134a97f8cSStefano Zampini #include <../src/ksp/pc/impls/bddc/bddc.h> 234a97f8cSStefano Zampini #include <../src/ksp/pc/impls/bddc/bddcprivate.h> 308122e43SStefano Zampini #include <petscblaslapack.h> 434a97f8cSStefano Zampini 53202ece2SStefano Zampini PETSC_STATIC_INLINE PetscErrorCode PCBDDCAdjGetNextLayer_Private(PetscInt*,PetscInt,PetscBT,PetscInt*,PetscInt*,PetscInt*); 65ec10c6aSStefano Zampini static PetscErrorCode PCBDDCComputeExplicitSchur(Mat,PetscBool,MatReuse,Mat*); 7df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolvers_Interior(PC,Vec,Vec); 8df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolvers_Correction(PC,Vec,Vec); 9d62866d3SStefano Zampini 10ca92afb2SStefano Zampini /* if v2 is not present, correction is done in-place */ 11ca92afb2SStefano Zampini #undef __FUNCT__ 125cbda25cSStefano Zampini #define __FUNCT__ "PCBDDCReuseSolversBenignAdapt" 135cbda25cSStefano Zampini PetscErrorCode PCBDDCReuseSolversBenignAdapt(PCBDDCReuseSolvers ctx, Vec v, Vec v2, PetscBool sol, PetscBool full) 14ca92afb2SStefano Zampini { 15ca92afb2SStefano Zampini PetscScalar *array; 16ca92afb2SStefano Zampini PetscScalar *array2; 17ca92afb2SStefano Zampini PetscErrorCode ierr; 18ca92afb2SStefano Zampini 19ca92afb2SStefano Zampini PetscFunctionBegin; 20ca92afb2SStefano Zampini if (!ctx->benign_n) PetscFunctionReturn(0); 215cbda25cSStefano Zampini if (sol && full) { 225cbda25cSStefano Zampini PetscInt n_I,size_schur; 235cbda25cSStefano Zampini 245cbda25cSStefano Zampini /* get sizes */ 25282d6408SStefano Zampini ierr = MatGetSize(ctx->benign_csAIB,&size_schur,NULL);CHKERRQ(ierr); 265cbda25cSStefano Zampini ierr = VecGetSize(v,&n_I);CHKERRQ(ierr); 275cbda25cSStefano Zampini n_I = n_I - size_schur; 285cbda25cSStefano Zampini /* get schur sol from array */ 295cbda25cSStefano Zampini ierr = VecGetArray(v,&array);CHKERRQ(ierr); 305cbda25cSStefano Zampini ierr = VecPlaceArray(ctx->benign_dummy_schur_vec,array+n_I);CHKERRQ(ierr); 315cbda25cSStefano Zampini ierr = VecRestoreArray(v,&array);CHKERRQ(ierr); 325cbda25cSStefano Zampini /* apply interior sol correction */ 33282d6408SStefano Zampini ierr = MatMultTranspose(ctx->benign_csAIB,ctx->benign_dummy_schur_vec,ctx->benign_corr_work);CHKERRQ(ierr); 345cbda25cSStefano Zampini ierr = VecResetArray(ctx->benign_dummy_schur_vec);CHKERRQ(ierr); 355cbda25cSStefano Zampini ierr = MatMultAdd(ctx->benign_AIIm1ones,ctx->benign_corr_work,v,v);CHKERRQ(ierr); 365cbda25cSStefano Zampini } 37ca92afb2SStefano Zampini if (v2) { 38ca92afb2SStefano Zampini PetscInt nl; 39ca92afb2SStefano Zampini 40ca92afb2SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 41ca92afb2SStefano Zampini ierr = VecGetLocalSize(v2,&nl);CHKERRQ(ierr); 42ca92afb2SStefano Zampini ierr = VecGetArray(v2,&array2);CHKERRQ(ierr); 43ca92afb2SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(PetscScalar));CHKERRQ(ierr); 44ca92afb2SStefano Zampini } else { 45ca92afb2SStefano Zampini ierr = VecGetArray(v,&array);CHKERRQ(ierr); 46ca92afb2SStefano Zampini array2 = array; 47ca92afb2SStefano Zampini } 48ca92afb2SStefano Zampini if (!sol) { /* change rhs */ 49ca92afb2SStefano Zampini PetscInt n; 50ca92afb2SStefano Zampini for (n=0;n<ctx->benign_n;n++) { 51ca92afb2SStefano Zampini PetscScalar sum = 0.; 52ca92afb2SStefano Zampini const PetscInt *cols; 53ca92afb2SStefano Zampini PetscInt nz,i; 54ca92afb2SStefano Zampini 55ca92afb2SStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[n],&nz);CHKERRQ(ierr); 56ca92afb2SStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[n],&cols);CHKERRQ(ierr); 57ca92afb2SStefano Zampini for (i=0;i<nz-1;i++) sum += array[cols[i]]; 5822db5ddcSStefano Zampini #if defined(PETSC_USE_COMPLEX) 5922db5ddcSStefano Zampini sum = -(PetscRealPart(sum)/nz + PETSC_i*(PetscImaginaryPart(sum)/nz)); 6022db5ddcSStefano Zampini #else 61ca92afb2SStefano Zampini sum = -sum/nz; 6222db5ddcSStefano Zampini #endif 63ca92afb2SStefano Zampini for (i=0;i<nz-1;i++) array2[cols[i]] += sum; 64ca92afb2SStefano Zampini ctx->benign_save_vals[n] = array2[cols[nz-1]]; 65ca92afb2SStefano Zampini array2[cols[nz-1]] = sum; 66ca92afb2SStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[n],&cols);CHKERRQ(ierr); 67ca92afb2SStefano Zampini } 68ca92afb2SStefano Zampini } else { 69ca92afb2SStefano Zampini PetscInt n; 70ca92afb2SStefano Zampini for (n=0;n<ctx->benign_n;n++) { 71ca92afb2SStefano Zampini PetscScalar sum = 0.; 72ca92afb2SStefano Zampini const PetscInt *cols; 73ca92afb2SStefano Zampini PetscInt nz,i; 74ca92afb2SStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[n],&nz);CHKERRQ(ierr); 75ca92afb2SStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[n],&cols);CHKERRQ(ierr); 76ca92afb2SStefano Zampini for (i=0;i<nz-1;i++) sum += array[cols[i]]; 7722db5ddcSStefano Zampini #if defined(PETSC_USE_COMPLEX) 7822db5ddcSStefano Zampini sum = -(PetscRealPart(sum)/nz + PETSC_i*(PetscImaginaryPart(sum)/nz)); 7922db5ddcSStefano Zampini #else 80ca92afb2SStefano Zampini sum = -sum/nz; 8122db5ddcSStefano Zampini #endif 82ca92afb2SStefano Zampini for (i=0;i<nz-1;i++) array2[cols[i]] += sum; 83ca92afb2SStefano Zampini array2[cols[nz-1]] = ctx->benign_save_vals[n]; 84ca92afb2SStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[n],&cols);CHKERRQ(ierr); 85ca92afb2SStefano Zampini } 86ca92afb2SStefano Zampini } 87ca92afb2SStefano Zampini if (v2) { 88ca92afb2SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 8972b8c272SStefano Zampini ierr = VecRestoreArray(v2,&array2);CHKERRQ(ierr); 90ca92afb2SStefano Zampini } else { 91ca92afb2SStefano Zampini ierr = VecRestoreArray(v,&array);CHKERRQ(ierr); 92ca92afb2SStefano Zampini } 935cbda25cSStefano Zampini if (!sol && full) { 945cbda25cSStefano Zampini Vec usedv; 955cbda25cSStefano Zampini PetscInt n_I,size_schur; 965cbda25cSStefano Zampini 975cbda25cSStefano Zampini /* get sizes */ 98282d6408SStefano Zampini ierr = MatGetSize(ctx->benign_csAIB,&size_schur,NULL);CHKERRQ(ierr); 995cbda25cSStefano Zampini ierr = VecGetSize(v,&n_I);CHKERRQ(ierr); 1005cbda25cSStefano Zampini n_I = n_I - size_schur; 1015cbda25cSStefano Zampini /* compute schur rhs correction */ 1025cbda25cSStefano Zampini if (v2) { 1035cbda25cSStefano Zampini usedv = v2; 1045cbda25cSStefano Zampini } else { 1055cbda25cSStefano Zampini usedv = v; 1065cbda25cSStefano Zampini } 1075cbda25cSStefano Zampini /* apply schur rhs correction */ 1085cbda25cSStefano Zampini ierr = MatMultTranspose(ctx->benign_AIIm1ones,usedv,ctx->benign_corr_work);CHKERRQ(ierr); 1095cbda25cSStefano Zampini ierr = VecGetArrayRead(usedv,(const PetscScalar**)&array);CHKERRQ(ierr); 1105cbda25cSStefano Zampini ierr = VecPlaceArray(ctx->benign_dummy_schur_vec,array+n_I);CHKERRQ(ierr); 1115cbda25cSStefano Zampini ierr = VecRestoreArrayRead(usedv,(const PetscScalar**)&array);CHKERRQ(ierr); 112282d6408SStefano Zampini ierr = MatMultAdd(ctx->benign_csAIB,ctx->benign_corr_work,ctx->benign_dummy_schur_vec,ctx->benign_dummy_schur_vec);CHKERRQ(ierr); 1135cbda25cSStefano Zampini ierr = VecResetArray(ctx->benign_dummy_schur_vec);CHKERRQ(ierr); 1145cbda25cSStefano Zampini } 115ca92afb2SStefano Zampini PetscFunctionReturn(0); 116ca92afb2SStefano Zampini } 117ca92afb2SStefano Zampini 118d62866d3SStefano Zampini #undef __FUNCT__ 119df4d28bfSStefano Zampini #define __FUNCT__ "PCBDDCReuseSolvers_Solve_Private" 120df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolvers_Solve_Private(PC pc, Vec rhs, Vec sol, PetscBool transpose, PetscBool full) 121d62866d3SStefano Zampini { 122df4d28bfSStefano Zampini PCBDDCReuseSolvers ctx; 123683d3df6SStefano Zampini PetscBool copy = PETSC_FALSE; 124d62866d3SStefano Zampini PetscErrorCode ierr; 125d62866d3SStefano Zampini 126d62866d3SStefano Zampini PetscFunctionBegin; 127d62866d3SStefano Zampini ierr = PCShellGetContext(pc,(void **)&ctx);CHKERRQ(ierr); 128683d3df6SStefano Zampini if (full) { 129d62866d3SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 130d62866d3SStefano Zampini ierr = MatMumpsSetIcntl(ctx->F,26,-1);CHKERRQ(ierr); 131d62866d3SStefano Zampini #endif 1325cbda25cSStefano Zampini #if defined(PETSC_HAVE_MKL_PARDISO) 1335cbda25cSStefano Zampini ierr = MatMkl_PardisoSetCntl(ctx->F,70,0);CHKERRQ(ierr); 1345cbda25cSStefano Zampini #endif 135683d3df6SStefano Zampini copy = ctx->has_vertices; 136d4933d67SStefano Zampini } else { /* interior solver */ 1376dba178dSStefano Zampini #if defined(PETSC_HAVE_MUMPS) 138683d3df6SStefano Zampini ierr = MatMumpsSetIcntl(ctx->F,26,0);CHKERRQ(ierr); 1396dba178dSStefano Zampini #endif 140d4933d67SStefano Zampini #if defined(PETSC_HAVE_MKL_PARDISO) 141d4933d67SStefano Zampini ierr = MatMkl_PardisoSetCntl(ctx->F,70,1);CHKERRQ(ierr); 142d4933d67SStefano Zampini #endif 143683d3df6SStefano Zampini copy = PETSC_TRUE; 144683d3df6SStefano Zampini } 145683d3df6SStefano Zampini /* copy rhs into factored matrix workspace */ 146683d3df6SStefano Zampini if (copy) { 147ca92afb2SStefano Zampini PetscInt n; 148df4d28bfSStefano Zampini PetscScalar *array,*array_solver; 149ca92afb2SStefano Zampini 150ca92afb2SStefano Zampini ierr = VecGetLocalSize(rhs,&n);CHKERRQ(ierr); 151683d3df6SStefano Zampini ierr = VecGetArrayRead(rhs,(const PetscScalar**)&array);CHKERRQ(ierr); 152df4d28bfSStefano Zampini ierr = VecGetArray(ctx->rhs,&array_solver);CHKERRQ(ierr); 153df4d28bfSStefano Zampini ierr = PetscMemcpy(array_solver,array,n*sizeof(PetscScalar));CHKERRQ(ierr); 154df4d28bfSStefano Zampini ierr = VecRestoreArray(ctx->rhs,&array_solver);CHKERRQ(ierr); 155683d3df6SStefano Zampini ierr = VecRestoreArrayRead(rhs,(const PetscScalar**)&array);CHKERRQ(ierr); 156683d3df6SStefano Zampini 1575cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(ctx,ctx->rhs,NULL,PETSC_FALSE,full);CHKERRQ(ierr); 158683d3df6SStefano Zampini if (transpose) { 159683d3df6SStefano Zampini ierr = MatSolveTranspose(ctx->F,ctx->rhs,ctx->sol);CHKERRQ(ierr); 160683d3df6SStefano Zampini } else { 161683d3df6SStefano Zampini ierr = MatSolve(ctx->F,ctx->rhs,ctx->sol);CHKERRQ(ierr); 162683d3df6SStefano Zampini } 1635cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(ctx,ctx->sol,NULL,PETSC_TRUE,full);CHKERRQ(ierr); 164683d3df6SStefano Zampini 165683d3df6SStefano Zampini /* get back data to caller worskpace */ 166df4d28bfSStefano Zampini ierr = VecGetArrayRead(ctx->sol,(const PetscScalar**)&array_solver);CHKERRQ(ierr); 167683d3df6SStefano Zampini ierr = VecGetArray(sol,&array);CHKERRQ(ierr); 168df4d28bfSStefano Zampini ierr = PetscMemcpy(array,array_solver,n*sizeof(PetscScalar));CHKERRQ(ierr); 169683d3df6SStefano Zampini ierr = VecRestoreArray(sol,&array);CHKERRQ(ierr); 170df4d28bfSStefano Zampini ierr = VecRestoreArrayRead(ctx->sol,(const PetscScalar**)&array_solver);CHKERRQ(ierr); 171683d3df6SStefano Zampini } else { 172ca92afb2SStefano Zampini if (ctx->benign_n) { 1735cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(ctx,rhs,ctx->rhs,PETSC_FALSE,full);CHKERRQ(ierr); 174ca92afb2SStefano Zampini if (transpose) { 175ca92afb2SStefano Zampini ierr = MatSolveTranspose(ctx->F,ctx->rhs,sol);CHKERRQ(ierr); 176ca92afb2SStefano Zampini } else { 177ca92afb2SStefano Zampini ierr = MatSolve(ctx->F,ctx->rhs,sol);CHKERRQ(ierr); 178ca92afb2SStefano Zampini } 1795cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(ctx,sol,NULL,PETSC_TRUE,full);CHKERRQ(ierr); 180ca92afb2SStefano Zampini } else { 181e28d306cSStefano Zampini if (transpose) { 182e28d306cSStefano Zampini ierr = MatSolveTranspose(ctx->F,rhs,sol);CHKERRQ(ierr); 183e28d306cSStefano Zampini } else { 1846816873aSStefano Zampini ierr = MatSolve(ctx->F,rhs,sol);CHKERRQ(ierr); 185e28d306cSStefano Zampini } 186683d3df6SStefano Zampini } 187ca92afb2SStefano Zampini } 1885cbda25cSStefano Zampini /* restore defaults */ 1895cbda25cSStefano Zampini #if defined(PETSC_HAVE_MUMPS) 1905cbda25cSStefano Zampini ierr = MatMumpsSetIcntl(ctx->F,26,-1);CHKERRQ(ierr); 1915cbda25cSStefano Zampini #endif 192d4933d67SStefano Zampini #if defined(PETSC_HAVE_MKL_PARDISO) 193d4933d67SStefano Zampini ierr = MatMkl_PardisoSetCntl(ctx->F,70,0);CHKERRQ(ierr); 194d4933d67SStefano Zampini #endif 195d62866d3SStefano Zampini PetscFunctionReturn(0); 196d62866d3SStefano Zampini } 197d62866d3SStefano Zampini 198d62866d3SStefano Zampini #undef __FUNCT__ 199df4d28bfSStefano Zampini #define __FUNCT__ "PCBDDCReuseSolvers_Correction" 200df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolvers_Correction(PC pc, Vec rhs, Vec sol) 201e28d306cSStefano Zampini { 202e28d306cSStefano Zampini PetscErrorCode ierr; 203e28d306cSStefano Zampini 204e28d306cSStefano Zampini PetscFunctionBegin; 205df4d28bfSStefano Zampini ierr = PCBDDCReuseSolvers_Solve_Private(pc,rhs,sol,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 206e28d306cSStefano Zampini PetscFunctionReturn(0); 207e28d306cSStefano Zampini } 208e28d306cSStefano Zampini 209e28d306cSStefano Zampini #undef __FUNCT__ 210df4d28bfSStefano Zampini #define __FUNCT__ "PCBDDCReuseSolvers_CorrectionTranspose" 211df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolvers_CorrectionTranspose(PC pc, Vec rhs, Vec sol) 212e28d306cSStefano Zampini { 213e28d306cSStefano Zampini PetscErrorCode ierr; 214e28d306cSStefano Zampini 215e28d306cSStefano Zampini PetscFunctionBegin; 216df4d28bfSStefano Zampini ierr = PCBDDCReuseSolvers_Solve_Private(pc,rhs,sol,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 217683d3df6SStefano Zampini PetscFunctionReturn(0); 218683d3df6SStefano Zampini } 219683d3df6SStefano Zampini 220683d3df6SStefano Zampini #undef __FUNCT__ 221df4d28bfSStefano Zampini #define __FUNCT__ "PCBDDCReuseSolvers_Interior" 222df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolvers_Interior(PC pc, Vec rhs, Vec sol) 223683d3df6SStefano Zampini { 224683d3df6SStefano Zampini PetscErrorCode ierr; 225683d3df6SStefano Zampini 226683d3df6SStefano Zampini PetscFunctionBegin; 227df4d28bfSStefano Zampini ierr = PCBDDCReuseSolvers_Solve_Private(pc,rhs,sol,PETSC_FALSE,PETSC_FALSE);CHKERRQ(ierr); 228683d3df6SStefano Zampini PetscFunctionReturn(0); 229683d3df6SStefano Zampini } 230683d3df6SStefano Zampini 231683d3df6SStefano Zampini #undef __FUNCT__ 232df4d28bfSStefano Zampini #define __FUNCT__ "PCBDDCReuseSolvers_InteriorTranspose" 233df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolvers_InteriorTranspose(PC pc, Vec rhs, Vec sol) 234683d3df6SStefano Zampini { 235683d3df6SStefano Zampini PetscErrorCode ierr; 236683d3df6SStefano Zampini 237683d3df6SStefano Zampini PetscFunctionBegin; 238df4d28bfSStefano Zampini ierr = PCBDDCReuseSolvers_Solve_Private(pc,rhs,sol,PETSC_TRUE,PETSC_FALSE);CHKERRQ(ierr); 239e28d306cSStefano Zampini PetscFunctionReturn(0); 240e28d306cSStefano Zampini } 241e28d306cSStefano Zampini 242e28d306cSStefano Zampini #undef __FUNCT__ 243df4d28bfSStefano Zampini #define __FUNCT__ "PCBDDCReuseSolversReset" 244df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolversReset(PCBDDCReuseSolvers reuse) 245d62866d3SStefano Zampini { 246ca92afb2SStefano Zampini PetscInt i; 247d62866d3SStefano Zampini PetscErrorCode ierr; 248d62866d3SStefano Zampini 249d62866d3SStefano Zampini PetscFunctionBegin; 250d62866d3SStefano Zampini ierr = MatDestroy(&reuse->F);CHKERRQ(ierr); 251d62866d3SStefano Zampini ierr = VecDestroy(&reuse->sol);CHKERRQ(ierr); 252d62866d3SStefano Zampini ierr = VecDestroy(&reuse->rhs);CHKERRQ(ierr); 253d62866d3SStefano Zampini ierr = PCDestroy(&reuse->interior_solver);CHKERRQ(ierr); 254d62866d3SStefano Zampini ierr = PCDestroy(&reuse->correction_solver);CHKERRQ(ierr); 25553892102SStefano Zampini ierr = ISDestroy(&reuse->is_R);CHKERRQ(ierr); 25653892102SStefano Zampini ierr = ISDestroy(&reuse->is_B);CHKERRQ(ierr); 257d62866d3SStefano Zampini ierr = VecScatterDestroy(&reuse->correction_scatter_B);CHKERRQ(ierr); 25853892102SStefano Zampini ierr = VecDestroy(&reuse->sol_B);CHKERRQ(ierr); 25953892102SStefano Zampini ierr = VecDestroy(&reuse->rhs_B);CHKERRQ(ierr); 260ca92afb2SStefano Zampini for (i=0;i<reuse->benign_n;i++) { 261ca92afb2SStefano Zampini ierr = ISDestroy(&reuse->benign_zerodiag_subs[i]);CHKERRQ(ierr); 262ca92afb2SStefano Zampini } 263ca92afb2SStefano Zampini ierr = PetscFree(reuse->benign_zerodiag_subs);CHKERRQ(ierr); 264ca92afb2SStefano Zampini ierr = PetscFree(reuse->benign_save_vals);CHKERRQ(ierr); 2655cbda25cSStefano Zampini ierr = MatDestroy(&reuse->benign_csAIB);CHKERRQ(ierr); 2665cbda25cSStefano Zampini ierr = MatDestroy(&reuse->benign_AIIm1ones);CHKERRQ(ierr); 2675cbda25cSStefano Zampini ierr = VecDestroy(&reuse->benign_corr_work);CHKERRQ(ierr); 2685cbda25cSStefano Zampini ierr = VecDestroy(&reuse->benign_dummy_schur_vec);CHKERRQ(ierr); 269d62866d3SStefano Zampini PetscFunctionReturn(0); 270d62866d3SStefano Zampini } 271d5574798SStefano Zampini 272d5574798SStefano Zampini #undef __FUNCT__ 2733202ece2SStefano Zampini #define __FUNCT__ "PCBDDCComputeExplicitSchur" 2745ec10c6aSStefano Zampini static PetscErrorCode PCBDDCComputeExplicitSchur(Mat M, PetscBool issym, MatReuse reuse, Mat *S) 2753202ece2SStefano Zampini { 2763202ece2SStefano Zampini Mat B, C, D, Bd, Cd, AinvBd; 2773202ece2SStefano Zampini KSP ksp; 2783202ece2SStefano Zampini PC pc; 2793202ece2SStefano Zampini PetscBool isLU, isILU, isCHOL, Bdense, Cdense; 2803202ece2SStefano Zampini PetscReal fill = 2.0; 281f11841e3SStefano Zampini PetscInt n_I; 2823202ece2SStefano Zampini PetscMPIInt size; 2833202ece2SStefano Zampini PetscErrorCode ierr; 2843202ece2SStefano Zampini 2853202ece2SStefano Zampini PetscFunctionBegin; 2863202ece2SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)M),&size);CHKERRQ(ierr); 2876c4ed002SBarry Smith if (size != 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not for parallel matrices"); 288f11841e3SStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 289f11841e3SStefano Zampini PetscBool Sdense; 290f11841e3SStefano Zampini 291f11841e3SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*S, MATSEQDENSE, &Sdense);CHKERRQ(ierr); 2926c4ed002SBarry Smith if (!Sdense) SETERRQ(PetscObjectComm((PetscObject)M),PETSC_ERR_SUP,"S should dense"); 293f11841e3SStefano Zampini } 2943202ece2SStefano Zampini ierr = MatSchurComplementGetSubMatrices(M, NULL, NULL, &B, &C, &D);CHKERRQ(ierr); 2953202ece2SStefano Zampini ierr = MatSchurComplementGetKSP(M, &ksp);CHKERRQ(ierr); 2963202ece2SStefano Zampini ierr = KSPGetPC(ksp, &pc);CHKERRQ(ierr); 2973202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject) pc, PCLU, &isLU);CHKERRQ(ierr); 2983202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject) pc, PCILU, &isILU);CHKERRQ(ierr); 2993202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject) pc, PCCHOLESKY, &isCHOL);CHKERRQ(ierr); 3003202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject) B, MATSEQDENSE, &Bdense);CHKERRQ(ierr); 3013202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject) C, MATSEQDENSE, &Cdense);CHKERRQ(ierr); 302f11841e3SStefano Zampini ierr = MatGetSize(B,&n_I,NULL);CHKERRQ(ierr); 303f11841e3SStefano Zampini if (n_I) { 3043202ece2SStefano Zampini if (!Bdense) { 3053202ece2SStefano Zampini ierr = MatConvert(B, MATSEQDENSE, MAT_INITIAL_MATRIX, &Bd);CHKERRQ(ierr); 3063202ece2SStefano Zampini } else { 3073202ece2SStefano Zampini Bd = B; 3083202ece2SStefano Zampini } 3093202ece2SStefano Zampini 3103202ece2SStefano Zampini if (isLU || isILU || isCHOL) { 3113202ece2SStefano Zampini Mat fact; 3123202ece2SStefano Zampini ierr = KSPSetUp(ksp);CHKERRQ(ierr); 3133202ece2SStefano Zampini ierr = PCFactorGetMatrix(pc, &fact);CHKERRQ(ierr); 3143202ece2SStefano Zampini ierr = MatDuplicate(Bd, MAT_DO_NOT_COPY_VALUES, &AinvBd);CHKERRQ(ierr); 3153202ece2SStefano Zampini ierr = MatMatSolve(fact, Bd, AinvBd);CHKERRQ(ierr); 3163202ece2SStefano Zampini } else { 31707b1e237SStefano Zampini PetscBool ex = PETSC_TRUE; 31807b1e237SStefano Zampini 31907b1e237SStefano Zampini if (ex) { 3203202ece2SStefano Zampini Mat Ainvd; 3213202ece2SStefano Zampini 3223202ece2SStefano Zampini ierr = PCComputeExplicitOperator(pc, &Ainvd);CHKERRQ(ierr); 3233202ece2SStefano Zampini ierr = MatMatMult(Ainvd, Bd, MAT_INITIAL_MATRIX, fill, &AinvBd);CHKERRQ(ierr); 3243202ece2SStefano Zampini ierr = MatDestroy(&Ainvd);CHKERRQ(ierr); 32507b1e237SStefano Zampini } else { 32607b1e237SStefano Zampini Vec sol,rhs; 32707b1e237SStefano Zampini PetscScalar *arrayrhs,*arraysol; 32807b1e237SStefano Zampini PetscInt i,nrhs,n; 32907b1e237SStefano Zampini 33007b1e237SStefano Zampini ierr = MatDuplicate(Bd, MAT_DO_NOT_COPY_VALUES, &AinvBd);CHKERRQ(ierr); 33107b1e237SStefano Zampini ierr = MatGetSize(Bd,&n,&nrhs);CHKERRQ(ierr); 33207b1e237SStefano Zampini ierr = MatDenseGetArray(Bd,&arrayrhs);CHKERRQ(ierr); 33307b1e237SStefano Zampini ierr = MatDenseGetArray(AinvBd,&arraysol);CHKERRQ(ierr); 33407b1e237SStefano Zampini ierr = KSPGetSolution(ksp,&sol);CHKERRQ(ierr); 33507b1e237SStefano Zampini ierr = KSPGetRhs(ksp,&rhs);CHKERRQ(ierr); 33607b1e237SStefano Zampini for (i=0;i<nrhs;i++) { 33707b1e237SStefano Zampini ierr = VecPlaceArray(rhs,arrayrhs+i*n);CHKERRQ(ierr); 33807b1e237SStefano Zampini ierr = VecPlaceArray(sol,arraysol+i*n);CHKERRQ(ierr); 33907b1e237SStefano Zampini ierr = KSPSolve(ksp,rhs,sol);CHKERRQ(ierr); 34007b1e237SStefano Zampini ierr = VecResetArray(rhs);CHKERRQ(ierr); 34107b1e237SStefano Zampini ierr = VecResetArray(sol);CHKERRQ(ierr); 34207b1e237SStefano Zampini } 34307b1e237SStefano Zampini ierr = MatDenseRestoreArray(Bd,&arrayrhs);CHKERRQ(ierr); 34407b1e237SStefano Zampini ierr = MatDenseRestoreArray(AinvBd,&arrayrhs);CHKERRQ(ierr); 34507b1e237SStefano Zampini } 3463202ece2SStefano Zampini } 3475ec10c6aSStefano Zampini if (!Bdense & !issym) { 3483202ece2SStefano Zampini ierr = MatDestroy(&Bd);CHKERRQ(ierr); 3493202ece2SStefano Zampini } 3505ec10c6aSStefano Zampini 3515ec10c6aSStefano Zampini if (!issym) { 3523202ece2SStefano Zampini if (!Cdense) { 3533202ece2SStefano Zampini ierr = MatConvert(C, MATSEQDENSE, MAT_INITIAL_MATRIX, &Cd);CHKERRQ(ierr); 3543202ece2SStefano Zampini } else { 3553202ece2SStefano Zampini Cd = C; 3563202ece2SStefano Zampini } 3575ec10c6aSStefano Zampini ierr = MatMatMult(Cd, AinvBd, reuse, fill, S);CHKERRQ(ierr); 3583202ece2SStefano Zampini if (!Cdense) { 3593202ece2SStefano Zampini ierr = MatDestroy(&Cd);CHKERRQ(ierr); 3603202ece2SStefano Zampini } 3615ec10c6aSStefano Zampini } else { 3625ec10c6aSStefano Zampini ierr = MatTransposeMatMult(Bd, AinvBd, reuse, fill, S);CHKERRQ(ierr); 3635ec10c6aSStefano Zampini if (!Bdense) { 3645ec10c6aSStefano Zampini ierr = MatDestroy(&Bd);CHKERRQ(ierr); 3655ec10c6aSStefano Zampini } 3665ec10c6aSStefano Zampini } 3675ec10c6aSStefano Zampini ierr = MatDestroy(&AinvBd);CHKERRQ(ierr); 368f11841e3SStefano Zampini } 3693202ece2SStefano Zampini 3703202ece2SStefano Zampini if (D) { 3713202ece2SStefano Zampini Mat Dd; 3723202ece2SStefano Zampini PetscBool Ddense; 3733202ece2SStefano Zampini 3743202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)D,MATSEQDENSE,&Ddense);CHKERRQ(ierr); 3753202ece2SStefano Zampini if (!Ddense) { 3763202ece2SStefano Zampini ierr = MatConvert(D, MATSEQDENSE, MAT_INITIAL_MATRIX, &Dd);CHKERRQ(ierr); 3773202ece2SStefano Zampini } else { 3783202ece2SStefano Zampini Dd = D; 3793202ece2SStefano Zampini } 380f11841e3SStefano Zampini if (n_I) { 3813202ece2SStefano Zampini ierr = MatAYPX(*S,-1.0,Dd,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 382f11841e3SStefano Zampini } else { 383f11841e3SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 384f11841e3SStefano Zampini ierr = MatDuplicate(Dd,MAT_COPY_VALUES,S);CHKERRQ(ierr); 385f11841e3SStefano Zampini } else { 386f11841e3SStefano Zampini ierr = MatCopy(Dd,*S,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 387f11841e3SStefano Zampini } 388f11841e3SStefano Zampini } 3893202ece2SStefano Zampini if (!Ddense) { 3903202ece2SStefano Zampini ierr = MatDestroy(&Dd);CHKERRQ(ierr); 3913202ece2SStefano Zampini } 3923202ece2SStefano Zampini } else { 3933202ece2SStefano Zampini ierr = MatScale(*S,-1.0);CHKERRQ(ierr); 3943202ece2SStefano Zampini } 3953202ece2SStefano Zampini PetscFunctionReturn(0); 3963202ece2SStefano Zampini } 39734a97f8cSStefano Zampini 39834a97f8cSStefano Zampini #undef __FUNCT__ 3991580ed26SStefano Zampini #define __FUNCT__ "PCBDDCSubSchursSetUp" 40091af6908SStefano Zampini PetscErrorCode PCBDDCSubSchursSetUp(PCBDDCSubSchurs sub_schurs, Mat Ain, Mat Sin, PetscBool exact_schur, PetscInt xadj[], PetscInt adjncy[], PetscInt nlayers, Vec scaling, PetscBool compute_Stilda, PetscBool reuse_solvers, PetscBool benign_trick, PetscInt benign_n, PetscInt benign_p0_lidx[], IS benign_zerodiag_subs[], Mat change, IS change_primal) 401b1b3d7a2SStefano Zampini { 402be83ff47SStefano Zampini Mat F,A_II,A_IB,A_BI,A_BB,AE_II; 403be83ff47SStefano Zampini Mat S_all; 4045a94c880SStefano Zampini Mat global_schur_subsets,work_mat,*submats; 4055db18549SStefano Zampini ISLocalToGlobalMapping l2gmap_subsets; 406b7ab4a40SStefano Zampini IS is_I,is_I_layer; 407dc456d91SStefano Zampini IS all_subsets,all_subsets_mult,all_subsets_n; 408dc456d91SStefano Zampini PetscInt *nnz,*all_local_idx_N; 409dc456d91SStefano Zampini PetscInt *auxnum1,*auxnum2; 4105a95e1ceSStefano Zampini PetscInt i,subset_size,max_subset_size; 411683d3df6SStefano Zampini PetscInt n_B,extra,local_size,global_size; 41208122e43SStefano Zampini PetscBLASInt B_N,B_ierr,B_lwork,*pivots; 41306a4b1faSStefano Zampini PetscScalar *Bwork; 4145a95e1ceSStefano Zampini PetscSubcomm subcomm; 4155a95e1ceSStefano Zampini PetscMPIInt color,rank; 4165a95e1ceSStefano Zampini MPI_Comm comm_n; 4179d54b7f4SStefano Zampini PetscBool deluxe = PETSC_TRUE, use_getr = PETSC_FALSE; 418b1b3d7a2SStefano Zampini PetscErrorCode ierr; 419b1b3d7a2SStefano Zampini 420b1b3d7a2SStefano Zampini PetscFunctionBegin; 421a64f4aa4SStefano Zampini /* update info in sub_schurs */ 422a64f4aa4SStefano Zampini ierr = MatDestroy(&sub_schurs->A);CHKERRQ(ierr); 423a64f4aa4SStefano Zampini ierr = MatDestroy(&sub_schurs->S);CHKERRQ(ierr); 4243301b35fSStefano Zampini sub_schurs->is_hermitian = PETSC_FALSE; 4253301b35fSStefano Zampini sub_schurs->is_posdef = PETSC_FALSE; 426a64f4aa4SStefano Zampini if (Ain) { 427a64f4aa4SStefano Zampini PetscBool isseqaij; 4283301b35fSStefano Zampini /* determine if we are dealing with hermitian positive definite problems */ 4293301b35fSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 430*8d92b15dSStefano Zampini if (Ain->symmetric_set) sub_schurs->is_hermitian = Ain->symmetric; 4313301b35fSStefano Zampini #else 432*8d92b15dSStefano Zampini if (Ain->hermitian_set) sub_schurs->is_hermitian = Ain->hermitian; 4333301b35fSStefano Zampini #endif 434*8d92b15dSStefano Zampini if (Ain->spd_set) sub_schurs->is_posdef = Ain->spd; 435a64f4aa4SStefano Zampini 4363301b35fSStefano Zampini /* check */ 437a64f4aa4SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)Ain,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 4381dac5d76SStefano Zampini if (compute_Stilda && (!Ain->hermitian_set || !Ain->spd_set)) { /* these are lazy checks, maybe I should throw an error if they are not set */ 4393301b35fSStefano Zampini PetscInt lsize; 4403301b35fSStefano Zampini 4413301b35fSStefano Zampini ierr = MatGetSize(Ain,&lsize,NULL);CHKERRQ(ierr); 4423301b35fSStefano Zampini if (lsize) { 4433301b35fSStefano Zampini PetscScalar val; 4443301b35fSStefano Zampini PetscReal norm; 4453301b35fSStefano Zampini Vec vec1,vec2,vec3; 4463301b35fSStefano Zampini 4473301b35fSStefano Zampini ierr = MatCreateVecs(Ain,&vec1,&vec2);CHKERRQ(ierr); 4483301b35fSStefano Zampini ierr = VecSetRandom(vec1,NULL);CHKERRQ(ierr); 4493301b35fSStefano Zampini ierr = MatMult(Ain,vec1,vec2);CHKERRQ(ierr); 450*8d92b15dSStefano Zampini if (!Ain->spd_set && !benign_trick) { 451*8d92b15dSStefano Zampini ierr = VecDot(vec1,vec2,&val);CHKERRQ(ierr); 452*8d92b15dSStefano Zampini if (PetscRealPart(val) > 0. && PetscAbsReal(PetscImaginaryPart(val)) < PETSC_SMALL) sub_schurs->is_posdef = PETSC_TRUE; 453*8d92b15dSStefano Zampini } 454*8d92b15dSStefano Zampini if (!Ain->hermitian_set) { 455*8d92b15dSStefano Zampini ierr = VecDuplicate(vec1,&vec3);CHKERRQ(ierr); 4563301b35fSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 4573301b35fSStefano Zampini ierr = MatMultTranspose(Ain,vec1,vec3);CHKERRQ(ierr); 4583301b35fSStefano Zampini #else 4593301b35fSStefano Zampini ierr = MatMultHermitianTranspose(Ain,vec1,vec3);CHKERRQ(ierr); 4603301b35fSStefano Zampini #endif 4613301b35fSStefano Zampini ierr = VecAXPY(vec3,-1.0,vec2);CHKERRQ(ierr); 4623301b35fSStefano Zampini ierr = VecNorm(vec3,NORM_INFINITY,&norm);CHKERRQ(ierr); 4631dac5d76SStefano Zampini if (!Ain->hermitian_set) { 464862806e4SStefano Zampini if (norm > PetscSqrtReal(PETSC_SMALL)) { 4653301b35fSStefano Zampini sub_schurs->is_hermitian = PETSC_FALSE; 4663301b35fSStefano Zampini } else { 4673301b35fSStefano Zampini sub_schurs->is_hermitian = PETSC_TRUE; 4683301b35fSStefano Zampini } 4691dac5d76SStefano Zampini } 470*8d92b15dSStefano Zampini ierr = VecDestroy(&vec3);CHKERRQ(ierr); 47106a4e24aSStefano Zampini } 4723301b35fSStefano Zampini ierr = VecDestroy(&vec1);CHKERRQ(ierr); 4733301b35fSStefano Zampini ierr = VecDestroy(&vec2);CHKERRQ(ierr); 4743301b35fSStefano Zampini } else { 4753301b35fSStefano Zampini sub_schurs->is_hermitian = PETSC_TRUE; 4763301b35fSStefano Zampini sub_schurs->is_posdef = PETSC_TRUE; 4773301b35fSStefano Zampini } 4783301b35fSStefano Zampini } 479a64f4aa4SStefano Zampini if (isseqaij) { 480a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)Ain);CHKERRQ(ierr); 481a64f4aa4SStefano Zampini sub_schurs->A = Ain; 4823301b35fSStefano Zampini } else { 483a64f4aa4SStefano Zampini ierr = MatConvert(Ain,MATSEQAIJ,MAT_INITIAL_MATRIX,&sub_schurs->A);CHKERRQ(ierr); 484a64f4aa4SStefano Zampini } 485a64f4aa4SStefano Zampini } 4863301b35fSStefano Zampini 487a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)Sin);CHKERRQ(ierr); 488a64f4aa4SStefano Zampini sub_schurs->S = Sin; 489df4d28bfSStefano Zampini if (sub_schurs->schur_explicit) { 490df4d28bfSStefano Zampini sub_schurs->schur_explicit = (PetscBool)(!!sub_schurs->A); 491a64f4aa4SStefano Zampini } 492a64f4aa4SStefano Zampini 4935a95e1ceSStefano Zampini /* preliminary checks */ 494af25d912SStefano Zampini if (!sub_schurs->schur_explicit && compute_Stilda) SETERRQ(PetscObjectComm((PetscObject)sub_schurs->l2gmap),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_PARDISO"); 4955a95e1ceSStefano Zampini 4965a95e1ceSStefano Zampini /* restrict work on active processes */ 4975a95e1ceSStefano Zampini color = 0; 4985a95e1ceSStefano Zampini if (!sub_schurs->n_subs) color = 1; /* this can happen if we are in a multilevel case or if the subdomain is disconnected */ 4995a95e1ceSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)sub_schurs->l2gmap),&rank);CHKERRQ(ierr); 5005a95e1ceSStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)sub_schurs->l2gmap),&subcomm);CHKERRQ(ierr); 5015a95e1ceSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 5025a95e1ceSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 5035a95e1ceSStefano Zampini ierr = PetscCommDuplicate(PetscSubcommChild(subcomm),&comm_n,NULL);CHKERRQ(ierr); 5045a95e1ceSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 5055a95e1ceSStefano Zampini if (!sub_schurs->n_subs) { 5065a95e1ceSStefano Zampini ierr = PetscCommDestroy(&comm_n);CHKERRQ(ierr); 5075a95e1ceSStefano Zampini PetscFunctionReturn(0); 5085a95e1ceSStefano Zampini } 50981ea8064SStefano Zampini /* ierr = PetscCommDuplicate(PetscObjectComm((PetscObject)sub_schurs->l2gmap),&comm_n,NULL);CHKERRQ(ierr); */ 5105a95e1ceSStefano Zampini 511b1b3d7a2SStefano Zampini /* get Schur complement matrices */ 512df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 513a64f4aa4SStefano Zampini Mat tA_IB,tA_BI,tA_BB; 5143301b35fSStefano Zampini PetscBool isseqsbaij; 515a64f4aa4SStefano Zampini ierr = MatSchurComplementGetSubMatrices(sub_schurs->S,&A_II,NULL,&tA_IB,&tA_BI,&tA_BB);CHKERRQ(ierr); 5163301b35fSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)tA_BB,MATSEQSBAIJ,&isseqsbaij);CHKERRQ(ierr); 5173301b35fSStefano Zampini if (isseqsbaij) { 518a64f4aa4SStefano Zampini ierr = MatConvert(tA_BB,MATSEQAIJ,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 519a64f4aa4SStefano Zampini ierr = MatConvert(tA_IB,MATSEQAIJ,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 520a64f4aa4SStefano Zampini ierr = MatConvert(tA_BI,MATSEQAIJ,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 521a64f4aa4SStefano Zampini } else { 522a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)tA_BB);CHKERRQ(ierr); 523a64f4aa4SStefano Zampini A_BB = tA_BB; 524a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)tA_IB);CHKERRQ(ierr); 525a64f4aa4SStefano Zampini A_IB = tA_IB; 526a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)tA_BI);CHKERRQ(ierr); 527a64f4aa4SStefano Zampini A_BI = tA_BI; 528f11841e3SStefano Zampini } 529a58a30b4SStefano Zampini } else { 5305a95e1ceSStefano Zampini A_II = NULL; 5315a95e1ceSStefano Zampini A_IB = NULL; 5325a95e1ceSStefano Zampini A_BI = NULL; 5335a95e1ceSStefano Zampini A_BB = NULL; 534b1b3d7a2SStefano Zampini } 5355a95e1ceSStefano Zampini S_all = NULL; 536b1b3d7a2SStefano Zampini 537b1b3d7a2SStefano Zampini /* determine interior problems */ 5383dc780c3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I,&i);CHKERRQ(ierr); 5393dc780c3SStefano Zampini if (nlayers >= 0 && i) { /* Interior problems can be different from the original one */ 540b1b3d7a2SStefano Zampini PetscBT touched; 541b1b3d7a2SStefano Zampini const PetscInt* idx_B; 542b1b3d7a2SStefano Zampini PetscInt n_I,n_B,n_local_dofs,n_prev_added,j,layer,*local_numbering; 543b1b3d7a2SStefano Zampini 5446c4ed002SBarry Smith if (!xadj || !adjncy) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot request layering without adjacency"); 545b1b3d7a2SStefano Zampini /* get sizes */ 546b1b3d7a2SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I,&n_I);CHKERRQ(ierr); 547b1b3d7a2SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_B,&n_B);CHKERRQ(ierr); 548b1b3d7a2SStefano Zampini 549b1b3d7a2SStefano Zampini ierr = PetscMalloc1(n_I+n_B,&local_numbering);CHKERRQ(ierr); 550b1b3d7a2SStefano Zampini ierr = PetscBTCreate(n_I+n_B,&touched);CHKERRQ(ierr); 551b1b3d7a2SStefano Zampini ierr = PetscBTMemzero(n_I+n_B,touched);CHKERRQ(ierr); 552b1b3d7a2SStefano Zampini 553b1b3d7a2SStefano Zampini /* all boundary dofs must be skipped when adding layers */ 554b1b3d7a2SStefano Zampini ierr = ISGetIndices(sub_schurs->is_B,&idx_B);CHKERRQ(ierr); 555b1b3d7a2SStefano Zampini for (j=0;j<n_B;j++) { 556b1b3d7a2SStefano Zampini ierr = PetscBTSet(touched,idx_B[j]);CHKERRQ(ierr); 557b1b3d7a2SStefano Zampini } 558b1b3d7a2SStefano Zampini ierr = PetscMemcpy(local_numbering,idx_B,n_B*sizeof(PetscInt));CHKERRQ(ierr); 559b1b3d7a2SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_B,&idx_B);CHKERRQ(ierr); 560b1b3d7a2SStefano Zampini 561b1b3d7a2SStefano Zampini /* add prescribed number of layers of dofs */ 562b1b3d7a2SStefano Zampini n_local_dofs = n_B; 563b1b3d7a2SStefano Zampini n_prev_added = n_B; 564b1b3d7a2SStefano Zampini for (layer=0;layer<nlayers;layer++) { 565b1b3d7a2SStefano Zampini PetscInt n_added; 566b1b3d7a2SStefano Zampini if (n_local_dofs == n_I+n_B) break; 5676c4ed002SBarry Smith if (n_local_dofs > n_I+n_B) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error querying layer %D. Out of bound access (%D > %D)",layer,n_local_dofs,n_I+n_B); 568b1b3d7a2SStefano Zampini ierr = PCBDDCAdjGetNextLayer_Private(local_numbering+n_local_dofs,n_prev_added,touched,xadj,adjncy,&n_added);CHKERRQ(ierr); 569b1b3d7a2SStefano Zampini n_prev_added = n_added; 570b1b3d7a2SStefano Zampini n_local_dofs += n_added; 571b1b3d7a2SStefano Zampini if (!n_added) break; 572b1b3d7a2SStefano Zampini } 573b1b3d7a2SStefano Zampini ierr = PetscBTDestroy(&touched);CHKERRQ(ierr); 574b1b3d7a2SStefano Zampini 575883469d8SStefano Zampini /* IS for I layer dofs in original numbering */ 576a9b99552SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)sub_schurs->is_I),n_local_dofs-n_B,local_numbering+n_B,PETSC_COPY_VALUES,&is_I_layer);CHKERRQ(ierr); 577b1b3d7a2SStefano Zampini ierr = PetscFree(local_numbering);CHKERRQ(ierr); 578a9b99552SStefano Zampini ierr = ISSort(is_I_layer);CHKERRQ(ierr); 579883469d8SStefano Zampini /* IS for I layer dofs in I numbering */ 580df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 581b1b3d7a2SStefano Zampini ISLocalToGlobalMapping ItoNmap; 582b1b3d7a2SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(sub_schurs->is_I,&ItoNmap);CHKERRQ(ierr); 583a9b99552SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(ItoNmap,IS_GTOLM_DROP,is_I_layer,&is_I);CHKERRQ(ierr); 584b1b3d7a2SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&ItoNmap);CHKERRQ(ierr); 585b1b3d7a2SStefano Zampini 586b1b3d7a2SStefano Zampini /* II block */ 587b1b3d7a2SStefano Zampini ierr = MatGetSubMatrix(A_II,is_I,is_I,MAT_INITIAL_MATRIX,&AE_II);CHKERRQ(ierr); 588b1b3d7a2SStefano Zampini } 589b1b3d7a2SStefano Zampini } else { 590b1b3d7a2SStefano Zampini PetscInt n_I; 591b1b3d7a2SStefano Zampini 592b1b3d7a2SStefano Zampini /* IS for I dofs in original numbering */ 593b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)sub_schurs->is_I);CHKERRQ(ierr); 594a9b99552SStefano Zampini is_I_layer = sub_schurs->is_I; 595b1b3d7a2SStefano Zampini 596b1b3d7a2SStefano Zampini /* IS for I dofs in I numbering (strided 1) */ 597df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 598b1b3d7a2SStefano Zampini ierr = ISGetSize(sub_schurs->is_I,&n_I);CHKERRQ(ierr); 599b1b3d7a2SStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)sub_schurs->is_I),n_I,0,1,&is_I);CHKERRQ(ierr); 600b1b3d7a2SStefano Zampini 601b1b3d7a2SStefano Zampini /* II block is the same */ 602b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)A_II);CHKERRQ(ierr); 603b1b3d7a2SStefano Zampini AE_II = A_II; 604b1b3d7a2SStefano Zampini } 605b1b3d7a2SStefano Zampini } 6065a95e1ceSStefano Zampini 607883469d8SStefano Zampini /* Get info on subset sizes and sum of all subsets sizes */ 6085a95e1ceSStefano Zampini max_subset_size = 0; 609883469d8SStefano Zampini local_size = 0; 6105a95e1ceSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 6115a95e1ceSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 6125a95e1ceSStefano Zampini max_subset_size = PetscMax(subset_size,max_subset_size); 613883469d8SStefano Zampini local_size += subset_size; 614883469d8SStefano Zampini } 615883469d8SStefano Zampini 616883469d8SStefano Zampini /* Work arrays for local indices */ 617883469d8SStefano Zampini extra = 0; 618683d3df6SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_B,&n_B);CHKERRQ(ierr); 619df4d28bfSStefano Zampini if (sub_schurs->schur_explicit && is_I_layer) { 620a9b99552SStefano Zampini ierr = ISGetLocalSize(is_I_layer,&extra);CHKERRQ(ierr); 621883469d8SStefano Zampini } 622683d3df6SStefano Zampini ierr = PetscMalloc1(n_B+extra,&all_local_idx_N);CHKERRQ(ierr); 623883469d8SStefano Zampini if (extra) { 624883469d8SStefano Zampini const PetscInt *idxs; 625a9b99552SStefano Zampini ierr = ISGetIndices(is_I_layer,&idxs);CHKERRQ(ierr); 626883469d8SStefano Zampini ierr = PetscMemcpy(all_local_idx_N,idxs,extra*sizeof(PetscInt));CHKERRQ(ierr); 627a9b99552SStefano Zampini ierr = ISRestoreIndices(is_I_layer,&idxs);CHKERRQ(ierr); 628883469d8SStefano Zampini } 629883469d8SStefano Zampini ierr = PetscMalloc1(local_size,&nnz);CHKERRQ(ierr); 630dc456d91SStefano Zampini ierr = PetscMalloc1(sub_schurs->n_subs,&auxnum1);CHKERRQ(ierr); 631dc456d91SStefano Zampini ierr = PetscMalloc1(sub_schurs->n_subs,&auxnum2);CHKERRQ(ierr); 632883469d8SStefano Zampini 633883469d8SStefano Zampini /* Get local indices in local numbering */ 634883469d8SStefano Zampini local_size = 0; 6355a95e1ceSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 636883469d8SStefano Zampini PetscInt j; 637883469d8SStefano Zampini const PetscInt *idxs; 638883469d8SStefano Zampini 6395a95e1ceSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 6405a95e1ceSStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 641eb595f79SStefano Zampini /* start (smallest in global ordering) and multiplicity */ 642eb595f79SStefano Zampini auxnum1[i] = idxs[0]; 643eb595f79SStefano Zampini auxnum2[i] = subset_size; 644883469d8SStefano Zampini /* subset indices in local numbering */ 645883469d8SStefano Zampini ierr = PetscMemcpy(all_local_idx_N+local_size+extra,idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 6465a95e1ceSStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 647883469d8SStefano Zampini for (j=0;j<subset_size;j++) nnz[local_size+j] = subset_size; 648883469d8SStefano Zampini local_size += subset_size; 649883469d8SStefano Zampini } 650883469d8SStefano Zampini 6515a95e1ceSStefano Zampini /* allocate extra workspace needed only for GETRI */ 652d2627357SStefano Zampini Bwork = NULL; 653d2627357SStefano Zampini pivots = NULL; 65406a4e24aSStefano Zampini if (local_size && !benign_trick && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) { 655d2627357SStefano Zampini PetscScalar lwork; 656d2627357SStefano Zampini 65706a4e24aSStefano Zampini use_getr = PETSC_TRUE; 658d2627357SStefano Zampini B_lwork = -1; 659d2627357SStefano Zampini ierr = PetscBLASIntCast(local_size,&B_N);CHKERRQ(ierr); 660d2627357SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 661d2627357SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,Bwork,&B_N,pivots,&lwork,&B_lwork,&B_ierr)); 662d2627357SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 663d2627357SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GETRI Lapack routine %d",(int)B_ierr); 664d2627357SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 665d2627357SStefano Zampini ierr = PetscMalloc2(B_lwork,&Bwork,B_N,&pivots);CHKERRQ(ierr); 666d2627357SStefano Zampini } 667d2627357SStefano Zampini 668d2627357SStefano Zampini /* prepare parallel matrices for summing up properly schurs on subsets */ 669dc456d91SStefano Zampini ierr = ISCreateGeneral(comm_n,sub_schurs->n_subs,auxnum1,PETSC_OWN_POINTER,&all_subsets_n);CHKERRQ(ierr); 670dc456d91SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(sub_schurs->l2gmap,all_subsets_n,&all_subsets);CHKERRQ(ierr); 671dc456d91SStefano Zampini ierr = ISDestroy(&all_subsets_n);CHKERRQ(ierr); 672dc456d91SStefano Zampini ierr = ISCreateGeneral(comm_n,sub_schurs->n_subs,auxnum2,PETSC_OWN_POINTER,&all_subsets_mult);CHKERRQ(ierr); 673dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(all_subsets,all_subsets_mult,&global_size,&all_subsets_n);CHKERRQ(ierr); 674dc456d91SStefano Zampini ierr = ISDestroy(&all_subsets);CHKERRQ(ierr); 675dc456d91SStefano Zampini ierr = ISDestroy(&all_subsets_mult);CHKERRQ(ierr); 676dc456d91SStefano Zampini ierr = ISGetLocalSize(all_subsets_n,&i);CHKERRQ(ierr); 6776c4ed002SBarry Smith if (i != local_size) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid size of new subset! %D != %D",i,local_size); 678dc456d91SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(all_subsets_n,&l2gmap_subsets);CHKERRQ(ierr); 679e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,1,PETSC_DECIDE,PETSC_DECIDE,global_size,global_size,l2gmap_subsets,NULL,&work_mat);CHKERRQ(ierr); 680d2627357SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_subsets);CHKERRQ(ierr); 681d2627357SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)work_mat),&global_schur_subsets);CHKERRQ(ierr); 682d2627357SStefano Zampini ierr = MatSetSizes(global_schur_subsets,PETSC_DECIDE,PETSC_DECIDE,global_size,global_size);CHKERRQ(ierr); 683d2627357SStefano Zampini ierr = MatSetType(global_schur_subsets,MATMPIAIJ);CHKERRQ(ierr); 6842972d61bSStefano Zampini 6855a95e1ceSStefano Zampini /* subset indices in local boundary numbering */ 6865a95e1ceSStefano Zampini if (!sub_schurs->is_Ej_all) { 6875a95e1ceSStefano Zampini PetscInt *all_local_idx_B; 6885a95e1ceSStefano Zampini 6895a95e1ceSStefano Zampini ierr = PetscMalloc1(local_size,&all_local_idx_B);CHKERRQ(ierr); 6905a95e1ceSStefano Zampini ierr = ISGlobalToLocalMappingApply(sub_schurs->BtoNmap,IS_GTOLM_DROP,local_size,all_local_idx_N+extra,&subset_size,all_local_idx_B);CHKERRQ(ierr); 6916c4ed002SBarry Smith if (subset_size != local_size) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in sub_schurs serial (BtoNmap)! %D != %D\n",subset_size,local_size); 6925a95e1ceSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,local_size,all_local_idx_B,PETSC_OWN_POINTER,&sub_schurs->is_Ej_all);CHKERRQ(ierr); 693b1b3d7a2SStefano Zampini } 694b1b3d7a2SStefano Zampini 69572b8c272SStefano Zampini if (change) { 69672b8c272SStefano Zampini ISLocalToGlobalMapping BtoS; 69772b8c272SStefano Zampini IS change_primal_B; 69872b8c272SStefano Zampini IS change_primal_all; 69972b8c272SStefano Zampini 700b7ab4a40SStefano Zampini if (sub_schurs->change_primal_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 701b7ab4a40SStefano Zampini if (sub_schurs->change) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 702b7ab4a40SStefano Zampini ierr = PetscMalloc1(sub_schurs->n_subs,&sub_schurs->change_primal_sub);CHKERRQ(ierr); 70372b8c272SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 70472b8c272SStefano Zampini ISLocalToGlobalMapping NtoS; 70572b8c272SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(sub_schurs->is_subs[i],&NtoS);CHKERRQ(ierr); 706b7ab4a40SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoS,IS_GTOLM_DROP,change_primal,&sub_schurs->change_primal_sub[i]);CHKERRQ(ierr); 70772b8c272SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoS);CHKERRQ(ierr); 70872b8c272SStefano Zampini } 70972b8c272SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(sub_schurs->BtoNmap,IS_GTOLM_DROP,change_primal,&change_primal_B);CHKERRQ(ierr); 71072b8c272SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(sub_schurs->is_Ej_all,&BtoS);CHKERRQ(ierr); 71172b8c272SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(BtoS,IS_GTOLM_DROP,change_primal_B,&change_primal_all);CHKERRQ(ierr); 71272b8c272SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&BtoS);CHKERRQ(ierr); 71372b8c272SStefano Zampini ierr = ISDestroy(&change_primal_B);CHKERRQ(ierr); 714b7ab4a40SStefano Zampini ierr = PetscMalloc1(sub_schurs->n_subs,&sub_schurs->change);CHKERRQ(ierr); 71572b8c272SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 71672b8c272SStefano Zampini Mat change_sub; 71772b8c272SStefano Zampini 718b7ab4a40SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 71972b8c272SStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&sub_schurs->change[i]);CHKERRQ(ierr); 72072b8c272SStefano Zampini ierr = KSPSetType(sub_schurs->change[i],KSPPREONLY);CHKERRQ(ierr); 72172b8c272SStefano Zampini if (!sub_schurs->change_with_qr) { 72272b8c272SStefano Zampini ierr = MatGetSubMatrix(change,sub_schurs->is_subs[i],sub_schurs->is_subs[i],MAT_INITIAL_MATRIX,&change_sub);CHKERRQ(ierr); 72372b8c272SStefano Zampini } else { 72472b8c272SStefano Zampini Mat change_subt; 72572b8c272SStefano Zampini ierr = MatGetSubMatrix(change,sub_schurs->is_subs[i],sub_schurs->is_subs[i],MAT_INITIAL_MATRIX,&change_subt);CHKERRQ(ierr); 72672b8c272SStefano Zampini ierr = MatConvert(change_subt,MATSEQDENSE,MAT_INITIAL_MATRIX,&change_sub);CHKERRQ(ierr); 72772b8c272SStefano Zampini ierr = MatDestroy(&change_subt);CHKERRQ(ierr); 72872b8c272SStefano Zampini } 72972b8c272SStefano Zampini ierr = KSPSetOperators(sub_schurs->change[i],change_sub,change_sub);CHKERRQ(ierr); 73072b8c272SStefano Zampini ierr = MatDestroy(&change_sub);CHKERRQ(ierr); 73172b8c272SStefano Zampini ierr = KSPSetOptionsPrefix(sub_schurs->change[i],"sub_schurs_change_");CHKERRQ(ierr); 73272b8c272SStefano Zampini } 73372b8c272SStefano Zampini ierr = ISDestroy(&change_primal_all);CHKERRQ(ierr); 73472b8c272SStefano Zampini } 73572b8c272SStefano Zampini 7365a95e1ceSStefano Zampini /* Local matrix of all local Schur on subsets (transposed) */ 7375a95e1ceSStefano Zampini if (!sub_schurs->S_Ej_all) { 7385a95e1ceSStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&sub_schurs->S_Ej_all);CHKERRQ(ierr); 7395a95e1ceSStefano Zampini ierr = MatSetSizes(sub_schurs->S_Ej_all,PETSC_DECIDE,PETSC_DECIDE,local_size,local_size);CHKERRQ(ierr); 7405a95e1ceSStefano Zampini ierr = MatSetType(sub_schurs->S_Ej_all,MATAIJ);CHKERRQ(ierr); 7415a95e1ceSStefano Zampini ierr = MatSeqAIJSetPreallocation(sub_schurs->S_Ej_all,0,nnz);CHKERRQ(ierr); 742aa83b6aeSStefano Zampini } 743b1b3d7a2SStefano Zampini 7445a95e1ceSStefano Zampini /* Compute Schur complements explicitly */ 745be83ff47SStefano Zampini F = NULL; 746df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { /* this code branch is used when MatFactor with Schur complemnent support is not present; it is not very efficient, unless the economic version of the scaling is required */ 7475a95e1ceSStefano Zampini Mat S_Ej_expl; 7485a95e1ceSStefano Zampini PetscScalar *work; 7495a95e1ceSStefano Zampini PetscInt j,*dummy_idx; 7505a95e1ceSStefano Zampini PetscBool Sdense; 7515a95e1ceSStefano Zampini 7525a95e1ceSStefano Zampini ierr = PetscMalloc2(max_subset_size,&dummy_idx,max_subset_size*max_subset_size,&work);CHKERRQ(ierr); 7535a95e1ceSStefano Zampini local_size = 0; 754b1b3d7a2SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 7555a95e1ceSStefano Zampini IS is_subset_B; 7565a95e1ceSStefano Zampini Mat AE_EE,AE_IE,AE_EI,S_Ej; 7575a95e1ceSStefano Zampini 7585a95e1ceSStefano Zampini /* subsets in original and boundary numbering */ 7595a95e1ceSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(sub_schurs->BtoNmap,IS_GTOLM_DROP,sub_schurs->is_subs[i],&is_subset_B);CHKERRQ(ierr); 7605a95e1ceSStefano Zampini /* EE block */ 7615a95e1ceSStefano Zampini ierr = MatGetSubMatrix(A_BB,is_subset_B,is_subset_B,MAT_INITIAL_MATRIX,&AE_EE);CHKERRQ(ierr); 7625a95e1ceSStefano Zampini /* IE block */ 7635a95e1ceSStefano Zampini ierr = MatGetSubMatrix(A_IB,is_I,is_subset_B,MAT_INITIAL_MATRIX,&AE_IE);CHKERRQ(ierr); 7645a95e1ceSStefano Zampini /* EI block */ 7655a95e1ceSStefano Zampini if (sub_schurs->is_hermitian) { 7665a95e1ceSStefano Zampini ierr = MatCreateTranspose(AE_IE,&AE_EI);CHKERRQ(ierr); 7675a95e1ceSStefano Zampini } else { 7685a95e1ceSStefano Zampini ierr = MatGetSubMatrix(A_BI,is_subset_B,is_I,MAT_INITIAL_MATRIX,&AE_EI);CHKERRQ(ierr); 7695a95e1ceSStefano Zampini } 770a64f4aa4SStefano Zampini ierr = ISDestroy(&is_subset_B);CHKERRQ(ierr); 7715a95e1ceSStefano Zampini ierr = MatCreateSchurComplement(AE_II,AE_II,AE_IE,AE_EI,AE_EE,&S_Ej);CHKERRQ(ierr); 7725a95e1ceSStefano Zampini ierr = MatDestroy(&AE_EE);CHKERRQ(ierr); 7735a95e1ceSStefano Zampini ierr = MatDestroy(&AE_IE);CHKERRQ(ierr); 7745a95e1ceSStefano Zampini ierr = MatDestroy(&AE_EI);CHKERRQ(ierr); 775b1b3d7a2SStefano Zampini if (AE_II == A_II) { /* we can reuse the same ksp */ 776b1b3d7a2SStefano Zampini KSP ksp; 777b1b3d7a2SStefano Zampini ierr = MatSchurComplementGetKSP(sub_schurs->S,&ksp);CHKERRQ(ierr); 7785a95e1ceSStefano Zampini ierr = MatSchurComplementSetKSP(S_Ej,ksp);CHKERRQ(ierr); 779b1b3d7a2SStefano Zampini } else { /* build new ksp object which inherits ksp and pc types from the original one */ 780b1b3d7a2SStefano Zampini KSP origksp,schurksp; 781b1b3d7a2SStefano Zampini PC origpc,schurpc; 782b1b3d7a2SStefano Zampini KSPType ksp_type; 783b1b3d7a2SStefano Zampini PetscInt n_internal; 7845a95e1ceSStefano Zampini PetscBool ispcnone; 785b1b3d7a2SStefano Zampini 786b1b3d7a2SStefano Zampini ierr = MatSchurComplementGetKSP(sub_schurs->S,&origksp);CHKERRQ(ierr); 7875a95e1ceSStefano Zampini ierr = MatSchurComplementGetKSP(S_Ej,&schurksp);CHKERRQ(ierr); 788b1b3d7a2SStefano Zampini ierr = KSPGetType(origksp,&ksp_type);CHKERRQ(ierr); 789b1b3d7a2SStefano Zampini ierr = KSPSetType(schurksp,ksp_type);CHKERRQ(ierr); 790b1b3d7a2SStefano Zampini ierr = KSPGetPC(schurksp,&schurpc);CHKERRQ(ierr); 791b1b3d7a2SStefano Zampini ierr = KSPGetPC(origksp,&origpc);CHKERRQ(ierr); 7925a95e1ceSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)origpc,PCNONE,&ispcnone);CHKERRQ(ierr); 7935a95e1ceSStefano Zampini if (!ispcnone) { 7945a95e1ceSStefano Zampini PCType pc_type; 795b1b3d7a2SStefano Zampini ierr = PCGetType(origpc,&pc_type);CHKERRQ(ierr); 796b1b3d7a2SStefano Zampini ierr = PCSetType(schurpc,pc_type);CHKERRQ(ierr); 7975a95e1ceSStefano Zampini } else { 7985a95e1ceSStefano Zampini ierr = PCSetType(schurpc,PCLU);CHKERRQ(ierr); 7995a95e1ceSStefano Zampini } 800b1b3d7a2SStefano Zampini ierr = ISGetSize(is_I,&n_internal);CHKERRQ(ierr); 801b1b3d7a2SStefano Zampini if (n_internal) { /* UMFPACK gives error with 0 sized problems */ 802b1b3d7a2SStefano Zampini MatSolverPackage solver=NULL; 803b1b3d7a2SStefano Zampini ierr = PCFactorGetMatSolverPackage(origpc,(const MatSolverPackage*)&solver);CHKERRQ(ierr); 804b1b3d7a2SStefano Zampini if (solver) { 805b1b3d7a2SStefano Zampini ierr = PCFactorSetMatSolverPackage(schurpc,solver);CHKERRQ(ierr); 806b1b3d7a2SStefano Zampini } 807b1b3d7a2SStefano Zampini } 808b1b3d7a2SStefano Zampini ierr = KSPSetUp(schurksp);CHKERRQ(ierr); 809b1b3d7a2SStefano Zampini } 8105a95e1ceSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 8115a95e1ceSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,subset_size,work,&S_Ej_expl);CHKERRQ(ierr); 8125a95e1ceSStefano Zampini ierr = PCBDDCComputeExplicitSchur(S_Ej,sub_schurs->is_hermitian,MAT_REUSE_MATRIX,&S_Ej_expl);CHKERRQ(ierr); 8135a95e1ceSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)S_Ej_expl,MATSEQDENSE,&Sdense);CHKERRQ(ierr); 8145a95e1ceSStefano Zampini if (Sdense) { 8155a95e1ceSStefano Zampini for (j=0;j<subset_size;j++) { 8165a95e1ceSStefano Zampini dummy_idx[j]=local_size+j; 817b1b3d7a2SStefano Zampini } 8185a95e1ceSStefano Zampini ierr = MatSetValues(sub_schurs->S_Ej_all,subset_size,dummy_idx,subset_size,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr); 8196c4ed002SBarry Smith } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented for sparse matrices"); 8205a95e1ceSStefano Zampini ierr = MatDestroy(&S_Ej);CHKERRQ(ierr); 821a64f4aa4SStefano Zampini ierr = MatDestroy(&S_Ej_expl);CHKERRQ(ierr); 8225a95e1ceSStefano Zampini local_size += subset_size; 8235a95e1ceSStefano Zampini } 8245a95e1ceSStefano Zampini ierr = PetscFree2(dummy_idx,work);CHKERRQ(ierr); 825b1b3d7a2SStefano Zampini /* free */ 826b1b3d7a2SStefano Zampini ierr = ISDestroy(&is_I);CHKERRQ(ierr); 827b1b3d7a2SStefano Zampini ierr = MatDestroy(&AE_II);CHKERRQ(ierr); 8285a95e1ceSStefano Zampini ierr = PetscFree(all_local_idx_N);CHKERRQ(ierr); 829883469d8SStefano Zampini } else { 8305cbda25cSStefano Zampini Mat A,cs_AIB_mat = NULL,benign_AIIm1_ones_mat = NULL; 8319d54b7f4SStefano Zampini Vec Dall = NULL; 832ca92afb2SStefano Zampini IS is_A_all,*is_p_r = NULL; 8335cbda25cSStefano Zampini PetscScalar *work,*S_data,*schur_factor,infty = PETSC_MAX_REAL; 8346dba178dSStefano Zampini PetscInt n,n_I,*dummy_idx,size_schur,size_active_schur,cum,cum2; 835df4d28bfSStefano Zampini PetscBool economic,solver_S,S_lower_triangular = PETSC_FALSE,factor_workaround; 836d4933d67SStefano Zampini char solver[256]; 837883469d8SStefano Zampini 838683d3df6SStefano Zampini /* get sizes */ 83981ea8064SStefano Zampini n_I = 0; 84081ea8064SStefano Zampini if (is_I_layer) { 841a9b99552SStefano Zampini ierr = ISGetLocalSize(is_I_layer,&n_I);CHKERRQ(ierr); 84281ea8064SStefano Zampini } 843683d3df6SStefano Zampini economic = PETSC_FALSE; 844683d3df6SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I,&cum);CHKERRQ(ierr); 845683d3df6SStefano Zampini if (cum != n_I) economic = PETSC_TRUE; 846683d3df6SStefano Zampini ierr = MatGetLocalSize(sub_schurs->A,&n,NULL);CHKERRQ(ierr); 8479d54b7f4SStefano Zampini size_active_schur = local_size; 8489d54b7f4SStefano Zampini 849f17d2ae1SStefano Zampini /* import scaling vector (wrong formulation if we have 3D edges) */ 8509d54b7f4SStefano Zampini if (scaling && compute_Stilda) { 8519d54b7f4SStefano Zampini const PetscScalar *array; 8529d54b7f4SStefano Zampini PetscScalar *array2; 8539d54b7f4SStefano Zampini const PetscInt *idxs; 8549d54b7f4SStefano Zampini PetscInt i; 8559d54b7f4SStefano Zampini 8569d54b7f4SStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs);CHKERRQ(ierr); 8579d54b7f4SStefano Zampini ierr = VecCreateSeq(PETSC_COMM_SELF,size_active_schur,&Dall);CHKERRQ(ierr); 8589d54b7f4SStefano Zampini ierr = VecGetArrayRead(scaling,&array);CHKERRQ(ierr); 8599d54b7f4SStefano Zampini ierr = VecGetArray(Dall,&array2);CHKERRQ(ierr); 8609d54b7f4SStefano Zampini for (i=0;i<size_active_schur;i++) array2[i] = array[idxs[i]]; 8619d54b7f4SStefano Zampini ierr = VecRestoreArray(Dall,&array2);CHKERRQ(ierr); 8629d54b7f4SStefano Zampini ierr = VecRestoreArrayRead(scaling,&array);CHKERRQ(ierr); 8639d54b7f4SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs);CHKERRQ(ierr); 8649d54b7f4SStefano Zampini deluxe = PETSC_FALSE; 8659d54b7f4SStefano Zampini } 866d62866d3SStefano Zampini 867683d3df6SStefano Zampini /* size active schurs does not count any dirichlet or vertex dof on the interface */ 868683d3df6SStefano Zampini cum = n_I+size_active_schur; 869683d3df6SStefano Zampini if (sub_schurs->is_dir) { 870683d3df6SStefano Zampini const PetscInt* idxs; 871683d3df6SStefano Zampini PetscInt n_dir; 872683d3df6SStefano Zampini 873683d3df6SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_dir,&n_dir);CHKERRQ(ierr); 874683d3df6SStefano Zampini ierr = ISGetIndices(sub_schurs->is_dir,&idxs);CHKERRQ(ierr); 875683d3df6SStefano Zampini ierr = PetscMemcpy(all_local_idx_N+cum,idxs,n_dir*sizeof(PetscInt));CHKERRQ(ierr); 876683d3df6SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_dir,&idxs);CHKERRQ(ierr); 877683d3df6SStefano Zampini cum += n_dir; 878d62866d3SStefano Zampini } 879683d3df6SStefano Zampini factor_workaround = PETSC_FALSE; 880683d3df6SStefano Zampini /* include the primal vertices in the Schur complement */ 881367aa537SStefano Zampini if (exact_schur && sub_schurs->is_vertices && (compute_Stilda || benign_n)) { 882683d3df6SStefano Zampini PetscInt n_v; 883683d3df6SStefano Zampini 884683d3df6SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_v);CHKERRQ(ierr); 885683d3df6SStefano Zampini if (n_v) { 886683d3df6SStefano Zampini const PetscInt* idxs; 887683d3df6SStefano Zampini 888683d3df6SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 889683d3df6SStefano Zampini ierr = PetscMemcpy(all_local_idx_N+cum,idxs,n_v*sizeof(PetscInt));CHKERRQ(ierr); 890683d3df6SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 891683d3df6SStefano Zampini cum += n_v; 892683d3df6SStefano Zampini factor_workaround = PETSC_TRUE; 893683d3df6SStefano Zampini } 894683d3df6SStefano Zampini } 895683d3df6SStefano Zampini size_schur = cum - n_I; 896683d3df6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,cum,all_local_idx_N,PETSC_OWN_POINTER,&is_A_all);CHKERRQ(ierr); 897683d3df6SStefano Zampini /* get working mat (TODO: factorize without actually permuting) */ 898683d3df6SStefano Zampini if (cum == n) { 899683d3df6SStefano Zampini ierr = ISSetPermutation(is_A_all);CHKERRQ(ierr); 900683d3df6SStefano Zampini ierr = MatPermute(sub_schurs->A,is_A_all,is_A_all,&A);CHKERRQ(ierr); 901683d3df6SStefano Zampini } else { 9026816873aSStefano Zampini ierr = MatGetSubMatrix(sub_schurs->A,is_A_all,is_A_all,MAT_INITIAL_MATRIX,&A);CHKERRQ(ierr); 903683d3df6SStefano Zampini } 904a9b99552SStefano Zampini ierr = MatSetOptionsPrefix(A,"sub_schurs_");CHKERRQ(ierr); 905ca92afb2SStefano Zampini 906ca92afb2SStefano Zampini /* if we actually change the basis for the pressures, LDL^T factors will use a lot of memory 907367aa537SStefano Zampini n^2 instead of n^1.5 or something. This is a workaround */ 908ca92afb2SStefano Zampini if (benign_n) { 909ca92afb2SStefano Zampini Vec v; 910ca92afb2SStefano Zampini ISLocalToGlobalMapping N_to_reor; 911ca92afb2SStefano Zampini IS is_p0,is_p0_p; 9125cbda25cSStefano Zampini PetscScalar *cs_AIB,*AIIm1_data; 9135cbda25cSStefano Zampini PetscInt sizeA; 914ca92afb2SStefano Zampini 915ca92afb2SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_A_all,&N_to_reor);CHKERRQ(ierr); 916ca92afb2SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,benign_n,benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 917ca92afb2SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(N_to_reor,IS_GTOLM_DROP,is_p0,&is_p0_p);CHKERRQ(ierr); 918ca92afb2SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 919ca92afb2SStefano Zampini ierr = MatCreateVecs(A,&v,NULL);CHKERRQ(ierr); 9205cbda25cSStefano Zampini ierr = VecGetSize(v,&sizeA);CHKERRQ(ierr); 9215cbda25cSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,sizeA,benign_n,NULL,&benign_AIIm1_ones_mat);CHKERRQ(ierr); 922282d6408SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,size_schur,benign_n,NULL,&cs_AIB_mat);CHKERRQ(ierr); 9235cbda25cSStefano Zampini ierr = MatDenseGetArray(cs_AIB_mat,&cs_AIB);CHKERRQ(ierr); 9245cbda25cSStefano Zampini ierr = MatDenseGetArray(benign_AIIm1_ones_mat,&AIIm1_data);CHKERRQ(ierr); 925ca92afb2SStefano Zampini ierr = PetscMalloc1(benign_n,&is_p_r);CHKERRQ(ierr); 926ca92afb2SStefano Zampini /* compute colsum of A_IB restricted to pressures */ 927ca92afb2SStefano Zampini for (i=0;i<benign_n;i++) { 9285cbda25cSStefano Zampini Vec benign_AIIm1_ones; 929ca92afb2SStefano Zampini PetscScalar *array; 930ca92afb2SStefano Zampini const PetscInt *idxs; 931ca92afb2SStefano Zampini PetscInt j,nz; 932ca92afb2SStefano Zampini 933ca92afb2SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(N_to_reor,IS_GTOLM_DROP,benign_zerodiag_subs[i],&is_p_r[i]);CHKERRQ(ierr); 934ca92afb2SStefano Zampini ierr = ISGetLocalSize(is_p_r[i],&nz);CHKERRQ(ierr); 935ca92afb2SStefano Zampini ierr = ISGetIndices(is_p_r[i],&idxs);CHKERRQ(ierr); 9365cbda25cSStefano Zampini for (j=0;j<nz;j++) AIIm1_data[idxs[j]+sizeA*i] = 1.; 937ca92afb2SStefano Zampini ierr = ISRestoreIndices(is_p_r[i],&idxs);CHKERRQ(ierr); 9385cbda25cSStefano Zampini ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,sizeA,AIIm1_data+sizeA*i,&benign_AIIm1_ones);CHKERRQ(ierr); 9395cbda25cSStefano Zampini ierr = MatMult(A,benign_AIIm1_ones,v);CHKERRQ(ierr); 9405cbda25cSStefano Zampini ierr = VecDestroy(&benign_AIIm1_ones);CHKERRQ(ierr); 941ca92afb2SStefano Zampini ierr = VecGetArray(v,&array);CHKERRQ(ierr); 94222db5ddcSStefano Zampini for (j=0;j<size_schur;j++) { 94322db5ddcSStefano Zampini #if defined(PETSC_USE_COMPLEX) 94422db5ddcSStefano Zampini cs_AIB[i*size_schur + j] = (PetscRealPart(array[j+n_I])/nz + PETSC_i*(PetscImaginaryPart(array[j+n_I])/nz)); 94522db5ddcSStefano Zampini #else 94622db5ddcSStefano Zampini cs_AIB[i*size_schur + j] = array[j+n_I]/nz; 94722db5ddcSStefano Zampini #endif 94822db5ddcSStefano Zampini } 949ca92afb2SStefano Zampini ierr = VecRestoreArray(v,&array);CHKERRQ(ierr); 950ca92afb2SStefano Zampini } 9515cbda25cSStefano Zampini ierr = MatDenseRestoreArray(cs_AIB_mat,&cs_AIB);CHKERRQ(ierr); 9525cbda25cSStefano Zampini ierr = MatDenseRestoreArray(benign_AIIm1_ones_mat,&AIIm1_data);CHKERRQ(ierr); 953ca92afb2SStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 954ca92afb2SStefano Zampini ierr = MatSetOption(A,MAT_KEEP_NONZERO_PATTERN,PETSC_FALSE);CHKERRQ(ierr); 955ca92afb2SStefano Zampini ierr = MatZeroRowsColumnsIS(A,is_p0_p,1.0,NULL,NULL);CHKERRQ(ierr); 956ca92afb2SStefano Zampini ierr = ISDestroy(&is_p0_p);CHKERRQ(ierr); 957ca92afb2SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&N_to_reor);CHKERRQ(ierr); 958ca92afb2SStefano Zampini } 9593301b35fSStefano Zampini ierr = MatSetOption(A,MAT_SYMMETRIC,sub_schurs->is_hermitian);CHKERRQ(ierr); 9603301b35fSStefano Zampini ierr = MatSetOption(A,MAT_HERMITIAN,sub_schurs->is_hermitian);CHKERRQ(ierr); 9613301b35fSStefano Zampini ierr = MatSetOption(A,MAT_SPD,sub_schurs->is_posdef);CHKERRQ(ierr); 962df4d28bfSStefano Zampini #if defined(PETSC_HAVE_MUMPS) 963d4933d67SStefano Zampini ierr = PetscStrcpy(solver,MATSOLVERMUMPS);CHKERRQ(ierr); 964df4d28bfSStefano Zampini #else 965df4d28bfSStefano Zampini ierr = PetscStrcpy(solver,MATSOLVERMKL_PARDISO);CHKERRQ(ierr); 966df4d28bfSStefano Zampini #endif 9678760537fSStefano Zampini ierr = PetscOptionsGetString(NULL,"sub_schurs_","-mat_solver_package",solver,256,NULL);CHKERRQ(ierr); 968883469d8SStefano Zampini 969683d3df6SStefano Zampini /* when using the benign subspace trick, the local Schur complements are SPD */ 970d47842beSStefano Zampini if (benign_trick) sub_schurs->is_posdef = PETSC_TRUE; 971d47842beSStefano Zampini 972683d3df6SStefano Zampini if (n_I) { /* TODO add ordering from options */ 9735a05ddb0SStefano Zampini IS is_schur; 9745a05ddb0SStefano Zampini 9759ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 976d4933d67SStefano Zampini ierr = MatGetFactor(A,solver,MAT_FACTOR_CHOLESKY,&F);CHKERRQ(ierr); 977883469d8SStefano Zampini } else { 978d4933d67SStefano Zampini ierr = MatGetFactor(A,solver,MAT_FACTOR_LU,&F);CHKERRQ(ierr); 979883469d8SStefano Zampini } 980883469d8SStefano Zampini /* subsets ordered last */ 9815a05ddb0SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,size_schur,n_I,1,&is_schur);CHKERRQ(ierr); 9825a05ddb0SStefano Zampini ierr = MatFactorSetSchurIS(F,is_schur);CHKERRQ(ierr); 9835a05ddb0SStefano Zampini ierr = ISDestroy(&is_schur);CHKERRQ(ierr); 984883469d8SStefano Zampini 985883469d8SStefano Zampini /* factorization step */ 9869ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 987883469d8SStefano Zampini ierr = MatCholeskyFactorSymbolic(F,A,NULL,NULL);CHKERRQ(ierr); 988be83ff47SStefano Zampini #if defined(PETSC_HAVE_MUMPS) /* be sure that icntl 19 is not set by command line */ 989be83ff47SStefano Zampini ierr = MatMumpsSetIcntl(F,19,2);CHKERRQ(ierr); 990be83ff47SStefano Zampini #endif 991e8ade678SStefano Zampini if (sub_schurs->is_posdef) { 992e8ade678SStefano Zampini ierr = MatFactorSetSchurComplementSolverType(F,1);CHKERRQ(ierr); 993e8ade678SStefano Zampini } else { 994e8ade678SStefano Zampini ierr = MatFactorSetSchurComplementSolverType(F,2);CHKERRQ(ierr); 995e8ade678SStefano Zampini } 996883469d8SStefano Zampini ierr = MatCholeskyFactorNumeric(F,A,NULL);CHKERRQ(ierr); 997a0b0af32SStefano Zampini S_lower_triangular = PETSC_TRUE; 998883469d8SStefano Zampini } else { 999883469d8SStefano Zampini ierr = MatLUFactorSymbolic(F,A,NULL,NULL,NULL);CHKERRQ(ierr); 1000be83ff47SStefano Zampini #if defined(PETSC_HAVE_MUMPS) /* be sure that icntl 19 is not set by command line */ 1001be83ff47SStefano Zampini ierr = MatMumpsSetIcntl(F,19,3);CHKERRQ(ierr); 1002be83ff47SStefano Zampini #endif 1003883469d8SStefano Zampini ierr = MatLUFactorNumeric(F,A,NULL);CHKERRQ(ierr); 1004883469d8SStefano Zampini } 1005883469d8SStefano Zampini 1006883469d8SStefano Zampini /* get explicit Schur Complement computed during numeric factorization */ 10075a05ddb0SStefano Zampini ierr = MatFactorGetSchurComplement(F,&S_all);CHKERRQ(ierr); 1008d62866d3SStefano Zampini /* we can reuse the solvers if we are not using the economic version */ 1009683d3df6SStefano Zampini reuse_solvers = (PetscBool)(reuse_solvers && !economic); 1010683d3df6SStefano Zampini factor_workaround = (PetscBool)(reuse_solvers && factor_workaround); 1011df4d28bfSStefano Zampini solver_S = PETSC_TRUE; 1012ca92afb2SStefano Zampini 101372b8c272SStefano Zampini /* update the Schur complement with the change of basis on the pressures */ 1014ca92afb2SStefano Zampini if (benign_n) { 10155cbda25cSStefano Zampini PetscScalar *S_data,*cs_AIB,*AIIm1_data; 1016ca92afb2SStefano Zampini Vec v; 10175cbda25cSStefano Zampini PetscInt sizeA; 1018ca92afb2SStefano Zampini 1019ca92afb2SStefano Zampini ierr = MatDenseGetArray(S_all,&S_data);CHKERRQ(ierr); 10205cbda25cSStefano Zampini ierr = MatCreateVecs(A,&v,NULL);CHKERRQ(ierr); 10215cbda25cSStefano Zampini ierr = VecGetSize(v,&sizeA);CHKERRQ(ierr); 1022ca92afb2SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 1023ca92afb2SStefano Zampini ierr = MatMumpsSetIcntl(F,26,0);CHKERRQ(ierr); 1024ca92afb2SStefano Zampini #endif 1025ca92afb2SStefano Zampini #if defined(PETSC_HAVE_MKL_PARDISO) 1026ca92afb2SStefano Zampini ierr = MatMkl_PardisoSetCntl(F,70,1);CHKERRQ(ierr); 1027ca92afb2SStefano Zampini #endif 10285cbda25cSStefano Zampini ierr = MatDenseGetArray(cs_AIB_mat,&cs_AIB);CHKERRQ(ierr); 10295cbda25cSStefano Zampini ierr = MatDenseGetArray(benign_AIIm1_ones_mat,&AIIm1_data);CHKERRQ(ierr); 1030ca92afb2SStefano Zampini for (i=0;i<benign_n;i++) { 10315cbda25cSStefano Zampini Vec benign_AIIm1_ones; 103247484b83SStefano Zampini PetscScalar *array,sum = 0.,one = 1.; 1033ca92afb2SStefano Zampini const PetscInt *idxs; 1034ca92afb2SStefano Zampini PetscInt j,nz; 103547484b83SStefano Zampini PetscBLASInt B_k,B_n; 1036ca92afb2SStefano Zampini 10375cbda25cSStefano Zampini ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,sizeA,AIIm1_data+sizeA*i,&benign_AIIm1_ones);CHKERRQ(ierr); 10385cbda25cSStefano Zampini ierr = VecCopy(benign_AIIm1_ones,v);CHKERRQ(ierr); 10395cbda25cSStefano Zampini ierr = MatSolve(F,v,benign_AIIm1_ones);CHKERRQ(ierr); 1040ca92afb2SStefano Zampini ierr = ISGetLocalSize(is_p_r[i],&nz);CHKERRQ(ierr); 1041ca92afb2SStefano Zampini ierr = ISGetIndices(is_p_r[i],&idxs);CHKERRQ(ierr); 1042282d6408SStefano Zampini /* p0 dof (eliminated) is excluded from the sum */ 1043282d6408SStefano Zampini for (j=0;j<nz-1;j++) sum -= AIIm1_data[idxs[j]+sizeA*i]; 1044ca92afb2SStefano Zampini ierr = ISRestoreIndices(is_p_r[i],&idxs);CHKERRQ(ierr); 10455cbda25cSStefano Zampini ierr = MatMult(A,benign_AIIm1_ones,v);CHKERRQ(ierr); 1046ca92afb2SStefano Zampini ierr = VecGetArray(v,&array);CHKERRQ(ierr); 104747484b83SStefano Zampini /* perform sparse rank updates on symmetric Schur (TODO: move outside of the loop?) */ 104847484b83SStefano Zampini /* cs_AIB already scaled by 1./nz */ 104947484b83SStefano Zampini B_k = 1; 105047484b83SStefano Zampini ierr = PetscBLASIntCast(size_schur,&B_n);CHKERRQ(ierr); 105147484b83SStefano Zampini PetscStackCallBLAS("BLASsyrk",BLASsyrk_("L","N",&B_n,&B_k,&sum,cs_AIB+i*size_schur,&B_n,&one,S_data,&B_n)); 10525cbda25cSStefano Zampini sum = 1.; 105347484b83SStefano Zampini PetscStackCallBLAS("BLASsyr2k",BLASsyr2k_("L","N",&B_n,&B_k,&sum,array+n_I,&B_n,cs_AIB+i*size_schur,&B_n,&one,S_data,&B_n)); 1054ca92afb2SStefano Zampini ierr = VecRestoreArray(v,&array);CHKERRQ(ierr); 10555cbda25cSStefano Zampini /* set p0 entry of AIIm1_ones to zero */ 10565cbda25cSStefano Zampini ierr = ISGetIndices(is_p_r[i],&idxs);CHKERRQ(ierr); 1057282d6408SStefano Zampini for (j=0;j<benign_n;j++) AIIm1_data[idxs[nz-1]+sizeA*j] = 0.; 10585cbda25cSStefano Zampini ierr = ISRestoreIndices(is_p_r[i],&idxs);CHKERRQ(ierr); 10595cbda25cSStefano Zampini ierr = VecDestroy(&benign_AIIm1_ones);CHKERRQ(ierr); 1060ca92afb2SStefano Zampini } 10615cbda25cSStefano Zampini /* restore defaults */ 10625cbda25cSStefano Zampini #if defined(PETSC_HAVE_MUMPS) 10635cbda25cSStefano Zampini ierr = MatMumpsSetIcntl(F,26,-1);CHKERRQ(ierr); 10645cbda25cSStefano Zampini #endif 10655cbda25cSStefano Zampini #if defined(PETSC_HAVE_MKL_PARDISO) 10665cbda25cSStefano Zampini ierr = MatMkl_PardisoSetCntl(F,70,0);CHKERRQ(ierr); 10675cbda25cSStefano Zampini #endif 10685cbda25cSStefano Zampini ierr = MatDenseRestoreArray(cs_AIB_mat,&cs_AIB);CHKERRQ(ierr); 10695cbda25cSStefano Zampini ierr = MatDenseRestoreArray(benign_AIIm1_ones_mat,&AIIm1_data);CHKERRQ(ierr); 1070ca92afb2SStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 1071ca92afb2SStefano Zampini ierr = MatDenseRestoreArray(S_all,&S_data);CHKERRQ(ierr); 1072ca92afb2SStefano Zampini } 1073a3df083aSStefano Zampini if (!reuse_solvers) { 1074a3df083aSStefano Zampini for (i=0;i<benign_n;i++) { 1075a3df083aSStefano Zampini ierr = ISDestroy(&is_p_r[i]);CHKERRQ(ierr); 1076a3df083aSStefano Zampini } 1077a3df083aSStefano Zampini ierr = PetscFree(is_p_r);CHKERRQ(ierr); 10785cbda25cSStefano Zampini ierr = MatDestroy(&cs_AIB_mat);CHKERRQ(ierr); 10795cbda25cSStefano Zampini ierr = MatDestroy(&benign_AIIm1_ones_mat);CHKERRQ(ierr); 1080a3df083aSStefano Zampini } 1081df4d28bfSStefano Zampini } else { /* we can't use MatFactor when size_schur == size_of_the_problem */ 1082be83ff47SStefano Zampini ierr = MatConvert(A,MATSEQDENSE,MAT_INITIAL_MATRIX,&S_all);CHKERRQ(ierr); 1083683d3df6SStefano Zampini reuse_solvers = PETSC_FALSE; /* TODO: why we can't reuse the solvers here? */ 1084df4d28bfSStefano Zampini solver_S = PETSC_FALSE; 1085be83ff47SStefano Zampini } 1086be83ff47SStefano Zampini 1087be83ff47SStefano Zampini if (reuse_solvers) { 1088a00504b5SStefano Zampini Mat A_II,Afake; 108953892102SStefano Zampini Vec vec1_B; 1090df4d28bfSStefano Zampini PCBDDCReuseSolvers msolv_ctx; 10913462e049SStefano Zampini PetscInt n_R; 1092d5574798SStefano Zampini 1093df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 1094df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversReset(sub_schurs->reuse_solver);CHKERRQ(ierr); 1095e28d306cSStefano Zampini } else { 1096df4d28bfSStefano Zampini ierr = PetscNew(&sub_schurs->reuse_solver);CHKERRQ(ierr); 1097d62866d3SStefano Zampini } 1098df4d28bfSStefano Zampini msolv_ctx = sub_schurs->reuse_solver; 1099be83ff47SStefano Zampini ierr = MatSchurComplementGetSubMatrices(sub_schurs->S,&A_II,NULL,NULL,NULL,NULL);CHKERRQ(ierr); 1100d5574798SStefano Zampini ierr = PetscObjectReference((PetscObject)F);CHKERRQ(ierr); 1101d5574798SStefano Zampini msolv_ctx->F = F; 1102683d3df6SStefano Zampini ierr = MatCreateVecs(F,&msolv_ctx->sol,NULL);CHKERRQ(ierr); 1103683d3df6SStefano Zampini /* currently PETSc has no support for MatSolve(F,x,x), so cheat and let rhs and sol share the same memory */ 1104683d3df6SStefano Zampini { 1105683d3df6SStefano Zampini PetscScalar *array; 1106683d3df6SStefano Zampini PetscInt n; 1107683d3df6SStefano Zampini 1108683d3df6SStefano Zampini ierr = VecGetLocalSize(msolv_ctx->sol,&n);CHKERRQ(ierr); 1109683d3df6SStefano Zampini ierr = VecGetArray(msolv_ctx->sol,&array);CHKERRQ(ierr); 1110683d3df6SStefano Zampini ierr = VecCreateSeqWithArray(PetscObjectComm((PetscObject)msolv_ctx->sol),1,n,array,&msolv_ctx->rhs);CHKERRQ(ierr); 1111683d3df6SStefano Zampini ierr = VecRestoreArray(msolv_ctx->sol,&array);CHKERRQ(ierr); 1112683d3df6SStefano Zampini } 1113683d3df6SStefano Zampini msolv_ctx->has_vertices = factor_workaround; 1114d62866d3SStefano Zampini 1115d62866d3SStefano Zampini /* interior solver */ 1116683d3df6SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)A_II),&msolv_ctx->interior_solver);CHKERRQ(ierr); 1117d62866d3SStefano Zampini ierr = PCSetOperators(msolv_ctx->interior_solver,A_II,A_II);CHKERRQ(ierr); 1118d62866d3SStefano Zampini ierr = PCSetType(msolv_ctx->interior_solver,PCSHELL);CHKERRQ(ierr); 1119d62866d3SStefano Zampini ierr = PCShellSetContext(msolv_ctx->interior_solver,msolv_ctx);CHKERRQ(ierr); 1120df4d28bfSStefano Zampini ierr = PCShellSetApply(msolv_ctx->interior_solver,PCBDDCReuseSolvers_Interior);CHKERRQ(ierr); 1121df4d28bfSStefano Zampini ierr = PCShellSetApplyTranspose(msolv_ctx->interior_solver,PCBDDCReuseSolvers_InteriorTranspose);CHKERRQ(ierr); 1122d62866d3SStefano Zampini 1123d62866d3SStefano Zampini /* correction solver */ 11243462e049SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)A_II),&msolv_ctx->correction_solver);CHKERRQ(ierr); 1125d62866d3SStefano Zampini ierr = PCSetType(msolv_ctx->correction_solver,PCSHELL);CHKERRQ(ierr); 1126d62866d3SStefano Zampini ierr = PCShellSetContext(msolv_ctx->correction_solver,msolv_ctx);CHKERRQ(ierr); 1127df4d28bfSStefano Zampini ierr = PCShellSetApply(msolv_ctx->correction_solver,PCBDDCReuseSolvers_Correction);CHKERRQ(ierr); 1128df4d28bfSStefano Zampini ierr = PCShellSetApplyTranspose(msolv_ctx->correction_solver,PCBDDCReuseSolvers_CorrectionTranspose);CHKERRQ(ierr); 112953892102SStefano Zampini 113053892102SStefano Zampini /* scatter and vecs for Schur complement solver */ 113153892102SStefano Zampini ierr = MatCreateVecs(S_all,&msolv_ctx->sol_B,&msolv_ctx->rhs_B);CHKERRQ(ierr); 113253892102SStefano Zampini ierr = MatCreateVecs(sub_schurs->S,&vec1_B,NULL);CHKERRQ(ierr); 1133683d3df6SStefano Zampini if (!factor_workaround) { 113453892102SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(sub_schurs->BtoNmap,IS_GTOLM_DROP,is_A_all,&msolv_ctx->is_B);CHKERRQ(ierr); 113553892102SStefano Zampini ierr = VecScatterCreate(vec1_B,msolv_ctx->is_B,msolv_ctx->sol_B,NULL,&msolv_ctx->correction_scatter_B);CHKERRQ(ierr); 113653892102SStefano Zampini ierr = PetscObjectReference((PetscObject)is_A_all);CHKERRQ(ierr); 113753892102SStefano Zampini msolv_ctx->is_R = is_A_all; 1138683d3df6SStefano Zampini } else { 1139683d3df6SStefano Zampini IS is_B_all; 1140683d3df6SStefano Zampini const PetscInt* idxs; 1141683d3df6SStefano Zampini PetscInt dual,n_v,n; 1142683d3df6SStefano Zampini 1143683d3df6SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_v);CHKERRQ(ierr); 1144683d3df6SStefano Zampini dual = size_schur - n_v; 1145683d3df6SStefano Zampini ierr = ISGetLocalSize(is_A_all,&n);CHKERRQ(ierr); 1146683d3df6SStefano Zampini ierr = ISGetIndices(is_A_all,&idxs);CHKERRQ(ierr); 1147683d3df6SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)is_A_all),dual,idxs+n_I,PETSC_COPY_VALUES,&is_B_all);CHKERRQ(ierr); 1148683d3df6SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(sub_schurs->BtoNmap,IS_GTOLM_DROP,is_B_all,&msolv_ctx->is_B);CHKERRQ(ierr); 1149683d3df6SStefano Zampini ierr = ISDestroy(&is_B_all);CHKERRQ(ierr); 1150683d3df6SStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)is_A_all),dual,0,1,&is_B_all);CHKERRQ(ierr); 1151683d3df6SStefano Zampini ierr = VecScatterCreate(vec1_B,msolv_ctx->is_B,msolv_ctx->sol_B,is_B_all,&msolv_ctx->correction_scatter_B);CHKERRQ(ierr); 1152683d3df6SStefano Zampini ierr = ISDestroy(&is_B_all);CHKERRQ(ierr); 1153683d3df6SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)is_A_all),n-n_v,idxs,PETSC_COPY_VALUES,&msolv_ctx->is_R);CHKERRQ(ierr); 1154683d3df6SStefano Zampini ierr = ISRestoreIndices(is_A_all,&idxs);CHKERRQ(ierr); 1155683d3df6SStefano Zampini } 11563462e049SStefano Zampini ierr = ISGetLocalSize(msolv_ctx->is_R,&n_R);CHKERRQ(ierr); 11573462e049SStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,n_R,n_R,0,NULL,&Afake);CHKERRQ(ierr); 11583462e049SStefano Zampini ierr = PCSetOperators(msolv_ctx->correction_solver,Afake,Afake);CHKERRQ(ierr); 11593462e049SStefano Zampini ierr = MatDestroy(&Afake);CHKERRQ(ierr); 1160683d3df6SStefano Zampini ierr = VecDestroy(&vec1_B);CHKERRQ(ierr); 1161ca92afb2SStefano Zampini 1162ca92afb2SStefano Zampini /* communicate benign info to solver context */ 1163ca92afb2SStefano Zampini if (benign_n) { 11645cbda25cSStefano Zampini PetscScalar *array; 11655cbda25cSStefano Zampini 1166ca92afb2SStefano Zampini msolv_ctx->benign_n = benign_n; 1167ca92afb2SStefano Zampini msolv_ctx->benign_zerodiag_subs = is_p_r; 1168ca92afb2SStefano Zampini ierr = PetscMalloc1(benign_n,&msolv_ctx->benign_save_vals);CHKERRQ(ierr); 11695cbda25cSStefano Zampini msolv_ctx->benign_csAIB = cs_AIB_mat; 1170282d6408SStefano Zampini ierr = MatCreateVecs(cs_AIB_mat,&msolv_ctx->benign_corr_work,NULL);CHKERRQ(ierr); 11715cbda25cSStefano Zampini ierr = VecGetArray(msolv_ctx->benign_corr_work,&array);CHKERRQ(ierr); 11725cbda25cSStefano Zampini ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,size_schur,array,&msolv_ctx->benign_dummy_schur_vec);CHKERRQ(ierr); 11735cbda25cSStefano Zampini ierr = VecRestoreArray(msolv_ctx->benign_corr_work,&array);CHKERRQ(ierr); 11745cbda25cSStefano Zampini msolv_ctx->benign_AIIm1ones = benign_AIIm1_ones_mat; 1175ca92afb2SStefano Zampini } 1176d5574798SStefano Zampini } 117708122e43SStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 117853892102SStefano Zampini ierr = ISDestroy(&is_A_all);CHKERRQ(ierr); 11795db18549SStefano Zampini 1180be83ff47SStefano Zampini /* Work arrays */ 1181be83ff47SStefano Zampini ierr = PetscMalloc2(max_subset_size,&dummy_idx,max_subset_size*max_subset_size,&work);CHKERRQ(ierr); 11825a94c880SStefano Zampini 11835a94c880SStefano Zampini /* matrices for adaptive selection */ 118412d906b1SStefano Zampini if (compute_Stilda) { 11855a94c880SStefano Zampini if (!sub_schurs->sum_S_Ej_tilda_all) { 11865a94c880SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 11875a94c880SStefano Zampini ierr = MatSetSizes(sub_schurs->sum_S_Ej_tilda_all,PETSC_DECIDE,PETSC_DECIDE,size_active_schur,size_active_schur);CHKERRQ(ierr); 11885a94c880SStefano Zampini ierr = MatSetType(sub_schurs->sum_S_Ej_tilda_all,MATAIJ);CHKERRQ(ierr); 11895a94c880SStefano Zampini ierr = MatSeqAIJSetPreallocation(sub_schurs->sum_S_Ej_tilda_all,0,nnz);CHKERRQ(ierr); 11905a94c880SStefano Zampini } 11919d54b7f4SStefano Zampini if (!sub_schurs->sum_S_Ej_inv_all && deluxe) { 11925a94c880SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 11935a94c880SStefano Zampini ierr = MatSetSizes(sub_schurs->sum_S_Ej_inv_all,PETSC_DECIDE,PETSC_DECIDE,size_active_schur,size_active_schur);CHKERRQ(ierr); 11945a94c880SStefano Zampini ierr = MatSetType(sub_schurs->sum_S_Ej_inv_all,MATAIJ);CHKERRQ(ierr); 11955a94c880SStefano Zampini ierr = MatSeqAIJSetPreallocation(sub_schurs->sum_S_Ej_inv_all,0,nnz);CHKERRQ(ierr); 11965a94c880SStefano Zampini } 119712d906b1SStefano Zampini } 1198d2627357SStefano Zampini 1199be83ff47SStefano Zampini /* S_Ej_all */ 1200be83ff47SStefano Zampini cum = cum2 = 0; 1201be83ff47SStefano Zampini ierr = MatDenseGetArray(S_all,&S_data);CHKERRQ(ierr); 120265d8bf0aSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 120365d8bf0aSStefano Zampini PetscInt j; 120465d8bf0aSStefano Zampini 12055a95e1ceSStefano Zampini /* get S_E */ 1206b96c3477SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 1207683d3df6SStefano Zampini if (S_lower_triangular) { /* I need to expand the upper triangular data (column oriented) */ 1208be83ff47SStefano Zampini PetscInt k; 1209be83ff47SStefano Zampini for (k=0;k<subset_size;k++) { 1210be83ff47SStefano Zampini for (j=k;j<subset_size;j++) { 1211be83ff47SStefano Zampini work[k*subset_size+j] = S_data[cum2+k*size_schur+j]; 1212683d3df6SStefano Zampini work[j*subset_size+k] = PetscConj(S_data[cum2+k*size_schur+j]); 1213be83ff47SStefano Zampini } 1214be83ff47SStefano Zampini } 121506a4e24aSStefano Zampini } else { /* just copy to workspace */ 1216be83ff47SStefano Zampini PetscInt k; 1217be83ff47SStefano Zampini for (k=0;k<subset_size;k++) { 1218be83ff47SStefano Zampini for (j=0;j<subset_size;j++) { 1219be83ff47SStefano Zampini work[k*subset_size+j] = S_data[cum2+k*size_schur+j]; 1220be83ff47SStefano Zampini } 1221be83ff47SStefano Zampini } 12229087bf02SStefano Zampini } 12235a95e1ceSStefano Zampini /* insert S_E values */ 1224be83ff47SStefano Zampini for (j=0;j<subset_size;j++) dummy_idx[j] = cum+j; 1225b7ab4a40SStefano Zampini if (sub_schurs->change) { 12268760537fSStefano Zampini Mat change_sub,SEj,T; 122772b8c272SStefano Zampini 122872b8c272SStefano Zampini /* change basis */ 122972b8c272SStefano Zampini ierr = KSPGetOperators(sub_schurs->change[i],&change_sub,NULL);CHKERRQ(ierr); 123072b8c272SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,subset_size,work,&SEj);CHKERRQ(ierr); 12318760537fSStefano Zampini if (!sub_schurs->change_with_qr) { /* currently there's no support for PtAP with P SeqAIJ */ 12328760537fSStefano Zampini Mat T2; 12338760537fSStefano Zampini ierr = MatTransposeMatMult(change_sub,SEj,MAT_INITIAL_MATRIX,1.0,&T2);CHKERRQ(ierr); 12348760537fSStefano Zampini ierr = MatMatMult(T2,change_sub,MAT_INITIAL_MATRIX,1.0,&T);CHKERRQ(ierr); 12358760537fSStefano Zampini ierr = MatConvert(T,MATSEQDENSE,MAT_INPLACE_MATRIX,&T);CHKERRQ(ierr); 123672b8c272SStefano Zampini ierr = MatDestroy(&T2);CHKERRQ(ierr); 12378760537fSStefano Zampini } else { 12388760537fSStefano Zampini ierr = MatPtAP(SEj,change_sub,MAT_INITIAL_MATRIX,1.0,&T);CHKERRQ(ierr); 123972b8c272SStefano Zampini } 124072b8c272SStefano Zampini ierr = MatCopy(T,SEj,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 124172b8c272SStefano Zampini ierr = MatDestroy(&T);CHKERRQ(ierr); 1242afa0f562SStefano Zampini ierr = MatZeroRowsColumnsIS(SEj,sub_schurs->change_primal_sub[i],1.0,NULL,NULL);CHKERRQ(ierr); 124372b8c272SStefano Zampini ierr = MatDestroy(&SEj);CHKERRQ(ierr); 124472b8c272SStefano Zampini } 12459d54b7f4SStefano Zampini if (deluxe) { 12465a95e1ceSStefano Zampini ierr = MatSetValues(sub_schurs->S_Ej_all,subset_size,dummy_idx,subset_size,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr); 1247683d3df6SStefano Zampini /* if adaptivity is requested, invert S_E blocks */ 1248862806e4SStefano Zampini if (compute_Stilda) { 124908122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 125008122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 125106a4e24aSStefano Zampini if (!use_getr) { /* TODO add sytrf/i for symmetric non hermitian */ 12526c3e6151SStefano Zampini PetscInt k; 12535a95e1ceSStefano Zampini PetscStackCallBLAS("LAPACKpotrf",LAPACKpotrf_("L",&B_N,work,&B_N,&B_ierr)); 12542972d61bSStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRF Lapack routine %d",(int)B_ierr); 12555a95e1ceSStefano Zampini PetscStackCallBLAS("LAPACKpotri",LAPACKpotri_("L",&B_N,work,&B_N,&B_ierr)); 12562972d61bSStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRI Lapack routine %d",(int)B_ierr); 12576c3e6151SStefano Zampini for (k=0;k<subset_size;k++) { 12586c3e6151SStefano Zampini for (j=k;j<subset_size;j++) { 12596c3e6151SStefano Zampini work[j*subset_size+k] = work[k*subset_size+j]; 12606c3e6151SStefano Zampini } 12616c3e6151SStefano Zampini } 1262d6462365SStefano Zampini } else { 1263d6462365SStefano Zampini PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,work,&B_N,pivots,&B_ierr)); 1264d6462365SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr); 1265d6462365SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,work,&B_N,pivots,Bwork,&B_lwork,&B_ierr)); 1266d6462365SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr); 12672972d61bSStefano Zampini } 126808122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 12695a94c880SStefano Zampini ierr = MatSetValues(sub_schurs->sum_S_Ej_inv_all,subset_size,dummy_idx,subset_size,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr); 12709087bf02SStefano Zampini } 12719d54b7f4SStefano Zampini } else if (compute_Stilda) { /* not using deluxe */ 12729d54b7f4SStefano Zampini Mat SEj; 12739d54b7f4SStefano Zampini Vec D; 12749d54b7f4SStefano Zampini PetscScalar *array; 12759d54b7f4SStefano Zampini 12769d54b7f4SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,subset_size,work,&SEj);CHKERRQ(ierr); 12779d54b7f4SStefano Zampini ierr = VecGetArray(Dall,&array);CHKERRQ(ierr); 12789d54b7f4SStefano Zampini ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,subset_size,array+cum,&D);CHKERRQ(ierr); 12799d54b7f4SStefano Zampini ierr = VecRestoreArray(Dall,&array);CHKERRQ(ierr); 1280f17d2ae1SStefano Zampini ierr = VecShift(D,-1.);CHKERRQ(ierr); 12819d54b7f4SStefano Zampini ierr = MatDiagonalScale(SEj,D,D);CHKERRQ(ierr); 12829d54b7f4SStefano Zampini ierr = MatDestroy(&SEj);CHKERRQ(ierr); 12839d54b7f4SStefano Zampini ierr = VecDestroy(&D);CHKERRQ(ierr); 12849d54b7f4SStefano Zampini ierr = MatSetValues(sub_schurs->S_Ej_all,subset_size,dummy_idx,subset_size,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr); 12859d54b7f4SStefano Zampini } 1286be83ff47SStefano Zampini cum += subset_size; 1287be83ff47SStefano Zampini cum2 += subset_size*(size_schur + 1); 1288be83ff47SStefano Zampini } 1289be83ff47SStefano Zampini ierr = MatDenseRestoreArray(S_all,&S_data);CHKERRQ(ierr); 12904a6c6b0dSStefano Zampini 1291df4d28bfSStefano Zampini if (solver_S) { 12925a05ddb0SStefano Zampini ierr = MatFactorRestoreSchurComplement(F,&S_all);CHKERRQ(ierr); 12934a6c6b0dSStefano Zampini } 1294683d3df6SStefano Zampini 1295683d3df6SStefano Zampini schur_factor = NULL; 129645951f25SStefano Zampini if (compute_Stilda && size_active_schur) { 1297683d3df6SStefano Zampini 12989d54b7f4SStefano Zampini if (sub_schurs->n_subs == 1 && size_schur == size_active_schur && deluxe) { /* we already computed the inverse */ 12994a6c6b0dSStefano Zampini PetscInt j; 13004a6c6b0dSStefano Zampini for (j=0;j<size_schur;j++) dummy_idx[j] = j; 13015a94c880SStefano Zampini ierr = MatSetValues(sub_schurs->sum_S_Ej_tilda_all,size_schur,dummy_idx,size_schur,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr); 13024a6c6b0dSStefano Zampini } else { 1303683d3df6SStefano Zampini Mat S_all_inv=NULL; 1304df4d28bfSStefano Zampini if (solver_S) { /* use MatFactor calls to invert S */ 130572b8c272SStefano Zampini /* let MatFactor factor the Schur complement */ 13066dba178dSStefano Zampini ierr = MatFactorFactorizeSchurComplement(F);CHKERRQ(ierr); 1307683d3df6SStefano Zampini /* for adaptive selection we need S^-1; for solver reusage we need S_\Delta\Delta^-1. 1308683d3df6SStefano Zampini The latter is not the principal subminor for S^-1. However, the factors can be reused since S_\Delta\Delta is the leading principal submatrix of S */ 1309683d3df6SStefano Zampini if (factor_workaround) { 1310683d3df6SStefano Zampini PetscScalar *data; 1311683d3df6SStefano Zampini PetscInt nd = 0; 13126dba178dSStefano Zampini 1313683d3df6SStefano Zampini if (!sub_schurs->is_posdef) { 1314683d3df6SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Factor update not yet implemented for non SPD matrices"); 1315683d3df6SStefano Zampini } 13166dba178dSStefano Zampini ierr = MatFactorGetSchurComplement(F,&S_all_inv);CHKERRQ(ierr); 1317683d3df6SStefano Zampini ierr = MatDenseGetArray(S_all_inv,&data);CHKERRQ(ierr); 1318683d3df6SStefano Zampini if (sub_schurs->is_dir) { /* dirichlet dofs could have different scalings */ 1319683d3df6SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_dir,&nd);CHKERRQ(ierr); 1320683d3df6SStefano Zampini } 1321683d3df6SStefano Zampini ierr = PetscMalloc1((size_active_schur*(size_active_schur+1))/2+nd,&schur_factor);CHKERRQ(ierr); 1322683d3df6SStefano Zampini cum = 0; 1323683d3df6SStefano Zampini for (i=0;i<size_active_schur;i++) { 1324683d3df6SStefano Zampini ierr = PetscMemcpy(schur_factor+cum,data+i*(size_schur+1),(size_active_schur-i)*sizeof(PetscScalar));CHKERRQ(ierr); 1325683d3df6SStefano Zampini cum += size_active_schur-i; 1326683d3df6SStefano Zampini } 1327683d3df6SStefano Zampini for (i=0;i<nd;i++) schur_factor[cum+i] = data[(i+size_active_schur)*(size_schur+1)]; 13286dba178dSStefano Zampini /* invert without calling MatFactorInvertSchurComplement, since we are hacking */ 1329683d3df6SStefano Zampini ierr = PetscBLASIntCast(size_schur,&B_N);CHKERRQ(ierr); 1330683d3df6SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1331683d3df6SStefano Zampini PetscStackCallBLAS("LAPACKpotri",LAPACKpotri_("L",&B_N,data,&B_N,&B_ierr)); 1332683d3df6SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 1333683d3df6SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRI Lapack routine %d",(int)B_ierr); 1334683d3df6SStefano Zampini ierr = MatDenseRestoreArray(S_all_inv,&data);CHKERRQ(ierr); 1335683d3df6SStefano Zampini } else { 13366dba178dSStefano Zampini ierr = MatFactorInvertSchurComplement(F);CHKERRQ(ierr); 13376dba178dSStefano Zampini ierr = MatFactorGetSchurComplement(F,&S_all_inv);CHKERRQ(ierr); 1338683d3df6SStefano Zampini } 1339df4d28bfSStefano Zampini } else { /* we need to invert explicitly since we are not using MatFactor for S */ 1340683d3df6SStefano Zampini ierr = PetscObjectReference((PetscObject)S_all);CHKERRQ(ierr); 1341683d3df6SStefano Zampini S_all_inv = S_all; 1342683d3df6SStefano Zampini ierr = MatDenseGetArray(S_all_inv,&S_data);CHKERRQ(ierr); 1343be83ff47SStefano Zampini ierr = PetscBLASIntCast(size_schur,&B_N);CHKERRQ(ierr); 1344be83ff47SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 134506a4e24aSStefano Zampini if (!use_getr) { /* TODO add sytrf/i for symmetric non hermitian */ 1346be83ff47SStefano Zampini PetscStackCallBLAS("LAPACKpotrf",LAPACKpotrf_("L",&B_N,S_data,&B_N,&B_ierr)); 1347be83ff47SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRF Lapack routine %d",(int)B_ierr); 1348be83ff47SStefano Zampini PetscStackCallBLAS("LAPACKpotri",LAPACKpotri_("L",&B_N,S_data,&B_N,&B_ierr)); 1349be83ff47SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRI Lapack routine %d",(int)B_ierr); 1350d6462365SStefano Zampini } else { 1351d6462365SStefano Zampini PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,S_data,&B_N,pivots,&B_ierr)); 1352d6462365SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr); 1353d6462365SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,S_data,&B_N,pivots,Bwork,&B_lwork,&B_ierr)); 1354d6462365SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr); 1355be83ff47SStefano Zampini } 1356be83ff47SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 1357683d3df6SStefano Zampini ierr = MatDenseRestoreArray(S_all_inv,&S_data);CHKERRQ(ierr); 1358be83ff47SStefano Zampini } 1359be83ff47SStefano Zampini /* S_Ej_tilda_all */ 1360be83ff47SStefano Zampini cum = cum2 = 0; 1361683d3df6SStefano Zampini ierr = MatDenseGetArray(S_all_inv,&S_data);CHKERRQ(ierr); 1362be83ff47SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 1363be83ff47SStefano Zampini PetscInt j; 1364862806e4SStefano Zampini 1365862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 1366be83ff47SStefano Zampini /* get (St^-1)_E */ 136772b8c272SStefano Zampini /* Unless we are changing the variables, I don't need to expand to upper triangular since St^-1 136806a4e24aSStefano Zampini will be properly accessed later during adaptive selection */ 1369a0b0af32SStefano Zampini if (S_lower_triangular) { 1370be83ff47SStefano Zampini PetscInt k; 1371b7ab4a40SStefano Zampini if (sub_schurs->change) { 1372be83ff47SStefano Zampini for (k=0;k<subset_size;k++) { 1373be83ff47SStefano Zampini for (j=k;j<subset_size;j++) { 1374be83ff47SStefano Zampini work[k*subset_size+j] = S_data[cum2+k*size_schur+j]; 13756c3e6151SStefano Zampini work[j*subset_size+k] = work[k*subset_size+j]; 1376be83ff47SStefano Zampini } 1377be83ff47SStefano Zampini } 137872b8c272SStefano Zampini } else { 137972b8c272SStefano Zampini for (k=0;k<subset_size;k++) { 138072b8c272SStefano Zampini for (j=k;j<subset_size;j++) { 138172b8c272SStefano Zampini work[k*subset_size+j] = S_data[cum2+k*size_schur+j]; 138272b8c272SStefano Zampini } 138372b8c272SStefano Zampini } 138472b8c272SStefano Zampini } 138572b8c272SStefano Zampini } else { 1386be83ff47SStefano Zampini PetscInt k; 1387be83ff47SStefano Zampini for (k=0;k<subset_size;k++) { 1388be83ff47SStefano Zampini for (j=0;j<subset_size;j++) { 1389be83ff47SStefano Zampini work[k*subset_size+j] = S_data[cum2+k*size_schur+j]; 1390be83ff47SStefano Zampini } 1391be83ff47SStefano Zampini } 1392be83ff47SStefano Zampini } 1393b7ab4a40SStefano Zampini if (sub_schurs->change) { 13948760537fSStefano Zampini Mat change_sub,SEj,T; 139572b8c272SStefano Zampini 139672b8c272SStefano Zampini /* change basis */ 139772b8c272SStefano Zampini ierr = KSPGetOperators(sub_schurs->change[i],&change_sub,NULL);CHKERRQ(ierr); 139872b8c272SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,subset_size,work,&SEj);CHKERRQ(ierr); 13998760537fSStefano Zampini if (!sub_schurs->change_with_qr) { /* currently there's no support for PtAP with P SeqAIJ */ 14008760537fSStefano Zampini Mat T2; 14018760537fSStefano Zampini ierr = MatTransposeMatMult(change_sub,SEj,MAT_INITIAL_MATRIX,1.0,&T2);CHKERRQ(ierr); 14028760537fSStefano Zampini ierr = MatMatMult(T2,change_sub,MAT_INITIAL_MATRIX,1.0,&T);CHKERRQ(ierr); 140372b8c272SStefano Zampini ierr = MatDestroy(&T2);CHKERRQ(ierr); 14048760537fSStefano Zampini ierr = MatConvert(T,MATSEQDENSE,MAT_INPLACE_MATRIX,&T);CHKERRQ(ierr); 14058760537fSStefano Zampini } else { 14068760537fSStefano Zampini ierr = MatPtAP(SEj,change_sub,MAT_INITIAL_MATRIX,1.0,&T);CHKERRQ(ierr); 140772b8c272SStefano Zampini } 14088760537fSStefano Zampini ierr = MatCopy(T,SEj,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 14098760537fSStefano Zampini ierr = MatDestroy(&T);CHKERRQ(ierr); 141072b8c272SStefano Zampini /* set diagonal entry to a very large value to pick the basis we are eliminating as the first eigenvectors with adaptive selection */ 1411afa0f562SStefano Zampini ierr = MatZeroRowsColumnsIS(SEj,sub_schurs->change_primal_sub[i],1./PETSC_SMALL,NULL,NULL);CHKERRQ(ierr); 141272b8c272SStefano Zampini ierr = MatDestroy(&SEj);CHKERRQ(ierr); 141372b8c272SStefano Zampini } 1414be83ff47SStefano Zampini for (j=0;j<subset_size;j++) dummy_idx[j] = cum+j; 14155a94c880SStefano Zampini ierr = MatSetValues(sub_schurs->sum_S_Ej_tilda_all,subset_size,dummy_idx,subset_size,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr); 1416be83ff47SStefano Zampini cum += subset_size; 1417be83ff47SStefano Zampini cum2 += subset_size*(size_schur + 1); 1418883469d8SStefano Zampini } 1419683d3df6SStefano Zampini ierr = MatDenseRestoreArray(S_all_inv,&S_data);CHKERRQ(ierr); 1420df4d28bfSStefano Zampini if (solver_S) { 14216dba178dSStefano Zampini ierr = MatFactorRestoreSchurComplement(F,&S_all_inv);CHKERRQ(ierr); 14225db18549SStefano Zampini } 1423683d3df6SStefano Zampini ierr = MatDestroy(&S_all_inv);CHKERRQ(ierr); 1424683d3df6SStefano Zampini } 1425683d3df6SStefano Zampini 1426683d3df6SStefano Zampini /* move back factors to Schur data into F */ 1427683d3df6SStefano Zampini if (factor_workaround) { 1428683d3df6SStefano Zampini Mat S_tmp; 1429683d3df6SStefano Zampini PetscInt nv,nd=0; 1430683d3df6SStefano Zampini 1431df4d28bfSStefano Zampini if (!solver_S) { 1432683d3df6SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 1433683d3df6SStefano Zampini } 1434683d3df6SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 14356dba178dSStefano Zampini ierr = MatFactorGetSchurComplement(F,&S_tmp);CHKERRQ(ierr); 1436683d3df6SStefano Zampini if (sub_schurs->is_posdef) { 1437683d3df6SStefano Zampini PetscScalar *data; 1438683d3df6SStefano Zampini 1439683d3df6SStefano Zampini ierr = MatZeroEntries(S_tmp);CHKERRQ(ierr); 1440683d3df6SStefano Zampini ierr = MatDenseGetArray(S_tmp,&data);CHKERRQ(ierr); 1441683d3df6SStefano Zampini 1442683d3df6SStefano Zampini if (S_lower_triangular) { 1443683d3df6SStefano Zampini cum = 0; 1444683d3df6SStefano Zampini for (i=0;i<size_active_schur;i++) { 1445683d3df6SStefano Zampini ierr = PetscMemcpy(data+i*(size_schur+1),schur_factor+cum,(size_active_schur-i)*sizeof(PetscScalar));CHKERRQ(ierr); 1446683d3df6SStefano Zampini cum += size_active_schur-i; 1447683d3df6SStefano Zampini } 1448683d3df6SStefano Zampini } else { 1449683d3df6SStefano Zampini ierr = PetscMemcpy(data,schur_factor,size_schur*size_schur*sizeof(PetscScalar));CHKERRQ(ierr); 1450683d3df6SStefano Zampini } 1451683d3df6SStefano Zampini if (sub_schurs->is_dir) { 1452683d3df6SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_dir,&nd);CHKERRQ(ierr); 1453683d3df6SStefano Zampini for (i=0;i<nd;i++) { 1454683d3df6SStefano Zampini data[(i+size_active_schur)*(size_schur+1)] = schur_factor[cum+i]; 1455683d3df6SStefano Zampini } 1456683d3df6SStefano Zampini } 14576dba178dSStefano Zampini /* workaround: since I cannot modify the matrices used inside the solvers for the forward and backward substitutions, 1458683d3df6SStefano Zampini set the diagonal entry of the Schur factor to a very large value */ 1459683d3df6SStefano Zampini for (i=size_active_schur+nd;i<size_schur;i++) { 14606c3e6151SStefano Zampini data[i*(size_schur+1)] = infty; 1461683d3df6SStefano Zampini } 1462683d3df6SStefano Zampini ierr = MatDenseRestoreArray(S_tmp,&data);CHKERRQ(ierr); 1463683d3df6SStefano Zampini } else { 1464683d3df6SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Factor update not yet implemented for non SPD matrices"); 1465683d3df6SStefano Zampini } 14666dba178dSStefano Zampini ierr = MatFactorRestoreSchurComplement(F,&S_tmp);CHKERRQ(ierr); 14679087bf02SStefano Zampini } 1468367aa537SStefano Zampini } else if (factor_workaround) { /* we need to eliminate any unneeded coupling */ 1469367aa537SStefano Zampini PetscScalar *data; 1470367aa537SStefano Zampini PetscInt nd = 0; 1471367aa537SStefano Zampini 1472367aa537SStefano Zampini if (sub_schurs->is_dir) { /* dirichlet dofs could have different scalings */ 1473367aa537SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_dir,&nd);CHKERRQ(ierr); 1474367aa537SStefano Zampini } 1475367aa537SStefano Zampini ierr = MatFactorGetSchurComplement(F,&S_all);CHKERRQ(ierr); 1476367aa537SStefano Zampini ierr = MatDenseGetArray(S_all,&data);CHKERRQ(ierr); 1477367aa537SStefano Zampini for (i=0;i<size_active_schur;i++) { 1478367aa537SStefano Zampini ierr = PetscMemzero(data+i*size_schur+size_active_schur,(size_schur-size_active_schur)*sizeof(PetscScalar));CHKERRQ(ierr); 1479367aa537SStefano Zampini } 1480367aa537SStefano Zampini for (i=size_active_schur+nd;i<size_schur;i++) { 1481367aa537SStefano Zampini ierr = PetscMemzero(data+i*size_schur+size_active_schur,(size_schur-size_active_schur)*sizeof(PetscScalar));CHKERRQ(ierr); 14826c3e6151SStefano Zampini data[i*(size_schur+1)] = infty; 1483367aa537SStefano Zampini } 1484367aa537SStefano Zampini ierr = MatDenseRestoreArray(S_all,&data);CHKERRQ(ierr); 1485367aa537SStefano Zampini ierr = MatFactorRestoreSchurComplement(F,&S_all);CHKERRQ(ierr); 14864a6c6b0dSStefano Zampini } 14874a6c6b0dSStefano Zampini ierr = PetscFree2(dummy_idx,work);CHKERRQ(ierr); 1488683d3df6SStefano Zampini ierr = PetscFree(schur_factor);CHKERRQ(ierr); 14899d54b7f4SStefano Zampini ierr = VecDestroy(&Dall);CHKERRQ(ierr); 14904a6c6b0dSStefano Zampini } 14915a94c880SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 1492be83ff47SStefano Zampini ierr = MatDestroy(&F);CHKERRQ(ierr); 1493a9b99552SStefano Zampini ierr = ISDestroy(&is_I_layer);CHKERRQ(ierr); 1494a1337663SStefano Zampini ierr = MatDestroy(&S_all);CHKERRQ(ierr); 1495a64f4aa4SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 1496a64f4aa4SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 1497a64f4aa4SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 14985db18549SStefano Zampini ierr = MatAssemblyBegin(sub_schurs->S_Ej_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 14995db18549SStefano Zampini ierr = MatAssemblyEnd(sub_schurs->S_Ej_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 15005a95e1ceSStefano Zampini if (compute_Stilda) { 15015a94c880SStefano Zampini ierr = MatAssemblyBegin(sub_schurs->sum_S_Ej_tilda_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 15025a94c880SStefano Zampini ierr = MatAssemblyEnd(sub_schurs->sum_S_Ej_tilda_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 15039d54b7f4SStefano Zampini if (deluxe) { 15045a94c880SStefano Zampini ierr = MatAssemblyBegin(sub_schurs->sum_S_Ej_inv_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 15055a94c880SStefano Zampini ierr = MatAssemblyEnd(sub_schurs->sum_S_Ej_inv_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 150608122e43SStefano Zampini } 15079d54b7f4SStefano Zampini } 1508a1337663SStefano Zampini 15095db18549SStefano Zampini /* Global matrix of all assembled Schur on subsets */ 15105db18549SStefano Zampini ierr = MatISSetLocalMat(work_mat,sub_schurs->S_Ej_all);CHKERRQ(ierr); 15113927de2eSStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(work_mat,global_schur_subsets,PETSC_TRUE);CHKERRQ(ierr); 15123927de2eSStefano Zampini ierr = MatISGetMPIXAIJ(work_mat,MAT_REUSE_MATRIX,&global_schur_subsets);CHKERRQ(ierr); 15135a95e1ceSStefano Zampini 15145db18549SStefano Zampini /* Get local part of (\sum_j S_Ej) */ 15155a94c880SStefano Zampini if (!sub_schurs->sum_S_Ej_all) { 1516dc456d91SStefano Zampini ierr = MatGetSubMatrices(global_schur_subsets,1,&all_subsets_n,&all_subsets_n,MAT_INITIAL_MATRIX,&submats);CHKERRQ(ierr); 15175a94c880SStefano Zampini sub_schurs->sum_S_Ej_all = submats[0]; 15185a94c880SStefano Zampini } else { 15195a94c880SStefano Zampini ierr = PetscMalloc1(1,&submats);CHKERRQ(ierr); 15205a94c880SStefano Zampini submats[0] = sub_schurs->sum_S_Ej_all; 1521dc456d91SStefano Zampini ierr = MatGetSubMatrices(global_schur_subsets,1,&all_subsets_n,&all_subsets_n,MAT_REUSE_MATRIX,&submats);CHKERRQ(ierr); 15225a94c880SStefano Zampini } 152308122e43SStefano Zampini 1524f6f667cfSStefano Zampini /* Get local part of (\sum_j S^-1_Ej) (\sum_j St^-1_Ej) */ 15255a95e1ceSStefano Zampini if (compute_Stilda) { 15265a94c880SStefano Zampini ierr = MatISSetLocalMat(work_mat,sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 1527a1337663SStefano Zampini ierr = MatISGetMPIXAIJ(work_mat,MAT_REUSE_MATRIX,&global_schur_subsets);CHKERRQ(ierr); 15285a94c880SStefano Zampini submats[0] = sub_schurs->sum_S_Ej_tilda_all; 1529dc456d91SStefano Zampini ierr = MatGetSubMatrices(global_schur_subsets,1,&all_subsets_n,&all_subsets_n,MAT_REUSE_MATRIX,&submats);CHKERRQ(ierr); 15309d54b7f4SStefano Zampini if (deluxe) { 15315a94c880SStefano Zampini ierr = MatISSetLocalMat(work_mat,sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 153208122e43SStefano Zampini ierr = MatISGetMPIXAIJ(work_mat,MAT_REUSE_MATRIX,&global_schur_subsets);CHKERRQ(ierr); 15335a94c880SStefano Zampini submats[0] = sub_schurs->sum_S_Ej_inv_all; 1534dc456d91SStefano Zampini ierr = MatGetSubMatrices(global_schur_subsets,1,&all_subsets_n,&all_subsets_n,MAT_REUSE_MATRIX,&submats);CHKERRQ(ierr); 15359d54b7f4SStefano Zampini } else { 15369d54b7f4SStefano Zampini PetscScalar *array; 15379d54b7f4SStefano Zampini PetscInt cum; 15389d54b7f4SStefano Zampini 15399d54b7f4SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&array);CHKERRQ(ierr); 15409d54b7f4SStefano Zampini cum = 0; 15419d54b7f4SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 15429d54b7f4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 15439d54b7f4SStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 15449d54b7f4SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 15459d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKpotrf",LAPACKpotrf_("L",&B_N,array+cum,&B_N,&B_ierr)); 15469d54b7f4SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRF Lapack routine %d",(int)B_ierr); 15479d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKpotri",LAPACKpotri_("L",&B_N,array+cum,&B_N,&B_ierr)); 15489d54b7f4SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRI Lapack routine %d",(int)B_ierr); 15499d54b7f4SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 15509d54b7f4SStefano Zampini cum += subset_size*subset_size; 15519d54b7f4SStefano Zampini } 15529d54b7f4SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&array);CHKERRQ(ierr); 15539d54b7f4SStefano Zampini ierr = PetscObjectReference((PetscObject)sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 15549d54b7f4SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 15559d54b7f4SStefano Zampini sub_schurs->sum_S_Ej_inv_all = sub_schurs->sum_S_Ej_all; 15569d54b7f4SStefano Zampini } 155708122e43SStefano Zampini } 15583202ece2SStefano Zampini 15595a95e1ceSStefano Zampini /* free workspace */ 15605a94c880SStefano Zampini ierr = PetscFree(submats);CHKERRQ(ierr); 156106a4b1faSStefano Zampini ierr = PetscFree2(Bwork,pivots);CHKERRQ(ierr); 1562a1337663SStefano Zampini ierr = MatDestroy(&global_schur_subsets);CHKERRQ(ierr); 15633202ece2SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 1564dc456d91SStefano Zampini ierr = ISDestroy(&all_subsets_n);CHKERRQ(ierr); 15655a95e1ceSStefano Zampini ierr = PetscCommDestroy(&comm_n);CHKERRQ(ierr); 1566b1b3d7a2SStefano Zampini PetscFunctionReturn(0); 1567b1b3d7a2SStefano Zampini } 1568b1b3d7a2SStefano Zampini 1569b1b3d7a2SStefano Zampini #undef __FUNCT__ 1570b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCSubSchursInit" 1571a64f4aa4SStefano Zampini PetscErrorCode PCBDDCSubSchursInit(PCBDDCSubSchurs sub_schurs, IS is_I, IS is_B, PCBDDCGraph graph, ISLocalToGlobalMapping BtoNmap) 1572b1b3d7a2SStefano Zampini { 15739bb4a8caSStefano Zampini IS *faces,*edges,*all_cc,vertices; 15745a95e1ceSStefano Zampini PetscInt i,n_faces,n_edges,n_all_cc; 1575b1b3d7a2SStefano Zampini PetscBool is_sorted; 1576b1b3d7a2SStefano Zampini PetscErrorCode ierr; 1577b1b3d7a2SStefano Zampini 1578b1b3d7a2SStefano Zampini PetscFunctionBegin; 1579b1b3d7a2SStefano Zampini ierr = ISSorted(is_I,&is_sorted);CHKERRQ(ierr); 15806c4ed002SBarry Smith if (!is_sorted) SETERRQ(PetscObjectComm((PetscObject)is_I),PETSC_ERR_PLIB,"IS for I dofs should be shorted"); 1581b1b3d7a2SStefano Zampini ierr = ISSorted(is_B,&is_sorted);CHKERRQ(ierr); 15826c4ed002SBarry Smith if (!is_sorted) SETERRQ(PetscObjectComm((PetscObject)is_B),PETSC_ERR_PLIB,"IS for B dofs should be shorted"); 1583b1b3d7a2SStefano Zampini 1584b1b3d7a2SStefano Zampini /* reset any previous data */ 1585b1b3d7a2SStefano Zampini ierr = PCBDDCSubSchursReset(sub_schurs);CHKERRQ(ierr); 1586b1b3d7a2SStefano Zampini 15875a95e1ceSStefano Zampini /* get index sets for faces and edges (already sorted by global ordering) */ 15889bb4a8caSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&n_faces,&faces,&n_edges,&edges,&vertices);CHKERRQ(ierr); 1589b1b3d7a2SStefano Zampini n_all_cc = n_faces+n_edges; 159008122e43SStefano Zampini ierr = PetscBTCreate(n_all_cc,&sub_schurs->is_edge);CHKERRQ(ierr); 1591b1b3d7a2SStefano Zampini ierr = PetscMalloc1(n_all_cc,&all_cc);CHKERRQ(ierr); 1592b1b3d7a2SStefano Zampini for (i=0;i<n_faces;i++) { 1593b1b3d7a2SStefano Zampini all_cc[i] = faces[i]; 1594b1b3d7a2SStefano Zampini } 1595b1b3d7a2SStefano Zampini for (i=0;i<n_edges;i++) { 1596b1b3d7a2SStefano Zampini all_cc[n_faces+i] = edges[i]; 159708122e43SStefano Zampini ierr = PetscBTSet(sub_schurs->is_edge,n_faces+i);CHKERRQ(ierr); 1598b1b3d7a2SStefano Zampini } 1599b1b3d7a2SStefano Zampini ierr = PetscFree(faces);CHKERRQ(ierr); 1600b1b3d7a2SStefano Zampini ierr = PetscFree(edges);CHKERRQ(ierr); 1601d62866d3SStefano Zampini sub_schurs->is_dir = NULL; 1602d62866d3SStefano Zampini ierr = PCBDDCGraphGetDirichletDofsB(graph,&sub_schurs->is_dir);CHKERRQ(ierr); 1603b1b3d7a2SStefano Zampini 1604df4d28bfSStefano Zampini /* Determine if MatFactor can be used */ 1605df4d28bfSStefano Zampini sub_schurs->schur_explicit = PETSC_FALSE; 1606883469d8SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 1607df4d28bfSStefano Zampini sub_schurs->schur_explicit = PETSC_TRUE; 1608df4d28bfSStefano Zampini #endif 1609df4d28bfSStefano Zampini #if defined(PETSC_HAVE_MKL_PARDISO) 1610df4d28bfSStefano Zampini sub_schurs->schur_explicit = PETSC_TRUE; 1611883469d8SStefano Zampini #endif 1612b1b3d7a2SStefano Zampini 1613b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)is_I);CHKERRQ(ierr); 1614b1b3d7a2SStefano Zampini sub_schurs->is_I = is_I; 1615b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)is_B);CHKERRQ(ierr); 1616b1b3d7a2SStefano Zampini sub_schurs->is_B = is_B; 16175db18549SStefano Zampini ierr = PetscObjectReference((PetscObject)graph->l2gmap);CHKERRQ(ierr); 16185db18549SStefano Zampini sub_schurs->l2gmap = graph->l2gmap; 16195db18549SStefano Zampini ierr = PetscObjectReference((PetscObject)BtoNmap);CHKERRQ(ierr); 16205db18549SStefano Zampini sub_schurs->BtoNmap = BtoNmap; 16215a95e1ceSStefano Zampini sub_schurs->n_subs = n_all_cc; 1622b1b3d7a2SStefano Zampini sub_schurs->is_subs = all_cc; 1623df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { /* sort by local ordering if we are not using MatFactor */ 1624b96c3477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 1625b96c3477SStefano Zampini ierr = ISSort(sub_schurs->is_subs[i]);CHKERRQ(ierr); 1626b96c3477SStefano Zampini } 16279bb4a8caSStefano Zampini } 1628d62866d3SStefano Zampini sub_schurs->is_vertices = vertices; 1629b96c3477SStefano Zampini sub_schurs->S_Ej_all = NULL; 1630b96c3477SStefano Zampini sub_schurs->sum_S_Ej_all = NULL; 163108122e43SStefano Zampini sub_schurs->sum_S_Ej_inv_all = NULL; 1632b96c3477SStefano Zampini sub_schurs->sum_S_Ej_tilda_all = NULL; 1633b96c3477SStefano Zampini sub_schurs->is_Ej_all = NULL; 1634b1b3d7a2SStefano Zampini PetscFunctionReturn(0); 1635b1b3d7a2SStefano Zampini } 1636b1b3d7a2SStefano Zampini 1637b1b3d7a2SStefano Zampini #undef __FUNCT__ 163834a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCSubSchursCreate" 163934a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursCreate(PCBDDCSubSchurs *sub_schurs) 164034a97f8cSStefano Zampini { 164134a97f8cSStefano Zampini PCBDDCSubSchurs schurs_ctx; 164234a97f8cSStefano Zampini PetscErrorCode ierr; 164334a97f8cSStefano Zampini 164434a97f8cSStefano Zampini PetscFunctionBegin; 164534a97f8cSStefano Zampini ierr = PetscNew(&schurs_ctx);CHKERRQ(ierr); 16465ff63025SStefano Zampini schurs_ctx->n_subs = 0; 164734a97f8cSStefano Zampini *sub_schurs = schurs_ctx; 164834a97f8cSStefano Zampini PetscFunctionReturn(0); 164934a97f8cSStefano Zampini } 165034a97f8cSStefano Zampini 165134a97f8cSStefano Zampini #undef __FUNCT__ 165234a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCSubSchursReset" 165334a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursReset(PCBDDCSubSchurs sub_schurs) 165434a97f8cSStefano Zampini { 165534a97f8cSStefano Zampini PetscInt i; 165634a97f8cSStefano Zampini PetscErrorCode ierr; 165734a97f8cSStefano Zampini 165834a97f8cSStefano Zampini PetscFunctionBegin; 16591e9c79c2SStefano Zampini ierr = MatDestroy(&sub_schurs->A);CHKERRQ(ierr); 1660b1b3d7a2SStefano Zampini ierr = MatDestroy(&sub_schurs->S);CHKERRQ(ierr); 1661b1b3d7a2SStefano Zampini ierr = ISDestroy(&sub_schurs->is_I);CHKERRQ(ierr); 1662b1b3d7a2SStefano Zampini ierr = ISDestroy(&sub_schurs->is_B);CHKERRQ(ierr); 16635db18549SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&sub_schurs->l2gmap);CHKERRQ(ierr); 16645db18549SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&sub_schurs->BtoNmap);CHKERRQ(ierr); 166541c3ba1bSStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 166641c3ba1bSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 166708122e43SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 1668a1337663SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 16695db18549SStefano Zampini ierr = ISDestroy(&sub_schurs->is_Ej_all);CHKERRQ(ierr); 1670d62866d3SStefano Zampini ierr = ISDestroy(&sub_schurs->is_vertices);CHKERRQ(ierr); 1671d62866d3SStefano Zampini ierr = ISDestroy(&sub_schurs->is_dir);CHKERRQ(ierr); 167208122e43SStefano Zampini ierr = PetscBTDestroy(&sub_schurs->is_edge);CHKERRQ(ierr); 167334a97f8cSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 1674b1b3d7a2SStefano Zampini ierr = ISDestroy(&sub_schurs->is_subs[i]);CHKERRQ(ierr); 167534a97f8cSStefano Zampini } 16765ff63025SStefano Zampini if (sub_schurs->n_subs) { 1677b1b3d7a2SStefano Zampini ierr = PetscFree(sub_schurs->is_subs);CHKERRQ(ierr); 16783dc780c3SStefano Zampini } 1679df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 1680df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversReset(sub_schurs->reuse_solver);CHKERRQ(ierr); 1681d62866d3SStefano Zampini } 1682df4d28bfSStefano Zampini ierr = PetscFree(sub_schurs->reuse_solver);CHKERRQ(ierr); 168372b8c272SStefano Zampini if (sub_schurs->change) { 168472b8c272SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 168572b8c272SStefano Zampini ierr = KSPDestroy(&sub_schurs->change[i]);CHKERRQ(ierr); 1686b7ab4a40SStefano Zampini ierr = ISDestroy(&sub_schurs->change_primal_sub[i]);CHKERRQ(ierr); 168772b8c272SStefano Zampini } 168872b8c272SStefano Zampini } 168972b8c272SStefano Zampini ierr = PetscFree(sub_schurs->change);CHKERRQ(ierr); 1690b7ab4a40SStefano Zampini ierr = PetscFree(sub_schurs->change_primal_sub);CHKERRQ(ierr); 169134a97f8cSStefano Zampini sub_schurs->n_subs = 0; 169234a97f8cSStefano Zampini PetscFunctionReturn(0); 169334a97f8cSStefano Zampini } 169434a97f8cSStefano Zampini 169534a97f8cSStefano Zampini #undef __FUNCT__ 169634a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCAdjGetNextLayer_Private" 16972a155e38SStefano Zampini PETSC_STATIC_INLINE PetscErrorCode PCBDDCAdjGetNextLayer_Private(PetscInt* queue_tip,PetscInt n_prev,PetscBT touched,PetscInt* xadj,PetscInt* adjncy,PetscInt* n_added) 169834a97f8cSStefano Zampini { 169934a97f8cSStefano Zampini PetscInt i,j,n; 170034a97f8cSStefano Zampini PetscErrorCode ierr; 170134a97f8cSStefano Zampini 170234a97f8cSStefano Zampini PetscFunctionBegin; 170334a97f8cSStefano Zampini n = 0; 170434a97f8cSStefano Zampini for (i=-n_prev;i<0;i++) { 170534a97f8cSStefano Zampini PetscInt start_dof = queue_tip[i]; 170634a97f8cSStefano Zampini for (j=xadj[start_dof];j<xadj[start_dof+1];j++) { 170734a97f8cSStefano Zampini PetscInt dof = adjncy[j]; 170834a97f8cSStefano Zampini if (!PetscBTLookup(touched,dof)) { 170934a97f8cSStefano Zampini ierr = PetscBTSet(touched,dof);CHKERRQ(ierr); 171034a97f8cSStefano Zampini queue_tip[n] = dof; 171134a97f8cSStefano Zampini n++; 171234a97f8cSStefano Zampini } 171334a97f8cSStefano Zampini } 171434a97f8cSStefano Zampini } 171534a97f8cSStefano Zampini *n_added = n; 171634a97f8cSStefano Zampini PetscFunctionReturn(0); 171734a97f8cSStefano Zampini } 1718