diff options
Diffstat (limited to 'src/invars_calc_4th.cc')
-rw-r--r-- | src/invars_calc_4th.cc | 123 |
1 files changed, 47 insertions, 76 deletions
diff --git a/src/invars_calc_4th.cc b/src/invars_calc_4th.cc index fe92321..b061941 100644 --- a/src/invars_calc_4th.cc +++ b/src/invars_calc_4th.cc @@ -12,6 +12,7 @@ #include "cctk_Parameters.h" #include "GenericFD.h" #include "Differencing.h" +#include "cctk_Loop.h" #include "loopcontrol.h" #include "vectors.h" @@ -61,8 +62,6 @@ static void invars_calc_4th_Body(cGH const * restrict const cctkGH, int const di DECLARE_CCTK_PARAMETERS; - /* Declare finite differencing variables */ - /* Include user-supplied include files */ /* Initialise finite differencing variables */ @@ -93,30 +92,30 @@ static void invars_calc_4th_Body(cGH const * restrict const cctkGH, int const di CCTK_REAL_VEC const p1o12dx = kmul(INV(dx),ToReal(0.0833333333333333333333333333333)); CCTK_REAL_VEC const p1o12dy = kmul(INV(dy),ToReal(0.0833333333333333333333333333333)); CCTK_REAL_VEC const p1o12dz = kmul(INV(dz),ToReal(0.0833333333333333333333333333333)); - CCTK_REAL_VEC const p1o144dxdy = kmul(INV(dx),kmul(INV(dy),ToReal(0.00694444444444444444444444444444))); - CCTK_REAL_VEC const p1o144dxdz = kmul(INV(dx),kmul(INV(dz),ToReal(0.00694444444444444444444444444444))); - CCTK_REAL_VEC const p1o144dydz = kmul(INV(dy),kmul(INV(dz),ToReal(0.00694444444444444444444444444444))); + CCTK_REAL_VEC const p1o144dxdy = kmul(INV(kmul(dx,dy)),ToReal(0.00694444444444444444444444444444)); + CCTK_REAL_VEC const p1o144dxdz = kmul(INV(kmul(dx,dz)),ToReal(0.00694444444444444444444444444444)); + CCTK_REAL_VEC const p1o144dydz = kmul(INV(kmul(dy,dz)),ToReal(0.00694444444444444444444444444444)); CCTK_REAL_VEC const p1o180dx2 = kmul(INV(SQR(dx)),ToReal(0.00555555555555555555555555555556)); CCTK_REAL_VEC const p1o180dy2 = kmul(INV(SQR(dy)),ToReal(0.00555555555555555555555555555556)); CCTK_REAL_VEC const p1o180dz2 = kmul(INV(SQR(dz)),ToReal(0.00555555555555555555555555555556)); CCTK_REAL_VEC const p1o2dx = kmul(INV(dx),ToReal(0.5)); CCTK_REAL_VEC const p1o2dy = kmul(INV(dy),ToReal(0.5)); CCTK_REAL_VEC const p1o2dz = kmul(INV(dz),ToReal(0.5)); - CCTK_REAL_VEC const p1o3600dxdy = kmul(INV(dx),kmul(INV(dy),ToReal(0.000277777777777777777777777777778))); - CCTK_REAL_VEC const p1o3600dxdz = kmul(INV(dx),kmul(INV(dz),ToReal(0.000277777777777777777777777777778))); - CCTK_REAL_VEC const p1o3600dydz = kmul(INV(dy),kmul(INV(dz),ToReal(0.000277777777777777777777777777778))); - CCTK_REAL_VEC const p1o4dxdy = kmul(INV(dx),kmul(INV(dy),ToReal(0.25))); - CCTK_REAL_VEC const p1o4dxdz = kmul(INV(dx),kmul(INV(dz),ToReal(0.25))); - CCTK_REAL_VEC const p1o4dydz = kmul(INV(dy),kmul(INV(dz),ToReal(0.25))); + CCTK_REAL_VEC const p1o3600dxdy = kmul(INV(kmul(dx,dy)),ToReal(0.000277777777777777777777777777778)); + CCTK_REAL_VEC const p1o3600dxdz = kmul(INV(kmul(dx,dz)),ToReal(0.000277777777777777777777777777778)); + CCTK_REAL_VEC const p1o3600dydz = kmul(INV(kmul(dy,dz)),ToReal(0.000277777777777777777777777777778)); + CCTK_REAL_VEC const p1o4dxdy = kmul(INV(kmul(dx,dy)),ToReal(0.25)); + CCTK_REAL_VEC const p1o4dxdz = kmul(INV(kmul(dx,dz)),ToReal(0.25)); + CCTK_REAL_VEC const p1o4dydz = kmul(INV(kmul(dy,dz)),ToReal(0.25)); CCTK_REAL_VEC const p1o5040dx2 = kmul(INV(SQR(dx)),ToReal(0.000198412698412698412698412698413)); CCTK_REAL_VEC const p1o5040dy2 = kmul(INV(SQR(dy)),ToReal(0.000198412698412698412698412698413)); CCTK_REAL_VEC const p1o5040dz2 = kmul(INV(SQR(dz)),ToReal(0.000198412698412698412698412698413)); CCTK_REAL_VEC const p1o60dx = kmul(INV(dx),ToReal(0.0166666666666666666666666666667)); CCTK_REAL_VEC const p1o60dy = kmul(INV(dy),ToReal(0.0166666666666666666666666666667)); CCTK_REAL_VEC const p1o60dz = kmul(INV(dz),ToReal(0.0166666666666666666666666666667)); - CCTK_REAL_VEC const p1o705600dxdy = kmul(INV(dx),kmul(INV(dy),ToReal(1.41723356009070294784580498866e-6))); - CCTK_REAL_VEC const p1o705600dxdz = kmul(INV(dx),kmul(INV(dz),ToReal(1.41723356009070294784580498866e-6))); - CCTK_REAL_VEC const p1o705600dydz = kmul(INV(dy),kmul(INV(dz),ToReal(1.41723356009070294784580498866e-6))); + CCTK_REAL_VEC const p1o705600dxdy = kmul(INV(kmul(dx,dy)),ToReal(1.41723356009070294784580498866e-6)); + CCTK_REAL_VEC const p1o705600dxdz = kmul(INV(kmul(dx,dz)),ToReal(1.41723356009070294784580498866e-6)); + CCTK_REAL_VEC const p1o705600dydz = kmul(INV(kmul(dy,dz)),ToReal(1.41723356009070294784580498866e-6)); CCTK_REAL_VEC const p1o840dx = kmul(INV(dx),ToReal(0.00119047619047619047619047619048)); CCTK_REAL_VEC const p1o840dy = kmul(INV(dy),ToReal(0.00119047619047619047619047619048)); CCTK_REAL_VEC const p1o840dz = kmul(INV(dz),ToReal(0.00119047619047619047619047619048)); @@ -182,7 +181,7 @@ static void invars_calc_4th_Body(cGH const * restrict const cctkGH, int const di /* Loop over the grid points */ #pragma omp parallel - LC_LOOP3VEC (invars_calc_4th, + LC_LOOP3VEC(invars_calc_4th, i,j,k, imin[0],imin[1],imin[2], imax[0],imax[1],imax[2], cctk_lsh[0],cctk_lsh[1],cctk_lsh[2], CCTK_REAL_VEC_SIZE) @@ -243,69 +242,23 @@ static void invars_calc_4th_Body(cGH const * restrict const cctkGH, int const di kmul(kadd(CUB(Psi2rL),kmadd(kmadd(Psi0iL,kmul(Psi3iL,Psi3rL),kmul(Psi1iL,kmul(Psi1rL,Psi4iL))),ToReal(-2),kmadd(Psi4rL,ksub(SQR(Psi1rL),SQR(Psi1iL)),kmadd(Psi0rL,ksub(SQR(Psi3rL),SQR(Psi3iL)),kmadd(Psi2rL,kmadd(Psi0iL,Psi4iL,kmadd(SQR(Psi2iL),ToReal(-3),kmadd(Psi1rL,kmul(Psi3rL,ToReal(-2)),kmsub(Psi1iL,kmul(Psi3iL,ToReal(2)),kmul(Psi0rL,Psi4rL))))),kmul(Psi2iL,kmadd(Psi0rL,Psi4iL,kmadd(Psi0iL,Psi4rL,kmul(kmadd(Psi1rL,Psi3iL,kmul(Psi1iL,Psi3rL)),ToReal(2)))))))))),ToReal(96)); CCTK_REAL_VEC curvJ3L = - kmul(kmadd(SQR(Psi4rL),ksub(SQR(Psi0rL),SQR(Psi0iL)),kmadd(SQR(Psi4iL),ksub(SQR(Psi0iL),SQR(Psi0rL)),kmadd(SQR(Psi2rL),kmul(kmadd(Psi0rL,Psi4rL,kmadd(Psi1rL,kmul(Psi3rL,ToReal(-4)),kmsub(Psi1iL,kmul(Psi3iL,ToReal(4)),kmul(Psi0iL,Psi4iL)))),ToReal(6)),kmadd(kmadd(Psi0rL,kmul(kmadd(Psi1rL,Psi3iL,kmul(Psi1iL,Psi3rL)),Psi4iL),kmul(Psi0iL,kmul(Psi1rL,kmadd(Psi3rL,Psi4iL,kmul(Psi3iL,Psi4rL))))),ToReal(8),kmadd(Psi1iL,kmul(Psi3iL,kmadd(Psi1rL,kmul(Psi3rL,ToReal(-64)),kmadd(Psi0iL,kmul(Psi4iL,ToReal(-8)),kmul(Psi0rL,kmul(Psi4rL,ToReal(8)))))),kmadd(Psi4rL,kmadd(Psi0iL,kmul(Psi0rL,kmul(Psi4iL,ToReal(-4))),kmul(Psi3rL,kmadd(Psi0rL,kmul(Psi1rL,ToReal(-8)),kmul(Psi0iL,kmul(Psi1iL,ToReal(8)))))),kmadd(kadd(QAD(Psi2iL),QAD(Psi2rL)),ToReal(9),kmadd(SQR(Psi2iL),kmul(ToReal(-6),kmadd(Psi0rL,Psi4rL,kmadd(Psi1rL,kmul(Psi3rL,ToReal(-4)),knmsub(Psi0iL,Psi4iL,kmadd(Psi1iL,kmul(Psi3iL,ToReal(4)),kmul(SQR(Psi2rL),ToReal(9))))))),kmadd(Psi2iL,kmul(Psi2rL,kmul(knmsub(Psi0rL,Psi4iL,kmsub(kmadd(Psi1rL,Psi3iL,kmul(Psi1iL,Psi3rL)),ToReal(4),kmul(Psi0iL,Psi4rL))),ToReal(12))),kmadd(SQR(Psi3iL),kmadd(SQR(Psi1rL),ToReal(-16),kmul(SQR(Psi1iL),ToReal(16))),kmul(SQR(Psi3rL),kmadd(SQR(Psi1iL),ToReal(-16),kmul(SQR(Psi1rL),ToReal(16)))))))))))))),ToReal(64)); + kmul(kadd(SQR(kmul(Psi0iL,Psi4iL)),kadd(SQR(kmul(Psi0rL,Psi4rL)),kmadd(kadd(SQR(kmul(Psi1rL,Psi3iL)),SQR(kmul(Psi1iL,Psi3rL))),ToReal(-16),kmadd(Psi0iL,kmul(Psi0rL,kmul(Psi4iL,kmul(Psi4rL,ToReal(-4)))),ksub(ksub(kmadd(SQR(Psi2rL),kmul(kmadd(Psi0rL,Psi4rL,kmadd(Psi1rL,kmul(Psi3rL,ToReal(-4)),kmsub(Psi1iL,kmul(Psi3iL,ToReal(4)),kmul(Psi0iL,Psi4iL)))),ToReal(6)),kmadd(kmadd(Psi0rL,kmul(Psi1rL,kmul(Psi3iL,Psi4iL)),kmadd(Psi0rL,kmul(Psi1iL,kmul(Psi3rL,Psi4iL)),kmadd(Psi0iL,kmul(Psi1rL,kmul(Psi3rL,Psi4iL)),kmul(Psi0iL,kmul(Psi1rL,kmul(Psi3iL,Psi4rL)))))),ToReal(8),kmadd(Psi3rL,kmadd(Psi0rL,kmul(Psi1rL,kmul(Psi4rL,ToReal(-8))),kmul(Psi0iL,kmul(Psi1iL,kmul(Psi4rL,ToReal(8))))),kmadd(Psi1iL,kmadd(Psi1rL,kmul(Psi3iL,kmul(Psi3rL,ToReal(-64))),kmadd(Psi0iL,kmul(Psi3iL,kmul(Psi4iL,ToReal(-8))),kmul(Psi0rL,kmul(Psi3iL,kmul(Psi4rL,ToReal(8)))))),kmadd(kadd(QAD(Psi2iL),QAD(Psi2rL)),ToReal(9),kmadd(SQR(Psi2iL),kmul(ToReal(-6),kmadd(Psi0rL,Psi4rL,kmadd(Psi1rL,kmul(Psi3rL,ToReal(-4)),knmsub(Psi0iL,Psi4iL,kmadd(Psi1iL,kmul(Psi3iL,ToReal(4)),kmul(SQR(Psi2rL),ToReal(9))))))),kmadd(Psi2iL,kmul(Psi2rL,kmul(knmsub(Psi0rL,Psi4iL,knmsub(Psi0iL,Psi4rL,kmadd(Psi1rL,kmul(Psi3iL,ToReal(4)),kmul(Psi1iL,kmul(Psi3rL,ToReal(4)))))),ToReal(12))),kmul(kadd(SQR(kmul(Psi1iL,Psi3iL)),SQR(kmul(Psi1rL,Psi3rL))),ToReal(16))))))))),SQR(kmul(Psi0iL,Psi4rL))),SQR(kmul(Psi0rL,Psi4iL))))))),ToReal(64)); CCTK_REAL_VEC curvJ4L = - kmul(ToReal(-640),kmadd(Psi1iL,kmul(kmadd(kmadd(Psi3rL,Psi4iL,kmul(Psi3iL,Psi4rL)),SQR(Psi1rL),kmul(Psi0rL,kmul(Psi3iL,SQR(Psi3rL)))),ToReal(-12)),kmadd(kmadd(Psi3iL,kmul(Psi4iL,CUB(Psi1rL)),kmul(Psi0iL,kmul(Psi1iL,CUB(Psi3rL)))),ToReal(-4),kmadd(kpow(Psi2rL,5),ToReal(-3),kmadd(Psi4rL,kmul(SQR(Psi3rL),ksub(SQR(Psi0iL),SQR(Psi0rL))),kmadd(Psi0rL,kmadd(SQR(Psi4iL),ksub(SQR(Psi1rL),SQR(Psi1iL)),kmul(SQR(Psi4rL),ksub(SQR(Psi1iL),SQR(Psi1rL)))),kmadd(kmadd(kmadd(Psi3rL,Psi4iL,kmul(Psi3iL,Psi4rL)),CUB(Psi1iL),kmadd(Psi3rL,kmul(Psi4rL,CUB(Psi1rL)),kmadd(Psi0rL,kmadd(kmadd(Psi0iL,kmul(Psi3iL,Psi3rL),kmul(Psi1iL,kmul(Psi1rL,Psi4iL))),Psi4rL,kmul(Psi1iL,CUB(Psi3iL))),kmul(Psi1rL,kmadd(Psi0iL,CUB(Psi3iL),kmul(Psi0rL,CUB(Psi3rL))))))),ToReal(4),kmadd(ToReal(-2),kmadd(Psi4iL,kmadd(Psi3iL,kmul(Psi3rL,SQR(Psi0iL)),kmul(Psi0iL,kmadd(Psi4rL,SQR(Psi1iL),kmul(Psi0rL,SQR(Psi3iL))))),kmadd(Psi0iL,kmul(Psi1iL,kmul(Psi1rL,SQR(Psi4iL))),kmul(CUB(Psi2rL),kmadd(Psi0iL,Psi4iL,kmadd(Psi1rL,kmul(Psi3rL,ToReal(-5)),kmsub(Psi1iL,kmul(Psi3iL,ToReal(5)),kmul(Psi0rL,Psi4rL))))))),kmadd(ToReal(2),kmadd(Psi4iL,kmadd(Psi3iL,kmul(Psi3rL,SQR(Psi0rL)),kmul(Psi0iL,kmadd(Psi4rL,SQR(Psi1rL),kmul(Psi0rL,SQR(Psi3rL))))),kmadd(CUB(Psi2iL),kmadd(Psi0rL,Psi4iL,kmadd(Psi0iL,Psi4rL,kmul(kmadd(Psi1rL,Psi3iL,kmul(Psi1iL,Psi3rL)),ToReal(5)))),kmul(Psi2iL,kmadd(Psi3rL,kmadd(Psi1iL,kmul(SQR(Psi2rL),ToReal(-15)),kmul(Psi3iL,kmul(SQR(Psi1iL),ToReal(-8)))),kmadd(Psi4iL,kmadd(Psi2rL,kmul(SQR(Psi1iL),ToReal(-3)),kmul(Psi4rL,ksub(SQR(Psi0iL),SQR(Psi0rL)))),knmsub(Psi0iL,kmadd(Psi1iL,kmadd(Psi3rL,Psi4iL,kmul(Psi3iL,Psi4rL)),kmadd(Psi2rL,kmul(SQR(Psi3rL),ToReal(-3)),kmadd(Psi0rL,ksub(SQR(Psi4rL),SQR(Psi4iL)),kmul(kmadd(Psi4rL,SQR(Psi2rL),kmul(Psi2rL,SQR(Psi3iL))),ToReal(3))))),kmadd(Psi0rL,kmadd(Psi1iL,kmul(Psi3rL,Psi4rL),kmadd(Psi4iL,kmul(SQR(Psi2rL),ToReal(-3)),kmul(Psi3iL,kmsub(Psi2rL,kmul(Psi3rL,ToReal(6)),kmul(Psi1iL,Psi4iL))))),kmadd(SQR(Psi1rL),kmadd(Psi2rL,kmul(Psi4iL,ToReal(3)),kmul(Psi3iL,kmul(Psi3rL,ToReal(8)))),kmul(Psi1rL,kmadd(Psi3iL,kmul(SQR(Psi2rL),ToReal(-15)),kmadd(Psi4iL,kmsub(Psi0rL,Psi3rL,kmul(Psi0iL,Psi3iL)),kmadd(Psi4rL,kmadd(Psi0rL,Psi3iL,kmadd(Psi0iL,Psi3rL,kmul(Psi1iL,kmul(Psi2rL,ToReal(6))))),kmul(Psi1iL,kmadd(SQR(Psi3iL),ToReal(-8),kmul(SQR(Psi3rL),ToReal(8)))))))))))))))),kmadd(ToReal(3),kmadd(SQR(Psi2rL),kmadd(Psi4rL,ksub(SQR(Psi1iL),SQR(Psi1rL)),kmadd(Psi0rL,ksub(SQR(Psi3iL),SQR(Psi3rL)),kmul(kmadd(Psi0iL,kmul(Psi3iL,Psi3rL),kmul(Psi1iL,kmul(Psi1rL,Psi4iL))),ToReal(2)))),kmul(SQR(Psi2iL),kmadd(kmadd(Psi0iL,kmul(Psi3iL,Psi3rL),kmul(Psi1iL,kmul(Psi1rL,Psi4iL))),ToReal(-2),kmadd(Psi4rL,ksub(SQR(Psi1rL),SQR(Psi1iL)),kmadd(Psi0rL,ksub(SQR(Psi3rL),SQR(Psi3iL)),kmadd(Psi2rL,kmul(ToReal(2),kmadd(Psi0iL,Psi4iL,kmadd(Psi1rL,kmul(Psi3rL,ToReal(-5)),kmsub(Psi1iL,kmul(Psi3iL,ToReal(5)),kmul(Psi0rL,Psi4rL))))),kmul(CUB(Psi2rL),ToReal(10)))))))),kmadd(SQR(Psi3iL),kmadd(Psi4rL,ksub(SQR(Psi0rL),SQR(Psi0iL)),kmul(Psi0iL,kmul(Psi1iL,kmul(Psi3rL,ToReal(12))))),kmadd(Psi1rL,kmadd(kmadd(Psi0rL,kmul(Psi3rL,SQR(Psi3iL)),kmul(Psi0iL,kmul(Psi3iL,SQR(Psi3rL)))),ToReal(-12),kmadd(Psi0iL,kmul(Psi1iL,kmul(SQR(Psi4rL),ToReal(2))),kmul(SQR(Psi1iL),kmadd(Psi3rL,kmul(Psi4rL,ToReal(-12)),kmul(Psi3iL,kmul(Psi4iL,ToReal(12))))))),kmul(Psi2rL,kmadd(QAD(Psi2iL),ToReal(-15),kmadd(Psi0iL,kmul(Psi0rL,kmul(Psi4iL,kmul(Psi4rL,ToReal(-4)))),kmadd(SQR(Psi4rL),ksub(SQR(Psi0rL),SQR(Psi0iL)),kmadd(SQR(Psi4iL),ksub(SQR(Psi0iL),SQR(Psi0rL)),kmadd(ksub(SQR(Psi3iL),SQR(Psi3rL)),kmadd(SQR(Psi1iL),ToReal(-8),kmul(SQR(Psi1rL),ToReal(8))),kmul(ToReal(2),kmadd(Psi1rL,kmadd(kmadd(Psi0rL,Psi3iL,kmul(Psi0iL,Psi3rL)),Psi4iL,kmul(Psi4rL,kmsub(Psi0iL,Psi3iL,kmul(Psi0rL,Psi3rL)))),kmul(Psi1iL,kmadd(Psi0iL,kmul(Psi3rL,Psi4rL),kmadd(Psi0rL,kmadd(Psi3rL,Psi4iL,kmul(Psi3iL,Psi4rL)),kmul(Psi3iL,kmsub(Psi1rL,kmul(Psi3rL,ToReal(16)),kmul(Psi0iL,Psi4iL)))))))))))))))))))))))))); - - /* If necessary, store only partial vectors after the first iteration */ - - if (CCTK_REAL_VEC_SIZE > 2 && CCTK_BUILTIN_EXPECT(i < lc_imin && i+CCTK_REAL_VEC_SIZE > lc_imax, 0)) - { - ptrdiff_t const elt_count_lo = lc_imin-i; - ptrdiff_t const elt_count_hi = lc_imax-i; - vec_store_nta_partial_mid(curvIi[index],curvIiL,elt_count_lo,elt_count_hi); - vec_store_nta_partial_mid(curvIr[index],curvIrL,elt_count_lo,elt_count_hi); - vec_store_nta_partial_mid(curvJ1[index],curvJ1L,elt_count_lo,elt_count_hi); - vec_store_nta_partial_mid(curvJ2[index],curvJ2L,elt_count_lo,elt_count_hi); - vec_store_nta_partial_mid(curvJ3[index],curvJ3L,elt_count_lo,elt_count_hi); - vec_store_nta_partial_mid(curvJ4[index],curvJ4L,elt_count_lo,elt_count_hi); - vec_store_nta_partial_mid(curvJi[index],curvJiL,elt_count_lo,elt_count_hi); - vec_store_nta_partial_mid(curvJr[index],curvJrL,elt_count_lo,elt_count_hi); - break; - } - - /* If necessary, store only partial vectors after the first iteration */ + kmul(ToReal(-640),kmadd(Psi1iL,kmul(kmadd(kmadd(Psi3rL,Psi4iL,kmul(Psi3iL,Psi4rL)),SQR(Psi1rL),kmul(Psi0rL,kmul(Psi3iL,SQR(Psi3rL)))),ToReal(-12)),kmadd(kmadd(Psi3iL,kmul(Psi4iL,CUB(Psi1rL)),kmul(Psi0iL,kmul(Psi1iL,CUB(Psi3rL)))),ToReal(-4),kmadd(kpow(Psi2rL,5),ToReal(-3),kmadd(Psi4rL,kadd(SQR(kmul(Psi0rL,Psi3iL)),ksub(SQR(kmul(Psi0iL,Psi3rL)),kadd(SQR(kmul(Psi0rL,Psi3rL)),SQR(kmul(Psi0iL,Psi3iL))))),kmadd(Psi0rL,kadd(SQR(kmul(Psi1rL,Psi4iL)),ksub(SQR(kmul(Psi1iL,Psi4rL)),kadd(SQR(kmul(Psi1rL,Psi4rL)),SQR(kmul(Psi1iL,Psi4iL))))),kmadd(kmadd(kmadd(Psi3rL,Psi4iL,kmul(Psi3iL,Psi4rL)),CUB(Psi1iL),kmadd(Psi3rL,kmul(Psi4rL,CUB(Psi1rL)),kmadd(Psi0rL,kmadd(kmadd(Psi0iL,kmul(Psi3iL,Psi3rL),kmul(Psi1iL,kmul(Psi1rL,Psi4iL))),Psi4rL,kmul(Psi1iL,CUB(Psi3iL))),kmul(Psi1rL,kmadd(Psi0iL,CUB(Psi3iL),kmul(Psi0rL,CUB(Psi3rL))))))),ToReal(4),kmadd(ToReal(-2),kmadd(Psi4iL,kmadd(Psi3iL,kmul(Psi3rL,SQR(Psi0iL)),kmul(Psi0iL,kmadd(Psi4rL,SQR(Psi1iL),kmul(Psi0rL,SQR(Psi3iL))))),kmul(CUB(Psi2rL),kmadd(Psi0iL,Psi4iL,kmadd(Psi1rL,kmul(Psi3rL,ToReal(-5)),kmsub(Psi1iL,kmul(Psi3iL,ToReal(5)),kmul(Psi0rL,Psi4rL)))))),kmadd(ToReal(2),kmadd(Psi4iL,kmadd(Psi3iL,kmul(Psi3rL,SQR(Psi0rL)),kmul(Psi0iL,kmadd(Psi4rL,SQR(Psi1rL),kmul(Psi0rL,SQR(Psi3rL))))),kmadd(CUB(Psi2iL),kmadd(Psi0rL,Psi4iL,kmadd(Psi0iL,Psi4rL,kmul(kmadd(Psi1rL,Psi3iL,kmul(Psi1iL,Psi3rL)),ToReal(5)))),kmul(Psi2iL,kmadd(Psi3rL,kmadd(Psi1iL,kmul(SQR(Psi2rL),ToReal(-15)),kmul(Psi3iL,kmul(SQR(Psi1iL),ToReal(-8)))),kmadd(Psi4iL,kmadd(Psi2rL,kmul(SQR(Psi1iL),ToReal(-3)),kmul(Psi4rL,ksub(SQR(Psi0iL),SQR(Psi0rL)))),knmsub(Psi0iL,kmadd(Psi1iL,kmadd(Psi3rL,Psi4iL,kmul(Psi3iL,Psi4rL)),kmadd(Psi2rL,kmul(SQR(Psi3rL),ToReal(-3)),kmadd(Psi0rL,ksub(SQR(Psi4rL),SQR(Psi4iL)),kmul(kmadd(Psi4rL,SQR(Psi2rL),kmul(Psi2rL,SQR(Psi3iL))),ToReal(3))))),kmadd(Psi0rL,kmadd(Psi1iL,kmul(Psi3rL,Psi4rL),kmadd(Psi4iL,kmul(SQR(Psi2rL),ToReal(-3)),kmul(Psi3iL,kmsub(Psi2rL,kmul(Psi3rL,ToReal(6)),kmul(Psi1iL,Psi4iL))))),kmadd(SQR(Psi1rL),kmadd(Psi2rL,kmul(Psi4iL,ToReal(3)),kmul(Psi3iL,kmul(Psi3rL,ToReal(8)))),kmul(Psi1rL,kmadd(Psi3iL,kmul(SQR(Psi2rL),ToReal(-15)),kmadd(Psi4iL,kmsub(Psi0rL,Psi3rL,kmul(Psi0iL,Psi3iL)),kmadd(Psi4rL,kmadd(Psi0rL,Psi3iL,kmadd(Psi0iL,Psi3rL,kmul(Psi1iL,kmul(Psi2rL,ToReal(6))))),kmul(Psi1iL,kmadd(SQR(Psi3iL),ToReal(-8),kmul(SQR(Psi3rL),ToReal(8)))))))))))))))),kmadd(ToReal(3),kmadd(SQR(Psi2rL),kmadd(Psi4rL,ksub(SQR(Psi1iL),SQR(Psi1rL)),kmadd(Psi0rL,ksub(SQR(Psi3iL),SQR(Psi3rL)),kmul(kmadd(Psi0iL,kmul(Psi3iL,Psi3rL),kmul(Psi1iL,kmul(Psi1rL,Psi4iL))),ToReal(2)))),kmul(SQR(Psi2iL),kmadd(kmadd(Psi0iL,kmul(Psi3iL,Psi3rL),kmul(Psi1iL,kmul(Psi1rL,Psi4iL))),ToReal(-2),kmadd(Psi4rL,ksub(SQR(Psi1rL),SQR(Psi1iL)),kmadd(Psi0rL,ksub(SQR(Psi3rL),SQR(Psi3iL)),kmadd(Psi2rL,kmul(ToReal(2),kmadd(Psi0iL,Psi4iL,kmadd(Psi1rL,kmul(Psi3rL,ToReal(-5)),kmsub(Psi1iL,kmul(Psi3iL,ToReal(5)),kmul(Psi0rL,Psi4rL))))),kmul(CUB(Psi2rL),ToReal(10)))))))),kmadd(Psi1rL,kmadd(Psi0rL,kmul(Psi3rL,kmul(SQR(Psi3iL),ToReal(-12))),kmadd(Psi0iL,kmul(Psi1iL,kmul(SQR(Psi4iL),ToReal(-2))),kmul(SQR(Psi1iL),kmadd(Psi3rL,kmul(Psi4rL,ToReal(-12)),kmul(Psi3iL,kmul(Psi4iL,ToReal(12))))))),kmadd(Psi0iL,kmadd(Psi1rL,kmul(Psi3iL,kmul(SQR(Psi3rL),ToReal(-12))),kmul(Psi1iL,kmadd(Psi1rL,kmul(SQR(Psi4rL),ToReal(2)),kmul(Psi3rL,kmul(SQR(Psi3iL),ToReal(12)))))),kmul(Psi2rL,kadd(SQR(kmul(Psi0iL,Psi4iL)),kadd(SQR(kmul(Psi0rL,Psi4rL)),kmadd(QAD(Psi2iL),ToReal(-15),kmadd(Psi0iL,kmul(Psi0rL,kmul(Psi4iL,kmul(Psi4rL,ToReal(-4)))),ksub(ksub(kmadd(ksub(SQR(Psi3iL),SQR(Psi3rL)),kmadd(SQR(Psi1iL),ToReal(-8),kmul(SQR(Psi1rL),ToReal(8))),kmul(ToReal(2),kmadd(Psi1rL,kmadd(Psi0iL,kmul(Psi3rL,Psi4iL),kmadd(Psi0iL,kmul(Psi3iL,Psi4rL),kmul(Psi0rL,kmsub(Psi3iL,Psi4iL,kmul(Psi3rL,Psi4rL))))),kmul(Psi1iL,kmadd(Psi0rL,kmul(Psi3rL,Psi4iL),kmadd(Psi0rL,kmul(Psi3iL,Psi4rL),kmadd(Psi0iL,kmul(Psi3rL,Psi4rL),kmul(Psi3iL,kmsub(Psi1rL,kmul(Psi3rL,ToReal(16)),kmul(Psi0iL,Psi4iL)))))))))),SQR(kmul(Psi0iL,Psi4rL))),SQR(kmul(Psi0rL,Psi4iL)))))))))))))))))))); - if (CCTK_REAL_VEC_SIZE > 1 && CCTK_BUILTIN_EXPECT(i < lc_imin, 0)) - { - ptrdiff_t const elt_count = lc_imin-i; - vec_store_nta_partial_hi(curvIi[index],curvIiL,elt_count); - vec_store_nta_partial_hi(curvIr[index],curvIrL,elt_count); - vec_store_nta_partial_hi(curvJ1[index],curvJ1L,elt_count); - vec_store_nta_partial_hi(curvJ2[index],curvJ2L,elt_count); - vec_store_nta_partial_hi(curvJ3[index],curvJ3L,elt_count); - vec_store_nta_partial_hi(curvJ4[index],curvJ4L,elt_count); - vec_store_nta_partial_hi(curvJi[index],curvJiL,elt_count); - vec_store_nta_partial_hi(curvJr[index],curvJrL,elt_count); - continue; - } - - /* If necessary, store only partial vectors after the last iteration */ - - if (CCTK_REAL_VEC_SIZE > 1 && CCTK_BUILTIN_EXPECT(i+CCTK_REAL_VEC_SIZE > lc_imax, 0)) - { - ptrdiff_t const elt_count = lc_imax-i; - vec_store_nta_partial_lo(curvIi[index],curvIiL,elt_count); - vec_store_nta_partial_lo(curvIr[index],curvIrL,elt_count); - vec_store_nta_partial_lo(curvJ1[index],curvJ1L,elt_count); - vec_store_nta_partial_lo(curvJ2[index],curvJ2L,elt_count); - vec_store_nta_partial_lo(curvJ3[index],curvJ3L,elt_count); - vec_store_nta_partial_lo(curvJ4[index],curvJ4L,elt_count); - vec_store_nta_partial_lo(curvJi[index],curvJiL,elt_count); - vec_store_nta_partial_lo(curvJr[index],curvJrL,elt_count); - break; - } - vec_store_nta(curvIi[index],curvIiL); - vec_store_nta(curvIr[index],curvIrL); - vec_store_nta(curvJ1[index],curvJ1L); - vec_store_nta(curvJ2[index],curvJ2L); - vec_store_nta(curvJ3[index],curvJ3L); - vec_store_nta(curvJ4[index],curvJ4L); - vec_store_nta(curvJi[index],curvJiL); - vec_store_nta(curvJr[index],curvJrL); + /* Copy local copies back to grid functions */ + vec_store_partial_prepare(i,lc_imin,lc_imax); + vec_store_nta_partial(curvIi[index],curvIiL); + vec_store_nta_partial(curvIr[index],curvIrL); + vec_store_nta_partial(curvJ1[index],curvJ1L); + vec_store_nta_partial(curvJ2[index],curvJ2L); + vec_store_nta_partial(curvJ3[index],curvJ3L); + vec_store_nta_partial(curvJ4[index],curvJ4L); + vec_store_nta_partial(curvJi[index],curvJiL); + vec_store_nta_partial(curvJr[index],curvJrL); } - LC_ENDLOOP3VEC (invars_calc_4th); + LC_ENDLOOP3VEC(invars_calc_4th); } extern "C" void invars_calc_4th(CCTK_ARGUMENTS) @@ -324,7 +277,25 @@ extern "C" void invars_calc_4th(CCTK_ARGUMENTS) return; } - const char *groups[] = {"WeylScal4::curvIi_group","WeylScal4::curvIr_group","WeylScal4::curvJ1_group","WeylScal4::curvJ2_group","WeylScal4::curvJ3_group","WeylScal4::curvJ4_group","WeylScal4::curvJi_group","WeylScal4::curvJr_group","WeylScal4::Psi0i_group","WeylScal4::Psi0r_group","WeylScal4::Psi1i_group","WeylScal4::Psi1r_group","WeylScal4::Psi2i_group","WeylScal4::Psi2r_group","WeylScal4::Psi3i_group","WeylScal4::Psi3r_group","WeylScal4::Psi4i_group","WeylScal4::Psi4r_group"}; + const char *const groups[] = { + "WeylScal4::curvIi_group", + "WeylScal4::curvIr_group", + "WeylScal4::curvJ1_group", + "WeylScal4::curvJ2_group", + "WeylScal4::curvJ3_group", + "WeylScal4::curvJ4_group", + "WeylScal4::curvJi_group", + "WeylScal4::curvJr_group", + "WeylScal4::Psi0i_group", + "WeylScal4::Psi0r_group", + "WeylScal4::Psi1i_group", + "WeylScal4::Psi1r_group", + "WeylScal4::Psi2i_group", + "WeylScal4::Psi2r_group", + "WeylScal4::Psi3i_group", + "WeylScal4::Psi3r_group", + "WeylScal4::Psi4i_group", + "WeylScal4::Psi4r_group"}; GenericFD_AssertGroupStorage(cctkGH, "invars_calc_4th", 18, groups); switch(fdOrder) @@ -342,7 +313,7 @@ extern "C" void invars_calc_4th(CCTK_ARGUMENTS) break; } - GenericFD_LoopOverInterior(cctkGH, &invars_calc_4th_Body); + GenericFD_LoopOverInterior(cctkGH, invars_calc_4th_Body); if (verbose > 1) { |