1 #include <../src/ksp/pc/impls/bddc/bddc.h> 2 #include <../src/ksp/pc/impls/bddc/bddcprivate.h> 3 #include <petscblaslapack.h> 4 5 /* prototypes for deluxe functions */ 6 static PetscErrorCode PCBDDCScalingCreate_Deluxe(PC); 7 static PetscErrorCode PCBDDCScalingDestroy_Deluxe(PC); 8 static PetscErrorCode PCBDDCScalingSetUp_Deluxe(PC); 9 static PetscErrorCode PCBDDCScalingSetUp_Deluxe_Private(PC); 10 static PetscErrorCode PCBDDCScalingReset_Deluxe_Solvers(PCBDDCDeluxeScaling); 11 12 #undef __FUNCT__ 13 #define __FUNCT__ "PCBDDCScalingExtension_Basic" 14 static PetscErrorCode PCBDDCScalingExtension_Basic(PC pc, Vec local_interface_vector, Vec global_vector) 15 { 16 PC_IS* pcis = (PC_IS*)pc->data; 17 PC_BDDC* pcbddc = (PC_BDDC*)pc->data; 18 PetscErrorCode ierr; 19 20 PetscFunctionBegin; 21 /* Apply partition of unity */ 22 ierr = VecPointwiseMult(pcbddc->work_scaling,pcis->D,local_interface_vector);CHKERRQ(ierr); 23 ierr = VecSet(global_vector,0.0);CHKERRQ(ierr); 24 ierr = VecScatterBegin(pcis->global_to_B,pcbddc->work_scaling,global_vector,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 25 ierr = VecScatterEnd(pcis->global_to_B,pcbddc->work_scaling,global_vector,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 26 PetscFunctionReturn(0); 27 } 28 29 #undef __FUNCT__ 30 #define __FUNCT__ "PCBDDCScalingExtension_Deluxe" 31 static PetscErrorCode PCBDDCScalingExtension_Deluxe(PC pc, Vec x, Vec y) 32 { 33 PC_IS* pcis=(PC_IS*)pc->data; 34 PC_BDDC* pcbddc=(PC_BDDC*)pc->data; 35 PCBDDCDeluxeScaling deluxe_ctx = pcbddc->deluxe_ctx; 36 PetscErrorCode ierr; 37 38 PetscFunctionBegin; 39 ierr = VecSet(pcbddc->work_scaling,0.0);CHKERRQ(ierr); 40 ierr = VecSet(y,0.0);CHKERRQ(ierr); 41 if (deluxe_ctx->n_simple) { /* scale deluxe vertices using diagonal scaling */ 42 PetscInt i; 43 const PetscScalar *array_x,*array_D; 44 PetscScalar *array; 45 ierr = VecGetArrayRead(x,&array_x);CHKERRQ(ierr); 46 ierr = VecGetArrayRead(pcis->D,&array_D);CHKERRQ(ierr); 47 ierr = VecGetArray(pcbddc->work_scaling,&array);CHKERRQ(ierr); 48 for (i=0;i<deluxe_ctx->n_simple;i++) { 49 array[deluxe_ctx->idx_simple_B[i]] = array_x[deluxe_ctx->idx_simple_B[i]]*array_D[deluxe_ctx->idx_simple_B[i]]; 50 } 51 ierr = VecRestoreArray(pcbddc->work_scaling,&array);CHKERRQ(ierr); 52 ierr = VecRestoreArrayRead(pcis->D,&array_D);CHKERRQ(ierr); 53 ierr = VecRestoreArrayRead(x,&array_x);CHKERRQ(ierr); 54 } 55 /* sequential part : all problems and Schur applications collapsed into a single matrix vector multiplication or a matvec and a solve */ 56 if (deluxe_ctx->seq_mat) { 57 PetscInt i; 58 for (i=0;i<deluxe_ctx->seq_n;i++) { 59 if (deluxe_ctx->change) { 60 ierr = VecScatterBegin(deluxe_ctx->seq_scctx[i],x,deluxe_ctx->seq_work2[i],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 61 ierr = VecScatterEnd(deluxe_ctx->seq_scctx[i],x,deluxe_ctx->seq_work2[i],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 62 if (deluxe_ctx->change_with_qr) { 63 Mat change; 64 65 ierr = KSPGetOperators(deluxe_ctx->change[i],&change,NULL);CHKERRQ(ierr); 66 ierr = MatMultTranspose(change,deluxe_ctx->seq_work2[i],deluxe_ctx->seq_work1[i]);CHKERRQ(ierr); 67 } else { 68 ierr = KSPSolve(deluxe_ctx->change[i],deluxe_ctx->seq_work2[i],deluxe_ctx->seq_work1[i]);CHKERRQ(ierr); 69 } 70 } else { 71 ierr = VecScatterBegin(deluxe_ctx->seq_scctx[i],x,deluxe_ctx->seq_work1[i],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 72 ierr = VecScatterEnd(deluxe_ctx->seq_scctx[i],x,deluxe_ctx->seq_work1[i],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 73 } 74 ierr = MatMultTranspose(deluxe_ctx->seq_mat[i],deluxe_ctx->seq_work1[i],deluxe_ctx->seq_work2[i]);CHKERRQ(ierr); 75 if (deluxe_ctx->seq_mat_inv_sum[i]) { 76 PetscScalar *x; 77 78 ierr = VecGetArray(deluxe_ctx->seq_work2[i],&x);CHKERRQ(ierr); 79 ierr = VecPlaceArray(deluxe_ctx->seq_work1[i],x);CHKERRQ(ierr); 80 ierr = VecRestoreArray(deluxe_ctx->seq_work2[i],&x);CHKERRQ(ierr); 81 ierr = MatSolveTranspose(deluxe_ctx->seq_mat_inv_sum[i],deluxe_ctx->seq_work1[i],deluxe_ctx->seq_work2[i]);CHKERRQ(ierr); 82 ierr = VecResetArray(deluxe_ctx->seq_work1[i]);CHKERRQ(ierr); 83 } 84 if (deluxe_ctx->change) { 85 Mat change; 86 87 ierr = KSPGetOperators(deluxe_ctx->change[i],&change,NULL);CHKERRQ(ierr); 88 ierr = MatMult(change,deluxe_ctx->seq_work2[i],deluxe_ctx->seq_work1[i]);CHKERRQ(ierr); 89 ierr = VecScatterBegin(deluxe_ctx->seq_scctx[i],deluxe_ctx->seq_work1[i],pcbddc->work_scaling,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 90 ierr = VecScatterEnd(deluxe_ctx->seq_scctx[i],deluxe_ctx->seq_work1[i],pcbddc->work_scaling,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 91 } else { 92 ierr = VecScatterBegin(deluxe_ctx->seq_scctx[i],deluxe_ctx->seq_work2[i],pcbddc->work_scaling,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 93 ierr = VecScatterEnd(deluxe_ctx->seq_scctx[i],deluxe_ctx->seq_work2[i],pcbddc->work_scaling,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 94 } 95 } 96 } 97 /* put local boundary part in global vector */ 98 ierr = VecScatterBegin(pcis->global_to_B,pcbddc->work_scaling,y,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 99 ierr = VecScatterEnd(pcis->global_to_B,pcbddc->work_scaling,y,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 100 PetscFunctionReturn(0); 101 } 102 103 #undef __FUNCT__ 104 #define __FUNCT__ "PCBDDCScalingExtension" 105 PetscErrorCode PCBDDCScalingExtension(PC pc, Vec local_interface_vector, Vec global_vector) 106 { 107 PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 108 PetscErrorCode ierr; 109 110 PetscFunctionBegin; 111 PetscValidHeaderSpecific(pc,PC_CLASSID,1); 112 PetscValidHeaderSpecific(local_interface_vector,VEC_CLASSID,2); 113 PetscValidHeaderSpecific(global_vector,VEC_CLASSID,3); 114 if (local_interface_vector == pcbddc->work_scaling) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Local vector cannot be pcbddc->work_scaling!\n"); 115 ierr = PetscUseMethod(pc,"PCBDDCScalingExtension_C",(PC,Vec,Vec),(pc,local_interface_vector,global_vector));CHKERRQ(ierr); 116 PetscFunctionReturn(0); 117 } 118 119 #undef __FUNCT__ 120 #define __FUNCT__ "PCBDDCScalingRestriction_Basic" 121 static PetscErrorCode PCBDDCScalingRestriction_Basic(PC pc, Vec global_vector, Vec local_interface_vector) 122 { 123 PetscErrorCode ierr; 124 PC_IS *pcis = (PC_IS*)pc->data; 125 126 PetscFunctionBegin; 127 ierr = VecScatterBegin(pcis->global_to_B,global_vector,local_interface_vector,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 128 ierr = VecScatterEnd(pcis->global_to_B,global_vector,local_interface_vector,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 129 /* Apply partition of unity */ 130 ierr = VecPointwiseMult(local_interface_vector,pcis->D,local_interface_vector);CHKERRQ(ierr); 131 PetscFunctionReturn(0); 132 } 133 134 #undef __FUNCT__ 135 #define __FUNCT__ "PCBDDCScalingRestriction_Deluxe" 136 static PetscErrorCode PCBDDCScalingRestriction_Deluxe(PC pc, Vec x, Vec y) 137 { 138 PC_IS* pcis=(PC_IS*)pc->data; 139 PC_BDDC* pcbddc=(PC_BDDC*)pc->data; 140 PCBDDCDeluxeScaling deluxe_ctx = pcbddc->deluxe_ctx; 141 PetscErrorCode ierr; 142 143 PetscFunctionBegin; 144 /* get local boundary part of global vector */ 145 ierr = VecScatterBegin(pcis->global_to_B,x,y,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 146 ierr = VecScatterEnd(pcis->global_to_B,x,y,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 147 if (deluxe_ctx->n_simple) { /* scale deluxe vertices using diagonal scaling */ 148 PetscInt i; 149 PetscScalar *array_y; 150 const PetscScalar *array_D; 151 ierr = VecGetArray(y,&array_y);CHKERRQ(ierr); 152 ierr = VecGetArrayRead(pcis->D,&array_D);CHKERRQ(ierr); 153 for (i=0;i<deluxe_ctx->n_simple;i++) { 154 array_y[deluxe_ctx->idx_simple_B[i]] *= array_D[deluxe_ctx->idx_simple_B[i]]; 155 } 156 ierr = VecRestoreArrayRead(pcis->D,&array_D);CHKERRQ(ierr); 157 ierr = VecRestoreArray(y,&array_y);CHKERRQ(ierr); 158 } 159 /* sequential part : all problems and Schur applications collapsed into a single matrix vector multiplication and ksp solution */ 160 if (deluxe_ctx->seq_mat) { 161 PetscInt i; 162 for (i=0;i<deluxe_ctx->seq_n;i++) { 163 if (deluxe_ctx->change) { 164 Mat change; 165 166 ierr = VecScatterBegin(deluxe_ctx->seq_scctx[i],y,deluxe_ctx->seq_work2[i],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 167 ierr = VecScatterEnd(deluxe_ctx->seq_scctx[i],y,deluxe_ctx->seq_work2[i],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 168 ierr = KSPGetOperators(deluxe_ctx->change[i],&change,NULL);CHKERRQ(ierr); 169 ierr = MatMultTranspose(change,deluxe_ctx->seq_work2[i],deluxe_ctx->seq_work1[i]);CHKERRQ(ierr); 170 } else { 171 ierr = VecScatterBegin(deluxe_ctx->seq_scctx[i],y,deluxe_ctx->seq_work1[i],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 172 ierr = VecScatterEnd(deluxe_ctx->seq_scctx[i],y,deluxe_ctx->seq_work1[i],INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 173 } 174 if (deluxe_ctx->seq_mat_inv_sum[i]) { 175 PetscScalar *x; 176 177 ierr = VecGetArray(deluxe_ctx->seq_work1[i],&x);CHKERRQ(ierr); 178 ierr = VecPlaceArray(deluxe_ctx->seq_work2[i],x);CHKERRQ(ierr); 179 ierr = VecRestoreArray(deluxe_ctx->seq_work1[i],&x);CHKERRQ(ierr); 180 ierr = MatSolve(deluxe_ctx->seq_mat_inv_sum[i],deluxe_ctx->seq_work1[i],deluxe_ctx->seq_work2[i]);CHKERRQ(ierr); 181 ierr = VecResetArray(deluxe_ctx->seq_work2[i]);CHKERRQ(ierr); 182 } 183 ierr = MatMult(deluxe_ctx->seq_mat[i],deluxe_ctx->seq_work1[i],deluxe_ctx->seq_work2[i]);CHKERRQ(ierr); 184 if (deluxe_ctx->change) { 185 if (deluxe_ctx->change_with_qr) { 186 Mat change; 187 188 ierr = KSPGetOperators(deluxe_ctx->change[i],&change,NULL);CHKERRQ(ierr); 189 ierr = MatMult(change,deluxe_ctx->seq_work2[i],deluxe_ctx->seq_work1[i]);CHKERRQ(ierr); 190 } else { 191 ierr = KSPSolveTranspose(deluxe_ctx->change[i],deluxe_ctx->seq_work2[i],deluxe_ctx->seq_work1[i]);CHKERRQ(ierr); 192 } 193 ierr = VecScatterBegin(deluxe_ctx->seq_scctx[i],deluxe_ctx->seq_work1[i],y,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 194 ierr = VecScatterEnd(deluxe_ctx->seq_scctx[i],deluxe_ctx->seq_work1[i],y,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 195 } else { 196 ierr = VecScatterBegin(deluxe_ctx->seq_scctx[i],deluxe_ctx->seq_work2[i],y,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 197 ierr = VecScatterEnd(deluxe_ctx->seq_scctx[i],deluxe_ctx->seq_work2[i],y,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 198 } 199 } 200 } 201 PetscFunctionReturn(0); 202 } 203 204 #undef __FUNCT__ 205 #define __FUNCT__ "PCBDDCScalingRestriction" 206 PetscErrorCode PCBDDCScalingRestriction(PC pc, Vec global_vector, Vec local_interface_vector) 207 { 208 PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 209 PetscErrorCode ierr; 210 211 PetscFunctionBegin; 212 PetscValidHeaderSpecific(pc,PC_CLASSID,1); 213 PetscValidHeaderSpecific(global_vector,VEC_CLASSID,2); 214 PetscValidHeaderSpecific(local_interface_vector,VEC_CLASSID,3); 215 if (local_interface_vector == pcbddc->work_scaling) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Local vector cannot be pcbddc->work_scaling!\n"); 216 ierr = PetscUseMethod(pc,"PCBDDCScalingRestriction_C",(PC,Vec,Vec),(pc,global_vector,local_interface_vector));CHKERRQ(ierr); 217 PetscFunctionReturn(0); 218 } 219 220 #undef __FUNCT__ 221 #define __FUNCT__ "PCBDDCScalingSetUp" 222 PetscErrorCode PCBDDCScalingSetUp(PC pc) 223 { 224 PC_IS* pcis=(PC_IS*)pc->data; 225 PC_BDDC* pcbddc=(PC_BDDC*)pc->data; 226 PetscErrorCode ierr; 227 228 PetscFunctionBegin; 229 PetscValidHeaderSpecific(pc,PC_CLASSID,1); 230 /* create work vector for the operator */ 231 ierr = VecDestroy(&pcbddc->work_scaling);CHKERRQ(ierr); 232 ierr = VecDuplicate(pcis->vec1_B,&pcbddc->work_scaling);CHKERRQ(ierr); 233 /* always rebuild pcis->D */ 234 if (pcis->use_stiffness_scaling) { 235 ierr = MatGetDiagonal(pcbddc->local_mat,pcis->vec1_N);CHKERRQ(ierr); 236 ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 237 ierr = VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 238 } 239 ierr = VecCopy(pcis->D,pcis->vec1_B);CHKERRQ(ierr); 240 ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 241 ierr = VecScatterBegin(pcis->global_to_B,pcis->vec1_B,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 242 ierr = VecScatterEnd(pcis->global_to_B,pcis->vec1_B,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 243 ierr = VecScatterBegin(pcis->global_to_B,pcis->vec1_global,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 244 ierr = VecScatterEnd(pcis->global_to_B,pcis->vec1_global,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 245 ierr = VecPointwiseDivide(pcis->D,pcis->D,pcis->vec1_B);CHKERRQ(ierr); 246 /* now setup */ 247 if (pcbddc->use_deluxe_scaling) { 248 if (!pcbddc->deluxe_ctx) { 249 ierr = PCBDDCScalingCreate_Deluxe(pc);CHKERRQ(ierr); 250 } 251 ierr = PCBDDCScalingSetUp_Deluxe(pc);CHKERRQ(ierr); 252 ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCScalingRestriction_C",PCBDDCScalingRestriction_Deluxe);CHKERRQ(ierr); 253 ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCScalingExtension_C",PCBDDCScalingExtension_Deluxe);CHKERRQ(ierr); 254 } else { 255 ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCScalingRestriction_C",PCBDDCScalingRestriction_Basic);CHKERRQ(ierr); 256 ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCScalingExtension_C",PCBDDCScalingExtension_Basic);CHKERRQ(ierr); 257 } 258 259 /* test */ 260 if (pcbddc->dbg_flag) { 261 Mat B0_B = NULL; 262 Vec B0_Bv = NULL, B0_Bv2 = NULL; 263 Vec vec2_global; 264 PetscViewer viewer = pcbddc->dbg_viewer; 265 PetscReal error; 266 267 /* extension -> from local to parallel */ 268 ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 269 ierr = VecSetRandom(pcis->vec1_B,NULL);CHKERRQ(ierr); 270 ierr = VecScatterBegin(pcis->global_to_B,pcis->vec1_B,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 271 ierr = VecScatterEnd(pcis->global_to_B,pcis->vec1_B,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 272 ierr = VecDuplicate(pcis->vec1_global,&vec2_global);CHKERRQ(ierr); 273 ierr = VecCopy(pcis->vec1_global,vec2_global);CHKERRQ(ierr); 274 ierr = VecScatterBegin(pcis->global_to_B,pcis->vec1_global,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 275 ierr = VecScatterEnd(pcis->global_to_B,pcis->vec1_global,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 276 if (pcbddc->benign_n) { 277 IS is_dummy; 278 279 ierr = ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);CHKERRQ(ierr); 280 ierr = MatGetSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);CHKERRQ(ierr); 281 ierr = ISDestroy(&is_dummy);CHKERRQ(ierr); 282 ierr = MatCreateVecs(B0_B,NULL,&B0_Bv);CHKERRQ(ierr); 283 ierr = VecDuplicate(B0_Bv,&B0_Bv2);CHKERRQ(ierr); 284 ierr = MatMult(B0_B,pcis->vec1_B,B0_Bv);CHKERRQ(ierr); 285 } 286 ierr = PCBDDCScalingExtension(pc,pcis->vec1_B,pcis->vec1_global);CHKERRQ(ierr); 287 if (pcbddc->benign_saddle_point) { 288 PetscReal errorl = 0.; 289 ierr = VecScatterBegin(pcis->global_to_B,pcis->vec1_global,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 290 ierr = VecScatterEnd(pcis->global_to_B,pcis->vec1_global,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 291 if (pcbddc->benign_n) { 292 ierr = MatMult(B0_B,pcis->vec1_B,B0_Bv2);CHKERRQ(ierr); 293 ierr = VecAXPY(B0_Bv,-1.0,B0_Bv2);CHKERRQ(ierr); 294 ierr = VecNorm(B0_Bv,NORM_INFINITY,&errorl);CHKERRQ(ierr); 295 } 296 ierr = MPI_Allreduce(&errorl,&error,1,MPIU_REAL,MPI_SUM,PetscObjectComm((PetscObject)pc));CHKERRQ(ierr); 297 ierr = PetscViewerASCIIPrintf(viewer,"Error benign extension %1.14e\n",error);CHKERRQ(ierr); 298 } 299 ierr = VecAXPY(pcis->vec1_global,-1.0,vec2_global);CHKERRQ(ierr); 300 ierr = VecNorm(pcis->vec1_global,NORM_INFINITY,&error);CHKERRQ(ierr); 301 ierr = PetscViewerASCIIPrintf(viewer,"Error scaling extension %1.14e\n",error);CHKERRQ(ierr); 302 ierr = VecDestroy(&vec2_global);CHKERRQ(ierr); 303 304 /* restriction -> from parallel to local */ 305 ierr = VecSet(pcis->vec1_global,0.0);CHKERRQ(ierr); 306 ierr = VecSetRandom(pcis->vec1_B,NULL);CHKERRQ(ierr); 307 ierr = VecScatterBegin(pcis->global_to_B,pcis->vec1_B,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 308 ierr = VecScatterEnd(pcis->global_to_B,pcis->vec1_B,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 309 ierr = PCBDDCScalingRestriction(pc,pcis->vec1_global,pcis->vec1_B);CHKERRQ(ierr); 310 ierr = VecScale(pcis->vec1_B,-1.0);CHKERRQ(ierr); 311 ierr = VecScatterBegin(pcis->global_to_B,pcis->vec1_B,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 312 ierr = VecScatterEnd(pcis->global_to_B,pcis->vec1_B,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 313 ierr = VecNorm(pcis->vec1_global,NORM_INFINITY,&error);CHKERRQ(ierr); 314 ierr = PetscViewerASCIIPrintf(viewer,"Error scaling restriction %1.14e\n",error);CHKERRQ(ierr); 315 ierr = MatDestroy(&B0_B);CHKERRQ(ierr); 316 ierr = VecDestroy(&B0_Bv);CHKERRQ(ierr); 317 ierr = VecDestroy(&B0_Bv2);CHKERRQ(ierr); 318 } 319 PetscFunctionReturn(0); 320 } 321 322 #undef __FUNCT__ 323 #define __FUNCT__ "PCBDDCScalingDestroy" 324 PetscErrorCode PCBDDCScalingDestroy(PC pc) 325 { 326 PC_BDDC* pcbddc=(PC_BDDC*)pc->data; 327 PetscErrorCode ierr; 328 329 PetscFunctionBegin; 330 if (pcbddc->deluxe_ctx) { 331 ierr = PCBDDCScalingDestroy_Deluxe(pc);CHKERRQ(ierr); 332 } 333 ierr = VecDestroy(&pcbddc->work_scaling);CHKERRQ(ierr); 334 ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCScalingRestriction_C",NULL);CHKERRQ(ierr); 335 ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCScalingExtension_C",NULL);CHKERRQ(ierr); 336 PetscFunctionReturn(0); 337 } 338 339 #undef __FUNCT__ 340 #define __FUNCT__ "PCBDDCScalingCreate_Deluxe" 341 static PetscErrorCode PCBDDCScalingCreate_Deluxe(PC pc) 342 { 343 PC_BDDC* pcbddc=(PC_BDDC*)pc->data; 344 PCBDDCDeluxeScaling deluxe_ctx; 345 PetscErrorCode ierr; 346 347 PetscFunctionBegin; 348 ierr = PetscNew(&deluxe_ctx);CHKERRQ(ierr); 349 pcbddc->deluxe_ctx = deluxe_ctx; 350 PetscFunctionReturn(0); 351 } 352 353 #undef __FUNCT__ 354 #define __FUNCT__ "PCBDDCScalingDestroy_Deluxe" 355 static PetscErrorCode PCBDDCScalingDestroy_Deluxe(PC pc) 356 { 357 PC_BDDC* pcbddc=(PC_BDDC*)pc->data; 358 PetscErrorCode ierr; 359 360 PetscFunctionBegin; 361 ierr = PCBDDCScalingReset_Deluxe_Solvers(pcbddc->deluxe_ctx);CHKERRQ(ierr); 362 ierr = PetscFree(pcbddc->deluxe_ctx);CHKERRQ(ierr); 363 PetscFunctionReturn(0); 364 } 365 366 #undef __FUNCT__ 367 #define __FUNCT__ "PCBDDCScalingReset_Deluxe_Solvers" 368 static PetscErrorCode PCBDDCScalingReset_Deluxe_Solvers(PCBDDCDeluxeScaling deluxe_ctx) 369 { 370 PetscInt i; 371 PetscErrorCode ierr; 372 373 PetscFunctionBegin; 374 ierr = PetscFree(deluxe_ctx->idx_simple_B);CHKERRQ(ierr); 375 deluxe_ctx->n_simple = 0; 376 for (i=0;i<deluxe_ctx->seq_n;i++) { 377 ierr = VecScatterDestroy(&deluxe_ctx->seq_scctx[i]);CHKERRQ(ierr); 378 ierr = VecDestroy(&deluxe_ctx->seq_work1[i]);CHKERRQ(ierr); 379 ierr = VecDestroy(&deluxe_ctx->seq_work2[i]);CHKERRQ(ierr); 380 ierr = MatDestroy(&deluxe_ctx->seq_mat[i]);CHKERRQ(ierr); 381 ierr = MatDestroy(&deluxe_ctx->seq_mat_inv_sum[i]);CHKERRQ(ierr); 382 } 383 ierr = PetscFree5(deluxe_ctx->seq_scctx,deluxe_ctx->seq_work1,deluxe_ctx->seq_work2,deluxe_ctx->seq_mat,deluxe_ctx->seq_mat_inv_sum);CHKERRQ(ierr); 384 ierr = PetscFree(deluxe_ctx->workspace);CHKERRQ(ierr); 385 deluxe_ctx->seq_n = 0; 386 PetscFunctionReturn(0); 387 } 388 389 #undef __FUNCT__ 390 #define __FUNCT__ "PCBDDCScalingSetUp_Deluxe" 391 static PetscErrorCode PCBDDCScalingSetUp_Deluxe(PC pc) 392 { 393 PC_IS *pcis=(PC_IS*)pc->data; 394 PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 395 PCBDDCDeluxeScaling deluxe_ctx=pcbddc->deluxe_ctx; 396 PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs; 397 PetscErrorCode ierr; 398 399 PetscFunctionBegin; 400 /* reset data structures if the topology has changed */ 401 if (pcbddc->recompute_topography) { 402 ierr = PCBDDCScalingReset_Deluxe_Solvers(deluxe_ctx);CHKERRQ(ierr); 403 } 404 405 /* Compute data structures to solve sequential problems */ 406 ierr = PCBDDCScalingSetUp_Deluxe_Private(pc);CHKERRQ(ierr); 407 408 /* diagonal scaling on interface dofs not contained in cc */ 409 if (sub_schurs->is_vertices || sub_schurs->is_dir) { 410 PetscInt n_com,n_dir; 411 n_com = 0; 412 if (sub_schurs->is_vertices) { 413 ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_com);CHKERRQ(ierr); 414 } 415 n_dir = 0; 416 if (sub_schurs->is_dir) { 417 ierr = ISGetLocalSize(sub_schurs->is_dir,&n_dir);CHKERRQ(ierr); 418 } 419 if (!deluxe_ctx->n_simple) { 420 deluxe_ctx->n_simple = n_dir + n_com; 421 ierr = PetscMalloc1(deluxe_ctx->n_simple,&deluxe_ctx->idx_simple_B);CHKERRQ(ierr); 422 if (sub_schurs->is_vertices) { 423 PetscInt nmap; 424 const PetscInt *idxs; 425 426 ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 427 ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_com,idxs,&nmap,deluxe_ctx->idx_simple_B);CHKERRQ(ierr); 428 if (nmap != n_com) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error when mapping simply scaled dofs (is_vertices)! %d != %d",nmap,n_com); 429 ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 430 } 431 if (sub_schurs->is_dir) { 432 PetscInt nmap; 433 const PetscInt *idxs; 434 435 ierr = ISGetIndices(sub_schurs->is_dir,&idxs);CHKERRQ(ierr); 436 ierr = ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_dir,idxs,&nmap,deluxe_ctx->idx_simple_B+n_com);CHKERRQ(ierr); 437 if (nmap != n_dir) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error when mapping simply scaled dofs (sub_schurs->is_dir)! %d != %d",nmap,n_dir); 438 ierr = ISRestoreIndices(sub_schurs->is_dir,&idxs);CHKERRQ(ierr); 439 } 440 ierr = PetscSortInt(deluxe_ctx->n_simple,deluxe_ctx->idx_simple_B);CHKERRQ(ierr); 441 } else { 442 if (deluxe_ctx->n_simple != n_dir + n_com) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Number of simply scaled dofs %d is different from the previous one computed %d",n_dir + n_com,deluxe_ctx->n_simple); 443 } 444 } else { 445 deluxe_ctx->n_simple = 0; 446 deluxe_ctx->idx_simple_B = 0; 447 } 448 PetscFunctionReturn(0); 449 } 450 451 #undef __FUNCT__ 452 #define __FUNCT__ "PCBDDCScalingSetUp_Deluxe_Private" 453 static PetscErrorCode PCBDDCScalingSetUp_Deluxe_Private(PC pc) 454 { 455 PC_BDDC *pcbddc=(PC_BDDC*)pc->data; 456 PCBDDCDeluxeScaling deluxe_ctx=pcbddc->deluxe_ctx; 457 PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs; 458 PetscScalar *matdata,*matdata2; 459 PetscInt i,max_subset_size,cum,cum2; 460 const PetscInt *idxs; 461 PetscBool newsetup = PETSC_FALSE; 462 PetscErrorCode ierr; 463 464 PetscFunctionBegin; 465 if (!sub_schurs->n_subs) { 466 PetscFunctionReturn(0); 467 } 468 469 /* Allocate arrays for subproblems */ 470 if (!deluxe_ctx->seq_n) { 471 deluxe_ctx->seq_n = sub_schurs->n_subs; 472 ierr = PetscCalloc5(deluxe_ctx->seq_n,&deluxe_ctx->seq_scctx,deluxe_ctx->seq_n,&deluxe_ctx->seq_work1,deluxe_ctx->seq_n,&deluxe_ctx->seq_work2,deluxe_ctx->seq_n,&deluxe_ctx->seq_mat,deluxe_ctx->seq_n,&deluxe_ctx->seq_mat_inv_sum);CHKERRQ(ierr); 473 newsetup = PETSC_TRUE; 474 } else if (deluxe_ctx->seq_n != sub_schurs->n_subs) { 475 SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Number of deluxe subproblems %d is different from the sub_schurs %d",deluxe_ctx->seq_n,sub_schurs->n_subs); 476 } 477 deluxe_ctx->change_with_qr = sub_schurs->change_with_qr; 478 479 /* Create objects for deluxe */ 480 max_subset_size = 0; 481 for (i=0;i<sub_schurs->n_subs;i++) { 482 PetscInt subset_size; 483 ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 484 max_subset_size = PetscMax(subset_size,max_subset_size); 485 } 486 if (newsetup) { 487 ierr = PetscMalloc1(2*max_subset_size,&deluxe_ctx->workspace);CHKERRQ(ierr); 488 } 489 cum = cum2 = 0; 490 ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs);CHKERRQ(ierr); 491 ierr = MatSeqAIJGetArray(sub_schurs->S_Ej_all,&matdata);CHKERRQ(ierr); 492 matdata2 = NULL; 493 if (sub_schurs->sum_S_Ej_all) { 494 ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_all,&matdata2);CHKERRQ(ierr); 495 } 496 for (i=0;i<deluxe_ctx->seq_n;i++) { 497 PetscInt subset_size; 498 499 ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 500 if (newsetup) { 501 IS sub; 502 /* work vectors */ 503 ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,subset_size,deluxe_ctx->workspace,&deluxe_ctx->seq_work1[i]);CHKERRQ(ierr); 504 ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,subset_size,deluxe_ctx->workspace+subset_size,&deluxe_ctx->seq_work2[i]);CHKERRQ(ierr); 505 506 /* scatters */ 507 ierr = ISCreateGeneral(PETSC_COMM_SELF,subset_size,idxs+cum,PETSC_COPY_VALUES,&sub);CHKERRQ(ierr); 508 ierr = VecScatterCreate(pcbddc->work_scaling,sub,deluxe_ctx->seq_work1[i],NULL,&deluxe_ctx->seq_scctx[i]);CHKERRQ(ierr); 509 ierr = ISDestroy(&sub);CHKERRQ(ierr); 510 511 /* S_E_j */ 512 ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,subset_size,matdata+cum2,&deluxe_ctx->seq_mat[i]);CHKERRQ(ierr); 513 } 514 /* \sum_k S^k_E_j */ 515 if (matdata2) { 516 ierr = MatDestroy(&deluxe_ctx->seq_mat_inv_sum[i]);CHKERRQ(ierr); 517 ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,subset_size,matdata2+cum2,&deluxe_ctx->seq_mat_inv_sum[i]);CHKERRQ(ierr); 518 if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 519 ierr = MatCholeskyFactor(deluxe_ctx->seq_mat_inv_sum[i],NULL,NULL);CHKERRQ(ierr); 520 } else { 521 ierr = MatLUFactor(deluxe_ctx->seq_mat_inv_sum[i],NULL,NULL,NULL);CHKERRQ(ierr); 522 } 523 } 524 cum += subset_size; 525 cum2 += subset_size*subset_size; 526 } 527 ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs);CHKERRQ(ierr); 528 ierr = MatSeqAIJRestoreArray(sub_schurs->S_Ej_all,&matdata);CHKERRQ(ierr); 529 if (sub_schurs->sum_S_Ej_all) { 530 ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_all,&matdata2);CHKERRQ(ierr); 531 } 532 533 /* the change of basis is just a reference to sub_schurs->change (if any) */ 534 deluxe_ctx->change = sub_schurs->change; 535 if (deluxe_ctx->change && !deluxe_ctx->change_with_qr) { 536 for (i=0;i<deluxe_ctx->seq_n;i++) { 537 if (newsetup) { 538 PC pc; 539 540 ierr = KSPGetPC(deluxe_ctx->change[i],&pc);CHKERRQ(ierr); 541 ierr = PCSetType(pc,PCLU);CHKERRQ(ierr); 542 ierr = KSPSetFromOptions(deluxe_ctx->change[i]);CHKERRQ(ierr); 543 } 544 ierr = KSPSetUp(deluxe_ctx->change[i]);CHKERRQ(ierr); 545 } 546 } 547 PetscFunctionReturn(0); 548 } 549