Ejemplo n.º 1
0
static void n1fv_13(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
{
     DVK(KP904176221, +0.904176221990848204433795481776887926501523162);
     DVK(KP575140729, +0.575140729474003121368385547455453388461001608);
     DVK(KP300462606, +0.300462606288665774426601772289207995520941381);
     DVK(KP516520780, +0.516520780623489722840901288569017135705033622);
     DVK(KP522026385, +0.522026385161275033714027226654165028300441940);
     DVK(KP957805992, +0.957805992594665126462521754605754580515587217);
     DVK(KP600477271, +0.600477271932665282925769253334763009352012849);
     DVK(KP251768516, +0.251768516431883313623436926934233488546674281);
     DVK(KP503537032, +0.503537032863766627246873853868466977093348562);
     DVK(KP769338817, +0.769338817572980603471413688209101117038278899);
     DVK(KP859542535, +0.859542535098774820163672132761689612766401925);
     DVK(KP581704778, +0.581704778510515730456870384989698884939833902);
     DVK(KP853480001, +0.853480001859823990758994934970528322872359049);
     DVK(KP083333333, +0.083333333333333333333333333333333333333333333);
     DVK(KP226109445, +0.226109445035782405468510155372505010481906348);
     DVK(KP301479260, +0.301479260047709873958013540496673347309208464);
     DVK(KP686558370, +0.686558370781754340655719594850823015421401653);
     DVK(KP514918778, +0.514918778086315755491789696138117261566051239);
     DVK(KP038632954, +0.038632954644348171955506895830342264440241080);
     DVK(KP612264650, +0.612264650376756543746494474777125408779395514);
     DVK(KP302775637, +0.302775637731994646559610633735247973125648287);
     DVK(KP866025403, +0.866025403784438646763723170752936183471402627);
     DVK(KP500000000, +0.500000000000000000000000000000000000000000000);
     INT i;
     const R *xi;
     R *xo;
     xi = ri;
     xo = ro;
     for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(is), MAKE_VOLATILE_STRIDE(os)) {
	  V T1, T7, T2, Tg, Tf, TN, Th, Tq, Ta, Tj, T5, Tr, Tk;
	  T1 = LD(&(xi[0]), ivs, &(xi[0]));
	  {
	       V Td, Te, T8, T9, T3, T4;
	       Td = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
	       Te = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
	       T7 = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
	       T8 = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
	       T9 = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
	       T2 = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
	       T3 = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
	       T4 = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
	       Tg = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
	       Tf = VADD(Td, Te);
	       TN = VSUB(Td, Te);
	       Th = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
	       Tq = VSUB(T8, T9);
	       Ta = VADD(T8, T9);
	       Tj = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
	       T5 = VADD(T3, T4);
	       Tr = VSUB(T4, T3);
	       Tk = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
	  }
	  {
	       V Tt, Ti, Ty, Tb, Ts, TQ, Tx, T6, Tu, Tl;
	       Tt = VSUB(Tg, Th);
	       Ti = VADD(Tg, Th);
	       Ty = VFMS(LDK(KP500000000), Ta, T7);
	       Tb = VADD(T7, Ta);
	       Ts = VSUB(Tq, Tr);
	       TQ = VADD(Tr, Tq);
	       Tx = VFNMS(LDK(KP500000000), T5, T2);
	       T6 = VADD(T2, T5);
	       Tu = VSUB(Tj, Tk);
	       Tl = VADD(Tj, Tk);
	       {
		    V TK, Tz, Tc, TX, Tv, TO, TL, Tm;
		    TK = VADD(Tx, Ty);
		    Tz = VSUB(Tx, Ty);
		    Tc = VADD(T6, Tb);
		    TX = VSUB(T6, Tb);
		    Tv = VSUB(Tt, Tu);
		    TO = VADD(Tt, Tu);
		    TL = VSUB(Ti, Tl);
		    Tm = VADD(Ti, Tl);
		    {
			 V TF, Tw, TP, TY, TT, TM, TA, Tn;
			 TF = VSUB(Ts, Tv);
			 Tw = VADD(Ts, Tv);
			 TP = VFNMS(LDK(KP500000000), TO, TN);
			 TY = VADD(TN, TO);
			 TT = VFNMS(LDK(KP866025403), TL, TK);
			 TM = VFMA(LDK(KP866025403), TL, TK);
			 TA = VFNMS(LDK(KP500000000), Tm, Tf);
			 Tn = VADD(Tf, Tm);
			 {
			      V T1f, T1n, TI, T18, T1k, T1c, TD, T17, T10, T1m, T16, T1e, TU, TR;
			      TU = VFNMS(LDK(KP866025403), TQ, TP);
			      TR = VFMA(LDK(KP866025403), TQ, TP);
			      {
				   V TZ, T15, TE, TB;
				   TZ = VFMA(LDK(KP302775637), TY, TX);
				   T15 = VFNMS(LDK(KP302775637), TX, TY);
				   TE = VSUB(Tz, TA);
				   TB = VADD(Tz, TA);
				   {
					V TH, To, TV, T13;
					TH = VSUB(Tc, Tn);
					To = VADD(Tc, Tn);
					TV = VFNMS(LDK(KP612264650), TU, TT);
					T13 = VFMA(LDK(KP612264650), TT, TU);
					{
					     V TS, T12, TG, T1b;
					     TS = VFNMS(LDK(KP038632954), TR, TM);
					     T12 = VFMA(LDK(KP038632954), TM, TR);
					     TG = VFNMS(LDK(KP514918778), TF, TE);
					     T1b = VFMA(LDK(KP686558370), TE, TF);
					     {
						  V TC, T1a, Tp, TW, T14;
						  TC = VFMA(LDK(KP301479260), TB, Tw);
						  T1a = VFNMS(LDK(KP226109445), Tw, TB);
						  Tp = VFNMS(LDK(KP083333333), To, T1);
						  ST(&(xo[0]), VADD(T1, To), ovs, &(xo[0]));
						  T1f = VFMA(LDK(KP853480001), TV, TS);
						  TW = VFNMS(LDK(KP853480001), TV, TS);
						  T1n = VFMA(LDK(KP853480001), T13, T12);
						  T14 = VFNMS(LDK(KP853480001), T13, T12);
						  TI = VFMA(LDK(KP581704778), TH, TG);
						  T18 = VFNMS(LDK(KP859542535), TG, TH);
						  T1k = VFMA(LDK(KP769338817), T1b, T1a);
						  T1c = VFNMS(LDK(KP769338817), T1b, T1a);
						  TD = VFMA(LDK(KP503537032), TC, Tp);
						  T17 = VFNMS(LDK(KP251768516), TC, Tp);
						  T10 = VMUL(LDK(KP600477271), VFMA(LDK(KP957805992), TZ, TW));
						  T1m = VFNMS(LDK(KP522026385), TW, TZ);
						  T16 = VMUL(LDK(KP600477271), VFMA(LDK(KP957805992), T15, T14));
						  T1e = VFNMS(LDK(KP522026385), T14, T15);
					     }
					}
				   }
			      }
			      {
				   V T1o, T1q, T1g, T1i, T1d, T1h, T1l, T1p;
				   {
					V T11, TJ, T19, T1j;
					T11 = VFMA(LDK(KP516520780), TI, TD);
					TJ = VFNMS(LDK(KP516520780), TI, TD);
					T19 = VFMA(LDK(KP300462606), T18, T17);
					T1j = VFNMS(LDK(KP300462606), T18, T17);
					T1o = VMUL(LDK(KP575140729), VFNMS(LDK(KP904176221), T1n, T1m));
					T1q = VMUL(LDK(KP575140729), VFMA(LDK(KP904176221), T1n, T1m));
					T1g = VMUL(LDK(KP575140729), VFMA(LDK(KP904176221), T1f, T1e));
					T1i = VMUL(LDK(KP575140729), VFNMS(LDK(KP904176221), T1f, T1e));
					ST(&(xo[WS(os, 12)]), VFNMSI(T16, T11), ovs, &(xo[0]));
					ST(&(xo[WS(os, 1)]), VFMAI(T16, T11), ovs, &(xo[WS(os, 1)]));
					ST(&(xo[WS(os, 8)]), VFMAI(T10, TJ), ovs, &(xo[0]));
					ST(&(xo[WS(os, 5)]), VFNMSI(T10, TJ), ovs, &(xo[WS(os, 1)]));
					T1d = VFNMS(LDK(KP503537032), T1c, T19);
					T1h = VFMA(LDK(KP503537032), T1c, T19);
					T1l = VFNMS(LDK(KP503537032), T1k, T1j);
					T1p = VFMA(LDK(KP503537032), T1k, T1j);
				   }
				   ST(&(xo[WS(os, 9)]), VFMAI(T1g, T1d), ovs, &(xo[WS(os, 1)]));
				   ST(&(xo[WS(os, 4)]), VFNMSI(T1g, T1d), ovs, &(xo[0]));
				   ST(&(xo[WS(os, 10)]), VFNMSI(T1i, T1h), ovs, &(xo[0]));
				   ST(&(xo[WS(os, 3)]), VFMAI(T1i, T1h), ovs, &(xo[WS(os, 1)]));
				   ST(&(xo[WS(os, 7)]), VFMAI(T1o, T1l), ovs, &(xo[WS(os, 1)]));
				   ST(&(xo[WS(os, 6)]), VFNMSI(T1o, T1l), ovs, &(xo[0]));
				   ST(&(xo[WS(os, 11)]), VFMAI(T1q, T1p), ovs, &(xo[WS(os, 1)]));
				   ST(&(xo[WS(os, 2)]), VFNMSI(T1q, T1p), ovs, &(xo[0]));
			      }
			 }
		    }
	       }
	  }
     }
}
Ejemplo n.º 2
0
static void n1fv_13(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
{
     DVK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
     DVK(KP083333333, +0.083333333333333333333333333333333333333333333);
     DVK(KP075902986, +0.075902986037193865983102897245103540356428373);
     DVK(KP251768516, +0.251768516431883313623436926934233488546674281);
     DVK(KP132983124, +0.132983124607418643793760531921092974399165133);
     DVK(KP258260390, +0.258260390311744861420450644284508567852516811);
     DVK(KP1_732050807, +1.732050807568877293527446341505872366942805254);
     DVK(KP300238635, +0.300238635966332641462884626667381504676006424);
     DVK(KP011599105, +0.011599105605768290721655456654083252189827041);
     DVK(KP156891391, +0.156891391051584611046832726756003269660212636);
     DVK(KP256247671, +0.256247671582936600958684654061725059144125175);
     DVK(KP174138601, +0.174138601152135905005660794929264742616964676);
     DVK(KP575140729, +0.575140729474003121368385547455453388461001608);
     DVK(KP503537032, +0.503537032863766627246873853868466977093348562);
     DVK(KP113854479, +0.113854479055790798974654345867655310534642560);
     DVK(KP265966249, +0.265966249214837287587521063842185948798330267);
     DVK(KP387390585, +0.387390585467617292130675966426762851778775217);
     DVK(KP300462606, +0.300462606288665774426601772289207995520941381);
     DVK(KP866025403, +0.866025403784438646763723170752936183471402627);
     DVK(KP500000000, +0.500000000000000000000000000000000000000000000);
     INT i;
     const R *xi;
     R *xo;
     xi = ri;
     xo = ro;
     for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(is), MAKE_VOLATILE_STRIDE(os)) {
	  V TW, Tb, Tm, Tu, TC, TR, TX, TK, TU, Tz, TB, TN, TT;
	  TW = LD(&(xi[0]), ivs, &(xi[0]));
	  {
	       V T3, TH, Tl, Tw, Tp, Tg, Tv, To, T6, Tr, T9, Ts, Ta, TI, T1;
	       V T2, Tq, Tt;
	       T1 = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
	       T2 = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
	       T3 = VSUB(T1, T2);
	       TH = VADD(T1, T2);
	       {
		    V Th, Ti, Tj, Tk;
		    Th = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
		    Ti = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
		    Tj = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
		    Tk = VADD(Ti, Tj);
		    Tl = VADD(Th, Tk);
		    Tw = VSUB(Ti, Tj);
		    Tp = VFNMS(LDK(KP500000000), Tk, Th);
	       }
	       {
		    V Tc, Td, Te, Tf;
		    Tc = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
		    Td = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
		    Te = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
		    Tf = VADD(Td, Te);
		    Tg = VADD(Tc, Tf);
		    Tv = VSUB(Td, Te);
		    To = VFNMS(LDK(KP500000000), Tf, Tc);
	       }
	       {
		    V T4, T5, T7, T8;
		    T4 = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
		    T5 = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
		    T6 = VSUB(T4, T5);
		    Tr = VADD(T4, T5);
		    T7 = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
		    T8 = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
		    T9 = VSUB(T7, T8);
		    Ts = VADD(T7, T8);
	       }
	       Ta = VADD(T6, T9);
	       TI = VADD(Tr, Ts);
	       Tb = VADD(T3, Ta);
	       Tm = VSUB(Tg, Tl);
	       Tq = VSUB(To, Tp);
	       Tt = VMUL(LDK(KP866025403), VSUB(Tr, Ts));
	       Tu = VADD(Tq, Tt);
	       TC = VSUB(Tq, Tt);
	       {
		    V TP, TQ, TG, TJ;
		    TP = VADD(Tg, Tl);
		    TQ = VADD(TH, TI);
		    TR = VMUL(LDK(KP300462606), VSUB(TP, TQ));
		    TX = VADD(TP, TQ);
		    TG = VADD(To, Tp);
		    TJ = VFNMS(LDK(KP500000000), TI, TH);
		    TK = VSUB(TG, TJ);
		    TU = VADD(TG, TJ);
	       }
	       {
		    V Tx, Ty, TL, TM;
		    Tx = VMUL(LDK(KP866025403), VSUB(Tv, Tw));
		    Ty = VFNMS(LDK(KP500000000), Ta, T3);
		    Tz = VSUB(Tx, Ty);
		    TB = VADD(Tx, Ty);
		    TL = VADD(Tv, Tw);
		    TM = VSUB(T6, T9);
		    TN = VSUB(TL, TM);
		    TT = VADD(TL, TM);
	       }
	  }
	  ST(&(xo[0]), VADD(TW, TX), ovs, &(xo[0]));
	  {
	       V T19, T1n, T14, T13, T1f, T1k, Tn, TE, T1e, T1j, TS, T1m, TZ, T1c, TA;
	       V TD;
	       {
		    V T17, T18, T11, T12;
		    T17 = VFMA(LDK(KP387390585), TN, VMUL(LDK(KP265966249), TK));
		    T18 = VFNMS(LDK(KP503537032), TU, VMUL(LDK(KP113854479), TT));
		    T19 = VSUB(T17, T18);
		    T1n = VADD(T17, T18);
		    T14 = VFMA(LDK(KP575140729), Tm, VMUL(LDK(KP174138601), Tb));
		    T11 = VFNMS(LDK(KP156891391), TB, VMUL(LDK(KP256247671), TC));
		    T12 = VFMA(LDK(KP011599105), Tz, VMUL(LDK(KP300238635), Tu));
		    T13 = VSUB(T11, T12);
		    T1f = VADD(T14, T13);
		    T1k = VMUL(LDK(KP1_732050807), VADD(T11, T12));
	       }
	       Tn = VFNMS(LDK(KP174138601), Tm, VMUL(LDK(KP575140729), Tb));
	       TA = VFNMS(LDK(KP300238635), Tz, VMUL(LDK(KP011599105), Tu));
	       TD = VFMA(LDK(KP256247671), TB, VMUL(LDK(KP156891391), TC));
	       TE = VSUB(TA, TD);
	       T1e = VMUL(LDK(KP1_732050807), VADD(TD, TA));
	       T1j = VSUB(Tn, TE);
	       {
		    V TO, T1b, TV, TY, T1a;
		    TO = VFNMS(LDK(KP132983124), TN, VMUL(LDK(KP258260390), TK));
		    T1b = VSUB(TR, TO);
		    TV = VFMA(LDK(KP251768516), TT, VMUL(LDK(KP075902986), TU));
		    TY = VFNMS(LDK(KP083333333), TX, TW);
		    T1a = VSUB(TY, TV);
		    TS = VFMA(LDK(KP2_000000000), TO, TR);
		    T1m = VADD(T1b, T1a);
		    TZ = VFMA(LDK(KP2_000000000), TV, TY);
		    T1c = VSUB(T1a, T1b);
	       }
	       {
		    V TF, T10, T1l, T1o;
		    TF = VBYI(VFMA(LDK(KP2_000000000), TE, Tn));
		    T10 = VADD(TS, TZ);
		    ST(&(xo[WS(os, 1)]), VADD(TF, T10), ovs, &(xo[WS(os, 1)]));
		    ST(&(xo[WS(os, 12)]), VSUB(T10, TF), ovs, &(xo[0]));
		    {
			 V T15, T16, T1p, T1q;
			 T15 = VBYI(VFMS(LDK(KP2_000000000), T13, T14));
			 T16 = VSUB(TZ, TS);
			 ST(&(xo[WS(os, 5)]), VADD(T15, T16), ovs, &(xo[WS(os, 1)]));
			 ST(&(xo[WS(os, 8)]), VSUB(T16, T15), ovs, &(xo[0]));
			 T1p = VADD(T1n, T1m);
			 T1q = VBYI(VADD(T1j, T1k));
			 ST(&(xo[WS(os, 4)]), VSUB(T1p, T1q), ovs, &(xo[0]));
			 ST(&(xo[WS(os, 9)]), VADD(T1q, T1p), ovs, &(xo[WS(os, 1)]));
		    }
		    T1l = VBYI(VSUB(T1j, T1k));
		    T1o = VSUB(T1m, T1n);
		    ST(&(xo[WS(os, 3)]), VADD(T1l, T1o), ovs, &(xo[WS(os, 1)]));
		    ST(&(xo[WS(os, 10)]), VSUB(T1o, T1l), ovs, &(xo[0]));
		    {
			 V T1h, T1i, T1d, T1g;
			 T1h = VBYI(VSUB(T1e, T1f));
			 T1i = VSUB(T1c, T19);
			 ST(&(xo[WS(os, 6)]), VADD(T1h, T1i), ovs, &(xo[0]));
			 ST(&(xo[WS(os, 7)]), VSUB(T1i, T1h), ovs, &(xo[WS(os, 1)]));
			 T1d = VADD(T19, T1c);
			 T1g = VBYI(VADD(T1e, T1f));
			 ST(&(xo[WS(os, 2)]), VSUB(T1d, T1g), ovs, &(xo[0]));
			 ST(&(xo[WS(os, 11)]), VADD(T1g, T1d), ovs, &(xo[WS(os, 1)]));
		    }
	       }
	  }
     }
}
Ejemplo n.º 3
0
static void t3fv_20(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
{
     DVK(KP587785252, +0.587785252292473129168705954639072768597652438);
     DVK(KP951056516, +0.951056516295153572116439333379382143405698634);
     DVK(KP250000000, +0.250000000000000000000000000000000000000000000);
     DVK(KP559016994, +0.559016994374947424102293417182819058860154590);
     {
	  INT m;
	  R *x;
	  x = ri;
	  for (m = mb, W = W + (mb * ((TWVL / VL) * 8)); m < me; m = m + VL, x = x + (VL * ms), W = W + (TWVL * 8), MAKE_VOLATILE_STRIDE(20, rs)) {
	       V T2, T8, T9, TA, T3, Tc, T4, TZ, T18, Tl, Tq, Tx, TU, Td, Te;
	       V T15, Ti, Tt, TJ;
	       T2 = LDW(&(W[0]));
	       T8 = LDW(&(W[TWVL * 2]));
	       T9 = VZMUL(T2, T8);
	       TA = VZMULJ(T2, T8);
	       T3 = LDW(&(W[TWVL * 4]));
	       Tc = VZMULJ(T9, T3);
	       T4 = VZMUL(T2, T3);
	       TZ = VZMUL(T9, T3);
	       T18 = VZMULJ(TA, T3);
	       Tl = VZMULJ(T8, T3);
	       Tq = VZMULJ(T2, T3);
	       Tx = VZMUL(T8, T3);
	       TU = VZMUL(TA, T3);
	       Td = LDW(&(W[TWVL * 6]));
	       Te = VZMULJ(Tc, Td);
	       T15 = VZMULJ(TA, Td);
	       Ti = VZMULJ(T8, Td);
	       Tt = VZMULJ(T2, Td);
	       TJ = VZMULJ(T9, Td);
	       {
		    V T7, TM, T1U, T2d, T1i, T1p, T1q, T1j, Tp, TE, TF, T26, T27, T2b, T1M;
		    V T1P, T1V, TY, T1c, T1d, T23, T24, T2a, T1F, T1I, T1W, TG, T1e;
		    {
			 V T1, TL, T6, TI, TK, T5, TH, T1S, T1T;
			 T1 = LD(&(x[0]), ms, &(x[0]));
			 TK = LD(&(x[WS(rs, 15)]), ms, &(x[WS(rs, 1)]));
			 TL = VZMULJ(TJ, TK);
			 T5 = LD(&(x[WS(rs, 10)]), ms, &(x[0]));
			 T6 = VZMULJ(T4, T5);
			 TH = LD(&(x[WS(rs, 5)]), ms, &(x[WS(rs, 1)]));
			 TI = VZMULJ(Tc, TH);
			 T7 = VSUB(T1, T6);
			 TM = VSUB(TI, TL);
			 T1S = VADD(T1, T6);
			 T1T = VADD(TI, TL);
			 T1U = VSUB(T1S, T1T);
			 T2d = VADD(T1S, T1T);
		    }
		    {
			 V Th, T1K, T14, T1E, T1b, T1H, To, T1N, Tw, T1D, TR, T1L, TX, T1O, TD;
			 V T1G;
			 {
			      V Tb, Tg, Ta, Tf;
			      Ta = LD(&(x[WS(rs, 4)]), ms, &(x[0]));
			      Tb = VZMULJ(T9, Ta);
			      Tf = LD(&(x[WS(rs, 14)]), ms, &(x[0]));
			      Tg = VZMULJ(Te, Tf);
			      Th = VSUB(Tb, Tg);
			      T1K = VADD(Tb, Tg);
			 }
			 {
			      V T11, T13, T10, T12;
			      T10 = LD(&(x[WS(rs, 13)]), ms, &(x[WS(rs, 1)]));
			      T11 = VZMULJ(TZ, T10);
			      T12 = LD(&(x[WS(rs, 3)]), ms, &(x[WS(rs, 1)]));
			      T13 = VZMULJ(T8, T12);
			      T14 = VSUB(T11, T13);
			      T1E = VADD(T11, T13);
			 }
			 {
			      V T17, T1a, T16, T19;
			      T16 = LD(&(x[WS(rs, 17)]), ms, &(x[WS(rs, 1)]));
			      T17 = VZMULJ(T15, T16);
			      T19 = LD(&(x[WS(rs, 7)]), ms, &(x[WS(rs, 1)]));
			      T1a = VZMULJ(T18, T19);
			      T1b = VSUB(T17, T1a);
			      T1H = VADD(T17, T1a);
			 }
			 {
			      V Tk, Tn, Tj, Tm;
			      Tj = LD(&(x[WS(rs, 16)]), ms, &(x[0]));
			      Tk = VZMULJ(Ti, Tj);
			      Tm = LD(&(x[WS(rs, 6)]), ms, &(x[0]));
			      Tn = VZMULJ(Tl, Tm);
			      To = VSUB(Tk, Tn);
			      T1N = VADD(Tk, Tn);
			 }
			 {
			      V Ts, Tv, Tr, Tu;
			      Tr = LD(&(x[WS(rs, 8)]), ms, &(x[0]));
			      Ts = VZMULJ(Tq, Tr);
			      Tu = LD(&(x[WS(rs, 18)]), ms, &(x[0]));
			      Tv = VZMULJ(Tt, Tu);
			      Tw = VSUB(Ts, Tv);
			      T1D = VADD(Ts, Tv);
			 }
			 {
			      V TO, TQ, TN, TP;
			      TN = LD(&(x[WS(rs, 9)]), ms, &(x[WS(rs, 1)]));
			      TO = VZMULJ(T3, TN);
			      TP = LD(&(x[WS(rs, 19)]), ms, &(x[WS(rs, 1)]));
			      TQ = VZMULJ(Td, TP);
			      TR = VSUB(TO, TQ);
			      T1L = VADD(TO, TQ);
			 }
			 {
			      V TT, TW, TS, TV;
			      TS = LD(&(x[WS(rs, 1)]), ms, &(x[WS(rs, 1)]));
			      TT = VZMULJ(T2, TS);
			      TV = LD(&(x[WS(rs, 11)]), ms, &(x[WS(rs, 1)]));
			      TW = VZMULJ(TU, TV);
			      TX = VSUB(TT, TW);
			      T1O = VADD(TT, TW);
			 }
			 {
			      V Tz, TC, Ty, TB;
			      Ty = LD(&(x[WS(rs, 12)]), ms, &(x[0]));
			      Tz = VZMULJ(Tx, Ty);
			      TB = LD(&(x[WS(rs, 2)]), ms, &(x[0]));
			      TC = VZMULJ(TA, TB);
			      TD = VSUB(Tz, TC);
			      T1G = VADD(Tz, TC);
			 }
			 T1i = VSUB(TX, TR);
			 T1p = VSUB(Th, To);
			 T1q = VSUB(Tw, TD);
			 T1j = VSUB(T1b, T14);
			 Tp = VADD(Th, To);
			 TE = VADD(Tw, TD);
			 TF = VADD(Tp, TE);
			 T26 = VADD(T1D, T1E);
			 T27 = VADD(T1G, T1H);
			 T2b = VADD(T26, T27);
			 T1M = VSUB(T1K, T1L);
			 T1P = VSUB(T1N, T1O);
			 T1V = VADD(T1M, T1P);
			 TY = VADD(TR, TX);
			 T1c = VADD(T14, T1b);
			 T1d = VADD(TY, T1c);
			 T23 = VADD(T1K, T1L);
			 T24 = VADD(T1N, T1O);
			 T2a = VADD(T23, T24);
			 T1F = VSUB(T1D, T1E);
			 T1I = VSUB(T1G, T1H);
			 T1W = VADD(T1F, T1I);
		    }
		    TG = VADD(T7, TF);
		    T1e = VBYI(VADD(TM, T1d));
		    ST(&(x[WS(rs, 5)]), VSUB(TG, T1e), ms, &(x[WS(rs, 1)]));
		    ST(&(x[WS(rs, 15)]), VADD(TG, T1e), ms, &(x[WS(rs, 1)]));
		    {
			 V T2c, T2e, T2f, T29, T2i, T25, T28, T2h, T2g;
			 T2c = VMUL(LDK(KP559016994), VSUB(T2a, T2b));
			 T2e = VADD(T2a, T2b);
			 T2f = VFNMS(LDK(KP250000000), T2e, T2d);
			 T25 = VSUB(T23, T24);
			 T28 = VSUB(T26, T27);
			 T29 = VBYI(VFMA(LDK(KP951056516), T25, VMUL(LDK(KP587785252), T28)));
			 T2i = VBYI(VFNMS(LDK(KP587785252), T25, VMUL(LDK(KP951056516), T28)));
			 ST(&(x[0]), VADD(T2d, T2e), ms, &(x[0]));
			 T2h = VSUB(T2f, T2c);
			 ST(&(x[WS(rs, 8)]), VSUB(T2h, T2i), ms, &(x[0]));
			 ST(&(x[WS(rs, 12)]), VADD(T2i, T2h), ms, &(x[0]));
			 T2g = VADD(T2c, T2f);
			 ST(&(x[WS(rs, 4)]), VADD(T29, T2g), ms, &(x[0]));
			 ST(&(x[WS(rs, 16)]), VSUB(T2g, T29), ms, &(x[0]));
		    }
		    {
			 V T1Z, T1X, T1Y, T1R, T22, T1J, T1Q, T21, T20;
			 T1Z = VMUL(LDK(KP559016994), VSUB(T1V, T1W));
			 T1X = VADD(T1V, T1W);
			 T1Y = VFNMS(LDK(KP250000000), T1X, T1U);
			 T1J = VSUB(T1F, T1I);
			 T1Q = VSUB(T1M, T1P);
			 T1R = VBYI(VFNMS(LDK(KP587785252), T1Q, VMUL(LDK(KP951056516), T1J)));
			 T22 = VBYI(VFMA(LDK(KP951056516), T1Q, VMUL(LDK(KP587785252), T1J)));
			 ST(&(x[WS(rs, 10)]), VADD(T1U, T1X), ms, &(x[0]));
			 T21 = VADD(T1Z, T1Y);
			 ST(&(x[WS(rs, 6)]), VSUB(T21, T22), ms, &(x[0]));
			 ST(&(x[WS(rs, 14)]), VADD(T22, T21), ms, &(x[0]));
			 T20 = VSUB(T1Y, T1Z);
			 ST(&(x[WS(rs, 2)]), VADD(T1R, T20), ms, &(x[0]));
			 ST(&(x[WS(rs, 18)]), VSUB(T20, T1R), ms, &(x[0]));
		    }
		    {
			 V T1k, T1r, T1z, T1w, T1o, T1y, T1h, T1v;
			 T1k = VFMA(LDK(KP951056516), T1i, VMUL(LDK(KP587785252), T1j));
			 T1r = VFMA(LDK(KP951056516), T1p, VMUL(LDK(KP587785252), T1q));
			 T1z = VFNMS(LDK(KP587785252), T1p, VMUL(LDK(KP951056516), T1q));
			 T1w = VFNMS(LDK(KP587785252), T1i, VMUL(LDK(KP951056516), T1j));
			 {
			      V T1m, T1n, T1f, T1g;
			      T1m = VFMS(LDK(KP250000000), T1d, TM);
			      T1n = VMUL(LDK(KP559016994), VSUB(T1c, TY));
			      T1o = VADD(T1m, T1n);
			      T1y = VSUB(T1n, T1m);
			      T1f = VMUL(LDK(KP559016994), VSUB(Tp, TE));
			      T1g = VFNMS(LDK(KP250000000), TF, T7);
			      T1h = VADD(T1f, T1g);
			      T1v = VSUB(T1g, T1f);
			 }
			 {
			      V T1l, T1s, T1B, T1C;
			      T1l = VADD(T1h, T1k);
			      T1s = VBYI(VSUB(T1o, T1r));
			      ST(&(x[WS(rs, 19)]), VSUB(T1l, T1s), ms, &(x[WS(rs, 1)]));
			      ST(&(x[WS(rs, 1)]), VADD(T1l, T1s), ms, &(x[WS(rs, 1)]));
			      T1B = VADD(T1v, T1w);
			      T1C = VBYI(VADD(T1z, T1y));
			      ST(&(x[WS(rs, 13)]), VSUB(T1B, T1C), ms, &(x[WS(rs, 1)]));
			      ST(&(x[WS(rs, 7)]), VADD(T1B, T1C), ms, &(x[WS(rs, 1)]));
			 }
			 {
			      V T1t, T1u, T1x, T1A;
			      T1t = VSUB(T1h, T1k);
			      T1u = VBYI(VADD(T1r, T1o));
			      ST(&(x[WS(rs, 11)]), VSUB(T1t, T1u), ms, &(x[WS(rs, 1)]));
			      ST(&(x[WS(rs, 9)]), VADD(T1t, T1u), ms, &(x[WS(rs, 1)]));
			      T1x = VSUB(T1v, T1w);
			      T1A = VBYI(VSUB(T1y, T1z));
			      ST(&(x[WS(rs, 17)]), VSUB(T1x, T1A), ms, &(x[WS(rs, 1)]));
			      ST(&(x[WS(rs, 3)]), VADD(T1x, T1A), ms, &(x[WS(rs, 1)]));
			 }
		    }
	       }
	  }
     }
     VLEAVE();
}
Ejemplo n.º 4
0
static void n2fv_20(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
{
     DVK(KP587785252, +0.587785252292473129168705954639072768597652438);
     DVK(KP951056516, +0.951056516295153572116439333379382143405698634);
     DVK(KP250000000, +0.250000000000000000000000000000000000000000000);
     DVK(KP559016994, +0.559016994374947424102293417182819058860154590);
     {
	  INT i;
	  const R *xi;
	  R *xo;
	  xi = ri;
	  xo = ro;
	  for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(40, is), MAKE_VOLATILE_STRIDE(40, os)) {
	       V T3, T1B, Tm, T1i, TG, TN, TO, TH, T13, T16, T1k, T1u, T1v, T1z, T1r;
	       V T1s, T1y, T1a, T1d, T1j, Ti, TD, TB, TL;
	       {
		    V T1, T2, T1g, Tk, Tl, T1h;
		    T1 = LD(&(xi[0]), ivs, &(xi[0]));
		    T2 = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
		    T1g = VADD(T1, T2);
		    Tk = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
		    Tl = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
		    T1h = VADD(Tk, Tl);
		    T3 = VSUB(T1, T2);
		    T1B = VADD(T1g, T1h);
		    Tm = VSUB(Tk, Tl);
		    T1i = VSUB(T1g, T1h);
	       }
	       {
		    V T6, T18, Tw, T12, Tz, T15, T9, T1b, Td, T11, Tp, T19, Ts, T1c, Tg;
		    V T14;
		    {
			 V T4, T5, Tu, Tv;
			 T4 = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
			 T5 = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
			 T6 = VSUB(T4, T5);
			 T18 = VADD(T4, T5);
			 Tu = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
			 Tv = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
			 Tw = VSUB(Tu, Tv);
			 T12 = VADD(Tu, Tv);
		    }
		    {
			 V Tx, Ty, T7, T8;
			 Tx = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
			 Ty = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
			 Tz = VSUB(Tx, Ty);
			 T15 = VADD(Tx, Ty);
			 T7 = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
			 T8 = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
			 T9 = VSUB(T7, T8);
			 T1b = VADD(T7, T8);
		    }
		    {
			 V Tb, Tc, Tn, To;
			 Tb = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
			 Tc = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
			 Td = VSUB(Tb, Tc);
			 T11 = VADD(Tb, Tc);
			 Tn = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
			 To = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
			 Tp = VSUB(Tn, To);
			 T19 = VADD(Tn, To);
		    }
		    {
			 V Tq, Tr, Te, Tf;
			 Tq = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
			 Tr = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
			 Ts = VSUB(Tq, Tr);
			 T1c = VADD(Tq, Tr);
			 Te = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
			 Tf = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
			 Tg = VSUB(Te, Tf);
			 T14 = VADD(Te, Tf);
		    }
		    TG = VSUB(Ts, Tp);
		    TN = VSUB(T6, T9);
		    TO = VSUB(Td, Tg);
		    TH = VSUB(Tz, Tw);
		    T13 = VSUB(T11, T12);
		    T16 = VSUB(T14, T15);
		    T1k = VADD(T13, T16);
		    T1u = VADD(T11, T12);
		    T1v = VADD(T14, T15);
		    T1z = VADD(T1u, T1v);
		    T1r = VADD(T18, T19);
		    T1s = VADD(T1b, T1c);
		    T1y = VADD(T1r, T1s);
		    T1a = VSUB(T18, T19);
		    T1d = VSUB(T1b, T1c);
		    T1j = VADD(T1a, T1d);
		    {
			 V Ta, Th, Tt, TA;
			 Ta = VADD(T6, T9);
			 Th = VADD(Td, Tg);
			 Ti = VADD(Ta, Th);
			 TD = VMUL(LDK(KP559016994), VSUB(Ta, Th));
			 Tt = VADD(Tp, Ts);
			 TA = VADD(Tw, Tz);
			 TB = VADD(Tt, TA);
			 TL = VMUL(LDK(KP559016994), VSUB(TA, Tt));
		    }
	       }
	       {
		    V T1I, T1J, T1K, T1L, T1N, T1H, Tj, TC;
		    Tj = VADD(T3, Ti);
		    TC = VBYI(VADD(Tm, TB));
		    T1H = VSUB(Tj, TC);
		    STM2(&(xo[10]), T1H, ovs, &(xo[2]));
		    T1I = VADD(Tj, TC);
		    STM2(&(xo[30]), T1I, ovs, &(xo[2]));
		    {
			 V T1A, T1C, T1D, T1x, T1G, T1t, T1w, T1F, T1E, T1M;
			 T1A = VMUL(LDK(KP559016994), VSUB(T1y, T1z));
			 T1C = VADD(T1y, T1z);
			 T1D = VFNMS(LDK(KP250000000), T1C, T1B);
			 T1t = VSUB(T1r, T1s);
			 T1w = VSUB(T1u, T1v);
			 T1x = VBYI(VFMA(LDK(KP951056516), T1t, VMUL(LDK(KP587785252), T1w)));
			 T1G = VBYI(VFNMS(LDK(KP587785252), T1t, VMUL(LDK(KP951056516), T1w)));
			 T1J = VADD(T1B, T1C);
			 STM2(&(xo[0]), T1J, ovs, &(xo[0]));
			 T1F = VSUB(T1D, T1A);
			 T1K = VSUB(T1F, T1G);
			 STM2(&(xo[16]), T1K, ovs, &(xo[0]));
			 T1L = VADD(T1G, T1F);
			 STM2(&(xo[24]), T1L, ovs, &(xo[0]));
			 T1E = VADD(T1A, T1D);
			 T1M = VADD(T1x, T1E);
			 STM2(&(xo[8]), T1M, ovs, &(xo[0]));
			 STN2(&(xo[8]), T1M, T1H, ovs);
			 T1N = VSUB(T1E, T1x);
			 STM2(&(xo[32]), T1N, ovs, &(xo[0]));
		    }
		    {
			 V T1O, T1P, T1R, T1S;
			 {
			      V T1n, T1l, T1m, T1f, T1q, T17, T1e, T1p, T1Q, T1o;
			      T1n = VMUL(LDK(KP559016994), VSUB(T1j, T1k));
			      T1l = VADD(T1j, T1k);
			      T1m = VFNMS(LDK(KP250000000), T1l, T1i);
			      T17 = VSUB(T13, T16);
			      T1e = VSUB(T1a, T1d);
			      T1f = VBYI(VFNMS(LDK(KP587785252), T1e, VMUL(LDK(KP951056516), T17)));
			      T1q = VBYI(VFMA(LDK(KP951056516), T1e, VMUL(LDK(KP587785252), T17)));
			      T1O = VADD(T1i, T1l);
			      STM2(&(xo[20]), T1O, ovs, &(xo[0]));
			      T1p = VADD(T1n, T1m);
			      T1P = VSUB(T1p, T1q);
			      STM2(&(xo[12]), T1P, ovs, &(xo[0]));
			      T1Q = VADD(T1q, T1p);
			      STM2(&(xo[28]), T1Q, ovs, &(xo[0]));
			      STN2(&(xo[28]), T1Q, T1I, ovs);
			      T1o = VSUB(T1m, T1n);
			      T1R = VADD(T1f, T1o);
			      STM2(&(xo[4]), T1R, ovs, &(xo[0]));
			      T1S = VSUB(T1o, T1f);
			      STM2(&(xo[36]), T1S, ovs, &(xo[0]));
			 }
			 {
			      V TI, TP, TX, TU, TM, TW, TF, TT, TK, TE;
			      TI = VFMA(LDK(KP951056516), TG, VMUL(LDK(KP587785252), TH));
			      TP = VFMA(LDK(KP951056516), TN, VMUL(LDK(KP587785252), TO));
			      TX = VFNMS(LDK(KP587785252), TN, VMUL(LDK(KP951056516), TO));
			      TU = VFNMS(LDK(KP587785252), TG, VMUL(LDK(KP951056516), TH));
			      TK = VFMS(LDK(KP250000000), TB, Tm);
			      TM = VADD(TK, TL);
			      TW = VSUB(TL, TK);
			      TE = VFNMS(LDK(KP250000000), Ti, T3);
			      TF = VADD(TD, TE);
			      TT = VSUB(TE, TD);
			      {
				   V TJ, TQ, T1T, T1U;
				   TJ = VADD(TF, TI);
				   TQ = VBYI(VSUB(TM, TP));
				   T1T = VSUB(TJ, TQ);
				   STM2(&(xo[38]), T1T, ovs, &(xo[2]));
				   STN2(&(xo[36]), T1S, T1T, ovs);
				   T1U = VADD(TJ, TQ);
				   STM2(&(xo[2]), T1U, ovs, &(xo[2]));
				   STN2(&(xo[0]), T1J, T1U, ovs);
			      }
			      {
				   V TZ, T10, T1V, T1W;
				   TZ = VADD(TT, TU);
				   T10 = VBYI(VADD(TX, TW));
				   T1V = VSUB(TZ, T10);
				   STM2(&(xo[26]), T1V, ovs, &(xo[2]));
				   STN2(&(xo[24]), T1L, T1V, ovs);
				   T1W = VADD(TZ, T10);
				   STM2(&(xo[14]), T1W, ovs, &(xo[2]));
				   STN2(&(xo[12]), T1P, T1W, ovs);
			      }
			      {
				   V TR, TS, T1X, T1Y;
				   TR = VSUB(TF, TI);
				   TS = VBYI(VADD(TP, TM));
				   T1X = VSUB(TR, TS);
				   STM2(&(xo[22]), T1X, ovs, &(xo[2]));
				   STN2(&(xo[20]), T1O, T1X, ovs);
				   T1Y = VADD(TR, TS);
				   STM2(&(xo[18]), T1Y, ovs, &(xo[2]));
				   STN2(&(xo[16]), T1K, T1Y, ovs);
			      }
			      {
				   V TV, TY, T1Z, T20;
				   TV = VSUB(TT, TU);
				   TY = VBYI(VSUB(TW, TX));
				   T1Z = VSUB(TV, TY);
				   STM2(&(xo[34]), T1Z, ovs, &(xo[2]));
				   STN2(&(xo[32]), T1N, T1Z, ovs);
				   T20 = VADD(TV, TY);
				   STM2(&(xo[6]), T20, ovs, &(xo[2]));
				   STN2(&(xo[4]), T1R, T20, ovs);
			      }
			 }
		    }
	       }
	  }
     }
     VLEAVE();
}
Ejemplo n.º 5
0
static void t1fv_20(R *ri, R *ii, const R *W, stride rs, INT mb, INT me, INT ms)
{
     DVK(KP587785252, +0.587785252292473129168705954639072768597652438);
     DVK(KP951056516, +0.951056516295153572116439333379382143405698634);
     DVK(KP250000000, +0.250000000000000000000000000000000000000000000);
     DVK(KP559016994, +0.559016994374947424102293417182819058860154590);
     {
	  INT m;
	  R *x;
	  x = ri;
	  for (m = mb, W = W + (mb * ((TWVL / VL) * 38)); m < me; m = m + VL, x = x + (VL * ms), W = W + (TWVL * 38), MAKE_VOLATILE_STRIDE(20, rs)) {
	       V T4, Tx, T1B, T1U, TZ, T16, T17, T10, Tf, Tq, Tr, T1N, T1O, T1S, T1t;
	       V T1w, T1C, TI, TT, TU, T1K, T1L, T1R, T1m, T1p, T1D, Ts, TV;
	       {
		    V T1, Tw, T3, Tu, Tv, T2, Tt, T1z, T1A;
		    T1 = LD(&(x[0]), ms, &(x[0]));
		    Tv = LD(&(x[WS(rs, 15)]), ms, &(x[WS(rs, 1)]));
		    Tw = BYTWJ(&(W[TWVL * 28]), Tv);
		    T2 = LD(&(x[WS(rs, 10)]), ms, &(x[0]));
		    T3 = BYTWJ(&(W[TWVL * 18]), T2);
		    Tt = LD(&(x[WS(rs, 5)]), ms, &(x[WS(rs, 1)]));
		    Tu = BYTWJ(&(W[TWVL * 8]), Tt);
		    T4 = VSUB(T1, T3);
		    Tx = VSUB(Tu, Tw);
		    T1z = VADD(T1, T3);
		    T1A = VADD(Tu, Tw);
		    T1B = VSUB(T1z, T1A);
		    T1U = VADD(T1z, T1A);
	       }
	       {
		    V T9, T1r, TN, T1l, TS, T1o, Te, T1u, Tk, T1k, TC, T1s, TH, T1v, Tp;
		    V T1n;
		    {
			 V T6, T8, T5, T7;
			 T5 = LD(&(x[WS(rs, 4)]), ms, &(x[0]));
			 T6 = BYTWJ(&(W[TWVL * 6]), T5);
			 T7 = LD(&(x[WS(rs, 14)]), ms, &(x[0]));
			 T8 = BYTWJ(&(W[TWVL * 26]), T7);
			 T9 = VSUB(T6, T8);
			 T1r = VADD(T6, T8);
		    }
		    {
			 V TK, TM, TJ, TL;
			 TJ = LD(&(x[WS(rs, 13)]), ms, &(x[WS(rs, 1)]));
			 TK = BYTWJ(&(W[TWVL * 24]), TJ);
			 TL = LD(&(x[WS(rs, 3)]), ms, &(x[WS(rs, 1)]));
			 TM = BYTWJ(&(W[TWVL * 4]), TL);
			 TN = VSUB(TK, TM);
			 T1l = VADD(TK, TM);
		    }
		    {
			 V TP, TR, TO, TQ;
			 TO = LD(&(x[WS(rs, 17)]), ms, &(x[WS(rs, 1)]));
			 TP = BYTWJ(&(W[TWVL * 32]), TO);
			 TQ = LD(&(x[WS(rs, 7)]), ms, &(x[WS(rs, 1)]));
			 TR = BYTWJ(&(W[TWVL * 12]), TQ);
			 TS = VSUB(TP, TR);
			 T1o = VADD(TP, TR);
		    }
		    {
			 V Tb, Td, Ta, Tc;
			 Ta = LD(&(x[WS(rs, 16)]), ms, &(x[0]));
			 Tb = BYTWJ(&(W[TWVL * 30]), Ta);
			 Tc = LD(&(x[WS(rs, 6)]), ms, &(x[0]));
			 Td = BYTWJ(&(W[TWVL * 10]), Tc);
			 Te = VSUB(Tb, Td);
			 T1u = VADD(Tb, Td);
		    }
		    {
			 V Th, Tj, Tg, Ti;
			 Tg = LD(&(x[WS(rs, 8)]), ms, &(x[0]));
			 Th = BYTWJ(&(W[TWVL * 14]), Tg);
			 Ti = LD(&(x[WS(rs, 18)]), ms, &(x[0]));
			 Tj = BYTWJ(&(W[TWVL * 34]), Ti);
			 Tk = VSUB(Th, Tj);
			 T1k = VADD(Th, Tj);
		    }
		    {
			 V Tz, TB, Ty, TA;
			 Ty = LD(&(x[WS(rs, 9)]), ms, &(x[WS(rs, 1)]));
			 Tz = BYTWJ(&(W[TWVL * 16]), Ty);
			 TA = LD(&(x[WS(rs, 19)]), ms, &(x[WS(rs, 1)]));
			 TB = BYTWJ(&(W[TWVL * 36]), TA);
			 TC = VSUB(Tz, TB);
			 T1s = VADD(Tz, TB);
		    }
		    {
			 V TE, TG, TD, TF;
			 TD = LD(&(x[WS(rs, 1)]), ms, &(x[WS(rs, 1)]));
			 TE = BYTWJ(&(W[0]), TD);
			 TF = LD(&(x[WS(rs, 11)]), ms, &(x[WS(rs, 1)]));
			 TG = BYTWJ(&(W[TWVL * 20]), TF);
			 TH = VSUB(TE, TG);
			 T1v = VADD(TE, TG);
		    }
		    {
			 V Tm, To, Tl, Tn;
			 Tl = LD(&(x[WS(rs, 12)]), ms, &(x[0]));
			 Tm = BYTWJ(&(W[TWVL * 22]), Tl);
			 Tn = LD(&(x[WS(rs, 2)]), ms, &(x[0]));
			 To = BYTWJ(&(W[TWVL * 2]), Tn);
			 Tp = VSUB(Tm, To);
			 T1n = VADD(Tm, To);
		    }
		    TZ = VSUB(TH, TC);
		    T16 = VSUB(T9, Te);
		    T17 = VSUB(Tk, Tp);
		    T10 = VSUB(TS, TN);
		    Tf = VADD(T9, Te);
		    Tq = VADD(Tk, Tp);
		    Tr = VADD(Tf, Tq);
		    T1N = VADD(T1k, T1l);
		    T1O = VADD(T1n, T1o);
		    T1S = VADD(T1N, T1O);
		    T1t = VSUB(T1r, T1s);
		    T1w = VSUB(T1u, T1v);
		    T1C = VADD(T1t, T1w);
		    TI = VADD(TC, TH);
		    TT = VADD(TN, TS);
		    TU = VADD(TI, TT);
		    T1K = VADD(T1r, T1s);
		    T1L = VADD(T1u, T1v);
		    T1R = VADD(T1K, T1L);
		    T1m = VSUB(T1k, T1l);
		    T1p = VSUB(T1n, T1o);
		    T1D = VADD(T1m, T1p);
	       }
	       Ts = VADD(T4, Tr);
	       TV = VBYI(VADD(Tx, TU));
	       ST(&(x[WS(rs, 5)]), VSUB(Ts, TV), ms, &(x[WS(rs, 1)]));
	       ST(&(x[WS(rs, 15)]), VADD(Ts, TV), ms, &(x[WS(rs, 1)]));
	       {
		    V T1T, T1V, T1W, T1Q, T1Z, T1M, T1P, T1Y, T1X;
		    T1T = VMUL(LDK(KP559016994), VSUB(T1R, T1S));
		    T1V = VADD(T1R, T1S);
		    T1W = VFNMS(LDK(KP250000000), T1V, T1U);
		    T1M = VSUB(T1K, T1L);
		    T1P = VSUB(T1N, T1O);
		    T1Q = VBYI(VFMA(LDK(KP951056516), T1M, VMUL(LDK(KP587785252), T1P)));
		    T1Z = VBYI(VFNMS(LDK(KP587785252), T1M, VMUL(LDK(KP951056516), T1P)));
		    ST(&(x[0]), VADD(T1U, T1V), ms, &(x[0]));
		    T1Y = VSUB(T1W, T1T);
		    ST(&(x[WS(rs, 8)]), VSUB(T1Y, T1Z), ms, &(x[0]));
		    ST(&(x[WS(rs, 12)]), VADD(T1Z, T1Y), ms, &(x[0]));
		    T1X = VADD(T1T, T1W);
		    ST(&(x[WS(rs, 4)]), VADD(T1Q, T1X), ms, &(x[0]));
		    ST(&(x[WS(rs, 16)]), VSUB(T1X, T1Q), ms, &(x[0]));
	       }
	       {
		    V T1G, T1E, T1F, T1y, T1J, T1q, T1x, T1I, T1H;
		    T1G = VMUL(LDK(KP559016994), VSUB(T1C, T1D));
		    T1E = VADD(T1C, T1D);
		    T1F = VFNMS(LDK(KP250000000), T1E, T1B);
		    T1q = VSUB(T1m, T1p);
		    T1x = VSUB(T1t, T1w);
		    T1y = VBYI(VFNMS(LDK(KP587785252), T1x, VMUL(LDK(KP951056516), T1q)));
		    T1J = VBYI(VFMA(LDK(KP951056516), T1x, VMUL(LDK(KP587785252), T1q)));
		    ST(&(x[WS(rs, 10)]), VADD(T1B, T1E), ms, &(x[0]));
		    T1I = VADD(T1G, T1F);
		    ST(&(x[WS(rs, 6)]), VSUB(T1I, T1J), ms, &(x[0]));
		    ST(&(x[WS(rs, 14)]), VADD(T1J, T1I), ms, &(x[0]));
		    T1H = VSUB(T1F, T1G);
		    ST(&(x[WS(rs, 2)]), VADD(T1y, T1H), ms, &(x[0]));
		    ST(&(x[WS(rs, 18)]), VSUB(T1H, T1y), ms, &(x[0]));
	       }
	       {
		    V T11, T18, T1g, T1d, T15, T1f, TY, T1c;
		    T11 = VFMA(LDK(KP951056516), TZ, VMUL(LDK(KP587785252), T10));
		    T18 = VFMA(LDK(KP951056516), T16, VMUL(LDK(KP587785252), T17));
		    T1g = VFNMS(LDK(KP587785252), T16, VMUL(LDK(KP951056516), T17));
		    T1d = VFNMS(LDK(KP587785252), TZ, VMUL(LDK(KP951056516), T10));
		    {
			 V T13, T14, TW, TX;
			 T13 = VFMS(LDK(KP250000000), TU, Tx);
			 T14 = VMUL(LDK(KP559016994), VSUB(TT, TI));
			 T15 = VADD(T13, T14);
			 T1f = VSUB(T14, T13);
			 TW = VMUL(LDK(KP559016994), VSUB(Tf, Tq));
			 TX = VFNMS(LDK(KP250000000), Tr, T4);
			 TY = VADD(TW, TX);
			 T1c = VSUB(TX, TW);
		    }
		    {
			 V T12, T19, T1i, T1j;
			 T12 = VADD(TY, T11);
			 T19 = VBYI(VSUB(T15, T18));
			 ST(&(x[WS(rs, 19)]), VSUB(T12, T19), ms, &(x[WS(rs, 1)]));
			 ST(&(x[WS(rs, 1)]), VADD(T12, T19), ms, &(x[WS(rs, 1)]));
			 T1i = VADD(T1c, T1d);
			 T1j = VBYI(VADD(T1g, T1f));
			 ST(&(x[WS(rs, 13)]), VSUB(T1i, T1j), ms, &(x[WS(rs, 1)]));
			 ST(&(x[WS(rs, 7)]), VADD(T1i, T1j), ms, &(x[WS(rs, 1)]));
		    }
		    {
			 V T1a, T1b, T1e, T1h;
			 T1a = VSUB(TY, T11);
			 T1b = VBYI(VADD(T18, T15));
			 ST(&(x[WS(rs, 11)]), VSUB(T1a, T1b), ms, &(x[WS(rs, 1)]));
			 ST(&(x[WS(rs, 9)]), VADD(T1a, T1b), ms, &(x[WS(rs, 1)]));
			 T1e = VSUB(T1c, T1d);
			 T1h = VBYI(VSUB(T1f, T1g));
			 ST(&(x[WS(rs, 17)]), VSUB(T1e, T1h), ms, &(x[WS(rs, 1)]));
			 ST(&(x[WS(rs, 3)]), VADD(T1e, T1h), ms, &(x[WS(rs, 1)]));
		    }
	       }
	  }
     }
     VLEAVE();
}
static void n1bv_25(const R *ri, const R *ii, R *ro, R *io, stride is, stride os, INT v, INT ivs, INT ovs)
{
     DVK(KP792626838, +0.792626838241819413632131824093538848057784557);
     DVK(KP876091699, +0.876091699473550838204498029706869638173524346);
     DVK(KP803003575, +0.803003575438660414833440593570376004635464850);
     DVK(KP617882369, +0.617882369114440893914546919006756321695042882);
     DVK(KP242145790, +0.242145790282157779872542093866183953459003101);
     DVK(KP968583161, +0.968583161128631119490168375464735813836012403);
     DVK(KP999544308, +0.999544308746292983948881682379742149196758193);
     DVK(KP683113946, +0.683113946453479238701949862233725244439656928);
     DVK(KP559154169, +0.559154169276087864842202529084232643714075927);
     DVK(KP904730450, +0.904730450839922351881287709692877908104763647);
     DVK(KP829049696, +0.829049696159252993975487806364305442437946767);
     DVK(KP831864738, +0.831864738706457140726048799369896829771167132);
     DVK(KP916574801, +0.916574801383451584742370439148878693530976769);
     DVK(KP894834959, +0.894834959464455102997960030820114611498661386);
     DVK(KP809385824, +0.809385824416008241660603814668679683846476688);
     DVK(KP447417479, +0.447417479732227551498980015410057305749330693);
     DVK(KP860541664, +0.860541664367944677098261680920518816412804187);
     DVK(KP897376177, +0.897376177523557693138608077137219684419427330);
     DVK(KP876306680, +0.876306680043863587308115903922062583399064238);
     DVK(KP681693190, +0.681693190061530575150324149145440022633095390);
     DVK(KP560319534, +0.560319534973832390111614715371676131169633784);
     DVK(KP855719849, +0.855719849902058969314654733608091555096772472);
     DVK(KP237294955, +0.237294955877110315393888866460840817927895961);
     DVK(KP949179823, +0.949179823508441261575555465843363271711583843);
     DVK(KP904508497, +0.904508497187473712051146708591409529430077295);
     DVK(KP997675361, +0.997675361079556513670859573984492383596555031);
     DVK(KP262346850, +0.262346850930607871785420028382979691334784273);
     DVK(KP763932022, +0.763932022500210303590826331268723764559381640);
     DVK(KP992114701, +0.992114701314477831049793042785778521453036709);
     DVK(KP690983005, +0.690983005625052575897706582817180941139845410);
     DVK(KP952936919, +0.952936919628306576880750665357914584765951388);
     DVK(KP998026728, +0.998026728428271561952336806863450553336905220);
     DVK(KP570584518, +0.570584518783621657366766175430996792655723863);
     DVK(KP669429328, +0.669429328479476605641803240971985825917022098);
     DVK(KP923225144, +0.923225144846402650453449441572664695995209956);
     DVK(KP906616052, +0.906616052148196230441134447086066874408359177);
     DVK(KP956723877, +0.956723877038460305821989399535483155872969262);
     DVK(KP522616830, +0.522616830205754336872861364785224694908468440);
     DVK(KP945422727, +0.945422727388575946270360266328811958657216298);
     DVK(KP912575812, +0.912575812670962425556968549836277086778922727);
     DVK(KP982009705, +0.982009705009746369461829878184175962711969869);
     DVK(KP921078979, +0.921078979742360627699756128143719920817673854);
     DVK(KP734762448, +0.734762448793050413546343770063151342619912334);
     DVK(KP951056516, +0.951056516295153572116439333379382143405698634);
     DVK(KP958953096, +0.958953096729998668045963838399037225970891871);
     DVK(KP867381224, +0.867381224396525206773171885031575671309956167);
     DVK(KP269969613, +0.269969613759572083574752974412347470060951301);
     DVK(KP244189809, +0.244189809627953270309879511234821255780225091);
     DVK(KP845997307, +0.845997307939530944175097360758058292389769300);
     DVK(KP772036680, +0.772036680810363904029489473607579825330539880);
     DVK(KP132830569, +0.132830569247582714407653942074819768844536507);
     DVK(KP120146378, +0.120146378570687701782758537356596213647956445);
     DVK(KP987388751, +0.987388751065621252324603216482382109400433949);
     DVK(KP893101515, +0.893101515366181661711202267938416198338079437);
     DVK(KP786782374, +0.786782374965295178365099601674911834788448471);
     DVK(KP869845200, +0.869845200362138853122720822420327157933056305);
     DVK(KP447533225, +0.447533225982656890041886979663652563063114397);
     DVK(KP494780565, +0.494780565770515410344588413655324772219443730);
     DVK(KP578046249, +0.578046249379945007321754579646815604023525655);
     DVK(KP522847744, +0.522847744331509716623755382187077770911012542);
     DVK(KP059835404, +0.059835404262124915169548397419498386427871950);
     DVK(KP066152395, +0.066152395967733048213034281011006031460903353);
     DVK(KP603558818, +0.603558818296015001454675132653458027918768137);
     DVK(KP667278218, +0.667278218140296670899089292254759909713898805);
     DVK(KP559016994, +0.559016994374947424102293417182819058860154590);
     DVK(KP250000000, +0.250000000000000000000000000000000000000000000);
     DVK(KP618033988, +0.618033988749894848204586834365638117720309180);
     INT i;
     const R *xi;
     R *xo;
     xi = ii;
     xo = io;
     for (i = v; i > 0; i = i - VL, xi = xi + (VL * ivs), xo = xo + (VL * ovs), MAKE_VOLATILE_STRIDE(is), MAKE_VOLATILE_STRIDE(os)) {
	  V T1g, T1k, T1I, T24, T2a, T1G, T1A, T1l, T1B, T1H, T1d;
	  {
	       V T2z, T1q, Ta, T9, T3n, Ty, Tl, T2O, T2W, T2l, T2s, TV, T1i, T1K, T1S;
	       V T3z, T3t, Tk, T3o, Tp, T2g, T2N, T2V, T2o, T2t, T1a, T1j, T1J, T1R, Tz;
	       V Tt, TA, Tw;
	       {
		    V T1, T5, T6, T2, T3;
		    T1 = LD(&(xi[0]), ivs, &(xi[0]));
		    T5 = LD(&(xi[WS(is, 10)]), ivs, &(xi[0]));
		    T6 = LD(&(xi[WS(is, 15)]), ivs, &(xi[WS(is, 1)]));
		    T2 = LD(&(xi[WS(is, 5)]), ivs, &(xi[WS(is, 1)]));
		    T3 = LD(&(xi[WS(is, 20)]), ivs, &(xi[0]));
		    {
			 V TH, TW, TK, TS, T10, T8, TN, TT, T17, TZ, T11;
			 TH = LD(&(xi[WS(is, 2)]), ivs, &(xi[0]));
			 TW = LD(&(xi[WS(is, 3)]), ivs, &(xi[WS(is, 1)]));
			 {
			      V TI, TJ, TL, T7, T1p, T4, T1o, TM, TX, TY;
			      TI = LD(&(xi[WS(is, 7)]), ivs, &(xi[WS(is, 1)]));
			      TJ = LD(&(xi[WS(is, 22)]), ivs, &(xi[0]));
			      TL = LD(&(xi[WS(is, 12)]), ivs, &(xi[0]));
			      T7 = VADD(T5, T6);
			      T1p = VSUB(T5, T6);
			      T4 = VADD(T2, T3);
			      T1o = VSUB(T2, T3);
			      TM = LD(&(xi[WS(is, 17)]), ivs, &(xi[WS(is, 1)]));
			      TX = LD(&(xi[WS(is, 8)]), ivs, &(xi[0]));
			      TK = VADD(TI, TJ);
			      TS = VSUB(TI, TJ);
			      TY = LD(&(xi[WS(is, 23)]), ivs, &(xi[WS(is, 1)]));
			      T10 = LD(&(xi[WS(is, 13)]), ivs, &(xi[WS(is, 1)]));
			      T2z = VFNMS(LDK(KP618033988), T1o, T1p);
			      T1q = VFMA(LDK(KP618033988), T1p, T1o);
			      Ta = VSUB(T4, T7);
			      T8 = VADD(T4, T7);
			      TN = VADD(TL, TM);
			      TT = VSUB(TM, TL);
			      T17 = VSUB(TX, TY);
			      TZ = VADD(TX, TY);
			      T11 = LD(&(xi[WS(is, 18)]), ivs, &(xi[0]));
			 }
			 {
			      V Tc, T2m, T19, Tn, To, Tr, Tj, T16, T2n, Ts, Tu, Tv;
			      {
				   V TU, T2j, TO, TQ, T12, T18;
				   Tc = LD(&(xi[WS(is, 1)]), ivs, &(xi[WS(is, 1)]));
				   T9 = VFNMS(LDK(KP250000000), T8, T1);
				   T3n = VADD(T1, T8);
				   TU = VFNMS(LDK(KP618033988), TT, TS);
				   T2j = VFMA(LDK(KP618033988), TS, TT);
				   TO = VADD(TK, TN);
				   TQ = VSUB(TN, TK);
				   T12 = VADD(T10, T11);
				   T18 = VSUB(T10, T11);
				   Ty = LD(&(xi[WS(is, 4)]), ivs, &(xi[0]));
				   {
					V T3r, T15, T13, Tf, Ti, T2k, TR, TP, T3s, T14;
					{
					     V Td, Te, Tg, Th;
					     Td = LD(&(xi[WS(is, 6)]), ivs, &(xi[0]));
					     Te = LD(&(xi[WS(is, 21)]), ivs, &(xi[WS(is, 1)]));
					     Tg = LD(&(xi[WS(is, 11)]), ivs, &(xi[WS(is, 1)]));
					     Th = LD(&(xi[WS(is, 16)]), ivs, &(xi[0]));
					     TP = VFNMS(LDK(KP250000000), TO, TH);
					     T3r = VADD(TH, TO);
					     T2m = VFNMS(LDK(KP618033988), T17, T18);
					     T19 = VFMA(LDK(KP618033988), T18, T17);
					     T15 = VSUB(T12, TZ);
					     T13 = VADD(TZ, T12);
					     Tf = VADD(Td, Te);
					     Tn = VSUB(Td, Te);
					     To = VSUB(Th, Tg);
					     Ti = VADD(Tg, Th);
					}
					T2k = VFMA(LDK(KP559016994), TQ, TP);
					TR = VFNMS(LDK(KP559016994), TQ, TP);
					Tr = LD(&(xi[WS(is, 24)]), ivs, &(xi[0]));
					T3s = VADD(TW, T13);
					T14 = VFNMS(LDK(KP250000000), T13, TW);
					Tj = VADD(Tf, Ti);
					Tl = VSUB(Tf, Ti);
					T2O = VFNMS(LDK(KP667278218), T2k, T2j);
					T2W = VFMA(LDK(KP603558818), T2j, T2k);
					T2l = VFMA(LDK(KP066152395), T2k, T2j);
					T2s = VFNMS(LDK(KP059835404), T2j, T2k);
					TV = VFNMS(LDK(KP522847744), TU, TR);
					T1i = VFMA(LDK(KP578046249), TR, TU);
					T1K = VFNMS(LDK(KP494780565), TR, TU);
					T1S = VFMA(LDK(KP447533225), TU, TR);
					T16 = VFNMS(LDK(KP559016994), T15, T14);
					T2n = VFMA(LDK(KP559016994), T15, T14);
					T3z = VSUB(T3r, T3s);
					T3t = VADD(T3r, T3s);
					Ts = LD(&(xi[WS(is, 9)]), ivs, &(xi[WS(is, 1)]));
					Tu = LD(&(xi[WS(is, 19)]), ivs, &(xi[WS(is, 1)]));
					Tv = LD(&(xi[WS(is, 14)]), ivs, &(xi[0]));
				   }
			      }
			      Tk = VFNMS(LDK(KP250000000), Tj, Tc);
			      T3o = VADD(Tc, Tj);
			      Tp = VFNMS(LDK(KP618033988), To, Tn);
			      T2g = VFMA(LDK(KP618033988), Tn, To);
			      T2N = VFMA(LDK(KP066152395), T2n, T2m);
			      T2V = VFNMS(LDK(KP059835404), T2m, T2n);
			      T2o = VFMA(LDK(KP869845200), T2n, T2m);
			      T2t = VFNMS(LDK(KP786782374), T2m, T2n);
			      T1a = VFNMS(LDK(KP893101515), T19, T16);
			      T1j = VFMA(LDK(KP987388751), T16, T19);
			      T1J = VFNMS(LDK(KP120146378), T19, T16);
			      T1R = VFMA(LDK(KP132830569), T16, T19);
			      Tz = VADD(Ts, Tr);
			      Tt = VSUB(Tr, Ts);
			      TA = VADD(Tv, Tu);
			      Tw = VSUB(Tu, Tv);
			 }
		    }
	       }
	       {
		    V T2p, T2I, T2u, T2C, Tx, T2d, T2X, T34, T2P, T3b, T2b, Tb, T2Q, T2Z, T2h;
		    V T2w, Tq, T1e, T1M, T1U, TE, T2c, T3q, T3y;
		    T2p = VFNMS(LDK(KP772036680), T2o, T2l);
		    T2I = VFMA(LDK(KP772036680), T2o, T2l);
		    T2u = VFMA(LDK(KP772036680), T2t, T2s);
		    T2C = VFNMS(LDK(KP772036680), T2t, T2s);
		    {
			 V TD, TB, Tm, T2f, T3p, TC;
			 Tx = VFMA(LDK(KP618033988), Tw, Tt);
			 T2d = VFNMS(LDK(KP618033988), Tt, Tw);
			 TD = VSUB(Tz, TA);
			 TB = VADD(Tz, TA);
			 Tm = VFMA(LDK(KP559016994), Tl, Tk);
			 T2f = VFNMS(LDK(KP559016994), Tl, Tk);
			 T2X = VFMA(LDK(KP845997307), T2W, T2V);
			 T34 = VFNMS(LDK(KP845997307), T2W, T2V);
			 T2P = VFNMS(LDK(KP845997307), T2O, T2N);
			 T3b = VFMA(LDK(KP845997307), T2O, T2N);
			 T2b = VFNMS(LDK(KP559016994), Ta, T9);
			 Tb = VFMA(LDK(KP559016994), Ta, T9);
			 T3p = VADD(Ty, TB);
			 TC = VFMS(LDK(KP250000000), TB, Ty);
			 T2Q = VFNMS(LDK(KP522847744), T2g, T2f);
			 T2Z = VFMA(LDK(KP578046249), T2f, T2g);
			 T2h = VFMA(LDK(KP893101515), T2g, T2f);
			 T2w = VFNMS(LDK(KP987388751), T2f, T2g);
			 Tq = VFNMS(LDK(KP244189809), Tp, Tm);
			 T1e = VFMA(LDK(KP269969613), Tm, Tp);
			 T1M = VFMA(LDK(KP667278218), Tm, Tp);
			 T1U = VFNMS(LDK(KP603558818), Tp, Tm);
			 TE = VFNMS(LDK(KP559016994), TD, TC);
			 T2c = VFMA(LDK(KP559016994), TD, TC);
			 T3q = VADD(T3o, T3p);
			 T3y = VSUB(T3o, T3p);
		    }
		    {
			 V T1Z, T25, T1P, T22, T1X, TG, T1b, T28, T1t, T1y, T1x, T1E, T1Q, T1Y;
			 {
			      V T26, T1L, T1T, TF, T1f, T1W, T3m, T3g, T2M, T2G, T39, T3j, T21, T1O, T20;
			      V T27;
			      T26 = VFMA(LDK(KP867381224), T1K, T1J);
			      T1L = VFNMS(LDK(KP867381224), T1K, T1J);
			      T20 = VFNMS(LDK(KP958953096), T1S, T1R);
			      T1T = VFMA(LDK(KP958953096), T1S, T1R);
			      {
				   V T2R, T2Y, T2e, T2v, T1N, T1V;
				   T2R = VFNMS(LDK(KP494780565), T2c, T2d);
				   T2Y = VFMA(LDK(KP447533225), T2d, T2c);
				   T2e = VFMA(LDK(KP120146378), T2d, T2c);
				   T2v = VFNMS(LDK(KP132830569), T2c, T2d);
				   TF = VFNMS(LDK(KP667278218), TE, Tx);
				   T1f = VFMA(LDK(KP603558818), Tx, TE);
				   T1N = VFMA(LDK(KP869845200), TE, Tx);
				   T1V = VFNMS(LDK(KP786782374), Tx, TE);
				   {
					V T3A, T3C, T3w, T3u;
					T3A = VMUL(LDK(KP951056516), VFMA(LDK(KP618033988), T3z, T3y));
					T3C = VMUL(LDK(KP951056516), VFNMS(LDK(KP618033988), T3y, T3z));
					T3w = VSUB(T3q, T3t);
					T3u = VADD(T3q, T3t);
					{
					     V T2B, T2x, T2H, T2i;
					     T2B = VFMA(LDK(KP734762448), T2w, T2v);
					     T2x = VFNMS(LDK(KP734762448), T2w, T2v);
					     T2H = VFNMS(LDK(KP734762448), T2h, T2e);
					     T2i = VFMA(LDK(KP734762448), T2h, T2e);
					     {
						  V T30, T35, T3c, T2S, T3v;
						  T30 = VFNMS(LDK(KP921078979), T2Z, T2Y);
						  T35 = VFMA(LDK(KP921078979), T2Z, T2Y);
						  T3c = VFMA(LDK(KP982009705), T2R, T2Q);
						  T2S = VFNMS(LDK(KP982009705), T2R, T2Q);
						  T1W = VFMA(LDK(KP912575812), T1V, T1U);
						  T1Z = VFNMS(LDK(KP912575812), T1V, T1U);
						  T1O = VFMA(LDK(KP912575812), T1N, T1M);
						  T25 = VFNMS(LDK(KP912575812), T1N, T1M);
						  ST(&(xo[0]), VADD(T3u, T3n), ovs, &(xo[0]));
						  T3v = VFNMS(LDK(KP250000000), T3u, T3n);
						  {
						       V T2y, T2J, T2q, T2D;
						       T2y = VFMA(LDK(KP945422727), T2x, T2u);
						       T2J = VFMA(LDK(KP522616830), T2x, T2I);
						       T2q = VFMA(LDK(KP956723877), T2p, T2i);
						       T2D = VFNMS(LDK(KP522616830), T2i, T2C);
						       {
							    V T3e, T31, T36, T2T;
							    T3e = VFMA(LDK(KP906616052), T30, T2X);
							    T31 = VFNMS(LDK(KP906616052), T30, T2X);
							    T36 = VFNMS(LDK(KP923225144), T2S, T2P);
							    T2T = VFMA(LDK(KP923225144), T2S, T2P);
							    {
								 V T3k, T3d, T3x, T3B;
								 T3k = VFNMS(LDK(KP669429328), T3b, T3c);
								 T3d = VFMA(LDK(KP570584518), T3c, T3b);
								 T3x = VFMA(LDK(KP559016994), T3w, T3v);
								 T3B = VFNMS(LDK(KP559016994), T3w, T3v);
								 {
								      V T2A, T2K, T2r, T2E;
								      T2A = VMUL(LDK(KP998026728), VFMA(LDK(KP952936919), T2z, T2y));
								      T2K = VFNMS(LDK(KP690983005), T2J, T2u);
								      T2r = VFMA(LDK(KP992114701), T2q, T2b);
								      T2E = VFMA(LDK(KP763932022), T2D, T2p);
								      {
									   V T32, T3a, T37, T3h;
									   T32 = VMUL(LDK(KP998026728), VFNMS(LDK(KP952936919), T2z, T31));
									   T3a = VFMA(LDK(KP262346850), T31, T2z);
									   T37 = VFNMS(LDK(KP997675361), T36, T35);
									   T3h = VFNMS(LDK(KP904508497), T36, T34);
									   {
										V T2U, T33, T3l, T3f;
										T2U = VFMA(LDK(KP949179823), T2T, T2b);
										T33 = VFNMS(LDK(KP237294955), T2T, T2b);
										T3l = VFNMS(LDK(KP669429328), T3e, T3k);
										T3f = VFMA(LDK(KP618033988), T3e, T3d);
										ST(&(xo[WS(os, 20)]), VFNMSI(T3A, T3x), ovs, &(xo[0]));
										ST(&(xo[WS(os, 5)]), VFMAI(T3A, T3x), ovs, &(xo[WS(os, 1)]));
										ST(&(xo[WS(os, 15)]), VFMAI(T3C, T3B), ovs, &(xo[WS(os, 1)]));
										ST(&(xo[WS(os, 10)]), VFNMSI(T3C, T3B), ovs, &(xo[0]));
										{
										     V T2L, T2F, T38, T3i;
										     T2L = VFMA(LDK(KP855719849), T2K, T2H);
										     ST(&(xo[WS(os, 3)]), VFMAI(T2A, T2r), ovs, &(xo[WS(os, 1)]));
										     ST(&(xo[WS(os, 22)]), VFNMSI(T2A, T2r), ovs, &(xo[0]));
										     T2F = VFNMS(LDK(KP855719849), T2E, T2B);
										     T38 = VFMA(LDK(KP560319534), T37, T34);
										     T3i = VFNMS(LDK(KP681693190), T3h, T35);
										     ST(&(xo[WS(os, 2)]), VFMAI(T32, T2U), ovs, &(xo[0]));
										     ST(&(xo[WS(os, 23)]), VFNMSI(T32, T2U), ovs, &(xo[WS(os, 1)]));
										     T3m = VMUL(LDK(KP951056516), VFNMS(LDK(KP876306680), T3l, T3a));
										     T3g = VMUL(LDK(KP951056516), VFNMS(LDK(KP949179823), T3f, T3a));
										     T2M = VMUL(LDK(KP951056516), VFNMS(LDK(KP992114701), T2L, T2z));
										     T2G = VFMA(LDK(KP897376177), T2F, T2b);
										     T39 = VFNMS(LDK(KP949179823), T38, T33);
										     T3j = VFNMS(LDK(KP860541664), T3i, T33);
										     T21 = VFMA(LDK(KP447417479), T1O, T20);
										}
									   }
								      }
								 }
							    }
						       }
						  }
					     }
					}
				   }
			      }
			      T1P = VFNMS(LDK(KP809385824), T1O, T1L);
			      ST(&(xo[WS(os, 17)]), VFNMSI(T2M, T2G), ovs, &(xo[WS(os, 1)]));
			      ST(&(xo[WS(os, 8)]), VFMAI(T2M, T2G), ovs, &(xo[0]));
			      ST(&(xo[WS(os, 13)]), VFMAI(T3g, T39), ovs, &(xo[WS(os, 1)]));
			      ST(&(xo[WS(os, 12)]), VFNMSI(T3g, T39), ovs, &(xo[0]));
			      ST(&(xo[WS(os, 18)]), VFMAI(T3m, T3j), ovs, &(xo[0]));
			      ST(&(xo[WS(os, 7)]), VFNMSI(T3m, T3j), ovs, &(xo[WS(os, 1)]));
			      T22 = VFMA(LDK(KP690983005), T21, T1L);
			      T27 = VFMA(LDK(KP447417479), T1W, T26);
			      T1X = VFMA(LDK(KP894834959), T1W, T1T);
			      {
				   V T1r, T1s, T1v, T1w;
				   T1r = VFNMS(LDK(KP916574801), T1f, T1e);
				   T1g = VFMA(LDK(KP916574801), T1f, T1e);
				   T1k = VFNMS(LDK(KP831864738), T1j, T1i);
				   T1s = VFMA(LDK(KP831864738), T1j, T1i);
				   T1v = VFNMS(LDK(KP829049696), TF, Tq);
				   TG = VFMA(LDK(KP829049696), TF, Tq);
				   T1b = VFMA(LDK(KP831864738), T1a, TV);
				   T1w = VFNMS(LDK(KP831864738), T1a, TV);
				   T28 = VFNMS(LDK(KP763932022), T27, T1T);
				   T1t = VFMA(LDK(KP904730450), T1s, T1r);
				   T1y = VFNMS(LDK(KP904730450), T1s, T1r);
				   T1x = VFMA(LDK(KP559154169), T1w, T1v);
				   T1E = VFNMS(LDK(KP683113946), T1v, T1w);
			      }
			 }
			 T1Q = VFNMS(LDK(KP992114701), T1P, Tb);
			 T1Y = VMUL(LDK(KP951056516), VFNMS(LDK(KP992114701), T1X, T1q));
			 {
			      V T1u, T1F, T1z, T1h, T1c, T23, T29;
			      T23 = VFNMS(LDK(KP999544308), T22, T1Z);
			      T29 = VFNMS(LDK(KP999544308), T28, T25);
			      T1I = VMUL(LDK(KP951056516), VFMA(LDK(KP968583161), T1t, T1q));
			      T1u = VFNMS(LDK(KP242145790), T1t, T1q);
			      T1F = VFMA(LDK(KP617882369), T1y, T1E);
			      T1z = VFMA(LDK(KP559016994), T1y, T1x);
			      T1h = VFNMS(LDK(KP904730450), T1b, TG);
			      T1c = VFMA(LDK(KP904730450), T1b, TG);
			      ST(&(xo[WS(os, 21)]), VFMAI(T1Y, T1Q), ovs, &(xo[WS(os, 1)]));
			      ST(&(xo[WS(os, 4)]), VFNMSI(T1Y, T1Q), ovs, &(xo[0]));
			      T24 = VFNMS(LDK(KP803003575), T23, Tb);
			      T2a = VMUL(LDK(KP951056516), VFNMS(LDK(KP803003575), T29, T1q));
			      T1G = VMUL(LDK(KP951056516), VFNMS(LDK(KP876306680), T1F, T1u));
			      T1A = VMUL(LDK(KP951056516), VFMA(LDK(KP968583161), T1z, T1u));
			      T1l = VFNMS(LDK(KP904730450), T1k, T1h);
			      T1B = VADD(T1g, T1h);
			      T1H = VFMA(LDK(KP968583161), T1c, Tb);
			      T1d = VFNMS(LDK(KP242145790), T1c, Tb);
			 }
		    }
	       }
	  }
	  ST(&(xo[WS(os, 16)]), VFMAI(T2a, T24), ovs, &(xo[0]));
	  ST(&(xo[WS(os, 9)]), VFNMSI(T2a, T24), ovs, &(xo[WS(os, 1)]));
	  {
	       V T1m, T1C, T1n, T1D;
	       T1m = VFNMS(LDK(KP618033988), T1l, T1g);
	       T1C = VFNMS(LDK(KP683113946), T1B, T1k);
	       ST(&(xo[WS(os, 24)]), VFNMSI(T1I, T1H), ovs, &(xo[0]));
	       ST(&(xo[WS(os, 1)]), VFMAI(T1I, T1H), ovs, &(xo[WS(os, 1)]));
	       T1n = VFNMS(LDK(KP876091699), T1m, T1d);
	       T1D = VFMA(LDK(KP792626838), T1C, T1d);
	       ST(&(xo[WS(os, 19)]), VFNMSI(T1A, T1n), ovs, &(xo[WS(os, 1)]));
	       ST(&(xo[WS(os, 6)]), VFMAI(T1A, T1n), ovs, &(xo[0]));
	       ST(&(xo[WS(os, 14)]), VFNMSI(T1G, T1D), ovs, &(xo[0]));
	       ST(&(xo[WS(os, 11)]), VFMAI(T1G, T1D), ovs, &(xo[WS(os, 1)]));
	  }
     }
}