Tbi = VADD(T5q, T5r);					     T5s = VSUB(T5q, T5r);					     T1X = LD(&(ri[WS(is, 43)]), ivs, &(ri[WS(is, 1)]));					     T5l = LD(&(ii[WS(is, 11)]), ivs, &(ii[WS(is, 1)]));					     T5m = LD(&(ii[WS(is, 43)]), ivs, &(ii[WS(is, 1)]));					}				   }				   Tbt = VADD(Tbf, Tbe);				   Tbg = VSUB(Tbe, Tbf);				   T5F = VFNMS(LDK(KP414213562), T5E, T5z);				   T5R = VFMA(LDK(KP414213562), T5z, T5E);				   T5p = VSUB(T1W, T1X);				   T1Y = VADD(T1W, T1X);				   Tbj = VADD(T5l, T5m);				   T5n = VSUB(T5l, T5m);				   T8z = VFMA(LDK(KP414213562), T8o, T8p);				   T8q = VFNMS(LDK(KP414213562), T8p, T8o);			      }			 }		    }	       }	       {		    V Tbm, Tbv, T9A, T8u, T9D, T8B, T6Z, T5T, T6W, T5G, TeL, TeM, TeN, TeO, TeR;		    V TeS, TeT, TeU, TeV, TeW, TeX, TeY, TeZ, Tf0, Tf1, Tf2, Tf3, Tf4, Tf5, Tf6;		    V Tf7, Tf8, Tf9, Tfa, Tfb, Tfc, TbE, Tao, Tfd, Tfe, Td7, Td8, Tff, Tfg, Tfh;		    V Tfi, Tfj, Tfk, Tfl, Tfm, Tfn, Tfo, Tfp, Tfq, Tfr, Tfs;		    {			 V Tel, Tdy, TdD, Tcu, Tcx, Teq, Tei, Ten, Tex, Teh, TeB, Tev, Te9, Tec;			 {			      V Tef, Teu, TeE, TeD, T11, TeF, T1w, T21, Tet, T2y, T33, Teg, T20;			      {				   V Tv, T8r, T5t, T1Z, Tbh, TdC, Tbk, T8s, T5o, T10, Tep, Tbl, Tbu;				   Tef = VSUB(Tf, Tu);				   Tv = VADD(Tf, Tu);				   T8r = VSUB(T5s, T5p);				   T5t = VADD(T5p, T5s);				   T1Z = VADD(T1V, T1Y);				   Tbh = VSUB(T1V, T1Y);				   TdC = VADD(Tbi, Tbj);				   Tbk = VSUB(Tbi, Tbj);				   T8s = VADD(T5k, T5n);				   T5o = VSUB(T5k, T5n);				   T10 = VADD(TK, TZ);				   Teu = VSUB(TZ, TK);				   Tel = VSUB(Tej, Tek);				   TeE = VADD(Tej, Tek);				   Tdy = VSUB(T1Z, T1S);				   T20 = VADD(T1S, T1Z);				   Tep = VADD(TdB, TdC);				   TdD = VSUB(TdB, TdC);				   Tbl = VADD(Tbh, Tbk);				   Tbu = VSUB(Tbh, Tbk);				   {					V T8t, T8A, T5S, T5u;					T8t = VFMA(LDK(KP414213562), T8s, T8r);					T8A = VFNMS(LDK(KP414213562), T8r, T8s);					T5S = VFNMS(LDK(KP414213562), T5o, T5t);					T5u = VFMA(LDK(KP414213562), T5t, T5o);					TeD = VSUB(Tv, T10);					T11 = VADD(Tv, T10);					Tcu = VSUB(Tbl, Tbg);					Tbm = VADD(Tbg, Tbl);					Tcx = VSUB(Tbt, Tbu);					Tbv = VADD(Tbt, Tbu);					T9A = VSUB(T8t, T8q);					T8u = VADD(T8q, T8t);					T9D = VSUB(T8z, T8A);					T8B = VADD(T8z, T8A);					T6Z = VADD(T5R, T5S);					T5T = VSUB(T5R, T5S);					T6W = VADD(T5F, T5u);					T5G = VSUB(T5u, T5F);					TeF = VADD(Teo, Tep);					Teq = VSUB(Teo, Tep);				   }			      }			      Tei = VSUB(T1g, T1v);			      T1w = VADD(T1g, T1v);			      T21 = VADD(T1L, T20);			      Ten = VSUB(T1L, T20);			      Tet = VSUB(T2i, T2x);			      T2y = VADD(T2i, T2x);			      T33 = VADD(T2N, T32);			      Teg = VSUB(T2N, T32);			      {				   V TeI, TeG, T23, T22, TeH, T34;				   TeI = VADD(TeE, TeF);				   TeG = VSUB(TeE, TeF);				   T23 = VSUB(T21, T1w);				   T22 = VADD(T1w, T21);				   TeH = VADD(T2y, T33);				   T34 = VSUB(T2y, T33);				   Tex = VSUB(Tef, Teg);				   Teh = VADD(Tef, Teg);				   TeJ = VSUB(TeD, TeG);				   STM4(&(ro[48]), TeJ, ovs, &(ro[0]));				   TeK = VADD(TeD, TeG);				   STM4(&(ro[16]), TeK, ovs, &(ro[0]));				   TeL = VADD(T11, T22);				   STM4(&(ro[0]), TeL, ovs, &(ro[0]));				   TeM = VSUB(T11, T22);				   STM4(&(ro[32]), TeM, ovs, &(ro[0]));				   TeN = VADD(TeH, TeI);				   STM4(&(io[0]), TeN, ovs, &(io[0]));				   TeO = VSUB(TeH, TeI);				   STM4(&(io[32]), TeO, ovs, &(io[0]));				   TeP = VSUB(T34, T23);				   STM4(&(io[48]), TeP, ovs, &(io[0]));				   TeQ = VADD(T23, T34);				   STM4(&(io[16]), TeQ, ovs, &(io[0]));				   TeB = VADD(Teu, Tet);				   Tev = VSUB(Tet, Teu);			      }			 }			 {			      V TdV, Tdb, TdJ, Te5, TdE, Tdz, TdZ, Tdo, Te6, Tdi, Teb, Te3, TdW, TdM, Tdt;			      V TdY;			      {				   V TdL, Tde, Tey, Tem, Tez, Ter, Tdh, TdK, Te1, Te2;				   TdV = VADD(Td9, Tda);				   Tdb = VSUB(Td9, Tda);				   TdJ = VSUB(TdH, TdI);				   Te5 = VADD(TdI, TdH);				   TdL = VADD(Tdd, Tdc);				   Tde = VSUB(Tdc, Tdd);				   Tey = VSUB(Tel, Tei);				   Tem = VADD(Tei, Tel);				   Tez = VADD(Ten, Teq);				   Ter = VSUB(Ten, Teq);				   Tdh = VADD(Tdf, Tdg);				   TdK = VSUB(Tdf, Tdg);				   TdE = VSUB(TdA, TdD);				   Te1 = VADD(TdA, TdD);				   Te2 = VADD(Tdy, Tdx);				   Tdz = VSUB(Tdx, Tdy);				   TdZ = VADD(Tdn, Tdm);				   Tdo = VSUB(Tdm, Tdn);				   {					V TeA, TeC, Tew, Tes;					TeA = VSUB(Tey, Tez);					TeC = VADD(Tey, Tez);					Tew = VSUB(Ter, Tem);					Tes = VADD(Tem, Ter);					Te6 = VADD(Tde, Tdh);					Tdi = VSUB(Tde, Tdh);					Teb = VFMA(LDK(KP414213562), Te1, Te2);					Te3 = VFNMS(LDK(KP414213562), Te2, Te1);					TdW = VADD(TdL, TdK);					TdM = VSUB(TdK, TdL);					TeR = VFMA(LDK(KP707106781), TeA, Tex);					STM4(&(ro[24]), TeR, ovs, &(ro[0]));					TeS = VFNMS(LDK(KP707106781), TeA, Tex);					STM4(&(ro[56]), TeS, ovs, &(ro[0]));					TeT = VFMA(LDK(KP707106781), TeC, TeB);					STM4(&(io[8]), TeT, ovs, &(io[0]));					TeU = VFNMS(LDK(KP707106781), TeC, TeB);					STM4(&(io[40]), TeU, ovs, &(io[0]));					TeV = VFMA(LDK(KP707106781), Tew, Tev);					STM4(&(io[24]), TeV, ovs, &(io[0]));					TeW = VFNMS(LDK(KP707106781), Tew, Tev);					STM4(&(io[56]), TeW, ovs, &(io[0]));					TeX = VFMA(LDK(KP707106781), Tes, Teh);					STM4(&(ro[8]), TeX, ovs, &(ro[0]));					TeY = VFNMS(LDK(KP707106781), Tes, Teh);					STM4(&(ro[40]), TeY, ovs, &(ro[0]));					Tdt = VSUB(Tdp, Tds);					TdY = VADD(Tdp, Tds);				   }			      }			      {				   V TdT, Tdj, TdP, TdN, TdR, Tdu, Tea, Te0, TdQ, TdF, TdX, Ted, Te7;				   TdT = VFNMS(LDK(KP707106781), Tdi, Tdb);				   Tdj = VFMA(LDK(KP707106781), Tdi, Tdb);				   TdP = VFMA(LDK(KP707106781), TdM, TdJ);				   TdN = VFNMS(LDK(KP707106781), TdM, TdJ);				   TdR = VFNMS(LDK(KP414213562), Tdo, Tdt);				   Tdu = VFMA(LDK(KP414213562), Tdt, Tdo);				   Tea = VFNMS(LDK(KP414213562), TdY, TdZ);				   Te0 = VFMA(LDK(KP414213562), TdZ, TdY);				   TdQ = VFMA(LDK(KP414213562), Tdz, TdE);				   TdF = VFNMS(LDK(KP414213562), TdE, Tdz);				   Te9 = VFNMS(LDK(KP707106781), TdW, TdV);				   TdX = VFMA(LDK(KP707106781), TdW, TdV);				   Ted = VFMA(LDK(KP707106781), Te6, Te5);				   Te7 = VFNMS(LDK(KP707106781), Te6, Te5);				   {					V Tee, Te8, Te4, TdU, TdS, TdO, TdG;					Tee = VADD(Tea, Teb);					Tec = VSUB(Tea, Teb);					Te8 = VSUB(Te3, Te0);					Te4 = VADD(Te0, Te3);					TdU = VADD(TdR, TdQ);					TdS = VSUB(TdQ, TdR);					TdO = VADD(Tdu, TdF);					TdG = VSUB(Tdu, TdF);					TeZ = VFMA(LDK(KP923879532), Tee, Ted);					STM4(&(io[4]), TeZ, ovs, &(io[0]));					Tf0 = VFNMS(LDK(KP923879532), Tee, Ted);					STM4(&(io[36]), Tf0, ovs, &(io[0]));					Tf1 = VFMA(LDK(KP923879532), Te4, TdX);					STM4(&(ro[4]), Tf1, ovs, &(ro[0]));					Tf2 = VFNMS(LDK(KP923879532), Te4, TdX);					STM4(&(ro[36]), Tf2, ovs, &(ro[0]));					Tf3 = VFMA(LDK(KP923879532), TdU, TdT);					STM4(&(ro[60]), Tf3, ovs, &(ro[0]));					Tf4 = VFNMS(LDK(KP923879532), TdU, TdT);					STM4(&(ro[28]), Tf4, ovs, &(ro[0]));					Tf5 = VFMA(LDK(KP923879532), TdS, TdP);					STM4(&(io[12]), Tf5, ovs, &(io[0]));					Tf6 = VFNMS(LDK(KP923879532), TdS, TdP);					STM4(&(io[44]), Tf6, ovs, &(io[0]));					Tf7 = VFMA(LDK(KP923879532), TdO, TdN);					STM4(&(io[60]), Tf7, ovs, &(io[0]));					Tf8 = VFNMS(LDK(KP923879532), TdO, TdN);					STM4(&(io[28]), Tf8, ovs, &(io[0]));					Tf9 = VFMA(LDK(KP923879532), TdG, Tdj);					STM4(&(ro[12]), Tf9, ovs, &(ro[0]));					Tfa = VFNMS(LDK(KP923879532), TdG, Tdj);					STM4(&(ro[44]), Tfa, ovs, &(ro[0]));					Tfb = VFMA(LDK(KP923879532), Te8, Te7);					STM4(&(io[20]), Tfb, ovs, &(io[0]));					Tfc = VFNMS(LDK(KP923879532), Te8, Te7);					STM4(&(io[52]), Tfc, ovs, &(io[0]));				   }			      }			 }			 {			      V TcF, TcE, Tcy, Tcv, TcT, Tco, TcP, Tcd, TcZ, TcD, Td0, Tck, Td4, TcX, Tcr;			      V TcS;			      {				   V Tcc, TcC, Tcg, Tcj, TcV, TcW;				   TbE = VADD(TbC, TbD);				   Tcc = VSUB(TbC, TbD);				   TcC = VSUB(Tan, Tak);				   Tao = VADD(Tak, Tan);				   TcF = VFNMS(LDK(KP414213562), Tce, Tcf);				   Tcg = VFMA(LDK(KP414213562), Tcf, Tce);				   Tcj = VFNMS(LDK(KP414213562), Tci, Tch);				   TcE = VFMA(LDK(KP414213562), Tch, Tci);				   Tcy = VFNMS(LDK(KP707106781), Tcx, Tcw);				   TcV = VFMA(LDK(KP707106781), Tcx, Tcw);				   TcW = VFMA(LDK(KP707106781), Tcu, Tct);				   Tcv = VFNMS(LDK(KP707106781), Tcu, Tct);				   TcT = VFMA(LDK(KP707106781), Tcn, Tcm);				   Tco = VFNMS(LDK(KP707106781), Tcn, Tcm);				   Tfd = VFMA(LDK(KP923879532), Tec, Te9);				   STM4(&(ro[20]), Tfd, ovs, &(ro[0]));				   Tfe = VFNMS(LDK(KP923879532), Tec, Te9);				   STM4(&(ro[52]), Tfe, ovs, &(ro[0]));				   TcP = VFNMS(LDK(KP707106781), Tcc, Tcb);				   Tcd = VFMA(LDK(KP707106781), Tcc, Tcb);				   TcZ = VFNMS(LDK(KP707106781), TcC, TcB);				   TcD = VFMA(LDK(KP707106781), TcC, TcB);				   Td0 = VADD(Tcg, Tcj);				   Tck = VSUB(Tcg, Tcj);				   Td4 = VFMA(LDK(KP198912367), TcV, TcW);				   TcX = VFNMS(LDK(KP198912367), TcW, TcV);				   Tcr = VFNMS(LDK(KP707106781), Tcq, Tcp);				   TcS = VFMA(LDK(KP707106781), Tcq, Tcp);			      }			      {				   V TcJ, Tcl, TcK, Tcs, TcQ, TcG, Td5, TcU, TcL, Tcz;				   TcJ = VFNMS(LDK(KP923879532), Tck, Tcd);				   Tcl = VFMA(LDK(KP923879532), Tck, Tcd);				   TcK = VFNMS(LDK(KP668178637), Tco, Tcr);				   Tcs = VFMA(LDK(KP668178637), Tcr, Tco);				   TcQ = VADD(TcF, TcE);				   TcG = VSUB(TcE, TcF);				   Td5 = VFNMS(LDK(KP198912367), TcS, TcT);				   TcU = VFMA(LDK(KP198912367), TcT, TcS);				   TcL = VFMA(LDK(KP668178637), Tcv, Tcy);				   Tcz = VFNMS(LDK(KP668178637), Tcy, Tcv);				   {					V Td1, Td3, TcR, TcN, TcH, Td2, TcY, TcM, TcO, TcI, TcA, Td6;					Td1 = VFMA(LDK(KP923879532), Td0, TcZ);					Td3 = VFNMS(LDK(KP923879532), Td0, TcZ);					TcR = VFNMS(LDK(KP923879532), TcQ, TcP);					Td7 = VFMA(LDK(KP923879532), TcQ, TcP);					TcN = VFMA(LDK(KP923879532), TcG, TcD);					TcH = VFNMS(LDK(KP923879532), TcG, TcD);					Td2 = VADD(TcU, TcX);					TcY = VSUB(TcU, TcX);					TcM = VSUB(TcK, TcL);					TcO = VADD(TcK, TcL);					TcI = VSUB(Tcz, Tcs);					TcA = VADD(Tcs, Tcz);					Td6 = VSUB(Td4, Td5);					Td8 = VADD(Td5, Td4);					Tff = VFMA(LDK(KP980785280), TcY, TcR);					STM4(&(ro[14]), Tff, ovs, &(ro[0]));					Tfg = VFNMS(LDK(KP980785280), TcY, TcR);					STM4(&(ro[46]), Tfg, ovs, &(ro[0]));					Tfh = VFMA(LDK(KP831469612), TcM, TcJ);					STM4(&(ro[22]), Tfh, ovs, &(ro[0]));					Tfi = VFNMS(LDK(KP831469612), TcM, TcJ);					STM4(&(ro[54]), Tfi, ovs, &(ro[0]));					Tfj = VFMA(LDK(KP831469612), TcO, TcN);					STM4(&(io[6]), Tfj, ovs, &(io[0]));					Tfk = VFNMS(LDK(KP831469612), TcO, TcN);					STM4(&(io[38]), Tfk, ovs, &(io[0]));					Tfl = VFMA(LDK(KP831469612), TcI, TcH);					STM4(&(io[22]), Tfl, ovs, &(io[0]));					Tfm = VFNMS(LDK(KP831469612), TcI, TcH);					STM4(&(io[54]), Tfm, ovs, &(io[0]));					Tfn = VFMA(LDK(KP831469612), TcA, Tcl);					STM4(&(ro[6]), Tfn, ovs, &(ro[0]));					Tfo = VFNMS(LDK(KP831469612), TcA, Tcl);					STM4(&(ro[38]), Tfo, ovs, &(ro[0]));					Tfp = VFMA(LDK(KP980785280), Td6, Td3);					STM4(&(io[14]), Tfp, ovs, &(io[0]));					Tfq = VFNMS(LDK(KP980785280), Td6, Td3);					STM4(&(io[46]), Tfq, ovs, &(io[0]));					Tfr = VFNMS(LDK(KP980785280), Td2, Td1);					STM4(&(io[30]), Tfr, ovs, &(io[0]));					Tfs = VFMA(LDK(KP980785280), Td2, Td1);					STM4(&(io[62]), Tfs, ovs, &(io[0]));				   }			      }			 }		    }		    {			 V Tft, Tfu, Tfv, Tfw, Tfx, Tfy, Tfz, TfA, TfB, TfC, TfD, TfE, TfF, TfG, T3f;			 V T66, T63, T3u, TfL, TfM, TfN, TfO, TfP, TfQ, TfR, TfS, TfT, TfU, TfV, TfW;			 V TfX, TfY, TfZ, Tg0, Tc5, Tc8;			 {			      V TbH, TbG, Tbw, Tbn, TbV, TaW, TbR, Tap, Tc1, TbF, Tc2, TaE, Tc7, TbZ, Tb5;			      V TbU;			      {				   V Taw, TaD, TbX, TbY;				   TbH = VFMA(LDK(KP414213562), Tas, Tav);				   Taw = VFNMS(LDK(KP414213562), Tav, Tas);				   TaD = VFMA(LDK(KP414213562), TaC, Taz);				   TbG = VFNMS(LDK(KP414213562), Taz, TaC);				   Tbw = VFNMS(LDK(KP707106781), Tbv, Tbs);				   TbX = VFMA(LDK(KP707106781), Tbv, Tbs);				   TbY = VFMA(LDK(KP707106781), Tbm, Tbb);				   Tbn = VFNMS(LDK(KP707106781), Tbm, Tbb);				   TbV = VFMA(LDK(KP707106781), TaV, TaK);				   TaW = VFNMS(LDK(KP707106781), TaV, TaK);				   Tft = VFMA(LDK(KP980785280), Td8, Td7);				   STM4(&(ro[62]), Tft, ovs, &(ro[0]));				   Tfu = VFNMS(LDK(KP980785280), Td8, Td7);				   STM4(&(ro[30]), Tfu, ovs, &(ro[0]));				   TbR = VFMA(LDK(KP707106781), Tao, Tah);				   Tap = VFNMS(LDK(KP707106781), Tao, Tah);				   Tc1 = VFMA(LDK(KP707106781), TbE, TbB);				   TbF = VFNMS(LDK(KP707106781), TbE, TbB);