397268);     int i;     for (i = v; i > 0; i = i - 1, ri = ri + ivs, ii = ii + ivs, O = O + ovs) {	  E Ta, Tc, T1, T4, T5, T6, Tb, T7;	  {	       E T8, T9, T2, T3;	       T8 = ii[WS(iis, 1)];	       T9 = ii[0];	       Ta = FMA(KP1_902113032, T8, KP1_175570504 * T9);	       Tc = FNMS(KP1_902113032, T9, KP1_175570504 * T8);	       T1 = ri[WS(ris, 2)];	       T2 = ri[WS(ris, 1)];	       T3 = ri[0];	       T4 = T2 + T3;	       T5 = FMS(KP500000000, T4, T1);	       T6 = KP1_118033988 * (T3 - T2);	  }	  O[0] = FMA(KP2_000000000, T4, T1);	  Tb = T6 - T5;	  O[WS(os, 2)] = Tb + Tc;	  O[WS(os, 3)] = Tc - Tb;	  T7 = T5 + T6;	  O[WS(os, 1)] = T7 - Ta;	  O[WS(os, 4)] = -(T7 + Ta);     }}

static void r2cfII_6(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs){     DK(KP866025403, +0.866025403784438646763723170752936183471402627);     DK(KP500000000, +0.500000000000000000000000000000000000000000000);     {	  INT i;	  for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(24, rs), MAKE_VOLATILE_STRIDE(24, csr), MAKE_VOLATILE_STRIDE(24, csi)) {	       E T1, T9, T2, T3, T6, T7;	       T1 = R0[0];	       T9 = R1[WS(rs, 1)];	       T2 = R0[WS(rs, 2)];	       T3 = R0[WS(rs, 1)];	       T6 = R1[WS(rs, 2)];	       T7 = R1[0];	       {		    E Tc, T4, Ta, T8, T5, Tb;		    Cr[WS(csr, 1)] = T1 + T2 - T3;		    Tc = T2 + T3;		    T4 = T3 - T2;		    Ta = T6 + T7;		    T8 = T6 - T7;		    T5 = FMA(KP500000000, T4, T1);		    Tb = FMA(KP500000000, Ta, T9);		    Ci[WS(csi, 1)] = T9 - Ta;		    Cr[WS(csr, 2)] = FMA(KP866025403, T8, T5);		    Cr[0] = FNMS(KP866025403, T8, T5);		    Ci[WS(csi, 2)] = FMS(KP866025403, Tc, Tb);		    Ci[0] = -(FMA(KP866025403, Tc, Tb));	       }	  }     }}

static void r2hcII_6(const R *I, R *ro, R *io, stride is, stride ros, stride ios, INT v, INT ivs, INT ovs){     DK(KP866025403, +0.866025403784438646763723170752936183471402627);     DK(KP500000000, +0.500000000000000000000000000000000000000000000);     INT i;     for (i = v; i > 0; i = i - 1, I = I + ivs, ro = ro + ovs, io = io + ovs, MAKE_VOLATILE_STRIDE(is), MAKE_VOLATILE_STRIDE(ros), MAKE_VOLATILE_STRIDE(ios)) {	  E T1, T9, T2, T3, T6, T7;	  T1 = I[0];	  T9 = I[WS(is, 3)];	  T2 = I[WS(is, 4)];	  T3 = I[WS(is, 2)];	  T6 = I[WS(is, 5)];	  T7 = I[WS(is, 1)];	  {	       E Tc, T4, Ta, T8, T5, Tb;	       ro[WS(ros, 1)] = T1 + T2 - T3;	       Tc = T2 + T3;	       T4 = T3 - T2;	       Ta = T6 + T7;	       T8 = T6 - T7;	       T5 = FMA(KP500000000, T4, T1);	       Tb = FMA(KP500000000, Ta, T9);	       io[WS(ios, 1)] = T9 - Ta;	       ro[WS(ros, 2)] = FMA(KP866025403, T8, T5);	       ro[0] = FNMS(KP866025403, T8, T5);	       io[WS(ios, 2)] = FMS(KP866025403, Tc, Tb);	       io[0] = -(FMA(KP866025403, Tc, Tb));	  }     }}

static void r2cbIII_7(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs){     DK(KP1_949855824, +1.949855824363647214036263365987862434465571601);     DK(KP1_801937735, +1.801937735804838252472204639014890102331838324);     DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);     DK(KP692021471, +0.692021471630095869627814897002069140197260599);     DK(KP801937735, +0.801937735804838252472204639014890102331838324);     DK(KP356895867, +0.356895867892209443894399510021300583399127187);     DK(KP554958132, +0.554958132087371191422194871006410481067288862);     {	  INT i;	  for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(rs), MAKE_VOLATILE_STRIDE(csr), MAKE_VOLATILE_STRIDE(csi)) {	       E Tn, Td, Tg, Ti, Tl, T8;	       {		    E T1, T9, Tb, Ta, T2, T4, Th, Tm, Tc, T3, Te;		    T1 = Cr[WS(csr, 3)];		    T9 = Ci[WS(csi, 1)];		    Tb = Ci[0];		    Ta = Ci[WS(csi, 2)];		    T2 = Cr[WS(csr, 2)];		    T4 = Cr[0];		    Th = FMA(KP554958132, T9, Tb);		    Tm = FNMS(KP554958132, Ta, T9);		    Tc = FMA(KP554958132, Tb, Ta);		    T3 = Cr[WS(csr, 1)];		    Te = FNMS(KP356895867, T2, T4);		    Tn = FNMS(KP801937735, Tm, Tb);		    {			 E Tf, Tk, T7, T5, Tj, T6;			 Td = FMA(KP801937735, Tc, T9);			 T5 = T2 + T3 + T4;			 Tj = FNMS(KP356895867, T4, T3);			 T6 = FNMS(KP356895867, T3, T2);			 Tf = FNMS(KP692021471, Te, T3);			 R0[0] = FMA(KP2_000000000, T5, T1);			 Tk = FNMS(KP692021471, Tj, T2);			 T7 = FNMS(KP692021471, T6, T4);			 Tg = FNMS(KP1_801937735, Tf, T1);			 Ti = FNMS(KP801937735, Th, Ta);			 Tl = FNMS(KP1_801937735, Tk, T1);			 T8 = FNMS(KP1_801937735, T7, T1);		    }	       }	       R1[WS(rs, 2)] = FMS(KP1_949855824, Ti, Tg);	       R0[WS(rs, 1)] = FMA(KP1_949855824, Ti, Tg);	       R0[WS(rs, 2)] = FNMS(KP1_949855824, Tn, Tl);	       R1[WS(rs, 1)] = -(FMA(KP1_949855824, Tn, Tl));	       R0[WS(rs, 3)] = FNMS(KP1_949855824, Td, T8);	       R1[0] = -(FMA(KP1_949855824, Td, T8));	  }     }}

static void r2cf_8(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs){     DK(KP707106781, +0.707106781186547524400844362104849039284835938);     {	  INT i;	  for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(rs), MAKE_VOLATILE_STRIDE(csr), MAKE_VOLATILE_STRIDE(csi)) {	       E T4, T7, T3, Tj, Td, T5, T8, T9;	       {		    E T1, T2, Tb, Tc;		    T1 = R0[0];		    T2 = R0[WS(rs, 2)];		    Tb = R1[WS(rs, 3)];		    Tc = R1[WS(rs, 1)];		    T4 = R0[WS(rs, 1)];		    T7 = T1 - T2;		    T3 = T1 + T2;		    Tj = Tb + Tc;		    Td = Tb - Tc;		    T5 = R0[WS(rs, 3)];		    T8 = R1[0];		    T9 = R1[WS(rs, 2)];	       }	       {		    E T6, Tf, Ta, Ti;		    T6 = T4 + T5;		    Tf = T4 - T5;		    Ta = T8 - T9;		    Ti = T8 + T9;		    {			 E Th, Tk, Te, Tg;			 Th = T3 + T6;			 Cr[WS(csr, 2)] = T3 - T6;			 Tk = Ti + Tj;			 Ci[WS(csi, 2)] = Tj - Ti;			 Te = Ta + Td;			 Tg = Td - Ta;			 Cr[0] = Th + Tk;			 Cr[WS(csr, 4)] = Th - Tk;			 Ci[WS(csi, 3)] = FMA(KP707106781, Tg, Tf);			 Ci[WS(csi, 1)] = FMS(KP707106781, Tg, Tf);			 Cr[WS(csr, 1)] = FMA(KP707106781, Te, T7);			 Cr[WS(csr, 3)] = FNMS(KP707106781, Te, T7);		    }	       }	  }     }}

//.........这里部分代码省略.........					T4u = FNMS(KP559016994, T4t, T4s);					T4C = FMA(KP559016994, T4t, T4s);				   }				   {					E T2x, T4Q, T4p, T4r, T4R, T2A, T51, T55;					{					     E T4h, T50, T4X, T4o, T4Z;					     T4E = T4d + T4g;					     T4h = T4d - T4g;					     Rm[WS(rs, 1)] = FMA(KP951056516, T4B, T4u);					     Rp[WS(rs, 2)] = FNMS(KP951056516, T4B, T4u);					     Rp[WS(rs, 6)] = FMA(KP951056516, T4D, T4C);					     Rm[WS(rs, 5)] = FNMS(KP951056516, T4D, T4C);					     T50 = T4W - T4V;					     T4X = T4V + T4W;					     T4o = T4k - T4n;					     T4F = T4k + T4n;					     T2x = T2v + T2w;					     T4Q = T2v - T2w;					     Im[WS(rs, 9)] = T4X - T4Y;					     T4Z = FMA(KP250000000, T4X, T4Y);					     T4p = FMA(KP618033988, T4o, T4h);					     T4r = FNMS(KP618033988, T4h, T4o);					     T4R = T2z - T2y;					     T2A = T2y + T2z;					     T51 = FNMS(KP559016994, T50, T4Z);					     T55 = FMA(KP559016994, T50, T4Z);					}					{					     E T49, T48, T2B, T4a, T4q;					     T2B = T2x + T2A;					     T49 = T2x - T2A;					     Ip[WS(rs, 2)] = FMA(KP951056516, T54, T51);					     Im[WS(rs, 1)] = FMS(KP951056516, T54, T51);					     Ip[WS(rs, 6)] = FMA(KP951056516, T56, T55);					     Im[WS(rs, 5)] = FMS(KP951056516, T56, T55);					     Rp[0] = T2u + T2B;					     T48 = FNMS(KP250000000, T2B, T2u);					     T4a = FMA(KP559016994, T49, T48);					     T4q = FNMS(KP559016994, T49, T48);					     T4U = FMA(KP618033988, T4Q, T4R);					     T4S = FNMS(KP618033988, T4R, T4Q);					     Rm[WS(rs, 3)] = FMA(KP951056516, T4p, T4a);					     Rp[WS(rs, 4)] = FNMS(KP951056516, T4p, T4a);					     Rp[WS(rs, 8)] = FMA(KP951056516, T4r, T4q);					     Rm[WS(rs, 7)] = FNMS(KP951056516, T4r, T4q);					}				   }			      }			      {				   E T3O, T5u, T5w, T5o, T5q, T5n;				   {					E T5m, T5l, T2I, T4O, T3N, T3L, T2X, T5s, T4N, T5t, T3c, T3v, T3K, T4G;					T5m = T3u + T3n;					T3v = T3n - T3u;					T3K = T3C - T3J;					T5l = T3J + T3C;					T3O = T2C + T2H;					T2I = T2C - T2H;					T4O = T4E - T4F;					T4G = T4E + T4F;					T3N = FMA(KP618033988, T3v, T3K);					T3L = FNMS(KP618033988, T3K, T3v);					T2X = T2P + T2W;					T5s = T2P - T2W;					Ip[0] = T4G + T4M;

//.........这里部分代码省略.........					T1V = FNMS(KP866025403, T1U, T1P);					T3c = FNMS(KP618033988, T3b, T38);					T3e = FMA(KP618033988, T38, T3b);					T3q = FNMS(KP618033988, T3p, T3o);					T3s = FMA(KP618033988, T3o, T3p);					T1A = TU + T1z;					T34 = TU - T1z;				   }			      }			      {				   E T2W, T33, T3m, T3h, T2Z, T3d, T35, T3l;				   T3J = T2U + T2V;				   T2W = T2U - T2V;				   rio[0] = Tf + T1A;				   T33 = FNMS(KP250000000, T1A, Tf);				   T3m = T3f - T3g;				   T3h = T3f + T3g;				   T2Z = T2X - T2Y;				   T3I = T2X + T2Y;				   T3d = FMA(KP559016994, T34, T33);				   T35 = FNMS(KP559016994, T34, T33);				   iio[0] = T3h + T3k;				   T3l = FNMS(KP250000000, T3h, T3k);				   rio[WS(ios, 3)] = FMA(KP951056516, T3c, T35);				   iio[-WS(ios, 12)] = FNMS(KP951056516, T3c, T35);				   rio[WS(ios, 6)] = FMA(KP951056516, T3e, T3d);				   iio[-WS(ios, 9)] = FNMS(KP951056516, T3e, T3d);				   T3r = FMA(KP559016994, T3m, T3l);				   T3n = FNMS(KP559016994, T3m, T3l);				   T32 = FMA(KP618033988, T2W, T2Z);				   T30 = FNMS(KP618033988, T2Z, T2W);			      }			      iio[-WS(ios, 3)] = FMA(KP951056516, T3q, T3n);			      rio[WS(ios, 12)] = FMS(KP951056516, T3q, T3n);			      iio[-WS(ios, 6)] = FMA(KP951056516, T3s, T3r);			      rio[WS(ios, 9)] = FMS(KP951056516, T3s, T3r);			      T2d = FMA(KP866025403, T1G, T1B);			      T1H = FNMS(KP866025403, T1G, T1B);			      T1W = T1O + T1V;			      T3O = T1O - T1V;			      T3P = T23 - T2a;			      T2b = T23 + T2a;			 }		    }		    {			 E T3H, T3y, T2S, T3Q, T3S, T2R, T2c;			 T3H = FNMS(KP866025403, T3x, T3w);			 T3y = FMA(KP866025403, T3x, T3w);			 T2c = T1W + T2b;			 T2S = T1W - T2b;			 T3Q = FMA(KP618033988, T3P, T3O);			 T3S = FNMS(KP618033988, T3O, T3P);			 rio[WS(ios, 5)] = T1H + T2c;			 T2R = FNMS(KP250000000, T2c, T1H);			 {			      E T2g, T2j, T3G, T3E, T2A, T2N, T3v, T3A, T3M, T3L, T3z, T3F, T3B;			      {				   E T3C, T3D, T31, T2T, T3K;				   T2g = T2e + T2f;				   T3C = T2e - T2f;				   T3D = T2h - T2i;				   T2j = T2h + T2i;				   T31 = FMA(KP559016994, T2S, T2R);				   T2T = FNMS(KP559016994, T2S, T2R);				   T3K = T3I + T3J;				   T3M = T3J - T3I;

//.........这里部分代码省略.........				   E T3m, T3a, T3J, T3H;				   Ip[0] = T3u + T3C;				   Im[WS(rs, 7)] = T3u - T3C;				   Rp[WS(rs, 4)] = T3q + T3t;				   Rm[WS(rs, 3)] = T3q - T3t;				   T3m = T36 - T39;				   T3a = T36 + T39;				   T3J = T3G - T3F;				   T3H = T3F + T3G;				   {					E T2Q, T20, T3N, T3T, T2J, T2C, T3O, T2f, T34, T30, T2W, T2V, T3U, T2T, T2N;					E T2v;					{					     E T2R, T27, T2e, T2S;					     {						  E T3n, T3f, T3o, T3k;						  T2Q = T1U + T1Z;						  T20 = T1U - T1Z;						  T3n = T3e - T3b;						  T3f = T3b + T3e;						  T3o = T3g + T3j;						  T3k = T3g - T3j;						  T3N = T3L - T3M;						  T3T = T3M + T3L;						  {						       E T3p, T3I, T3K, T3l;						       T3p = T3n - T3o;						       T3I = T3n + T3o;						       T3K = T3k - T3f;						       T3l = T3f + T3k;						       Rp[WS(rs, 6)] = FMA(KP707106781, T3p, T3m);						       Rm[WS(rs, 1)] = FNMS(KP707106781, T3p, T3m);						       Ip[WS(rs, 2)] = FMA(KP707106781, T3I, T3H);						       Im[WS(rs, 5)] = FMS(KP707106781, T3I, T3H);						       Ip[WS(rs, 6)] = FMA(KP707106781, T3K, T3J);						       Im[WS(rs, 1)] = FMS(KP707106781, T3K, T3J);						       Rp[WS(rs, 2)] = FMA(KP707106781, T3l, T3a);						       Rm[WS(rs, 5)] = FNMS(KP707106781, T3l, T3a);						       T2R = T26 + T25;						       T27 = T25 - T26;						       T2e = T28 + T2d;						       T2S = T28 - T2d;						  }					     }					     {						  E T2Y, T2Z, T2n, T2u;						  T2J = T2D - T2I;						  T2Y = T2D + T2I;						  T2Z = T2B + T2w;						  T2C = T2w - T2B;						  T3O = T27 + T2e;						  T2f = T27 - T2e;						  T34 = FMA(KP414213562, T2Y, T2Z);						  T30 = FNMS(KP414213562, T2Z, T2Y);						  T2W = T2l - T2m;						  T2n = T2l + T2m;						  T2u = T2o - T2t;						  T2V = T2o + T2t;						  T3U = T2S - T2R;						  T2T = T2R + T2S;						  T2N = FNMS(KP414213562, T2n, T2u);						  T2v = FMA(KP414213562, T2u, T2n);					     }					}					{					     E T33, T2X, T3X, T3Y;

//.........这里部分代码省略.........			      TQ = iio[-WS(ios, 4)];			      Td = FMA(Tb, Tc, T9);			      T1q = T7 * Tc;			      TL = Tl * TK;			      T1b = T3 * TQ;			      TM = iio[0];			      T1r = FNMS(Tb, T8, T1q);			      TO = rio[WS(ios, 3)];			      {				   E TN, T1c, T1a, TR, T19, TP;				   Tx = rio[WS(ios, 1)];				   TN = FMA(Tn, TM, TL);				   T19 = Tl * TM;				   T1c = FNMS(T6, TO, T1b);				   TP = T3 * TO;				   Ty = T2 * Tx;				   T1a = FNMS(Tn, TK, T19);				   TH = iio[-WS(ios, 2)];				   TR = FMA(T6, TQ, TP);				   Tz = iio[-WS(ios, 6)];				   T1m = T1a + T1c;				   T1d = T1a - T1c;				   T14 = TC * TH;				   TS = TN + TR;				   T18 = TN - TR;				   T12 = T2 * Tz;				   TD = rio[WS(ios, 5)];			      }			      TA = FMA(T5, Tz, Ty);			 }		    }		    {			 E Te, T1p, T1t, T10, T1g, Tv, TJ, T1i, T1e, T1z, T1B, T17, T1h;			 {			      E T1x, T16, T1y, T11;			      {				   E TU, T13, T15, TE, TZ, TI;				   Te = T1 + Td;				   TU = T1 - Td;				   T13 = FNMS(T5, Tx, T12);				   T15 = FNMS(TG, TD, T14);				   TE = TC * TD;				   T1p = TW + TY;				   TZ = TW - TY;				   T1x = T1s - T1r;				   T1t = T1r + T1s;				   T1l = T13 + T15;				   T16 = T13 - T15;				   TI = FMA(TG, TH, TE);				   T10 = TU + TZ;				   T1g = TU - TZ;				   Tv = Tk + Tu;				   T1y = Tk - Tu;				   T11 = TA - TI;				   TJ = TA + TI;			      }			      T1i = T18 + T1d;			      T1e = T18 - T1d;			      T1z = T1x - T1y;			      T1B = T1y + T1x;			      T17 = T11 + T16;			      T1h = T16 - T11;			 }			 {			      E T1j, T1A, T1f, T1C;			      T1j = T1h - T1i;			      T1A = T1h + T1i;			      T1f = T17 + T1e;			      T1C = T1e - T17;			      iio[-WS(ios, 1)] = FMA(KP707106781, T1A, T1z);			      rio[WS(ios, 5)] = FMS(KP707106781, T1A, T1z);			      rio[WS(ios, 3)] = FMA(KP707106781, T1j, T1g);			      iio[-WS(ios, 7)] = FNMS(KP707106781, T1j, T1g);			      iio[-WS(ios, 3)] = FMA(KP707106781, T1C, T1B);			      rio[WS(ios, 7)] = FMS(KP707106781, T1C, T1B);			      rio[WS(ios, 1)] = FMA(KP707106781, T1f, T10);			      iio[-WS(ios, 5)] = FNMS(KP707106781, T1f, T10);			      T1k = Te - Tv;			      Tw = Te + Tv;			 }			 T1w = T1t - T1p;			 T1u = T1p + T1t;			 T1v = TS - TJ;			 TT = TJ + TS;		    }	       }	  }	  iio[-WS(ios, 2)] = T1v + T1w;	  rio[WS(ios, 6)] = T1v - T1w;	  rio[0] = Tw + TT;	  iio[-WS(ios, 4)] = Tw - TT;	  T1n = T1l - T1m;	  T1o = T1l + T1m;	  iio[0] = T1o + T1u;	  rio[WS(ios, 4)] = T1o - T1u;	  rio[WS(ios, 2)] = T1k + T1n;	  iio[-WS(ios, 6)] = T1k - T1n;     }     return W;}

//.........这里部分代码省略.........			 T1t = FMA(KP481753674, T1b, KP1_752613360 * T1e);			 T1u = FMA(KP684547105, T1i, KP1_457937254 * T1l);			 T1B = T1t + T1u;			 T1r = T1p - T1q;			 T1C = T1A + T1B;			 T1v = T1t - T1u;		    }		    {			 E T10, T17, T1f, T1m;			 T10 = FNMS(KP497379774, TZ, KP968583161 * TW);			 T17 = FNMS(KP1_688655851, T16, KP535826794 * T13);			 T18 = T10 + T17;			 T1f = FNMS(KP963507348, T1e, KP876306680 * T1b);			 T1m = FNMS(KP1_369094211, T1l, KP728968627 * T1i);			 T1n = T1f + T1m;			 T1o = T18 + T1n;			 T1G = T10 - T17;			 T1D = T1f - T1m;		    }		    {			 E T1R, T1Q, T20, T1Z;			 Cr[0] = TT + T1o;			 Ci[0] = -(T1z + T1C);			 T1R = KP559016994 * (T1P + T1M);			 T1Q = FMA(KP250000000, T1M - T1P, TT);			 Cr[WS(csr, 4)] = FMA(KP951056516, T1J, T1Q) + FMA(KP587785252, T1U, T1R);			 Cr[WS(csr, 9)] = FMA(KP951056516, T1U, T1Q) + FNMA(KP587785252, T1J, T1R);			 T20 = KP559016994 * (T1Y + T1X);			 T1Z = FMA(KP250000000, T1X - T1Y, T1z);			 Ci[WS(csi, 9)] = FMA(KP587785252, T1V, KP951056516 * T1W) + T1Z - T20;			 Ci[WS(csi, 4)] = FMA(KP587785252, T1W, T1Z) + FNMS(KP951056516, T1V, T20);			 {			      E T1E, T1F, T1s, T1w;			      T1E = FMS(KP250000000, T1C, T1z);			      T1F = KP559016994 * (T1B - T1A);			      Ci[WS(csi, 5)] = FMA(KP951056516, T1D, T1E) + FNMA(KP587785252, T1G, T1F);			      Ci[WS(csi, 10)] = FMA(KP951056516, T1G, KP587785252 * T1D) + T1E + T1F;			      T1s = FNMS(KP250000000, T1o, TT);			      T1w = KP559016994 * (T18 - T1n);			      Cr[WS(csr, 5)] = FMA(KP587785252, T1r, T1s) + FMS(KP951056516, T1v, T1w);			      Cr[WS(csr, 10)] = T1w + FMA(KP587785252, T1v, T1s) - (KP951056516 * T1r);			 }		    }	       }	       {		    E T21, T2z, T2L, T2K, T2M, T2F, T2P, T2C, T2Q, T2l, T2o, T2p, T2w, T2u, T28;		    E T2f, T2g, T2s, T2h;		    T21 = TS - TR;		    {			 E T2x, T2y, T2I, T2J;			 T2x = FNMS(KP844327925, T29, KP1_071653589 * T2a);			 T2y = FNMS(KP125581039, T2d, KP998026728 * T2c);			 T2z = T2x + T2y;			 T2L = T2y - T2x;			 T2I = FNMS(KP481753674, T22, KP1_752613360 * T23);			 T2J = FMA(KP904827052, T25, KP851558583 * T26);			 T2K = T2I + T2J;			 T2M = T2I - T2J;		    }		    {			 E T2D, T2E, T2A, T2B;			 T2D = FMA(KP535826794, T29, KP1_688655851 * T2a);			 T2E = FMA(KP062790519, T2c, KP1_996053456 * T2d);			 T2F = T2D + T2E;			 T2P = T2E - T2D;			 T2A = FMA(KP876306680, T22, KP963507348 * T23);

//.........这里部分代码省略.........					T1i = Tf * Tj;					Th = Tf * Tg;					Tm = W[14];					Tp = W[15];					T1j = FNMS(Ti, Tg, T1i);					Tk = FMA(Ti, Tj, Th);					T1f = Tm * Tq;					To = Tm * Tn;				   }			      }			      {				   E T1m, T1g, Tr, T1h;				   T1m = T1j - T1l;				   T1I = T1l + T1j;				   Tl = Te - Tk;				   T13 = Te + Tk;				   T1g = FNMS(Tp, Tn, T1f);				   Tr = FMA(Tp, Tq, To);				   T1J = T1g + T1e;				   T1h = T1e - T1g;				   Ty = Tr - Tx;				   T14 = Tr + Tx;				   T1n = T1h - T1m;				   T1O = T1m + T1h;			      }			 }		    }	       }	       T24 = T1I + T1J;	       T1K = T1I - T1J;	       {		    E T2a, T15, Tz, T1Z;		    T2a = T13 - T14;		    T15 = T13 + T14;		    Tz = Tl + Ty;		    T1Z = Ty - Tl;		    {			 E T1L, T1N, T1E, T1M;			 {			      E T19, T1D, T1C, T11, T1b;			      T19 = T15 + T18;			      T1D = T15 - T18;			      T11 = Tz + T10;			      T1b = Tz - T10;			      {				   E T1B, T1z, T1a, T1A, T1c;				   T1B = FNMS(KP618033988, T1n, T1y);				   T1z = FMA(KP618033988, T1y, T1n);				   ci[WS(rs, 4)] = T8 + T11;				   T1a = FNMS(KP250000000, T11, T8);				   T1A = FNMS(KP559016994, T1b, T1a);				   T1c = FMA(KP559016994, T1b, T1a);				   T1C = FNMS(KP250000000, T19, T12);				   T1L = FNMS(KP618033988, T1K, T1H);				   T1N = FMA(KP618033988, T1H, T1K);				   cr[WS(rs, 1)] = FMA(KP951056516, T1z, T1c);				   ci[0] = FNMS(KP951056516, T1z, T1c);				   cr[WS(rs, 3)] = FMA(KP951056516, T1B, T1A);				   ci[WS(rs, 2)] = FNMS(KP951056516, T1B, T1A);			      }			      cr[0] = T12 + T19;			      T1E = FNMS(KP559016994, T1D, T1C);			      T1M = FMA(KP559016994, T1D, T1C);			 }			 {			      E T1X, T21, T20, T22, T1Q, T1W, T1V, T26, T28, T27;			      T1Q = T1O + T1P;			      T1W = T1P - T1O;			      ci[WS(rs, 3)] = FMA(KP951056516, T1N, T1M);			      cr[WS(rs, 4)] = FNMS(KP951056516, T1N, T1M);			      ci[WS(rs, 1)] = FMA(KP951056516, T1L, T1E);			      cr[WS(rs, 2)] = FNMS(KP951056516, T1L, T1E);			      T1V = FMA(KP250000000, T1Q, T1U);			      cr[WS(rs, 5)] = T1Q - T1U;			      T1X = FNMS(KP559016994, T1W, T1V);			      T21 = FMA(KP559016994, T1W, T1V);			      T20 = FNMS(KP618033988, T1Z, T1Y);			      T22 = FMA(KP618033988, T1Y, T1Z);			      T26 = T24 + T25;			      T28 = T24 - T25;			      ci[WS(rs, 8)] = FMA(KP951056516, T22, T21);			      cr[WS(rs, 9)] = FMS(KP951056516, T22, T21);			      ci[WS(rs, 6)] = FMA(KP951056516, T20, T1X);			      cr[WS(rs, 7)] = FMS(KP951056516, T20, T1X);			      T27 = FNMS(KP250000000, T26, T23);			      ci[WS(rs, 9)] = T26 + T23;			      T29 = FMA(KP559016994, T28, T27);			      T2d = FNMS(KP559016994, T28, T27);			      T2c = FMA(KP618033988, T2b, T2a);			      T2e = FNMS(KP618033988, T2a, T2b);			 }		    }	       }	  }	  ci[WS(rs, 7)] = FMA(KP951056516, T2e, T2d);	  cr[WS(rs, 8)] = FMS(KP951056516, T2e, T2d);	  ci[WS(rs, 5)] = FMA(KP951056516, T2c, T29);	  cr[WS(rs, 6)] = FMS(KP951056516, T2c, T29);     }}

//.........这里部分代码省略.........				   T17 = Tx - Ty;				   Tz = Tx + Ty;				   T3 = R0[WS(rs, 12)];				   T5 = R0[WS(rs, 7)];				   T1L = FMA(KP618033988, T16, T17);				   T18 = FNMS(KP618033988, T17, T16);				   T1a = Tz - Tw;				   TA = Tw + Tz;				   T4 = T2 + T3;				   TU = T3 - T2;				   T6 = R1[WS(rs, 9)];			      }			      TB = Tt + TA;			      T19 = FNMS(KP250000000, TA, Tt);			      {				   E T7, TV, T1b, T1K, T8;				   T7 = T5 + T6;				   TV = T5 - T6;				   T1b = FNMS(KP559016994, T1a, T19);				   T1K = FMA(KP559016994, T1a, T19);				   T1P = FMA(KP618033988, TU, TV);				   TW = FNMS(KP618033988, TV, TU);				   TS = T4 - T7;				   T8 = T4 + T7;				   T21 = FMA(KP869845200, T1K, T1L);				   T1M = FNMS(KP786782374, T1L, T1K);				   T2e = FMA(KP066152395, T1K, T1L);				   T26 = FNMS(KP059835404, T1L, T1K);				   T1B = FMA(KP120146378, T18, T1b);				   T1r = FNMS(KP132830569, T1b, T18);				   T1k = FMA(KP893101515, T18, T1b);				   T1c = FNMS(KP987388751, T1b, T18);				   T9 = T1 + T8;				   TR = FMS(KP250000000, T8, T1);			      }			 }			 {			      E Ta, Te, TK, Td, Tf;			      Ta = R1[0];			      {				   E Tb, Tc, T1Q, TT;				   Tb = R0[WS(rs, 3)];				   Tc = R1[WS(rs, 10)];				   T1Q = FMA(KP559016994, TS, TR);				   TT = FNMS(KP559016994, TS, TR);				   Te = R1[WS(rs, 5)];				   TK = Tb - Tc;				   Td = Tb + Tc;				   T1X = FNMS(KP120146378, T1P, T1Q);				   T1R = FMA(KP132830569, T1Q, T1P);				   T2k = FMA(KP494780565, T1Q, T1P);				   T29 = FNMS(KP447533225, T1P, T1Q);				   T1z = FMA(KP869845200, TT, TW);				   T1v = FNMS(KP786782374, TW, TT);				   T1h = FNMS(KP667278218, TT, TW);				   TX = FMA(KP603558818, TW, TT);				   Tf = R0[WS(rs, 8)];			      }			      {				   E Tk, T1S, TM, TO, Tn, TZ, TN, T10, Tq, To, Th, Tp, TP, T1T, Tr;				   Tk = R0[WS(rs, 1)];				   {					E Tl, TL, Tg, Tm;					Tl = R1[WS(rs, 3)];					TL = Tf - Te;					Tg = Te + Tf;

//.........这里部分代码省略.........		    Tj = W[6];		    Tl = W[7];		    Tn = FMA(Tj, Tk, Tl * Tm);		    TT = FNMS(Tl, Tk, Tj * Tm);	       }	       {		    E Tp, Tr, To, Tq;		    Tp = cr[WS(rs, 7)];		    Tr = ci[WS(rs, 7)];		    To = W[12];		    Tq = W[13];		    Ts = FMA(To, Tp, Tq * Tr);		    TU = FNMS(Tq, Tp, To * Tr);	       }	       Tt = Tn + Ts;	       TY = TT + TU;	       Tu = Ti + Tt;	       T1w = TX + TY;	       {		    E TS, TV, TZ, T10;		    TS = FNMS(KP500000000, Tt, Ti);		    TV = KP866025403 * (TT - TU);		    TW = TS - TV;		    T1k = TS + TV;		    TZ = FNMS(KP500000000, TY, TX);		    T10 = KP866025403 * (Ts - Tn);		    T11 = TZ - T10;		    T1l = T10 + TZ;	       }	  }	  {	       E T1y, Td, TM, T1v;	       T1y = KP866025403 * (T1w - T1x);	       Td = T1 + Tc;	       TM = Tu + TL;	       T1v = FNMS(KP500000000, TM, Td);	       cr[0] = Td + TM;	       cr[WS(rs, 3)] = T1v + T1y;	       ci[WS(rs, 2)] = T1v - T1y;	  }	  {	       E TR, T1I, T1e, T1K, T1i, T1H, T1f, T1J;	       TR = TN - TQ;	       T1I = T1D - T1A;	       {		    E T12, T1d, T1g, T1h;		    T12 = FMA(KP173648177, TW, KP984807753 * T11);		    T1d = FNMS(KP939692620, T1c, KP342020143 * T19);		    T1e = T12 + T1d;		    T1K = KP866025403 * (T1d - T12);		    T1g = FNMS(KP984807753, TW, KP173648177 * T11);		    T1h = FMA(KP342020143, T1c, KP939692620 * T19);		    T1i = KP866025403 * (T1g + T1h);		    T1H = T1g - T1h;	       }	       cr[WS(rs, 2)] = TR + T1e;	       ci[WS(rs, 6)] = T1H + T1I;	       T1f = FNMS(KP500000000, T1e, TR);	       ci[0] = T1f - T1i;	       ci[WS(rs, 3)] = T1f + T1i;	       T1J = FMS(KP500000000, T1H, T1I);	       cr[WS(rs, 5)] = T1J - T1K;	       cr[WS(rs, 8)] = T1K + T1J;	  }	  {	       E T1L, T1M, T1N, T1O;	       T1L = KP866025403 * (TL - Tu);	       T1M = T1C + T1B;	       T1N = T1w + T1x;	       T1O = FNMS(KP500000000, T1N, T1M);	       cr[WS(rs, 6)] = T1L - T1O;	       ci[WS(rs, 8)] = T1N + T1M;	       ci[WS(rs, 5)] = T1L + T1O;	  }	  {	       E T1j, T1E, T1q, T1z, T1u, T1F, T1r, T1G;	       T1j = TN + TQ;	       T1E = T1A + T1D;	       {		    E T1m, T1p, T1s, T1t;		    T1m = FMA(KP766044443, T1k, KP642787609 * T1l);		    T1p = FMA(KP173648177, T1n, KP984807753 * T1o);		    T1q = T1m + T1p;		    T1z = KP866025403 * (T1p - T1m);		    T1s = FNMS(KP642787609, T1k, KP766044443 * T1l);		    T1t = FNMS(KP984807753, T1n, KP173648177 * T1o);		    T1u = KP866025403 * (T1s - T1t);		    T1F = T1s + T1t;	       }	       cr[WS(rs, 1)] = T1j + T1q;	       T1r = FNMS(KP500000000, T1q, T1j);	       ci[WS(rs, 1)] = T1r - T1u;	       cr[WS(rs, 4)] = T1r + T1u;	       ci[WS(rs, 7)] = T1F + T1E;	       T1G = FNMS(KP500000000, T1F, T1E);	       cr[WS(rs, 7)] = T1z - T1G;	       ci[WS(rs, 4)] = T1z + T1G;	  }     }}

static void r2cbIII_25(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs){     DK(KP968583161, +0.968583161128631119490168375464735813836012403);     DK(KP248689887, +0.248689887164854788242283746006447968417567406);     DK(KP684547105, +0.684547105928688673732283357621209269889519233);     DK(KP728968627, +0.728968627421411523146730319055259111372571664);     DK(KP062790519, +0.062790519529313376076178224565631133122484832);     DK(KP998026728, +0.998026728428271561952336806863450553336905220);     DK(KP876306680, +0.876306680043863587308115903922062583399064238);     DK(KP481753674, +0.481753674101715274987191502872129653528542010);     DK(KP535826794, +0.535826794978996618271308767867639978063575346);     DK(KP844327925, +0.844327925502015078548558063966681505381659241);     DK(KP904827052, +0.904827052466019527713668647932697593970413911);     DK(KP425779291, +0.425779291565072648862502445744251703979973042);     DK(KP250000000, +0.250000000000000000000000000000000000000000000);     DK(KP951056516, +0.951056516295153572116439333379382143405698634);     DK(KP587785252, +0.587785252292473129168705954639072768597652438);     DK(KP559016994, +0.559016994374947424102293417182819058860154590);     DK(KP500000000, +0.500000000000000000000000000000000000000000000);     DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);     DK(KP1_118033988, +1.118033988749894848204586834365638117720309180);     DK(KP1_175570504, +1.175570504584946258337411909278145537195304875);     DK(KP1_902113032, +1.902113032590307144232878666758764286811397268);     {	  INT i;	  for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(100, rs), MAKE_VOLATILE_STRIDE(100, csr), MAKE_VOLATILE_STRIDE(100, csi)) {	       E TS, T1O, T5, TP, T1N, TI, TH, Te, T17, T2h, T1y, T1V, T10, T2g, T1x;	       E T1S, Tz, Ty, Tn, T1m, T2e, T1B, T22, T1f, T2d, T1A, T1Z, TQ, TR;	       TQ = Ci[WS(csi, 2)];	       TR = Ci[WS(csi, 7)];	       TS = FNMS(KP1_175570504, TR, KP1_902113032 * TQ);	       T1O = FMA(KP1_902113032, TR, KP1_175570504 * TQ);	       {		    E T1, T4, TN, T2, T3, TO;		    T1 = Cr[WS(csr, 12)];		    T2 = Cr[WS(csr, 7)];		    T3 = Cr[WS(csr, 2)];		    T4 = T2 + T3;		    TN = KP1_118033988 * (T3 - T2);		    T5 = FMA(KP2_000000000, T4, T1);		    TO = FMS(KP500000000, T4, T1);		    TP = TN - TO;		    T1N = TO + TN;	       }	       {		    E T6, Td, T15, TU, T14, T11, TX, TY;		    T6 = Cr[WS(csr, 11)];		    TI = Ci[WS(csi, 11)];		    {			 E T7, T8, T9, Ta, Tb, Tc;			 T7 = Cr[WS(csr, 6)];			 T8 = Cr[WS(csr, 8)];			 T9 = T7 + T8;			 Ta = Cr[WS(csr, 1)];			 Tb = Cr[WS(csr, 3)];			 Tc = Ta + Tb;			 Td = T9 + Tc;			 T15 = Ta - Tb;			 TU = KP559016994 * (Tc - T9);			 T14 = T8 - T7;		    }		    {			 E TB, TC, TD, TE, TF, TG;			 TB = Ci[WS(csi, 6)];			 TC = Ci[WS(csi, 8)];			 TD = TB - TC;			 TE = Ci[WS(csi, 1)];			 TF = Ci[WS(csi, 3)];			 TG = TE - TF;			 TH = TD + TG;			 T11 = KP559016994 * (TD - TG);			 TX = TB + TC;			 TY = TE + TF;		    }		    Te = T6 + Td;		    {			 E T16, T1T, T13, T1U, T12;			 T16 = FMA(KP587785252, T14, KP951056516 * T15);			 T1T = FNMS(KP587785252, T15, KP951056516 * T14);			 T12 = FNMS(KP250000000, TH, TI);			 T13 = T11 - T12;			 T1U = T11 + T12;			 T17 = T13 - T16;			 T2h = T1T - T1U;			 T1y = T16 + T13;			 T1V = T1T + T1U;		    }		    {			 E TZ, T1R, TW, T1Q, TV;			 TZ = FNMS(KP951056516, TY, KP587785252 * TX);			 T1R = FMA(KP951056516, TX, KP587785252 * TY);			 TV = FMS(KP250000000, Td, T6);			 TW = TU - TV;			 T1Q = TV + TU;			 T10 = TW + TZ;			 T2g = T1Q + T1R;			 T1x = TZ - TW;			 T1S = T1Q - T1R;		    }	       }//.........这里部分代码省略.........

//.........这里部分代码省略.........			      TQ = FNMS(Tc, Ta, TP);			      Te = FMA(Tc, Td, Tb);			      T13 = TA * TE;			      TC = TA * TB;			      TK = Im[WS(rs, 1)];			      TG = W[4];			      TD = W[13];			      TJ = W[5];			      {				   E T14, TF, T16, TL, T15, TI;				   To = Ip[0];				   T15 = TG * TK;				   TI = TG * TH;				   T14 = FNMS(TD, TB, T13);				   TF = FMA(TD, TE, TC);				   T16 = FNMS(TJ, TH, T15);				   TL = FMA(TJ, TK, TI);				   Tr = Im[0];				   Tn = W[0];				   T17 = T14 - T16;				   T1g = T14 + T16;				   TM = TF + TL;				   T12 = TF - TL;			      }			      Tu = Ip[WS(rs, 2)];			      TW = Tn * Tr;			      Tp = Tn * To;			      Tx = Im[WS(rs, 2)];			      Tt = W[8];			      Tq = W[1];			      Tw = W[9];			 }		    }	       }	       {		    E T8, T1j, T1n, Tz, T1a, TU, Tl, T1b, T1c, T1v, T1t, T1w, T19, T1u, T1d;		    {			 E T1r, T10, TV, T1s, T11, T18;			 {			      E TO, TX, Ts, TZ, Ty, TT, TY, Tv;			      T8 = T1 + T7;			      TO = T1 - T7;			      TY = Tt * Tx;			      Tv = Tt * Tu;			      TX = FNMS(Tq, To, TW);			      Ts = FMA(Tq, Tr, Tp);			      TZ = FNMS(Tw, Tu, TY);			      Ty = FMA(Tw, Tx, Tv);			      TT = TQ - TS;			      T1j = TQ + TS;			      T1n = T1l + T1m;			      T1r = T1m - T1l;			      T10 = TX - TZ;			      T1f = TX + TZ;			      Tz = Ts + Ty;			      TV = Ts - Ty;			      T1a = TO - TT;			      TU = TO + TT;			      T1s = Te - Tk;			      Tl = Te + Tk;			 }			 T1b = T10 - TV;			 T11 = TV + T10;			 T18 = T12 - T17;			 T1c = T12 + T17;			 T1v = T1s + T1r;			 T1t = T1r - T1s;			 T1w = T18 - T11;			 T19 = T11 + T18;		    }		    Ip[WS(rs, 3)] = FMA(KP707106781, T1w, T1v);		    Im[0] = FMS(KP707106781, T1w, T1v);		    Rp[WS(rs, 1)] = FMA(KP707106781, T19, TU);		    Rm[WS(rs, 2)] = FNMS(KP707106781, T19, TU);		    T1u = T1b + T1c;		    T1d = T1b - T1c;		    Ip[WS(rs, 1)] = FMA(KP707106781, T1u, T1t);		    Im[WS(rs, 2)] = FMS(KP707106781, T1u, T1t);		    Rp[WS(rs, 3)] = FMA(KP707106781, T1d, T1a);		    Rm[0] = FNMS(KP707106781, T1d, T1a);		    T1e = T8 - Tl;		    Tm = T8 + Tl;		    T1q = T1n - T1j;		    T1o = T1j + T1n;		    T1p = TM - Tz;		    TN = Tz + TM;	       }	  }	  Ip[WS(rs, 2)] = T1p + T1q;	  Im[WS(rs, 1)] = T1p - T1q;	  Rp[0] = Tm + TN;	  Rm[WS(rs, 3)] = Tm - TN;	  T1h = T1f - T1g;	  T1i = T1f + T1g;	  Ip[0] = T1i + T1o;	  Im[WS(rs, 3)] = T1i - T1o;	  Rp[WS(rs, 2)] = T1e + T1h;	  Rm[WS(rs, 1)] = T1e - T1h;     }}

//.........这里部分代码省略.........                        T2h = T1U - T1T;                        T1V = T1T + T1U;                        TR = TN + TQ;                        TW = TN - TQ;                        T27 = Tf + Tu;                        T1W = T1S - T1V;                        T28 = T1S + T1V;                        T1y = FNMS(KP707106781, TR, TK);                        TS = FMA(KP707106781, TR, TK);                        T1z = FNMS(KP707106781, TW, TV);                        TX = FMA(KP707106781, TW, TV);                        T2a = T28 + T29;                    }                    Cr[WS(csr, 8)] = Tf - Tu;                    Ci[WS(csi, 8)] = T29 - T28;                }                Cr[0] = T27 + T2a;                Cr[WS(csr, 16)] = T27 - T2a;                {                    E T2s, T2i, T2v, T2f, T2r, T2p, T2l, T2t;                    {                        E T2o, T2e, T26, T24;                        T2o = T2d - T2c;                        T2e = T2c + T2d;                        T2s = FNMS(KP414213562, T2g, T2h);                        T2i = FMA(KP414213562, T2h, T2g);                        T26 = T23 - T1W;                        T24 = T1W + T23;                        T2v = FNMS(KP707106781, T2e, T2b);                        T2f = FMA(KP707106781, T2e, T2b);                        T2r = FMA(KP707106781, T2o, T2n);                        T2p = FNMS(KP707106781, T2o, T2n);                        Ci[WS(csi, 4)] = FMA(KP707106781, T26, T25);                        Ci[WS(csi, 12)] = FMS(KP707106781, T26, T25);                        Cr[WS(csr, 4)] = FMA(KP707106781, T24, T1P);                        Cr[WS(csr, 12)] = FNMS(KP707106781, T24, T1P);                        T2l = FNMS(KP414213562, T2k, T2j);                        T2t = FMA(KP414213562, T2j, T2k);                    }                    {                        E T1v, T1G, TH, T1s, T1F, T1w, T1o, T1g, T1p, T1n;                        {                            E T1f, TY, T1t, T1u, T1j, T1m;                            {                                E Tz, TG, T1q, T1r;                                T1v = FNMS(KP707106781, Ty, Tv);                                Tz = FMA(KP707106781, Ty, Tv);                                {                                    E T2q, T2m, T2w, T2u;                                    T2q = T2l - T2i;                                    T2m = T2i + T2l;                                    T2w = T2t - T2s;                                    T2u = T2s + T2t;                                    Ci[WS(csi, 10)] = FMA(KP923879532, T2q, T2p);                                    Ci[WS(csi, 6)] = FMS(KP923879532, T2q, T2p);                                    Cr[WS(csr, 2)] = FMA(KP923879532, T2m, T2f);                                    Cr[WS(csr, 14)] = FNMS(KP923879532, T2m, T2f);                                    Cr[WS(csr, 10)] = FNMS(KP923879532, T2w, T2v);                                    Cr[WS(csr, 6)] = FMA(KP923879532, T2w, T2v);                                    Ci[WS(csi, 2)] = FMA(KP923879532, T2u, T2r);                                    Ci[WS(csi, 14)] = FMS(KP923879532, T2u, T2r);                                    TG = TC + TF;                                    T1G = TF - TC;                                }                                T1f = FNMS(KP198912367, T1e, T19);                                T1q = FMA(KP198912367, T19, T1e);

//.........这里部分代码省略.........					     T7y = FNMS(KP831864738, T5q, T5n);					     T5r = FMA(KP831864738, T5q, T5n);					     T7w = FNMS(KP904730450, T7t, T7s);					     T7u = FMA(KP904730450, T7t, T7s);					}					ci[WS(rs, 20)] = FNMS(KP992114701, T7G, T7D);					T5u = FNMS(KP904730450, T5r, T5k);					T5s = FMA(KP904730450, T5r, T5k);				   }				   {					E T5a, T5c, T7A, T7C, T7v, T53, T5b, T51, T7H, T7x, T7B;					T5a = FNMS(KP726211448, T59, T56);					T5c = FMA(KP525970792, T56, T59);					ci[WS(rs, 23)] = FMA(KP968583161, T7u, T7r);					cr[WS(rs, 1)] = FMA(KP968583161, T5s, T5d);					T51 = FMA(KP248028675, T50, T3Z);					T7A = FNMS(KP683113946, T7z, T7y);					T7C = FMA(KP559154169, T7y, T7z);					T7v = FNMS(KP242145790, T7u, T7r);					T53 = FMA(KP554608978, T52, T51);					T5b = FNMS(KP554608978, T52, T51);					T7M = FNMS(KP525970792, T7L, T7K);					T7O = FMA(KP726211448, T7K, T7L);					ci[WS(rs, 10)] = FNMS(KP943557151, T5c, T5b);					ci[WS(rs, 5)] = FMA(KP943557151, T5c, T5b);					ci[0] = FMA(KP803003575, T5a, T53);					cr[WS(rs, 9)] = FNMS(KP803003575, T5a, T53);					T7x = FNMS(KP541454447, T7w, T7v);					T7B = FMA(KP541454447, T7w, T7v);					T7H = FMA(KP248028675, T7G, T7D);					cr[WS(rs, 21)] = -(FMA(KP921177326, T7C, T7B));					ci[WS(rs, 18)] = FNMS(KP921177326, T7C, T7B);					ci[WS(rs, 13)] = FMA(KP833417178, T7A, T7x);					cr[WS(rs, 16)] = FMS(KP833417178, T7A, T7x);					T5C = FMA(KP559154169, T5B, T5y);					T5E = FNMS(KP683113946, T5y, T5B);					T5t = FNMS(KP242145790, T5s, T5d);					T7J = FNMS(KP554608978, T7I, T7H);					T7N = FMA(KP554608978, T7I, T7H);				   }			      }			 }		    }		    {			 E T7Y, T80, T5v, T5D;			 cr[WS(rs, 24)] = -(FMA(KP803003575, T7O, T7N));			 ci[WS(rs, 15)] = FNMS(KP803003575, T7O, T7N);			 cr[WS(rs, 19)] = FMS(KP943557151, T7M, T7J);			 cr[WS(rs, 14)] = -(FMA(KP943557151, T7M, T7J));			 T5v = FMA(KP541454447, T5u, T5t);			 T5D = FNMS(KP541454447, T5u, T5t);			 cr[WS(rs, 11)] = FNMS(KP833417178, T5E, T5D);			 ci[WS(rs, 8)] = FMA(KP833417178, T5E, T5D);			 cr[WS(rs, 6)] = FMA(KP921177326, T5C, T5v);			 ci[WS(rs, 3)] = FNMS(KP921177326, T5C, T5v);			 T7Y = FMA(KP618033988, T7X, T7W);			 T80 = FNMS(KP618033988, T7W, T7X);			 {			      E T6t, T6p, T5H, T7d, T71, T6u, T6y, T6x, T6l, T7k, T6i, T7l, T7g, T6c, T6e;			      E T6s, T6L, T6J, T6C;			      {				   E T6A, T6B, T5O, T6j, T6h, T6a, T6q, T5R, T5U, T6r, T5Z, T62;				   {					E T5K, T7U, T7T, T5N, T7S;					T6t = FNMS(KP951056516, T5J, T5I);					T5K = FMA(KP951056516, T5J, T5I);

static void hb_9(R *cr, R *ci, const R *W, stride rs, INT mb, INT me, INT ms){     DK(KP954188894, +0.954188894138671133499268364187245676532219158);     DK(KP852868531, +0.852868531952443209628250963940074071936020296);     DK(KP492403876, +0.492403876506104029683371512294761506835321626);     DK(KP984807753, +0.984807753012208059366743024589523013670643252);     DK(KP777861913, +0.777861913430206160028177977318626690410586096);     DK(KP839099631, +0.839099631177280011763127298123181364687434283);     DK(KP363970234, +0.363970234266202361351047882776834043890471784);     DK(KP176326980, +0.176326980708464973471090386868618986121633062);     DK(KP866025403, +0.866025403784438646763723170752936183471402627);     DK(KP500000000, +0.500000000000000000000000000000000000000000000);     {	  INT m;	  for (m = mb, W = W + ((mb - 1) * 16); m < me; m = m + 1, cr = cr + ms, ci = ci - ms, W = W + 16, MAKE_VOLATILE_STRIDE(18, rs)) {	       E T1X, T1S, T1U, T1P, T1Y, T1T;	       {		    E T5, Tl, TQ, T1y, T1b, T1J, Tg, TE, TW, T13, T10, Tz, Tw, TT, T1K;		    E T1B, T1L, T1E;		    {			 E T1, Th, T2, T3, Ti, Tj;			 T1 = cr[0];			 Th = ci[WS(rs, 8)];			 T2 = cr[WS(rs, 3)];			 T3 = ci[WS(rs, 2)];			 Ti = ci[WS(rs, 5)];			 Tj = cr[WS(rs, 6)];			 {			      E T12, Tb, TZ, TY, Ta, Tq, T11, Tr, Ts, TS, Te, Tt;			      {				   E T6, Tm, Tn, To, T9, Tc, Td, Tp;				   {					E T7, T8, T1a, T4;					T6 = cr[WS(rs, 1)];					T1a = T2 - T3;					T4 = T2 + T3;					{					     E TP, Tk, TO, T19;					     TP = Ti + Tj;					     Tk = Ti - Tj;					     T7 = cr[WS(rs, 4)];					     T5 = T1 + T4;					     TO = FNMS(KP500000000, T4, T1);					     Tl = Th + Tk;					     T19 = FNMS(KP500000000, Tk, Th);					     TQ = FNMS(KP866025403, TP, TO);					     T1y = FMA(KP866025403, TP, TO);					     T1b = FMA(KP866025403, T1a, T19);					     T1J = FNMS(KP866025403, T1a, T19);					     T8 = ci[WS(rs, 1)];					}					Tm = ci[WS(rs, 7)];					Tn = ci[WS(rs, 4)];					To = cr[WS(rs, 7)];					T9 = T7 + T8;					T12 = T7 - T8;				   }				   Tb = cr[WS(rs, 2)];				   TZ = Tn + To;				   Tp = Tn - To;				   TY = FNMS(KP500000000, T9, T6);				   Ta = T6 + T9;				   Tc = ci[WS(rs, 3)];				   Td = ci[0];				   Tq = Tm + Tp;				   T11 = FMS(KP500000000, Tp, Tm);				   Tr = ci[WS(rs, 6)];				   Ts = cr[WS(rs, 5)];				   TS = Td - Tc;				   Te = Tc + Td;				   Tt = cr[WS(rs, 8)];			      }			      {				   E T1C, Tv, TR, T1D, T1z, T1A;				   {					E TU, Tu, TV, Tf;					TU = FNMS(KP500000000, Te, Tb);					Tf = Tb + Te;					Tu = Ts + Tt;					TV = Ts - Tt;					Tg = Ta + Tf;					TE = Ta - Tf;					TW = FMA(KP866025403, TV, TU);					T1C = FNMS(KP866025403, TV, TU);					Tv = Tr - Tu;					TR = FMA(KP500000000, Tu, Tr);				   }				   T1z = FMA(KP866025403, T12, T11);				   T13 = FNMS(KP866025403, T12, T11);				   T10 = FNMS(KP866025403, TZ, TY);				   T1A = FMA(KP866025403, TZ, TY);				   Tz = Tv - Tq;				   Tw = Tq + Tv;				   T1D = FMA(KP866025403, TS, TR);				   TT = FNMS(KP866025403, TS, TR);				   T1K = FNMS(KP176326980, T1z, T1A);				   T1B = FMA(KP176326980, T1A, T1z);				   T1L = FNMS(KP363970234, T1C, T1D);				   T1E = FMA(KP363970234, T1D, T1C);			      }//.........这里部分代码省略.........

static void e01_8(const R *I, R *O, stride is, stride os, INT v, INT ivs, INT ovs){     DK(KP1_961570560, +1.961570560806460898252364472268478073947867462);     DK(KP1_662939224, +1.662939224605090474157576755235811513477121624);     DK(KP198912367, +0.198912367379658006911597622644676228597850501);     DK(KP668178637, +0.668178637919298919997757686523080761552472251);     DK(KP707106781, +0.707106781186547524400844362104849039284835938);     DK(KP1_847759065, +1.847759065022573512256366378793576573644833252);     DK(KP414213562, +0.414213562373095048801688724209698078569671875);     DK(KP1_414213562, +1.414213562373095048801688724209698078569671875);     INT i;     for (i = v; i > 0; i = i - 1, I = I + ivs, O = O + ovs, MAKE_VOLATILE_STRIDE(is), MAKE_VOLATILE_STRIDE(os)) {	  E T8, Td, Th, T7, Tp, Tl, Te, Tb;	  {	       E Tj, T3, Tk, T6, T9, Ta;	       {		    E T1, T2, T4, T5;		    T1 = I[0];		    T2 = I[WS(is, 4)];		    T4 = I[WS(is, 2)];		    T5 = I[WS(is, 6)];		    T8 = I[WS(is, 1)];		    Tj = FNMS(KP1_414213562, T2, T1);		    T3 = FMA(KP1_414213562, T2, T1);		    Tk = FMS(KP414213562, T4, T5);		    T6 = FMA(KP414213562, T5, T4);		    Td = I[WS(is, 7)];		    T9 = I[WS(is, 5)];		    Ta = I[WS(is, 3)];	       }	       Th = FNMS(KP1_847759065, T6, T3);	       T7 = FMA(KP1_847759065, T6, T3);	       Tp = FNMS(KP1_847759065, Tk, Tj);	       Tl = FMA(KP1_847759065, Tk, Tj);	       Te = Ta - T9;	       Tb = T9 + Ta;	  }	  {	       E Tn, Tf, Tc, Tm;	       Tn = FNMS(KP707106781, Te, Td);	       Tf = FMA(KP707106781, Te, Td);	       Tc = FMA(KP707106781, Tb, T8);	       Tm = FNMS(KP707106781, Tb, T8);	       {		    E Tq, To, Tg, Ti;		    Tq = FMA(KP668178637, Tm, Tn);		    To = FNMS(KP668178637, Tn, Tm);		    Tg = FMA(KP198912367, Tf, Tc);		    Ti = FNMS(KP198912367, Tc, Tf);		    O[WS(os, 1)] = FMA(KP1_662939224, To, Tl);		    O[WS(os, 6)] = FNMS(KP1_662939224, To, Tl);		    O[WS(os, 2)] = FMA(KP1_662939224, Tq, Tp);		    O[WS(os, 5)] = FNMS(KP1_662939224, Tq, Tp);		    O[WS(os, 4)] = FMA(KP1_961570560, Ti, Th);		    O[WS(os, 3)] = FNMS(KP1_961570560, Ti, Th);		    O[0] = FMA(KP1_961570560, Tg, T7);		    O[WS(os, 7)] = FNMS(KP1_961570560, Tg, T7);	       }	  }     }}

static void r2cfII_10(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs){     DK(KP951056516, +0.951056516295153572116439333379382143405698634);     DK(KP559016994, +0.559016994374947424102293417182819058860154590);     DK(KP250000000, +0.250000000000000000000000000000000000000000000);     DK(KP618033988, +0.618033988749894848204586834365638117720309180);     INT i;     for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(rs), MAKE_VOLATILE_STRIDE(csr), MAKE_VOLATILE_STRIDE(csi)) {	  E Tq, Ti, Tk, Tu, Tw, Tp, Tb, Tj, Tr, Tv;	  {	       E T1, To, Ts, Tt, T8, Ta, Te, Tm, Tl, Th, Tn, T9;	       T1 = R0[0];	       To = R1[WS(rs, 2)];	       {		    E T2, T3, T5, T6;		    T2 = R0[WS(rs, 2)];		    T3 = R0[WS(rs, 3)];		    T5 = R0[WS(rs, 4)];		    T6 = R0[WS(rs, 1)];		    {			 E Tc, T4, T7, Td, Tf, Tg;			 Tc = R1[0];			 Ts = T2 + T3;			 T4 = T2 - T3;			 Tt = T5 + T6;			 T7 = T5 - T6;			 Td = R1[WS(rs, 4)];			 Tf = R1[WS(rs, 1)];			 Tg = R1[WS(rs, 3)];			 T8 = T4 + T7;			 Ta = T4 - T7;			 Te = Tc - Td;			 Tm = Tc + Td;			 Tl = Tf + Tg;			 Th = Tf - Tg;		    }	       }	       Cr[WS(csr, 2)] = T1 + T8;	       Tn = Tl - Tm;	       Tq = Tm + Tl;	       Ti = FMA(KP618033988, Th, Te);	       Tk = FNMS(KP618033988, Te, Th);	       Ci[WS(csi, 2)] = Tn - To;	       T9 = FNMS(KP250000000, T8, T1);	       Tu = FMA(KP618033988, Tt, Ts);	       Tw = FNMS(KP618033988, Ts, Tt);	       Tp = FMA(KP250000000, Tn, To);	       Tb = FMA(KP559016994, Ta, T9);	       Tj = FNMS(KP559016994, Ta, T9);	  }	  Tr = FMA(KP559016994, Tq, Tp);	  Tv = FNMS(KP559016994, Tq, Tp);	  Cr[WS(csr, 1)] = FNMS(KP951056516, Tk, Tj);	  Cr[WS(csr, 3)] = FMA(KP951056516, Tk, Tj);	  Cr[0] = FMA(KP951056516, Ti, Tb);	  Cr[WS(csr, 4)] = FNMS(KP951056516, Ti, Tb);	  Ci[WS(csi, 1)] = FNMS(KP951056516, Tw, Tv);	  Ci[WS(csi, 3)] = FMA(KP951056516, Tw, Tv);	  Ci[WS(csi, 4)] = FMS(KP951056516, Tu, Tr);	  Ci[0] = -(FMA(KP951056516, Tu, Tr));     }}

static void r2cfII_12(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs){     DK(KP707106781, +0.707106781186547524400844362104849039284835938);     DK(KP866025403, +0.866025403784438646763723170752936183471402627);     DK(KP500000000, +0.500000000000000000000000000000000000000000000);     INT i;     for (i = v; i > 0; i = i - 1, R0 = R0 + ivs, R1 = R1 + ivs, Cr = Cr + ovs, Ci = Ci + ovs, MAKE_VOLATILE_STRIDE(rs), MAKE_VOLATILE_STRIDE(csr), MAKE_VOLATILE_STRIDE(csi)) {	  E TD, TB, Tp, T9, Tq, Tr, TE, To, Ts, TC;	  {	       E T8, T1, Tv, Tm, TF, Tz, Tl, Ta, Tb, Tt, TA, T4, Tc;	       {		    E Tx, Th, Ti, Tj, Ty, T6, T7, T2, T3, Tk;		    Tx = R0[WS(rs, 3)];		    T6 = R0[WS(rs, 5)];		    T7 = R0[WS(rs, 1)];		    Th = R1[WS(rs, 4)];		    Ti = R1[WS(rs, 2)];		    Tj = R1[0];		    Ty = T6 + T7;		    T8 = T6 - T7;		    T1 = R0[0];		    Tv = Ti - Tj - Th;		    Tk = Ti - Tj;		    Tm = Ti + Tj;		    TF = Tx - Ty;		    Tz = FMA(KP500000000, Ty, Tx);		    T2 = R0[WS(rs, 2)];		    T3 = R0[WS(rs, 4)];		    Tl = FMA(KP500000000, Tk, Th);		    Ta = R1[WS(rs, 1)];		    Tb = R1[WS(rs, 3)];		    Tt = T1 + T3 - T2;		    TA = T3 + T2;		    T4 = T2 - T3;		    Tc = R1[WS(rs, 5)];	       }	       {		    E Tn, Tg, T5, Tu;		    TD = FNMS(KP866025403, TA, Tz);		    TB = FMA(KP866025403, TA, Tz);		    T5 = FMA(KP500000000, T4, T1);		    Tu = Ta + Tc - Tb;		    {			 E Td, Tf, TG, Tw, Te;			 Td = Tb - Tc;			 Tf = Tc + Tb;			 Tp = FMA(KP866025403, T8, T5);			 T9 = FNMS(KP866025403, T8, T5);			 TG = Tv - Tu;			 Tw = Tu + Tv;			 Te = FMA(KP500000000, Td, Ta);			 Tq = FMA(KP866025403, Tm, Tl);			 Tn = FNMS(KP866025403, Tm, Tl);			 Ci[WS(csi, 1)] = FMA(KP707106781, TG, TF);			 Ci[WS(csi, 4)] = FMS(KP707106781, TG, TF);			 Cr[WS(csr, 4)] = FMA(KP707106781, Tw, Tt);			 Cr[WS(csr, 1)] = FNMS(KP707106781, Tw, Tt);			 Tg = FNMS(KP866025403, Tf, Te);			 Tr = FMA(KP866025403, Tf, Te);		    }		    TE = Tg + Tn;		    To = Tg - Tn;	       }	  }	  Ci[WS(csi, 2)] = FMS(KP707106781, TE, TD);	  Ci[WS(csi, 3)] = FMA(KP707106781, TE, TD);	  Cr[0] = FMA(KP707106781, To, T9);	  Cr[WS(csr, 5)] = FNMS(KP707106781, To, T9);	  Ts = Tq - Tr;	  TC = Tr + Tq;	  Ci[0] = -(FMA(KP707106781, TC, TB));	  Ci[WS(csi, 5)] = FNMS(KP707106781, TC, TB);	  Cr[WS(csr, 2)] = FMA(KP707106781, Ts, Tp);	  Cr[WS(csr, 3)] = FNMS(KP707106781, Ts, Tp);     }}