134a97f8cSStefano Zampini #include <../src/ksp/pc/impls/bddc/bddc.h> 234a97f8cSStefano Zampini #include <../src/ksp/pc/impls/bddc/bddcprivate.h> 308122e43SStefano Zampini #include <petscblaslapack.h> 434a97f8cSStefano Zampini 53202ece2SStefano Zampini PETSC_STATIC_INLINE PetscErrorCode PCBDDCAdjGetNextLayer_Private(PetscInt*,PetscInt,PetscBT,PetscInt*,PetscInt*,PetscInt*); 65ec10c6aSStefano Zampini static PetscErrorCode PCBDDCComputeExplicitSchur(Mat,PetscBool,MatReuse,Mat*); 7d5574798SStefano Zampini static PetscErrorCode PCBDDCMumpsInteriorSolve(PC,Vec,Vec); 8d62866d3SStefano Zampini static PetscErrorCode PCBDDCMumpsCorrectionSolve(PC,Vec,Vec); 9d62866d3SStefano Zampini 10*ca92afb2SStefano Zampini 11*ca92afb2SStefano Zampini #undef __FUNCT__ 12*ca92afb2SStefano Zampini #define __FUNCT__ "SparseRankOneUpdate" 13*ca92afb2SStefano Zampini PETSC_STATIC_INLINE PetscErrorCode SparseRankOneUpdate(PetscScalar *S,PetscInt n,PetscScalar a,PetscScalar *v1,PetscScalar *v2) 14*ca92afb2SStefano Zampini { 15*ca92afb2SStefano Zampini PetscInt i; 16*ca92afb2SStefano Zampini 17*ca92afb2SStefano Zampini PetscFunctionBegin; 18*ca92afb2SStefano Zampini for (i=0;i<n;i++) { 19*ca92afb2SStefano Zampini if (PetscUnlikely(PetscAbsScalar(v2[i]) > PETSC_SMALL)) { 20*ca92afb2SStefano Zampini PetscScalar v2v = a*v2[i]; 21*ca92afb2SStefano Zampini PetscBLASInt B_N = n-i,B_one = 1; 22*ca92afb2SStefano Zampini PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&v2v,v1+i,&B_one,S+i*(n+1),&B_one)); 23*ca92afb2SStefano Zampini } 24*ca92afb2SStefano Zampini } 25*ca92afb2SStefano Zampini PetscFunctionReturn(0); 26*ca92afb2SStefano Zampini } 27*ca92afb2SStefano Zampini 28*ca92afb2SStefano Zampini /* if v2 is not present, correction is done in-place */ 29*ca92afb2SStefano Zampini #undef __FUNCT__ 30*ca92afb2SStefano Zampini #define __FUNCT__ "PCBDDCReuseSolversChangeInterior" 31*ca92afb2SStefano Zampini static PetscErrorCode PCBDDCReuseSolversChangeInterior(PCBDDCReuseMumps ctx, Vec v, Vec v2, PetscBool sol) 32*ca92afb2SStefano Zampini { 33*ca92afb2SStefano Zampini PetscScalar *array; 34*ca92afb2SStefano Zampini PetscScalar *array2; 35*ca92afb2SStefano Zampini PetscErrorCode ierr; 36*ca92afb2SStefano Zampini 37*ca92afb2SStefano Zampini PetscFunctionBegin; 38*ca92afb2SStefano Zampini if (!ctx->benign_n) PetscFunctionReturn(0); 39*ca92afb2SStefano Zampini if (v2) { 40*ca92afb2SStefano Zampini PetscInt nl; 41*ca92afb2SStefano Zampini 42*ca92afb2SStefano Zampini ierr = VecGetArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 43*ca92afb2SStefano Zampini ierr = VecGetLocalSize(v2,&nl);CHKERRQ(ierr); 44*ca92afb2SStefano Zampini ierr = VecGetArray(v2,&array2);CHKERRQ(ierr); 45*ca92afb2SStefano Zampini ierr = PetscMemcpy(array2,array,nl*sizeof(PetscScalar));CHKERRQ(ierr); 46*ca92afb2SStefano Zampini } else { 47*ca92afb2SStefano Zampini ierr = VecGetArray(v,&array);CHKERRQ(ierr); 48*ca92afb2SStefano Zampini array2 = array; 49*ca92afb2SStefano Zampini } 50*ca92afb2SStefano Zampini if (!sol) { /* change rhs */ 51*ca92afb2SStefano Zampini PetscInt n; 52*ca92afb2SStefano Zampini for (n=0;n<ctx->benign_n;n++) { 53*ca92afb2SStefano Zampini PetscScalar sum = 0.; 54*ca92afb2SStefano Zampini const PetscInt *cols; 55*ca92afb2SStefano Zampini PetscInt nz,i; 56*ca92afb2SStefano Zampini 57*ca92afb2SStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[n],&nz);CHKERRQ(ierr); 58*ca92afb2SStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[n],&cols);CHKERRQ(ierr); 59*ca92afb2SStefano Zampini for (i=0;i<nz-1;i++) sum += array[cols[i]]; 60*ca92afb2SStefano Zampini sum = -sum/nz; 61*ca92afb2SStefano Zampini for (i=0;i<nz-1;i++) array2[cols[i]] += sum; 62*ca92afb2SStefano Zampini ctx->benign_save_vals[n] = array2[cols[nz-1]]; 63*ca92afb2SStefano Zampini array2[cols[nz-1]] = sum; 64*ca92afb2SStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[n],&cols);CHKERRQ(ierr); 65*ca92afb2SStefano Zampini } 66*ca92afb2SStefano Zampini } else { 67*ca92afb2SStefano Zampini PetscInt n; 68*ca92afb2SStefano Zampini for (n=0;n<ctx->benign_n;n++) { 69*ca92afb2SStefano Zampini PetscScalar sum = 0.; 70*ca92afb2SStefano Zampini const PetscInt *cols; 71*ca92afb2SStefano Zampini PetscInt nz,i; 72*ca92afb2SStefano Zampini 73*ca92afb2SStefano Zampini ierr = ISGetLocalSize(ctx->benign_zerodiag_subs[n],&nz);CHKERRQ(ierr); 74*ca92afb2SStefano Zampini ierr = ISGetIndices(ctx->benign_zerodiag_subs[n],&cols);CHKERRQ(ierr); 75*ca92afb2SStefano Zampini for (i=0;i<nz-1;i++) sum += array[cols[i]]; 76*ca92afb2SStefano Zampini sum = -sum/nz; 77*ca92afb2SStefano Zampini for (i=0;i<nz-1;i++) array2[cols[i]] += sum; 78*ca92afb2SStefano Zampini array2[cols[nz-1]] = ctx->benign_save_vals[n]; 79*ca92afb2SStefano Zampini ierr = ISRestoreIndices(ctx->benign_zerodiag_subs[n],&cols);CHKERRQ(ierr); 80*ca92afb2SStefano Zampini } 81*ca92afb2SStefano Zampini } 82*ca92afb2SStefano Zampini if (v2) { 83*ca92afb2SStefano Zampini ierr = VecRestoreArrayRead(v,(const PetscScalar**)&array);CHKERRQ(ierr); 84*ca92afb2SStefano Zampini ierr = VecRestoreArray(v2,&array);CHKERRQ(ierr); 85*ca92afb2SStefano Zampini } else { 86*ca92afb2SStefano Zampini ierr = VecRestoreArray(v,&array);CHKERRQ(ierr); 87*ca92afb2SStefano Zampini } 88*ca92afb2SStefano Zampini PetscFunctionReturn(0); 89*ca92afb2SStefano Zampini } 90*ca92afb2SStefano Zampini 91d62866d3SStefano Zampini #undef __FUNCT__ 92683d3df6SStefano Zampini #define __FUNCT__ "PCBDDCMumpsReuseSolve_Private" 93683d3df6SStefano Zampini static PetscErrorCode PCBDDCMumpsReuseSolve_Private(PC pc, Vec rhs, Vec sol, PetscBool transpose, PetscBool full) 94d62866d3SStefano Zampini { 95d62866d3SStefano Zampini PCBDDCReuseMumps ctx; 96683d3df6SStefano Zampini PetscBool copy = PETSC_FALSE; 97d62866d3SStefano Zampini PetscErrorCode ierr; 98d62866d3SStefano Zampini 99d62866d3SStefano Zampini PetscFunctionBegin; 100d62866d3SStefano Zampini ierr = PCShellGetContext(pc,(void **)&ctx);CHKERRQ(ierr); 101683d3df6SStefano Zampini if (full) { 102d62866d3SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 103d62866d3SStefano Zampini ierr = MatMumpsSetIcntl(ctx->F,26,-1);CHKERRQ(ierr); 104d62866d3SStefano Zampini #endif 105683d3df6SStefano Zampini copy = ctx->has_vertices; 106d4933d67SStefano Zampini } else { /* interior solver */ 1076dba178dSStefano Zampini #if defined(PETSC_HAVE_MUMPS) 108683d3df6SStefano Zampini ierr = MatMumpsSetIcntl(ctx->F,26,0);CHKERRQ(ierr); 1096dba178dSStefano Zampini #endif 110d4933d67SStefano Zampini #if defined(PETSC_HAVE_MKL_PARDISO) 111d4933d67SStefano Zampini ierr = MatMkl_PardisoSetCntl(ctx->F,70,1);CHKERRQ(ierr); 112d4933d67SStefano Zampini #endif 113683d3df6SStefano Zampini copy = PETSC_TRUE; 114683d3df6SStefano Zampini } 115683d3df6SStefano Zampini /* copy rhs into factored matrix workspace */ 116683d3df6SStefano Zampini if (copy) { 117*ca92afb2SStefano Zampini PetscInt n; 118*ca92afb2SStefano Zampini PetscScalar *array,*array_mumps; 119*ca92afb2SStefano Zampini 120*ca92afb2SStefano Zampini ierr = VecGetLocalSize(rhs,&n);CHKERRQ(ierr); 121683d3df6SStefano Zampini ierr = VecGetArrayRead(rhs,(const PetscScalar**)&array);CHKERRQ(ierr); 122683d3df6SStefano Zampini ierr = VecGetArray(ctx->rhs,&array_mumps);CHKERRQ(ierr); 123683d3df6SStefano Zampini ierr = PetscMemcpy(array_mumps,array,n*sizeof(PetscScalar));CHKERRQ(ierr); 124683d3df6SStefano Zampini ierr = VecRestoreArray(ctx->rhs,&array_mumps);CHKERRQ(ierr); 125683d3df6SStefano Zampini ierr = VecRestoreArrayRead(rhs,(const PetscScalar**)&array);CHKERRQ(ierr); 126683d3df6SStefano Zampini 127*ca92afb2SStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(ctx,ctx->rhs,NULL,PETSC_FALSE);CHKERRQ(ierr); 128683d3df6SStefano Zampini if (transpose) { 129683d3df6SStefano Zampini ierr = MatSolveTranspose(ctx->F,ctx->rhs,ctx->sol);CHKERRQ(ierr); 130683d3df6SStefano Zampini } else { 131683d3df6SStefano Zampini ierr = MatSolve(ctx->F,ctx->rhs,ctx->sol);CHKERRQ(ierr); 132683d3df6SStefano Zampini } 133*ca92afb2SStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(ctx,ctx->sol,NULL,PETSC_TRUE);CHKERRQ(ierr); 134683d3df6SStefano Zampini 135683d3df6SStefano Zampini /* get back data to caller worskpace */ 136683d3df6SStefano Zampini ierr = VecGetArrayRead(ctx->sol,(const PetscScalar**)&array_mumps);CHKERRQ(ierr); 137683d3df6SStefano Zampini ierr = VecGetArray(sol,&array);CHKERRQ(ierr); 138683d3df6SStefano Zampini ierr = PetscMemcpy(array,array_mumps,n*sizeof(PetscScalar));CHKERRQ(ierr); 139683d3df6SStefano Zampini ierr = VecRestoreArray(sol,&array);CHKERRQ(ierr); 140683d3df6SStefano Zampini ierr = VecRestoreArrayRead(ctx->sol,(const PetscScalar**)&array_mumps);CHKERRQ(ierr); 141683d3df6SStefano Zampini } else { 142*ca92afb2SStefano Zampini if (ctx->benign_n) { 143*ca92afb2SStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(ctx,rhs,ctx->rhs,PETSC_FALSE);CHKERRQ(ierr); 144*ca92afb2SStefano Zampini if (transpose) { 145*ca92afb2SStefano Zampini ierr = MatSolveTranspose(ctx->F,ctx->rhs,sol);CHKERRQ(ierr); 146*ca92afb2SStefano Zampini } else { 147*ca92afb2SStefano Zampini ierr = MatSolve(ctx->F,ctx->rhs,sol);CHKERRQ(ierr); 148*ca92afb2SStefano Zampini } 149*ca92afb2SStefano Zampini ierr = PCBDDCReuseSolversChangeInterior(ctx,sol,NULL,PETSC_TRUE);CHKERRQ(ierr); 150*ca92afb2SStefano Zampini } else { 151e28d306cSStefano Zampini if (transpose) { 152e28d306cSStefano Zampini ierr = MatSolveTranspose(ctx->F,rhs,sol);CHKERRQ(ierr); 153e28d306cSStefano Zampini } else { 1546816873aSStefano Zampini ierr = MatSolve(ctx->F,rhs,sol);CHKERRQ(ierr); 155e28d306cSStefano Zampini } 156683d3df6SStefano Zampini } 157*ca92afb2SStefano Zampini } 158d4933d67SStefano Zampini #if defined(PETSC_HAVE_MKL_PARDISO) 159d4933d67SStefano Zampini ierr = MatMkl_PardisoSetCntl(ctx->F,70,0);CHKERRQ(ierr); 160d4933d67SStefano Zampini #endif 161d62866d3SStefano Zampini PetscFunctionReturn(0); 162d62866d3SStefano Zampini } 163d62866d3SStefano Zampini 164d62866d3SStefano Zampini #undef __FUNCT__ 165e28d306cSStefano Zampini #define __FUNCT__ "PCBDDCMumpsCorrectionSolve" 166e28d306cSStefano Zampini static PetscErrorCode PCBDDCMumpsCorrectionSolve(PC pc, Vec rhs, Vec sol) 167e28d306cSStefano Zampini { 168e28d306cSStefano Zampini PetscErrorCode ierr; 169e28d306cSStefano Zampini 170e28d306cSStefano Zampini PetscFunctionBegin; 171683d3df6SStefano Zampini ierr = PCBDDCMumpsReuseSolve_Private(pc,rhs,sol,PETSC_FALSE,PETSC_TRUE);CHKERRQ(ierr); 172e28d306cSStefano Zampini PetscFunctionReturn(0); 173e28d306cSStefano Zampini } 174e28d306cSStefano Zampini 175e28d306cSStefano Zampini #undef __FUNCT__ 176e28d306cSStefano Zampini #define __FUNCT__ "PCBDDCMumpsCorrectionSolveTranspose" 177e28d306cSStefano Zampini static PetscErrorCode PCBDDCMumpsCorrectionSolveTranspose(PC pc, Vec rhs, Vec sol) 178e28d306cSStefano Zampini { 179e28d306cSStefano Zampini PetscErrorCode ierr; 180e28d306cSStefano Zampini 181e28d306cSStefano Zampini PetscFunctionBegin; 182683d3df6SStefano Zampini ierr = PCBDDCMumpsReuseSolve_Private(pc,rhs,sol,PETSC_TRUE,PETSC_TRUE);CHKERRQ(ierr); 183683d3df6SStefano Zampini PetscFunctionReturn(0); 184683d3df6SStefano Zampini } 185683d3df6SStefano Zampini 186683d3df6SStefano Zampini #undef __FUNCT__ 187683d3df6SStefano Zampini #define __FUNCT__ "PCBDDCMumpsInteriorSolve" 188683d3df6SStefano Zampini static PetscErrorCode PCBDDCMumpsInteriorSolve(PC pc, Vec rhs, Vec sol) 189683d3df6SStefano Zampini { 190683d3df6SStefano Zampini PetscErrorCode ierr; 191683d3df6SStefano Zampini 192683d3df6SStefano Zampini PetscFunctionBegin; 193683d3df6SStefano Zampini ierr = PCBDDCMumpsReuseSolve_Private(pc,rhs,sol,PETSC_FALSE,PETSC_FALSE);CHKERRQ(ierr); 194683d3df6SStefano Zampini PetscFunctionReturn(0); 195683d3df6SStefano Zampini } 196683d3df6SStefano Zampini 197683d3df6SStefano Zampini #undef __FUNCT__ 198683d3df6SStefano Zampini #define __FUNCT__ "PCBDDCMumpsInteriorSolveTranspose" 199683d3df6SStefano Zampini static PetscErrorCode PCBDDCMumpsInteriorSolveTranspose(PC pc, Vec rhs, Vec sol) 200683d3df6SStefano Zampini { 201683d3df6SStefano Zampini PetscErrorCode ierr; 202683d3df6SStefano Zampini 203683d3df6SStefano Zampini PetscFunctionBegin; 204683d3df6SStefano Zampini ierr = PCBDDCMumpsReuseSolve_Private(pc,rhs,sol,PETSC_TRUE,PETSC_FALSE);CHKERRQ(ierr); 205e28d306cSStefano Zampini PetscFunctionReturn(0); 206e28d306cSStefano Zampini } 207e28d306cSStefano Zampini 208e28d306cSStefano Zampini #undef __FUNCT__ 209d62866d3SStefano Zampini #define __FUNCT__ "PCBDDCReuseMumpsReset" 210d62866d3SStefano Zampini static PetscErrorCode PCBDDCReuseMumpsReset(PCBDDCReuseMumps reuse) 211d62866d3SStefano Zampini { 212*ca92afb2SStefano Zampini PetscInt i; 213d62866d3SStefano Zampini PetscErrorCode ierr; 214d62866d3SStefano Zampini 215d62866d3SStefano Zampini PetscFunctionBegin; 216d62866d3SStefano Zampini ierr = MatDestroy(&reuse->F);CHKERRQ(ierr); 217d62866d3SStefano Zampini ierr = VecDestroy(&reuse->sol);CHKERRQ(ierr); 218d62866d3SStefano Zampini ierr = VecDestroy(&reuse->rhs);CHKERRQ(ierr); 219d62866d3SStefano Zampini ierr = PCDestroy(&reuse->interior_solver);CHKERRQ(ierr); 220d62866d3SStefano Zampini ierr = PCDestroy(&reuse->correction_solver);CHKERRQ(ierr); 22153892102SStefano Zampini ierr = ISDestroy(&reuse->is_R);CHKERRQ(ierr); 22253892102SStefano Zampini ierr = ISDestroy(&reuse->is_B);CHKERRQ(ierr); 223d62866d3SStefano Zampini ierr = VecScatterDestroy(&reuse->correction_scatter_B);CHKERRQ(ierr); 22453892102SStefano Zampini ierr = VecDestroy(&reuse->sol_B);CHKERRQ(ierr); 22553892102SStefano Zampini ierr = VecDestroy(&reuse->rhs_B);CHKERRQ(ierr); 226*ca92afb2SStefano Zampini for (i=0;i<reuse->benign_n;i++) { 227*ca92afb2SStefano Zampini ierr = ISDestroy(&reuse->benign_zerodiag_subs[i]);CHKERRQ(ierr); 228*ca92afb2SStefano Zampini } 229*ca92afb2SStefano Zampini ierr = PetscFree(reuse->benign_zerodiag_subs);CHKERRQ(ierr); 230*ca92afb2SStefano Zampini ierr = PetscFree(reuse->benign_save_vals);CHKERRQ(ierr); 231d62866d3SStefano Zampini PetscFunctionReturn(0); 232d62866d3SStefano Zampini } 233d5574798SStefano Zampini 234d5574798SStefano Zampini #undef __FUNCT__ 2353202ece2SStefano Zampini #define __FUNCT__ "PCBDDCComputeExplicitSchur" 2365ec10c6aSStefano Zampini static PetscErrorCode PCBDDCComputeExplicitSchur(Mat M, PetscBool issym, MatReuse reuse, Mat *S) 2373202ece2SStefano Zampini { 2383202ece2SStefano Zampini Mat B, C, D, Bd, Cd, AinvBd; 2393202ece2SStefano Zampini KSP ksp; 2403202ece2SStefano Zampini PC pc; 2413202ece2SStefano Zampini PetscBool isLU, isILU, isCHOL, Bdense, Cdense; 2423202ece2SStefano Zampini PetscReal fill = 2.0; 243f11841e3SStefano Zampini PetscInt n_I; 2443202ece2SStefano Zampini PetscMPIInt size; 2453202ece2SStefano Zampini PetscErrorCode ierr; 2463202ece2SStefano Zampini 2473202ece2SStefano Zampini PetscFunctionBegin; 2483202ece2SStefano Zampini ierr = MPI_Comm_size(PetscObjectComm((PetscObject)M),&size);CHKERRQ(ierr); 2493202ece2SStefano Zampini if (size != 1) { 2503202ece2SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not for parallel matrices"); 2513202ece2SStefano Zampini } 252f11841e3SStefano Zampini if (reuse == MAT_REUSE_MATRIX) { 253f11841e3SStefano Zampini PetscBool Sdense; 254f11841e3SStefano Zampini 255f11841e3SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)*S, MATSEQDENSE, &Sdense);CHKERRQ(ierr); 256f11841e3SStefano Zampini if (!Sdense) { 257683d3df6SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)M),PETSC_ERR_SUP,"S should be dense"); 258f11841e3SStefano Zampini } 259f11841e3SStefano Zampini } 2603202ece2SStefano Zampini ierr = MatSchurComplementGetSubMatrices(M, NULL, NULL, &B, &C, &D);CHKERRQ(ierr); 2613202ece2SStefano Zampini ierr = MatSchurComplementGetKSP(M, &ksp);CHKERRQ(ierr); 2623202ece2SStefano Zampini ierr = KSPGetPC(ksp, &pc);CHKERRQ(ierr); 2633202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject) pc, PCLU, &isLU);CHKERRQ(ierr); 2643202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject) pc, PCILU, &isILU);CHKERRQ(ierr); 2653202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject) pc, PCCHOLESKY, &isCHOL);CHKERRQ(ierr); 2663202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject) B, MATSEQDENSE, &Bdense);CHKERRQ(ierr); 2673202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject) C, MATSEQDENSE, &Cdense);CHKERRQ(ierr); 268f11841e3SStefano Zampini ierr = MatGetSize(B,&n_I,NULL);CHKERRQ(ierr); 269f11841e3SStefano Zampini if (n_I) { 2703202ece2SStefano Zampini if (!Bdense) { 2713202ece2SStefano Zampini ierr = MatConvert(B, MATSEQDENSE, MAT_INITIAL_MATRIX, &Bd);CHKERRQ(ierr); 2723202ece2SStefano Zampini } else { 2733202ece2SStefano Zampini Bd = B; 2743202ece2SStefano Zampini } 2753202ece2SStefano Zampini 2763202ece2SStefano Zampini if (isLU || isILU || isCHOL) { 2773202ece2SStefano Zampini Mat fact; 2783202ece2SStefano Zampini ierr = KSPSetUp(ksp);CHKERRQ(ierr); 2793202ece2SStefano Zampini ierr = PCFactorGetMatrix(pc, &fact);CHKERRQ(ierr); 2803202ece2SStefano Zampini ierr = MatDuplicate(Bd, MAT_DO_NOT_COPY_VALUES, &AinvBd);CHKERRQ(ierr); 2813202ece2SStefano Zampini ierr = MatMatSolve(fact, Bd, AinvBd);CHKERRQ(ierr); 2823202ece2SStefano Zampini } else { 28307b1e237SStefano Zampini PetscBool ex = PETSC_TRUE; 28407b1e237SStefano Zampini 28507b1e237SStefano Zampini if (ex) { 2863202ece2SStefano Zampini Mat Ainvd; 2873202ece2SStefano Zampini 2883202ece2SStefano Zampini ierr = PCComputeExplicitOperator(pc, &Ainvd);CHKERRQ(ierr); 2893202ece2SStefano Zampini ierr = MatMatMult(Ainvd, Bd, MAT_INITIAL_MATRIX, fill, &AinvBd);CHKERRQ(ierr); 2903202ece2SStefano Zampini ierr = MatDestroy(&Ainvd);CHKERRQ(ierr); 29107b1e237SStefano Zampini } else { 29207b1e237SStefano Zampini Vec sol,rhs; 29307b1e237SStefano Zampini PetscScalar *arrayrhs,*arraysol; 29407b1e237SStefano Zampini PetscInt i,nrhs,n; 29507b1e237SStefano Zampini 29607b1e237SStefano Zampini ierr = MatDuplicate(Bd, MAT_DO_NOT_COPY_VALUES, &AinvBd);CHKERRQ(ierr); 29707b1e237SStefano Zampini ierr = MatGetSize(Bd,&n,&nrhs);CHKERRQ(ierr); 29807b1e237SStefano Zampini ierr = MatDenseGetArray(Bd,&arrayrhs);CHKERRQ(ierr); 29907b1e237SStefano Zampini ierr = MatDenseGetArray(AinvBd,&arraysol);CHKERRQ(ierr); 30007b1e237SStefano Zampini ierr = KSPGetSolution(ksp,&sol);CHKERRQ(ierr); 30107b1e237SStefano Zampini ierr = KSPGetRhs(ksp,&rhs);CHKERRQ(ierr); 30207b1e237SStefano Zampini for (i=0;i<nrhs;i++) { 30307b1e237SStefano Zampini ierr = VecPlaceArray(rhs,arrayrhs+i*n);CHKERRQ(ierr); 30407b1e237SStefano Zampini ierr = VecPlaceArray(sol,arraysol+i*n);CHKERRQ(ierr); 30507b1e237SStefano Zampini ierr = KSPSolve(ksp,rhs,sol);CHKERRQ(ierr); 30607b1e237SStefano Zampini ierr = VecResetArray(rhs);CHKERRQ(ierr); 30707b1e237SStefano Zampini ierr = VecResetArray(sol);CHKERRQ(ierr); 30807b1e237SStefano Zampini } 30907b1e237SStefano Zampini ierr = MatDenseRestoreArray(Bd,&arrayrhs);CHKERRQ(ierr); 31007b1e237SStefano Zampini ierr = MatDenseRestoreArray(AinvBd,&arrayrhs);CHKERRQ(ierr); 31107b1e237SStefano Zampini } 3123202ece2SStefano Zampini } 3135ec10c6aSStefano Zampini if (!Bdense & !issym) { 3143202ece2SStefano Zampini ierr = MatDestroy(&Bd);CHKERRQ(ierr); 3153202ece2SStefano Zampini } 3165ec10c6aSStefano Zampini 3175ec10c6aSStefano Zampini if (!issym) { 3183202ece2SStefano Zampini if (!Cdense) { 3193202ece2SStefano Zampini ierr = MatConvert(C, MATSEQDENSE, MAT_INITIAL_MATRIX, &Cd);CHKERRQ(ierr); 3203202ece2SStefano Zampini } else { 3213202ece2SStefano Zampini Cd = C; 3223202ece2SStefano Zampini } 3235ec10c6aSStefano Zampini ierr = MatMatMult(Cd, AinvBd, reuse, fill, S);CHKERRQ(ierr); 3243202ece2SStefano Zampini if (!Cdense) { 3253202ece2SStefano Zampini ierr = MatDestroy(&Cd);CHKERRQ(ierr); 3263202ece2SStefano Zampini } 3275ec10c6aSStefano Zampini } else { 3285ec10c6aSStefano Zampini ierr = MatTransposeMatMult(Bd, AinvBd, reuse, fill, S);CHKERRQ(ierr); 3295ec10c6aSStefano Zampini if (!Bdense) { 3305ec10c6aSStefano Zampini ierr = MatDestroy(&Bd);CHKERRQ(ierr); 3315ec10c6aSStefano Zampini } 3325ec10c6aSStefano Zampini } 3335ec10c6aSStefano Zampini ierr = MatDestroy(&AinvBd);CHKERRQ(ierr); 334f11841e3SStefano Zampini } 3353202ece2SStefano Zampini 3363202ece2SStefano Zampini if (D) { 3373202ece2SStefano Zampini Mat Dd; 3383202ece2SStefano Zampini PetscBool Ddense; 3393202ece2SStefano Zampini 3403202ece2SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)D,MATSEQDENSE,&Ddense);CHKERRQ(ierr); 3413202ece2SStefano Zampini if (!Ddense) { 3423202ece2SStefano Zampini ierr = MatConvert(D, MATSEQDENSE, MAT_INITIAL_MATRIX, &Dd);CHKERRQ(ierr); 3433202ece2SStefano Zampini } else { 3443202ece2SStefano Zampini Dd = D; 3453202ece2SStefano Zampini } 346f11841e3SStefano Zampini if (n_I) { 3473202ece2SStefano Zampini ierr = MatAYPX(*S,-1.0,Dd,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 348f11841e3SStefano Zampini } else { 349f11841e3SStefano Zampini if (reuse == MAT_INITIAL_MATRIX) { 350f11841e3SStefano Zampini ierr = MatDuplicate(Dd,MAT_COPY_VALUES,S);CHKERRQ(ierr); 351f11841e3SStefano Zampini } else { 352f11841e3SStefano Zampini ierr = MatCopy(Dd,*S,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 353f11841e3SStefano Zampini } 354f11841e3SStefano Zampini } 3553202ece2SStefano Zampini if (!Ddense) { 3563202ece2SStefano Zampini ierr = MatDestroy(&Dd);CHKERRQ(ierr); 3573202ece2SStefano Zampini } 3583202ece2SStefano Zampini } else { 3593202ece2SStefano Zampini ierr = MatScale(*S,-1.0);CHKERRQ(ierr); 3603202ece2SStefano Zampini } 3613202ece2SStefano Zampini PetscFunctionReturn(0); 3623202ece2SStefano Zampini } 36334a97f8cSStefano Zampini 36434a97f8cSStefano Zampini #undef __FUNCT__ 3651580ed26SStefano Zampini #define __FUNCT__ "PCBDDCSubSchursSetUp" 366*ca92afb2SStefano Zampini PetscErrorCode PCBDDCSubSchursSetUp(PCBDDCSubSchurs sub_schurs, Mat Ain, Mat Sin, PetscBool exact_schur, PetscInt xadj[], PetscInt adjncy[], PetscInt nlayers, PetscBool faster_deluxe, PetscBool compute_Stilda, PetscBool reuse_solvers, PetscBool benign_trick, PetscInt benign_n, PetscInt benign_p0_lidx[], IS benign_zerodiag_subs[]) 367b1b3d7a2SStefano Zampini { 368be83ff47SStefano Zampini Mat F,A_II,A_IB,A_BI,A_BB,AE_II; 369be83ff47SStefano Zampini Mat S_all; 3705a94c880SStefano Zampini Mat global_schur_subsets,work_mat,*submats; 3715db18549SStefano Zampini ISLocalToGlobalMapping l2gmap_subsets; 372dc456d91SStefano Zampini IS is_I,is_I_layer; 373dc456d91SStefano Zampini IS all_subsets,all_subsets_mult,all_subsets_n; 374dc456d91SStefano Zampini PetscInt *nnz,*all_local_idx_N; 375dc456d91SStefano Zampini PetscInt *auxnum1,*auxnum2; 3765a95e1ceSStefano Zampini PetscInt i,subset_size,max_subset_size; 377683d3df6SStefano Zampini PetscInt n_B,extra,local_size,global_size; 37808122e43SStefano Zampini PetscBLASInt B_N,B_ierr,B_lwork,*pivots; 37906a4b1faSStefano Zampini PetscScalar *Bwork; 3805a95e1ceSStefano Zampini PetscSubcomm subcomm; 3815a95e1ceSStefano Zampini PetscMPIInt color,rank; 3825a95e1ceSStefano Zampini MPI_Comm comm_n; 38306a4e24aSStefano Zampini PetscBool use_getr = PETSC_FALSE; 384b1b3d7a2SStefano Zampini PetscErrorCode ierr; 385b1b3d7a2SStefano Zampini 386b1b3d7a2SStefano Zampini PetscFunctionBegin; 387a64f4aa4SStefano Zampini /* update info in sub_schurs */ 388a64f4aa4SStefano Zampini ierr = MatDestroy(&sub_schurs->A);CHKERRQ(ierr); 389a64f4aa4SStefano Zampini ierr = MatDestroy(&sub_schurs->S);CHKERRQ(ierr); 3903301b35fSStefano Zampini sub_schurs->is_hermitian = PETSC_FALSE; 3913301b35fSStefano Zampini sub_schurs->is_posdef = PETSC_FALSE; 392a64f4aa4SStefano Zampini if (Ain) { 393a64f4aa4SStefano Zampini PetscBool isseqaij; 3943301b35fSStefano Zampini /* determine if we are dealing with hermitian positive definite problems */ 3953301b35fSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 3963301b35fSStefano Zampini if (Ain->symmetric_set) { 3973301b35fSStefano Zampini sub_schurs->is_hermitian = Ain->symmetric; 3983301b35fSStefano Zampini } 3993301b35fSStefano Zampini #else 4003301b35fSStefano Zampini if (Ain->hermitian_set) { 4013301b35fSStefano Zampini sub_schurs->is_hermitian = Ain->hermitian; 4023301b35fSStefano Zampini } 4033301b35fSStefano Zampini #endif 4043301b35fSStefano Zampini if (Ain->spd_set) { 4053301b35fSStefano Zampini sub_schurs->is_posdef = Ain->spd; 4063301b35fSStefano Zampini } 407a64f4aa4SStefano Zampini 4083301b35fSStefano Zampini /* check */ 409a64f4aa4SStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)Ain,MATSEQAIJ,&isseqaij);CHKERRQ(ierr); 4101dac5d76SStefano Zampini if (compute_Stilda && (!Ain->hermitian_set || !Ain->spd_set)) { /* these are lazy checks, maybe I should throw an error if they are not set */ 4113301b35fSStefano Zampini PetscInt lsize; 4123301b35fSStefano Zampini 4133301b35fSStefano Zampini ierr = MatGetSize(Ain,&lsize,NULL);CHKERRQ(ierr); 4143301b35fSStefano Zampini if (lsize) { 4153301b35fSStefano Zampini PetscScalar val; 4163301b35fSStefano Zampini PetscReal norm; 4173301b35fSStefano Zampini Vec vec1,vec2,vec3; 4183301b35fSStefano Zampini 4193301b35fSStefano Zampini ierr = MatCreateVecs(Ain,&vec1,&vec2);CHKERRQ(ierr); 4203301b35fSStefano Zampini ierr = VecDuplicate(vec1,&vec3);CHKERRQ(ierr); 4213301b35fSStefano Zampini ierr = VecSetRandom(vec1,NULL);CHKERRQ(ierr); 4223301b35fSStefano Zampini ierr = MatMult(Ain,vec1,vec2);CHKERRQ(ierr); 4233301b35fSStefano Zampini #if !defined(PETSC_USE_COMPLEX) 4243301b35fSStefano Zampini ierr = MatMultTranspose(Ain,vec1,vec3);CHKERRQ(ierr); 4253301b35fSStefano Zampini #else 4263301b35fSStefano Zampini ierr = MatMultHermitianTranspose(Ain,vec1,vec3);CHKERRQ(ierr); 4273301b35fSStefano Zampini #endif 4283301b35fSStefano Zampini ierr = VecAXPY(vec3,-1.0,vec2);CHKERRQ(ierr); 4293301b35fSStefano Zampini ierr = VecNorm(vec3,NORM_INFINITY,&norm);CHKERRQ(ierr); 4301dac5d76SStefano Zampini if (!Ain->hermitian_set) { 431862806e4SStefano Zampini if (norm > PetscSqrtReal(PETSC_SMALL)) { 4323301b35fSStefano Zampini sub_schurs->is_hermitian = PETSC_FALSE; 4333301b35fSStefano Zampini } else { 4343301b35fSStefano Zampini sub_schurs->is_hermitian = PETSC_TRUE; 4353301b35fSStefano Zampini } 4361dac5d76SStefano Zampini } 4371dac5d76SStefano Zampini if (!Ain->spd_set && !benign_trick) { 4383301b35fSStefano Zampini ierr = VecDot(vec1,vec2,&val);CHKERRQ(ierr); 4393301b35fSStefano Zampini if (PetscRealPart(val) > 0. && PetscAbsReal(PetscImaginaryPart(val)) < PETSC_SMALL) sub_schurs->is_posdef = PETSC_TRUE; 44006a4e24aSStefano Zampini } 4413301b35fSStefano Zampini ierr = VecDestroy(&vec1);CHKERRQ(ierr); 4423301b35fSStefano Zampini ierr = VecDestroy(&vec2);CHKERRQ(ierr); 4433301b35fSStefano Zampini ierr = VecDestroy(&vec3);CHKERRQ(ierr); 4443301b35fSStefano Zampini } else { 4453301b35fSStefano Zampini sub_schurs->is_hermitian = PETSC_TRUE; 4463301b35fSStefano Zampini sub_schurs->is_posdef = PETSC_TRUE; 4473301b35fSStefano Zampini } 4483301b35fSStefano Zampini } 449a64f4aa4SStefano Zampini if (isseqaij) { 450a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)Ain);CHKERRQ(ierr); 451a64f4aa4SStefano Zampini sub_schurs->A = Ain; 4523301b35fSStefano Zampini } else { 453a64f4aa4SStefano Zampini ierr = MatConvert(Ain,MATSEQAIJ,MAT_INITIAL_MATRIX,&sub_schurs->A);CHKERRQ(ierr); 454a64f4aa4SStefano Zampini } 455a64f4aa4SStefano Zampini } 4563301b35fSStefano Zampini 457a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)Sin);CHKERRQ(ierr); 458a64f4aa4SStefano Zampini sub_schurs->S = Sin; 459a64f4aa4SStefano Zampini if (sub_schurs->use_mumps) { 460a64f4aa4SStefano Zampini sub_schurs->use_mumps = (PetscBool)(!!sub_schurs->A); 461a64f4aa4SStefano Zampini } 462a64f4aa4SStefano Zampini 4635a95e1ceSStefano Zampini /* preliminary checks */ 4645a95e1ceSStefano Zampini if (!sub_schurs->use_mumps && compute_Stilda) { 4655a95e1ceSStefano Zampini SETERRQ(PetscObjectComm((PetscObject)sub_schurs->l2gmap),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS"); 4665a95e1ceSStefano Zampini } 4675a95e1ceSStefano Zampini 4685a95e1ceSStefano Zampini /* restrict work on active processes */ 4695a95e1ceSStefano Zampini color = 0; 4705a95e1ceSStefano Zampini if (!sub_schurs->n_subs) color = 1; /* this can happen if we are in a multilevel case or if the subdomain is disconnected */ 4715a95e1ceSStefano Zampini ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)sub_schurs->l2gmap),&rank);CHKERRQ(ierr); 4725a95e1ceSStefano Zampini ierr = PetscSubcommCreate(PetscObjectComm((PetscObject)sub_schurs->l2gmap),&subcomm);CHKERRQ(ierr); 4735a95e1ceSStefano Zampini ierr = PetscSubcommSetNumber(subcomm,2);CHKERRQ(ierr); 4745a95e1ceSStefano Zampini ierr = PetscSubcommSetTypeGeneral(subcomm,color,rank);CHKERRQ(ierr); 4755a95e1ceSStefano Zampini ierr = PetscCommDuplicate(PetscSubcommChild(subcomm),&comm_n,NULL);CHKERRQ(ierr); 4765a95e1ceSStefano Zampini ierr = PetscSubcommDestroy(&subcomm);CHKERRQ(ierr); 4775a95e1ceSStefano Zampini if (!sub_schurs->n_subs) { 4785a95e1ceSStefano Zampini ierr = PetscCommDestroy(&comm_n);CHKERRQ(ierr); 4795a95e1ceSStefano Zampini PetscFunctionReturn(0); 4805a95e1ceSStefano Zampini } 48181ea8064SStefano Zampini /* ierr = PetscCommDuplicate(PetscObjectComm((PetscObject)sub_schurs->l2gmap),&comm_n,NULL);CHKERRQ(ierr); */ 4825a95e1ceSStefano Zampini 483b1b3d7a2SStefano Zampini /* get Schur complement matrices */ 484883469d8SStefano Zampini if (!sub_schurs->use_mumps) { 485a64f4aa4SStefano Zampini Mat tA_IB,tA_BI,tA_BB; 4863301b35fSStefano Zampini PetscBool isseqsbaij; 487a64f4aa4SStefano Zampini ierr = MatSchurComplementGetSubMatrices(sub_schurs->S,&A_II,NULL,&tA_IB,&tA_BI,&tA_BB);CHKERRQ(ierr); 4883301b35fSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)tA_BB,MATSEQSBAIJ,&isseqsbaij);CHKERRQ(ierr); 4893301b35fSStefano Zampini if (isseqsbaij) { 490a64f4aa4SStefano Zampini ierr = MatConvert(tA_BB,MATSEQAIJ,MAT_INITIAL_MATRIX,&A_BB);CHKERRQ(ierr); 491a64f4aa4SStefano Zampini ierr = MatConvert(tA_IB,MATSEQAIJ,MAT_INITIAL_MATRIX,&A_IB);CHKERRQ(ierr); 492a64f4aa4SStefano Zampini ierr = MatConvert(tA_BI,MATSEQAIJ,MAT_INITIAL_MATRIX,&A_BI);CHKERRQ(ierr); 493a64f4aa4SStefano Zampini } else { 494a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)tA_BB);CHKERRQ(ierr); 495a64f4aa4SStefano Zampini A_BB = tA_BB; 496a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)tA_IB);CHKERRQ(ierr); 497a64f4aa4SStefano Zampini A_IB = tA_IB; 498a64f4aa4SStefano Zampini ierr = PetscObjectReference((PetscObject)tA_BI);CHKERRQ(ierr); 499a64f4aa4SStefano Zampini A_BI = tA_BI; 500f11841e3SStefano Zampini } 501a58a30b4SStefano Zampini } else { 5025a95e1ceSStefano Zampini A_II = NULL; 5035a95e1ceSStefano Zampini A_IB = NULL; 5045a95e1ceSStefano Zampini A_BI = NULL; 5055a95e1ceSStefano Zampini A_BB = NULL; 506b1b3d7a2SStefano Zampini } 5075a95e1ceSStefano Zampini S_all = NULL; 508b1b3d7a2SStefano Zampini 509b1b3d7a2SStefano Zampini /* determine interior problems */ 5103dc780c3SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I,&i);CHKERRQ(ierr); 5113dc780c3SStefano Zampini if (nlayers >= 0 && i) { /* Interior problems can be different from the original one */ 512b1b3d7a2SStefano Zampini PetscBT touched; 513b1b3d7a2SStefano Zampini const PetscInt* idx_B; 514b1b3d7a2SStefano Zampini PetscInt n_I,n_B,n_local_dofs,n_prev_added,j,layer,*local_numbering; 515b1b3d7a2SStefano Zampini 5163dc780c3SStefano Zampini if (xadj == NULL || adjncy == NULL) { 5173dc780c3SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot request layering without adjacency"); 5183dc780c3SStefano Zampini } 519b1b3d7a2SStefano Zampini /* get sizes */ 520b1b3d7a2SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I,&n_I);CHKERRQ(ierr); 521b1b3d7a2SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_B,&n_B);CHKERRQ(ierr); 522b1b3d7a2SStefano Zampini 523b1b3d7a2SStefano Zampini ierr = PetscMalloc1(n_I+n_B,&local_numbering);CHKERRQ(ierr); 524b1b3d7a2SStefano Zampini ierr = PetscBTCreate(n_I+n_B,&touched);CHKERRQ(ierr); 525b1b3d7a2SStefano Zampini ierr = PetscBTMemzero(n_I+n_B,touched);CHKERRQ(ierr); 526b1b3d7a2SStefano Zampini 527b1b3d7a2SStefano Zampini /* all boundary dofs must be skipped when adding layers */ 528b1b3d7a2SStefano Zampini ierr = ISGetIndices(sub_schurs->is_B,&idx_B);CHKERRQ(ierr); 529b1b3d7a2SStefano Zampini for (j=0;j<n_B;j++) { 530b1b3d7a2SStefano Zampini ierr = PetscBTSet(touched,idx_B[j]);CHKERRQ(ierr); 531b1b3d7a2SStefano Zampini } 532b1b3d7a2SStefano Zampini ierr = PetscMemcpy(local_numbering,idx_B,n_B*sizeof(PetscInt));CHKERRQ(ierr); 533b1b3d7a2SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_B,&idx_B);CHKERRQ(ierr); 534b1b3d7a2SStefano Zampini 535b1b3d7a2SStefano Zampini /* add prescribed number of layers of dofs */ 536b1b3d7a2SStefano Zampini n_local_dofs = n_B; 537b1b3d7a2SStefano Zampini n_prev_added = n_B; 538b1b3d7a2SStefano Zampini for (layer=0;layer<nlayers;layer++) { 539b1b3d7a2SStefano Zampini PetscInt n_added; 540b1b3d7a2SStefano Zampini if (n_local_dofs == n_I+n_B) break; 541b1b3d7a2SStefano Zampini if (n_local_dofs > n_I+n_B) { 542b1b3d7a2SStefano Zampini SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error querying layer %d. Out of bound access (%d > %d)",layer,n_local_dofs,n_I+n_B); 543b1b3d7a2SStefano Zampini } 544b1b3d7a2SStefano Zampini ierr = PCBDDCAdjGetNextLayer_Private(local_numbering+n_local_dofs,n_prev_added,touched,xadj,adjncy,&n_added);CHKERRQ(ierr); 545b1b3d7a2SStefano Zampini n_prev_added = n_added; 546b1b3d7a2SStefano Zampini n_local_dofs += n_added; 547b1b3d7a2SStefano Zampini if (!n_added) break; 548b1b3d7a2SStefano Zampini } 549b1b3d7a2SStefano Zampini ierr = PetscBTDestroy(&touched);CHKERRQ(ierr); 550b1b3d7a2SStefano Zampini 551883469d8SStefano Zampini /* IS for I layer dofs in original numbering */ 552a9b99552SStefano 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); 553b1b3d7a2SStefano Zampini ierr = PetscFree(local_numbering);CHKERRQ(ierr); 554a9b99552SStefano Zampini ierr = ISSort(is_I_layer);CHKERRQ(ierr); 555883469d8SStefano Zampini /* IS for I layer dofs in I numbering */ 556883469d8SStefano Zampini if (!sub_schurs->use_mumps) { 557b1b3d7a2SStefano Zampini ISLocalToGlobalMapping ItoNmap; 558b1b3d7a2SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(sub_schurs->is_I,&ItoNmap);CHKERRQ(ierr); 559a9b99552SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(ItoNmap,IS_GTOLM_DROP,is_I_layer,&is_I);CHKERRQ(ierr); 560b1b3d7a2SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&ItoNmap);CHKERRQ(ierr); 561b1b3d7a2SStefano Zampini 562b1b3d7a2SStefano Zampini /* II block */ 563b1b3d7a2SStefano Zampini ierr = MatGetSubMatrix(A_II,is_I,is_I,MAT_INITIAL_MATRIX,&AE_II);CHKERRQ(ierr); 564b1b3d7a2SStefano Zampini } 565b1b3d7a2SStefano Zampini } else { 566b1b3d7a2SStefano Zampini PetscInt n_I; 567b1b3d7a2SStefano Zampini 568b1b3d7a2SStefano Zampini /* IS for I dofs in original numbering */ 569b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)sub_schurs->is_I);CHKERRQ(ierr); 570a9b99552SStefano Zampini is_I_layer = sub_schurs->is_I; 571b1b3d7a2SStefano Zampini 572b1b3d7a2SStefano Zampini /* IS for I dofs in I numbering (strided 1) */ 573883469d8SStefano Zampini if (!sub_schurs->use_mumps) { 574b1b3d7a2SStefano Zampini ierr = ISGetSize(sub_schurs->is_I,&n_I);CHKERRQ(ierr); 575b1b3d7a2SStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)sub_schurs->is_I),n_I,0,1,&is_I);CHKERRQ(ierr); 576b1b3d7a2SStefano Zampini 577b1b3d7a2SStefano Zampini /* II block is the same */ 578b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)A_II);CHKERRQ(ierr); 579b1b3d7a2SStefano Zampini AE_II = A_II; 580b1b3d7a2SStefano Zampini } 581b1b3d7a2SStefano Zampini } 5825a95e1ceSStefano Zampini 583883469d8SStefano Zampini /* Get info on subset sizes and sum of all subsets sizes */ 5845a95e1ceSStefano Zampini max_subset_size = 0; 585883469d8SStefano Zampini local_size = 0; 5865a95e1ceSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 5875a95e1ceSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 5885a95e1ceSStefano Zampini max_subset_size = PetscMax(subset_size,max_subset_size); 589883469d8SStefano Zampini local_size += subset_size; 590883469d8SStefano Zampini } 591883469d8SStefano Zampini 592883469d8SStefano Zampini /* Work arrays for local indices */ 593883469d8SStefano Zampini extra = 0; 594683d3df6SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_B,&n_B);CHKERRQ(ierr); 59581ea8064SStefano Zampini if (sub_schurs->use_mumps && is_I_layer) { 596a9b99552SStefano Zampini ierr = ISGetLocalSize(is_I_layer,&extra);CHKERRQ(ierr); 597883469d8SStefano Zampini } 598683d3df6SStefano Zampini ierr = PetscMalloc1(n_B+extra,&all_local_idx_N);CHKERRQ(ierr); 599883469d8SStefano Zampini if (extra) { 600883469d8SStefano Zampini const PetscInt *idxs; 601a9b99552SStefano Zampini ierr = ISGetIndices(is_I_layer,&idxs);CHKERRQ(ierr); 602883469d8SStefano Zampini ierr = PetscMemcpy(all_local_idx_N,idxs,extra*sizeof(PetscInt));CHKERRQ(ierr); 603a9b99552SStefano Zampini ierr = ISRestoreIndices(is_I_layer,&idxs);CHKERRQ(ierr); 604883469d8SStefano Zampini } 605883469d8SStefano Zampini ierr = PetscMalloc1(local_size,&nnz);CHKERRQ(ierr); 606dc456d91SStefano Zampini ierr = PetscMalloc1(sub_schurs->n_subs,&auxnum1);CHKERRQ(ierr); 607dc456d91SStefano Zampini ierr = PetscMalloc1(sub_schurs->n_subs,&auxnum2);CHKERRQ(ierr); 608883469d8SStefano Zampini 609883469d8SStefano Zampini /* Get local indices in local numbering */ 610883469d8SStefano Zampini local_size = 0; 6115a95e1ceSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 612883469d8SStefano Zampini PetscInt j; 613883469d8SStefano Zampini const PetscInt *idxs; 614883469d8SStefano Zampini 6155a95e1ceSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 6165a95e1ceSStefano Zampini ierr = ISGetIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 617eb595f79SStefano Zampini /* start (smallest in global ordering) and multiplicity */ 618eb595f79SStefano Zampini auxnum1[i] = idxs[0]; 619eb595f79SStefano Zampini auxnum2[i] = subset_size; 620883469d8SStefano Zampini /* subset indices in local numbering */ 621883469d8SStefano Zampini ierr = PetscMemcpy(all_local_idx_N+local_size+extra,idxs,subset_size*sizeof(PetscInt));CHKERRQ(ierr); 6225a95e1ceSStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_subs[i],&idxs);CHKERRQ(ierr); 623883469d8SStefano Zampini for (j=0;j<subset_size;j++) nnz[local_size+j] = subset_size; 624883469d8SStefano Zampini local_size += subset_size; 625883469d8SStefano Zampini } 626883469d8SStefano Zampini 6275a95e1ceSStefano Zampini /* allocate extra workspace needed only for GETRI */ 628d2627357SStefano Zampini Bwork = NULL; 629d2627357SStefano Zampini pivots = NULL; 63006a4e24aSStefano Zampini if (local_size && !benign_trick && (!sub_schurs->is_hermitian || !sub_schurs->is_posdef)) { 631d2627357SStefano Zampini PetscScalar lwork; 632d2627357SStefano Zampini 63306a4e24aSStefano Zampini use_getr = PETSC_TRUE; 634d2627357SStefano Zampini B_lwork = -1; 635d2627357SStefano Zampini ierr = PetscBLASIntCast(local_size,&B_N);CHKERRQ(ierr); 636d2627357SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 637d2627357SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,Bwork,&B_N,pivots,&lwork,&B_lwork,&B_ierr)); 638d2627357SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 639d2627357SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GETRI Lapack routine %d",(int)B_ierr); 640d2627357SStefano Zampini ierr = PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);CHKERRQ(ierr); 641d2627357SStefano Zampini ierr = PetscMalloc2(B_lwork,&Bwork,B_N,&pivots);CHKERRQ(ierr); 642d2627357SStefano Zampini } 643d2627357SStefano Zampini 644d2627357SStefano Zampini /* prepare parallel matrices for summing up properly schurs on subsets */ 645dc456d91SStefano Zampini ierr = ISCreateGeneral(comm_n,sub_schurs->n_subs,auxnum1,PETSC_OWN_POINTER,&all_subsets_n);CHKERRQ(ierr); 646dc456d91SStefano Zampini ierr = ISLocalToGlobalMappingApplyIS(sub_schurs->l2gmap,all_subsets_n,&all_subsets);CHKERRQ(ierr); 647dc456d91SStefano Zampini ierr = ISDestroy(&all_subsets_n);CHKERRQ(ierr); 648dc456d91SStefano Zampini ierr = ISCreateGeneral(comm_n,sub_schurs->n_subs,auxnum2,PETSC_OWN_POINTER,&all_subsets_mult);CHKERRQ(ierr); 649dc456d91SStefano Zampini ierr = PCBDDCSubsetNumbering(all_subsets,all_subsets_mult,&global_size,&all_subsets_n);CHKERRQ(ierr); 650dc456d91SStefano Zampini ierr = ISDestroy(&all_subsets);CHKERRQ(ierr); 651dc456d91SStefano Zampini ierr = ISDestroy(&all_subsets_mult);CHKERRQ(ierr); 652dc456d91SStefano Zampini ierr = ISGetLocalSize(all_subsets_n,&i);CHKERRQ(ierr); 653dc456d91SStefano Zampini if (i != local_size) { 654dc456d91SStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid size of new subset! %d != %d",i,local_size); 655eb595f79SStefano Zampini } 656dc456d91SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(all_subsets_n,&l2gmap_subsets);CHKERRQ(ierr); 657e176bc59SStefano Zampini ierr = MatCreateIS(comm_n,1,PETSC_DECIDE,PETSC_DECIDE,global_size,global_size,l2gmap_subsets,NULL,&work_mat);CHKERRQ(ierr); 658d2627357SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&l2gmap_subsets);CHKERRQ(ierr); 659d2627357SStefano Zampini ierr = MatCreate(PetscObjectComm((PetscObject)work_mat),&global_schur_subsets);CHKERRQ(ierr); 660d2627357SStefano Zampini ierr = MatSetSizes(global_schur_subsets,PETSC_DECIDE,PETSC_DECIDE,global_size,global_size);CHKERRQ(ierr); 661d2627357SStefano Zampini ierr = MatSetType(global_schur_subsets,MATMPIAIJ);CHKERRQ(ierr); 6622972d61bSStefano Zampini 6635a95e1ceSStefano Zampini /* subset indices in local boundary numbering */ 6645a95e1ceSStefano Zampini if (!sub_schurs->is_Ej_all) { 6655a95e1ceSStefano Zampini PetscInt *all_local_idx_B; 6665a95e1ceSStefano Zampini 6675a95e1ceSStefano Zampini ierr = PetscMalloc1(local_size,&all_local_idx_B);CHKERRQ(ierr); 6685a95e1ceSStefano Zampini ierr = ISGlobalToLocalMappingApply(sub_schurs->BtoNmap,IS_GTOLM_DROP,local_size,all_local_idx_N+extra,&subset_size,all_local_idx_B);CHKERRQ(ierr); 6695a95e1ceSStefano Zampini if (subset_size != local_size) { 6705a95e1ceSStefano Zampini SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in sub_schurs serial (BtoNmap)! %d != %d\n",subset_size,local_size); 6715a95e1ceSStefano Zampini } 6725a95e1ceSStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,local_size,all_local_idx_B,PETSC_OWN_POINTER,&sub_schurs->is_Ej_all);CHKERRQ(ierr); 673b1b3d7a2SStefano Zampini } 674b1b3d7a2SStefano Zampini 6755a95e1ceSStefano Zampini /* Local matrix of all local Schur on subsets (transposed) */ 6765a95e1ceSStefano Zampini if (!sub_schurs->S_Ej_all) { 6775a95e1ceSStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&sub_schurs->S_Ej_all);CHKERRQ(ierr); 6785a95e1ceSStefano Zampini ierr = MatSetSizes(sub_schurs->S_Ej_all,PETSC_DECIDE,PETSC_DECIDE,local_size,local_size);CHKERRQ(ierr); 6795a95e1ceSStefano Zampini ierr = MatSetType(sub_schurs->S_Ej_all,MATAIJ);CHKERRQ(ierr); 6805a95e1ceSStefano Zampini ierr = MatSeqAIJSetPreallocation(sub_schurs->S_Ej_all,0,nnz);CHKERRQ(ierr); 681aa83b6aeSStefano Zampini } 682b1b3d7a2SStefano Zampini 6835a95e1ceSStefano Zampini /* Compute Schur complements explicitly */ 684be83ff47SStefano Zampini F = NULL; 685683d3df6SStefano Zampini if (!sub_schurs->use_mumps) { /* this code branch is used when MUMPS is not present; it is not very efficient, unless the economic version of the scaling is required */ 6865a95e1ceSStefano Zampini Mat S_Ej_expl; 6875a95e1ceSStefano Zampini PetscScalar *work; 6885a95e1ceSStefano Zampini PetscInt j,*dummy_idx; 6895a95e1ceSStefano Zampini PetscBool Sdense; 6905a95e1ceSStefano Zampini 6915a95e1ceSStefano Zampini ierr = PetscMalloc2(max_subset_size,&dummy_idx,max_subset_size*max_subset_size,&work);CHKERRQ(ierr); 6925a95e1ceSStefano Zampini local_size = 0; 693b1b3d7a2SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 6945a95e1ceSStefano Zampini IS is_subset_B; 6955a95e1ceSStefano Zampini Mat AE_EE,AE_IE,AE_EI,S_Ej; 6965a95e1ceSStefano Zampini 6975a95e1ceSStefano Zampini /* subsets in original and boundary numbering */ 6985a95e1ceSStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(sub_schurs->BtoNmap,IS_GTOLM_DROP,sub_schurs->is_subs[i],&is_subset_B);CHKERRQ(ierr); 6995a95e1ceSStefano Zampini /* EE block */ 7005a95e1ceSStefano Zampini ierr = MatGetSubMatrix(A_BB,is_subset_B,is_subset_B,MAT_INITIAL_MATRIX,&AE_EE);CHKERRQ(ierr); 7015a95e1ceSStefano Zampini /* IE block */ 7025a95e1ceSStefano Zampini ierr = MatGetSubMatrix(A_IB,is_I,is_subset_B,MAT_INITIAL_MATRIX,&AE_IE);CHKERRQ(ierr); 7035a95e1ceSStefano Zampini /* EI block */ 7045a95e1ceSStefano Zampini if (sub_schurs->is_hermitian) { 7055a95e1ceSStefano Zampini ierr = MatCreateTranspose(AE_IE,&AE_EI);CHKERRQ(ierr); 7065a95e1ceSStefano Zampini } else { 7075a95e1ceSStefano Zampini ierr = MatGetSubMatrix(A_BI,is_subset_B,is_I,MAT_INITIAL_MATRIX,&AE_EI);CHKERRQ(ierr); 7085a95e1ceSStefano Zampini } 709a64f4aa4SStefano Zampini ierr = ISDestroy(&is_subset_B);CHKERRQ(ierr); 7105a95e1ceSStefano Zampini ierr = MatCreateSchurComplement(AE_II,AE_II,AE_IE,AE_EI,AE_EE,&S_Ej);CHKERRQ(ierr); 7115a95e1ceSStefano Zampini ierr = MatDestroy(&AE_EE);CHKERRQ(ierr); 7125a95e1ceSStefano Zampini ierr = MatDestroy(&AE_IE);CHKERRQ(ierr); 7135a95e1ceSStefano Zampini ierr = MatDestroy(&AE_EI);CHKERRQ(ierr); 714b1b3d7a2SStefano Zampini if (AE_II == A_II) { /* we can reuse the same ksp */ 715b1b3d7a2SStefano Zampini KSP ksp; 716b1b3d7a2SStefano Zampini ierr = MatSchurComplementGetKSP(sub_schurs->S,&ksp);CHKERRQ(ierr); 7175a95e1ceSStefano Zampini ierr = MatSchurComplementSetKSP(S_Ej,ksp);CHKERRQ(ierr); 718b1b3d7a2SStefano Zampini } else { /* build new ksp object which inherits ksp and pc types from the original one */ 719b1b3d7a2SStefano Zampini KSP origksp,schurksp; 720b1b3d7a2SStefano Zampini PC origpc,schurpc; 721b1b3d7a2SStefano Zampini KSPType ksp_type; 722b1b3d7a2SStefano Zampini PetscInt n_internal; 7235a95e1ceSStefano Zampini PetscBool ispcnone; 724b1b3d7a2SStefano Zampini 725b1b3d7a2SStefano Zampini ierr = MatSchurComplementGetKSP(sub_schurs->S,&origksp);CHKERRQ(ierr); 7265a95e1ceSStefano Zampini ierr = MatSchurComplementGetKSP(S_Ej,&schurksp);CHKERRQ(ierr); 727b1b3d7a2SStefano Zampini ierr = KSPGetType(origksp,&ksp_type);CHKERRQ(ierr); 728b1b3d7a2SStefano Zampini ierr = KSPSetType(schurksp,ksp_type);CHKERRQ(ierr); 729b1b3d7a2SStefano Zampini ierr = KSPGetPC(schurksp,&schurpc);CHKERRQ(ierr); 730b1b3d7a2SStefano Zampini ierr = KSPGetPC(origksp,&origpc);CHKERRQ(ierr); 7315a95e1ceSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)origpc,PCNONE,&ispcnone);CHKERRQ(ierr); 7325a95e1ceSStefano Zampini if (!ispcnone) { 7335a95e1ceSStefano Zampini PCType pc_type; 734b1b3d7a2SStefano Zampini ierr = PCGetType(origpc,&pc_type);CHKERRQ(ierr); 735b1b3d7a2SStefano Zampini ierr = PCSetType(schurpc,pc_type);CHKERRQ(ierr); 7365a95e1ceSStefano Zampini } else { 7375a95e1ceSStefano Zampini ierr = PCSetType(schurpc,PCLU);CHKERRQ(ierr); 7385a95e1ceSStefano Zampini } 739b1b3d7a2SStefano Zampini ierr = ISGetSize(is_I,&n_internal);CHKERRQ(ierr); 740b1b3d7a2SStefano Zampini if (n_internal) { /* UMFPACK gives error with 0 sized problems */ 741b1b3d7a2SStefano Zampini MatSolverPackage solver=NULL; 742b1b3d7a2SStefano Zampini ierr = PCFactorGetMatSolverPackage(origpc,(const MatSolverPackage*)&solver);CHKERRQ(ierr); 743b1b3d7a2SStefano Zampini if (solver) { 744b1b3d7a2SStefano Zampini ierr = PCFactorSetMatSolverPackage(schurpc,solver);CHKERRQ(ierr); 745b1b3d7a2SStefano Zampini } 746b1b3d7a2SStefano Zampini } 747b1b3d7a2SStefano Zampini ierr = KSPSetUp(schurksp);CHKERRQ(ierr); 748b1b3d7a2SStefano Zampini } 7495a95e1ceSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 7505a95e1ceSStefano Zampini ierr = MatCreateSeqDense(PETSC_COMM_SELF,subset_size,subset_size,work,&S_Ej_expl);CHKERRQ(ierr); 7515a95e1ceSStefano Zampini ierr = PCBDDCComputeExplicitSchur(S_Ej,sub_schurs->is_hermitian,MAT_REUSE_MATRIX,&S_Ej_expl);CHKERRQ(ierr); 7525a95e1ceSStefano Zampini ierr = PetscObjectTypeCompare((PetscObject)S_Ej_expl,MATSEQDENSE,&Sdense);CHKERRQ(ierr); 7535a95e1ceSStefano Zampini if (Sdense) { 7545a95e1ceSStefano Zampini for (j=0;j<subset_size;j++) { 7555a95e1ceSStefano Zampini dummy_idx[j]=local_size+j; 756b1b3d7a2SStefano Zampini } 7575a95e1ceSStefano Zampini ierr = MatSetValues(sub_schurs->S_Ej_all,subset_size,dummy_idx,subset_size,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr); 7585a95e1ceSStefano Zampini } else { 7595a95e1ceSStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented for sparse matrices"); 7605a95e1ceSStefano Zampini } 7615a95e1ceSStefano Zampini ierr = MatDestroy(&S_Ej);CHKERRQ(ierr); 762a64f4aa4SStefano Zampini ierr = MatDestroy(&S_Ej_expl);CHKERRQ(ierr); 7635a95e1ceSStefano Zampini local_size += subset_size; 7645a95e1ceSStefano Zampini } 7655a95e1ceSStefano Zampini ierr = PetscFree2(dummy_idx,work);CHKERRQ(ierr); 766b1b3d7a2SStefano Zampini /* free */ 767b1b3d7a2SStefano Zampini ierr = ISDestroy(&is_I);CHKERRQ(ierr); 768b1b3d7a2SStefano Zampini ierr = MatDestroy(&AE_II);CHKERRQ(ierr); 7695a95e1ceSStefano Zampini ierr = PetscFree(all_local_idx_N);CHKERRQ(ierr); 770883469d8SStefano Zampini } else { 771be83ff47SStefano Zampini Mat A; 772*ca92afb2SStefano Zampini Vec *benign_AIIm1_ones = NULL; 773*ca92afb2SStefano Zampini IS is_A_all,*is_p_r = NULL; 774*ca92afb2SStefano Zampini PetscScalar *work,*S_data,*schur_factor,*cs_AIB = NULL; 7756dba178dSStefano Zampini PetscInt n,n_I,*dummy_idx,size_schur,size_active_schur,cum,cum2; 776683d3df6SStefano Zampini PetscBool economic,mumps_S,S_lower_triangular = PETSC_FALSE,factor_workaround; 777d4933d67SStefano Zampini char solver[256]; 778883469d8SStefano Zampini 779683d3df6SStefano Zampini /* get sizes */ 78081ea8064SStefano Zampini n_I = 0; 78181ea8064SStefano Zampini if (is_I_layer) { 782a9b99552SStefano Zampini ierr = ISGetLocalSize(is_I_layer,&n_I);CHKERRQ(ierr); 78381ea8064SStefano Zampini } 784683d3df6SStefano Zampini economic = PETSC_FALSE; 785683d3df6SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_I,&cum);CHKERRQ(ierr); 786683d3df6SStefano Zampini if (cum != n_I) economic = PETSC_TRUE; 787683d3df6SStefano Zampini ierr = MatGetLocalSize(sub_schurs->A,&n,NULL);CHKERRQ(ierr); 788d62866d3SStefano Zampini 789683d3df6SStefano Zampini /* size active schurs does not count any dirichlet or vertex dof on the interface */ 790683d3df6SStefano Zampini size_active_schur = local_size; 791683d3df6SStefano Zampini cum = n_I+size_active_schur; 792683d3df6SStefano Zampini if (sub_schurs->is_dir) { 793683d3df6SStefano Zampini const PetscInt* idxs; 794683d3df6SStefano Zampini PetscInt n_dir; 795683d3df6SStefano Zampini 796683d3df6SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_dir,&n_dir);CHKERRQ(ierr); 797683d3df6SStefano Zampini ierr = ISGetIndices(sub_schurs->is_dir,&idxs);CHKERRQ(ierr); 798683d3df6SStefano Zampini ierr = PetscMemcpy(all_local_idx_N+cum,idxs,n_dir*sizeof(PetscInt));CHKERRQ(ierr); 799683d3df6SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_dir,&idxs);CHKERRQ(ierr); 800683d3df6SStefano Zampini cum += n_dir; 801d62866d3SStefano Zampini } 802683d3df6SStefano Zampini factor_workaround = PETSC_FALSE; 803683d3df6SStefano Zampini /* include the primal vertices in the Schur complement */ 804683d3df6SStefano Zampini if (exact_schur && sub_schurs->is_vertices && compute_Stilda) { 805683d3df6SStefano Zampini PetscInt n_v; 806683d3df6SStefano Zampini 807683d3df6SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_v);CHKERRQ(ierr); 808683d3df6SStefano Zampini if (n_v) { 809683d3df6SStefano Zampini const PetscInt* idxs; 810683d3df6SStefano Zampini 811683d3df6SStefano Zampini ierr = ISGetIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 812683d3df6SStefano Zampini ierr = PetscMemcpy(all_local_idx_N+cum,idxs,n_v*sizeof(PetscInt));CHKERRQ(ierr); 813683d3df6SStefano Zampini ierr = ISRestoreIndices(sub_schurs->is_vertices,&idxs);CHKERRQ(ierr); 814683d3df6SStefano Zampini cum += n_v; 815683d3df6SStefano Zampini factor_workaround = PETSC_TRUE; 816683d3df6SStefano Zampini } 817683d3df6SStefano Zampini } 818683d3df6SStefano Zampini size_schur = cum - n_I; 819683d3df6SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,cum,all_local_idx_N,PETSC_OWN_POINTER,&is_A_all);CHKERRQ(ierr); 820683d3df6SStefano Zampini /* get working mat (TODO: factorize without actually permuting) */ 821683d3df6SStefano Zampini if (cum == n) { 822683d3df6SStefano Zampini ierr = ISSetPermutation(is_A_all);CHKERRQ(ierr); 823683d3df6SStefano Zampini ierr = MatPermute(sub_schurs->A,is_A_all,is_A_all,&A);CHKERRQ(ierr); 824683d3df6SStefano Zampini } else { 8256816873aSStefano Zampini ierr = MatGetSubMatrix(sub_schurs->A,is_A_all,is_A_all,MAT_INITIAL_MATRIX,&A);CHKERRQ(ierr); 826683d3df6SStefano Zampini } 827a9b99552SStefano Zampini ierr = MatSetOptionsPrefix(A,"sub_schurs_");CHKERRQ(ierr); 828*ca92afb2SStefano Zampini 829*ca92afb2SStefano Zampini /* if we actually change the basis for the pressures, LDL^T factors will use a lot of memory 830*ca92afb2SStefano Zampini more than n^2 instead of n^1.5 or something. This is a workaround */ 831*ca92afb2SStefano Zampini if (benign_n) { 832*ca92afb2SStefano Zampini Vec v; 833*ca92afb2SStefano Zampini ISLocalToGlobalMapping N_to_reor; 834*ca92afb2SStefano Zampini IS is_p0,is_p0_p; 835*ca92afb2SStefano Zampini 836*ca92afb2SStefano Zampini ierr = ISLocalToGlobalMappingCreateIS(is_A_all,&N_to_reor);CHKERRQ(ierr); 837*ca92afb2SStefano Zampini ierr = ISCreateGeneral(PETSC_COMM_SELF,benign_n,benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);CHKERRQ(ierr); 838*ca92afb2SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(N_to_reor,IS_GTOLM_DROP,is_p0,&is_p0_p);CHKERRQ(ierr); 839*ca92afb2SStefano Zampini ierr = ISDestroy(&is_p0);CHKERRQ(ierr); 840*ca92afb2SStefano Zampini ierr = MatCreateVecs(A,&v,NULL);CHKERRQ(ierr); 841*ca92afb2SStefano Zampini ierr = VecDuplicateVecs(v,benign_n,&benign_AIIm1_ones);CHKERRQ(ierr); 842*ca92afb2SStefano Zampini ierr = PetscMalloc1(size_schur*benign_n,&cs_AIB);CHKERRQ(ierr); 843*ca92afb2SStefano Zampini ierr = PetscMalloc1(benign_n,&is_p_r);CHKERRQ(ierr); 844*ca92afb2SStefano Zampini /* compute colsum of A_IB restricted to pressures */ 845*ca92afb2SStefano Zampini for (i=0;i<benign_n;i++) { 846*ca92afb2SStefano Zampini PetscScalar *array; 847*ca92afb2SStefano Zampini const PetscInt *idxs; 848*ca92afb2SStefano Zampini PetscInt j,nz; 849*ca92afb2SStefano Zampini 850*ca92afb2SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(N_to_reor,IS_GTOLM_DROP,benign_zerodiag_subs[i],&is_p_r[i]);CHKERRQ(ierr); 851*ca92afb2SStefano Zampini ierr = VecGetArray(benign_AIIm1_ones[i],&array);CHKERRQ(ierr); 852*ca92afb2SStefano Zampini ierr = ISGetLocalSize(is_p_r[i],&nz);CHKERRQ(ierr); 853*ca92afb2SStefano Zampini ierr = ISGetIndices(is_p_r[i],&idxs);CHKERRQ(ierr); 854*ca92afb2SStefano Zampini for (j=0;j<nz;j++) array[idxs[j]] = 1.; 855*ca92afb2SStefano Zampini ierr = ISRestoreIndices(is_p_r[i],&idxs);CHKERRQ(ierr); 856*ca92afb2SStefano Zampini ierr = VecRestoreArray(benign_AIIm1_ones[i],&array);CHKERRQ(ierr); 857*ca92afb2SStefano Zampini ierr = MatMult(A,benign_AIIm1_ones[i],v);CHKERRQ(ierr); 858*ca92afb2SStefano Zampini ierr = VecGetArray(v,&array);CHKERRQ(ierr); 859*ca92afb2SStefano Zampini for (j=0;j<size_schur;j++) cs_AIB[i*size_schur + j] = array[j+n_I]; 860*ca92afb2SStefano Zampini ierr = VecRestoreArray(v,&array);CHKERRQ(ierr); 861*ca92afb2SStefano Zampini } 862*ca92afb2SStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 863*ca92afb2SStefano Zampini ierr = MatSetOption(A,MAT_KEEP_NONZERO_PATTERN,PETSC_FALSE);CHKERRQ(ierr); 864*ca92afb2SStefano Zampini ierr = MatZeroRowsColumnsIS(A,is_p0_p,1.0,NULL,NULL);CHKERRQ(ierr); 865*ca92afb2SStefano Zampini ierr = ISDestroy(&is_p0_p);CHKERRQ(ierr); 866*ca92afb2SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&N_to_reor);CHKERRQ(ierr); 867*ca92afb2SStefano Zampini } 8683301b35fSStefano Zampini ierr = MatSetOption(A,MAT_SYMMETRIC,sub_schurs->is_hermitian);CHKERRQ(ierr); 8693301b35fSStefano Zampini ierr = MatSetOption(A,MAT_HERMITIAN,sub_schurs->is_hermitian);CHKERRQ(ierr); 8703301b35fSStefano Zampini ierr = MatSetOption(A,MAT_SPD,sub_schurs->is_posdef);CHKERRQ(ierr); 871d4933d67SStefano Zampini ierr = PetscStrcpy(solver,MATSOLVERMUMPS);CHKERRQ(ierr); 872d4933d67SStefano Zampini ierr = PetscOptionsGetString("sub_schurs_","-mat_solver_package",solver,256,NULL);CHKERRQ(ierr); 873883469d8SStefano Zampini 874683d3df6SStefano Zampini /* when using the benign subspace trick, the local Schur complements are SPD */ 875d47842beSStefano Zampini if (benign_trick) sub_schurs->is_posdef = PETSC_TRUE; 876d47842beSStefano Zampini 877683d3df6SStefano Zampini if (n_I) { /* TODO add ordering from options */ 8785a05ddb0SStefano Zampini IS is_schur; 8795a05ddb0SStefano Zampini 8809ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 881d4933d67SStefano Zampini ierr = MatGetFactor(A,solver,MAT_FACTOR_CHOLESKY,&F);CHKERRQ(ierr); 882883469d8SStefano Zampini } else { 883d4933d67SStefano Zampini ierr = MatGetFactor(A,solver,MAT_FACTOR_LU,&F);CHKERRQ(ierr); 884883469d8SStefano Zampini } 885883469d8SStefano Zampini /* subsets ordered last */ 8865a05ddb0SStefano Zampini ierr = ISCreateStride(PETSC_COMM_SELF,size_schur,n_I,1,&is_schur);CHKERRQ(ierr); 8875a05ddb0SStefano Zampini ierr = MatFactorSetSchurIS(F,is_schur);CHKERRQ(ierr); 8885a05ddb0SStefano Zampini ierr = ISDestroy(&is_schur);CHKERRQ(ierr); 889883469d8SStefano Zampini 890883469d8SStefano Zampini /* factorization step */ 8919ab7bb16SStefano Zampini if (sub_schurs->is_hermitian && sub_schurs->is_posdef) { 892883469d8SStefano Zampini ierr = MatCholeskyFactorSymbolic(F,A,NULL,NULL);CHKERRQ(ierr); 893be83ff47SStefano Zampini #if defined(PETSC_HAVE_MUMPS) /* be sure that icntl 19 is not set by command line */ 894be83ff47SStefano Zampini ierr = MatMumpsSetIcntl(F,19,2);CHKERRQ(ierr); 895be83ff47SStefano Zampini #endif 896e8ade678SStefano Zampini if (sub_schurs->is_posdef) { 897e8ade678SStefano Zampini ierr = MatFactorSetSchurComplementSolverType(F,1);CHKERRQ(ierr); 898e8ade678SStefano Zampini } else { 899e8ade678SStefano Zampini ierr = MatFactorSetSchurComplementSolverType(F,2);CHKERRQ(ierr); 900e8ade678SStefano Zampini } 901883469d8SStefano Zampini ierr = MatCholeskyFactorNumeric(F,A,NULL);CHKERRQ(ierr); 902a0b0af32SStefano Zampini S_lower_triangular = PETSC_TRUE; 903883469d8SStefano Zampini } else { 904883469d8SStefano Zampini ierr = MatLUFactorSymbolic(F,A,NULL,NULL,NULL);CHKERRQ(ierr); 905be83ff47SStefano Zampini #if defined(PETSC_HAVE_MUMPS) /* be sure that icntl 19 is not set by command line */ 906be83ff47SStefano Zampini ierr = MatMumpsSetIcntl(F,19,3);CHKERRQ(ierr); 907be83ff47SStefano Zampini #endif 908883469d8SStefano Zampini ierr = MatLUFactorNumeric(F,A,NULL);CHKERRQ(ierr); 909883469d8SStefano Zampini } 910883469d8SStefano Zampini 911883469d8SStefano Zampini /* get explicit Schur Complement computed during numeric factorization */ 9125a05ddb0SStefano Zampini ierr = MatFactorGetSchurComplement(F,&S_all);CHKERRQ(ierr); 913d62866d3SStefano Zampini /* we can reuse the solvers if we are not using the economic version */ 914683d3df6SStefano Zampini reuse_solvers = (PetscBool)(reuse_solvers && !economic); 915683d3df6SStefano Zampini factor_workaround = (PetscBool)(reuse_solvers && factor_workaround); 9164a6c6b0dSStefano Zampini mumps_S = PETSC_TRUE; 917*ca92afb2SStefano Zampini 918*ca92afb2SStefano Zampini /* update the Schur complement with the change of basis */ 919*ca92afb2SStefano Zampini if (benign_n) { 920*ca92afb2SStefano Zampini PetscScalar *S_data; 921*ca92afb2SStefano Zampini Vec v; 922*ca92afb2SStefano Zampini 923*ca92afb2SStefano Zampini ierr = MatDenseGetArray(S_all,&S_data);CHKERRQ(ierr); 924*ca92afb2SStefano Zampini ierr = VecDuplicate(benign_AIIm1_ones[0],&v);CHKERRQ(ierr); 925*ca92afb2SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 926*ca92afb2SStefano Zampini ierr = MatMumpsSetIcntl(F,26,0);CHKERRQ(ierr); 927*ca92afb2SStefano Zampini #endif 928*ca92afb2SStefano Zampini #if defined(PETSC_HAVE_MKL_PARDISO) 929*ca92afb2SStefano Zampini ierr = MatMkl_PardisoSetCntl(F,70,1);CHKERRQ(ierr); 930*ca92afb2SStefano Zampini #endif 931*ca92afb2SStefano Zampini for (i=0;i<benign_n;i++) { 932*ca92afb2SStefano Zampini PetscScalar *array,sum = 0.; 933*ca92afb2SStefano Zampini const PetscInt *idxs; 934*ca92afb2SStefano Zampini PetscInt j,nz; 935*ca92afb2SStefano Zampini 936*ca92afb2SStefano Zampini ierr = VecCopy(benign_AIIm1_ones[i],v);CHKERRQ(ierr); 937*ca92afb2SStefano Zampini ierr = MatSolve(F,v,benign_AIIm1_ones[i]);CHKERRQ(ierr); 938*ca92afb2SStefano Zampini ierr = VecGetArray(benign_AIIm1_ones[i],&array);CHKERRQ(ierr); 939*ca92afb2SStefano Zampini ierr = ISGetLocalSize(is_p_r[i],&nz);CHKERRQ(ierr); 940*ca92afb2SStefano Zampini ierr = ISGetIndices(is_p_r[i],&idxs);CHKERRQ(ierr); 941*ca92afb2SStefano Zampini for (j=0;j<nz;j++) sum += array[idxs[j]]; 942*ca92afb2SStefano Zampini sum -= 1.; /* p0 dof (eliminated) is excluded from the sum */ 943*ca92afb2SStefano Zampini ierr = ISRestoreIndices(is_p_r[i],&idxs);CHKERRQ(ierr); 944*ca92afb2SStefano Zampini ierr = VecRestoreArray(benign_AIIm1_ones[i],&array);CHKERRQ(ierr); 945*ca92afb2SStefano Zampini ierr = MatMult(A,benign_AIIm1_ones[i],v);CHKERRQ(ierr); 946*ca92afb2SStefano Zampini ierr = VecGetArray(v,&array);CHKERRQ(ierr); 947*ca92afb2SStefano Zampini /* perform sparse rank-1 updates on symmetric Schur */ 948*ca92afb2SStefano Zampini sum = -sum/(1.*nz*nz); 949*ca92afb2SStefano Zampini ierr = SparseRankOneUpdate(S_data,size_schur,sum,cs_AIB+i*size_schur,cs_AIB+i*size_schur);CHKERRQ(ierr); 950*ca92afb2SStefano Zampini sum = 1./nz; 951*ca92afb2SStefano Zampini ierr = SparseRankOneUpdate(S_data,size_schur,sum,array+n_I,cs_AIB+i*size_schur);CHKERRQ(ierr); 952*ca92afb2SStefano Zampini ierr = SparseRankOneUpdate(S_data,size_schur,sum,cs_AIB+i*size_schur,array+n_I);CHKERRQ(ierr); 953*ca92afb2SStefano Zampini ierr = VecRestoreArray(v,&array);CHKERRQ(ierr); 954*ca92afb2SStefano Zampini } 955*ca92afb2SStefano Zampini ierr = VecDestroy(&v);CHKERRQ(ierr); 956*ca92afb2SStefano Zampini ierr = MatDenseRestoreArray(S_all,&S_data);CHKERRQ(ierr); 957*ca92afb2SStefano Zampini } 958*ca92afb2SStefano Zampini ierr = PetscFree(cs_AIB);CHKERRQ(ierr); 959*ca92afb2SStefano Zampini ierr = VecDestroyVecs(benign_n,&benign_AIIm1_ones);CHKERRQ(ierr); 960be83ff47SStefano Zampini } else { /* we can't use MUMPS when size_schur == size_of_the_problem */ 961be83ff47SStefano Zampini ierr = MatConvert(A,MATSEQDENSE,MAT_INITIAL_MATRIX,&S_all);CHKERRQ(ierr); 962683d3df6SStefano Zampini reuse_solvers = PETSC_FALSE; /* TODO: why we can't reuse the solvers here? */ 9634a6c6b0dSStefano Zampini mumps_S = PETSC_FALSE; 964be83ff47SStefano Zampini } 965be83ff47SStefano Zampini 966be83ff47SStefano Zampini if (reuse_solvers) { 967d62866d3SStefano Zampini Mat A_II; 96853892102SStefano Zampini Vec vec1_B; 969d62866d3SStefano Zampini PCBDDCReuseMumps msolv_ctx; 970d5574798SStefano Zampini 971d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 9726816873aSStefano Zampini ierr = PCBDDCReuseMumpsReset(sub_schurs->reuse_mumps);CHKERRQ(ierr); 973e28d306cSStefano Zampini } else { 974e28d306cSStefano Zampini ierr = PetscNew(&sub_schurs->reuse_mumps);CHKERRQ(ierr); 975d62866d3SStefano Zampini } 976e28d306cSStefano Zampini msolv_ctx = sub_schurs->reuse_mumps; 977be83ff47SStefano Zampini ierr = MatSchurComplementGetSubMatrices(sub_schurs->S,&A_II,NULL,NULL,NULL,NULL);CHKERRQ(ierr); 978d5574798SStefano Zampini ierr = PetscObjectReference((PetscObject)F);CHKERRQ(ierr); 979d5574798SStefano Zampini msolv_ctx->F = F; 980683d3df6SStefano Zampini ierr = MatCreateVecs(F,&msolv_ctx->sol,NULL);CHKERRQ(ierr); 981683d3df6SStefano Zampini /* currently PETSc has no support for MatSolve(F,x,x), so cheat and let rhs and sol share the same memory */ 982683d3df6SStefano Zampini { 983683d3df6SStefano Zampini PetscScalar *array; 984683d3df6SStefano Zampini PetscInt n; 985683d3df6SStefano Zampini 986683d3df6SStefano Zampini ierr = VecGetLocalSize(msolv_ctx->sol,&n);CHKERRQ(ierr); 987683d3df6SStefano Zampini ierr = VecGetArray(msolv_ctx->sol,&array);CHKERRQ(ierr); 988683d3df6SStefano Zampini ierr = VecCreateSeqWithArray(PetscObjectComm((PetscObject)msolv_ctx->sol),1,n,array,&msolv_ctx->rhs);CHKERRQ(ierr); 989683d3df6SStefano Zampini ierr = VecRestoreArray(msolv_ctx->sol,&array);CHKERRQ(ierr); 990683d3df6SStefano Zampini } 991683d3df6SStefano Zampini msolv_ctx->has_vertices = factor_workaround; 992d62866d3SStefano Zampini 993d62866d3SStefano Zampini /* interior solver */ 994683d3df6SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)A_II),&msolv_ctx->interior_solver);CHKERRQ(ierr); 995d62866d3SStefano Zampini ierr = PCSetOperators(msolv_ctx->interior_solver,A_II,A_II);CHKERRQ(ierr); 996d62866d3SStefano Zampini ierr = PCSetType(msolv_ctx->interior_solver,PCSHELL);CHKERRQ(ierr); 997d62866d3SStefano Zampini ierr = PCShellSetContext(msolv_ctx->interior_solver,msolv_ctx);CHKERRQ(ierr); 998d62866d3SStefano Zampini ierr = PCShellSetApply(msolv_ctx->interior_solver,PCBDDCMumpsInteriorSolve);CHKERRQ(ierr); 999e28d306cSStefano Zampini ierr = PCShellSetApplyTranspose(msolv_ctx->interior_solver,PCBDDCMumpsInteriorSolveTranspose);CHKERRQ(ierr); 1000d62866d3SStefano Zampini 1001d62866d3SStefano Zampini /* correction solver */ 1002683d3df6SStefano Zampini ierr = PCCreate(PetscObjectComm((PetscObject)A),&msolv_ctx->correction_solver);CHKERRQ(ierr); 1003d62866d3SStefano Zampini ierr = PCSetOperators(msolv_ctx->correction_solver,A,A);CHKERRQ(ierr); 1004d62866d3SStefano Zampini ierr = PCSetType(msolv_ctx->correction_solver,PCSHELL);CHKERRQ(ierr); 1005d62866d3SStefano Zampini ierr = PCShellSetContext(msolv_ctx->correction_solver,msolv_ctx);CHKERRQ(ierr); 1006d62866d3SStefano Zampini ierr = PCShellSetApply(msolv_ctx->correction_solver,PCBDDCMumpsCorrectionSolve);CHKERRQ(ierr); 1007e28d306cSStefano Zampini ierr = PCShellSetApplyTranspose(msolv_ctx->correction_solver,PCBDDCMumpsCorrectionSolveTranspose);CHKERRQ(ierr); 100853892102SStefano Zampini 100953892102SStefano Zampini /* scatter and vecs for Schur complement solver */ 101053892102SStefano Zampini ierr = MatCreateVecs(S_all,&msolv_ctx->sol_B,&msolv_ctx->rhs_B);CHKERRQ(ierr); 101153892102SStefano Zampini ierr = MatCreateVecs(sub_schurs->S,&vec1_B,NULL);CHKERRQ(ierr); 1012683d3df6SStefano Zampini if (!factor_workaround) { 101353892102SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(sub_schurs->BtoNmap,IS_GTOLM_DROP,is_A_all,&msolv_ctx->is_B);CHKERRQ(ierr); 101453892102SStefano Zampini ierr = VecScatterCreate(vec1_B,msolv_ctx->is_B,msolv_ctx->sol_B,NULL,&msolv_ctx->correction_scatter_B);CHKERRQ(ierr); 101553892102SStefano Zampini ierr = PetscObjectReference((PetscObject)is_A_all);CHKERRQ(ierr); 101653892102SStefano Zampini msolv_ctx->is_R = is_A_all; 1017683d3df6SStefano Zampini } else { 1018683d3df6SStefano Zampini IS is_B_all; 1019683d3df6SStefano Zampini const PetscInt* idxs; 1020683d3df6SStefano Zampini PetscInt dual,n_v,n; 1021683d3df6SStefano Zampini 1022683d3df6SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&n_v);CHKERRQ(ierr); 1023683d3df6SStefano Zampini dual = size_schur - n_v; 1024683d3df6SStefano Zampini ierr = ISGetLocalSize(is_A_all,&n);CHKERRQ(ierr); 1025683d3df6SStefano Zampini ierr = ISGetIndices(is_A_all,&idxs);CHKERRQ(ierr); 1026683d3df6SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)is_A_all),dual,idxs+n_I,PETSC_COPY_VALUES,&is_B_all);CHKERRQ(ierr); 1027683d3df6SStefano Zampini ierr = ISGlobalToLocalMappingApplyIS(sub_schurs->BtoNmap,IS_GTOLM_DROP,is_B_all,&msolv_ctx->is_B);CHKERRQ(ierr); 1028683d3df6SStefano Zampini ierr = ISDestroy(&is_B_all);CHKERRQ(ierr); 1029683d3df6SStefano Zampini ierr = ISCreateStride(PetscObjectComm((PetscObject)is_A_all),dual,0,1,&is_B_all);CHKERRQ(ierr); 1030683d3df6SStefano Zampini ierr = VecScatterCreate(vec1_B,msolv_ctx->is_B,msolv_ctx->sol_B,is_B_all,&msolv_ctx->correction_scatter_B);CHKERRQ(ierr); 1031683d3df6SStefano Zampini ierr = ISDestroy(&is_B_all);CHKERRQ(ierr); 1032683d3df6SStefano Zampini ierr = ISCreateGeneral(PetscObjectComm((PetscObject)is_A_all),n-n_v,idxs,PETSC_COPY_VALUES,&msolv_ctx->is_R);CHKERRQ(ierr); 1033683d3df6SStefano Zampini ierr = ISRestoreIndices(is_A_all,&idxs);CHKERRQ(ierr); 1034683d3df6SStefano Zampini } 1035683d3df6SStefano Zampini ierr = VecDestroy(&vec1_B);CHKERRQ(ierr); 1036*ca92afb2SStefano Zampini 1037*ca92afb2SStefano Zampini /* communicate benign info to solver context */ 1038*ca92afb2SStefano Zampini if (benign_n) { 1039*ca92afb2SStefano Zampini msolv_ctx->benign_n = benign_n; 1040*ca92afb2SStefano Zampini msolv_ctx->benign_zerodiag_subs = is_p_r; 1041*ca92afb2SStefano Zampini ierr = PetscMalloc1(benign_n,&msolv_ctx->benign_save_vals);CHKERRQ(ierr); 1042*ca92afb2SStefano Zampini } 1043d5574798SStefano Zampini } 104408122e43SStefano Zampini ierr = MatDestroy(&A);CHKERRQ(ierr); 104553892102SStefano Zampini ierr = ISDestroy(&is_A_all);CHKERRQ(ierr); 10465db18549SStefano Zampini 1047be83ff47SStefano Zampini /* Work arrays */ 1048be83ff47SStefano Zampini ierr = PetscMalloc2(max_subset_size,&dummy_idx,max_subset_size*max_subset_size,&work);CHKERRQ(ierr); 10495a94c880SStefano Zampini 10505a94c880SStefano Zampini /* matrices for adaptive selection */ 105112d906b1SStefano Zampini if (compute_Stilda) { 10525a94c880SStefano Zampini if (!sub_schurs->sum_S_Ej_tilda_all) { 10535a94c880SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 10545a94c880SStefano Zampini ierr = MatSetSizes(sub_schurs->sum_S_Ej_tilda_all,PETSC_DECIDE,PETSC_DECIDE,size_active_schur,size_active_schur);CHKERRQ(ierr); 10555a94c880SStefano Zampini ierr = MatSetType(sub_schurs->sum_S_Ej_tilda_all,MATAIJ);CHKERRQ(ierr); 10565a94c880SStefano Zampini ierr = MatSeqAIJSetPreallocation(sub_schurs->sum_S_Ej_tilda_all,0,nnz);CHKERRQ(ierr); 10575a94c880SStefano Zampini } 10585a94c880SStefano Zampini if (!sub_schurs->sum_S_Ej_inv_all) { 10595a94c880SStefano Zampini ierr = MatCreate(PETSC_COMM_SELF,&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 10605a94c880SStefano Zampini ierr = MatSetSizes(sub_schurs->sum_S_Ej_inv_all,PETSC_DECIDE,PETSC_DECIDE,size_active_schur,size_active_schur);CHKERRQ(ierr); 10615a94c880SStefano Zampini ierr = MatSetType(sub_schurs->sum_S_Ej_inv_all,MATAIJ);CHKERRQ(ierr); 10625a94c880SStefano Zampini ierr = MatSeqAIJSetPreallocation(sub_schurs->sum_S_Ej_inv_all,0,nnz);CHKERRQ(ierr); 10635a94c880SStefano Zampini } 106412d906b1SStefano Zampini } 1065d2627357SStefano Zampini 1066be83ff47SStefano Zampini /* S_Ej_all */ 1067be83ff47SStefano Zampini cum = cum2 = 0; 1068be83ff47SStefano Zampini ierr = MatDenseGetArray(S_all,&S_data);CHKERRQ(ierr); 106965d8bf0aSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 107065d8bf0aSStefano Zampini PetscInt j; 107165d8bf0aSStefano Zampini 10725a95e1ceSStefano Zampini /* get S_E */ 1073b96c3477SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 1074683d3df6SStefano Zampini if (S_lower_triangular) { /* I need to expand the upper triangular data (column oriented) */ 1075be83ff47SStefano Zampini PetscInt k; 1076be83ff47SStefano Zampini for (k=0;k<subset_size;k++) { 1077be83ff47SStefano Zampini for (j=k;j<subset_size;j++) { 1078be83ff47SStefano Zampini work[k*subset_size+j] = S_data[cum2+k*size_schur+j]; 1079683d3df6SStefano Zampini work[j*subset_size+k] = PetscConj(S_data[cum2+k*size_schur+j]); 1080be83ff47SStefano Zampini } 1081be83ff47SStefano Zampini } 108206a4e24aSStefano Zampini } else { /* just copy to workspace */ 1083be83ff47SStefano Zampini PetscInt k; 1084be83ff47SStefano Zampini for (k=0;k<subset_size;k++) { 1085be83ff47SStefano Zampini for (j=0;j<subset_size;j++) { 1086be83ff47SStefano Zampini work[k*subset_size+j] = S_data[cum2+k*size_schur+j]; 1087be83ff47SStefano Zampini } 1088be83ff47SStefano Zampini } 10899087bf02SStefano Zampini } 10905a95e1ceSStefano Zampini /* insert S_E values */ 1091be83ff47SStefano Zampini for (j=0;j<subset_size;j++) dummy_idx[j] = cum+j; 10925a95e1ceSStefano Zampini ierr = MatSetValues(sub_schurs->S_Ej_all,subset_size,dummy_idx,subset_size,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr); 1093a1337663SStefano Zampini 1094683d3df6SStefano Zampini /* if adaptivity is requested, invert S_E blocks */ 1095862806e4SStefano Zampini if (compute_Stilda) { 109608122e43SStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 109708122e43SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 109806a4e24aSStefano Zampini if (!use_getr) { /* TODO add sytrf/i for symmetric non hermitian */ 10995a95e1ceSStefano Zampini PetscStackCallBLAS("LAPACKpotrf",LAPACKpotrf_("L",&B_N,work,&B_N,&B_ierr)); 11002972d61bSStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRF Lapack routine %d",(int)B_ierr); 11015a95e1ceSStefano Zampini PetscStackCallBLAS("LAPACKpotri",LAPACKpotri_("L",&B_N,work,&B_N,&B_ierr)); 11022972d61bSStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRI Lapack routine %d",(int)B_ierr); 1103d6462365SStefano Zampini } else { 1104d6462365SStefano Zampini PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,work,&B_N,pivots,&B_ierr)); 1105d6462365SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr); 1106d6462365SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,work,&B_N,pivots,Bwork,&B_lwork,&B_ierr)); 1107d6462365SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr); 11082972d61bSStefano Zampini } 110908122e43SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 11105a94c880SStefano Zampini ierr = MatSetValues(sub_schurs->sum_S_Ej_inv_all,subset_size,dummy_idx,subset_size,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr); 11119087bf02SStefano Zampini } 1112be83ff47SStefano Zampini cum += subset_size; 1113be83ff47SStefano Zampini cum2 += subset_size*(size_schur + 1); 1114be83ff47SStefano Zampini } 1115be83ff47SStefano Zampini ierr = MatDenseRestoreArray(S_all,&S_data);CHKERRQ(ierr); 11164a6c6b0dSStefano Zampini 11174a6c6b0dSStefano Zampini if (mumps_S) { 11185a05ddb0SStefano Zampini ierr = MatFactorRestoreSchurComplement(F,&S_all);CHKERRQ(ierr); 11194a6c6b0dSStefano Zampini } 1120683d3df6SStefano Zampini 1121683d3df6SStefano Zampini schur_factor = NULL; 112245951f25SStefano Zampini if (compute_Stilda && size_active_schur) { 1123683d3df6SStefano Zampini 11244a6c6b0dSStefano Zampini if (sub_schurs->n_subs == 1 && size_schur == size_active_schur) { /* we already computed the inverse */ 11254a6c6b0dSStefano Zampini PetscInt j; 11264a6c6b0dSStefano Zampini for (j=0;j<size_schur;j++) dummy_idx[j] = j; 11275a94c880SStefano Zampini ierr = MatSetValues(sub_schurs->sum_S_Ej_tilda_all,size_schur,dummy_idx,size_schur,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr); 11284a6c6b0dSStefano Zampini } else { 1129683d3df6SStefano Zampini Mat S_all_inv=NULL; 11305a05ddb0SStefano Zampini if (mumps_S) { /* use MatFactor calls to invert S */ 11316dba178dSStefano Zampini /* let MatFactor factorize the Schur complement */ 11326dba178dSStefano Zampini ierr = MatFactorFactorizeSchurComplement(F);CHKERRQ(ierr); 1133683d3df6SStefano Zampini /* for adaptive selection we need S^-1; for solver reusage we need S_\Delta\Delta^-1. 1134683d3df6SStefano 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 */ 1135683d3df6SStefano Zampini if (factor_workaround) { 1136683d3df6SStefano Zampini PetscScalar *data; 1137683d3df6SStefano Zampini PetscInt nd = 0; 11386dba178dSStefano Zampini 1139683d3df6SStefano Zampini if (!sub_schurs->is_posdef) { 1140683d3df6SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Factor update not yet implemented for non SPD matrices"); 1141683d3df6SStefano Zampini } 11426dba178dSStefano Zampini ierr = MatFactorGetSchurComplement(F,&S_all_inv);CHKERRQ(ierr); 1143683d3df6SStefano Zampini ierr = MatDenseGetArray(S_all_inv,&data);CHKERRQ(ierr); 1144683d3df6SStefano Zampini if (sub_schurs->is_dir) { /* dirichlet dofs could have different scalings */ 1145683d3df6SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_dir,&nd);CHKERRQ(ierr); 1146683d3df6SStefano Zampini } 1147683d3df6SStefano Zampini ierr = PetscMalloc1((size_active_schur*(size_active_schur+1))/2+nd,&schur_factor);CHKERRQ(ierr); 1148683d3df6SStefano Zampini cum = 0; 1149683d3df6SStefano Zampini for (i=0;i<size_active_schur;i++) { 1150683d3df6SStefano Zampini ierr = PetscMemcpy(schur_factor+cum,data+i*(size_schur+1),(size_active_schur-i)*sizeof(PetscScalar));CHKERRQ(ierr); 1151683d3df6SStefano Zampini cum += size_active_schur-i; 1152683d3df6SStefano Zampini } 1153683d3df6SStefano Zampini for (i=0;i<nd;i++) schur_factor[cum+i] = data[(i+size_active_schur)*(size_schur+1)]; 11546dba178dSStefano Zampini /* invert without calling MatFactorInvertSchurComplement, since we are hacking */ 1155683d3df6SStefano Zampini ierr = PetscBLASIntCast(size_schur,&B_N);CHKERRQ(ierr); 1156683d3df6SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 1157683d3df6SStefano Zampini PetscStackCallBLAS("LAPACKpotri",LAPACKpotri_("L",&B_N,data,&B_N,&B_ierr)); 1158683d3df6SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 1159683d3df6SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRI Lapack routine %d",(int)B_ierr); 1160683d3df6SStefano Zampini ierr = MatDenseRestoreArray(S_all_inv,&data);CHKERRQ(ierr); 1161683d3df6SStefano Zampini } else { 11626dba178dSStefano Zampini ierr = MatFactorInvertSchurComplement(F);CHKERRQ(ierr); 11636dba178dSStefano Zampini ierr = MatFactorGetSchurComplement(F,&S_all_inv);CHKERRQ(ierr); 1164683d3df6SStefano Zampini } 11654a6c6b0dSStefano Zampini } else { /* we need to invert explicitly since we are not using MUMPS for S */ 1166683d3df6SStefano Zampini ierr = PetscObjectReference((PetscObject)S_all);CHKERRQ(ierr); 1167683d3df6SStefano Zampini S_all_inv = S_all; 1168683d3df6SStefano Zampini ierr = MatDenseGetArray(S_all_inv,&S_data);CHKERRQ(ierr); 1169be83ff47SStefano Zampini ierr = PetscBLASIntCast(size_schur,&B_N);CHKERRQ(ierr); 1170be83ff47SStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 117106a4e24aSStefano Zampini if (!use_getr) { /* TODO add sytrf/i for symmetric non hermitian */ 1172be83ff47SStefano Zampini PetscStackCallBLAS("LAPACKpotrf",LAPACKpotrf_("L",&B_N,S_data,&B_N,&B_ierr)); 1173be83ff47SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRF Lapack routine %d",(int)B_ierr); 1174be83ff47SStefano Zampini PetscStackCallBLAS("LAPACKpotri",LAPACKpotri_("L",&B_N,S_data,&B_N,&B_ierr)); 1175be83ff47SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRI Lapack routine %d",(int)B_ierr); 1176d6462365SStefano Zampini } else { 1177d6462365SStefano Zampini PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,S_data,&B_N,pivots,&B_ierr)); 1178d6462365SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr); 1179d6462365SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,S_data,&B_N,pivots,Bwork,&B_lwork,&B_ierr)); 1180d6462365SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr); 1181be83ff47SStefano Zampini } 1182be83ff47SStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 1183683d3df6SStefano Zampini ierr = MatDenseRestoreArray(S_all_inv,&S_data);CHKERRQ(ierr); 1184be83ff47SStefano Zampini } 1185be83ff47SStefano Zampini /* S_Ej_tilda_all */ 1186be83ff47SStefano Zampini cum = cum2 = 0; 1187683d3df6SStefano Zampini ierr = MatDenseGetArray(S_all_inv,&S_data);CHKERRQ(ierr); 1188be83ff47SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 1189be83ff47SStefano Zampini PetscInt j; 1190862806e4SStefano Zampini 1191862806e4SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 1192be83ff47SStefano Zampini /* get (St^-1)_E */ 119306a4e24aSStefano Zampini /* Here I don't need to expand to upper triangular since St^-1 119406a4e24aSStefano Zampini will be properly accessed later during adaptive selection */ 1195a0b0af32SStefano Zampini if (S_lower_triangular) { 1196be83ff47SStefano Zampini PetscInt k; 1197be83ff47SStefano Zampini for (k=0;k<subset_size;k++) { 1198be83ff47SStefano Zampini for (j=k;j<subset_size;j++) { 1199be83ff47SStefano Zampini work[k*subset_size+j] = S_data[cum2+k*size_schur+j]; 1200be83ff47SStefano Zampini } 1201be83ff47SStefano Zampini } 1202be83ff47SStefano Zampini } else { /* copy to workspace */ 1203be83ff47SStefano Zampini PetscInt k; 1204be83ff47SStefano Zampini for (k=0;k<subset_size;k++) { 1205be83ff47SStefano Zampini for (j=0;j<subset_size;j++) { 1206be83ff47SStefano Zampini work[k*subset_size+j] = S_data[cum2+k*size_schur+j]; 1207be83ff47SStefano Zampini } 1208be83ff47SStefano Zampini } 1209be83ff47SStefano Zampini } 1210be83ff47SStefano Zampini for (j=0;j<subset_size;j++) dummy_idx[j] = cum+j; 12115a94c880SStefano Zampini ierr = MatSetValues(sub_schurs->sum_S_Ej_tilda_all,subset_size,dummy_idx,subset_size,dummy_idx,work,INSERT_VALUES);CHKERRQ(ierr); 1212be83ff47SStefano Zampini cum += subset_size; 1213be83ff47SStefano Zampini cum2 += subset_size*(size_schur + 1); 1214883469d8SStefano Zampini } 1215683d3df6SStefano Zampini ierr = MatDenseRestoreArray(S_all_inv,&S_data);CHKERRQ(ierr); 12164a6c6b0dSStefano Zampini if (mumps_S) { 12176dba178dSStefano Zampini ierr = MatFactorRestoreSchurComplement(F,&S_all_inv);CHKERRQ(ierr); 12185db18549SStefano Zampini } 1219683d3df6SStefano Zampini ierr = MatDestroy(&S_all_inv);CHKERRQ(ierr); 1220683d3df6SStefano Zampini } 1221683d3df6SStefano Zampini 1222683d3df6SStefano Zampini /* move back factors to Schur data into F */ 1223683d3df6SStefano Zampini if (factor_workaround) { 1224683d3df6SStefano Zampini Mat S_tmp; 1225683d3df6SStefano Zampini PetscInt nv,nd=0; 1226683d3df6SStefano Zampini 1227683d3df6SStefano Zampini if (!mumps_S) { 1228683d3df6SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen"); 1229683d3df6SStefano Zampini } 1230683d3df6SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_vertices,&nv);CHKERRQ(ierr); 12316dba178dSStefano Zampini ierr = MatFactorGetSchurComplement(F,&S_tmp);CHKERRQ(ierr); 1232683d3df6SStefano Zampini if (sub_schurs->is_posdef) { 1233683d3df6SStefano Zampini PetscScalar *data; 1234683d3df6SStefano Zampini 1235683d3df6SStefano Zampini ierr = MatZeroEntries(S_tmp);CHKERRQ(ierr); 1236683d3df6SStefano Zampini ierr = MatDenseGetArray(S_tmp,&data);CHKERRQ(ierr); 1237683d3df6SStefano Zampini 1238683d3df6SStefano Zampini if (S_lower_triangular) { 1239683d3df6SStefano Zampini cum = 0; 1240683d3df6SStefano Zampini for (i=0;i<size_active_schur;i++) { 1241683d3df6SStefano Zampini ierr = PetscMemcpy(data+i*(size_schur+1),schur_factor+cum,(size_active_schur-i)*sizeof(PetscScalar));CHKERRQ(ierr); 1242683d3df6SStefano Zampini cum += size_active_schur-i; 1243683d3df6SStefano Zampini } 1244683d3df6SStefano Zampini } else { 1245683d3df6SStefano Zampini ierr = PetscMemcpy(data,schur_factor,size_schur*size_schur*sizeof(PetscScalar));CHKERRQ(ierr); 1246683d3df6SStefano Zampini } 1247683d3df6SStefano Zampini if (sub_schurs->is_dir) { 1248683d3df6SStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_dir,&nd);CHKERRQ(ierr); 1249683d3df6SStefano Zampini for (i=0;i<nd;i++) { 1250683d3df6SStefano Zampini data[(i+size_active_schur)*(size_schur+1)] = schur_factor[cum+i]; 1251683d3df6SStefano Zampini } 1252683d3df6SStefano Zampini } 12536dba178dSStefano Zampini /* workaround: since I cannot modify the matrices used inside the solvers for the forward and backward substitutions, 1254683d3df6SStefano Zampini set the diagonal entry of the Schur factor to a very large value */ 1255683d3df6SStefano Zampini for (i=size_active_schur+nd;i<size_schur;i++) { 1256683d3df6SStefano Zampini data[i*(size_schur+1)] = PETSC_MAX_REAL; 1257683d3df6SStefano Zampini } 1258683d3df6SStefano Zampini ierr = MatDenseRestoreArray(S_tmp,&data);CHKERRQ(ierr); 1259683d3df6SStefano Zampini } else { 1260683d3df6SStefano Zampini SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Factor update not yet implemented for non SPD matrices"); 1261683d3df6SStefano Zampini } 12626dba178dSStefano Zampini ierr = MatFactorRestoreSchurComplement(F,&S_tmp);CHKERRQ(ierr); 12639087bf02SStefano Zampini } 12644a6c6b0dSStefano Zampini } 12654a6c6b0dSStefano Zampini ierr = PetscFree2(dummy_idx,work);CHKERRQ(ierr); 1266683d3df6SStefano Zampini ierr = PetscFree(schur_factor);CHKERRQ(ierr); 12674a6c6b0dSStefano Zampini } 12685a94c880SStefano Zampini ierr = PetscFree(nnz);CHKERRQ(ierr); 1269be83ff47SStefano Zampini ierr = MatDestroy(&F);CHKERRQ(ierr); 1270a9b99552SStefano Zampini ierr = ISDestroy(&is_I_layer);CHKERRQ(ierr); 1271a1337663SStefano Zampini ierr = MatDestroy(&S_all);CHKERRQ(ierr); 1272a64f4aa4SStefano Zampini ierr = MatDestroy(&A_BB);CHKERRQ(ierr); 1273a64f4aa4SStefano Zampini ierr = MatDestroy(&A_IB);CHKERRQ(ierr); 1274a64f4aa4SStefano Zampini ierr = MatDestroy(&A_BI);CHKERRQ(ierr); 12755db18549SStefano Zampini ierr = MatAssemblyBegin(sub_schurs->S_Ej_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 12765db18549SStefano Zampini ierr = MatAssemblyEnd(sub_schurs->S_Ej_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 12775a95e1ceSStefano Zampini if (compute_Stilda) { 12785a94c880SStefano Zampini ierr = MatAssemblyBegin(sub_schurs->sum_S_Ej_tilda_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 12795a94c880SStefano Zampini ierr = MatAssemblyEnd(sub_schurs->sum_S_Ej_tilda_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 12805a94c880SStefano Zampini ierr = MatAssemblyBegin(sub_schurs->sum_S_Ej_inv_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 12815a94c880SStefano Zampini ierr = MatAssemblyEnd(sub_schurs->sum_S_Ej_inv_all,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 128208122e43SStefano Zampini } 1283a1337663SStefano Zampini 12845db18549SStefano Zampini /* Global matrix of all assembled Schur on subsets */ 12855db18549SStefano Zampini ierr = MatISSetLocalMat(work_mat,sub_schurs->S_Ej_all);CHKERRQ(ierr); 12863927de2eSStefano Zampini ierr = MatISSetMPIXAIJPreallocation_Private(work_mat,global_schur_subsets,PETSC_TRUE);CHKERRQ(ierr); 12873927de2eSStefano Zampini ierr = MatISGetMPIXAIJ(work_mat,MAT_REUSE_MATRIX,&global_schur_subsets);CHKERRQ(ierr); 12885a95e1ceSStefano Zampini 12895db18549SStefano Zampini /* Get local part of (\sum_j S_Ej) */ 12905a94c880SStefano Zampini if (!sub_schurs->sum_S_Ej_all) { 1291dc456d91SStefano Zampini ierr = MatGetSubMatrices(global_schur_subsets,1,&all_subsets_n,&all_subsets_n,MAT_INITIAL_MATRIX,&submats);CHKERRQ(ierr); 12925a94c880SStefano Zampini sub_schurs->sum_S_Ej_all = submats[0]; 12935a94c880SStefano Zampini } else { 12945a94c880SStefano Zampini ierr = PetscMalloc1(1,&submats);CHKERRQ(ierr); 12955a94c880SStefano Zampini submats[0] = sub_schurs->sum_S_Ej_all; 1296dc456d91SStefano Zampini ierr = MatGetSubMatrices(global_schur_subsets,1,&all_subsets_n,&all_subsets_n,MAT_REUSE_MATRIX,&submats);CHKERRQ(ierr); 12975a94c880SStefano Zampini } 129808122e43SStefano Zampini 1299ac632422SStefano Zampini /* Compute explicitly (\sum_j S_Ej)^-1 (faster scaling during PCApply, needs extra work when doing setup) */ 1300ac632422SStefano Zampini if (faster_deluxe) { 13015a95e1ceSStefano Zampini Mat tmpmat; 13025a95e1ceSStefano Zampini PetscScalar *array; 13035a95e1ceSStefano Zampini PetscInt cum; 13045a95e1ceSStefano Zampini 13055a95e1ceSStefano Zampini ierr = MatSeqAIJGetArray(sub_schurs->sum_S_Ej_all,&array);CHKERRQ(ierr); 13065a95e1ceSStefano Zampini cum = 0; 13075a95e1ceSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 13085a95e1ceSStefano Zampini ierr = ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);CHKERRQ(ierr); 13095a95e1ceSStefano Zampini ierr = PetscBLASIntCast(subset_size,&B_N);CHKERRQ(ierr); 13105a95e1ceSStefano Zampini ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); 131106a4e24aSStefano Zampini if (!use_getr) { 13125a95e1ceSStefano Zampini PetscInt j,k; 13135a95e1ceSStefano Zampini 13145a95e1ceSStefano Zampini PetscStackCallBLAS("LAPACKpotrf",LAPACKpotrf_("L",&B_N,array+cum,&B_N,&B_ierr)); 13155a95e1ceSStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRF Lapack routine %d",(int)B_ierr); 13165a95e1ceSStefano Zampini PetscStackCallBLAS("LAPACKpotri",LAPACKpotri_("L",&B_N,array+cum,&B_N,&B_ierr)); 13175a95e1ceSStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in POTRI Lapack routine %d",(int)B_ierr); 13185a95e1ceSStefano Zampini for (j=0;j<B_N;j++) { 13195a95e1ceSStefano Zampini for (k=j+1;k<B_N;k++) { 13205a95e1ceSStefano Zampini array[k*B_N+j+cum] = array[j*B_N+k+cum]; 13215a95e1ceSStefano Zampini } 13225a95e1ceSStefano Zampini } 1323d6462365SStefano Zampini } else { 1324d6462365SStefano Zampini PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&B_N,&B_N,array+cum,&B_N,pivots,&B_ierr)); 1325d6462365SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRF Lapack routine %d",(int)B_ierr); 1326d6462365SStefano Zampini PetscStackCallBLAS("LAPACKgetri",LAPACKgetri_(&B_N,array+cum,&B_N,pivots,Bwork,&B_lwork,&B_ierr)); 1327d6462365SStefano Zampini if (B_ierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GETRI Lapack routine %d",(int)B_ierr); 13285a95e1ceSStefano Zampini } 13295a95e1ceSStefano Zampini ierr = PetscFPTrapPop();CHKERRQ(ierr); 13305a95e1ceSStefano Zampini cum += subset_size*subset_size; 13315a95e1ceSStefano Zampini } 13325a95e1ceSStefano Zampini ierr = MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_all,&array);CHKERRQ(ierr); 13335a95e1ceSStefano Zampini ierr = MatMatMult(sub_schurs->S_Ej_all,sub_schurs->sum_S_Ej_all,MAT_INITIAL_MATRIX,1.0,&tmpmat);CHKERRQ(ierr); 13345a95e1ceSStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 1335ac632422SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 13365a95e1ceSStefano Zampini sub_schurs->S_Ej_all = tmpmat; 13375a95e1ceSStefano Zampini } 13385a95e1ceSStefano Zampini 1339f6f667cfSStefano Zampini /* Get local part of (\sum_j S^-1_Ej) (\sum_j St^-1_Ej) */ 13405a95e1ceSStefano Zampini if (compute_Stilda) { 13415a94c880SStefano Zampini ierr = MatISSetLocalMat(work_mat,sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 1342a1337663SStefano Zampini ierr = MatISGetMPIXAIJ(work_mat,MAT_REUSE_MATRIX,&global_schur_subsets);CHKERRQ(ierr); 13435a94c880SStefano Zampini submats[0] = sub_schurs->sum_S_Ej_tilda_all; 1344dc456d91SStefano Zampini ierr = MatGetSubMatrices(global_schur_subsets,1,&all_subsets_n,&all_subsets_n,MAT_REUSE_MATRIX,&submats);CHKERRQ(ierr); 13455a94c880SStefano Zampini ierr = MatISSetLocalMat(work_mat,sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 134608122e43SStefano Zampini ierr = MatISGetMPIXAIJ(work_mat,MAT_REUSE_MATRIX,&global_schur_subsets);CHKERRQ(ierr); 13475a94c880SStefano Zampini submats[0] = sub_schurs->sum_S_Ej_inv_all; 1348dc456d91SStefano Zampini ierr = MatGetSubMatrices(global_schur_subsets,1,&all_subsets_n,&all_subsets_n,MAT_REUSE_MATRIX,&submats);CHKERRQ(ierr); 134908122e43SStefano Zampini } 13503202ece2SStefano Zampini 13515a95e1ceSStefano Zampini /* free workspace */ 13525a94c880SStefano Zampini ierr = PetscFree(submats);CHKERRQ(ierr); 135306a4b1faSStefano Zampini ierr = PetscFree2(Bwork,pivots);CHKERRQ(ierr); 1354a1337663SStefano Zampini ierr = MatDestroy(&global_schur_subsets);CHKERRQ(ierr); 13553202ece2SStefano Zampini ierr = MatDestroy(&work_mat);CHKERRQ(ierr); 1356dc456d91SStefano Zampini ierr = ISDestroy(&all_subsets_n);CHKERRQ(ierr); 13575a95e1ceSStefano Zampini ierr = PetscCommDestroy(&comm_n);CHKERRQ(ierr); 1358b1b3d7a2SStefano Zampini PetscFunctionReturn(0); 1359b1b3d7a2SStefano Zampini } 1360b1b3d7a2SStefano Zampini 1361b1b3d7a2SStefano Zampini #undef __FUNCT__ 1362b1b3d7a2SStefano Zampini #define __FUNCT__ "PCBDDCSubSchursInit" 1363a64f4aa4SStefano Zampini PetscErrorCode PCBDDCSubSchursInit(PCBDDCSubSchurs sub_schurs, IS is_I, IS is_B, PCBDDCGraph graph, ISLocalToGlobalMapping BtoNmap) 1364b1b3d7a2SStefano Zampini { 13659bb4a8caSStefano Zampini IS *faces,*edges,*all_cc,vertices; 13665a95e1ceSStefano Zampini PetscInt i,n_faces,n_edges,n_all_cc; 1367b1b3d7a2SStefano Zampini PetscBool is_sorted; 1368b1b3d7a2SStefano Zampini PetscErrorCode ierr; 1369b1b3d7a2SStefano Zampini 1370b1b3d7a2SStefano Zampini PetscFunctionBegin; 1371b1b3d7a2SStefano Zampini ierr = ISSorted(is_I,&is_sorted);CHKERRQ(ierr); 1372b1b3d7a2SStefano Zampini if (!is_sorted) { 1373b1b3d7a2SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)is_I),PETSC_ERR_PLIB,"IS for I dofs should be shorted"); 1374b1b3d7a2SStefano Zampini } 1375b1b3d7a2SStefano Zampini ierr = ISSorted(is_B,&is_sorted);CHKERRQ(ierr); 1376b1b3d7a2SStefano Zampini if (!is_sorted) { 1377b1b3d7a2SStefano Zampini SETERRQ(PetscObjectComm((PetscObject)is_B),PETSC_ERR_PLIB,"IS for B dofs should be shorted"); 1378b1b3d7a2SStefano Zampini } 1379b1b3d7a2SStefano Zampini 1380b1b3d7a2SStefano Zampini /* reset any previous data */ 1381b1b3d7a2SStefano Zampini ierr = PCBDDCSubSchursReset(sub_schurs);CHKERRQ(ierr); 1382b1b3d7a2SStefano Zampini 13835a95e1ceSStefano Zampini /* get index sets for faces and edges (already sorted by global ordering) */ 13849bb4a8caSStefano Zampini ierr = PCBDDCGraphGetCandidatesIS(graph,&n_faces,&faces,&n_edges,&edges,&vertices);CHKERRQ(ierr); 1385b1b3d7a2SStefano Zampini n_all_cc = n_faces+n_edges; 138608122e43SStefano Zampini ierr = PetscBTCreate(n_all_cc,&sub_schurs->is_edge);CHKERRQ(ierr); 1387b1b3d7a2SStefano Zampini ierr = PetscMalloc1(n_all_cc,&all_cc);CHKERRQ(ierr); 1388b1b3d7a2SStefano Zampini for (i=0;i<n_faces;i++) { 1389b1b3d7a2SStefano Zampini all_cc[i] = faces[i]; 1390b1b3d7a2SStefano Zampini } 1391b1b3d7a2SStefano Zampini for (i=0;i<n_edges;i++) { 1392b1b3d7a2SStefano Zampini all_cc[n_faces+i] = edges[i]; 139308122e43SStefano Zampini ierr = PetscBTSet(sub_schurs->is_edge,n_faces+i);CHKERRQ(ierr); 1394b1b3d7a2SStefano Zampini } 1395b1b3d7a2SStefano Zampini ierr = PetscFree(faces);CHKERRQ(ierr); 1396b1b3d7a2SStefano Zampini ierr = PetscFree(edges);CHKERRQ(ierr); 1397d62866d3SStefano Zampini sub_schurs->is_dir = NULL; 1398d62866d3SStefano Zampini ierr = PCBDDCGraphGetDirichletDofsB(graph,&sub_schurs->is_dir);CHKERRQ(ierr); 1399b1b3d7a2SStefano Zampini 14006816873aSStefano Zampini /* Determine if MUMPS can be used */ 1401883469d8SStefano Zampini sub_schurs->use_mumps = PETSC_FALSE; 1402883469d8SStefano Zampini #if defined(PETSC_HAVE_MUMPS) 1403a64f4aa4SStefano Zampini sub_schurs->use_mumps = PETSC_TRUE; 1404883469d8SStefano Zampini #endif 1405b1b3d7a2SStefano Zampini 1406b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)is_I);CHKERRQ(ierr); 1407b1b3d7a2SStefano Zampini sub_schurs->is_I = is_I; 1408b1b3d7a2SStefano Zampini ierr = PetscObjectReference((PetscObject)is_B);CHKERRQ(ierr); 1409b1b3d7a2SStefano Zampini sub_schurs->is_B = is_B; 14105db18549SStefano Zampini ierr = PetscObjectReference((PetscObject)graph->l2gmap);CHKERRQ(ierr); 14115db18549SStefano Zampini sub_schurs->l2gmap = graph->l2gmap; 14125db18549SStefano Zampini ierr = PetscObjectReference((PetscObject)BtoNmap);CHKERRQ(ierr); 14135db18549SStefano Zampini sub_schurs->BtoNmap = BtoNmap; 14145a95e1ceSStefano Zampini sub_schurs->n_subs = n_all_cc; 1415b1b3d7a2SStefano Zampini sub_schurs->is_subs = all_cc; 1416a64f4aa4SStefano Zampini if (!sub_schurs->use_mumps) { /* sort by local ordering mumps is not present */ 1417b96c3477SStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 1418b96c3477SStefano Zampini ierr = ISSort(sub_schurs->is_subs[i]);CHKERRQ(ierr); 1419b96c3477SStefano Zampini } 14209bb4a8caSStefano Zampini } 1421d62866d3SStefano Zampini sub_schurs->is_vertices = vertices; 1422b96c3477SStefano Zampini sub_schurs->S_Ej_all = NULL; 1423b96c3477SStefano Zampini sub_schurs->sum_S_Ej_all = NULL; 142408122e43SStefano Zampini sub_schurs->sum_S_Ej_inv_all = NULL; 1425b96c3477SStefano Zampini sub_schurs->sum_S_Ej_tilda_all = NULL; 1426b96c3477SStefano Zampini sub_schurs->is_Ej_all = NULL; 1427b1b3d7a2SStefano Zampini PetscFunctionReturn(0); 1428b1b3d7a2SStefano Zampini } 1429b1b3d7a2SStefano Zampini 1430b1b3d7a2SStefano Zampini #undef __FUNCT__ 143134a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCSubSchursCreate" 143234a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursCreate(PCBDDCSubSchurs *sub_schurs) 143334a97f8cSStefano Zampini { 143434a97f8cSStefano Zampini PCBDDCSubSchurs schurs_ctx; 143534a97f8cSStefano Zampini PetscErrorCode ierr; 143634a97f8cSStefano Zampini 143734a97f8cSStefano Zampini PetscFunctionBegin; 143834a97f8cSStefano Zampini ierr = PetscNew(&schurs_ctx);CHKERRQ(ierr); 14395ff63025SStefano Zampini schurs_ctx->n_subs = 0; 144034a97f8cSStefano Zampini *sub_schurs = schurs_ctx; 144134a97f8cSStefano Zampini PetscFunctionReturn(0); 144234a97f8cSStefano Zampini } 144334a97f8cSStefano Zampini 144434a97f8cSStefano Zampini #undef __FUNCT__ 144534a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCSubSchursDestroy" 144634a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursDestroy(PCBDDCSubSchurs *sub_schurs) 144734a97f8cSStefano Zampini { 144834a97f8cSStefano Zampini PetscErrorCode ierr; 144934a97f8cSStefano Zampini 145034a97f8cSStefano Zampini PetscFunctionBegin; 145134a97f8cSStefano Zampini ierr = PCBDDCSubSchursReset(*sub_schurs);CHKERRQ(ierr); 145234a97f8cSStefano Zampini ierr = PetscFree(*sub_schurs);CHKERRQ(ierr); 145334a97f8cSStefano Zampini PetscFunctionReturn(0); 145434a97f8cSStefano Zampini } 145534a97f8cSStefano Zampini 145634a97f8cSStefano Zampini #undef __FUNCT__ 145734a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCSubSchursReset" 145834a97f8cSStefano Zampini PetscErrorCode PCBDDCSubSchursReset(PCBDDCSubSchurs sub_schurs) 145934a97f8cSStefano Zampini { 146034a97f8cSStefano Zampini PetscInt i; 146134a97f8cSStefano Zampini PetscErrorCode ierr; 146234a97f8cSStefano Zampini 146334a97f8cSStefano Zampini PetscFunctionBegin; 14641e9c79c2SStefano Zampini ierr = MatDestroy(&sub_schurs->A);CHKERRQ(ierr); 1465b1b3d7a2SStefano Zampini ierr = MatDestroy(&sub_schurs->S);CHKERRQ(ierr); 1466b1b3d7a2SStefano Zampini ierr = ISDestroy(&sub_schurs->is_I);CHKERRQ(ierr); 1467b1b3d7a2SStefano Zampini ierr = ISDestroy(&sub_schurs->is_B);CHKERRQ(ierr); 14685db18549SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&sub_schurs->l2gmap);CHKERRQ(ierr); 14695db18549SStefano Zampini ierr = ISLocalToGlobalMappingDestroy(&sub_schurs->BtoNmap);CHKERRQ(ierr); 147041c3ba1bSStefano Zampini ierr = MatDestroy(&sub_schurs->S_Ej_all);CHKERRQ(ierr); 147141c3ba1bSStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_all);CHKERRQ(ierr); 147208122e43SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_inv_all);CHKERRQ(ierr); 1473a1337663SStefano Zampini ierr = MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);CHKERRQ(ierr); 14745db18549SStefano Zampini ierr = ISDestroy(&sub_schurs->is_Ej_all);CHKERRQ(ierr); 1475d62866d3SStefano Zampini ierr = ISDestroy(&sub_schurs->is_vertices);CHKERRQ(ierr); 1476d62866d3SStefano Zampini ierr = ISDestroy(&sub_schurs->is_dir);CHKERRQ(ierr); 147708122e43SStefano Zampini ierr = PetscBTDestroy(&sub_schurs->is_edge);CHKERRQ(ierr); 147834a97f8cSStefano Zampini for (i=0;i<sub_schurs->n_subs;i++) { 1479b1b3d7a2SStefano Zampini ierr = ISDestroy(&sub_schurs->is_subs[i]);CHKERRQ(ierr); 148034a97f8cSStefano Zampini } 14815ff63025SStefano Zampini if (sub_schurs->n_subs) { 1482b1b3d7a2SStefano Zampini ierr = PetscFree(sub_schurs->is_subs);CHKERRQ(ierr); 14833dc780c3SStefano Zampini } 1484d62866d3SStefano Zampini if (sub_schurs->reuse_mumps) { 1485d62866d3SStefano Zampini ierr = PCBDDCReuseMumpsReset(sub_schurs->reuse_mumps);CHKERRQ(ierr); 1486d62866d3SStefano Zampini } 1487d62866d3SStefano Zampini ierr = PetscFree(sub_schurs->reuse_mumps);CHKERRQ(ierr); 148834a97f8cSStefano Zampini sub_schurs->n_subs = 0; 148934a97f8cSStefano Zampini PetscFunctionReturn(0); 149034a97f8cSStefano Zampini } 149134a97f8cSStefano Zampini 149234a97f8cSStefano Zampini #undef __FUNCT__ 149334a97f8cSStefano Zampini #define __FUNCT__ "PCBDDCAdjGetNextLayer_Private" 14942a155e38SStefano Zampini PETSC_STATIC_INLINE PetscErrorCode PCBDDCAdjGetNextLayer_Private(PetscInt* queue_tip,PetscInt n_prev,PetscBT touched,PetscInt* xadj,PetscInt* adjncy,PetscInt* n_added) 149534a97f8cSStefano Zampini { 149634a97f8cSStefano Zampini PetscInt i,j,n; 149734a97f8cSStefano Zampini PetscErrorCode ierr; 149834a97f8cSStefano Zampini 149934a97f8cSStefano Zampini PetscFunctionBegin; 150034a97f8cSStefano Zampini n = 0; 150134a97f8cSStefano Zampini for (i=-n_prev;i<0;i++) { 150234a97f8cSStefano Zampini PetscInt start_dof = queue_tip[i]; 150334a97f8cSStefano Zampini for (j=xadj[start_dof];j<xadj[start_dof+1];j++) { 150434a97f8cSStefano Zampini PetscInt dof = adjncy[j]; 150534a97f8cSStefano Zampini if (!PetscBTLookup(touched,dof)) { 150634a97f8cSStefano Zampini ierr = PetscBTSet(touched,dof);CHKERRQ(ierr); 150734a97f8cSStefano Zampini queue_tip[n] = dof; 150834a97f8cSStefano Zampini n++; 150934a97f8cSStefano Zampini } 151034a97f8cSStefano Zampini } 151134a97f8cSStefano Zampini } 151234a97f8cSStefano Zampini *n_added = n; 151334a97f8cSStefano Zampini PetscFunctionReturn(0); 151434a97f8cSStefano Zampini } 1515