134a97f8cSStefano Zampini #include <../src/ksp/pc/impls/bddc/bddc.h> 234a97f8cSStefano Zampini #include <../src/ksp/pc/impls/bddc/bddcprivate.h> 305709791SSatish Balay #include <../src/mat/impls/dense/seq/dense.h> 408122e43SStefano Zampini #include <petscblaslapack.h> 534a97f8cSStefano Zampini 69fbee547SJacob Faibussowitsch static inline PetscErrorCode PCBDDCAdjGetNextLayer_Private(PetscInt*,PetscInt,PetscBT,PetscInt*,PetscInt*,PetscInt*); 75ec10c6aSStefano Zampini static PetscErrorCode PCBDDCComputeExplicitSchur(Mat,PetscBool,MatReuse,Mat*); 8df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolvers_Interior(PC,Vec,Vec); 9df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolvers_Correction(PC,Vec,Vec); 10d62866d3SStefano Zampini 11ca92afb2SStefano Zampini /* if v2 is not present, correction is done in-place */ 125cbda25cSStefano Zampini PetscErrorCode PCBDDCReuseSolversBenignAdapt(PCBDDCReuseSolvers ctx, Vec v, Vec v2, PetscBool sol, PetscBool full) 13ca92afb2SStefano Zampini { 14ca92afb2SStefano Zampini PetscScalar *array; 15ca92afb2SStefano Zampini PetscScalar *array2; 16ca92afb2SStefano Zampini PetscErrorCode ierr; 17ca92afb2SStefano Zampini 18ca92afb2SStefano Zampini PetscFunctionBegin; 19ca92afb2SStefano Zampini if (!ctx->benign_n) PetscFunctionReturn(0); 205cbda25cSStefano Zampini if (sol && full) { 215cbda25cSStefano Zampini PetscInt n_I,size_schur; 225cbda25cSStefano Zampini 235cbda25cSStefano Zampini /* get sizes */ 24282d6408SStefano Zampini ierr = MatGetSize(ctx->benign_csAIB,&size_schur,NULL);CHKERRQ(ierr); 255cbda25cSStefano Zampini ierr = VecGetSize(v,&n_I);CHKERRQ(ierr); 265cbda25cSStefano Zampini n_I = n_I - size_schur; 275cbda25cSStefano Zampini /* get schur sol from array */ 285cbda25cSStefano Zampini ierr = VecGetArray(v,&array);CHKERRQ(ierr); 295cbda25cSStefano Zampini ierr = VecPlaceArray(ctx->benign_dummy_schur_vec,array+n_I);CHKERRQ(ierr); 305cbda25cSStefano Zampini ierr = VecRestoreArray(v,&array);CHKERRQ(ierr); 315cbda25cSStefano Zampini /* apply interior sol correction */ 32282d6408SStefano Zampini ierr = MatMultTranspose(ctx->benign_csAIB,ctx->benign_dummy_schur_vec,ctx->benign_corr_work);CHKERRQ(ierr); 335cbda25cSStefano Zampini ierr = VecResetArray(ctx->benign_dummy_schur_vec);CHKERRQ(ierr); 345cbda25cSStefano Zampini ierr = MatMultAdd(ctx->benign_AIIm1ones,ctx->benign_corr_work,v,v);CHKERRQ(ierr); 355cbda25cSStefano Zampini } 36ca92afb2SStefano Zampini if (v2) { 37ca92afb2SStefano Zampini PetscInt nl; 38ca92afb2SStefano Zampini 39ca92afb2SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 40ca92afb2SStefano Zampini ierr = VecGetLocalSize(v2,&nl);CHKERRQ(ierr); 41ca92afb2SStefano Zampini ierr = VecGetArray(v2,&array2);CHKERRQ(ierr); 42580bdb30SBarry Smith ierr = PetscArraycpy(array2,array,nl);CHKERRQ(ierr); 43ca92afb2SStefano Zampini } else { 44ca92afb2SStefano Zampini ierr = VecGetArray(v,&array);CHKERRQ(ierr); 45ca92afb2SStefano Zampini array2 = array; 46ca92afb2SStefano Zampini } 47ca92afb2SStefano Zampini if (!sol) { /* change rhs */ 48ca92afb2SStefano Zampini PetscInt n; 49ca92afb2SStefano Zampini for (n=0;n<ctx->benign_n;n++) { 50ca92afb2SStefano Zampini PetscScalar sum = 0.; 51ca92afb2SStefano Zampini const PetscInt *cols; 52ca92afb2SStefano Zampini PetscInt nz,i; 53ca92afb2SStefano Zampini 54ca92afb2SStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[n],&nz);CHKERRQ(ierr); 55ca92afb2SStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[n],&cols);CHKERRQ(ierr); 56ca92afb2SStefano Zampini for (i=0;i<nz-1;i++) sum += array[cols[i]]; 5722db5ddcSStefano Zampini #if defined(PETSC_USE_COMPLEX) 5822db5ddcSStefano Zampini sum = -(PetscRealPart(sum)/nz + PETSC_i*(PetscImaginaryPart(sum)/nz)); 5922db5ddcSStefano Zampini #else 60ca92afb2SStefano Zampini sum = -sum/nz; 6122db5ddcSStefano Zampini #endif 62ca92afb2SStefano Zampini for (i=0;i<nz-1;i++) array2[cols[i]] += sum; 63ca92afb2SStefano Zampini ctx->benign_save_vals[n] = array2[cols[nz-1]]; 64ca92afb2SStefano Zampini array2[cols[nz-1]] = sum; 65ca92afb2SStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[n],&cols);CHKERRQ(ierr); 66ca92afb2SStefano Zampini } 67ca92afb2SStefano Zampini } else { 68ca92afb2SStefano Zampini PetscInt n; 69ca92afb2SStefano Zampini for (n=0;n<ctx->benign_n;n++) { 70ca92afb2SStefano Zampini PetscScalar sum = 0.; 71ca92afb2SStefano Zampini const PetscInt *cols; 72ca92afb2SStefano Zampini PetscInt nz,i; 73ca92afb2SStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[n],&nz);CHKERRQ(ierr); 74ca92afb2SStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[n],&cols);CHKERRQ(ierr); 75ca92afb2SStefano Zampini for (i=0;i<nz-1;i++) sum += array[cols[i]]; 7622db5ddcSStefano Zampini #if defined(PETSC_USE_COMPLEX) 7722db5ddcSStefano Zampini sum = -(PetscRealPart(sum)/nz + PETSC_i*(PetscImaginaryPart(sum)/nz)); 7822db5ddcSStefano Zampini #else 79ca92afb2SStefano Zampini sum = -sum/nz; 8022db5ddcSStefano Zampini #endif 81ca92afb2SStefano Zampini for (i=0;i<nz-1;i++) array2[cols[i]] += sum; 82ca92afb2SStefano Zampini array2[cols[nz-1]] = ctx->benign_save_vals[n]; 83ca92afb2SStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[n],&cols);CHKERRQ(ierr); 84ca92afb2SStefano Zampini } 85ca92afb2SStefano Zampini } 86ca92afb2SStefano Zampini if (v2) { 87ca92afb2SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 8872b8c272SStefano Zampini ierr = VecRestoreArray(v2,&array2);CHKERRQ(ierr); 89ca92afb2SStefano Zampini } else { 90ca92afb2SStefano Zampini ierr = VecRestoreArray(v,&array);CHKERRQ(ierr); 91ca92afb2SStefano Zampini } 925cbda25cSStefano Zampini if (!sol && full) { 935cbda25cSStefano Zampini Vec usedv; 945cbda25cSStefano Zampini PetscInt n_I,size_schur; 955cbda25cSStefano Zampini 965cbda25cSStefano Zampini /* get sizes */ 97282d6408SStefano Zampini ierr = MatGetSize(ctx->benign_csAIB,&size_schur,NULL);CHKERRQ(ierr); 985cbda25cSStefano Zampini ierr = VecGetSize(v,&n_I);CHKERRQ(ierr); 995cbda25cSStefano Zampini n_I = n_I - size_schur; 1005cbda25cSStefano Zampini /* compute schur rhs correction */ 1015cbda25cSStefano Zampini if (v2) { 1025cbda25cSStefano Zampini usedv = v2; 1035cbda25cSStefano Zampini } else { 1045cbda25cSStefano Zampini usedv = v; 1055cbda25cSStefano Zampini } 1065cbda25cSStefano Zampini /* apply schur rhs correction */ 1075cbda25cSStefano Zampini ierr = MatMultTranspose(ctx->benign_AIIm1ones,usedv,ctx->benign_corr_work);CHKERRQ(ierr); 1085cbda25cSStefano Zampini ierr = VecGetArrayRead(usedv,(const PetscScalar**)&array);CHKERRQ(ierr); 1095cbda25cSStefano Zampini ierr = VecPlaceArray(ctx->benign_dummy_schur_vec,array+n_I);CHKERRQ(ierr); 1105cbda25cSStefano Zampini ierr = VecRestoreArrayRead(usedv,(const PetscScalar**)&array);CHKERRQ(ierr); 111282d6408SStefano Zampini ierr = MatMultAdd(ctx->benign_csAIB,ctx->benign_corr_work,ctx->benign_dummy_schur_vec,ctx->benign_dummy_schur_vec);CHKERRQ(ierr); 1125cbda25cSStefano Zampini ierr = VecResetArray(ctx->benign_dummy_schur_vec);CHKERRQ(ierr); 1135cbda25cSStefano Zampini } 114ca92afb2SStefano Zampini PetscFunctionReturn(0); 115ca92afb2SStefano Zampini } 116ca92afb2SStefano Zampini 117df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolvers_Solve_Private(PC pc, Vec rhs, Vec sol, PetscBool transpose, PetscBool full) 118d62866d3SStefano Zampini { 119df4d28bfSStefano Zampini PCBDDCReuseSolvers ctx; 120683d3df6SStefano Zampini PetscBool copy = PETSC_FALSE; 121d62866d3SStefano Zampini PetscErrorCode ierr; 122d62866d3SStefano Zampini 123d62866d3SStefano Zampini PetscFunctionBegin; 1243ec1f749SStefano Zampini ierr = PCShellGetContext(pc,&ctx);CHKERRQ(ierr); 125683d3df6SStefano Zampini if (full) { 126d62866d3SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 127d62866d3SStefano Zampini ierr = MatMumpsSetIcntl(ctx->F,26,-1);CHKERRQ(ierr); 128d62866d3SStefano Zampini #endif 1295cbda25cSStefano Zampini #if defined(PETSC_HAVE_MKL_PARDISO) 1305cbda25cSStefano Zampini ierr = MatMkl_PardisoSetCntl(ctx->F,70,0);CHKERRQ(ierr); 1315cbda25cSStefano Zampini #endif 132683d3df6SStefano Zampini copy = ctx->has_vertices; 133d4933d67SStefano Zampini } else { /* interior solver */ 1346dba178dSStefano Zampini #if defined(PETSC_HAVE_MUMPS) 135683d3df6SStefano Zampini ierr = MatMumpsSetIcntl(ctx->F,26,0);CHKERRQ(ierr); 1366dba178dSStefano Zampini #endif 137d4933d67SStefano Zampini #if defined(PETSC_HAVE_MKL_PARDISO) 138d4933d67SStefano Zampini ierr = MatMkl_PardisoSetCntl(ctx->F,70,1);CHKERRQ(ierr); 139d4933d67SStefano Zampini #endif 140683d3df6SStefano Zampini copy = PETSC_TRUE; 141683d3df6SStefano Zampini } 142683d3df6SStefano Zampini /* copy rhs into factored matrix workspace */ 143683d3df6SStefano Zampini if (copy) { 144ca92afb2SStefano Zampini PetscInt n; 145df4d28bfSStefano Zampini PetscScalar *array,*array_solver; 146ca92afb2SStefano Zampini 147ca92afb2SStefano Zampini ierr = VecGetLocalSize(rhs,&n);CHKERRQ(ierr); 148683d3df6SStefano Zampini ierr = VecGetArrayRead(rhs,(const PetscScalar**)&array);CHKERRQ(ierr); 149df4d28bfSStefano Zampini ierr = VecGetArray(ctx->rhs,&array_solver);CHKERRQ(ierr); 150580bdb30SBarry Smith ierr = PetscArraycpy(array_solver,array,n);CHKERRQ(ierr); 151df4d28bfSStefano Zampini ierr = VecRestoreArray(ctx->rhs,&array_solver);CHKERRQ(ierr); 152683d3df6SStefano Zampini ierr = VecRestoreArrayRead(rhs,(const PetscScalar**)&array);CHKERRQ(ierr); 153683d3df6SStefano Zampini 1545cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(ctx,ctx->rhs,NULL,PETSC_FALSE,full);CHKERRQ(ierr); 155683d3df6SStefano Zampini if (transpose) { 156683d3df6SStefano Zampini ierr = MatSolveTranspose(ctx->F,ctx->rhs,ctx->sol);CHKERRQ(ierr); 157683d3df6SStefano Zampini } else { 158683d3df6SStefano Zampini ierr = MatSolve(ctx->F,ctx->rhs,ctx->sol);CHKERRQ(ierr); 159683d3df6SStefano Zampini } 1605cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(ctx,ctx->sol,NULL,PETSC_TRUE,full);CHKERRQ(ierr); 161683d3df6SStefano Zampini 162683d3df6SStefano Zampini /* get back data to caller worskpace */ 163df4d28bfSStefano Zampini ierr = VecGetArrayRead(ctx->sol,(const PetscScalar**)&array_solver);CHKERRQ(ierr); 164683d3df6SStefano Zampini ierr = VecGetArray(sol,&array);CHKERRQ(ierr); 165580bdb30SBarry Smith ierr = PetscArraycpy(array,array_solver,n);CHKERRQ(ierr); 166683d3df6SStefano Zampini ierr = VecRestoreArray(sol,&array);CHKERRQ(ierr); 167df4d28bfSStefano Zampini ierr = VecRestoreArrayRead(ctx->sol,(const PetscScalar**)&array_solver);CHKERRQ(ierr); 168683d3df6SStefano Zampini } else { 169ca92afb2SStefano Zampini if (ctx->benign_n) { 1705cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(ctx,rhs,ctx->rhs,PETSC_FALSE,full);CHKERRQ(ierr); 171ca92afb2SStefano Zampini if (transpose) { 172ca92afb2SStefano Zampini ierr = MatSolveTranspose(ctx->F,ctx->rhs,sol);CHKERRQ(ierr); 173ca92afb2SStefano Zampini } else { 174ca92afb2SStefano Zampini ierr = MatSolve(ctx->F,ctx->rhs,sol);CHKERRQ(ierr); 175ca92afb2SStefano Zampini } 1765cbda25cSStefano Zampini ierr = PCBDDCReuseSolversBenignAdapt(ctx,sol,NULL,PETSC_TRUE,full);CHKERRQ(ierr); 177ca92afb2SStefano Zampini } else { 178e28d306cSStefano Zampini if (transpose) { 179e28d306cSStefano Zampini ierr = MatSolveTranspose(ctx->F,rhs,sol);CHKERRQ(ierr); 180e28d306cSStefano Zampini } else { 1816816873aSStefano Zampini ierr = MatSolve(ctx->F,rhs,sol);CHKERRQ(ierr); 182e28d306cSStefano Zampini } 183683d3df6SStefano Zampini } 184ca92afb2SStefano Zampini } 1855cbda25cSStefano Zampini /* restore defaults */ 1865cbda25cSStefano Zampini #if defined(PETSC_HAVE_MUMPS) 1875cbda25cSStefano Zampini ierr = MatMumpsSetIcntl(ctx->F,26,-1);CHKERRQ(ierr); 1885cbda25cSStefano Zampini #endif 189d4933d67SStefano Zampini #if defined(PETSC_HAVE_MKL_PARDISO) 190d4933d67SStefano Zampini ierr = MatMkl_PardisoSetCntl(ctx->F,70,0);CHKERRQ(ierr); 191d4933d67SStefano Zampini #endif 192d62866d3SStefano Zampini PetscFunctionReturn(0); 193d62866d3SStefano Zampini } 194d62866d3SStefano Zampini 195df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolvers_Correction(PC pc, Vec rhs, Vec sol) 196e28d306cSStefano Zampini { 197e28d306cSStefano Zampini PetscErrorCode ierr; 198e28d306cSStefano Zampini 199e28d306cSStefano Zampini PetscFunctionBegin; 200df4d28bfSStefano Zampini ierr = PCBDDCReuseSolvers_Solve_Private(pc,rhs,sol,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 201e28d306cSStefano Zampini PetscFunctionReturn(0); 202e28d306cSStefano Zampini } 203e28d306cSStefano Zampini 204df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolvers_CorrectionTranspose(PC pc, Vec rhs, Vec sol) 205e28d306cSStefano Zampini { 206e28d306cSStefano Zampini PetscErrorCode ierr; 207e28d306cSStefano Zampini 208e28d306cSStefano Zampini PetscFunctionBegin; 209df4d28bfSStefano Zampini ierr = PCBDDCReuseSolvers_Solve_Private(pc,rhs,sol,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 210683d3df6SStefano Zampini PetscFunctionReturn(0); 211683d3df6SStefano Zampini } 212683d3df6SStefano Zampini 213df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolvers_Interior(PC pc, Vec rhs, Vec sol) 214683d3df6SStefano Zampini { 215683d3df6SStefano Zampini PetscErrorCode ierr; 216683d3df6SStefano Zampini 217683d3df6SStefano Zampini PetscFunctionBegin; 218df4d28bfSStefano Zampini ierr = PCBDDCReuseSolvers_Solve_Private(pc,rhs,sol,PETSC_FALSE,PETSC_FALSE);CHKERRQ(ierr); 219683d3df6SStefano Zampini PetscFunctionReturn(0); 220683d3df6SStefano Zampini } 221683d3df6SStefano Zampini 222df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolvers_InteriorTranspose(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_TRUE,PETSC_FALSE);CHKERRQ(ierr); 228e28d306cSStefano Zampini PetscFunctionReturn(0); 229e28d306cSStefano Zampini } 230e28d306cSStefano Zampini 23115579a77SStefano Zampini static PetscErrorCode PCBDDCReuseSolvers_View(PC pc, PetscViewer viewer) 23215579a77SStefano Zampini { 23315579a77SStefano Zampini PCBDDCReuseSolvers ctx; 23415579a77SStefano Zampini PetscBool iascii; 23515579a77SStefano Zampini PetscErrorCode ierr; 23615579a77SStefano Zampini 23715579a77SStefano Zampini PetscFunctionBegin; 2383ec1f749SStefano Zampini ierr = PCShellGetContext(pc,&ctx);CHKERRQ(ierr); 23915579a77SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 24015579a77SStefano Zampini if (iascii) { 24115579a77SStefano Zampini ierr = PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_INFO);CHKERRQ(ierr); 24215579a77SStefano Zampini } 24315579a77SStefano Zampini ierr = MatView(ctx->F,viewer);CHKERRQ(ierr); 24415579a77SStefano Zampini if (iascii) { 24515579a77SStefano Zampini ierr = PetscViewerPopFormat(viewer);CHKERRQ(ierr); 24615579a77SStefano Zampini } 24715579a77SStefano Zampini PetscFunctionReturn(0); 24815579a77SStefano Zampini } 24915579a77SStefano Zampini 250df4d28bfSStefano Zampini static PetscErrorCode PCBDDCReuseSolversReset(PCBDDCReuseSolvers reuse) 251d62866d3SStefano Zampini { 252ca92afb2SStefano Zampini PetscInt i; 253d62866d3SStefano Zampini PetscErrorCode ierr; 254d62866d3SStefano Zampini 255d62866d3SStefano Zampini PetscFunctionBegin; 256d62866d3SStefano Zampini ierr = MatDestroy(&reuse->F);CHKERRQ(ierr); 257d62866d3SStefano Zampini ierr = VecDestroy(&reuse->sol);CHKERRQ(ierr); 258d62866d3SStefano Zampini ierr = VecDestroy(&reuse->rhs);CHKERRQ(ierr); 259d62866d3SStefano Zampini ierr = PCDestroy(&reuse->interior_solver);CHKERRQ(ierr); 260d62866d3SStefano Zampini ierr = PCDestroy(&reuse->correction_solver);CHKERRQ(ierr); 26153892102SStefano Zampini ierr = ISDestroy(&reuse->is_R);CHKERRQ(ierr); 26253892102SStefano Zampini ierr = ISDestroy(&reuse->is_B);CHKERRQ(ierr); 263d62866d3SStefano Zampini ierr = VecScatterDestroy(&reuse->correction_scatter_B);CHKERRQ(ierr); 26453892102SStefano Zampini ierr = VecDestroy(&reuse->sol_B);CHKERRQ(ierr); 26553892102SStefano Zampini ierr = VecDestroy(&reuse->rhs_B);CHKERRQ(ierr); 266ca92afb2SStefano Zampini for (i=0;i<reuse->benign_n;i++) { 267ca92afb2SStefano Zampini ierr = ISDestroy(&reuse->benign_zerodiag_subs[i]);CHKERRQ(ierr); 268ca92afb2SStefano Zampini } 269ca92afb2SStefano Zampini ierr = PetscFree(reuse->benign_zerodiag_subs);CHKERRQ(ierr); 270ca92afb2SStefano Zampini ierr = PetscFree(reuse->benign_save_vals);CHKERRQ(ierr); 2715cbda25cSStefano Zampini ierr = MatDestroy(&reuse->benign_csAIB);CHKERRQ(ierr); 2725cbda25cSStefano Zampini ierr = MatDestroy(&reuse->benign_AIIm1ones);CHKERRQ(ierr); 2735cbda25cSStefano Zampini ierr = VecDestroy(&reuse->benign_corr_work);CHKERRQ(ierr); 2745cbda25cSStefano Zampini ierr = VecDestroy(&reuse->benign_dummy_schur_vec);CHKERRQ(ierr); 275d62866d3SStefano Zampini PetscFunctionReturn(0); 276d62866d3SStefano Zampini } 277d5574798SStefano Zampini 2785ec10c6aSStefano Zampini static PetscErrorCode PCBDDCComputeExplicitSchur(Mat M, PetscBool issym, MatReuse reuse, Mat *S) 2793202ece2SStefano Zampini { 2803202ece2SStefano Zampini Mat B, C, D, Bd, Cd, AinvBd; 2813202ece2SStefano Zampini KSP ksp; 2823202ece2SStefano Zampini PC pc; 2833202ece2SStefano Zampini PetscBool isLU, isILU, isCHOL, Bdense, Cdense; 2843202ece2SStefano Zampini PetscReal fill = 2.0; 285f11841e3SStefano Zampini PetscInt n_I; 2863202ece2SStefano Zampini PetscMPIInt size; 2873202ece2SStefano Zampini PetscErrorCode ierr; 2883202ece2SStefano Zampini 2893202ece2SStefano Zampini PetscFunctionBegin; 290ffc4695bSBarry Smith ierr = MPI_Comm_size(PetscObjectComm((PetscObject)M),&size);CHKERRMPI(ierr); 291*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(size != 1,PETSC_COMM_SELF,PETSC_ERR_SUP,"Not for parallel matrices"); 292f11841e3SStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 293f11841e3SStefano Zampini PetscBool Sdense; 294f11841e3SStefano Zampini 295f11841e3SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*S, MATSEQDENSE, &Sdense);CHKERRQ(ierr); 296*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(!Sdense,PetscObjectComm((PetscObject)M),PETSC_ERR_SUP,"S should dense"); 297f11841e3SStefano Zampini } 2983202ece2SStefano Zampini ierr = MatSchurComplementGetSubMatrices(M, NULL, NULL, &B, &C, &D);CHKERRQ(ierr); 2993202ece2SStefano Zampini ierr = MatSchurComplementGetKSP(M, &ksp);CHKERRQ(ierr); 3003202ece2SStefano Zampini ierr = KSPGetPC(ksp, &pc);CHKERRQ(ierr); 3013202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject) pc, PCLU, &isLU);CHKERRQ(ierr); 3023202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject) pc, PCILU, &isILU);CHKERRQ(ierr); 3033202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject) pc, PCCHOLESKY, &isCHOL);CHKERRQ(ierr); 3043202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject) B, MATSEQDENSE, &Bdense);CHKERRQ(ierr); 3053202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject) C, MATSEQDENSE, &Cdense);CHKERRQ(ierr); 306f11841e3SStefano Zampini ierr = MatGetSize(B,&n_I,NULL);CHKERRQ(ierr); 307f11841e3SStefano Zampini if (n_I) { 3083202ece2SStefano Zampini if (!Bdense) { 3093202ece2SStefano Zampini ierr = MatConvert(B, MATSEQDENSE, MAT_INITIAL_MATRIX, &Bd);CHKERRQ(ierr); 3103202ece2SStefano Zampini } else { 3113202ece2SStefano Zampini Bd = B; 3123202ece2SStefano Zampini } 3133202ece2SStefano Zampini 3143202ece2SStefano Zampini if (isLU || isILU || isCHOL) { 3153202ece2SStefano Zampini Mat fact; 3163202ece2SStefano Zampini ierr = KSPSetUp(ksp);CHKERRQ(ierr); 3173202ece2SStefano Zampini ierr = PCFactorGetMatrix(pc, &fact);CHKERRQ(ierr); 3183202ece2SStefano Zampini ierr = MatDuplicate(Bd, MAT_DO_NOT_COPY_VALUES, &AinvBd);CHKERRQ(ierr); 3193202ece2SStefano Zampini ierr = MatMatSolve(fact, Bd, AinvBd);CHKERRQ(ierr); 3203202ece2SStefano Zampini } else { 32107b1e237SStefano Zampini PetscBool ex = PETSC_TRUE; 32207b1e237SStefano Zampini 32307b1e237SStefano Zampini if (ex) { 3243202ece2SStefano Zampini Mat Ainvd; 3253202ece2SStefano Zampini 3260bacdadaSStefano Zampini ierr = PCComputeOperator(pc, MATDENSE, &Ainvd);CHKERRQ(ierr); 3273202ece2SStefano Zampini ierr = MatMatMult(Ainvd, Bd, MAT_INITIAL_MATRIX, fill, &AinvBd);CHKERRQ(ierr); 3283202ece2SStefano Zampini ierr = MatDestroy(&Ainvd);CHKERRQ(ierr); 32907b1e237SStefano Zampini } else { 33007b1e237SStefano Zampini Vec sol,rhs; 33107b1e237SStefano Zampini PetscScalar *arrayrhs,*arraysol; 33207b1e237SStefano Zampini PetscInt i,nrhs,n; 33307b1e237SStefano Zampini 33407b1e237SStefano Zampini ierr = MatDuplicate(Bd, MAT_DO_NOT_COPY_VALUES, &AinvBd);CHKERRQ(ierr); 33507b1e237SStefano Zampini ierr = MatGetSize(Bd,&n,&nrhs);CHKERRQ(ierr); 33607b1e237SStefano Zampini ierr = MatDenseGetArray(Bd,&arrayrhs);CHKERRQ(ierr); 33707b1e237SStefano Zampini ierr = MatDenseGetArray(AinvBd,&arraysol);CHKERRQ(ierr); 33807b1e237SStefano Zampini ierr = KSPGetSolution(ksp,&sol);CHKERRQ(ierr); 33907b1e237SStefano Zampini ierr = KSPGetRhs(ksp,&rhs);CHKERRQ(ierr); 34007b1e237SStefano Zampini for (i=0;i<nrhs;i++) { 34107b1e237SStefano Zampini ierr = VecPlaceArray(rhs,arrayrhs+i*n);CHKERRQ(ierr); 34207b1e237SStefano Zampini ierr = VecPlaceArray(sol,arraysol+i*n);CHKERRQ(ierr); 34307b1e237SStefano Zampini ierr = KSPSolve(ksp,rhs,sol);CHKERRQ(ierr); 34407b1e237SStefano Zampini ierr = VecResetArray(rhs);CHKERRQ(ierr); 34507b1e237SStefano Zampini ierr = VecResetArray(sol);CHKERRQ(ierr); 34607b1e237SStefano Zampini } 34707b1e237SStefano Zampini ierr = MatDenseRestoreArray(Bd,&arrayrhs);CHKERRQ(ierr); 34807b1e237SStefano Zampini ierr = MatDenseRestoreArray(AinvBd,&arrayrhs);CHKERRQ(ierr); 34907b1e237SStefano Zampini } 3503202ece2SStefano Zampini } 3515ec10c6aSStefano Zampini if (!Bdense & !issym) { 3523202ece2SStefano Zampini ierr = MatDestroy(&Bd);CHKERRQ(ierr); 3533202ece2SStefano Zampini } 3545ec10c6aSStefano Zampini 3555ec10c6aSStefano Zampini if (!issym) { 3563202ece2SStefano Zampini if (!Cdense) { 3573202ece2SStefano Zampini ierr = MatConvert(C, MATSEQDENSE, MAT_INITIAL_MATRIX, &Cd);CHKERRQ(ierr); 3583202ece2SStefano Zampini } else { 3593202ece2SStefano Zampini Cd = C; 3603202ece2SStefano Zampini } 3615ec10c6aSStefano Zampini ierr = MatMatMult(Cd, AinvBd, reuse, fill, S);CHKERRQ(ierr); 3623202ece2SStefano Zampini if (!Cdense) { 3633202ece2SStefano Zampini ierr = MatDestroy(&Cd);CHKERRQ(ierr); 3643202ece2SStefano Zampini } 3655ec10c6aSStefano Zampini } else { 3665ec10c6aSStefano Zampini ierr = MatTransposeMatMult(Bd, AinvBd, reuse, fill, S);CHKERRQ(ierr); 3675ec10c6aSStefano Zampini if (!Bdense) { 3685ec10c6aSStefano Zampini ierr = MatDestroy(&Bd);CHKERRQ(ierr); 3695ec10c6aSStefano Zampini } 3705ec10c6aSStefano Zampini } 3715ec10c6aSStefano Zampini ierr = MatDestroy(&AinvBd);CHKERRQ(ierr); 372f11841e3SStefano Zampini } 3733202ece2SStefano Zampini 3743202ece2SStefano Zampini if (D) { 3753202ece2SStefano Zampini Mat Dd; 3763202ece2SStefano Zampini PetscBool Ddense; 3773202ece2SStefano Zampini 3783202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)D,MATSEQDENSE,&Ddense);CHKERRQ(ierr); 3793202ece2SStefano Zampini if (!Ddense) { 3803202ece2SStefano Zampini ierr = MatConvert(D, MATSEQDENSE, MAT_INITIAL_MATRIX, &Dd);CHKERRQ(ierr); 3813202ece2SStefano Zampini } else { 3823202ece2SStefano Zampini Dd = D; 3833202ece2SStefano Zampini } 384f11841e3SStefano Zampini if (n_I) { 3853202ece2SStefano Zampini ierr = MatAYPX(*S,-1.0,Dd,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 386f11841e3SStefano Zampini } else { 387f11841e3SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 388f11841e3SStefano Zampini ierr = MatDuplicate(Dd,MAT_COPY_VALUES,S);CHKERRQ(ierr); 389f11841e3SStefano Zampini } else { 390f11841e3SStefano Zampini ierr = MatCopy(Dd,*S,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 391f11841e3SStefano Zampini } 392f11841e3SStefano Zampini } 3933202ece2SStefano Zampini if (!Ddense) { 3943202ece2SStefano Zampini ierr = MatDestroy(&Dd);CHKERRQ(ierr); 3953202ece2SStefano Zampini } 3963202ece2SStefano Zampini } else { 3973202ece2SStefano Zampini ierr = MatScale(*S,-1.0);CHKERRQ(ierr); 3983202ece2SStefano Zampini } 3993202ece2SStefano Zampini PetscFunctionReturn(0); 4003202ece2SStefano Zampini } 40134a97f8cSStefano Zampini 40291af6908SStefano 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) 403b1b3d7a2SStefano Zampini { 404be83ff47SStefano Zampini Mat F,A_II,A_IB,A_BI,A_BB,AE_II; 405be83ff47SStefano Zampini Mat S_all; 40657a87bf3SStefano Zampini Vec gstash,lstash; 40757a87bf3SStefano Zampini VecScatter sstash; 408b7ab4a40SStefano Zampini IS is_I,is_I_layer; 409dc456d91SStefano Zampini IS all_subsets,all_subsets_mult,all_subsets_n; 41057a87bf3SStefano Zampini PetscScalar *stasharray,*Bwork; 411dc456d91SStefano Zampini PetscInt *nnz,*all_local_idx_N; 412dc456d91SStefano Zampini PetscInt *auxnum1,*auxnum2; 4135a95e1ceSStefano Zampini PetscInt i,subset_size,max_subset_size; 414683d3df6SStefano Zampini PetscInt n_B,extra,local_size,global_size; 41557a87bf3SStefano Zampini PetscInt local_stash_size; 41608122e43SStefano Zampini PetscBLASInt B_N,B_ierr,B_lwork,*pivots; 4175a95e1ceSStefano Zampini MPI_Comm comm_n; 418f4f7d9d6SStefano Zampini PetscBool deluxe = PETSC_TRUE; 419f4f7d9d6SStefano Zampini PetscBool use_potr = PETSC_FALSE, use_sytr = PETSC_FALSE; 4203b03f7bbSStefano Zampini PetscViewer matl_dbg_viewer = NULL; 421b1b3d7a2SStefano Zampini PetscErrorCode ierr; 42235d0533cSStefano Zampini PetscBool flg; 423b1b3d7a2SStefano Zampini 424b1b3d7a2SStefano Zampini PetscFunctionBegin; 425a64f4aa4SStefano Zampini ierr = MatDestroy(&sub_schurs->A);CHKERRQ(ierr); 426a64f4aa4SStefano Zampini ierr = MatDestroy(&sub_schurs->S);CHKERRQ(ierr); 427e62b6521Sstefano_zampini if (Ain) { 428a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)Ain);CHKERRQ(ierr); 429a64f4aa4SStefano Zampini sub_schurs->A = Ain; 430a64f4aa4SStefano Zampini } 4313301b35fSStefano Zampini 432a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)Sin);CHKERRQ(ierr); 433a64f4aa4SStefano Zampini sub_schurs->S = Sin; 434df4d28bfSStefano Zampini if (sub_schurs->schur_explicit) { 435df4d28bfSStefano Zampini sub_schurs->schur_explicit = (PetscBool)(!!sub_schurs->A); 436a64f4aa4SStefano Zampini } 437a64f4aa4SStefano Zampini 4385a95e1ceSStefano Zampini /* preliminary checks */ 439*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(!sub_schurs->schur_explicit && compute_Stilda,PetscObjectComm((PetscObject)sub_schurs->l2gmap),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_PARDISO"); 4405a95e1ceSStefano Zampini 44188113c35SStefano Zampini if (benign_trick) sub_schurs->is_posdef = PETSC_FALSE; 44288113c35SStefano Zampini 4433b03f7bbSStefano Zampini /* debug (MATLAB) */ 4447f9db97bSStefano Zampini if (sub_schurs->debug) { 4457f9db97bSStefano Zampini PetscMPIInt size,rank; 4467ebab0bbSStefano Zampini PetscInt nr,*print_schurs_ranks,print_schurs = PETSC_FALSE; 4477f9db97bSStefano Zampini 448ffc4695bSBarry Smith ierr = MPI_Comm_size(PetscObjectComm((PetscObject)sub_schurs->l2gmap),&size);CHKERRMPI(ierr); 449ffc4695bSBarry Smith ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)sub_schurs->l2gmap),&rank);CHKERRMPI(ierr); 4507f9db97bSStefano Zampini nr = size; 4517f9db97bSStefano Zampini ierr = PetscMalloc1(nr,&print_schurs_ranks);CHKERRQ(ierr); 4527f9db97bSStefano Zampini ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)sub_schurs->l2gmap),sub_schurs->prefix,"BDDC sub_schurs options","PC");CHKERRQ(ierr); 4537f9db97bSStefano Zampini ierr = PetscOptionsIntArray("-sub_schurs_debug_ranks","Ranks to debug (all if the option is not used)",NULL,print_schurs_ranks,&nr,&flg);CHKERRQ(ierr); 4547f9db97bSStefano Zampini if (!flg) print_schurs = PETSC_TRUE; 4557f9db97bSStefano Zampini else { 4567ebab0bbSStefano Zampini print_schurs = PETSC_FALSE; 4577f9db97bSStefano Zampini for (i=0;i<nr;i++) if (print_schurs_ranks[i] == (PetscInt)rank) { print_schurs = PETSC_TRUE; break; } 4587f9db97bSStefano Zampini } 4597f9db97bSStefano Zampini ierr = PetscOptionsEnd();CHKERRQ(ierr); 4607f9db97bSStefano Zampini ierr = PetscFree(print_schurs_ranks);CHKERRQ(ierr); 4613b03f7bbSStefano Zampini if (print_schurs) { 4623b03f7bbSStefano Zampini char filename[256]; 4633b03f7bbSStefano Zampini 4643b03f7bbSStefano Zampini ierr = PetscSNPrintf(filename,sizeof(filename),"sub_schurs_Schur_r%d.m",PetscGlobalRank);CHKERRQ(ierr); 4653b03f7bbSStefano Zampini ierr = PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&matl_dbg_viewer);CHKERRQ(ierr); 4663b03f7bbSStefano Zampini ierr = PetscViewerPushFormat(matl_dbg_viewer,PETSC_VIEWER_ASCII_MATLAB);CHKERRQ(ierr); 4673b03f7bbSStefano Zampini } 4687f9db97bSStefano Zampini } 4697f9db97bSStefano Zampini 4705a95e1ceSStefano Zampini /* restrict work on active processes */ 471991c41b4SStefano Zampini if (sub_schurs->restrict_comm) { 472991c41b4SStefano Zampini PetscSubcomm subcomm; 473991c41b4SStefano Zampini PetscMPIInt color,rank; 474991c41b4SStefano Zampini 4755a95e1ceSStefano Zampini color = 0; 4765a95e1ceSStefano Zampini if (!sub_schurs->n_subs) color = 1; /* this can happen if we are in a multilevel case or if the subdomain is disconnected */ 477ffc4695bSBarry Smith ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)sub_schurs->l2gmap),&rank);CHKERRMPI(ierr); 4785a95e1ceSStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)sub_schurs->l2gmap),&subcomm);CHKERRQ(ierr); 4795a95e1ceSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 4805a95e1ceSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4815a95e1ceSStefano Zampini ierr = PetscCommDuplicate(PetscSubcommChild(subcomm),&comm_n,NULL);CHKERRQ(ierr); 4825a95e1ceSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 4835a95e1ceSStefano Zampini if (!sub_schurs->n_subs) { 4845a95e1ceSStefano Zampini ierr = PetscCommDestroy(&comm_n);CHKERRQ(ierr); 4855a95e1ceSStefano Zampini PetscFunctionReturn(0); 4865a95e1ceSStefano Zampini } 487991c41b4SStefano Zampini } else { 488991c41b4SStefano Zampini ierr = PetscCommDuplicate(PetscObjectComm((PetscObject)sub_schurs->l2gmap),&comm_n,NULL);CHKERRQ(ierr); 489991c41b4SStefano Zampini } 4905a95e1ceSStefano Zampini 491b1b3d7a2SStefano Zampini /* get Schur complement matrices */ 492df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 493a64f4aa4SStefano Zampini Mat tA_IB,tA_BI,tA_BB; 4943301b35fSStefano Zampini PetscBool isseqsbaij; 495a64f4aa4SStefano Zampini ierr = MatSchurComplementGetSubMatrices(sub_schurs->S,&A_II,NULL,&tA_IB,&tA_BI,&tA_BB);CHKERRQ(ierr); 4963301b35fSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)tA_BB,MATSEQSBAIJ,&isseqsbaij);CHKERRQ(ierr); 4973301b35fSStefano Zampini if (isseqsbaij) { 498a64f4aa4SStefano Zampini ierr = MatConvert(tA_BB,MATSEQAIJ,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 499a64f4aa4SStefano Zampini ierr = MatConvert(tA_IB,MATSEQAIJ,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 500a64f4aa4SStefano Zampini ierr = MatConvert(tA_BI,MATSEQAIJ,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 501a64f4aa4SStefano Zampini } else { 502a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)tA_BB);CHKERRQ(ierr); 503a64f4aa4SStefano Zampini A_BB = tA_BB; 504a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)tA_IB);CHKERRQ(ierr); 505a64f4aa4SStefano Zampini A_IB = tA_IB; 506a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)tA_BI);CHKERRQ(ierr); 507a64f4aa4SStefano Zampini A_BI = tA_BI; 508f11841e3SStefano Zampini } 509a58a30b4SStefano Zampini } else { 5105a95e1ceSStefano Zampini A_II = NULL; 5115a95e1ceSStefano Zampini A_IB = NULL; 5125a95e1ceSStefano Zampini A_BI = NULL; 5135a95e1ceSStefano Zampini A_BB = NULL; 514b1b3d7a2SStefano Zampini } 5155a95e1ceSStefano Zampini S_all = NULL; 516b1b3d7a2SStefano Zampini 517b1b3d7a2SStefano Zampini /* determine interior problems */ 5183dc780c3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I,&i);CHKERRQ(ierr); 5193dc780c3SStefano Zampini if (nlayers >= 0 && i) { /* Interior problems can be different from the original one */ 520b1b3d7a2SStefano Zampini PetscBT touched; 521b1b3d7a2SStefano Zampini const PetscInt* idx_B; 522b1b3d7a2SStefano Zampini PetscInt n_I,n_B,n_local_dofs,n_prev_added,j,layer,*local_numbering; 523b1b3d7a2SStefano Zampini 524*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(!xadj,PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot request layering without adjacency"); 525b1b3d7a2SStefano Zampini /* get sizes */ 526b1b3d7a2SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I,&n_I);CHKERRQ(ierr); 527b1b3d7a2SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_B,&n_B);CHKERRQ(ierr); 528b1b3d7a2SStefano Zampini 529b1b3d7a2SStefano Zampini ierr = PetscMalloc1(n_I+n_B,&local_numbering);CHKERRQ(ierr); 530b1b3d7a2SStefano Zampini ierr = PetscBTCreate(n_I+n_B,&touched);CHKERRQ(ierr); 531b1b3d7a2SStefano Zampini ierr = PetscBTMemzero(n_I+n_B,touched);CHKERRQ(ierr); 532b1b3d7a2SStefano Zampini 533b1b3d7a2SStefano Zampini /* all boundary dofs must be skipped when adding layers */ 534b1b3d7a2SStefano Zampini ierr = ISGetIndices(sub_schurs->is_B,&idx_B);CHKERRQ(ierr); 535b1b3d7a2SStefano Zampini for (j=0;j<n_B;j++) { 536b1b3d7a2SStefano Zampini ierr = PetscBTSet(touched,idx_B[j]);CHKERRQ(ierr); 537b1b3d7a2SStefano Zampini } 538580bdb30SBarry Smith ierr = PetscArraycpy(local_numbering,idx_B,n_B);CHKERRQ(ierr); 539b1b3d7a2SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_B,&idx_B);CHKERRQ(ierr); 540b1b3d7a2SStefano Zampini 541b1b3d7a2SStefano Zampini /* add prescribed number of layers of dofs */ 542b1b3d7a2SStefano Zampini n_local_dofs = n_B; 543b1b3d7a2SStefano Zampini n_prev_added = n_B; 544b1b3d7a2SStefano Zampini for (layer=0;layer<nlayers;layer++) { 545b6bace71SJacob Faibussowitsch PetscInt n_added = 0; 546b1b3d7a2SStefano Zampini if (n_local_dofs == n_I+n_B) break; 547*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(n_local_dofs > n_I+n_B,PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error querying layer %D. Out of bound access (%D > %D)",layer,n_local_dofs,n_I+n_B); 548b1b3d7a2SStefano Zampini ierr = PCBDDCAdjGetNextLayer_Private(local_numbering+n_local_dofs,n_prev_added,touched,xadj,adjncy,&n_added);CHKERRQ(ierr); 549b1b3d7a2SStefano Zampini n_prev_added = n_added; 550b1b3d7a2SStefano Zampini n_local_dofs += n_added; 551b1b3d7a2SStefano Zampini if (!n_added) break; 552b1b3d7a2SStefano Zampini } 553b1b3d7a2SStefano Zampini ierr = PetscBTDestroy(&touched);CHKERRQ(ierr); 554b1b3d7a2SStefano Zampini 555883469d8SStefano Zampini /* IS for I layer dofs in original numbering */ 556a9b99552SStefano 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); 557b1b3d7a2SStefano Zampini ierr = PetscFree(local_numbering);CHKERRQ(ierr); 558a9b99552SStefano Zampini ierr = ISSort(is_I_layer);CHKERRQ(ierr); 559883469d8SStefano Zampini /* IS for I layer dofs in I numbering */ 560df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 561b1b3d7a2SStefano Zampini ISLocalToGlobalMapping ItoNmap; 562b1b3d7a2SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(sub_schurs->is_I,&ItoNmap);CHKERRQ(ierr); 563a9b99552SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(ItoNmap,IS_GTOLM_DROP,is_I_layer,&is_I);CHKERRQ(ierr); 564b1b3d7a2SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&ItoNmap);CHKERRQ(ierr); 565b1b3d7a2SStefano Zampini 566b1b3d7a2SStefano Zampini /* II block */ 5677dae84e0SHong Zhang ierr = MatCreateSubMatrix(A_II,is_I,is_I,MAT_INITIAL_MATRIX,&AE_II);CHKERRQ(ierr); 568b1b3d7a2SStefano Zampini } 569b1b3d7a2SStefano Zampini } else { 570b1b3d7a2SStefano Zampini PetscInt n_I; 571b1b3d7a2SStefano Zampini 572b1b3d7a2SStefano Zampini /* IS for I dofs in original numbering */ 573b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)sub_schurs->is_I);CHKERRQ(ierr); 574a9b99552SStefano Zampini is_I_layer = sub_schurs->is_I; 575b1b3d7a2SStefano Zampini 576b1b3d7a2SStefano Zampini /* IS for I dofs in I numbering (strided 1) */ 577df4d28bfSStefano Zampini if (!sub_schurs->schur_explicit) { 578b1b3d7a2SStefano Zampini ierr = ISGetSize(sub_schurs->is_I,&n_I);CHKERRQ(ierr); 579b1b3d7a2SStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)sub_schurs->is_I),n_I,0,1,&is_I);CHKERRQ(ierr); 580b1b3d7a2SStefano Zampini 581b1b3d7a2SStefano Zampini /* II block is the same */ 582b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)A_II);CHKERRQ(ierr); 583b1b3d7a2SStefano Zampini AE_II = A_II; 584b1b3d7a2SStefano Zampini } 585b1b3d7a2SStefano Zampini } 5865a95e1ceSStefano Zampini 587883469d8SStefano Zampini /* Get info on subset sizes and sum of all subsets sizes */ 5885a95e1ceSStefano Zampini max_subset_size = 0; 589883469d8SStefano Zampini local_size = 0; 5905a95e1ceSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 5915a95e1ceSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 5925a95e1ceSStefano Zampini max_subset_size = PetscMax(subset_size,max_subset_size); 593883469d8SStefano Zampini local_size += subset_size; 594883469d8SStefano Zampini } 595883469d8SStefano Zampini 596883469d8SStefano Zampini /* Work arrays for local indices */ 597883469d8SStefano Zampini extra = 0; 598683d3df6SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_B,&n_B);CHKERRQ(ierr); 599df4d28bfSStefano Zampini if (sub_schurs->schur_explicit && is_I_layer) { 600a9b99552SStefano Zampini ierr = ISGetLocalSize(is_I_layer,&extra);CHKERRQ(ierr); 601883469d8SStefano Zampini } 602683d3df6SStefano Zampini ierr = PetscMalloc1(n_B+extra,&all_local_idx_N);CHKERRQ(ierr); 603883469d8SStefano Zampini if (extra) { 604883469d8SStefano Zampini const PetscInt *idxs; 605a9b99552SStefano Zampini ierr = ISGetIndices(is_I_layer,&idxs);CHKERRQ(ierr); 606580bdb30SBarry Smith ierr = PetscArraycpy(all_local_idx_N,idxs,extra);CHKERRQ(ierr); 607a9b99552SStefano Zampini ierr = ISRestoreIndices(is_I_layer,&idxs);CHKERRQ(ierr); 608883469d8SStefano Zampini } 609dc456d91SStefano Zampini ierr = PetscMalloc1(sub_schurs->n_subs,&auxnum1);CHKERRQ(ierr); 610dc456d91SStefano Zampini ierr = PetscMalloc1(sub_schurs->n_subs,&auxnum2);CHKERRQ(ierr); 611883469d8SStefano Zampini 612883469d8SStefano Zampini /* Get local indices in local numbering */ 613883469d8SStefano Zampini local_size = 0; 61457a87bf3SStefano Zampini local_stash_size = 0; 6155a95e1ceSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 616883469d8SStefano Zampini const PetscInt *idxs; 617883469d8SStefano Zampini 6185a95e1ceSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 6195a95e1ceSStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 620eb595f79SStefano Zampini /* start (smallest in global ordering) and multiplicity */ 621eb595f79SStefano Zampini auxnum1[i] = idxs[0]; 62257a87bf3SStefano Zampini auxnum2[i] = subset_size*subset_size; 623883469d8SStefano Zampini /* subset indices in local numbering */ 624580bdb30SBarry Smith ierr = PetscArraycpy(all_local_idx_N+local_size+extra,idxs,subset_size);CHKERRQ(ierr); 6255a95e1ceSStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 626883469d8SStefano Zampini local_size += subset_size; 62757a87bf3SStefano Zampini local_stash_size += subset_size*subset_size; 628883469d8SStefano Zampini } 629883469d8SStefano Zampini 630f4f7d9d6SStefano Zampini /* allocate extra workspace needed only for GETRI or SYTRF */ 63111955456SStefano Zampini use_potr = use_sytr = PETSC_FALSE; 63211955456SStefano Zampini if (benign_trick || (sub_schurs->is_hermitian && sub_schurs->is_posdef)) { 633f4f7d9d6SStefano Zampini use_potr = PETSC_TRUE; 63411955456SStefano Zampini } else if (sub_schurs->is_symmetric) { 63511955456SStefano Zampini use_sytr = PETSC_TRUE; 63611955456SStefano Zampini } 63711955456SStefano Zampini if (local_size && !use_potr) { 63859ac4de7SStefano Zampini PetscScalar lwork,dummyscalar = 0.; 63959ac4de7SStefano Zampini PetscBLASInt dummyint = 0; 640d2627357SStefano Zampini 641d2627357SStefano Zampini B_lwork = -1; 642d2627357SStefano Zampini ierr = PetscBLASIntCast(local_size,&B_N);CHKERRQ(ierr); 643d2627357SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 644f4f7d9d6SStefano Zampini if (use_sytr) { 645f4f7d9d6SStefano Zampini PetscStackCallBLAS("LAPACKsytrf",LAPACKsytrf_("L",&B_N,&dummyscalar,&B_N,&dummyint,&lwork,&B_lwork,&B_ierr)); 646*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(B_ierr,PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYTRF Lapack routine %d",(int)B_ierr); 647f4f7d9d6SStefano Zampini } else { 64859ac4de7SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,&dummyscalar,&B_N,&dummyint,&lwork,&B_lwork,&B_ierr)); 649*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(B_ierr,PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GETRI Lapack routine %d",(int)B_ierr); 650f4f7d9d6SStefano Zampini } 651f4f7d9d6SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 652d2627357SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 653d2627357SStefano Zampini ierr = PetscMalloc2(B_lwork,&Bwork,B_N,&pivots);CHKERRQ(ierr); 654056290a2SStefano Zampini } else { 655056290a2SStefano Zampini Bwork = NULL; 656056290a2SStefano Zampini pivots = NULL; 657d2627357SStefano Zampini } 658d2627357SStefano Zampini 65957a87bf3SStefano Zampini /* prepare data for summing up properly schurs on subsets */ 660dc456d91SStefano Zampini ierr = ISCreateGeneral(comm_n,sub_schurs->n_subs,auxnum1,PETSC_OWN_POINTER,&all_subsets_n);CHKERRQ(ierr); 661dc456d91SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(sub_schurs->l2gmap,all_subsets_n,&all_subsets);CHKERRQ(ierr); 662dc456d91SStefano Zampini ierr = ISDestroy(&all_subsets_n);CHKERRQ(ierr); 663dc456d91SStefano Zampini ierr = ISCreateGeneral(comm_n,sub_schurs->n_subs,auxnum2,PETSC_OWN_POINTER,&all_subsets_mult);CHKERRQ(ierr); 6646583bcc1SStefano Zampini ierr = ISRenumber(all_subsets,all_subsets_mult,&global_size,&all_subsets_n);CHKERRQ(ierr); 665dc456d91SStefano Zampini ierr = ISDestroy(&all_subsets);CHKERRQ(ierr); 666dc456d91SStefano Zampini ierr = ISDestroy(&all_subsets_mult);CHKERRQ(ierr); 667dc456d91SStefano Zampini ierr = ISGetLocalSize(all_subsets_n,&i);CHKERRQ(ierr); 668*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(i != local_stash_size,PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid size of new subset! %D != %D",i,local_stash_size); 66957a87bf3SStefano Zampini ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,local_stash_size,NULL,&lstash);CHKERRQ(ierr); 67057a87bf3SStefano Zampini ierr = VecCreateMPI(comm_n,PETSC_DECIDE,global_size,&gstash);CHKERRQ(ierr); 67157a87bf3SStefano Zampini ierr = VecScatterCreate(lstash,NULL,gstash,all_subsets_n,&sstash);CHKERRQ(ierr); 67257a87bf3SStefano Zampini ierr = ISDestroy(&all_subsets_n);CHKERRQ(ierr); 6732972d61bSStefano Zampini 6745a95e1ceSStefano Zampini /* subset indices in local boundary numbering */ 6755a95e1ceSStefano Zampini if (!sub_schurs->is_Ej_all) { 6765a95e1ceSStefano Zampini PetscInt *all_local_idx_B; 6775a95e1ceSStefano Zampini 6785a95e1ceSStefano Zampini ierr = PetscMalloc1(local_size,&all_local_idx_B);CHKERRQ(ierr); 6795a95e1ceSStefano Zampini ierr = ISGlobalToLocalMappingApply(sub_schurs->BtoNmap,IS_GTOLM_DROP,local_size,all_local_idx_N+extra,&subset_size,all_local_idx_B);CHKERRQ(ierr); 680*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(subset_size != local_size,PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in sub_schurs serial (BtoNmap)! %D != %D",subset_size,local_size); 6815a95e1ceSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,local_size,all_local_idx_B,PETSC_OWN_POINTER,&sub_schurs->is_Ej_all);CHKERRQ(ierr); 682b1b3d7a2SStefano Zampini } 683b1b3d7a2SStefano Zampini 68472b8c272SStefano Zampini if (change) { 68572b8c272SStefano Zampini ISLocalToGlobalMapping BtoS; 68672b8c272SStefano Zampini IS change_primal_B; 68772b8c272SStefano Zampini IS change_primal_all; 68872b8c272SStefano Zampini 689*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(sub_schurs->change_primal_sub,PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 690*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(sub_schurs->change,PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 691b7ab4a40SStefano Zampini ierr = PetscMalloc1(sub_schurs->n_subs,&sub_schurs->change_primal_sub);CHKERRQ(ierr); 69272b8c272SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 69372b8c272SStefano Zampini ISLocalToGlobalMapping NtoS; 69472b8c272SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(sub_schurs->is_subs[i],&NtoS);CHKERRQ(ierr); 695b7ab4a40SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(NtoS,IS_GTOLM_DROP,change_primal,&sub_schurs->change_primal_sub[i]);CHKERRQ(ierr); 69672b8c272SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&NtoS);CHKERRQ(ierr); 69772b8c272SStefano Zampini } 69872b8c272SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(sub_schurs->BtoNmap,IS_GTOLM_DROP,change_primal,&change_primal_B);CHKERRQ(ierr); 69972b8c272SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(sub_schurs->is_Ej_all,&BtoS);CHKERRQ(ierr); 70072b8c272SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(BtoS,IS_GTOLM_DROP,change_primal_B,&change_primal_all);CHKERRQ(ierr); 70172b8c272SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&BtoS);CHKERRQ(ierr); 70272b8c272SStefano Zampini ierr = ISDestroy(&change_primal_B);CHKERRQ(ierr); 703b7ab4a40SStefano Zampini ierr = PetscMalloc1(sub_schurs->n_subs,&sub_schurs->change);CHKERRQ(ierr); 70472b8c272SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 70572b8c272SStefano Zampini Mat change_sub; 70672b8c272SStefano Zampini 707b7ab4a40SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 70872b8c272SStefano Zampini ierr = KSPCreate(PETSC_COMM_SELF,&sub_schurs->change[i]);CHKERRQ(ierr); 70972b8c272SStefano Zampini ierr = KSPSetType(sub_schurs->change[i],KSPPREONLY);CHKERRQ(ierr); 71072b8c272SStefano Zampini if (!sub_schurs->change_with_qr) { 7117dae84e0SHong Zhang ierr = MatCreateSubMatrix(change,sub_schurs->is_subs[i],sub_schurs->is_subs[i],MAT_INITIAL_MATRIX,&change_sub);CHKERRQ(ierr); 71272b8c272SStefano Zampini } else { 71372b8c272SStefano Zampini Mat change_subt; 7147dae84e0SHong Zhang ierr = MatCreateSubMatrix(change,sub_schurs->is_subs[i],sub_schurs->is_subs[i],MAT_INITIAL_MATRIX,&change_subt);CHKERRQ(ierr); 71572b8c272SStefano Zampini ierr = MatConvert(change_subt,MATSEQDENSE,MAT_INITIAL_MATRIX,&change_sub);CHKERRQ(ierr); 71672b8c272SStefano Zampini ierr = MatDestroy(&change_subt);CHKERRQ(ierr); 71772b8c272SStefano Zampini } 71872b8c272SStefano Zampini ierr = KSPSetOperators(sub_schurs->change[i],change_sub,change_sub);CHKERRQ(ierr); 71972b8c272SStefano Zampini ierr = MatDestroy(&change_sub);CHKERRQ(ierr); 720e62b6521Sstefano_zampini ierr = KSPSetOptionsPrefix(sub_schurs->change[i],sub_schurs->prefix);CHKERRQ(ierr); 721e62b6521Sstefano_zampini ierr = KSPAppendOptionsPrefix(sub_schurs->change[i],"sub_schurs_change_");CHKERRQ(ierr); 72272b8c272SStefano Zampini } 72372b8c272SStefano Zampini ierr = ISDestroy(&change_primal_all);CHKERRQ(ierr); 72472b8c272SStefano Zampini } 72572b8c272SStefano Zampini 7265a95e1ceSStefano Zampini /* Local matrix of all local Schur on subsets (transposed) */ 7275a95e1ceSStefano Zampini if (!sub_schurs->S_Ej_all) { 72804c5b2e6SStefano Zampini Mat T; 72904c5b2e6SStefano Zampini PetscScalar *v; 73004c5b2e6SStefano Zampini PetscInt *ii,*jj; 73104c5b2e6SStefano Zampini PetscInt cum,i,j,k; 73204c5b2e6SStefano Zampini 73304c5b2e6SStefano Zampini /* MatSeqAIJSetPreallocation + MatSetValues is slow for these kind of matrices (may have large blocks) 73404c5b2e6SStefano Zampini Allocate properly a representative matrix and duplicate */ 73504c5b2e6SStefano Zampini ierr = PetscMalloc3(local_size+1,&ii,local_stash_size,&jj,local_stash_size,&v);CHKERRQ(ierr); 73604c5b2e6SStefano Zampini ii[0] = 0; 73704c5b2e6SStefano Zampini cum = 0; 73804c5b2e6SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 73904c5b2e6SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 74004c5b2e6SStefano Zampini for (j=0;j<subset_size;j++) { 74104c5b2e6SStefano Zampini const PetscInt row = cum+j; 74204c5b2e6SStefano Zampini PetscInt col = cum; 74304c5b2e6SStefano Zampini 74404c5b2e6SStefano Zampini ii[row+1] = ii[row] + subset_size; 74504c5b2e6SStefano Zampini for (k=ii[row];k<ii[row+1];k++) { 74604c5b2e6SStefano Zampini jj[k] = col; 74704c5b2e6SStefano Zampini col++; 74804c5b2e6SStefano Zampini } 74904c5b2e6SStefano Zampini } 75004c5b2e6SStefano Zampini cum += subset_size; 75104c5b2e6SStefano Zampini } 75204c5b2e6SStefano Zampini ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,local_size,local_size,ii,jj,v,&T);CHKERRQ(ierr); 75304c5b2e6SStefano Zampini ierr = MatDuplicate(T,MAT_DO_NOT_COPY_VALUES,&sub_schurs->S_Ej_all);CHKERRQ(ierr); 75404c5b2e6SStefano Zampini ierr = MatDestroy(&T);CHKERRQ(ierr); 75504c5b2e6SStefano Zampini ierr = PetscFree3(ii,jj,v);CHKERRQ(ierr); 75604c5b2e6SStefano Zampini } 75704c5b2e6SStefano Zampini /* matrices for deluxe scaling and adaptive selection */ 75804c5b2e6SStefano Zampini if (compute_Stilda) { 75904c5b2e6SStefano Zampini if (!sub_schurs->sum_S_Ej_tilda_all) { 76004c5b2e6SStefano Zampini ierr = MatDuplicate(sub_schurs->S_Ej_all,MAT_DO_NOT_COPY_VALUES,&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 76104c5b2e6SStefano Zampini } 76204c5b2e6SStefano Zampini if (!sub_schurs->sum_S_Ej_inv_all && deluxe) { 76304c5b2e6SStefano Zampini ierr = MatDuplicate(sub_schurs->S_Ej_all,MAT_DO_NOT_COPY_VALUES,&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 76404c5b2e6SStefano Zampini } 765aa83b6aeSStefano Zampini } 766b1b3d7a2SStefano Zampini 7675a95e1ceSStefano Zampini /* Compute Schur complements explicitly */ 768be83ff47SStefano Zampini F = NULL; 769d943a642SStefano Zampini if (!sub_schurs->schur_explicit) { 770d943a642SStefano Zampini /* this code branch is used when MatFactor with Schur complement support is not present or when explicitly requested; 771d943a642SStefano Zampini it is not efficient, unless the economic version of the scaling is used */ 7725a95e1ceSStefano Zampini Mat S_Ej_expl; 7735a95e1ceSStefano Zampini PetscScalar *work; 7745a95e1ceSStefano Zampini PetscInt j,*dummy_idx; 7755a95e1ceSStefano Zampini PetscBool Sdense; 7765a95e1ceSStefano Zampini 7775a95e1ceSStefano Zampini ierr = PetscMalloc2(max_subset_size,&dummy_idx,max_subset_size*max_subset_size,&work);CHKERRQ(ierr); 7785a95e1ceSStefano Zampini local_size = 0; 779b1b3d7a2SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 7805a95e1ceSStefano Zampini IS is_subset_B; 7815a95e1ceSStefano Zampini Mat AE_EE,AE_IE,AE_EI,S_Ej; 7825a95e1ceSStefano Zampini 7835a95e1ceSStefano Zampini /* subsets in original and boundary numbering */ 7845a95e1ceSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(sub_schurs->BtoNmap,IS_GTOLM_DROP,sub_schurs->is_subs[i],&is_subset_B);CHKERRQ(ierr); 7855a95e1ceSStefano Zampini /* EE block */ 7867dae84e0SHong Zhang ierr = MatCreateSubMatrix(A_BB,is_subset_B,is_subset_B,MAT_INITIAL_MATRIX,&AE_EE);CHKERRQ(ierr); 7875a95e1ceSStefano Zampini /* IE block */ 7887dae84e0SHong Zhang ierr = MatCreateSubMatrix(A_IB,is_I,is_subset_B,MAT_INITIAL_MATRIX,&AE_IE);CHKERRQ(ierr); 7895a95e1ceSStefano Zampini /* EI block */ 790d943a642SStefano Zampini if (sub_schurs->is_symmetric) { 7915a95e1ceSStefano Zampini ierr = MatCreateTranspose(AE_IE,&AE_EI);CHKERRQ(ierr); 792d943a642SStefano Zampini } else if (sub_schurs->is_hermitian) { 793d943a642SStefano Zampini ierr = MatCreateHermitianTranspose(AE_IE,&AE_EI);CHKERRQ(ierr); 7945a95e1ceSStefano Zampini } else { 7957dae84e0SHong Zhang ierr = MatCreateSubMatrix(A_BI,is_subset_B,is_I,MAT_INITIAL_MATRIX,&AE_EI);CHKERRQ(ierr); 7965a95e1ceSStefano Zampini } 797a64f4aa4SStefano Zampini ierr = ISDestroy(&is_subset_B);CHKERRQ(ierr); 7985a95e1ceSStefano Zampini ierr = MatCreateSchurComplement(AE_II,AE_II,AE_IE,AE_EI,AE_EE,&S_Ej);CHKERRQ(ierr); 7995a95e1ceSStefano Zampini ierr = MatDestroy(&AE_EE);CHKERRQ(ierr); 8005a95e1ceSStefano Zampini ierr = MatDestroy(&AE_IE);CHKERRQ(ierr); 8015a95e1ceSStefano Zampini ierr = MatDestroy(&AE_EI);CHKERRQ(ierr); 802b1b3d7a2SStefano Zampini if (AE_II == A_II) { /* we can reuse the same ksp */ 803b1b3d7a2SStefano Zampini KSP ksp; 804b1b3d7a2SStefano Zampini ierr = MatSchurComplementGetKSP(sub_schurs->S,&ksp);CHKERRQ(ierr); 8055a95e1ceSStefano Zampini ierr = MatSchurComplementSetKSP(S_Ej,ksp);CHKERRQ(ierr); 806b1b3d7a2SStefano Zampini } else { /* build new ksp object which inherits ksp and pc types from the original one */ 807b1b3d7a2SStefano Zampini KSP origksp,schurksp; 808b1b3d7a2SStefano Zampini PC origpc,schurpc; 809b1b3d7a2SStefano Zampini KSPType ksp_type; 810b1b3d7a2SStefano Zampini PetscInt n_internal; 8115a95e1ceSStefano Zampini PetscBool ispcnone; 812b1b3d7a2SStefano Zampini 813b1b3d7a2SStefano Zampini ierr = MatSchurComplementGetKSP(sub_schurs->S,&origksp);CHKERRQ(ierr); 8145a95e1ceSStefano Zampini ierr = MatSchurComplementGetKSP(S_Ej,&schurksp);CHKERRQ(ierr); 815b1b3d7a2SStefano Zampini ierr = KSPGetType(origksp,&ksp_type);CHKERRQ(ierr); 816b1b3d7a2SStefano Zampini ierr = KSPSetType(schurksp,ksp_type);CHKERRQ(ierr); 817b1b3d7a2SStefano Zampini ierr = KSPGetPC(schurksp,&schurpc);CHKERRQ(ierr); 818b1b3d7a2SStefano Zampini ierr = KSPGetPC(origksp,&origpc);CHKERRQ(ierr); 8195a95e1ceSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)origpc,PCNONE,&ispcnone);CHKERRQ(ierr); 8205a95e1ceSStefano Zampini if (!ispcnone) { 8215a95e1ceSStefano Zampini PCType pc_type; 822b1b3d7a2SStefano Zampini ierr = PCGetType(origpc,&pc_type);CHKERRQ(ierr); 823b1b3d7a2SStefano Zampini ierr = PCSetType(schurpc,pc_type);CHKERRQ(ierr); 8245a95e1ceSStefano Zampini } else { 8255a95e1ceSStefano Zampini ierr = PCSetType(schurpc,PCLU);CHKERRQ(ierr); 8265a95e1ceSStefano Zampini } 827b1b3d7a2SStefano Zampini ierr = ISGetSize(is_I,&n_internal);CHKERRQ(ierr); 828365a3a41SStefano Zampini if (!n_internal) { /* UMFPACK gives error with 0 sized problems */ 8293ca39a21SBarry Smith MatSolverType solver = NULL; 830ea799195SBarry Smith ierr = PCFactorGetMatSolverType(origpc,(MatSolverType*)&solver);CHKERRQ(ierr); 831b1b3d7a2SStefano Zampini if (solver) { 8323ca39a21SBarry Smith ierr = PCFactorSetMatSolverType(schurpc,solver);CHKERRQ(ierr); 833b1b3d7a2SStefano Zampini } 834b1b3d7a2SStefano Zampini } 835b1b3d7a2SStefano Zampini ierr = KSPSetUp(schurksp);CHKERRQ(ierr); 836b1b3d7a2SStefano Zampini } 8375a95e1ceSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 8385a95e1ceSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,subset_size,work,&S_Ej_expl);CHKERRQ(ierr); 839d943a642SStefano Zampini ierr = PCBDDCComputeExplicitSchur(S_Ej,sub_schurs->is_symmetric,MAT_REUSE_MATRIX,&S_Ej_expl);CHKERRQ(ierr); 8405a95e1ceSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)S_Ej_expl,MATSEQDENSE,&Sdense);CHKERRQ(ierr); 8415a95e1ceSStefano Zampini if (Sdense) { 8425a95e1ceSStefano Zampini for (j=0;j<subset_size;j++) { 8435a95e1ceSStefano Zampini dummy_idx[j]=local_size+j; 844b1b3d7a2SStefano Zampini } 8455a95e1ceSStefano Zampini ierr = MatSetValues(sub_schurs->S_Ej_all,subset_size,dummy_idx,subset_size,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr); 8466c4ed002SBarry Smith } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented for sparse matrices"); 8475a95e1ceSStefano Zampini ierr = MatDestroy(&S_Ej);CHKERRQ(ierr); 848a64f4aa4SStefano Zampini ierr = MatDestroy(&S_Ej_expl);CHKERRQ(ierr); 8495a95e1ceSStefano Zampini local_size += subset_size; 8505a95e1ceSStefano Zampini } 8515a95e1ceSStefano Zampini ierr = PetscFree2(dummy_idx,work);CHKERRQ(ierr); 852b1b3d7a2SStefano Zampini /* free */ 853b1b3d7a2SStefano Zampini ierr = ISDestroy(&is_I);CHKERRQ(ierr); 854b1b3d7a2SStefano Zampini ierr = MatDestroy(&AE_II);CHKERRQ(ierr); 8555a95e1ceSStefano Zampini ierr = PetscFree(all_local_idx_N);CHKERRQ(ierr); 856883469d8SStefano Zampini } else { 8575cbda25cSStefano Zampini Mat A,cs_AIB_mat = NULL,benign_AIIm1_ones_mat = NULL; 8589d54b7f4SStefano Zampini Vec Dall = NULL; 859ca92afb2SStefano Zampini IS is_A_all,*is_p_r = NULL; 8607ebab0bbSStefano Zampini MatType Stype; 8615cbda25cSStefano Zampini PetscScalar *work,*S_data,*schur_factor,infty = PETSC_MAX_REAL; 86204c5b2e6SStefano Zampini PetscScalar *SEj_arr = NULL,*SEjinv_arr = NULL; 8631683a169SBarry Smith const PetscScalar *rS_data; 86404c5b2e6SStefano Zampini PetscInt n,n_I,size_schur,size_active_schur,cum,cum2; 8653fc34f97SStefano Zampini PetscBool economic,solver_S,S_lower_triangular = PETSC_FALSE; 8663fc34f97SStefano Zampini PetscBool schur_has_vertices,factor_workaround; 86711955456SStefano Zampini PetscBool use_cholesky; 8687ebab0bbSStefano Zampini #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 8697ebab0bbSStefano Zampini PetscBool oldpin; 8707ebab0bbSStefano Zampini #endif 871883469d8SStefano Zampini 872683d3df6SStefano Zampini /* get sizes */ 87381ea8064SStefano Zampini n_I = 0; 87481ea8064SStefano Zampini if (is_I_layer) { 875a9b99552SStefano Zampini ierr = ISGetLocalSize(is_I_layer,&n_I);CHKERRQ(ierr); 87681ea8064SStefano Zampini } 877683d3df6SStefano Zampini economic = PETSC_FALSE; 878683d3df6SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I,&cum);CHKERRQ(ierr); 879683d3df6SStefano Zampini if (cum != n_I) economic = PETSC_TRUE; 880683d3df6SStefano Zampini ierr = MatGetLocalSize(sub_schurs->A,&n,NULL);CHKERRQ(ierr); 8819d54b7f4SStefano Zampini size_active_schur = local_size; 8829d54b7f4SStefano Zampini 883f17d2ae1SStefano Zampini /* import scaling vector (wrong formulation if we have 3D edges) */ 8849d54b7f4SStefano Zampini if (scaling && compute_Stilda) { 8859d54b7f4SStefano Zampini const PetscScalar *array; 8869d54b7f4SStefano Zampini PetscScalar *array2; 8879d54b7f4SStefano Zampini const PetscInt *idxs; 8889d54b7f4SStefano Zampini PetscInt i; 8899d54b7f4SStefano Zampini 8909d54b7f4SStefano Zampini ierr = ISGetIndices(sub_schurs->is_Ej_all,&idxs);CHKERRQ(ierr); 8919d54b7f4SStefano Zampini ierr = VecCreateSeq(PETSC_COMM_SELF,size_active_schur,&Dall);CHKERRQ(ierr); 8929d54b7f4SStefano Zampini ierr = VecGetArrayRead(scaling,&array);CHKERRQ(ierr); 8939d54b7f4SStefano Zampini ierr = VecGetArray(Dall,&array2);CHKERRQ(ierr); 8949d54b7f4SStefano Zampini for (i=0;i<size_active_schur;i++) array2[i] = array[idxs[i]]; 8959d54b7f4SStefano Zampini ierr = VecRestoreArray(Dall,&array2);CHKERRQ(ierr); 8969d54b7f4SStefano Zampini ierr = VecRestoreArrayRead(scaling,&array);CHKERRQ(ierr); 8979d54b7f4SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_Ej_all,&idxs);CHKERRQ(ierr); 8989d54b7f4SStefano Zampini deluxe = PETSC_FALSE; 8999d54b7f4SStefano Zampini } 900d62866d3SStefano Zampini 901683d3df6SStefano Zampini /* size active schurs does not count any dirichlet or vertex dof on the interface */ 9023fc34f97SStefano Zampini factor_workaround = PETSC_FALSE; 9033fc34f97SStefano Zampini schur_has_vertices = PETSC_FALSE; 904683d3df6SStefano Zampini cum = n_I+size_active_schur; 905683d3df6SStefano Zampini if (sub_schurs->is_dir) { 906683d3df6SStefano Zampini const PetscInt* idxs; 907683d3df6SStefano Zampini PetscInt n_dir; 908683d3df6SStefano Zampini 909683d3df6SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_dir,&n_dir);CHKERRQ(ierr); 910683d3df6SStefano Zampini ierr = ISGetIndices(sub_schurs->is_dir,&idxs);CHKERRQ(ierr); 911580bdb30SBarry Smith ierr = PetscArraycpy(all_local_idx_N+cum,idxs,n_dir);CHKERRQ(ierr); 912683d3df6SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_dir,&idxs);CHKERRQ(ierr); 913683d3df6SStefano Zampini cum += n_dir; 9143fc34f97SStefano Zampini factor_workaround = PETSC_TRUE; 915d62866d3SStefano Zampini } 916683d3df6SStefano Zampini /* include the primal vertices in the Schur complement */ 917367aa537SStefano Zampini if (exact_schur && sub_schurs->is_vertices && (compute_Stilda || benign_n)) { 918683d3df6SStefano Zampini PetscInt n_v; 919683d3df6SStefano Zampini 920683d3df6SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_v);CHKERRQ(ierr); 921683d3df6SStefano Zampini if (n_v) { 922683d3df6SStefano Zampini const PetscInt* idxs; 923683d3df6SStefano Zampini 924683d3df6SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 925580bdb30SBarry Smith ierr = PetscArraycpy(all_local_idx_N+cum,idxs,n_v);CHKERRQ(ierr); 926683d3df6SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 927683d3df6SStefano Zampini cum += n_v; 928683d3df6SStefano Zampini factor_workaround = PETSC_TRUE; 9293fc34f97SStefano Zampini schur_has_vertices = PETSC_TRUE; 930683d3df6SStefano Zampini } 931683d3df6SStefano Zampini } 932683d3df6SStefano Zampini size_schur = cum - n_I; 933683d3df6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,cum,all_local_idx_N,PETSC_OWN_POINTER,&is_A_all);CHKERRQ(ierr); 9347ebab0bbSStefano Zampini #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 935b470e4b4SRichard Tran Mills oldpin = sub_schurs->A->boundtocpu; 936b470e4b4SRichard Tran Mills ierr = MatBindToCPU(sub_schurs->A,PETSC_TRUE);CHKERRQ(ierr); 9377ebab0bbSStefano Zampini #endif 938683d3df6SStefano Zampini if (cum == n) { 939683d3df6SStefano Zampini ierr = ISSetPermutation(is_A_all);CHKERRQ(ierr); 940683d3df6SStefano Zampini ierr = MatPermute(sub_schurs->A,is_A_all,is_A_all,&A);CHKERRQ(ierr); 941683d3df6SStefano Zampini } else { 9427dae84e0SHong Zhang ierr = MatCreateSubMatrix(sub_schurs->A,is_A_all,is_A_all,MAT_INITIAL_MATRIX,&A);CHKERRQ(ierr); 943683d3df6SStefano Zampini } 9447ebab0bbSStefano Zampini #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 945b470e4b4SRichard Tran Mills ierr = MatBindToCPU(sub_schurs->A,oldpin);CHKERRQ(ierr); 9467ebab0bbSStefano Zampini #endif 947e62b6521Sstefano_zampini ierr = MatSetOptionsPrefix(A,sub_schurs->prefix);CHKERRQ(ierr); 948e62b6521Sstefano_zampini ierr = MatAppendOptionsPrefix(A,"sub_schurs_");CHKERRQ(ierr); 949ca92afb2SStefano Zampini 950ca92afb2SStefano Zampini /* if we actually change the basis for the pressures, LDL^T factors will use a lot of memory 9517ebab0bbSStefano Zampini this is a workaround */ 952ca92afb2SStefano Zampini if (benign_n) { 9537ebab0bbSStefano Zampini Vec v,benign_AIIm1_ones; 954ca92afb2SStefano Zampini ISLocalToGlobalMapping N_to_reor; 955ca92afb2SStefano Zampini IS is_p0,is_p0_p; 9565cbda25cSStefano Zampini PetscScalar *cs_AIB,*AIIm1_data; 9575cbda25cSStefano Zampini PetscInt sizeA; 958ca92afb2SStefano Zampini 959ca92afb2SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_A_all,&N_to_reor);CHKERRQ(ierr); 960ca92afb2SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,benign_n,benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 961ca92afb2SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(N_to_reor,IS_GTOLM_DROP,is_p0,&is_p0_p);CHKERRQ(ierr); 962ca92afb2SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 9637ebab0bbSStefano Zampini ierr = MatCreateVecs(A,&v,&benign_AIIm1_ones);CHKERRQ(ierr); 9645cbda25cSStefano Zampini ierr = VecGetSize(v,&sizeA);CHKERRQ(ierr); 9655cbda25cSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,sizeA,benign_n,NULL,&benign_AIIm1_ones_mat);CHKERRQ(ierr); 966282d6408SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,size_schur,benign_n,NULL,&cs_AIB_mat);CHKERRQ(ierr); 9675cbda25cSStefano Zampini ierr = MatDenseGetArray(cs_AIB_mat,&cs_AIB);CHKERRQ(ierr); 9685cbda25cSStefano Zampini ierr = MatDenseGetArray(benign_AIIm1_ones_mat,&AIIm1_data);CHKERRQ(ierr); 969ca92afb2SStefano Zampini ierr = PetscMalloc1(benign_n,&is_p_r);CHKERRQ(ierr); 970ca92afb2SStefano Zampini /* compute colsum of A_IB restricted to pressures */ 971ca92afb2SStefano Zampini for (i=0;i<benign_n;i++) { 9727ebab0bbSStefano Zampini const PetscScalar *array; 973ca92afb2SStefano Zampini const PetscInt *idxs; 974ca92afb2SStefano Zampini PetscInt j,nz; 975ca92afb2SStefano Zampini 976ca92afb2SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(N_to_reor,IS_GTOLM_DROP,benign_zerodiag_subs[i],&is_p_r[i]);CHKERRQ(ierr); 977ca92afb2SStefano Zampini ierr = ISGetLocalSize(is_p_r[i],&nz);CHKERRQ(ierr); 978ca92afb2SStefano Zampini ierr = ISGetIndices(is_p_r[i],&idxs);CHKERRQ(ierr); 9795cbda25cSStefano Zampini for (j=0;j<nz;j++) AIIm1_data[idxs[j]+sizeA*i] = 1.; 980ca92afb2SStefano Zampini ierr = ISRestoreIndices(is_p_r[i],&idxs);CHKERRQ(ierr); 9817ebab0bbSStefano Zampini ierr = VecPlaceArray(benign_AIIm1_ones,AIIm1_data+sizeA*i);CHKERRQ(ierr); 9825cbda25cSStefano Zampini ierr = MatMult(A,benign_AIIm1_ones,v);CHKERRQ(ierr); 9837ebab0bbSStefano Zampini ierr = VecResetArray(benign_AIIm1_ones);CHKERRQ(ierr); 9847ebab0bbSStefano Zampini ierr = VecGetArrayRead(v,&array);CHKERRQ(ierr); 98522db5ddcSStefano Zampini for (j=0;j<size_schur;j++) { 98622db5ddcSStefano Zampini #if defined(PETSC_USE_COMPLEX) 98722db5ddcSStefano Zampini cs_AIB[i*size_schur + j] = (PetscRealPart(array[j+n_I])/nz + PETSC_i*(PetscImaginaryPart(array[j+n_I])/nz)); 98822db5ddcSStefano Zampini #else 98922db5ddcSStefano Zampini cs_AIB[i*size_schur + j] = array[j+n_I]/nz; 99022db5ddcSStefano Zampini #endif 99122db5ddcSStefano Zampini } 9927ebab0bbSStefano Zampini ierr = VecRestoreArrayRead(v,&array);CHKERRQ(ierr); 993ca92afb2SStefano Zampini } 9945cbda25cSStefano Zampini ierr = MatDenseRestoreArray(cs_AIB_mat,&cs_AIB);CHKERRQ(ierr); 9955cbda25cSStefano Zampini ierr = MatDenseRestoreArray(benign_AIIm1_ones_mat,&AIIm1_data);CHKERRQ(ierr); 996ca92afb2SStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 9977ebab0bbSStefano Zampini ierr = VecDestroy(&benign_AIIm1_ones);CHKERRQ(ierr); 998ca92afb2SStefano Zampini ierr = MatSetOption(A,MAT_KEEP_NONZERO_PATTERN,PETSC_FALSE);CHKERRQ(ierr); 999ec968b6dSstefano_zampini ierr = MatSetOption(A,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 1000ec968b6dSstefano_zampini ierr = MatSetOption(A,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 1001ca92afb2SStefano Zampini ierr = MatZeroRowsColumnsIS(A,is_p0_p,1.0,NULL,NULL);CHKERRQ(ierr); 1002ca92afb2SStefano Zampini ierr = ISDestroy(&is_p0_p);CHKERRQ(ierr); 1003ca92afb2SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&N_to_reor);CHKERRQ(ierr); 1004ca92afb2SStefano Zampini } 100511955456SStefano Zampini ierr = MatSetOption(A,MAT_SYMMETRIC,sub_schurs->is_symmetric);CHKERRQ(ierr); 10063301b35fSStefano Zampini ierr = MatSetOption(A,MAT_HERMITIAN,sub_schurs->is_hermitian);CHKERRQ(ierr); 10073301b35fSStefano Zampini ierr = MatSetOption(A,MAT_SPD,sub_schurs->is_posdef);CHKERRQ(ierr); 1008883469d8SStefano Zampini 100911955456SStefano Zampini /* for complexes, symmetric and hermitian at the same time implies null imaginary part */ 101011955456SStefano Zampini use_cholesky = (PetscBool)((use_potr || use_sytr) && sub_schurs->is_hermitian && sub_schurs->is_symmetric); 101111955456SStefano Zampini 1012683d3df6SStefano Zampini /* when using the benign subspace trick, the local Schur complements are SPD */ 101335d0533cSStefano Zampini /* MKL_PARDISO does not handle well the computation of a Schur complement from a symmetric indefinite factorization 101435d0533cSStefano Zampini Use LU and adapt pivoting perturbation (still, solution is not as accurate as with using MUMPS) */ 101535d0533cSStefano Zampini if (benign_trick) { 101635d0533cSStefano Zampini sub_schurs->is_posdef = PETSC_TRUE; 101735d0533cSStefano Zampini ierr = PetscStrcmp(sub_schurs->mat_solver_type,MATSOLVERMKL_PARDISO,&flg);CHKERRQ(ierr); 101835d0533cSStefano Zampini if (flg) use_cholesky = PETSC_FALSE; 101935d0533cSStefano Zampini } 1020d47842beSStefano Zampini 1021f4f7d9d6SStefano Zampini if (n_I) { 10220aa714b2SStefano Zampini IS is_schur; 10237ebab0bbSStefano Zampini char stype[64]; 10244ba54290SStefano Zampini PetscBool gpu = PETSC_FALSE; 10255a05ddb0SStefano Zampini 102611955456SStefano Zampini if (use_cholesky) { 102788113c35SStefano Zampini ierr = MatGetFactor(A,sub_schurs->mat_solver_type,MAT_FACTOR_CHOLESKY,&F);CHKERRQ(ierr); 1028883469d8SStefano Zampini } else { 102988113c35SStefano Zampini ierr = MatGetFactor(A,sub_schurs->mat_solver_type,MAT_FACTOR_LU,&F);CHKERRQ(ierr); 1030883469d8SStefano Zampini } 10318c09ecd8Sstefano_zampini ierr = MatSetErrorIfFailure(A,PETSC_TRUE);CHKERRQ(ierr); 103235d0533cSStefano Zampini #if defined(PETSC_HAVE_MKL_PARDISO) 103335d0533cSStefano Zampini if (benign_trick) { ierr = MatMkl_PardisoSetCntl(F,10,10);CHKERRQ(ierr); } 103435d0533cSStefano Zampini #endif 1035883469d8SStefano Zampini /* subsets ordered last */ 10365a05ddb0SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,size_schur,n_I,1,&is_schur);CHKERRQ(ierr); 10375a05ddb0SStefano Zampini ierr = MatFactorSetSchurIS(F,is_schur);CHKERRQ(ierr); 10385a05ddb0SStefano Zampini ierr = ISDestroy(&is_schur);CHKERRQ(ierr); 1039883469d8SStefano Zampini 1040883469d8SStefano Zampini /* factorization step */ 104111955456SStefano Zampini if (use_cholesky) { 1042883469d8SStefano Zampini ierr = MatCholeskyFactorSymbolic(F,A,NULL,NULL);CHKERRQ(ierr); 1043be83ff47SStefano Zampini #if defined(PETSC_HAVE_MUMPS) /* be sure that icntl 19 is not set by command line */ 1044be83ff47SStefano Zampini ierr = MatMumpsSetIcntl(F,19,2);CHKERRQ(ierr); 1045be83ff47SStefano Zampini #endif 1046883469d8SStefano Zampini ierr = MatCholeskyFactorNumeric(F,A,NULL);CHKERRQ(ierr); 1047a0b0af32SStefano Zampini S_lower_triangular = PETSC_TRUE; 1048883469d8SStefano Zampini } else { 1049883469d8SStefano Zampini ierr = MatLUFactorSymbolic(F,A,NULL,NULL,NULL);CHKERRQ(ierr); 1050be83ff47SStefano Zampini #if defined(PETSC_HAVE_MUMPS) /* be sure that icntl 19 is not set by command line */ 1051be83ff47SStefano Zampini ierr = MatMumpsSetIcntl(F,19,3);CHKERRQ(ierr); 1052be83ff47SStefano Zampini #endif 1053883469d8SStefano Zampini ierr = MatLUFactorNumeric(F,A,NULL);CHKERRQ(ierr); 1054883469d8SStefano Zampini } 1055ec968b6dSstefano_zampini ierr = MatViewFromOptions(F,(PetscObject)A,"-mat_factor_view");CHKERRQ(ierr); 1056883469d8SStefano Zampini 10573b03f7bbSStefano Zampini if (matl_dbg_viewer) { 105811955456SStefano Zampini Mat S; 105911955456SStefano Zampini IS is; 106011955456SStefano Zampini 106111955456SStefano Zampini ierr = PetscObjectSetName((PetscObject)A,"A");CHKERRQ(ierr); 10623b03f7bbSStefano Zampini ierr = MatView(A,matl_dbg_viewer);CHKERRQ(ierr); 106311955456SStefano Zampini ierr = MatFactorCreateSchurComplement(F,&S,NULL);CHKERRQ(ierr); 106411955456SStefano Zampini ierr = PetscObjectSetName((PetscObject)S,"S");CHKERRQ(ierr); 10653b03f7bbSStefano Zampini ierr = MatView(S,matl_dbg_viewer);CHKERRQ(ierr); 106611955456SStefano Zampini ierr = MatDestroy(&S);CHKERRQ(ierr); 106711955456SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,n_I,0,1,&is);CHKERRQ(ierr); 106811955456SStefano Zampini ierr = PetscObjectSetName((PetscObject)is,"I");CHKERRQ(ierr); 10693b03f7bbSStefano Zampini ierr = ISView(is,matl_dbg_viewer);CHKERRQ(ierr); 107011955456SStefano Zampini ierr = ISDestroy(&is);CHKERRQ(ierr); 107111955456SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,size_schur,n_I,1,&is);CHKERRQ(ierr); 107211955456SStefano Zampini ierr = PetscObjectSetName((PetscObject)is,"B");CHKERRQ(ierr); 10733b03f7bbSStefano Zampini ierr = ISView(is,matl_dbg_viewer);CHKERRQ(ierr); 107411955456SStefano Zampini ierr = ISDestroy(&is);CHKERRQ(ierr); 10753b03f7bbSStefano Zampini ierr = PetscObjectSetName((PetscObject)is_A_all,"IA");CHKERRQ(ierr); 10763b03f7bbSStefano Zampini ierr = ISView(is_A_all,matl_dbg_viewer);CHKERRQ(ierr); 107711955456SStefano Zampini } 107811955456SStefano Zampini 1079883469d8SStefano Zampini /* get explicit Schur Complement computed during numeric factorization */ 10807c2f51b8SStefano Zampini ierr = MatFactorGetSchurComplement(F,&S_all,NULL);CHKERRQ(ierr); 10817ebab0bbSStefano Zampini ierr = PetscStrncpy(stype,MATSEQDENSE,sizeof(stype));CHKERRQ(ierr); 10824ba54290SStefano Zampini #if defined(PETSC_HAVE_CUDA) 10837ebab0bbSStefano Zampini ierr = PetscObjectTypeCompareAny((PetscObject)A,&gpu,MATSEQAIJVIENNACL,MATSEQAIJCUSPARSE,"");CHKERRQ(ierr); 10844ba54290SStefano Zampini #endif 10857ebab0bbSStefano Zampini if (gpu) { 10867ebab0bbSStefano Zampini ierr = PetscStrncpy(stype,MATSEQDENSECUDA,sizeof(stype));CHKERRQ(ierr); 10877ebab0bbSStefano Zampini } 10887ebab0bbSStefano Zampini ierr = PetscOptionsGetString(NULL,sub_schurs->prefix,"-sub_schurs_schur_mat_type",stype,sizeof(stype),NULL);CHKERRQ(ierr); 10897ebab0bbSStefano Zampini ierr = MatConvert(S_all,stype,MAT_INPLACE_MATRIX,&S_all);CHKERRQ(ierr); 1090b3cb21ddSStefano Zampini ierr = MatSetOption(S_all,MAT_SPD,sub_schurs->is_posdef);CHKERRQ(ierr); 1091b3cb21ddSStefano Zampini ierr = MatSetOption(S_all,MAT_HERMITIAN,sub_schurs->is_hermitian);CHKERRQ(ierr); 10927ebab0bbSStefano Zampini ierr = MatGetType(S_all,&Stype);CHKERRQ(ierr); 1093b3cb21ddSStefano Zampini 1094d62866d3SStefano Zampini /* we can reuse the solvers if we are not using the economic version */ 1095683d3df6SStefano Zampini reuse_solvers = (PetscBool)(reuse_solvers && !economic); 1096683d3df6SStefano Zampini factor_workaround = (PetscBool)(reuse_solvers && factor_workaround); 109703dfb2d7SStefano Zampini if (!sub_schurs->is_posdef && factor_workaround && compute_Stilda && size_active_schur) 109803dfb2d7SStefano Zampini reuse_solvers = factor_workaround = PETSC_FALSE; 109903dfb2d7SStefano Zampini 1100df4d28bfSStefano Zampini solver_S = PETSC_TRUE; 1101ca92afb2SStefano Zampini 110272b8c272SStefano Zampini /* update the Schur complement with the change of basis on the pressures */ 1103ca92afb2SStefano Zampini if (benign_n) { 11047ebab0bbSStefano Zampini const PetscScalar *cs_AIB; 11057ebab0bbSStefano Zampini PetscScalar *S_data,*AIIm1_data; 11063b03f7bbSStefano Zampini Mat S2 = NULL,S3 = NULL; /* dbg */ 11073b03f7bbSStefano Zampini PetscScalar *S2_data,*S3_data; /* dbg */ 11087ebab0bbSStefano Zampini Vec v,benign_AIIm1_ones; 11095cbda25cSStefano Zampini PetscInt sizeA; 1110ca92afb2SStefano Zampini 1111ca92afb2SStefano Zampini ierr = MatDenseGetArray(S_all,&S_data);CHKERRQ(ierr); 11127ebab0bbSStefano Zampini ierr = MatCreateVecs(A,&v,&benign_AIIm1_ones);CHKERRQ(ierr); 11135cbda25cSStefano Zampini ierr = VecGetSize(v,&sizeA);CHKERRQ(ierr); 1114ca92afb2SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 1115ca92afb2SStefano Zampini ierr = MatMumpsSetIcntl(F,26,0);CHKERRQ(ierr); 1116ca92afb2SStefano Zampini #endif 1117ca92afb2SStefano Zampini #if defined(PETSC_HAVE_MKL_PARDISO) 1118ca92afb2SStefano Zampini ierr = MatMkl_PardisoSetCntl(F,70,1);CHKERRQ(ierr); 1119ca92afb2SStefano Zampini #endif 11207ebab0bbSStefano Zampini ierr = MatDenseGetArrayRead(cs_AIB_mat,&cs_AIB);CHKERRQ(ierr); 11215cbda25cSStefano Zampini ierr = MatDenseGetArray(benign_AIIm1_ones_mat,&AIIm1_data);CHKERRQ(ierr); 11223b03f7bbSStefano Zampini if (matl_dbg_viewer) { 11233b03f7bbSStefano Zampini ierr = MatDuplicate(S_all,MAT_DO_NOT_COPY_VALUES,&S2);CHKERRQ(ierr); 11243b03f7bbSStefano Zampini ierr = MatDuplicate(S_all,MAT_DO_NOT_COPY_VALUES,&S3);CHKERRQ(ierr); 11253b03f7bbSStefano Zampini ierr = MatDenseGetArray(S2,&S2_data);CHKERRQ(ierr); 11263b03f7bbSStefano Zampini ierr = MatDenseGetArray(S3,&S3_data);CHKERRQ(ierr); 11273b03f7bbSStefano Zampini } 1128ca92afb2SStefano Zampini for (i=0;i<benign_n;i++) { 11293b03f7bbSStefano Zampini PetscScalar *array,sum = 0.,one = 1.,*sums; 1130ca92afb2SStefano Zampini const PetscInt *idxs; 11313b03f7bbSStefano Zampini PetscInt k,j,nz; 113247484b83SStefano Zampini PetscBLASInt B_k,B_n; 1133ca92afb2SStefano Zampini 11343b03f7bbSStefano Zampini ierr = PetscCalloc1(benign_n,&sums);CHKERRQ(ierr); 11357ebab0bbSStefano Zampini ierr = VecPlaceArray(benign_AIIm1_ones,AIIm1_data+sizeA*i);CHKERRQ(ierr); 11365cbda25cSStefano Zampini ierr = VecCopy(benign_AIIm1_ones,v);CHKERRQ(ierr); 11375cbda25cSStefano Zampini ierr = MatSolve(F,v,benign_AIIm1_ones);CHKERRQ(ierr); 11387ebab0bbSStefano Zampini ierr = MatMult(A,benign_AIIm1_ones,v);CHKERRQ(ierr); 11397ebab0bbSStefano Zampini ierr = VecResetArray(benign_AIIm1_ones);CHKERRQ(ierr); 11403b03f7bbSStefano Zampini /* p0 dofs (eliminated) are excluded from the sums */ 11413b03f7bbSStefano Zampini for (k=0;k<benign_n;k++) { 11423b03f7bbSStefano Zampini ierr = ISGetLocalSize(is_p_r[k],&nz);CHKERRQ(ierr); 11433b03f7bbSStefano Zampini ierr = ISGetIndices(is_p_r[k],&idxs);CHKERRQ(ierr); 11443b03f7bbSStefano Zampini for (j=0;j<nz-1;j++) sums[k] -= AIIm1_data[idxs[j]+sizeA*i]; 11453b03f7bbSStefano Zampini ierr = ISRestoreIndices(is_p_r[k],&idxs);CHKERRQ(ierr); 11463b03f7bbSStefano Zampini } 11477ebab0bbSStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 11483b03f7bbSStefano Zampini if (matl_dbg_viewer) { 11493b03f7bbSStefano Zampini Vec vv; 11503b03f7bbSStefano Zampini char name[16]; 11513b03f7bbSStefano Zampini 11523b03f7bbSStefano Zampini ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,size_schur,array+n_I,&vv);CHKERRQ(ierr); 11533b03f7bbSStefano Zampini ierr = PetscSNPrintf(name,sizeof(name),"Pvs%D",i);CHKERRQ(ierr); 11543b03f7bbSStefano Zampini ierr = PetscObjectSetName((PetscObject)vv,name);CHKERRQ(ierr); 11553b03f7bbSStefano Zampini ierr = VecView(vv,matl_dbg_viewer);CHKERRQ(ierr); 11563b03f7bbSStefano Zampini } 115747484b83SStefano Zampini /* perform sparse rank updates on symmetric Schur (TODO: move outside of the loop?) */ 115847484b83SStefano Zampini /* cs_AIB already scaled by 1./nz */ 115947484b83SStefano Zampini B_k = 1; 11603b03f7bbSStefano Zampini for (k=0;k<benign_n;k++) { 11613b03f7bbSStefano Zampini sum = sums[k]; 116247484b83SStefano Zampini ierr = PetscBLASIntCast(size_schur,&B_n);CHKERRQ(ierr); 11633b03f7bbSStefano Zampini 11643b03f7bbSStefano Zampini if (PetscAbsScalar(sum) == 0.0) continue; 11653b03f7bbSStefano Zampini if (k == i) { 116647484b83SStefano Zampini PetscStackCallBLAS("BLASsyrk",BLASsyrk_("L","N",&B_n,&B_k,&sum,cs_AIB+i*size_schur,&B_n,&one,S_data,&B_n)); 11673b03f7bbSStefano Zampini if (matl_dbg_viewer) { 11683b03f7bbSStefano Zampini PetscStackCallBLAS("BLASsyrk",BLASsyrk_("L","N",&B_n,&B_k,&sum,cs_AIB+i*size_schur,&B_n,&one,S3_data,&B_n)); 11693b03f7bbSStefano Zampini } 11703b03f7bbSStefano Zampini } else { /* XXX Is it correct to use symmetric rank-2 update with half of the sum? */ 11713b03f7bbSStefano Zampini sum /= 2.0; 11723b03f7bbSStefano Zampini PetscStackCallBLAS("BLASsyr2k",BLASsyr2k_("L","N",&B_n,&B_k,&sum,cs_AIB+k*size_schur,&B_n,cs_AIB+i*size_schur,&B_n,&one,S_data,&B_n)); 11733b03f7bbSStefano Zampini if (matl_dbg_viewer) { 11743b03f7bbSStefano Zampini PetscStackCallBLAS("BLASsyr2k",BLASsyr2k_("L","N",&B_n,&B_k,&sum,cs_AIB+k*size_schur,&B_n,cs_AIB+i*size_schur,&B_n,&one,S3_data,&B_n)); 11753b03f7bbSStefano Zampini } 11763b03f7bbSStefano Zampini } 11773b03f7bbSStefano Zampini } 11785cbda25cSStefano Zampini sum = 1.; 117947484b83SStefano 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)); 11803b03f7bbSStefano Zampini if (matl_dbg_viewer) { 11813b03f7bbSStefano Zampini PetscStackCallBLAS("BLASsyr2k",BLASsyr2k_("L","N",&B_n,&B_k,&sum,array+n_I,&B_n,cs_AIB+i*size_schur,&B_n,&one,S2_data,&B_n)); 11823b03f7bbSStefano Zampini } 11837ebab0bbSStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 11845cbda25cSStefano Zampini /* set p0 entry of AIIm1_ones to zero */ 11853b03f7bbSStefano Zampini ierr = ISGetLocalSize(is_p_r[i],&nz);CHKERRQ(ierr); 11865cbda25cSStefano Zampini ierr = ISGetIndices(is_p_r[i],&idxs);CHKERRQ(ierr); 1187282d6408SStefano Zampini for (j=0;j<benign_n;j++) AIIm1_data[idxs[nz-1]+sizeA*j] = 0.; 11885cbda25cSStefano Zampini ierr = ISRestoreIndices(is_p_r[i],&idxs);CHKERRQ(ierr); 11893b03f7bbSStefano Zampini ierr = PetscFree(sums);CHKERRQ(ierr); 11903b03f7bbSStefano Zampini } 11917ebab0bbSStefano Zampini ierr = VecDestroy(&benign_AIIm1_ones);CHKERRQ(ierr); 11923b03f7bbSStefano Zampini if (matl_dbg_viewer) { 11933b03f7bbSStefano Zampini ierr = MatDenseRestoreArray(S2,&S2_data);CHKERRQ(ierr); 11943b03f7bbSStefano Zampini ierr = MatDenseRestoreArray(S3,&S3_data);CHKERRQ(ierr); 1195ca92afb2SStefano Zampini } 1196a7414863SStefano Zampini if (!S_lower_triangular) { /* I need to expand the upper triangular data (column oriented) */ 1197a7414863SStefano Zampini PetscInt k,j; 1198a7414863SStefano Zampini for (k=0;k<size_schur;k++) { 1199a7414863SStefano Zampini for (j=k;j<size_schur;j++) { 1200a7414863SStefano Zampini S_data[j*size_schur+k] = PetscConj(S_data[k*size_schur+j]); 1201a7414863SStefano Zampini } 1202a7414863SStefano Zampini } 1203a7414863SStefano Zampini } 1204a7414863SStefano Zampini 12055cbda25cSStefano Zampini /* restore defaults */ 12065cbda25cSStefano Zampini #if defined(PETSC_HAVE_MUMPS) 12075cbda25cSStefano Zampini ierr = MatMumpsSetIcntl(F,26,-1);CHKERRQ(ierr); 12085cbda25cSStefano Zampini #endif 12095cbda25cSStefano Zampini #if defined(PETSC_HAVE_MKL_PARDISO) 12105cbda25cSStefano Zampini ierr = MatMkl_PardisoSetCntl(F,70,0);CHKERRQ(ierr); 12115cbda25cSStefano Zampini #endif 12127ebab0bbSStefano Zampini ierr = MatDenseRestoreArrayRead(cs_AIB_mat,&cs_AIB);CHKERRQ(ierr); 12135cbda25cSStefano Zampini ierr = MatDenseRestoreArray(benign_AIIm1_ones_mat,&AIIm1_data);CHKERRQ(ierr); 1214ca92afb2SStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 1215ca92afb2SStefano Zampini ierr = MatDenseRestoreArray(S_all,&S_data);CHKERRQ(ierr); 12163b03f7bbSStefano Zampini if (matl_dbg_viewer) { 12173b03f7bbSStefano Zampini Mat S; 12183b03f7bbSStefano Zampini 12193b03f7bbSStefano Zampini ierr = MatFactorRestoreSchurComplement(F,&S_all,MAT_FACTOR_SCHUR_UNFACTORED);CHKERRQ(ierr); 12203b03f7bbSStefano Zampini ierr = MatFactorCreateSchurComplement(F,&S,NULL);CHKERRQ(ierr); 12213b03f7bbSStefano Zampini ierr = PetscObjectSetName((PetscObject)S,"Sb");CHKERRQ(ierr); 12223b03f7bbSStefano Zampini ierr = MatView(S,matl_dbg_viewer);CHKERRQ(ierr); 12233b03f7bbSStefano Zampini ierr = MatDestroy(&S);CHKERRQ(ierr); 12243b03f7bbSStefano Zampini ierr = PetscObjectSetName((PetscObject)S2,"S2P");CHKERRQ(ierr); 12253b03f7bbSStefano Zampini ierr = MatView(S2,matl_dbg_viewer);CHKERRQ(ierr); 12263b03f7bbSStefano Zampini ierr = PetscObjectSetName((PetscObject)S3,"S3P");CHKERRQ(ierr); 12273b03f7bbSStefano Zampini ierr = MatView(S3,matl_dbg_viewer);CHKERRQ(ierr); 12283b03f7bbSStefano Zampini ierr = PetscObjectSetName((PetscObject)cs_AIB_mat,"cs");CHKERRQ(ierr); 12293b03f7bbSStefano Zampini ierr = MatView(cs_AIB_mat,matl_dbg_viewer);CHKERRQ(ierr); 12303b03f7bbSStefano Zampini ierr = MatFactorGetSchurComplement(F,&S_all,NULL);CHKERRQ(ierr); 12313b03f7bbSStefano Zampini } 12323b03f7bbSStefano Zampini ierr = MatDestroy(&S2);CHKERRQ(ierr); 12333b03f7bbSStefano Zampini ierr = MatDestroy(&S3);CHKERRQ(ierr); 1234ca92afb2SStefano Zampini } 1235a3df083aSStefano Zampini if (!reuse_solvers) { 1236a3df083aSStefano Zampini for (i=0;i<benign_n;i++) { 1237a3df083aSStefano Zampini ierr = ISDestroy(&is_p_r[i]);CHKERRQ(ierr); 1238a3df083aSStefano Zampini } 1239a3df083aSStefano Zampini ierr = PetscFree(is_p_r);CHKERRQ(ierr); 12405cbda25cSStefano Zampini ierr = MatDestroy(&cs_AIB_mat);CHKERRQ(ierr); 12415cbda25cSStefano Zampini ierr = MatDestroy(&benign_AIIm1_ones_mat);CHKERRQ(ierr); 1242a3df083aSStefano Zampini } 1243df4d28bfSStefano Zampini } else { /* we can't use MatFactor when size_schur == size_of_the_problem */ 1244be83ff47SStefano Zampini ierr = MatConvert(A,MATSEQDENSE,MAT_INITIAL_MATRIX,&S_all);CHKERRQ(ierr); 12457ebab0bbSStefano Zampini ierr = MatGetType(S_all,&Stype);CHKERRQ(ierr); 1246683d3df6SStefano Zampini reuse_solvers = PETSC_FALSE; /* TODO: why we can't reuse the solvers here? */ 1247166598c1SStefano Zampini factor_workaround = PETSC_FALSE; 1248df4d28bfSStefano Zampini solver_S = PETSC_FALSE; 1249be83ff47SStefano Zampini } 1250be83ff47SStefano Zampini 1251be83ff47SStefano Zampini if (reuse_solvers) { 1252a00504b5SStefano Zampini Mat A_II,Afake; 125353892102SStefano Zampini Vec vec1_B; 1254df4d28bfSStefano Zampini PCBDDCReuseSolvers msolv_ctx; 12553462e049SStefano Zampini PetscInt n_R; 1256d5574798SStefano Zampini 1257df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 1258df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversReset(sub_schurs->reuse_solver);CHKERRQ(ierr); 1259e28d306cSStefano Zampini } else { 1260df4d28bfSStefano Zampini ierr = PetscNew(&sub_schurs->reuse_solver);CHKERRQ(ierr); 1261d62866d3SStefano Zampini } 1262df4d28bfSStefano Zampini msolv_ctx = sub_schurs->reuse_solver; 1263be83ff47SStefano Zampini ierr = MatSchurComplementGetSubMatrices(sub_schurs->S,&A_II,NULL,NULL,NULL,NULL);CHKERRQ(ierr); 1264d5574798SStefano Zampini ierr = PetscObjectReference((PetscObject)F);CHKERRQ(ierr); 1265d5574798SStefano Zampini msolv_ctx->F = F; 1266683d3df6SStefano Zampini ierr = MatCreateVecs(F,&msolv_ctx->sol,NULL);CHKERRQ(ierr); 1267683d3df6SStefano Zampini /* currently PETSc has no support for MatSolve(F,x,x), so cheat and let rhs and sol share the same memory */ 1268683d3df6SStefano Zampini { 1269683d3df6SStefano Zampini PetscScalar *array; 1270683d3df6SStefano Zampini PetscInt n; 1271683d3df6SStefano Zampini 1272683d3df6SStefano Zampini ierr = VecGetLocalSize(msolv_ctx->sol,&n);CHKERRQ(ierr); 1273683d3df6SStefano Zampini ierr = VecGetArray(msolv_ctx->sol,&array);CHKERRQ(ierr); 1274683d3df6SStefano Zampini ierr = VecCreateSeqWithArray(PetscObjectComm((PetscObject)msolv_ctx->sol),1,n,array,&msolv_ctx->rhs);CHKERRQ(ierr); 1275683d3df6SStefano Zampini ierr = VecRestoreArray(msolv_ctx->sol,&array);CHKERRQ(ierr); 1276683d3df6SStefano Zampini } 12773fc34f97SStefano Zampini msolv_ctx->has_vertices = schur_has_vertices; 1278d62866d3SStefano Zampini 1279d62866d3SStefano Zampini /* interior solver */ 1280683d3df6SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)A_II),&msolv_ctx->interior_solver);CHKERRQ(ierr); 1281d62866d3SStefano Zampini ierr = PCSetOperators(msolv_ctx->interior_solver,A_II,A_II);CHKERRQ(ierr); 1282d62866d3SStefano Zampini ierr = PCSetType(msolv_ctx->interior_solver,PCSHELL);CHKERRQ(ierr); 128315579a77SStefano Zampini ierr = PCShellSetName(msolv_ctx->interior_solver,"Interior solver (w/o Schur factorization)");CHKERRQ(ierr); 1284d62866d3SStefano Zampini ierr = PCShellSetContext(msolv_ctx->interior_solver,msolv_ctx);CHKERRQ(ierr); 128515579a77SStefano Zampini ierr = PCShellSetView(msolv_ctx->interior_solver,PCBDDCReuseSolvers_View);CHKERRQ(ierr); 1286df4d28bfSStefano Zampini ierr = PCShellSetApply(msolv_ctx->interior_solver,PCBDDCReuseSolvers_Interior);CHKERRQ(ierr); 1287df4d28bfSStefano Zampini ierr = PCShellSetApplyTranspose(msolv_ctx->interior_solver,PCBDDCReuseSolvers_InteriorTranspose);CHKERRQ(ierr); 1288d62866d3SStefano Zampini 1289d62866d3SStefano Zampini /* correction solver */ 12903462e049SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)A_II),&msolv_ctx->correction_solver);CHKERRQ(ierr); 1291d62866d3SStefano Zampini ierr = PCSetType(msolv_ctx->correction_solver,PCSHELL);CHKERRQ(ierr); 129215579a77SStefano Zampini ierr = PCShellSetName(msolv_ctx->correction_solver,"Correction solver (with Schur factorization)");CHKERRQ(ierr); 1293d62866d3SStefano Zampini ierr = PCShellSetContext(msolv_ctx->correction_solver,msolv_ctx);CHKERRQ(ierr); 129415579a77SStefano Zampini ierr = PCShellSetView(msolv_ctx->interior_solver,PCBDDCReuseSolvers_View);CHKERRQ(ierr); 1295df4d28bfSStefano Zampini ierr = PCShellSetApply(msolv_ctx->correction_solver,PCBDDCReuseSolvers_Correction);CHKERRQ(ierr); 1296df4d28bfSStefano Zampini ierr = PCShellSetApplyTranspose(msolv_ctx->correction_solver,PCBDDCReuseSolvers_CorrectionTranspose);CHKERRQ(ierr); 129753892102SStefano Zampini 129853892102SStefano Zampini /* scatter and vecs for Schur complement solver */ 129953892102SStefano Zampini ierr = MatCreateVecs(S_all,&msolv_ctx->sol_B,&msolv_ctx->rhs_B);CHKERRQ(ierr); 130053892102SStefano Zampini ierr = MatCreateVecs(sub_schurs->S,&vec1_B,NULL);CHKERRQ(ierr); 13013fc34f97SStefano Zampini if (!schur_has_vertices) { 130253892102SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(sub_schurs->BtoNmap,IS_GTOLM_DROP,is_A_all,&msolv_ctx->is_B);CHKERRQ(ierr); 13039448b7f1SJunchao Zhang ierr = VecScatterCreate(vec1_B,msolv_ctx->is_B,msolv_ctx->sol_B,NULL,&msolv_ctx->correction_scatter_B);CHKERRQ(ierr); 130453892102SStefano Zampini ierr = PetscObjectReference((PetscObject)is_A_all);CHKERRQ(ierr); 130553892102SStefano Zampini msolv_ctx->is_R = is_A_all; 1306683d3df6SStefano Zampini } else { 1307683d3df6SStefano Zampini IS is_B_all; 1308683d3df6SStefano Zampini const PetscInt* idxs; 1309683d3df6SStefano Zampini PetscInt dual,n_v,n; 1310683d3df6SStefano Zampini 1311683d3df6SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_v);CHKERRQ(ierr); 1312683d3df6SStefano Zampini dual = size_schur - n_v; 1313683d3df6SStefano Zampini ierr = ISGetLocalSize(is_A_all,&n);CHKERRQ(ierr); 1314683d3df6SStefano Zampini ierr = ISGetIndices(is_A_all,&idxs);CHKERRQ(ierr); 1315683d3df6SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)is_A_all),dual,idxs+n_I,PETSC_COPY_VALUES,&is_B_all);CHKERRQ(ierr); 1316683d3df6SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(sub_schurs->BtoNmap,IS_GTOLM_DROP,is_B_all,&msolv_ctx->is_B);CHKERRQ(ierr); 1317683d3df6SStefano Zampini ierr = ISDestroy(&is_B_all);CHKERRQ(ierr); 1318683d3df6SStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)is_A_all),dual,0,1,&is_B_all);CHKERRQ(ierr); 13199448b7f1SJunchao Zhang ierr = VecScatterCreate(vec1_B,msolv_ctx->is_B,msolv_ctx->sol_B,is_B_all,&msolv_ctx->correction_scatter_B);CHKERRQ(ierr); 1320683d3df6SStefano Zampini ierr = ISDestroy(&is_B_all);CHKERRQ(ierr); 1321683d3df6SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)is_A_all),n-n_v,idxs,PETSC_COPY_VALUES,&msolv_ctx->is_R);CHKERRQ(ierr); 1322683d3df6SStefano Zampini ierr = ISRestoreIndices(is_A_all,&idxs);CHKERRQ(ierr); 1323683d3df6SStefano Zampini } 13243462e049SStefano Zampini ierr = ISGetLocalSize(msolv_ctx->is_R,&n_R);CHKERRQ(ierr); 13253462e049SStefano Zampini ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,n_R,n_R,0,NULL,&Afake);CHKERRQ(ierr); 13268d692618SStefano Zampini ierr = MatAssemblyBegin(Afake,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 13278d692618SStefano Zampini ierr = MatAssemblyEnd(Afake,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 13283462e049SStefano Zampini ierr = PCSetOperators(msolv_ctx->correction_solver,Afake,Afake);CHKERRQ(ierr); 13293462e049SStefano Zampini ierr = MatDestroy(&Afake);CHKERRQ(ierr); 1330683d3df6SStefano Zampini ierr = VecDestroy(&vec1_B);CHKERRQ(ierr); 1331ca92afb2SStefano Zampini 1332ca92afb2SStefano Zampini /* communicate benign info to solver context */ 1333ca92afb2SStefano Zampini if (benign_n) { 13345cbda25cSStefano Zampini PetscScalar *array; 13355cbda25cSStefano Zampini 1336ca92afb2SStefano Zampini msolv_ctx->benign_n = benign_n; 1337ca92afb2SStefano Zampini msolv_ctx->benign_zerodiag_subs = is_p_r; 1338ca92afb2SStefano Zampini ierr = PetscMalloc1(benign_n,&msolv_ctx->benign_save_vals);CHKERRQ(ierr); 13395cbda25cSStefano Zampini msolv_ctx->benign_csAIB = cs_AIB_mat; 1340282d6408SStefano Zampini ierr = MatCreateVecs(cs_AIB_mat,&msolv_ctx->benign_corr_work,NULL);CHKERRQ(ierr); 13415cbda25cSStefano Zampini ierr = VecGetArray(msolv_ctx->benign_corr_work,&array);CHKERRQ(ierr); 13425cbda25cSStefano Zampini ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,size_schur,array,&msolv_ctx->benign_dummy_schur_vec);CHKERRQ(ierr); 13435cbda25cSStefano Zampini ierr = VecRestoreArray(msolv_ctx->benign_corr_work,&array);CHKERRQ(ierr); 13445cbda25cSStefano Zampini msolv_ctx->benign_AIIm1ones = benign_AIIm1_ones_mat; 1345ca92afb2SStefano Zampini } 1346ada6e2d7SStefano Zampini } else { 1347ada6e2d7SStefano Zampini if (sub_schurs->reuse_solver) { 1348ada6e2d7SStefano Zampini ierr = PCBDDCReuseSolversReset(sub_schurs->reuse_solver);CHKERRQ(ierr); 1349ada6e2d7SStefano Zampini } 1350ada6e2d7SStefano Zampini ierr = PetscFree(sub_schurs->reuse_solver);CHKERRQ(ierr); 1351d5574798SStefano Zampini } 135208122e43SStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 135353892102SStefano Zampini ierr = ISDestroy(&is_A_all);CHKERRQ(ierr); 13545db18549SStefano Zampini 1355be83ff47SStefano Zampini /* Work arrays */ 135604c5b2e6SStefano Zampini ierr = PetscMalloc1(max_subset_size*max_subset_size,&work);CHKERRQ(ierr); 1357d2627357SStefano Zampini 1358be83ff47SStefano Zampini /* S_Ej_all */ 1359be83ff47SStefano Zampini cum = cum2 = 0; 13601683a169SBarry Smith ierr = MatDenseGetArrayRead(S_all,&rS_data);CHKERRQ(ierr); 136104c5b2e6SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->S_Ej_all,&SEj_arr);CHKERRQ(ierr); 136204c5b2e6SStefano Zampini if (compute_Stilda) { 136304c5b2e6SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&SEjinv_arr);CHKERRQ(ierr); 136404c5b2e6SStefano Zampini } 136565d8bf0aSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 136665d8bf0aSStefano Zampini PetscInt j; 136765d8bf0aSStefano Zampini 13685a95e1ceSStefano Zampini /* get S_E */ 1369b96c3477SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 1370683d3df6SStefano Zampini if (S_lower_triangular) { /* I need to expand the upper triangular data (column oriented) */ 1371be83ff47SStefano Zampini PetscInt k; 1372be83ff47SStefano Zampini for (k=0;k<subset_size;k++) { 1373be83ff47SStefano Zampini for (j=k;j<subset_size;j++) { 13741683a169SBarry Smith work[k*subset_size+j] = rS_data[cum2+k*size_schur+j]; 13751683a169SBarry Smith work[j*subset_size+k] = PetscConj(rS_data[cum2+k*size_schur+j]); 1376be83ff47SStefano Zampini } 1377be83ff47SStefano Zampini } 137806a4e24aSStefano Zampini } else { /* just copy to workspace */ 1379be83ff47SStefano Zampini PetscInt k; 1380be83ff47SStefano Zampini for (k=0;k<subset_size;k++) { 1381be83ff47SStefano Zampini for (j=0;j<subset_size;j++) { 13821683a169SBarry Smith work[k*subset_size+j] = rS_data[cum2+k*size_schur+j]; 1383be83ff47SStefano Zampini } 1384be83ff47SStefano Zampini } 13859087bf02SStefano Zampini } 13865a95e1ceSStefano Zampini /* insert S_E values */ 1387b7ab4a40SStefano Zampini if (sub_schurs->change) { 13888760537fSStefano Zampini Mat change_sub,SEj,T; 138972b8c272SStefano Zampini 139072b8c272SStefano Zampini /* change basis */ 139172b8c272SStefano Zampini ierr = KSPGetOperators(sub_schurs->change[i],&change_sub,NULL);CHKERRQ(ierr); 139272b8c272SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,subset_size,work,&SEj);CHKERRQ(ierr); 13938760537fSStefano Zampini if (!sub_schurs->change_with_qr) { /* currently there's no support for PtAP with P SeqAIJ */ 13948760537fSStefano Zampini Mat T2; 13958760537fSStefano Zampini ierr = MatTransposeMatMult(change_sub,SEj,MAT_INITIAL_MATRIX,1.0,&T2);CHKERRQ(ierr); 13968760537fSStefano Zampini ierr = MatMatMult(T2,change_sub,MAT_INITIAL_MATRIX,1.0,&T);CHKERRQ(ierr); 13978760537fSStefano Zampini ierr = MatConvert(T,MATSEQDENSE,MAT_INPLACE_MATRIX,&T);CHKERRQ(ierr); 139872b8c272SStefano Zampini ierr = MatDestroy(&T2);CHKERRQ(ierr); 13998760537fSStefano Zampini } else { 14008760537fSStefano Zampini ierr = MatPtAP(SEj,change_sub,MAT_INITIAL_MATRIX,1.0,&T);CHKERRQ(ierr); 140172b8c272SStefano Zampini } 140272b8c272SStefano Zampini ierr = MatCopy(T,SEj,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 140372b8c272SStefano Zampini ierr = MatDestroy(&T);CHKERRQ(ierr); 1404afa0f562SStefano Zampini ierr = MatZeroRowsColumnsIS(SEj,sub_schurs->change_primal_sub[i],1.0,NULL,NULL);CHKERRQ(ierr); 140572b8c272SStefano Zampini ierr = MatDestroy(&SEj);CHKERRQ(ierr); 140672b8c272SStefano Zampini } 14079d54b7f4SStefano Zampini if (deluxe) { 140804c5b2e6SStefano Zampini ierr = PetscArraycpy(SEj_arr,work,subset_size*subset_size);CHKERRQ(ierr); 1409683d3df6SStefano Zampini /* if adaptivity is requested, invert S_E blocks */ 1410862806e4SStefano Zampini if (compute_Stilda) { 14117ebab0bbSStefano Zampini Mat M; 14127ebab0bbSStefano Zampini const PetscScalar *vals; 14137ebab0bbSStefano Zampini PetscBool isdense,isdensecuda; 1414f4f7d9d6SStefano Zampini 14157ebab0bbSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,subset_size,work,&M);CHKERRQ(ierr); 14167ebab0bbSStefano Zampini ierr = MatSetOption(M,MAT_SPD,sub_schurs->is_posdef);CHKERRQ(ierr); 14177ebab0bbSStefano Zampini ierr = MatSetOption(M,MAT_HERMITIAN,sub_schurs->is_hermitian);CHKERRQ(ierr); 14187ebab0bbSStefano Zampini if (!PetscBTLookup(sub_schurs->is_edge,i)) { 14197ebab0bbSStefano Zampini ierr = MatSetType(M,Stype);CHKERRQ(ierr); 14206c3e6151SStefano Zampini } 14217ebab0bbSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)M,MATSEQDENSE,&isdense);CHKERRQ(ierr); 14227ebab0bbSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)M,MATSEQDENSECUDA,&isdensecuda);CHKERRQ(ierr); 14237ebab0bbSStefano Zampini if (use_cholesky) { 14247ebab0bbSStefano Zampini ierr = MatCholeskyFactor(M,NULL,NULL);CHKERRQ(ierr); 1425d6462365SStefano Zampini } else { 14267ebab0bbSStefano Zampini ierr = MatLUFactor(M,NULL,NULL,NULL);CHKERRQ(ierr); 14272972d61bSStefano Zampini } 14287ebab0bbSStefano Zampini if (isdense) { 14297ebab0bbSStefano Zampini ierr = MatSeqDenseInvertFactors_Private(M);CHKERRQ(ierr); 14307ebab0bbSStefano Zampini #if defined(PETSC_HAVE_CUDA) 14317ebab0bbSStefano Zampini } else if (isdensecuda) { 14327ebab0bbSStefano Zampini ierr = MatSeqDenseCUDAInvertFactors_Private(M);CHKERRQ(ierr); 14337ebab0bbSStefano Zampini #endif 143498921bdaSJacob Faibussowitsch } else SETERRQ(PetscObjectComm((PetscObject)M),PETSC_ERR_SUP,"Not implemented for type %s",Stype); 14357ebab0bbSStefano Zampini ierr = MatDenseGetArrayRead(M,&vals);CHKERRQ(ierr); 143604c5b2e6SStefano Zampini ierr = PetscArraycpy(SEjinv_arr,vals,subset_size*subset_size);CHKERRQ(ierr); 14377ebab0bbSStefano Zampini ierr = MatDenseRestoreArrayRead(M,&vals);CHKERRQ(ierr); 14387ebab0bbSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 14399087bf02SStefano Zampini } 14409d54b7f4SStefano Zampini } else if (compute_Stilda) { /* not using deluxe */ 14419d54b7f4SStefano Zampini Mat SEj; 14429d54b7f4SStefano Zampini Vec D; 14439d54b7f4SStefano Zampini PetscScalar *array; 14449d54b7f4SStefano Zampini 14459d54b7f4SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,subset_size,work,&SEj);CHKERRQ(ierr); 14469d54b7f4SStefano Zampini ierr = VecGetArray(Dall,&array);CHKERRQ(ierr); 14479d54b7f4SStefano Zampini ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,subset_size,array+cum,&D);CHKERRQ(ierr); 14489d54b7f4SStefano Zampini ierr = VecRestoreArray(Dall,&array);CHKERRQ(ierr); 1449f17d2ae1SStefano Zampini ierr = VecShift(D,-1.);CHKERRQ(ierr); 14509d54b7f4SStefano Zampini ierr = MatDiagonalScale(SEj,D,D);CHKERRQ(ierr); 14519d54b7f4SStefano Zampini ierr = MatDestroy(&SEj);CHKERRQ(ierr); 14529d54b7f4SStefano Zampini ierr = VecDestroy(&D);CHKERRQ(ierr); 145304c5b2e6SStefano Zampini ierr = PetscArraycpy(SEj_arr,work,subset_size*subset_size);CHKERRQ(ierr); 14549d54b7f4SStefano Zampini } 1455be83ff47SStefano Zampini cum += subset_size; 1456be83ff47SStefano Zampini cum2 += subset_size*(size_schur + 1); 145704c5b2e6SStefano Zampini SEj_arr += subset_size*subset_size; 145804c5b2e6SStefano Zampini if (SEjinv_arr) SEjinv_arr += subset_size*subset_size; 1459be83ff47SStefano Zampini } 14601683a169SBarry Smith ierr = MatDenseRestoreArrayRead(S_all,&rS_data);CHKERRQ(ierr); 146104c5b2e6SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->S_Ej_all,&SEj_arr);CHKERRQ(ierr); 146204c5b2e6SStefano Zampini if (compute_Stilda) { 146304c5b2e6SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&SEjinv_arr);CHKERRQ(ierr); 146404c5b2e6SStefano Zampini } 1465df4d28bfSStefano Zampini if (solver_S) { 14667c2f51b8SStefano Zampini ierr = MatFactorRestoreSchurComplement(F,&S_all,MAT_FACTOR_SCHUR_UNFACTORED);CHKERRQ(ierr); 14674a6c6b0dSStefano Zampini } 1468683d3df6SStefano Zampini 14697ebab0bbSStefano Zampini /* may prevent from unneeded copies, since MUMPS or MKL_Pardiso always use CPU memory 14707ebab0bbSStefano Zampini however, preliminary tests indicate using GPUs is still faster in the solve phase */ 14717ebab0bbSStefano Zampini #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 14727ebab0bbSStefano Zampini if (reuse_solvers) { 14737ebab0bbSStefano Zampini Mat St; 14747ebab0bbSStefano Zampini MatFactorSchurStatus st; 14757ebab0bbSStefano Zampini 147635d0533cSStefano Zampini flg = PETSC_FALSE; 14777ebab0bbSStefano Zampini ierr = PetscOptionsGetBool(NULL,sub_schurs->prefix,"-sub_schurs_schur_pin_to_cpu",&flg,NULL);CHKERRQ(ierr); 14787ebab0bbSStefano Zampini ierr = MatFactorGetSchurComplement(F,&St,&st);CHKERRQ(ierr); 1479b470e4b4SRichard Tran Mills ierr = MatBindToCPU(St,flg);CHKERRQ(ierr); 14807ebab0bbSStefano Zampini ierr = MatFactorRestoreSchurComplement(F,&St,st);CHKERRQ(ierr); 14817ebab0bbSStefano Zampini } 14827ebab0bbSStefano Zampini #endif 14837ebab0bbSStefano Zampini 1484683d3df6SStefano Zampini schur_factor = NULL; 148545951f25SStefano Zampini if (compute_Stilda && size_active_schur) { 1486683d3df6SStefano Zampini 148704c5b2e6SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&SEjinv_arr);CHKERRQ(ierr); 14889d54b7f4SStefano Zampini if (sub_schurs->n_subs == 1 && size_schur == size_active_schur && deluxe) { /* we already computed the inverse */ 148904c5b2e6SStefano Zampini ierr = PetscArraycpy(SEjinv_arr,work,size_schur*size_schur);CHKERRQ(ierr); 14904a6c6b0dSStefano Zampini } else { 1491683d3df6SStefano Zampini Mat S_all_inv=NULL; 14927ebab0bbSStefano Zampini 14933fc34f97SStefano Zampini if (solver_S) { 1494683d3df6SStefano Zampini /* for adaptive selection we need S^-1; for solver reusage we need S_\Delta\Delta^-1. 1495683d3df6SStefano 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 */ 14963fc34f97SStefano Zampini if (factor_workaround) {/* invert without calling MatFactorInvertSchurComplement, since we are hacking */ 1497683d3df6SStefano Zampini PetscScalar *data; 1498683d3df6SStefano Zampini PetscInt nd = 0; 14996dba178dSStefano Zampini 1500f4f7d9d6SStefano Zampini if (!use_potr) { 1501683d3df6SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Factor update not yet implemented for non SPD matrices"); 1502683d3df6SStefano Zampini } 15037c2f51b8SStefano Zampini ierr = MatFactorGetSchurComplement(F,&S_all_inv,NULL);CHKERRQ(ierr); 1504683d3df6SStefano Zampini ierr = MatDenseGetArray(S_all_inv,&data);CHKERRQ(ierr); 1505683d3df6SStefano Zampini if (sub_schurs->is_dir) { /* dirichlet dofs could have different scalings */ 1506683d3df6SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_dir,&nd);CHKERRQ(ierr); 1507683d3df6SStefano Zampini } 15083fc34f97SStefano Zampini 15093fc34f97SStefano Zampini /* factor and invert activedofs and vertices (dirichlet dofs does not contribute) */ 15103fc34f97SStefano Zampini if (schur_has_vertices) { 15113fc34f97SStefano Zampini Mat M; 15123fc34f97SStefano Zampini PetscScalar *tdata; 15133fc34f97SStefano Zampini PetscInt nv = 0, news; 15143fc34f97SStefano Zampini 15153fc34f97SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 15163fc34f97SStefano Zampini news = size_active_schur + nv; 15173fc34f97SStefano Zampini ierr = PetscCalloc1(news*news,&tdata);CHKERRQ(ierr); 1518683d3df6SStefano Zampini for (i=0;i<size_active_schur;i++) { 1519580bdb30SBarry Smith ierr = PetscArraycpy(tdata+i*(news+1),data+i*(size_schur+1),size_active_schur-i);CHKERRQ(ierr); 1520580bdb30SBarry Smith ierr = PetscArraycpy(tdata+i*(news+1)+size_active_schur-i,data+i*size_schur+size_active_schur+nd,nv);CHKERRQ(ierr); 15213fc34f97SStefano Zampini } 15223fc34f97SStefano Zampini for (i=0;i<nv;i++) { 15233fc34f97SStefano Zampini PetscInt k = i+size_active_schur; 1524580bdb30SBarry Smith ierr = PetscArraycpy(tdata+k*(news+1),data+(k+nd)*(size_schur+1),nv-i);CHKERRQ(ierr); 15253fc34f97SStefano Zampini } 15263fc34f97SStefano Zampini 15273fc34f97SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,news,news,tdata,&M);CHKERRQ(ierr); 15283fc34f97SStefano Zampini ierr = MatSetOption(M,MAT_SPD,PETSC_TRUE);CHKERRQ(ierr); 15293fc34f97SStefano Zampini ierr = MatCholeskyFactor(M,NULL,NULL);CHKERRQ(ierr); 15303fc34f97SStefano Zampini /* save the factors */ 15313fc34f97SStefano Zampini cum = 0; 15323fc34f97SStefano Zampini ierr = PetscMalloc1((size_active_schur*(size_active_schur +1))/2+nd,&schur_factor);CHKERRQ(ierr); 15333fc34f97SStefano Zampini for (i=0;i<size_active_schur;i++) { 1534580bdb30SBarry Smith ierr = PetscArraycpy(schur_factor+cum,tdata+i*(news+1),size_active_schur-i);CHKERRQ(ierr); 1535683d3df6SStefano Zampini cum += size_active_schur - i; 1536683d3df6SStefano Zampini } 15373fc34f97SStefano Zampini for (i=0;i<nd;i++) schur_factor[cum+i] = PetscSqrtReal(PetscRealPart(data[(i+size_active_schur)*(size_schur+1)])); 15383fc34f97SStefano Zampini ierr = MatSeqDenseInvertFactors_Private(M);CHKERRQ(ierr); 15393fc34f97SStefano Zampini /* move back just the active dofs to the Schur complement */ 15403fc34f97SStefano Zampini for (i=0;i<size_active_schur;i++) { 1541580bdb30SBarry Smith ierr = PetscArraycpy(data+i*size_schur,tdata+i*news,size_active_schur);CHKERRQ(ierr); 15423fc34f97SStefano Zampini } 15433fc34f97SStefano Zampini ierr = PetscFree(tdata);CHKERRQ(ierr); 15443fc34f97SStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 15453fc34f97SStefano Zampini } else { /* we can factorize and invert just the activedofs */ 15463fc34f97SStefano Zampini Mat M; 15475002105bSStefano Zampini PetscScalar *aux; 15483fc34f97SStefano Zampini 15495002105bSStefano Zampini ierr = PetscMalloc1(nd,&aux);CHKERRQ(ierr); 15505002105bSStefano Zampini for (i=0;i<nd;i++) aux[i] = 1.0/data[(i+size_active_schur)*(size_schur+1)]; 15513fc34f97SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,size_active_schur,size_active_schur,data,&M);CHKERRQ(ierr); 1552ad16ce7aSStefano Zampini ierr = MatDenseSetLDA(M,size_schur);CHKERRQ(ierr); 15533fc34f97SStefano Zampini ierr = MatSetOption(M,MAT_SPD,PETSC_TRUE);CHKERRQ(ierr); 15543fc34f97SStefano Zampini ierr = MatCholeskyFactor(M,NULL,NULL);CHKERRQ(ierr); 15553fc34f97SStefano Zampini ierr = MatSeqDenseInvertFactors_Private(M);CHKERRQ(ierr); 15563fc34f97SStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 15575002105bSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,size_schur,nd,data+size_active_schur*size_schur,&M);CHKERRQ(ierr); 15585002105bSStefano Zampini ierr = MatZeroEntries(M);CHKERRQ(ierr); 15595002105bSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 15605002105bSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,nd,size_schur,data+size_active_schur,&M);CHKERRQ(ierr); 1561ad16ce7aSStefano Zampini ierr = MatDenseSetLDA(M,size_schur);CHKERRQ(ierr); 15625002105bSStefano Zampini ierr = MatZeroEntries(M);CHKERRQ(ierr); 15635002105bSStefano Zampini ierr = MatDestroy(&M);CHKERRQ(ierr); 15645002105bSStefano Zampini for (i=0;i<nd;i++) data[(i+size_active_schur)*(size_schur+1)] = aux[i]; 15655002105bSStefano Zampini ierr = PetscFree(aux);CHKERRQ(ierr); 15663fc34f97SStefano Zampini } 1567683d3df6SStefano Zampini ierr = MatDenseRestoreArray(S_all_inv,&data);CHKERRQ(ierr); 15683fc34f97SStefano Zampini } else { /* use MatFactor calls to invert S */ 15696dba178dSStefano Zampini ierr = MatFactorInvertSchurComplement(F);CHKERRQ(ierr); 15707c2f51b8SStefano Zampini ierr = MatFactorGetSchurComplement(F,&S_all_inv,NULL);CHKERRQ(ierr); 1571683d3df6SStefano Zampini } 1572df4d28bfSStefano Zampini } else { /* we need to invert explicitly since we are not using MatFactor for S */ 1573683d3df6SStefano Zampini ierr = PetscObjectReference((PetscObject)S_all);CHKERRQ(ierr); 1574683d3df6SStefano Zampini S_all_inv = S_all; 1575683d3df6SStefano Zampini ierr = MatDenseGetArray(S_all_inv,&S_data);CHKERRQ(ierr); 1576be83ff47SStefano Zampini ierr = PetscBLASIntCast(size_schur,&B_N);CHKERRQ(ierr); 1577be83ff47SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1578f4f7d9d6SStefano Zampini if (use_potr) { 1579be83ff47SStefano Zampini PetscStackCallBLAS("LAPACKpotrf",LAPACKpotrf_("L",&B_N,S_data,&B_N,&B_ierr)); 1580*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(B_ierr,PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRF Lapack routine %d",(int)B_ierr); 1581be83ff47SStefano Zampini PetscStackCallBLAS("LAPACKpotri",LAPACKpotri_("L",&B_N,S_data,&B_N,&B_ierr)); 1582*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(B_ierr,PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRI Lapack routine %d",(int)B_ierr); 1583f4f7d9d6SStefano Zampini } else if (use_sytr) { 1584f4f7d9d6SStefano Zampini PetscStackCallBLAS("LAPACKsytrf",LAPACKsytrf_("L",&B_N,S_data,&B_N,pivots,Bwork,&B_lwork,&B_ierr)); 1585*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(B_ierr,PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYTRF Lapack routine %d",(int)B_ierr); 1586f4f7d9d6SStefano Zampini PetscStackCallBLAS("LAPACKsytri",LAPACKsytri_("L",&B_N,S_data,&B_N,pivots,Bwork,&B_ierr)); 1587*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(B_ierr,PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYTRI Lapack routine %d",(int)B_ierr); 1588d6462365SStefano Zampini } else { 1589d6462365SStefano Zampini PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,S_data,&B_N,pivots,&B_ierr)); 1590*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(B_ierr,PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr); 1591d6462365SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,S_data,&B_N,pivots,Bwork,&B_lwork,&B_ierr)); 1592*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(B_ierr,PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr); 1593be83ff47SStefano Zampini } 159443371fb9SStefano Zampini ierr = PetscLogFlops(1.0*size_schur*size_schur*size_schur);CHKERRQ(ierr); 1595be83ff47SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 1596683d3df6SStefano Zampini ierr = MatDenseRestoreArray(S_all_inv,&S_data);CHKERRQ(ierr); 1597be83ff47SStefano Zampini } 1598be83ff47SStefano Zampini /* S_Ej_tilda_all */ 1599be83ff47SStefano Zampini cum = cum2 = 0; 16001683a169SBarry Smith ierr = MatDenseGetArrayRead(S_all_inv,&rS_data);CHKERRQ(ierr); 1601be83ff47SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 1602be83ff47SStefano Zampini PetscInt j; 1603862806e4SStefano Zampini 1604862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 1605be83ff47SStefano Zampini /* get (St^-1)_E */ 160672b8c272SStefano Zampini /* Unless we are changing the variables, I don't need to expand to upper triangular since St^-1 160706a4e24aSStefano Zampini will be properly accessed later during adaptive selection */ 1608a0b0af32SStefano Zampini if (S_lower_triangular) { 1609be83ff47SStefano Zampini PetscInt k; 1610b7ab4a40SStefano Zampini if (sub_schurs->change) { 1611be83ff47SStefano Zampini for (k=0;k<subset_size;k++) { 1612be83ff47SStefano Zampini for (j=k;j<subset_size;j++) { 16131683a169SBarry Smith work[k*subset_size+j] = rS_data[cum2+k*size_schur+j]; 16146c3e6151SStefano Zampini work[j*subset_size+k] = work[k*subset_size+j]; 1615be83ff47SStefano Zampini } 1616be83ff47SStefano Zampini } 161772b8c272SStefano Zampini } else { 161872b8c272SStefano Zampini for (k=0;k<subset_size;k++) { 161972b8c272SStefano Zampini for (j=k;j<subset_size;j++) { 16201683a169SBarry Smith work[k*subset_size+j] = rS_data[cum2+k*size_schur+j]; 162172b8c272SStefano Zampini } 162272b8c272SStefano Zampini } 162372b8c272SStefano Zampini } 162472b8c272SStefano Zampini } else { 1625be83ff47SStefano Zampini PetscInt k; 1626be83ff47SStefano Zampini for (k=0;k<subset_size;k++) { 1627be83ff47SStefano Zampini for (j=0;j<subset_size;j++) { 16281683a169SBarry Smith work[k*subset_size+j] = rS_data[cum2+k*size_schur+j]; 1629be83ff47SStefano Zampini } 1630be83ff47SStefano Zampini } 1631be83ff47SStefano Zampini } 1632b7ab4a40SStefano Zampini if (sub_schurs->change) { 16338760537fSStefano Zampini Mat change_sub,SEj,T; 163472b8c272SStefano Zampini 163572b8c272SStefano Zampini /* change basis */ 163672b8c272SStefano Zampini ierr = KSPGetOperators(sub_schurs->change[i],&change_sub,NULL);CHKERRQ(ierr); 163772b8c272SStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,subset_size,work,&SEj);CHKERRQ(ierr); 16388760537fSStefano Zampini if (!sub_schurs->change_with_qr) { /* currently there's no support for PtAP with P SeqAIJ */ 16398760537fSStefano Zampini Mat T2; 16408760537fSStefano Zampini ierr = MatTransposeMatMult(change_sub,SEj,MAT_INITIAL_MATRIX,1.0,&T2);CHKERRQ(ierr); 16418760537fSStefano Zampini ierr = MatMatMult(T2,change_sub,MAT_INITIAL_MATRIX,1.0,&T);CHKERRQ(ierr); 164272b8c272SStefano Zampini ierr = MatDestroy(&T2);CHKERRQ(ierr); 16438760537fSStefano Zampini ierr = MatConvert(T,MATSEQDENSE,MAT_INPLACE_MATRIX,&T);CHKERRQ(ierr); 16448760537fSStefano Zampini } else { 16458760537fSStefano Zampini ierr = MatPtAP(SEj,change_sub,MAT_INITIAL_MATRIX,1.0,&T);CHKERRQ(ierr); 164672b8c272SStefano Zampini } 16478760537fSStefano Zampini ierr = MatCopy(T,SEj,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 16488760537fSStefano Zampini ierr = MatDestroy(&T);CHKERRQ(ierr); 164972b8c272SStefano Zampini /* set diagonal entry to a very large value to pick the basis we are eliminating as the first eigenvectors with adaptive selection */ 1650afa0f562SStefano Zampini ierr = MatZeroRowsColumnsIS(SEj,sub_schurs->change_primal_sub[i],1./PETSC_SMALL,NULL,NULL);CHKERRQ(ierr); 165172b8c272SStefano Zampini ierr = MatDestroy(&SEj);CHKERRQ(ierr); 165272b8c272SStefano Zampini } 165304c5b2e6SStefano Zampini ierr = PetscArraycpy(SEjinv_arr,work,subset_size*subset_size);CHKERRQ(ierr); 1654be83ff47SStefano Zampini cum += subset_size; 1655be83ff47SStefano Zampini cum2 += subset_size*(size_schur + 1); 165604c5b2e6SStefano Zampini SEjinv_arr += subset_size*subset_size; 1657883469d8SStefano Zampini } 16581683a169SBarry Smith ierr = MatDenseRestoreArrayRead(S_all_inv,&rS_data);CHKERRQ(ierr); 1659df4d28bfSStefano Zampini if (solver_S) { 16603fc34f97SStefano Zampini if (schur_has_vertices) { 16613fc34f97SStefano Zampini ierr = MatFactorRestoreSchurComplement(F,&S_all_inv,MAT_FACTOR_SCHUR_FACTORED);CHKERRQ(ierr); 16623fc34f97SStefano Zampini } else { 16637c2f51b8SStefano Zampini ierr = MatFactorRestoreSchurComplement(F,&S_all_inv,MAT_FACTOR_SCHUR_INVERTED);CHKERRQ(ierr); 16645db18549SStefano Zampini } 16653fc34f97SStefano Zampini } 1666683d3df6SStefano Zampini ierr = MatDestroy(&S_all_inv);CHKERRQ(ierr); 1667683d3df6SStefano Zampini } 166804c5b2e6SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&SEjinv_arr);CHKERRQ(ierr); 1669683d3df6SStefano Zampini 16703fc34f97SStefano Zampini /* move back factors if needed */ 16713fc34f97SStefano Zampini if (schur_has_vertices) { 1672683d3df6SStefano Zampini Mat S_tmp; 16733fc34f97SStefano Zampini PetscInt nd = 0; 1674683d3df6SStefano Zampini 1675*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(!solver_S,PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 16767c2f51b8SStefano Zampini ierr = MatFactorGetSchurComplement(F,&S_tmp,NULL);CHKERRQ(ierr); 1677f4f7d9d6SStefano Zampini if (use_potr) { 1678683d3df6SStefano Zampini PetscScalar *data; 1679683d3df6SStefano Zampini 1680683d3df6SStefano Zampini ierr = MatDenseGetArray(S_tmp,&data);CHKERRQ(ierr); 1681580bdb30SBarry Smith ierr = PetscArrayzero(data,size_schur*size_schur);CHKERRQ(ierr); 1682683d3df6SStefano Zampini 1683683d3df6SStefano Zampini if (S_lower_triangular) { 1684683d3df6SStefano Zampini cum = 0; 1685683d3df6SStefano Zampini for (i=0;i<size_active_schur;i++) { 1686580bdb30SBarry Smith ierr = PetscArraycpy(data+i*(size_schur+1),schur_factor+cum,size_active_schur-i);CHKERRQ(ierr); 1687683d3df6SStefano Zampini cum += size_active_schur-i; 1688683d3df6SStefano Zampini } 1689683d3df6SStefano Zampini } else { 1690580bdb30SBarry Smith ierr = PetscArraycpy(data,schur_factor,size_schur*size_schur);CHKERRQ(ierr); 1691683d3df6SStefano Zampini } 1692683d3df6SStefano Zampini if (sub_schurs->is_dir) { 1693683d3df6SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_dir,&nd);CHKERRQ(ierr); 1694683d3df6SStefano Zampini for (i=0;i<nd;i++) { 1695683d3df6SStefano Zampini data[(i+size_active_schur)*(size_schur+1)] = schur_factor[cum+i]; 1696683d3df6SStefano Zampini } 1697683d3df6SStefano Zampini } 16986dba178dSStefano Zampini /* workaround: since I cannot modify the matrices used inside the solvers for the forward and backward substitutions, 1699683d3df6SStefano Zampini set the diagonal entry of the Schur factor to a very large value */ 1700683d3df6SStefano Zampini for (i=size_active_schur+nd;i<size_schur;i++) { 17016c3e6151SStefano Zampini data[i*(size_schur+1)] = infty; 1702683d3df6SStefano Zampini } 1703683d3df6SStefano Zampini ierr = MatDenseRestoreArray(S_tmp,&data);CHKERRQ(ierr); 17046542c05fSStefano Zampini } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Factor update not yet implemented for non SPD matrices"); 17053fc34f97SStefano Zampini ierr = MatFactorRestoreSchurComplement(F,&S_tmp,MAT_FACTOR_SCHUR_FACTORED);CHKERRQ(ierr); 17069087bf02SStefano Zampini } 1707367aa537SStefano Zampini } else if (factor_workaround) { /* we need to eliminate any unneeded coupling */ 1708367aa537SStefano Zampini PetscScalar *data; 1709367aa537SStefano Zampini PetscInt nd = 0; 1710367aa537SStefano Zampini 1711367aa537SStefano Zampini if (sub_schurs->is_dir) { /* dirichlet dofs could have different scalings */ 1712367aa537SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_dir,&nd);CHKERRQ(ierr); 1713367aa537SStefano Zampini } 17147c2f51b8SStefano Zampini ierr = MatFactorGetSchurComplement(F,&S_all,NULL);CHKERRQ(ierr); 1715367aa537SStefano Zampini ierr = MatDenseGetArray(S_all,&data);CHKERRQ(ierr); 1716367aa537SStefano Zampini for (i=0;i<size_active_schur;i++) { 1717580bdb30SBarry Smith ierr = PetscArrayzero(data+i*size_schur+size_active_schur,size_schur-size_active_schur);CHKERRQ(ierr); 1718367aa537SStefano Zampini } 1719367aa537SStefano Zampini for (i=size_active_schur+nd;i<size_schur;i++) { 1720580bdb30SBarry Smith ierr = PetscArrayzero(data+i*size_schur+size_active_schur,size_schur-size_active_schur);CHKERRQ(ierr); 17216c3e6151SStefano Zampini data[i*(size_schur+1)] = infty; 1722367aa537SStefano Zampini } 1723367aa537SStefano Zampini ierr = MatDenseRestoreArray(S_all,&data);CHKERRQ(ierr); 17243fc34f97SStefano Zampini ierr = MatFactorRestoreSchurComplement(F,&S_all,MAT_FACTOR_SCHUR_UNFACTORED);CHKERRQ(ierr); 17254a6c6b0dSStefano Zampini } 172604c5b2e6SStefano Zampini ierr = PetscFree(work);CHKERRQ(ierr); 1727683d3df6SStefano Zampini ierr = PetscFree(schur_factor);CHKERRQ(ierr); 17289d54b7f4SStefano Zampini ierr = VecDestroy(&Dall);CHKERRQ(ierr); 17294a6c6b0dSStefano Zampini } 1730a9b99552SStefano Zampini ierr = ISDestroy(&is_I_layer);CHKERRQ(ierr); 1731a1337663SStefano Zampini ierr = MatDestroy(&S_all);CHKERRQ(ierr); 1732a64f4aa4SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 1733a64f4aa4SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 1734a64f4aa4SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 17357ebab0bbSStefano Zampini ierr = MatDestroy(&F);CHKERRQ(ierr); 17366afe12f5SStefano Zampini 17376afe12f5SStefano Zampini ierr = PetscMalloc1(sub_schurs->n_subs,&nnz);CHKERRQ(ierr); 17386afe12f5SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 17396afe12f5SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&nnz[i]);CHKERRQ(ierr); 17406afe12f5SStefano Zampini } 17416afe12f5SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,sub_schurs->n_subs,nnz,PETSC_OWN_POINTER,&is_I_layer);CHKERRQ(ierr); 17426afe12f5SStefano Zampini ierr = MatSetVariableBlockSizes(sub_schurs->S_Ej_all,sub_schurs->n_subs,nnz);CHKERRQ(ierr); 17435db18549SStefano Zampini ierr = MatAssemblyBegin(sub_schurs->S_Ej_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 17445db18549SStefano Zampini ierr = MatAssemblyEnd(sub_schurs->S_Ej_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 17455a95e1ceSStefano Zampini if (compute_Stilda) { 17466afe12f5SStefano Zampini ierr = MatSetVariableBlockSizes(sub_schurs->sum_S_Ej_tilda_all,sub_schurs->n_subs,nnz);CHKERRQ(ierr); 17475a94c880SStefano Zampini ierr = MatAssemblyBegin(sub_schurs->sum_S_Ej_tilda_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 17485a94c880SStefano Zampini ierr = MatAssemblyEnd(sub_schurs->sum_S_Ej_tilda_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 17499d54b7f4SStefano Zampini if (deluxe) { 17506afe12f5SStefano Zampini ierr = MatSetVariableBlockSizes(sub_schurs->sum_S_Ej_inv_all,sub_schurs->n_subs,nnz);CHKERRQ(ierr); 17515a94c880SStefano Zampini ierr = MatAssemblyBegin(sub_schurs->sum_S_Ej_inv_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 17525a94c880SStefano Zampini ierr = MatAssemblyEnd(sub_schurs->sum_S_Ej_inv_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 175308122e43SStefano Zampini } 17549d54b7f4SStefano Zampini } 17556afe12f5SStefano Zampini ierr = ISDestroy(&is_I_layer);CHKERRQ(ierr); 17566afe12f5SStefano Zampini 17575db18549SStefano Zampini /* Get local part of (\sum_j S_Ej) */ 175841fd5443SStefano Zampini if (!sub_schurs->sum_S_Ej_all) { 175957a87bf3SStefano Zampini ierr = MatDuplicate(sub_schurs->S_Ej_all,MAT_DO_NOT_COPY_VALUES,&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 176041fd5443SStefano Zampini } 176157a87bf3SStefano Zampini ierr = VecSet(gstash,0.0);CHKERRQ(ierr); 176257a87bf3SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->S_Ej_all,&stasharray);CHKERRQ(ierr); 176357a87bf3SStefano Zampini ierr = VecPlaceArray(lstash,stasharray);CHKERRQ(ierr); 176457a87bf3SStefano Zampini ierr = VecScatterBegin(sstash,lstash,gstash,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 176557a87bf3SStefano Zampini ierr = VecScatterEnd(sstash,lstash,gstash,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 176657a87bf3SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->S_Ej_all,&stasharray);CHKERRQ(ierr); 176757a87bf3SStefano Zampini ierr = VecResetArray(lstash);CHKERRQ(ierr); 176857a87bf3SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_all,&stasharray);CHKERRQ(ierr); 176957a87bf3SStefano Zampini ierr = VecPlaceArray(lstash,stasharray);CHKERRQ(ierr); 177057a87bf3SStefano Zampini ierr = VecScatterBegin(sstash,gstash,lstash,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 177157a87bf3SStefano Zampini ierr = VecScatterEnd(sstash,gstash,lstash,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 177257a87bf3SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_all,&stasharray);CHKERRQ(ierr); 177357a87bf3SStefano Zampini ierr = VecResetArray(lstash);CHKERRQ(ierr); 177408122e43SStefano Zampini 1775f6f667cfSStefano Zampini /* Get local part of (\sum_j S^-1_Ej) (\sum_j St^-1_Ej) */ 17765a95e1ceSStefano Zampini if (compute_Stilda) { 177757a87bf3SStefano Zampini ierr = VecSet(gstash,0.0);CHKERRQ(ierr); 177857a87bf3SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&stasharray);CHKERRQ(ierr); 177957a87bf3SStefano Zampini ierr = VecPlaceArray(lstash,stasharray);CHKERRQ(ierr); 178057a87bf3SStefano Zampini ierr = VecScatterBegin(sstash,lstash,gstash,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 178157a87bf3SStefano Zampini ierr = VecScatterEnd(sstash,lstash,gstash,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 178257a87bf3SStefano Zampini ierr = VecScatterBegin(sstash,gstash,lstash,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 178357a87bf3SStefano Zampini ierr = VecScatterEnd(sstash,gstash,lstash,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 178457a87bf3SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&stasharray);CHKERRQ(ierr); 178557a87bf3SStefano Zampini ierr = VecResetArray(lstash);CHKERRQ(ierr); 17869d54b7f4SStefano Zampini if (deluxe) { 178757a87bf3SStefano Zampini ierr = VecSet(gstash,0.0);CHKERRQ(ierr); 178857a87bf3SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&stasharray);CHKERRQ(ierr); 178957a87bf3SStefano Zampini ierr = VecPlaceArray(lstash,stasharray);CHKERRQ(ierr); 179057a87bf3SStefano Zampini ierr = VecScatterBegin(sstash,lstash,gstash,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179157a87bf3SStefano Zampini ierr = VecScatterEnd(sstash,lstash,gstash,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 179257a87bf3SStefano Zampini ierr = VecScatterBegin(sstash,gstash,lstash,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 179357a87bf3SStefano Zampini ierr = VecScatterEnd(sstash,gstash,lstash,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 179457a87bf3SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&stasharray);CHKERRQ(ierr); 179557a87bf3SStefano Zampini ierr = VecResetArray(lstash);CHKERRQ(ierr); 17969d54b7f4SStefano Zampini } else { 17979d54b7f4SStefano Zampini PetscScalar *array; 17989d54b7f4SStefano Zampini PetscInt cum; 17999d54b7f4SStefano Zampini 18009d54b7f4SStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&array);CHKERRQ(ierr); 18019d54b7f4SStefano Zampini cum = 0; 18029d54b7f4SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 18039d54b7f4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 18049d54b7f4SStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 18059d54b7f4SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1806f4f7d9d6SStefano Zampini if (use_potr) { 18079d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKpotrf",LAPACKpotrf_("L",&B_N,array+cum,&B_N,&B_ierr)); 1808*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(B_ierr,PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRF Lapack routine %d",(int)B_ierr); 18099d54b7f4SStefano Zampini PetscStackCallBLAS("LAPACKpotri",LAPACKpotri_("L",&B_N,array+cum,&B_N,&B_ierr)); 1810*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(B_ierr,PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRI Lapack routine %d",(int)B_ierr); 1811f4f7d9d6SStefano Zampini } else if (use_sytr) { 1812f4f7d9d6SStefano Zampini PetscStackCallBLAS("LAPACKsytrf",LAPACKsytrf_("L",&B_N,array+cum,&B_N,pivots,Bwork,&B_lwork,&B_ierr)); 1813*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(B_ierr,PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYTRF Lapack routine %d",(int)B_ierr); 1814f4f7d9d6SStefano Zampini PetscStackCallBLAS("LAPACKsytri",LAPACKsytri_("L",&B_N,array+cum,&B_N,pivots,Bwork,&B_ierr)); 1815*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(B_ierr,PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYTRI Lapack routine %d",(int)B_ierr); 1816f4f7d9d6SStefano Zampini } else { 1817f4f7d9d6SStefano Zampini PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,array+cum,&B_N,pivots,&B_ierr)); 1818*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(B_ierr,PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr); 1819f4f7d9d6SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,array+cum,&B_N,pivots,Bwork,&B_lwork,&B_ierr)); 1820*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(B_ierr,PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr); 1821f4f7d9d6SStefano Zampini } 182243371fb9SStefano Zampini ierr = PetscLogFlops(1.0*subset_size*subset_size*subset_size);CHKERRQ(ierr); 18239d54b7f4SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 18249d54b7f4SStefano Zampini cum += subset_size*subset_size; 18259d54b7f4SStefano Zampini } 18269d54b7f4SStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&array);CHKERRQ(ierr); 18279d54b7f4SStefano Zampini ierr = PetscObjectReference((PetscObject)sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 18289d54b7f4SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 18299d54b7f4SStefano Zampini sub_schurs->sum_S_Ej_inv_all = sub_schurs->sum_S_Ej_all; 18309d54b7f4SStefano Zampini } 183108122e43SStefano Zampini } 183257a87bf3SStefano Zampini ierr = VecDestroy(&lstash);CHKERRQ(ierr); 183357a87bf3SStefano Zampini ierr = VecDestroy(&gstash);CHKERRQ(ierr); 183457a87bf3SStefano Zampini ierr = VecScatterDestroy(&sstash);CHKERRQ(ierr); 183557a87bf3SStefano Zampini 18363b03f7bbSStefano Zampini if (matl_dbg_viewer) { 183711955456SStefano Zampini PetscInt cum; 183811955456SStefano Zampini 183911955456SStefano Zampini if (sub_schurs->S_Ej_all) { 184011955456SStefano Zampini ierr = PetscObjectSetName((PetscObject)sub_schurs->S_Ej_all,"SE");CHKERRQ(ierr); 18413b03f7bbSStefano Zampini ierr = MatView(sub_schurs->S_Ej_all,matl_dbg_viewer);CHKERRQ(ierr); 184211955456SStefano Zampini } 184311955456SStefano Zampini if (sub_schurs->sum_S_Ej_all) { 184411955456SStefano Zampini ierr = PetscObjectSetName((PetscObject)sub_schurs->sum_S_Ej_all,"SSE");CHKERRQ(ierr); 18453b03f7bbSStefano Zampini ierr = MatView(sub_schurs->sum_S_Ej_all,matl_dbg_viewer);CHKERRQ(ierr); 184611955456SStefano Zampini } 184711955456SStefano Zampini if (sub_schurs->sum_S_Ej_inv_all) { 184811955456SStefano Zampini ierr = PetscObjectSetName((PetscObject)sub_schurs->sum_S_Ej_inv_all,"SSEm");CHKERRQ(ierr); 18493b03f7bbSStefano Zampini ierr = MatView(sub_schurs->sum_S_Ej_inv_all,matl_dbg_viewer);CHKERRQ(ierr); 185011955456SStefano Zampini } 185111955456SStefano Zampini if (sub_schurs->sum_S_Ej_tilda_all) { 185211955456SStefano Zampini ierr = PetscObjectSetName((PetscObject)sub_schurs->sum_S_Ej_tilda_all,"SSEt");CHKERRQ(ierr); 18533b03f7bbSStefano Zampini ierr = MatView(sub_schurs->sum_S_Ej_tilda_all,matl_dbg_viewer);CHKERRQ(ierr); 185411955456SStefano Zampini } 185511955456SStefano Zampini for (i=0,cum=0;i<sub_schurs->n_subs;i++) { 185611955456SStefano Zampini IS is; 185711955456SStefano Zampini char name[16]; 185811955456SStefano Zampini 18597f9db97bSStefano Zampini ierr = PetscSNPrintf(name,sizeof(name),"IE%D",i);CHKERRQ(ierr); 186011955456SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 186111955456SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,subset_size,cum,1,&is);CHKERRQ(ierr); 186211955456SStefano Zampini ierr = PetscObjectSetName((PetscObject)is,name);CHKERRQ(ierr); 18633b03f7bbSStefano Zampini ierr = ISView(is,matl_dbg_viewer);CHKERRQ(ierr); 186411955456SStefano Zampini ierr = ISDestroy(&is);CHKERRQ(ierr); 186511955456SStefano Zampini cum += subset_size; 186611955456SStefano Zampini } 186711955456SStefano Zampini } 18683202ece2SStefano Zampini 18695a95e1ceSStefano Zampini /* free workspace */ 18703b03f7bbSStefano Zampini ierr = PetscViewerDestroy(&matl_dbg_viewer);CHKERRQ(ierr); 187106a4b1faSStefano Zampini ierr = PetscFree2(Bwork,pivots);CHKERRQ(ierr); 18725a95e1ceSStefano Zampini ierr = PetscCommDestroy(&comm_n);CHKERRQ(ierr); 1873b1b3d7a2SStefano Zampini PetscFunctionReturn(0); 1874b1b3d7a2SStefano Zampini } 1875b1b3d7a2SStefano Zampini 187688113c35SStefano Zampini PetscErrorCode PCBDDCSubSchursInit(PCBDDCSubSchurs sub_schurs, const char* prefix, IS is_I, IS is_B, PCBDDCGraph graph, ISLocalToGlobalMapping BtoNmap, PetscBool copycc) 1877b1b3d7a2SStefano Zampini { 18789bb4a8caSStefano Zampini IS *faces,*edges,*all_cc,vertices; 18795a95e1ceSStefano Zampini PetscInt i,n_faces,n_edges,n_all_cc; 1880365a3a41SStefano Zampini PetscBool is_sorted,ispardiso,ismumps; 1881b1b3d7a2SStefano Zampini PetscErrorCode ierr; 1882b1b3d7a2SStefano Zampini 1883b1b3d7a2SStefano Zampini PetscFunctionBegin; 1884b1b3d7a2SStefano Zampini ierr = ISSorted(is_I,&is_sorted);CHKERRQ(ierr); 1885*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(!is_sorted,PetscObjectComm((PetscObject)is_I),PETSC_ERR_PLIB,"IS for I dofs should be shorted"); 1886b1b3d7a2SStefano Zampini ierr = ISSorted(is_B,&is_sorted);CHKERRQ(ierr); 1887*2c71b3e2SJacob Faibussowitsch PetscCheckFalse(!is_sorted,PetscObjectComm((PetscObject)is_B),PETSC_ERR_PLIB,"IS for B dofs should be shorted"); 1888b1b3d7a2SStefano Zampini 1889b1b3d7a2SStefano Zampini /* reset any previous data */ 1890b1b3d7a2SStefano Zampini ierr = PCBDDCSubSchursReset(sub_schurs);CHKERRQ(ierr); 1891b1b3d7a2SStefano Zampini 18925a95e1ceSStefano Zampini /* get index sets for faces and edges (already sorted by global ordering) */ 18939bb4a8caSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&n_faces,&faces,&n_edges,&edges,&vertices);CHKERRQ(ierr); 1894b1b3d7a2SStefano Zampini n_all_cc = n_faces+n_edges; 189508122e43SStefano Zampini ierr = PetscBTCreate(n_all_cc,&sub_schurs->is_edge);CHKERRQ(ierr); 1896b1b3d7a2SStefano Zampini ierr = PetscMalloc1(n_all_cc,&all_cc);CHKERRQ(ierr); 1897b1b3d7a2SStefano Zampini for (i=0;i<n_faces;i++) { 18988b6046baSStefano Zampini if (copycc) { 18998b6046baSStefano Zampini ierr = ISDuplicate(faces[i],&all_cc[i]);CHKERRQ(ierr); 19008b6046baSStefano Zampini } else { 1901c8272957SStefano Zampini ierr = PetscObjectReference((PetscObject)faces[i]);CHKERRQ(ierr); 1902b1b3d7a2SStefano Zampini all_cc[i] = faces[i]; 1903b1b3d7a2SStefano Zampini } 19048b6046baSStefano Zampini } 1905b1b3d7a2SStefano Zampini for (i=0;i<n_edges;i++) { 19068b6046baSStefano Zampini if (copycc) { 19078b6046baSStefano Zampini ierr = ISDuplicate(edges[i],&all_cc[n_faces+i]);CHKERRQ(ierr); 19088b6046baSStefano Zampini } else { 1909c8272957SStefano Zampini ierr = PetscObjectReference((PetscObject)edges[i]);CHKERRQ(ierr); 1910b1b3d7a2SStefano Zampini all_cc[n_faces+i] = edges[i]; 19118b6046baSStefano Zampini } 191208122e43SStefano Zampini ierr = PetscBTSet(sub_schurs->is_edge,n_faces+i);CHKERRQ(ierr); 1913b1b3d7a2SStefano Zampini } 1914c8272957SStefano Zampini ierr = PetscObjectReference((PetscObject)vertices);CHKERRQ(ierr); 1915c8272957SStefano Zampini sub_schurs->is_vertices = vertices; 1916c8272957SStefano Zampini ierr = PCBDDCGraphRestoreCandidatesIS(graph,&n_faces,&faces,&n_edges,&edges,&vertices);CHKERRQ(ierr); 1917d62866d3SStefano Zampini sub_schurs->is_dir = NULL; 1918d62866d3SStefano Zampini ierr = PCBDDCGraphGetDirichletDofsB(graph,&sub_schurs->is_dir);CHKERRQ(ierr); 1919b1b3d7a2SStefano Zampini 1920df4d28bfSStefano Zampini /* Determine if MatFactor can be used */ 192188113c35SStefano Zampini ierr = PetscStrallocpy(prefix,&sub_schurs->prefix);CHKERRQ(ierr); 1922883469d8SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 1923589a23caSBarry Smith ierr = PetscStrncpy(sub_schurs->mat_solver_type,MATSOLVERMUMPS,sizeof(sub_schurs->mat_solver_type));CHKERRQ(ierr); 192488113c35SStefano Zampini #elif defined(PETSC_HAVE_MKL_PARDISO) 1925589a23caSBarry Smith ierr = PetscStrncpy(sub_schurs->mat_solver_type,MATSOLVERMKL_PARDISO,sizeof(sub_schurs->mat_solver_type));CHKERRQ(ierr); 192688113c35SStefano Zampini #else 1927589a23caSBarry Smith ierr = PetscStrncpy(sub_schurs->mat_solver_type,MATSOLVERPETSC,sizeof(sub_schurs->mat_solver_type));CHKERRQ(ierr); 1928df4d28bfSStefano Zampini #endif 192988113c35SStefano Zampini #if defined(PETSC_USE_COMPLEX) 193088113c35SStefano Zampini sub_schurs->is_hermitian = PETSC_FALSE; /* Hermitian Cholesky is not supported by PETSc and external packages */ 193188113c35SStefano Zampini #else 193288113c35SStefano Zampini sub_schurs->is_hermitian = PETSC_TRUE; 1933883469d8SStefano Zampini #endif 193488113c35SStefano Zampini sub_schurs->is_posdef = PETSC_TRUE; 193511955456SStefano Zampini sub_schurs->is_symmetric = PETSC_TRUE; 19367f9db97bSStefano Zampini sub_schurs->debug = PETSC_FALSE; 1937991c41b4SStefano Zampini sub_schurs->restrict_comm = PETSC_FALSE; 193888113c35SStefano Zampini ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)graph->l2gmap),sub_schurs->prefix,"BDDC sub_schurs options","PC");CHKERRQ(ierr); 1939589a23caSBarry Smith ierr = PetscOptionsString("-sub_schurs_mat_solver_type","Specific direct solver to use",NULL,sub_schurs->mat_solver_type,sub_schurs->mat_solver_type,sizeof(sub_schurs->mat_solver_type),NULL);CHKERRQ(ierr); 194011955456SStefano Zampini ierr = PetscOptionsBool("-sub_schurs_symmetric","Symmetric problem",NULL,sub_schurs->is_symmetric,&sub_schurs->is_symmetric,NULL);CHKERRQ(ierr); 194188113c35SStefano Zampini ierr = PetscOptionsBool("-sub_schurs_hermitian","Hermitian problem",NULL,sub_schurs->is_hermitian,&sub_schurs->is_hermitian,NULL);CHKERRQ(ierr); 194288113c35SStefano Zampini ierr = PetscOptionsBool("-sub_schurs_posdef","Positive definite problem",NULL,sub_schurs->is_posdef,&sub_schurs->is_posdef,NULL);CHKERRQ(ierr); 1943991c41b4SStefano Zampini ierr = PetscOptionsBool("-sub_schurs_restrictcomm","Restrict communicator on active processes only",NULL,sub_schurs->restrict_comm,&sub_schurs->restrict_comm,NULL);CHKERRQ(ierr); 19447f9db97bSStefano Zampini ierr = PetscOptionsBool("-sub_schurs_debug","Debug output",NULL,sub_schurs->debug,&sub_schurs->debug,NULL);CHKERRQ(ierr); 194588113c35SStefano Zampini ierr = PetscOptionsEnd();CHKERRQ(ierr); 1946365a3a41SStefano Zampini ierr = PetscStrcmp(sub_schurs->mat_solver_type,MATSOLVERMUMPS,&ismumps);CHKERRQ(ierr); 1947365a3a41SStefano Zampini ierr = PetscStrcmp(sub_schurs->mat_solver_type,MATSOLVERMKL_PARDISO,&ispardiso);CHKERRQ(ierr); 1948365a3a41SStefano Zampini sub_schurs->schur_explicit = (PetscBool)(ispardiso || ismumps); 1949b1b3d7a2SStefano Zampini 195011955456SStefano Zampini /* for reals, symmetric and hermitian are synonims */ 195111955456SStefano Zampini #if !defined(PETSC_USE_COMPLEX) 195211955456SStefano Zampini sub_schurs->is_symmetric = (PetscBool)(sub_schurs->is_symmetric && sub_schurs->is_hermitian); 195311955456SStefano Zampini sub_schurs->is_hermitian = sub_schurs->is_symmetric; 195411955456SStefano Zampini #endif 195511955456SStefano Zampini 1956b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)is_I);CHKERRQ(ierr); 1957b1b3d7a2SStefano Zampini sub_schurs->is_I = is_I; 1958b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)is_B);CHKERRQ(ierr); 1959b1b3d7a2SStefano Zampini sub_schurs->is_B = is_B; 19605db18549SStefano Zampini ierr = PetscObjectReference((PetscObject)graph->l2gmap);CHKERRQ(ierr); 19615db18549SStefano Zampini sub_schurs->l2gmap = graph->l2gmap; 19625db18549SStefano Zampini ierr = PetscObjectReference((PetscObject)BtoNmap);CHKERRQ(ierr); 19635db18549SStefano Zampini sub_schurs->BtoNmap = BtoNmap; 19645a95e1ceSStefano Zampini sub_schurs->n_subs = n_all_cc; 1965b1b3d7a2SStefano Zampini sub_schurs->is_subs = all_cc; 1966b96c3477SStefano Zampini sub_schurs->S_Ej_all = NULL; 1967b96c3477SStefano Zampini sub_schurs->sum_S_Ej_all = NULL; 196808122e43SStefano Zampini sub_schurs->sum_S_Ej_inv_all = NULL; 1969b96c3477SStefano Zampini sub_schurs->sum_S_Ej_tilda_all = NULL; 1970b96c3477SStefano Zampini sub_schurs->is_Ej_all = NULL; 1971b1b3d7a2SStefano Zampini PetscFunctionReturn(0); 1972b1b3d7a2SStefano Zampini } 1973b1b3d7a2SStefano Zampini 197434a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursCreate(PCBDDCSubSchurs *sub_schurs) 197534a97f8cSStefano Zampini { 197634a97f8cSStefano Zampini PCBDDCSubSchurs schurs_ctx; 197734a97f8cSStefano Zampini PetscErrorCode ierr; 197834a97f8cSStefano Zampini 197934a97f8cSStefano Zampini PetscFunctionBegin; 198034a97f8cSStefano Zampini ierr = PetscNew(&schurs_ctx);CHKERRQ(ierr); 19815ff63025SStefano Zampini schurs_ctx->n_subs = 0; 198234a97f8cSStefano Zampini *sub_schurs = schurs_ctx; 198334a97f8cSStefano Zampini PetscFunctionReturn(0); 198434a97f8cSStefano Zampini } 198534a97f8cSStefano Zampini 198634a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursReset(PCBDDCSubSchurs sub_schurs) 198734a97f8cSStefano Zampini { 198834a97f8cSStefano Zampini PetscInt i; 198934a97f8cSStefano Zampini PetscErrorCode ierr; 199034a97f8cSStefano Zampini 199134a97f8cSStefano Zampini PetscFunctionBegin; 1992aea80f77Sstefano_zampini if (!sub_schurs) PetscFunctionReturn(0); 199388113c35SStefano Zampini ierr = PetscFree(sub_schurs->prefix);CHKERRQ(ierr); 19941e9c79c2SStefano Zampini ierr = MatDestroy(&sub_schurs->A);CHKERRQ(ierr); 1995b1b3d7a2SStefano Zampini ierr = MatDestroy(&sub_schurs->S);CHKERRQ(ierr); 1996b1b3d7a2SStefano Zampini ierr = ISDestroy(&sub_schurs->is_I);CHKERRQ(ierr); 1997b1b3d7a2SStefano Zampini ierr = ISDestroy(&sub_schurs->is_B);CHKERRQ(ierr); 19985db18549SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&sub_schurs->l2gmap);CHKERRQ(ierr); 19995db18549SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&sub_schurs->BtoNmap);CHKERRQ(ierr); 200041c3ba1bSStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 200141c3ba1bSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 200208122e43SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 2003a1337663SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 20045db18549SStefano Zampini ierr = ISDestroy(&sub_schurs->is_Ej_all);CHKERRQ(ierr); 2005d62866d3SStefano Zampini ierr = ISDestroy(&sub_schurs->is_vertices);CHKERRQ(ierr); 2006d62866d3SStefano Zampini ierr = ISDestroy(&sub_schurs->is_dir);CHKERRQ(ierr); 200708122e43SStefano Zampini ierr = PetscBTDestroy(&sub_schurs->is_edge);CHKERRQ(ierr); 200834a97f8cSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 2009b1b3d7a2SStefano Zampini ierr = ISDestroy(&sub_schurs->is_subs[i]);CHKERRQ(ierr); 201034a97f8cSStefano Zampini } 20115ff63025SStefano Zampini if (sub_schurs->n_subs) { 2012b1b3d7a2SStefano Zampini ierr = PetscFree(sub_schurs->is_subs);CHKERRQ(ierr); 20133dc780c3SStefano Zampini } 2014df4d28bfSStefano Zampini if (sub_schurs->reuse_solver) { 2015df4d28bfSStefano Zampini ierr = PCBDDCReuseSolversReset(sub_schurs->reuse_solver);CHKERRQ(ierr); 2016d62866d3SStefano Zampini } 2017df4d28bfSStefano Zampini ierr = PetscFree(sub_schurs->reuse_solver);CHKERRQ(ierr); 201872b8c272SStefano Zampini if (sub_schurs->change) { 201972b8c272SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 202072b8c272SStefano Zampini ierr = KSPDestroy(&sub_schurs->change[i]);CHKERRQ(ierr); 2021b7ab4a40SStefano Zampini ierr = ISDestroy(&sub_schurs->change_primal_sub[i]);CHKERRQ(ierr); 202272b8c272SStefano Zampini } 202372b8c272SStefano Zampini } 202472b8c272SStefano Zampini ierr = PetscFree(sub_schurs->change);CHKERRQ(ierr); 2025b7ab4a40SStefano Zampini ierr = PetscFree(sub_schurs->change_primal_sub);CHKERRQ(ierr); 202634a97f8cSStefano Zampini sub_schurs->n_subs = 0; 202734a97f8cSStefano Zampini PetscFunctionReturn(0); 202834a97f8cSStefano Zampini } 202934a97f8cSStefano Zampini 2030aea80f77Sstefano_zampini PetscErrorCode PCBDDCSubSchursDestroy(PCBDDCSubSchurs* sub_schurs) 2031aea80f77Sstefano_zampini { 2032aea80f77Sstefano_zampini PetscErrorCode ierr; 2033aea80f77Sstefano_zampini 2034aea80f77Sstefano_zampini PetscFunctionBegin; 2035aea80f77Sstefano_zampini ierr = PCBDDCSubSchursReset(*sub_schurs);CHKERRQ(ierr); 2036aea80f77Sstefano_zampini ierr = PetscFree(*sub_schurs);CHKERRQ(ierr); 2037aea80f77Sstefano_zampini PetscFunctionReturn(0); 2038aea80f77Sstefano_zampini } 2039aea80f77Sstefano_zampini 20409fbee547SJacob Faibussowitsch static inline PetscErrorCode PCBDDCAdjGetNextLayer_Private(PetscInt* queue_tip,PetscInt n_prev,PetscBT touched,PetscInt* xadj,PetscInt* adjncy,PetscInt* n_added) 204134a97f8cSStefano Zampini { 204234a97f8cSStefano Zampini PetscInt i,j,n; 204334a97f8cSStefano Zampini PetscErrorCode ierr; 204434a97f8cSStefano Zampini 204534a97f8cSStefano Zampini PetscFunctionBegin; 204634a97f8cSStefano Zampini n = 0; 204734a97f8cSStefano Zampini for (i=-n_prev;i<0;i++) { 204834a97f8cSStefano Zampini PetscInt start_dof = queue_tip[i]; 204934a97f8cSStefano Zampini for (j=xadj[start_dof];j<xadj[start_dof+1];j++) { 205034a97f8cSStefano Zampini PetscInt dof = adjncy[j]; 205134a97f8cSStefano Zampini if (!PetscBTLookup(touched,dof)) { 205234a97f8cSStefano Zampini ierr = PetscBTSet(touched,dof);CHKERRQ(ierr); 205334a97f8cSStefano Zampini queue_tip[n] = dof; 205434a97f8cSStefano Zampini n++; 205534a97f8cSStefano Zampini } 205634a97f8cSStefano Zampini } 205734a97f8cSStefano Zampini } 205834a97f8cSStefano Zampini *n_added = n; 205934a97f8cSStefano Zampini PetscFunctionReturn(0); 206034a97f8cSStefano Zampini } 2061