1 2 /* 3 Include files needed for the PBJacobi preconditioner: 4 pcimpl.h - private include file intended for use by all preconditioners 5 */ 6 7 #include <petsc/private/pcimpl.h> /*I "petscpc.h" I*/ 8 9 /* 10 Private context (data structure) for the PBJacobi preconditioner. 11 */ 12 typedef struct { 13 const MatScalar *diag; 14 PetscInt bs, mbs; 15 } PC_PBJacobi; 16 17 static PetscErrorCode PCApply_PBJacobi(PC pc, Vec x, Vec y) 18 { 19 PC_PBJacobi *jac = (PC_PBJacobi *)pc->data; 20 PetscInt i, ib, jb; 21 const PetscInt m = jac->mbs; 22 const PetscInt bs = jac->bs; 23 const MatScalar *diag = jac->diag; 24 PetscScalar *yy, x0, x1, x2, x3, x4, x5, x6; 25 const PetscScalar *xx; 26 27 PetscFunctionBegin; 28 PetscCall(VecGetArrayRead(x, &xx)); 29 PetscCall(VecGetArray(y, &yy)); 30 switch (bs) { 31 case 1: 32 for (i = 0; i < m; i++) yy[i] = diag[i] * xx[i]; 33 break; 34 case 2: 35 for (i = 0; i < m; i++) { 36 x0 = xx[2 * i]; 37 x1 = xx[2 * i + 1]; 38 yy[2 * i] = diag[0] * x0 + diag[2] * x1; 39 yy[2 * i + 1] = diag[1] * x0 + diag[3] * x1; 40 diag += 4; 41 } 42 break; 43 case 3: 44 for (i = 0; i < m; i++) { 45 x0 = xx[3 * i]; 46 x1 = xx[3 * i + 1]; 47 x2 = xx[3 * i + 2]; 48 49 yy[3 * i] = diag[0] * x0 + diag[3] * x1 + diag[6] * x2; 50 yy[3 * i + 1] = diag[1] * x0 + diag[4] * x1 + diag[7] * x2; 51 yy[3 * i + 2] = diag[2] * x0 + diag[5] * x1 + diag[8] * x2; 52 diag += 9; 53 } 54 break; 55 case 4: 56 for (i = 0; i < m; i++) { 57 x0 = xx[4 * i]; 58 x1 = xx[4 * i + 1]; 59 x2 = xx[4 * i + 2]; 60 x3 = xx[4 * i + 3]; 61 62 yy[4 * i] = diag[0] * x0 + diag[4] * x1 + diag[8] * x2 + diag[12] * x3; 63 yy[4 * i + 1] = diag[1] * x0 + diag[5] * x1 + diag[9] * x2 + diag[13] * x3; 64 yy[4 * i + 2] = diag[2] * x0 + diag[6] * x1 + diag[10] * x2 + diag[14] * x3; 65 yy[4 * i + 3] = diag[3] * x0 + diag[7] * x1 + diag[11] * x2 + diag[15] * x3; 66 diag += 16; 67 } 68 break; 69 case 5: 70 for (i = 0; i < m; i++) { 71 x0 = xx[5 * i]; 72 x1 = xx[5 * i + 1]; 73 x2 = xx[5 * i + 2]; 74 x3 = xx[5 * i + 3]; 75 x4 = xx[5 * i + 4]; 76 77 yy[5 * i] = diag[0] * x0 + diag[5] * x1 + diag[10] * x2 + diag[15] * x3 + diag[20] * x4; 78 yy[5 * i + 1] = diag[1] * x0 + diag[6] * x1 + diag[11] * x2 + diag[16] * x3 + diag[21] * x4; 79 yy[5 * i + 2] = diag[2] * x0 + diag[7] * x1 + diag[12] * x2 + diag[17] * x3 + diag[22] * x4; 80 yy[5 * i + 3] = diag[3] * x0 + diag[8] * x1 + diag[13] * x2 + diag[18] * x3 + diag[23] * x4; 81 yy[5 * i + 4] = diag[4] * x0 + diag[9] * x1 + diag[14] * x2 + diag[19] * x3 + diag[24] * x4; 82 diag += 25; 83 } 84 break; 85 case 6: 86 for (i = 0; i < m; i++) { 87 x0 = xx[6 * i]; 88 x1 = xx[6 * i + 1]; 89 x2 = xx[6 * i + 2]; 90 x3 = xx[6 * i + 3]; 91 x4 = xx[6 * i + 4]; 92 x5 = xx[6 * i + 5]; 93 94 yy[6 * i] = diag[0] * x0 + diag[6] * x1 + diag[12] * x2 + diag[18] * x3 + diag[24] * x4 + diag[30] * x5; 95 yy[6 * i + 1] = diag[1] * x0 + diag[7] * x1 + diag[13] * x2 + diag[19] * x3 + diag[25] * x4 + diag[31] * x5; 96 yy[6 * i + 2] = diag[2] * x0 + diag[8] * x1 + diag[14] * x2 + diag[20] * x3 + diag[26] * x4 + diag[32] * x5; 97 yy[6 * i + 3] = diag[3] * x0 + diag[9] * x1 + diag[15] * x2 + diag[21] * x3 + diag[27] * x4 + diag[33] * x5; 98 yy[6 * i + 4] = diag[4] * x0 + diag[10] * x1 + diag[16] * x2 + diag[22] * x3 + diag[28] * x4 + diag[34] * x5; 99 yy[6 * i + 5] = diag[5] * x0 + diag[11] * x1 + diag[17] * x2 + diag[23] * x3 + diag[29] * x4 + diag[35] * x5; 100 diag += 36; 101 } 102 break; 103 case 7: 104 for (i = 0; i < m; i++) { 105 x0 = xx[7 * i]; 106 x1 = xx[7 * i + 1]; 107 x2 = xx[7 * i + 2]; 108 x3 = xx[7 * i + 3]; 109 x4 = xx[7 * i + 4]; 110 x5 = xx[7 * i + 5]; 111 x6 = xx[7 * i + 6]; 112 113 yy[7 * i] = diag[0] * x0 + diag[7] * x1 + diag[14] * x2 + diag[21] * x3 + diag[28] * x4 + diag[35] * x5 + diag[42] * x6; 114 yy[7 * i + 1] = diag[1] * x0 + diag[8] * x1 + diag[15] * x2 + diag[22] * x3 + diag[29] * x4 + diag[36] * x5 + diag[43] * x6; 115 yy[7 * i + 2] = diag[2] * x0 + diag[9] * x1 + diag[16] * x2 + diag[23] * x3 + diag[30] * x4 + diag[37] * x5 + diag[44] * x6; 116 yy[7 * i + 3] = diag[3] * x0 + diag[10] * x1 + diag[17] * x2 + diag[24] * x3 + diag[31] * x4 + diag[38] * x5 + diag[45] * x6; 117 yy[7 * i + 4] = diag[4] * x0 + diag[11] * x1 + diag[18] * x2 + diag[25] * x3 + diag[32] * x4 + diag[39] * x5 + diag[46] * x6; 118 yy[7 * i + 5] = diag[5] * x0 + diag[12] * x1 + diag[19] * x2 + diag[26] * x3 + diag[33] * x4 + diag[40] * x5 + diag[47] * x6; 119 yy[7 * i + 6] = diag[6] * x0 + diag[13] * x1 + diag[20] * x2 + diag[27] * x3 + diag[34] * x4 + diag[41] * x5 + diag[48] * x6; 120 diag += 49; 121 } 122 break; 123 default: 124 for (i = 0; i < m; i++) { 125 for (ib = 0; ib < bs; ib++) { 126 PetscScalar rowsum = 0; 127 for (jb = 0; jb < bs; jb++) rowsum += diag[ib + jb * bs] * xx[bs * i + jb]; 128 yy[bs * i + ib] = rowsum; 129 } 130 diag += bs * bs; 131 } 132 } 133 PetscCall(VecRestoreArrayRead(x, &xx)); 134 PetscCall(VecRestoreArray(y, &yy)); 135 PetscCall(PetscLogFlops((2.0 * bs * bs - bs) * m)); /* 2*bs2 - bs */ 136 PetscFunctionReturn(0); 137 } 138 139 static PetscErrorCode PCApplyTranspose_PBJacobi_N(PC pc, Vec x, Vec y) 140 { 141 PC_PBJacobi *jac = (PC_PBJacobi *)pc->data; 142 PetscInt i, j, k, m = jac->mbs, bs = jac->bs; 143 const MatScalar *diag = jac->diag; 144 const PetscScalar *xx; 145 PetscScalar *yy; 146 147 PetscFunctionBegin; 148 PetscCall(VecGetArrayRead(x, &xx)); 149 PetscCall(VecGetArray(y, &yy)); 150 for (i = 0; i < m; i++) { 151 for (j = 0; j < bs; j++) yy[i * bs + j] = 0.; 152 for (j = 0; j < bs; j++) { 153 for (k = 0; k < bs; k++) yy[i * bs + k] += diag[k * bs + j] * xx[i * bs + j]; 154 } 155 diag += bs * bs; 156 } 157 PetscCall(VecRestoreArrayRead(x, &xx)); 158 PetscCall(VecRestoreArray(y, &yy)); 159 PetscCall(PetscLogFlops(m * bs * (2 * bs - 1))); 160 PetscFunctionReturn(0); 161 } 162 163 static PetscErrorCode PCSetUp_PBJacobi(PC pc) 164 { 165 PC_PBJacobi *jac = (PC_PBJacobi *)pc->data; 166 Mat A = pc->pmat; 167 MatFactorError err; 168 PetscInt nlocal; 169 170 PetscFunctionBegin; 171 PetscCall(MatInvertBlockDiagonal(A, &jac->diag)); 172 PetscCall(MatFactorGetError(A, &err)); 173 if (err) pc->failedreason = (PCFailedReason)err; 174 175 PetscCall(MatGetBlockSize(A, &jac->bs)); 176 PetscCall(MatGetLocalSize(A, &nlocal, NULL)); 177 jac->mbs = nlocal / jac->bs; 178 pc->ops->apply = PCApply_PBJacobi; 179 pc->ops->applytranspose = PCApplyTranspose_PBJacobi_N; 180 PetscFunctionReturn(0); 181 } 182 183 static PetscErrorCode PCDestroy_PBJacobi(PC pc) 184 { 185 PetscFunctionBegin; 186 /* 187 Free the private data structure that was hanging off the PC 188 */ 189 PetscCall(PetscFree(pc->data)); 190 PetscFunctionReturn(0); 191 } 192 193 static PetscErrorCode PCView_PBJacobi(PC pc, PetscViewer viewer) 194 { 195 PC_PBJacobi *jac = (PC_PBJacobi *)pc->data; 196 PetscBool iascii; 197 198 PetscFunctionBegin; 199 PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &iascii)); 200 if (iascii) PetscCall(PetscViewerASCIIPrintf(viewer, " point-block size %" PetscInt_FMT "\n", jac->bs)); 201 PetscFunctionReturn(0); 202 } 203 204 /*MC 205 PCPBJACOBI - Point block Jacobi preconditioner 206 207 Notes: 208 See `PCJACOBI` for diagonal Jacobi, `PCVPBJACOBI` for variable-size point block, and `PCBJACOBI` for large size blocks 209 210 This works for `MATAIJ` and `MATBAIJ` matrices and uses the blocksize provided to the matrix 211 212 Uses dense LU factorization with partial pivoting to invert the blocks; if a zero pivot 213 is detected a PETSc error is generated. 214 215 Developer Notes: 216 This should support the `PCSetErrorIfFailure()` flag set to `PETSC_TRUE` to allow 217 the factorization to continue even after a zero pivot is found resulting in a Nan and hence 218 terminating `KSP` with a `KSP_DIVERGED_NANORIF` allowing 219 a nonlinear solver/ODE integrator to recover without stopping the program as currently happens. 220 221 Perhaps should provide an option that allows generation of a valid preconditioner 222 even if a block is singular as the `PCJACOBI` does. 223 224 Level: beginner 225 226 .seealso: `PCCreate()`, `PCSetType()`, `PCType`, `PC`, `PCJACOBI`, `PCVPBJACOBI`, `PCBJACOBI` 227 M*/ 228 229 PETSC_EXTERN PetscErrorCode PCCreate_PBJacobi(PC pc) 230 { 231 PC_PBJacobi *jac; 232 233 PetscFunctionBegin; 234 /* 235 Creates the private data structure for this preconditioner and 236 attach it to the PC object. 237 */ 238 PetscCall(PetscNew(&jac)); 239 pc->data = (void *)jac; 240 241 /* 242 Initialize the pointers to vectors to ZERO; these will be used to store 243 diagonal entries of the matrix for fast preconditioner application. 244 */ 245 jac->diag = NULL; 246 247 /* 248 Set the pointers for the functions that are provided above. 249 Now when the user-level routines (such as PCApply(), PCDestroy(), etc.) 250 are called, they will automatically call these functions. Note we 251 choose not to provide a couple of these functions since they are 252 not needed. 253 */ 254 pc->ops->apply = NULL; /*set depending on the block size */ 255 pc->ops->applytranspose = NULL; 256 pc->ops->setup = PCSetUp_PBJacobi; 257 pc->ops->destroy = PCDestroy_PBJacobi; 258 pc->ops->setfromoptions = NULL; 259 pc->ops->view = PCView_PBJacobi; 260 pc->ops->applyrichardson = NULL; 261 pc->ops->applysymmetricleft = NULL; 262 pc->ops->applysymmetricright = NULL; 263 PetscFunctionReturn(0); 264 } 265