diff options
Diffstat (limited to 'ML_BSSN_O2/src/ML_BSSN_O2_Advect.cc')
| -rw-r--r-- | ML_BSSN_O2/src/ML_BSSN_O2_Advect.cc | 222 |
1 files changed, 71 insertions, 151 deletions
diff --git a/ML_BSSN_O2/src/ML_BSSN_O2_Advect.cc b/ML_BSSN_O2/src/ML_BSSN_O2_Advect.cc index 48b0610..e0c80a1 100644 --- a/ML_BSSN_O2/src/ML_BSSN_O2_Advect.cc +++ b/ML_BSSN_O2/src/ML_BSSN_O2_Advect.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" @@ -64,8 +65,6 @@ static void ML_BSSN_O2_Advect_Body(cGH const * restrict const cctkGH, int const DECLARE_CCTK_PARAMETERS; - /* Declare finite differencing variables */ - /* Include user-supplied include files */ /* Initialise finite differencing variables */ @@ -102,9 +101,9 @@ static void ML_BSSN_O2_Advect_Body(cGH const * restrict const cctkGH, int const 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 p1o1680dx = kmul(INV(dx),ToReal(0.000595238095238095238095238095238)); CCTK_REAL_VEC const p1o1680dy = kmul(INV(dy),ToReal(0.000595238095238095238095238095238)); CCTK_REAL_VEC const p1o1680dz = kmul(INV(dz),ToReal(0.000595238095238095238095238095238)); @@ -123,14 +122,14 @@ static void ML_BSSN_O2_Advect_Body(cGH const * restrict const cctkGH, int const 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 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 p1o4dx = kmul(INV(dx),ToReal(0.25)); - 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 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 p1o4dy = kmul(INV(dy),ToReal(0.25)); - CCTK_REAL_VEC const p1o4dydz = kmul(INV(dy),kmul(INV(dz),ToReal(0.25))); + CCTK_REAL_VEC const p1o4dydz = kmul(INV(kmul(dy,dz)),ToReal(0.25)); CCTK_REAL_VEC const p1o4dz = kmul(INV(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)); @@ -144,9 +143,9 @@ static void ML_BSSN_O2_Advect_Body(cGH const * restrict const cctkGH, int const CCTK_REAL_VEC const p1o64dx = kmul(INV(dx),ToReal(0.015625)); CCTK_REAL_VEC const p1o64dy = kmul(INV(dy),ToReal(0.015625)); CCTK_REAL_VEC const p1o64dz = kmul(INV(dz),ToReal(0.015625)); - 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)); @@ -227,7 +226,7 @@ static void ML_BSSN_O2_Advect_Body(cGH const * restrict const cctkGH, int const /* Loop over the grid points */ #pragma omp parallel - LC_LOOP3VEC (ML_BSSN_O2_Advect, + LC_LOOP3VEC(ML_BSSN_O2_Advect, 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) @@ -1986,7 +1985,7 @@ static void ML_BSSN_O2_Advect_Body(cGH const * restrict const cctkGH, int const } phirhsL = - kmadd(beta1L,JacPDupwindNthAnti1phi,kmadd(beta2L,JacPDupwindNthAnti2phi,kmadd(beta3L,JacPDupwindNthAnti3phi,kadd(phirhsL,kmadd(JacPDupwindNthSymm1phi,kfabs(beta1L),kmadd(JacPDupwindNthSymm2phi,kfabs(beta2L),kmul(JacPDupwindNthSymm3phi,kfabs(beta3L)))))))); + kadd(phirhsL,kmadd(beta1L,JacPDupwindNthAnti1phi,kmadd(beta2L,JacPDupwindNthAnti2phi,kmadd(beta3L,JacPDupwindNthAnti3phi,kmadd(JacPDupwindNthSymm1phi,kfabs(beta1L),kmadd(JacPDupwindNthSymm2phi,kfabs(beta2L),kmul(JacPDupwindNthSymm3phi,kfabs(beta3L)))))))); gt11rhsL = kadd(gt11rhsL,kmadd(beta1L,JacPDupwindNthAnti1gt11,kmadd(beta2L,JacPDupwindNthAnti2gt11,kmadd(beta3L,JacPDupwindNthAnti3gt11,kmadd(JacPDupwindNthSymm1gt11,kfabs(beta1L),kmadd(JacPDupwindNthSymm2gt11,kfabs(beta2L),kmul(JacPDupwindNthSymm3gt11,kfabs(beta3L)))))))); @@ -2007,16 +2006,16 @@ static void ML_BSSN_O2_Advect_Body(cGH const * restrict const cctkGH, int const kadd(gt33rhsL,kmadd(beta1L,JacPDupwindNthAnti1gt33,kmadd(beta2L,JacPDupwindNthAnti2gt33,kmadd(beta3L,JacPDupwindNthAnti3gt33,kmadd(JacPDupwindNthSymm1gt33,kfabs(beta1L),kmadd(JacPDupwindNthSymm2gt33,kfabs(beta2L),kmul(JacPDupwindNthSymm3gt33,kfabs(beta3L)))))))); Xt1rhsL = - kmadd(beta1L,JacPDupwindNthAnti1Xt1,kmadd(beta2L,JacPDupwindNthAnti2Xt1,kmadd(beta3L,JacPDupwindNthAnti3Xt1,kadd(Xt1rhsL,kmadd(JacPDupwindNthSymm1Xt1,kfabs(beta1L),kmadd(JacPDupwindNthSymm2Xt1,kfabs(beta2L),kmul(JacPDupwindNthSymm3Xt1,kfabs(beta3L)))))))); + kadd(Xt1rhsL,kmadd(beta1L,JacPDupwindNthAnti1Xt1,kmadd(beta2L,JacPDupwindNthAnti2Xt1,kmadd(beta3L,JacPDupwindNthAnti3Xt1,kmadd(JacPDupwindNthSymm1Xt1,kfabs(beta1L),kmadd(JacPDupwindNthSymm2Xt1,kfabs(beta2L),kmul(JacPDupwindNthSymm3Xt1,kfabs(beta3L)))))))); Xt2rhsL = - kmadd(beta1L,JacPDupwindNthAnti1Xt2,kmadd(beta2L,JacPDupwindNthAnti2Xt2,kmadd(beta3L,JacPDupwindNthAnti3Xt2,kadd(Xt2rhsL,kmadd(JacPDupwindNthSymm1Xt2,kfabs(beta1L),kmadd(JacPDupwindNthSymm2Xt2,kfabs(beta2L),kmul(JacPDupwindNthSymm3Xt2,kfabs(beta3L)))))))); + kadd(Xt2rhsL,kmadd(beta1L,JacPDupwindNthAnti1Xt2,kmadd(beta2L,JacPDupwindNthAnti2Xt2,kmadd(beta3L,JacPDupwindNthAnti3Xt2,kmadd(JacPDupwindNthSymm1Xt2,kfabs(beta1L),kmadd(JacPDupwindNthSymm2Xt2,kfabs(beta2L),kmul(JacPDupwindNthSymm3Xt2,kfabs(beta3L)))))))); Xt3rhsL = - kmadd(beta1L,JacPDupwindNthAnti1Xt3,kmadd(beta2L,JacPDupwindNthAnti2Xt3,kmadd(beta3L,JacPDupwindNthAnti3Xt3,kadd(Xt3rhsL,kmadd(JacPDupwindNthSymm1Xt3,kfabs(beta1L),kmadd(JacPDupwindNthSymm2Xt3,kfabs(beta2L),kmul(JacPDupwindNthSymm3Xt3,kfabs(beta3L)))))))); + kadd(Xt3rhsL,kmadd(beta1L,JacPDupwindNthAnti1Xt3,kmadd(beta2L,JacPDupwindNthAnti2Xt3,kmadd(beta3L,JacPDupwindNthAnti3Xt3,kmadd(JacPDupwindNthSymm1Xt3,kfabs(beta1L),kmadd(JacPDupwindNthSymm2Xt3,kfabs(beta2L),kmul(JacPDupwindNthSymm3Xt3,kfabs(beta3L)))))))); trKrhsL = - kmadd(beta1L,JacPDupwindNthAnti1trK,kmadd(beta2L,JacPDupwindNthAnti2trK,kmadd(beta3L,JacPDupwindNthAnti3trK,kadd(trKrhsL,kmadd(JacPDupwindNthSymm1trK,kfabs(beta1L),kmadd(JacPDupwindNthSymm2trK,kfabs(beta2L),kmul(JacPDupwindNthSymm3trK,kfabs(beta3L)))))))); + kadd(trKrhsL,kmadd(beta1L,JacPDupwindNthAnti1trK,kmadd(beta2L,JacPDupwindNthAnti2trK,kmadd(beta3L,JacPDupwindNthAnti3trK,kmadd(JacPDupwindNthSymm1trK,kfabs(beta1L),kmadd(JacPDupwindNthSymm2trK,kfabs(beta2L),kmul(JacPDupwindNthSymm3trK,kfabs(beta3L)))))))); At11rhsL = kadd(At11rhsL,kmadd(beta1L,JacPDupwindNthAnti1At11,kmadd(beta2L,JacPDupwindNthAnti2At11,kmadd(beta3L,JacPDupwindNthAnti3At11,kmadd(JacPDupwindNthSymm1At11,kfabs(beta1L),kmadd(JacPDupwindNthSymm2At11,kfabs(beta2L),kmul(JacPDupwindNthSymm3At11,kfabs(beta3L)))))))); @@ -2040,7 +2039,7 @@ static void ML_BSSN_O2_Advect_Body(cGH const * restrict const cctkGH, int const kmadd(kmadd(beta1L,JacPDupwindNthAnti1alpha,kmadd(beta2L,JacPDupwindNthAnti2alpha,kmadd(beta3L,JacPDupwindNthAnti3alpha,kmadd(JacPDupwindNthSymm1alpha,kfabs(beta1L),kmadd(JacPDupwindNthSymm2alpha,kfabs(beta2L),kmul(JacPDupwindNthSymm3alpha,kfabs(beta3L))))))),ToReal(LapseAdvectionCoeff),alpharhsL); ArhsL = - kmadd(kmadd(beta1L,JacPDupwindNthAnti1A,kmadd(beta2L,JacPDupwindNthAnti2A,kmadd(beta3L,JacPDupwindNthAnti3A,kmadd(JacPDupwindNthSymm1A,kfabs(beta1L),kmadd(JacPDupwindNthSymm2A,kfabs(beta2L),kmul(JacPDupwindNthSymm3A,kfabs(beta3L))))))),ToReal(LapseAdvectionCoeff),ArhsL); + kmadd(ToReal(LapseACoeff),kmsub(kmadd(beta1L,JacPDupwindNthAnti1A,kmadd(beta2L,JacPDupwindNthAnti2A,kmadd(beta3L,JacPDupwindNthAnti3A,kmadd(JacPDupwindNthSymm1A,kfabs(beta1L),kmadd(JacPDupwindNthSymm2A,kfabs(beta2L),kmul(JacPDupwindNthSymm3A,kfabs(beta3L))))))),ToReal(LapseAdvectionCoeff),kmul(kmadd(beta1L,JacPDupwindNthAnti1trK,kmadd(beta2L,JacPDupwindNthAnti2trK,kmadd(beta3L,JacPDupwindNthAnti3trK,kmadd(JacPDupwindNthSymm1trK,kfabs(beta1L),kmadd(JacPDupwindNthSymm2trK,kfabs(beta2L),kmul(JacPDupwindNthSymm3trK,kfabs(beta3L))))))),kadd(ToReal(-1),ToReal(LapseAdvectionCoeff)))),ArhsL); beta1rhsL = kmadd(kmadd(beta1L,JacPDupwindNthAnti1beta1,kmadd(beta2L,JacPDupwindNthAnti2beta1,kmadd(beta3L,JacPDupwindNthAnti3beta1,kmadd(JacPDupwindNthSymm1beta1,kfabs(beta1L),kmadd(JacPDupwindNthSymm2beta1,kfabs(beta2L),kmul(JacPDupwindNthSymm3beta1,kfabs(beta3L))))))),ToReal(ShiftAdvectionCoeff),beta1rhsL); @@ -2052,140 +2051,43 @@ static void ML_BSSN_O2_Advect_Body(cGH const * restrict const cctkGH, int const kmadd(kmadd(beta1L,JacPDupwindNthAnti1beta3,kmadd(beta2L,JacPDupwindNthAnti2beta3,kmadd(beta3L,JacPDupwindNthAnti3beta3,kmadd(JacPDupwindNthSymm1beta3,kfabs(beta1L),kmadd(JacPDupwindNthSymm2beta3,kfabs(beta2L),kmul(JacPDupwindNthSymm3beta3,kfabs(beta3L))))))),ToReal(ShiftAdvectionCoeff),beta3rhsL); B1rhsL = - kadd(B1rhsL,kmadd(kmadd(beta1L,ksub(JacPDupwindNthAnti1B1,JacPDupwindNthAnti1Xt1),kmadd(beta2L,ksub(JacPDupwindNthAnti2B1,JacPDupwindNthAnti2Xt1),kmadd(beta3L,ksub(JacPDupwindNthAnti3B1,JacPDupwindNthAnti3Xt1),kmadd(kfabs(beta1L),ksub(JacPDupwindNthSymm1B1,JacPDupwindNthSymm1Xt1),kmadd(kfabs(beta2L),ksub(JacPDupwindNthSymm2B1,JacPDupwindNthSymm2Xt1),kmul(kfabs(beta3L),ksub(JacPDupwindNthSymm3B1,JacPDupwindNthSymm3Xt1))))))),ToReal(ShiftAdvectionCoeff),kmul(kmadd(beta1L,JacPDupwindNthAnti1Xt1,kmadd(beta2L,JacPDupwindNthAnti2Xt1,kmadd(beta3L,JacPDupwindNthAnti3Xt1,kmadd(JacPDupwindNthSymm1Xt1,kfabs(beta1L),kmadd(JacPDupwindNthSymm2Xt1,kfabs(beta2L),kmul(JacPDupwindNthSymm3Xt1,kfabs(beta3L))))))),ToReal(ShiftBCoeff)))); + kmadd(kmsub(kmadd(beta1L,JacPDupwindNthAnti1B1,kmadd(beta2L,JacPDupwindNthAnti2B1,kmadd(beta3L,JacPDupwindNthAnti3B1,kmadd(JacPDupwindNthSymm1B1,kfabs(beta1L),kmadd(JacPDupwindNthSymm2B1,kfabs(beta2L),kmul(JacPDupwindNthSymm3B1,kfabs(beta3L))))))),ToReal(ShiftAdvectionCoeff),kmul(kmadd(beta1L,JacPDupwindNthAnti1Xt1,kmadd(beta2L,JacPDupwindNthAnti2Xt1,kmadd(beta3L,JacPDupwindNthAnti3Xt1,kmadd(JacPDupwindNthSymm1Xt1,kfabs(beta1L),kmadd(JacPDupwindNthSymm2Xt1,kfabs(beta2L),kmul(JacPDupwindNthSymm3Xt1,kfabs(beta3L))))))),kadd(ToReal(-1),ToReal(ShiftAdvectionCoeff)))),ToReal(ShiftBCoeff),B1rhsL); B2rhsL = - kadd(B2rhsL,kmadd(kmadd(beta1L,ksub(JacPDupwindNthAnti1B2,JacPDupwindNthAnti1Xt2),kmadd(beta2L,ksub(JacPDupwindNthAnti2B2,JacPDupwindNthAnti2Xt2),kmadd(beta3L,ksub(JacPDupwindNthAnti3B2,JacPDupwindNthAnti3Xt2),kmadd(kfabs(beta1L),ksub(JacPDupwindNthSymm1B2,JacPDupwindNthSymm1Xt2),kmadd(kfabs(beta2L),ksub(JacPDupwindNthSymm2B2,JacPDupwindNthSymm2Xt2),kmul(kfabs(beta3L),ksub(JacPDupwindNthSymm3B2,JacPDupwindNthSymm3Xt2))))))),ToReal(ShiftAdvectionCoeff),kmul(kmadd(beta1L,JacPDupwindNthAnti1Xt2,kmadd(beta2L,JacPDupwindNthAnti2Xt2,kmadd(beta3L,JacPDupwindNthAnti3Xt2,kmadd(JacPDupwindNthSymm1Xt2,kfabs(beta1L),kmadd(JacPDupwindNthSymm2Xt2,kfabs(beta2L),kmul(JacPDupwindNthSymm3Xt2,kfabs(beta3L))))))),ToReal(ShiftBCoeff)))); + kmadd(kmsub(kmadd(beta1L,JacPDupwindNthAnti1B2,kmadd(beta2L,JacPDupwindNthAnti2B2,kmadd(beta3L,JacPDupwindNthAnti3B2,kmadd(JacPDupwindNthSymm1B2,kfabs(beta1L),kmadd(JacPDupwindNthSymm2B2,kfabs(beta2L),kmul(JacPDupwindNthSymm3B2,kfabs(beta3L))))))),ToReal(ShiftAdvectionCoeff),kmul(kmadd(beta1L,JacPDupwindNthAnti1Xt2,kmadd(beta2L,JacPDupwindNthAnti2Xt2,kmadd(beta3L,JacPDupwindNthAnti3Xt2,kmadd(JacPDupwindNthSymm1Xt2,kfabs(beta1L),kmadd(JacPDupwindNthSymm2Xt2,kfabs(beta2L),kmul(JacPDupwindNthSymm3Xt2,kfabs(beta3L))))))),kadd(ToReal(-1),ToReal(ShiftAdvectionCoeff)))),ToReal(ShiftBCoeff),B2rhsL); B3rhsL = - kadd(B3rhsL,kmadd(kmadd(beta1L,ksub(JacPDupwindNthAnti1B3,JacPDupwindNthAnti1Xt3),kmadd(beta2L,ksub(JacPDupwindNthAnti2B3,JacPDupwindNthAnti2Xt3),kmadd(beta3L,ksub(JacPDupwindNthAnti3B3,JacPDupwindNthAnti3Xt3),kmadd(kfabs(beta1L),ksub(JacPDupwindNthSymm1B3,JacPDupwindNthSymm1Xt3),kmadd(kfabs(beta2L),ksub(JacPDupwindNthSymm2B3,JacPDupwindNthSymm2Xt3),kmul(kfabs(beta3L),ksub(JacPDupwindNthSymm3B3,JacPDupwindNthSymm3Xt3))))))),ToReal(ShiftAdvectionCoeff),kmul(kmadd(beta1L,JacPDupwindNthAnti1Xt3,kmadd(beta2L,JacPDupwindNthAnti2Xt3,kmadd(beta3L,JacPDupwindNthAnti3Xt3,kmadd(JacPDupwindNthSymm1Xt3,kfabs(beta1L),kmadd(JacPDupwindNthSymm2Xt3,kfabs(beta2L),kmul(JacPDupwindNthSymm3Xt3,kfabs(beta3L))))))),ToReal(ShiftBCoeff)))); - - /* 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(alpharhs[index],alpharhsL,elt_count_lo,elt_count_hi); - vec_store_nta_partial_mid(Arhs[index],ArhsL,elt_count_lo,elt_count_hi); - vec_store_nta_partial_mid(At11rhs[index],At11rhsL,elt_count_lo,elt_count_hi); - vec_store_nta_partial_mid(At12rhs[index],At12rhsL,elt_count_lo,elt_count_hi); - vec_store_nta_partial_mid(At13rhs[index],At13rhsL,elt_count_lo,elt_count_hi); - vec_store_nta_partial_mid(At22rhs[index],At22rhsL,elt_count_lo,elt_count_hi); - vec_store_nta_partial_mid(At23rhs[index],At23rhsL,elt_count_lo,elt_count_hi); - vec_store_nta_partial_mid(At33rhs[index],At33rhsL,elt_count_lo,elt_count_hi); - vec_store_nta_partial_mid(B1rhs[index],B1rhsL,elt_count_lo,elt_count_hi); - vec_store_nta_partial_mid(B2rhs[index],B2rhsL,elt_count_lo,elt_count_hi); - vec_store_nta_partial_mid(B3rhs[index],B3rhsL,elt_count_lo,elt_count_hi); - vec_store_nta_partial_mid(beta1rhs[index],beta1rhsL,elt_count_lo,elt_count_hi); - vec_store_nta_partial_mid(beta2rhs[index],beta2rhsL,elt_count_lo,elt_count_hi); - vec_store_nta_partial_mid(beta3rhs[index],beta3rhsL,elt_count_lo,elt_count_hi); - vec_store_nta_partial_mid(gt11rhs[index],gt11rhsL,elt_count_lo,elt_count_hi); - vec_store_nta_partial_mid(gt12rhs[index],gt12rhsL,elt_count_lo,elt_count_hi); - vec_store_nta_partial_mid(gt13rhs[index],gt13rhsL,elt_count_lo,elt_count_hi); - vec_store_nta_partial_mid(gt22rhs[index],gt22rhsL,elt_count_lo,elt_count_hi); - vec_store_nta_partial_mid(gt23rhs[index],gt23rhsL,elt_count_lo,elt_count_hi); - vec_store_nta_partial_mid(gt33rhs[index],gt33rhsL,elt_count_lo,elt_count_hi); - vec_store_nta_partial_mid(phirhs[index],phirhsL,elt_count_lo,elt_count_hi); - vec_store_nta_partial_mid(trKrhs[index],trKrhsL,elt_count_lo,elt_count_hi); - vec_store_nta_partial_mid(Xt1rhs[index],Xt1rhsL,elt_count_lo,elt_count_hi); - vec_store_nta_partial_mid(Xt2rhs[index],Xt2rhsL,elt_count_lo,elt_count_hi); - vec_store_nta_partial_mid(Xt3rhs[index],Xt3rhsL,elt_count_lo,elt_count_hi); - break; - } - - /* If necessary, store only partial vectors after the first iteration */ + kmadd(kmsub(kmadd(beta1L,JacPDupwindNthAnti1B3,kmadd(beta2L,JacPDupwindNthAnti2B3,kmadd(beta3L,JacPDupwindNthAnti3B3,kmadd(JacPDupwindNthSymm1B3,kfabs(beta1L),kmadd(JacPDupwindNthSymm2B3,kfabs(beta2L),kmul(JacPDupwindNthSymm3B3,kfabs(beta3L))))))),ToReal(ShiftAdvectionCoeff),kmul(kmadd(beta1L,JacPDupwindNthAnti1Xt3,kmadd(beta2L,JacPDupwindNthAnti2Xt3,kmadd(beta3L,JacPDupwindNthAnti3Xt3,kmadd(JacPDupwindNthSymm1Xt3,kfabs(beta1L),kmadd(JacPDupwindNthSymm2Xt3,kfabs(beta2L),kmul(JacPDupwindNthSymm3Xt3,kfabs(beta3L))))))),kadd(ToReal(-1),ToReal(ShiftAdvectionCoeff)))),ToReal(ShiftBCoeff),B3rhsL); - 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(alpharhs[index],alpharhsL,elt_count); - vec_store_nta_partial_hi(Arhs[index],ArhsL,elt_count); - vec_store_nta_partial_hi(At11rhs[index],At11rhsL,elt_count); - vec_store_nta_partial_hi(At12rhs[index],At12rhsL,elt_count); - vec_store_nta_partial_hi(At13rhs[index],At13rhsL,elt_count); - vec_store_nta_partial_hi(At22rhs[index],At22rhsL,elt_count); - vec_store_nta_partial_hi(At23rhs[index],At23rhsL,elt_count); - vec_store_nta_partial_hi(At33rhs[index],At33rhsL,elt_count); - vec_store_nta_partial_hi(B1rhs[index],B1rhsL,elt_count); - vec_store_nta_partial_hi(B2rhs[index],B2rhsL,elt_count); - vec_store_nta_partial_hi(B3rhs[index],B3rhsL,elt_count); - vec_store_nta_partial_hi(beta1rhs[index],beta1rhsL,elt_count); - vec_store_nta_partial_hi(beta2rhs[index],beta2rhsL,elt_count); - vec_store_nta_partial_hi(beta3rhs[index],beta3rhsL,elt_count); - vec_store_nta_partial_hi(gt11rhs[index],gt11rhsL,elt_count); - vec_store_nta_partial_hi(gt12rhs[index],gt12rhsL,elt_count); - vec_store_nta_partial_hi(gt13rhs[index],gt13rhsL,elt_count); - vec_store_nta_partial_hi(gt22rhs[index],gt22rhsL,elt_count); - vec_store_nta_partial_hi(gt23rhs[index],gt23rhsL,elt_count); - vec_store_nta_partial_hi(gt33rhs[index],gt33rhsL,elt_count); - vec_store_nta_partial_hi(phirhs[index],phirhsL,elt_count); - vec_store_nta_partial_hi(trKrhs[index],trKrhsL,elt_count); - vec_store_nta_partial_hi(Xt1rhs[index],Xt1rhsL,elt_count); - vec_store_nta_partial_hi(Xt2rhs[index],Xt2rhsL,elt_count); - vec_store_nta_partial_hi(Xt3rhs[index],Xt3rhsL,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(alpharhs[index],alpharhsL,elt_count); - vec_store_nta_partial_lo(Arhs[index],ArhsL,elt_count); - vec_store_nta_partial_lo(At11rhs[index],At11rhsL,elt_count); - vec_store_nta_partial_lo(At12rhs[index],At12rhsL,elt_count); - vec_store_nta_partial_lo(At13rhs[index],At13rhsL,elt_count); - vec_store_nta_partial_lo(At22rhs[index],At22rhsL,elt_count); - vec_store_nta_partial_lo(At23rhs[index],At23rhsL,elt_count); - vec_store_nta_partial_lo(At33rhs[index],At33rhsL,elt_count); - vec_store_nta_partial_lo(B1rhs[index],B1rhsL,elt_count); - vec_store_nta_partial_lo(B2rhs[index],B2rhsL,elt_count); - vec_store_nta_partial_lo(B3rhs[index],B3rhsL,elt_count); - vec_store_nta_partial_lo(beta1rhs[index],beta1rhsL,elt_count); - vec_store_nta_partial_lo(beta2rhs[index],beta2rhsL,elt_count); - vec_store_nta_partial_lo(beta3rhs[index],beta3rhsL,elt_count); - vec_store_nta_partial_lo(gt11rhs[index],gt11rhsL,elt_count); - vec_store_nta_partial_lo(gt12rhs[index],gt12rhsL,elt_count); - vec_store_nta_partial_lo(gt13rhs[index],gt13rhsL,elt_count); - vec_store_nta_partial_lo(gt22rhs[index],gt22rhsL,elt_count); - vec_store_nta_partial_lo(gt23rhs[index],gt23rhsL,elt_count); - vec_store_nta_partial_lo(gt33rhs[index],gt33rhsL,elt_count); - vec_store_nta_partial_lo(phirhs[index],phirhsL,elt_count); - vec_store_nta_partial_lo(trKrhs[index],trKrhsL,elt_count); - vec_store_nta_partial_lo(Xt1rhs[index],Xt1rhsL,elt_count); - vec_store_nta_partial_lo(Xt2rhs[index],Xt2rhsL,elt_count); - vec_store_nta_partial_lo(Xt3rhs[index],Xt3rhsL,elt_count); - break; - } - vec_store_nta(alpharhs[index],alpharhsL); - vec_store_nta(Arhs[index],ArhsL); - vec_store_nta(At11rhs[index],At11rhsL); - vec_store_nta(At12rhs[index],At12rhsL); - vec_store_nta(At13rhs[index],At13rhsL); - vec_store_nta(At22rhs[index],At22rhsL); - vec_store_nta(At23rhs[index],At23rhsL); - vec_store_nta(At33rhs[index],At33rhsL); - vec_store_nta(B1rhs[index],B1rhsL); - vec_store_nta(B2rhs[index],B2rhsL); - vec_store_nta(B3rhs[index],B3rhsL); - vec_store_nta(beta1rhs[index],beta1rhsL); - vec_store_nta(beta2rhs[index],beta2rhsL); - vec_store_nta(beta3rhs[index],beta3rhsL); - vec_store_nta(gt11rhs[index],gt11rhsL); - vec_store_nta(gt12rhs[index],gt12rhsL); - vec_store_nta(gt13rhs[index],gt13rhsL); - vec_store_nta(gt22rhs[index],gt22rhsL); - vec_store_nta(gt23rhs[index],gt23rhsL); - vec_store_nta(gt33rhs[index],gt33rhsL); - vec_store_nta(phirhs[index],phirhsL); - vec_store_nta(trKrhs[index],trKrhsL); - vec_store_nta(Xt1rhs[index],Xt1rhsL); - vec_store_nta(Xt2rhs[index],Xt2rhsL); - vec_store_nta(Xt3rhs[index],Xt3rhsL); + /* Copy local copies back to grid functions */ + vec_store_partial_prepare(i,lc_imin,lc_imax); + vec_store_nta_partial(alpharhs[index],alpharhsL); + vec_store_nta_partial(Arhs[index],ArhsL); + vec_store_nta_partial(At11rhs[index],At11rhsL); + vec_store_nta_partial(At12rhs[index],At12rhsL); + vec_store_nta_partial(At13rhs[index],At13rhsL); + vec_store_nta_partial(At22rhs[index],At22rhsL); + vec_store_nta_partial(At23rhs[index],At23rhsL); + vec_store_nta_partial(At33rhs[index],At33rhsL); + vec_store_nta_partial(B1rhs[index],B1rhsL); + vec_store_nta_partial(B2rhs[index],B2rhsL); + vec_store_nta_partial(B3rhs[index],B3rhsL); + vec_store_nta_partial(beta1rhs[index],beta1rhsL); + vec_store_nta_partial(beta2rhs[index],beta2rhsL); + vec_store_nta_partial(beta3rhs[index],beta3rhsL); + vec_store_nta_partial(gt11rhs[index],gt11rhsL); + vec_store_nta_partial(gt12rhs[index],gt12rhsL); + vec_store_nta_partial(gt13rhs[index],gt13rhsL); + vec_store_nta_partial(gt22rhs[index],gt22rhsL); + vec_store_nta_partial(gt23rhs[index],gt23rhsL); + vec_store_nta_partial(gt33rhs[index],gt33rhsL); + vec_store_nta_partial(phirhs[index],phirhsL); + vec_store_nta_partial(trKrhs[index],trKrhsL); + vec_store_nta_partial(Xt1rhs[index],Xt1rhsL); + vec_store_nta_partial(Xt2rhs[index],Xt2rhsL); + vec_store_nta_partial(Xt3rhs[index],Xt3rhsL); } - LC_ENDLOOP3VEC (ML_BSSN_O2_Advect); + LC_ENDLOOP3VEC(ML_BSSN_O2_Advect); } extern "C" void ML_BSSN_O2_Advect(CCTK_ARGUMENTS) @@ -2204,7 +2106,25 @@ extern "C" void ML_BSSN_O2_Advect(CCTK_ARGUMENTS) return; } - const char *groups[] = {"ML_BSSN_O2::ML_curv","ML_BSSN_O2::ML_curvrhs","ML_BSSN_O2::ML_dtlapse","ML_BSSN_O2::ML_dtlapserhs","ML_BSSN_O2::ML_dtshift","ML_BSSN_O2::ML_dtshiftrhs","ML_BSSN_O2::ML_Gamma","ML_BSSN_O2::ML_Gammarhs","ML_BSSN_O2::ML_lapse","ML_BSSN_O2::ML_lapserhs","ML_BSSN_O2::ML_log_confac","ML_BSSN_O2::ML_log_confacrhs","ML_BSSN_O2::ML_metric","ML_BSSN_O2::ML_metricrhs","ML_BSSN_O2::ML_shift","ML_BSSN_O2::ML_shiftrhs","ML_BSSN_O2::ML_trace_curv","ML_BSSN_O2::ML_trace_curvrhs"}; + const char *const groups[] = { + "ML_BSSN_O2::ML_curv", + "ML_BSSN_O2::ML_curvrhs", + "ML_BSSN_O2::ML_dtlapse", + "ML_BSSN_O2::ML_dtlapserhs", + "ML_BSSN_O2::ML_dtshift", + "ML_BSSN_O2::ML_dtshiftrhs", + "ML_BSSN_O2::ML_Gamma", + "ML_BSSN_O2::ML_Gammarhs", + "ML_BSSN_O2::ML_lapse", + "ML_BSSN_O2::ML_lapserhs", + "ML_BSSN_O2::ML_log_confac", + "ML_BSSN_O2::ML_log_confacrhs", + "ML_BSSN_O2::ML_metric", + "ML_BSSN_O2::ML_metricrhs", + "ML_BSSN_O2::ML_shift", + "ML_BSSN_O2::ML_shiftrhs", + "ML_BSSN_O2::ML_trace_curv", + "ML_BSSN_O2::ML_trace_curvrhs"}; GenericFD_AssertGroupStorage(cctkGH, "ML_BSSN_O2_Advect", 18, groups); switch(fdOrder) @@ -2226,7 +2146,7 @@ extern "C" void ML_BSSN_O2_Advect(CCTK_ARGUMENTS) break; } - GenericFD_LoopOverInterior(cctkGH, &ML_BSSN_O2_Advect_Body); + GenericFD_LoopOverInterior(cctkGH, ML_BSSN_O2_Advect_Body); if (verbose > 1) { |