1dba47a55SKris Buschelman 24b9ad928SBarry Smith /* 34b9ad928SBarry Smith Include files needed for the PBJacobi preconditioner: 44b9ad928SBarry Smith pcimpl.h - private include file intended for use by all preconditioners 54b9ad928SBarry Smith */ 64b9ad928SBarry Smith 79fcdc22bSBarry Smith #include <private/matimpl.h> 8c6db04a5SJed Brown #include <private/pcimpl.h> /*I "petscpc.h" I*/ 94b9ad928SBarry Smith 104b9ad928SBarry Smith /* 114b9ad928SBarry Smith Private context (data structure) for the PBJacobi preconditioner. 124b9ad928SBarry Smith */ 134b9ad928SBarry Smith typedef struct { 1485f4f44aSBarry Smith MatScalar *diag; 15c1ac3661SBarry Smith PetscInt bs,mbs; 164b9ad928SBarry Smith } PC_PBJacobi; 174b9ad928SBarry Smith 184b9ad928SBarry Smith 194b9ad928SBarry Smith #undef __FUNCT__ 20bbead8a2SBarry Smith #define __FUNCT__ "PCApply_PBJacobi_1" 21bbead8a2SBarry Smith static PetscErrorCode PCApply_PBJacobi_1(PC pc,Vec x,Vec y) 22bbead8a2SBarry Smith { 23bbead8a2SBarry Smith PC_PBJacobi *jac = (PC_PBJacobi*)pc->data; 24bbead8a2SBarry Smith PetscErrorCode ierr; 25bbead8a2SBarry Smith PetscInt i,m = jac->mbs; 26bbead8a2SBarry Smith const MatScalar *diag = jac->diag; 27bbead8a2SBarry Smith const PetscScalar *xx; 28bbead8a2SBarry Smith PetscScalar *yy; 29bbead8a2SBarry Smith 30bbead8a2SBarry Smith PetscFunctionBegin; 31bbead8a2SBarry Smith ierr = VecGetArrayRead(x,&xx);CHKERRQ(ierr); 32bbead8a2SBarry Smith ierr = VecGetArray(y,&yy);CHKERRQ(ierr); 33bbead8a2SBarry Smith for (i=0; i<m; i++) { 34bbead8a2SBarry Smith yy[i] = diag[i]*xx[i]; 35bbead8a2SBarry Smith } 36bbead8a2SBarry Smith ierr = VecRestoreArrayRead(x,&xx);CHKERRQ(ierr); 37bbead8a2SBarry Smith ierr = VecRestoreArray(y,&yy);CHKERRQ(ierr); 38bbead8a2SBarry Smith ierr = PetscLogFlops(2.0*m);CHKERRQ(ierr); 39bbead8a2SBarry Smith PetscFunctionReturn(0); 40bbead8a2SBarry Smith } 41bbead8a2SBarry Smith 42bbead8a2SBarry Smith #undef __FUNCT__ 434b9ad928SBarry Smith #define __FUNCT__ "PCApply_PBJacobi_2" 446849ba73SBarry Smith static PetscErrorCode PCApply_PBJacobi_2(PC pc,Vec x,Vec y) 454b9ad928SBarry Smith { 464b9ad928SBarry Smith PC_PBJacobi *jac = (PC_PBJacobi*)pc->data; 47dfbe8321SBarry Smith PetscErrorCode ierr; 48c1ac3661SBarry Smith PetscInt i,m = jac->mbs; 4985f4f44aSBarry Smith const MatScalar *diag = jac->diag; 5085f4f44aSBarry Smith PetscScalar x0,x1,*xx,*yy; 514b9ad928SBarry Smith 524b9ad928SBarry Smith PetscFunctionBegin; 534b9ad928SBarry Smith ierr = VecGetArray(x,&xx);CHKERRQ(ierr); 544b9ad928SBarry Smith ierr = VecGetArray(y,&yy);CHKERRQ(ierr); 554b9ad928SBarry Smith for (i=0; i<m; i++) { 564b9ad928SBarry Smith x0 = xx[2*i]; x1 = xx[2*i+1]; 574b9ad928SBarry Smith yy[2*i] = diag[0]*x0 + diag[2]*x1; 584b9ad928SBarry Smith yy[2*i+1] = diag[1]*x0 + diag[3]*x1; 594b9ad928SBarry Smith diag += 4; 604b9ad928SBarry Smith } 614b9ad928SBarry Smith ierr = VecRestoreArray(x,&xx);CHKERRQ(ierr); 624b9ad928SBarry Smith ierr = VecRestoreArray(y,&yy);CHKERRQ(ierr); 63dc0b31edSSatish Balay ierr = PetscLogFlops(6.0*m);CHKERRQ(ierr); 644b9ad928SBarry Smith PetscFunctionReturn(0); 654b9ad928SBarry Smith } 664b9ad928SBarry Smith #undef __FUNCT__ 674b9ad928SBarry Smith #define __FUNCT__ "PCApply_PBJacobi_3" 686849ba73SBarry Smith static PetscErrorCode PCApply_PBJacobi_3(PC pc,Vec x,Vec y) 694b9ad928SBarry Smith { 704b9ad928SBarry Smith PC_PBJacobi *jac = (PC_PBJacobi*)pc->data; 71dfbe8321SBarry Smith PetscErrorCode ierr; 72c1ac3661SBarry Smith PetscInt i,m = jac->mbs; 7385f4f44aSBarry Smith const MatScalar *diag = jac->diag; 7485f4f44aSBarry Smith PetscScalar x0,x1,x2,*xx,*yy; 754b9ad928SBarry Smith 764b9ad928SBarry Smith PetscFunctionBegin; 774b9ad928SBarry Smith ierr = VecGetArray(x,&xx);CHKERRQ(ierr); 784b9ad928SBarry Smith ierr = VecGetArray(y,&yy);CHKERRQ(ierr); 794b9ad928SBarry Smith for (i=0; i<m; i++) { 804b9ad928SBarry Smith x0 = xx[3*i]; x1 = xx[3*i+1]; x2 = xx[3*i+2]; 814b9ad928SBarry Smith yy[3*i] = diag[0]*x0 + diag[3]*x1 + diag[6]*x2; 824b9ad928SBarry Smith yy[3*i+1] = diag[1]*x0 + diag[4]*x1 + diag[7]*x2; 834b9ad928SBarry Smith yy[3*i+2] = diag[2]*x0 + diag[5]*x1 + diag[8]*x2; 844b9ad928SBarry Smith diag += 9; 854b9ad928SBarry Smith } 864b9ad928SBarry Smith ierr = VecRestoreArray(x,&xx);CHKERRQ(ierr); 874b9ad928SBarry Smith ierr = VecRestoreArray(y,&yy);CHKERRQ(ierr); 88dc0b31edSSatish Balay ierr = PetscLogFlops(15.0*m);CHKERRQ(ierr); 894b9ad928SBarry Smith PetscFunctionReturn(0); 904b9ad928SBarry Smith } 914b9ad928SBarry Smith #undef __FUNCT__ 924b9ad928SBarry Smith #define __FUNCT__ "PCApply_PBJacobi_4" 936849ba73SBarry Smith static PetscErrorCode PCApply_PBJacobi_4(PC pc,Vec x,Vec y) 944b9ad928SBarry Smith { 954b9ad928SBarry Smith PC_PBJacobi *jac = (PC_PBJacobi*)pc->data; 96dfbe8321SBarry Smith PetscErrorCode ierr; 97c1ac3661SBarry Smith PetscInt i,m = jac->mbs; 9885f4f44aSBarry Smith const MatScalar *diag = jac->diag; 9985f4f44aSBarry Smith PetscScalar x0,x1,x2,x3,*xx,*yy; 1004b9ad928SBarry Smith 1014b9ad928SBarry Smith PetscFunctionBegin; 1024b9ad928SBarry Smith ierr = VecGetArray(x,&xx);CHKERRQ(ierr); 1034b9ad928SBarry Smith ierr = VecGetArray(y,&yy);CHKERRQ(ierr); 1044b9ad928SBarry Smith for (i=0; i<m; i++) { 1054b9ad928SBarry Smith x0 = xx[4*i]; x1 = xx[4*i+1]; x2 = xx[4*i+2]; x3 = xx[4*i+3]; 1064b9ad928SBarry Smith yy[4*i] = diag[0]*x0 + diag[4]*x1 + diag[8]*x2 + diag[12]*x3; 1074b9ad928SBarry Smith yy[4*i+1] = diag[1]*x0 + diag[5]*x1 + diag[9]*x2 + diag[13]*x3; 1084b9ad928SBarry Smith yy[4*i+2] = diag[2]*x0 + diag[6]*x1 + diag[10]*x2 + diag[14]*x3; 1094b9ad928SBarry Smith yy[4*i+3] = diag[3]*x0 + diag[7]*x1 + diag[11]*x2 + diag[15]*x3; 1104b9ad928SBarry Smith diag += 16; 1114b9ad928SBarry Smith } 1124b9ad928SBarry Smith ierr = VecRestoreArray(x,&xx);CHKERRQ(ierr); 1134b9ad928SBarry Smith ierr = VecRestoreArray(y,&yy);CHKERRQ(ierr); 114dc0b31edSSatish Balay ierr = PetscLogFlops(28.0*m);CHKERRQ(ierr); 1154b9ad928SBarry Smith PetscFunctionReturn(0); 1164b9ad928SBarry Smith } 1174b9ad928SBarry Smith #undef __FUNCT__ 1184b9ad928SBarry Smith #define __FUNCT__ "PCApply_PBJacobi_5" 1196849ba73SBarry Smith static PetscErrorCode PCApply_PBJacobi_5(PC pc,Vec x,Vec y) 1204b9ad928SBarry Smith { 1214b9ad928SBarry Smith PC_PBJacobi *jac = (PC_PBJacobi*)pc->data; 122dfbe8321SBarry Smith PetscErrorCode ierr; 123c1ac3661SBarry Smith PetscInt i,m = jac->mbs; 12485f4f44aSBarry Smith const MatScalar *diag = jac->diag; 12585f4f44aSBarry Smith PetscScalar x0,x1,x2,x3,x4,*xx,*yy; 1264b9ad928SBarry Smith 1274b9ad928SBarry Smith PetscFunctionBegin; 1284b9ad928SBarry Smith ierr = VecGetArray(x,&xx);CHKERRQ(ierr); 1294b9ad928SBarry Smith ierr = VecGetArray(y,&yy);CHKERRQ(ierr); 1304b9ad928SBarry Smith for (i=0; i<m; i++) { 1314b9ad928SBarry Smith x0 = xx[5*i]; x1 = xx[5*i+1]; x2 = xx[5*i+2]; x3 = xx[5*i+3]; x4 = xx[5*i+4]; 1324b9ad928SBarry Smith yy[5*i] = diag[0]*x0 + diag[5]*x1 + diag[10]*x2 + diag[15]*x3 + diag[20]*x4; 1334b9ad928SBarry Smith yy[5*i+1] = diag[1]*x0 + diag[6]*x1 + diag[11]*x2 + diag[16]*x3 + diag[21]*x4; 1344b9ad928SBarry Smith yy[5*i+2] = diag[2]*x0 + diag[7]*x1 + diag[12]*x2 + diag[17]*x3 + diag[22]*x4; 1354b9ad928SBarry Smith yy[5*i+3] = diag[3]*x0 + diag[8]*x1 + diag[13]*x2 + diag[18]*x3 + diag[23]*x4; 1364b9ad928SBarry Smith yy[5*i+4] = diag[4]*x0 + diag[9]*x1 + diag[14]*x2 + diag[19]*x3 + diag[24]*x4; 1374b9ad928SBarry Smith diag += 25; 1384b9ad928SBarry Smith } 1394b9ad928SBarry Smith ierr = VecRestoreArray(x,&xx);CHKERRQ(ierr); 1404b9ad928SBarry Smith ierr = VecRestoreArray(y,&yy);CHKERRQ(ierr); 141dc0b31edSSatish Balay ierr = PetscLogFlops(45.0*m);CHKERRQ(ierr); 1424b9ad928SBarry Smith PetscFunctionReturn(0); 1434b9ad928SBarry Smith } 144*0e1b4bd6SMark F. Adams #undef __FUNCT__ 145*0e1b4bd6SMark F. Adams #define __FUNCT__ "PCApply_PBJacobi_6" 146*0e1b4bd6SMark F. Adams static PetscErrorCode PCApply_PBJacobi_6(PC pc,Vec x,Vec y) 147*0e1b4bd6SMark F. Adams { 148*0e1b4bd6SMark F. Adams PC_PBJacobi *jac = (PC_PBJacobi*)pc->data; 149*0e1b4bd6SMark F. Adams PetscErrorCode ierr; 150*0e1b4bd6SMark F. Adams PetscInt i,m = jac->mbs; 151*0e1b4bd6SMark F. Adams const MatScalar *diag = jac->diag; 152*0e1b4bd6SMark F. Adams PetscScalar x0,x1,x2,x3,x4,x5,*xx,*yy; 153*0e1b4bd6SMark F. Adams 154*0e1b4bd6SMark F. Adams PetscFunctionBegin; 155*0e1b4bd6SMark F. Adams ierr = VecGetArray(x,&xx);CHKERRQ(ierr); 156*0e1b4bd6SMark F. Adams ierr = VecGetArray(y,&yy);CHKERRQ(ierr); 157*0e1b4bd6SMark F. Adams for (i=0; i<m; i++) { 158*0e1b4bd6SMark F. Adams x0 = xx[6*i]; x1 = xx[6*i+1]; x2 = xx[6*i+2]; x3 = xx[6*i+3]; x4 = xx[6*i+4]; x5 = xx[6*i+5]; 159*0e1b4bd6SMark F. Adams yy[6*i] = diag[0]*x0 + diag[6]*x1 + diag[12]*x2 + diag[18]*x3 + diag[24]*x4 + diag[30]*x5; 160*0e1b4bd6SMark F. Adams yy[6*i+1] = diag[1]*x0 + diag[7]*x1 + diag[13]*x2 + diag[19]*x3 + diag[25]*x4 + diag[31]*x5; 161*0e1b4bd6SMark F. Adams yy[6*i+2] = diag[2]*x0 + diag[8]*x1 + diag[14]*x2 + diag[20]*x3 + diag[26]*x4 + diag[32]*x5; 162*0e1b4bd6SMark F. Adams yy[6*i+3] = diag[3]*x0 + diag[9]*x1 + diag[15]*x2 + diag[21]*x3 + diag[27]*x4 + diag[33]*x5; 163*0e1b4bd6SMark F. Adams yy[6*i+4] = diag[4]*x0 + diag[10]*x1 + diag[16]*x2 + diag[22]*x3 + diag[28]*x4 + diag[34]*x5; 164*0e1b4bd6SMark F. Adams yy[6*i+5] = diag[5]*x0 + diag[11]*x1 + diag[17]*x2 + diag[23]*x3 + diag[29]*x4 + diag[35]*x5; 165*0e1b4bd6SMark F. Adams diag += 36; 166*0e1b4bd6SMark F. Adams } 167*0e1b4bd6SMark F. Adams ierr = VecRestoreArray(x,&xx);CHKERRQ(ierr); 168*0e1b4bd6SMark F. Adams ierr = VecRestoreArray(y,&yy);CHKERRQ(ierr); 169*0e1b4bd6SMark F. Adams ierr = PetscLogFlops(66.0*m);CHKERRQ(ierr); 170*0e1b4bd6SMark F. Adams PetscFunctionReturn(0); 171*0e1b4bd6SMark F. Adams } 1724b9ad928SBarry Smith /* -------------------------------------------------------------------------- */ 1734b9ad928SBarry Smith #undef __FUNCT__ 1744b9ad928SBarry Smith #define __FUNCT__ "PCSetUp_PBJacobi" 1756849ba73SBarry Smith static PetscErrorCode PCSetUp_PBJacobi(PC pc) 1764b9ad928SBarry Smith { 1774b9ad928SBarry Smith PC_PBJacobi *jac = (PC_PBJacobi*)pc->data; 178dfbe8321SBarry Smith PetscErrorCode ierr; 1794b9ad928SBarry Smith Mat A = pc->pmat; 1804b9ad928SBarry Smith 1814b9ad928SBarry Smith PetscFunctionBegin; 182e32f2f54SBarry Smith if (A->rmap->n != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Supported only for square matrices and square storage"); 183a1d92eedSBarry Smith 184bbead8a2SBarry Smith ierr = MatInvertBlockDiagonal(A,&jac->diag);CHKERRQ(ierr); 185d0f46423SBarry Smith jac->bs = A->rmap->bs; 186bbead8a2SBarry Smith jac->mbs = A->rmap->n/A->rmap->bs; 187521d7252SBarry Smith switch (jac->bs){ 188bbead8a2SBarry Smith case 1: 189bbead8a2SBarry Smith pc->ops->apply = PCApply_PBJacobi_1; 190bbead8a2SBarry Smith break; 1914b9ad928SBarry Smith case 2: 1924b9ad928SBarry Smith pc->ops->apply = PCApply_PBJacobi_2; 1934b9ad928SBarry Smith break; 1944b9ad928SBarry Smith case 3: 1954b9ad928SBarry Smith pc->ops->apply = PCApply_PBJacobi_3; 1964b9ad928SBarry Smith break; 1974b9ad928SBarry Smith case 4: 1984b9ad928SBarry Smith pc->ops->apply = PCApply_PBJacobi_4; 1994b9ad928SBarry Smith break; 2004b9ad928SBarry Smith case 5: 2014b9ad928SBarry Smith pc->ops->apply = PCApply_PBJacobi_5; 2024b9ad928SBarry Smith break; 203*0e1b4bd6SMark F. Adams case 6: 204*0e1b4bd6SMark F. Adams pc->ops->apply = PCApply_PBJacobi_6; 205*0e1b4bd6SMark F. Adams break; 2064b9ad928SBarry Smith default: 20765e19b50SBarry Smith SETERRQ1(((PetscObject)pc)->comm,PETSC_ERR_SUP,"not supported for block size %D",jac->bs); 2084b9ad928SBarry Smith } 2094b9ad928SBarry Smith 2104b9ad928SBarry Smith PetscFunctionReturn(0); 2114b9ad928SBarry Smith } 2124b9ad928SBarry Smith /* -------------------------------------------------------------------------- */ 2134b9ad928SBarry Smith #undef __FUNCT__ 2144b9ad928SBarry Smith #define __FUNCT__ "PCDestroy_PBJacobi" 2156849ba73SBarry Smith static PetscErrorCode PCDestroy_PBJacobi(PC pc) 2164b9ad928SBarry Smith { 217dfbe8321SBarry Smith PetscErrorCode ierr; 2184b9ad928SBarry Smith 2194b9ad928SBarry Smith PetscFunctionBegin; 2204b9ad928SBarry Smith /* 2214b9ad928SBarry Smith Free the private data structure that was hanging off the PC 2224b9ad928SBarry Smith */ 223c31cb41cSBarry Smith ierr = PetscFree(pc->data);CHKERRQ(ierr); 2244b9ad928SBarry Smith PetscFunctionReturn(0); 2254b9ad928SBarry Smith } 2264b9ad928SBarry Smith /* -------------------------------------------------------------------------- */ 22737a17b4dSBarry Smith /*MC 22837a17b4dSBarry Smith PCPBJACOBI - Point block Jacobi 22937a17b4dSBarry Smith 23037a17b4dSBarry Smith Level: beginner 23137a17b4dSBarry Smith 23237a17b4dSBarry Smith Concepts: point block Jacobi 23337a17b4dSBarry Smith 23437a17b4dSBarry Smith 23537a17b4dSBarry Smith .seealso: PCCreate(), PCSetType(), PCType (for list of available types), PC 23637a17b4dSBarry Smith 23737a17b4dSBarry Smith M*/ 23837a17b4dSBarry Smith 2394b9ad928SBarry Smith EXTERN_C_BEGIN 2404b9ad928SBarry Smith #undef __FUNCT__ 2414b9ad928SBarry Smith #define __FUNCT__ "PCCreate_PBJacobi" 2427087cfbeSBarry Smith PetscErrorCode PCCreate_PBJacobi(PC pc) 2434b9ad928SBarry Smith { 2444b9ad928SBarry Smith PC_PBJacobi *jac; 245dfbe8321SBarry Smith PetscErrorCode ierr; 2464b9ad928SBarry Smith 2474b9ad928SBarry Smith PetscFunctionBegin; 2484b9ad928SBarry Smith 2494b9ad928SBarry Smith /* 2504b9ad928SBarry Smith Creates the private data structure for this preconditioner and 2514b9ad928SBarry Smith attach it to the PC object. 2524b9ad928SBarry Smith */ 25338f2d2fdSLisandro Dalcin ierr = PetscNewLog(pc,PC_PBJacobi,&jac);CHKERRQ(ierr); 2544b9ad928SBarry Smith pc->data = (void*)jac; 2554b9ad928SBarry Smith 2564b9ad928SBarry Smith /* 2574b9ad928SBarry Smith Initialize the pointers to vectors to ZERO; these will be used to store 2584b9ad928SBarry Smith diagonal entries of the matrix for fast preconditioner application. 2594b9ad928SBarry Smith */ 2604b9ad928SBarry Smith jac->diag = 0; 2614b9ad928SBarry Smith 2624b9ad928SBarry Smith /* 2634b9ad928SBarry Smith Set the pointers for the functions that are provided above. 2644b9ad928SBarry Smith Now when the user-level routines (such as PCApply(), PCDestroy(), etc.) 2654b9ad928SBarry Smith are called, they will automatically call these functions. Note we 2664b9ad928SBarry Smith choose not to provide a couple of these functions since they are 2674b9ad928SBarry Smith not needed. 2684b9ad928SBarry Smith */ 2694b9ad928SBarry Smith pc->ops->apply = 0; /*set depending on the block size */ 2704b9ad928SBarry Smith pc->ops->applytranspose = 0; 2714b9ad928SBarry Smith pc->ops->setup = PCSetUp_PBJacobi; 2724b9ad928SBarry Smith pc->ops->destroy = PCDestroy_PBJacobi; 2734b9ad928SBarry Smith pc->ops->setfromoptions = 0; 2744b9ad928SBarry Smith pc->ops->view = 0; 2754b9ad928SBarry Smith pc->ops->applyrichardson = 0; 2764b9ad928SBarry Smith pc->ops->applysymmetricleft = 0; 2774b9ad928SBarry Smith pc->ops->applysymmetricright = 0; 2784b9ad928SBarry Smith PetscFunctionReturn(0); 2794b9ad928SBarry Smith } 2804b9ad928SBarry Smith EXTERN_C_END 2814b9ad928SBarry Smith 2824b9ad928SBarry Smith 283