std::shared_ptr<GenericMatrix> MatTransposeMatMult(GenericMatrix& A, GenericMatrix& B)
{
const dolfin::PETScMatrix* Ap = &as_type<const dolfin::PETScMatrix>(A);
const dolfin::PETScMatrix* Bp = &as_type<const dolfin::PETScMatrix>(B);
Mat CC;
PetscErrorCode ierr = MatTransposeMatMult(Ap->mat(), Bp->mat(), MAT_INITIAL_MATRIX, PETSC_DEFAULT, &CC);
dolfin::PETScMatrix CCC = PETScMatrix(CC);
return CCC.copy();
}
static PetscErrorCode getGIDsOnSquareGraph(const PetscInt nselected_1,const PetscInt clid_lid_1[],const Mat Gmat1,IS *a_selected_2,Mat *a_Gmat_2,PetscInt **a_crsGID)
{
PetscErrorCode ierr;
PetscMPIInt rank,size;
PetscInt *crsGID, kk,my0,Iend,nloc;
MPI_Comm comm;
PetscFunctionBegin;
ierr = PetscObjectGetComm((PetscObject)Gmat1,&comm);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr);
ierr = MatGetOwnershipRange(Gmat1,&my0,&Iend);CHKERRQ(ierr); /* AIJ */
nloc = Iend - my0; /* this does not change */
if (size == 1) { /* not much to do in serial */
ierr = PetscMalloc1(nselected_1, &crsGID);CHKERRQ(ierr);
for (kk=0; kk<nselected_1; kk++) crsGID[kk] = kk;
*a_Gmat_2 = 0;
ierr = ISCreateGeneral(PETSC_COMM_SELF,nselected_1,clid_lid_1,PETSC_COPY_VALUES,a_selected_2);CHKERRQ(ierr);
} else {
PetscInt idx,num_fine_ghosts,num_crs_ghost,myCrs0;
Mat_MPIAIJ *mpimat2;
Mat Gmat2;
Vec locState;
PetscScalar *cpcol_state;
/* scan my coarse zero gid, set 'lid_state' with coarse GID */
kk = nselected_1;
MPI_Scan(&kk, &myCrs0, 1, MPIU_INT, MPIU_SUM, comm);
myCrs0 -= nselected_1;
if (a_Gmat_2) { /* output */
/* grow graph to get wider set of selected vertices to cover fine grid, invalidates 'llist' */
ierr = MatTransposeMatMult(Gmat1, Gmat1, MAT_INITIAL_MATRIX, PETSC_DEFAULT, &Gmat2);CHKERRQ(ierr);
*a_Gmat_2 = Gmat2; /* output */
} else Gmat2 = Gmat1; /* use local to get crsGIDs at least */
/* get coarse grid GIDS for selected (locals and ghosts) */
mpimat2 = (Mat_MPIAIJ*)Gmat2->data;
ierr = MatGetVecs(Gmat2, &locState, 0);CHKERRQ(ierr);
ierr = VecSet(locState, (PetscScalar)(PetscReal)(-1));CHKERRQ(ierr); /* set with UNKNOWN state */
for (kk=0; kk<nselected_1; kk++) {
PetscInt fgid = clid_lid_1[kk] + my0;
PetscScalar v = (PetscScalar)(kk+myCrs0);
ierr = VecSetValues(locState, 1, &fgid, &v, INSERT_VALUES);CHKERRQ(ierr); /* set with PID */
}
ierr = VecAssemblyBegin(locState);CHKERRQ(ierr);
ierr = VecAssemblyEnd(locState);CHKERRQ(ierr);
ierr = VecScatterBegin(mpimat2->Mvctx,locState,mpimat2->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd(mpimat2->Mvctx,locState,mpimat2->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecGetLocalSize(mpimat2->lvec, &num_fine_ghosts);CHKERRQ(ierr);
ierr = VecGetArray(mpimat2->lvec, &cpcol_state);CHKERRQ(ierr);
for (kk=0,num_crs_ghost=0; kk<num_fine_ghosts; kk++) {
if ((PetscInt)PetscRealPart(cpcol_state[kk]) != -1) num_crs_ghost++;
}
ierr = PetscMalloc1((nselected_1+num_crs_ghost), &crsGID);CHKERRQ(ierr); /* output */
{
PetscInt *selected_set;
ierr = PetscMalloc1((nselected_1+num_crs_ghost), &selected_set);CHKERRQ(ierr);
/* do ghost of 'crsGID' */
for (kk=0,idx=nselected_1; kk<num_fine_ghosts; kk++) {
if ((PetscInt)PetscRealPart(cpcol_state[kk]) != -1) {
PetscInt cgid = (PetscInt)PetscRealPart(cpcol_state[kk]);
selected_set[idx] = nloc + kk;
crsGID[idx++] = cgid;
}
}
if (idx != (nselected_1+num_crs_ghost)) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"idx %D != (nselected_1 %D + num_crs_ghost %D)",idx,nselected_1,num_crs_ghost);
ierr = VecRestoreArray(mpimat2->lvec, &cpcol_state);CHKERRQ(ierr);
/* do locals in 'crsGID' */
ierr = VecGetArray(locState, &cpcol_state);CHKERRQ(ierr);
for (kk=0,idx=0; kk<nloc; kk++) {
if ((PetscInt)PetscRealPart(cpcol_state[kk]) != -1) {
PetscInt cgid = (PetscInt)PetscRealPart(cpcol_state[kk]);
selected_set[idx] = kk;
crsGID[idx++] = cgid;
}
}
if (idx != nselected_1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"idx %D != nselected_1 %D",idx,nselected_1);
ierr = VecRestoreArray(locState, &cpcol_state);CHKERRQ(ierr);
if (a_selected_2 != 0) { /* output */
ierr = ISCreateGeneral(PETSC_COMM_SELF,(nselected_1+num_crs_ghost),selected_set,PETSC_OWN_POINTER,a_selected_2);CHKERRQ(ierr);
} else {
ierr = PetscFree(selected_set);CHKERRQ(ierr);
}
}
ierr = VecDestroy(&locState);CHKERRQ(ierr);
}
*a_crsGID = crsGID; /* output */
PetscFunctionReturn(0);
}
PetscErrorCode PCBDDCNullSpaceAssembleCorrection(PC pc, PetscBool isdir, IS local_dofs)
{
PC_BDDC *pcbddc = (PC_BDDC*)pc->data;
PC_IS *pcis = (PC_IS*)pc->data;
Mat_IS* matis = (Mat_IS*)pc->pmat->data;
KSP local_ksp;
PC newpc;
NullSpaceCorrection_ctx shell_ctx;
Mat local_mat,local_pmat,small_mat,inv_small_mat;
Vec work1,work2;
const Vec *nullvecs;
VecScatter scatter_ctx;
IS is_aux;
MatFactorInfo matinfo;
PetscScalar *basis_mat,*Kbasis_mat,*array,*array_mat;
PetscScalar one = 1.0,zero = 0.0, m_one = -1.0;
PetscInt basis_dofs,basis_size,nnsp_size,i,k;
PetscBool nnsp_has_cnst;
PetscErrorCode ierr;
PetscFunctionBegin;
/* Infer the local solver */
ierr = ISGetSize(local_dofs,&basis_dofs);CHKERRQ(ierr);
if (isdir) {
/* Dirichlet solver */
local_ksp = pcbddc->ksp_D;
} else {
/* Neumann solver */
local_ksp = pcbddc->ksp_R;
}
ierr = KSPGetOperators(local_ksp,&local_mat,&local_pmat);CHKERRQ(ierr);
/* Get null space vecs */
ierr = MatNullSpaceGetVecs(pcbddc->NullSpace,&nnsp_has_cnst,&nnsp_size,&nullvecs);CHKERRQ(ierr);
basis_size = nnsp_size;
if (nnsp_has_cnst) {
basis_size++;
}
if (basis_dofs) {
/* Create shell ctx */
ierr = PetscNew(&shell_ctx);CHKERRQ(ierr);
/* Create work vectors in shell context */
ierr = VecCreate(PETSC_COMM_SELF,&shell_ctx->work_small_1);CHKERRQ(ierr);
ierr = VecSetSizes(shell_ctx->work_small_1,basis_size,basis_size);CHKERRQ(ierr);
ierr = VecSetType(shell_ctx->work_small_1,VECSEQ);CHKERRQ(ierr);
ierr = VecDuplicate(shell_ctx->work_small_1,&shell_ctx->work_small_2);CHKERRQ(ierr);
ierr = VecCreate(PETSC_COMM_SELF,&shell_ctx->work_full_1);CHKERRQ(ierr);
ierr = VecSetSizes(shell_ctx->work_full_1,basis_dofs,basis_dofs);CHKERRQ(ierr);
ierr = VecSetType(shell_ctx->work_full_1,VECSEQ);CHKERRQ(ierr);
ierr = VecDuplicate(shell_ctx->work_full_1,&shell_ctx->work_full_2);CHKERRQ(ierr);
/* Allocate workspace */
ierr = MatCreateSeqDense(PETSC_COMM_SELF,basis_dofs,basis_size,NULL,&shell_ctx->basis_mat );CHKERRQ(ierr);
ierr = MatCreateSeqDense(PETSC_COMM_SELF,basis_dofs,basis_size,NULL,&shell_ctx->Kbasis_mat);CHKERRQ(ierr);
ierr = MatDenseGetArray(shell_ctx->basis_mat,&basis_mat);CHKERRQ(ierr);
ierr = MatDenseGetArray(shell_ctx->Kbasis_mat,&Kbasis_mat);CHKERRQ(ierr);
/* Restrict local null space on selected dofs (Dirichlet or Neumann)
and compute matrices N and K*N */
ierr = VecDuplicate(shell_ctx->work_full_1,&work1);CHKERRQ(ierr);
ierr = VecDuplicate(shell_ctx->work_full_1,&work2);CHKERRQ(ierr);
ierr = VecScatterCreate(pcis->vec1_N,local_dofs,work1,(IS)0,&scatter_ctx);CHKERRQ(ierr);
}
for (k=0;k<nnsp_size;k++) {
ierr = VecScatterBegin(matis->rctx,nullvecs[k],pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd(matis->rctx,nullvecs[k],pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
if (basis_dofs) {
ierr = VecPlaceArray(work1,(const PetscScalar*)&basis_mat[k*basis_dofs]);CHKERRQ(ierr);
ierr = VecScatterBegin(scatter_ctx,pcis->vec1_N,work1,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd(scatter_ctx,pcis->vec1_N,work1,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecPlaceArray(work2,(const PetscScalar*)&Kbasis_mat[k*basis_dofs]);CHKERRQ(ierr);
ierr = MatMult(local_mat,work1,work2);CHKERRQ(ierr);
ierr = VecResetArray(work1);CHKERRQ(ierr);
ierr = VecResetArray(work2);CHKERRQ(ierr);
}
}
if (basis_dofs) {
if (nnsp_has_cnst) {
ierr = VecPlaceArray(work1,(const PetscScalar*)&basis_mat[k*basis_dofs]);CHKERRQ(ierr);
ierr = VecSet(work1,one);CHKERRQ(ierr);
ierr = VecPlaceArray(work2,(const PetscScalar*)&Kbasis_mat[k*basis_dofs]);CHKERRQ(ierr);
ierr = MatMult(local_mat,work1,work2);CHKERRQ(ierr);
ierr = VecResetArray(work1);CHKERRQ(ierr);
ierr = VecResetArray(work2);CHKERRQ(ierr);
}
ierr = VecDestroy(&work1);CHKERRQ(ierr);
ierr = VecDestroy(&work2);CHKERRQ(ierr);
ierr = VecScatterDestroy(&scatter_ctx);CHKERRQ(ierr);
ierr = MatDenseRestoreArray(shell_ctx->basis_mat,&basis_mat);CHKERRQ(ierr);
ierr = MatDenseRestoreArray(shell_ctx->Kbasis_mat,&Kbasis_mat);CHKERRQ(ierr);
/* Assemble another Mat object in shell context */
ierr = MatTransposeMatMult(shell_ctx->basis_mat,shell_ctx->Kbasis_mat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&small_mat);CHKERRQ(ierr);
ierr = MatFactorInfoInitialize(&matinfo);CHKERRQ(ierr);
ierr = ISCreateStride(PETSC_COMM_SELF,basis_size,0,1,&is_aux);CHKERRQ(ierr);
ierr = MatLUFactor(small_mat,is_aux,is_aux,&matinfo);CHKERRQ(ierr);
ierr = ISDestroy(&is_aux);CHKERRQ(ierr);
ierr = PetscMalloc1(basis_size*basis_size,&array_mat);CHKERRQ(ierr);
for (k=0;k<basis_size;k++) {
ierr = VecSet(shell_ctx->work_small_1,zero);CHKERRQ(ierr);
ierr = VecSetValue(shell_ctx->work_small_1,k,one,INSERT_VALUES);CHKERRQ(ierr);
ierr = VecAssemblyBegin(shell_ctx->work_small_1);CHKERRQ(ierr);
ierr = VecAssemblyEnd(shell_ctx->work_small_1);CHKERRQ(ierr);
ierr = MatSolve(small_mat,shell_ctx->work_small_1,shell_ctx->work_small_2);CHKERRQ(ierr);
ierr = VecGetArrayRead(shell_ctx->work_small_2,(const PetscScalar**)&array);CHKERRQ(ierr);
for (i=0;i<basis_size;i++) {
array_mat[i*basis_size+k]=array[i];
}
ierr = VecRestoreArrayRead(shell_ctx->work_small_2,(const PetscScalar**)&array);CHKERRQ(ierr);
}
ierr = MatCreateSeqDense(PETSC_COMM_SELF,basis_size,basis_size,array_mat,&inv_small_mat);CHKERRQ(ierr);
ierr = MatMatMult(shell_ctx->basis_mat,inv_small_mat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&shell_ctx->Lbasis_mat);CHKERRQ(ierr);
ierr = PetscFree(array_mat);CHKERRQ(ierr);
ierr = MatDestroy(&inv_small_mat);CHKERRQ(ierr);
ierr = MatDestroy(&small_mat);CHKERRQ(ierr);
ierr = MatScale(shell_ctx->Kbasis_mat,m_one);CHKERRQ(ierr);
/* Rebuild local PC */
ierr = KSPGetPC(local_ksp,&shell_ctx->local_pc);CHKERRQ(ierr);
ierr = PetscObjectReference((PetscObject)shell_ctx->local_pc);CHKERRQ(ierr);
ierr = PCCreate(PETSC_COMM_SELF,&newpc);CHKERRQ(ierr);
ierr = PCSetOperators(newpc,local_mat,local_mat);CHKERRQ(ierr);
ierr = PCSetType(newpc,PCSHELL);CHKERRQ(ierr);
ierr = PCShellSetContext(newpc,shell_ctx);CHKERRQ(ierr);
ierr = PCShellSetApply(newpc,PCBDDCApplyNullSpaceCorrectionPC);CHKERRQ(ierr);
ierr = PCShellSetDestroy(newpc,PCBDDCDestroyNullSpaceCorrectionPC);CHKERRQ(ierr);
ierr = PCSetUp(newpc);CHKERRQ(ierr);
ierr = KSPSetPC(local_ksp,newpc);CHKERRQ(ierr);
ierr = PCDestroy(&newpc);CHKERRQ(ierr);
ierr = KSPSetUp(local_ksp);CHKERRQ(ierr);
}
/* test */
if (pcbddc->dbg_flag && basis_dofs) {
KSP check_ksp;
PC check_pc;
Mat test_mat;
Vec work3;
PetscReal test_err,lambda_min,lambda_max;
PetscBool setsym,issym=PETSC_FALSE;
PetscInt tabs;
ierr = PetscViewerASCIIGetTab(pcbddc->dbg_viewer,&tabs);CHKERRQ(ierr);
ierr = KSPGetPC(local_ksp,&check_pc);CHKERRQ(ierr);
ierr = VecDuplicate(shell_ctx->work_full_1,&work1);CHKERRQ(ierr);
ierr = VecDuplicate(shell_ctx->work_full_1,&work2);CHKERRQ(ierr);
ierr = VecDuplicate(shell_ctx->work_full_1,&work3);CHKERRQ(ierr);
ierr = VecSetRandom(shell_ctx->work_small_1,NULL);CHKERRQ(ierr);
ierr = MatMult(shell_ctx->basis_mat,shell_ctx->work_small_1,work1);CHKERRQ(ierr);
ierr = VecCopy(work1,work2);CHKERRQ(ierr);
ierr = MatMult(local_mat,work1,work3);CHKERRQ(ierr);
ierr = PCApply(check_pc,work3,work1);CHKERRQ(ierr);
ierr = VecAXPY(work1,m_one,work2);CHKERRQ(ierr);
ierr = VecNorm(work1,NORM_INFINITY,&test_err);CHKERRQ(ierr);
ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d error for nullspace correction for ",PetscGlobalRank);CHKERRQ(ierr);
ierr = PetscViewerASCIIUseTabs(pcbddc->dbg_viewer,PETSC_FALSE);CHKERRQ(ierr);
if (isdir) {
ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Dirichlet ");CHKERRQ(ierr);
} else {
ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Neumann ");CHKERRQ(ierr);
}
ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"solver is :%1.14e\n",test_err);CHKERRQ(ierr);
ierr = PetscViewerASCIISetTab(pcbddc->dbg_viewer,tabs);CHKERRQ(ierr);
ierr = PetscViewerASCIIUseTabs(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr);
ierr = MatTransposeMatMult(shell_ctx->Lbasis_mat,shell_ctx->Kbasis_mat,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&test_mat);CHKERRQ(ierr);
ierr = MatShift(test_mat,one);CHKERRQ(ierr);
ierr = MatNorm(test_mat,NORM_INFINITY,&test_err);CHKERRQ(ierr);
ierr = MatDestroy(&test_mat);CHKERRQ(ierr);
ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d error for nullspace matrices is :%1.14e\n",PetscGlobalRank,test_err);CHKERRQ(ierr);
/* Create ksp object suitable for extreme eigenvalues' estimation */
ierr = KSPCreate(PETSC_COMM_SELF,&check_ksp);CHKERRQ(ierr);
ierr = KSPSetErrorIfNotConverged(check_ksp,pc->erroriffailure);CHKERRQ(ierr);
ierr = KSPSetOperators(check_ksp,local_mat,local_mat);CHKERRQ(ierr);
ierr = KSPSetTolerances(check_ksp,1.e-8,1.e-8,PETSC_DEFAULT,basis_dofs);CHKERRQ(ierr);
ierr = KSPSetComputeSingularValues(check_ksp,PETSC_TRUE);CHKERRQ(ierr);
ierr = MatIsSymmetricKnown(pc->pmat,&setsym,&issym);CHKERRQ(ierr);
if (issym) {
ierr = KSPSetType(check_ksp,KSPCG);CHKERRQ(ierr);
}
ierr = KSPSetPC(check_ksp,check_pc);CHKERRQ(ierr);
ierr = KSPSetUp(check_ksp);CHKERRQ(ierr);
ierr = VecSetRandom(work1,NULL);CHKERRQ(ierr);
ierr = MatMult(local_mat,work1,work2);CHKERRQ(ierr);
ierr = KSPSolve(check_ksp,work2,work2);CHKERRQ(ierr);
ierr = VecAXPY(work2,m_one,work1);CHKERRQ(ierr);
ierr = VecNorm(work2,NORM_INFINITY,&test_err);CHKERRQ(ierr);
ierr = KSPComputeExtremeSingularValues(check_ksp,&lambda_max,&lambda_min);CHKERRQ(ierr);
ierr = KSPGetIterationNumber(check_ksp,&k);CHKERRQ(ierr);
ierr = PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d error for adapted KSP %1.14e (it %d, eigs %1.6e %1.6e)\n",PetscGlobalRank,test_err,k,lambda_min,lambda_max);CHKERRQ(ierr);
ierr = KSPDestroy(&check_ksp);CHKERRQ(ierr);
ierr = VecDestroy(&work1);CHKERRQ(ierr);
ierr = VecDestroy(&work2);CHKERRQ(ierr);
ierr = VecDestroy(&work3);CHKERRQ(ierr);
}
/* all processes shoud call this, even the void ones */
if (pcbddc->dbg_flag) {
ierr = PetscViewerFlush(pcbddc->dbg_viewer);CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
int main(int argc,char **argv)
{
Mat A,B,C,D;
PetscInt i,M=10,N=5,j,nrows,ncols,am,an,rstart,rend;
PetscErrorCode ierr;
PetscRandom r;
PetscBool equal,iselemental;
PetscReal fill = 1.0;
IS isrows,iscols;
const PetscInt *rows,*cols;
PetscScalar *v,rval;
#if defined(PETSC_HAVE_ELEMENTAL)
PetscBool Test_MatMatMult=PETSC_TRUE;
#else
PetscBool Test_MatMatMult=PETSC_FALSE;
#endif
PetscMPIInt size;
ierr = PetscInitialize(&argc,&argv,(char*)0,help);if (ierr) return ierr;
ierr = MPI_Comm_size(PETSC_COMM_WORLD,&size);CHKERRQ(ierr);
ierr = PetscOptionsGetInt(NULL,NULL,"-M",&M,NULL);CHKERRQ(ierr);
ierr = PetscOptionsGetInt(NULL,NULL,"-N",&N,NULL);CHKERRQ(ierr);
ierr = MatCreate(PETSC_COMM_WORLD,&A);CHKERRQ(ierr);
ierr = MatSetSizes(A,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr);
ierr = MatSetType(A,MATDENSE);CHKERRQ(ierr);
ierr = MatSetFromOptions(A);CHKERRQ(ierr);
ierr = MatSetUp(A);CHKERRQ(ierr);
ierr = PetscRandomCreate(PETSC_COMM_WORLD,&r);CHKERRQ(ierr);
ierr = PetscRandomSetFromOptions(r);CHKERRQ(ierr);
/* Set local matrix entries */
ierr = MatGetOwnershipIS(A,&isrows,&iscols);CHKERRQ(ierr);
ierr = ISGetLocalSize(isrows,&nrows);CHKERRQ(ierr);
ierr = ISGetIndices(isrows,&rows);CHKERRQ(ierr);
ierr = ISGetLocalSize(iscols,&ncols);CHKERRQ(ierr);
ierr = ISGetIndices(iscols,&cols);CHKERRQ(ierr);
ierr = PetscMalloc1(nrows*ncols,&v);CHKERRQ(ierr);
for (i=0; i<nrows; i++) {
for (j=0; j<ncols; j++) {
ierr = PetscRandomGetValue(r,&rval);CHKERRQ(ierr);
v[i*ncols+j] = rval;
}
}
ierr = MatSetValues(A,nrows,rows,ncols,cols,v,INSERT_VALUES);CHKERRQ(ierr);
ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = ISRestoreIndices(isrows,&rows);CHKERRQ(ierr);
ierr = ISRestoreIndices(iscols,&cols);CHKERRQ(ierr);
ierr = ISDestroy(&isrows);CHKERRQ(ierr);
ierr = ISDestroy(&iscols);CHKERRQ(ierr);
ierr = PetscRandomDestroy(&r);CHKERRQ(ierr);
/* Test MatTranspose() */
ierr = MatCreateTranspose(A,&C);CHKERRQ(ierr);
ierr = MatTranspose(A,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); /* B = A^T */
ierr = MatMultEqual(C,B,10,&equal);CHKERRQ(ierr);
if (!equal) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"A^T*x != (x^T*A)^T");
ierr = MatTranspose(A,MAT_REUSE_MATRIX,&B);CHKERRQ(ierr); /* B = A^T */
ierr = MatMultEqual(C,B,10,&equal);CHKERRQ(ierr);
if (!equal) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"A^T*x != (x^T*A)^T");
ierr = MatDestroy(&B);CHKERRQ(ierr);
ierr = MatDuplicate(A,MAT_COPY_VALUES,&B);CHKERRQ(ierr);
ierr = MatTranspose(B,MAT_INPLACE_MATRIX,&B);CHKERRQ(ierr);
ierr = MatMultEqual(C,B,10,&equal);CHKERRQ(ierr);
if (!equal) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"A^T*x != (x^T*A)^T");
ierr = MatDestroy(&B);CHKERRQ(ierr);
ierr = MatDestroy(&C);CHKERRQ(ierr);
/* Test MatMatMult() */
if (Test_MatMatMult) {
#if !defined(PETSC_HAVE_ELEMENTAL)
if (size > 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"This test requires ELEMENTAL");
#endif
ierr = MatTranspose(A,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); /* B = A^T */
ierr = MatMatMult(B,A,MAT_INITIAL_MATRIX,fill,&C);CHKERRQ(ierr); /* C = B*A = A^T*A */
ierr = MatMatMult(B,A,MAT_REUSE_MATRIX,fill,&C);CHKERRQ(ierr);
/* Test MatDuplicate for matrix product */
ierr = MatDuplicate(C,MAT_COPY_VALUES,&D);CHKERRQ(ierr);
ierr = MatDestroy(&D);CHKERRQ(ierr);
/* Test B*A*x = C*x for n random vector x */
ierr = MatMatMultEqual(B,A,C,10,&equal);CHKERRQ(ierr);
if (!equal) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"B*A*x != C*x");
ierr = MatDestroy(&C);CHKERRQ(ierr);
ierr = MatMatMultSymbolic(B,A,fill,&C);CHKERRQ(ierr);
for (i=0; i<2; i++) {
/* Repeat the numeric product to test reuse of the previous symbolic product */
ierr = MatMatMultNumeric(B,A,C);CHKERRQ(ierr);
ierr = MatMatMultEqual(B,A,C,10,&equal);CHKERRQ(ierr);
if (!equal) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"B*A*x != C*x");
}
ierr = MatDestroy(&C);CHKERRQ(ierr);
ierr = MatDestroy(&B);CHKERRQ(ierr);
}
/* Test MatTransposeMatMult() */
ierr = PetscObjectTypeCompare((PetscObject)A,MATELEMENTAL,&iselemental);CHKERRQ(ierr);
if (!iselemental) {
ierr = MatTransposeMatMult(A,A,MAT_INITIAL_MATRIX,fill,&D);CHKERRQ(ierr); /* D = A^T*A */
ierr = MatTransposeMatMult(A,A,MAT_REUSE_MATRIX,fill,&D);CHKERRQ(ierr);
/* Test MatDuplicate for matrix product */
ierr = MatDuplicate(D,MAT_COPY_VALUES,&C);CHKERRQ(ierr);
ierr = MatDestroy(&C);CHKERRQ(ierr);
/* ierr = MatView(D,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr); */
ierr = MatTransposeMatMultEqual(A,A,D,10,&equal);CHKERRQ(ierr);
if (!equal) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"D*x != A^T*A*x");
ierr = MatDestroy(&D);CHKERRQ(ierr);
/* Test D*x = A^T*C*A*x, where C is in AIJ format */
ierr = MatGetLocalSize(A,&am,&an);CHKERRQ(ierr);
ierr = MatCreate(PETSC_COMM_WORLD,&C);CHKERRQ(ierr);
if (size == 1) {
ierr = MatSetSizes(C,PETSC_DECIDE,PETSC_DECIDE,am,am);CHKERRQ(ierr);
} else {
ierr = MatSetSizes(C,am,am,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
}
ierr = MatSetFromOptions(C);CHKERRQ(ierr);
ierr = MatSetUp(C);CHKERRQ(ierr);
ierr = MatGetOwnershipRange(C,&rstart,&rend);CHKERRQ(ierr);
v[0] = 1.0;
for (i=rstart; i<rend; i++) {
ierr = MatSetValues(C,1,&i,1,&i,v,INSERT_VALUES);CHKERRQ(ierr);
}
ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
/* B = C*A, D = A^T*B */
ierr = MatMatMult(C,A,MAT_INITIAL_MATRIX,1.0,&B);CHKERRQ(ierr);
ierr = MatTransposeMatMult(A,B,MAT_INITIAL_MATRIX,fill,&D);CHKERRQ(ierr);
ierr = MatTransposeMatMultEqual(A,B,D,10,&equal);CHKERRQ(ierr);
if (!equal) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"D*x != A^T*B*x");
ierr = MatDestroy(&D);CHKERRQ(ierr);
ierr = MatDestroy(&C);CHKERRQ(ierr);
ierr = MatDestroy(&B);CHKERRQ(ierr);
}
ierr = MatDestroy(&A);CHKERRQ(ierr);
ierr = PetscFree(v);CHKERRQ(ierr);
ierr = PetscFinalize();
return ierr;
}
int main(int argc,char **argv)
{
Mat A,B,C,D;
PetscInt i,M=10,N=5,j,nrows,ncols;
PetscErrorCode ierr;
PetscRandom r;
PetscBool equal,iselemental;
PetscReal fill = 1.0;
IS isrows,iscols;
const PetscInt *rows,*cols;
PetscScalar *v,rval;
PetscInitialize(&argc,&argv,(char*)0,help);
ierr = PetscOptionsGetInt(NULL,"-M",&M,NULL);CHKERRQ(ierr);
ierr = PetscOptionsGetInt(NULL,"-N",&N,NULL);CHKERRQ(ierr);
ierr = MatCreate(PETSC_COMM_WORLD,&A);CHKERRQ(ierr);
ierr = MatSetSizes(A,PETSC_DECIDE,PETSC_DECIDE,M,N);CHKERRQ(ierr);
ierr = MatSetType(A,MATDENSE);CHKERRQ(ierr);
ierr = MatSetFromOptions(A);CHKERRQ(ierr);
ierr = MatSetUp(A);CHKERRQ(ierr);
ierr = PetscRandomCreate(PETSC_COMM_WORLD,&r);CHKERRQ(ierr);
ierr = PetscRandomSetFromOptions(r);CHKERRQ(ierr);
/* Set local matrix entries */
ierr = MatGetOwnershipIS(A,&isrows,&iscols);CHKERRQ(ierr);
ierr = ISGetLocalSize(isrows,&nrows);CHKERRQ(ierr);
ierr = ISGetIndices(isrows,&rows);CHKERRQ(ierr);
ierr = ISGetLocalSize(iscols,&ncols);CHKERRQ(ierr);
ierr = ISGetIndices(iscols,&cols);CHKERRQ(ierr);
ierr = PetscMalloc(nrows*ncols*sizeof(*v),&v);CHKERRQ(ierr);
for (i=0; i<nrows; i++) {
for (j=0; j<ncols; j++) {
ierr = PetscRandomGetValue(r,&rval);CHKERRQ(ierr);
v[i*ncols+j] = rval;
}
}
ierr = MatSetValues(A,nrows,rows,ncols,cols,v,INSERT_VALUES);CHKERRQ(ierr);
ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = ISRestoreIndices(isrows,&rows);CHKERRQ(ierr);
ierr = ISRestoreIndices(iscols,&cols);CHKERRQ(ierr);
ierr = ISDestroy(&isrows);CHKERRQ(ierr);
ierr = ISDestroy(&iscols);CHKERRQ(ierr);
ierr = PetscFree(v);CHKERRQ(ierr);
ierr = PetscRandomDestroy(&r);CHKERRQ(ierr);
/* Test MatMatMult() */
ierr = MatTranspose(A,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr); /* B = A^T */
ierr = MatMatMult(B,A,MAT_INITIAL_MATRIX,fill,&C);CHKERRQ(ierr); /* C = B*A = A^T*A */
ierr = MatMatMultSymbolic(B,A,fill,&D);CHKERRQ(ierr); /* D = B*A = A^T*A */
for (i=0; i<2; i++) {
/* Repeat the numeric product to test reuse of the previous symbolic product */
ierr = MatMatMultNumeric(B,A,D);CHKERRQ(ierr);
}
ierr = MatMultEqual(C,D,10,&equal);CHKERRQ(ierr);
if (!equal) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"C != D");
ierr = MatDestroy(&D);CHKERRQ(ierr);
/* Test MatTransposeMatMult() */
ierr = PetscObjectTypeCompare((PetscObject)A,MATELEMENTAL,&iselemental);CHKERRQ(ierr);
if (!iselemental) {
ierr = MatTransposeMatMult(A,A,MAT_INITIAL_MATRIX,fill,&D);CHKERRQ(ierr); /* D = A^T*A */
ierr = MatEqual(C,D,&equal);CHKERRQ(ierr);
if (!equal) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"C != D");
ierr = MatDestroy(&D);CHKERRQ(ierr);
}
ierr = MatDestroy(&C);CHKERRQ(ierr);
ierr = MatDestroy(&B);
ierr = MatDestroy(&A);
PetscFinalize();
return(0);
}
PetscErrorCode bsscr_buildK2(KSP ksp){
KSP_BSSCR *bsscr = (KSP_BSSCR *)ksp->data;
Mat K,G,M, K2=0;
Vec f2=0;
PetscReal minD,maxD;
//MatStokesBlockScaling BA=bsscr->BA;
Stokes_SLE* stokesSLE = (Stokes_SLE*)bsscr->solver->st_sle;
StokesBlockKSPInterface* Solver = bsscr->solver;
//PetscErrorCode ierr;
PetscFunctionBegin;
K = stokesSLE->kStiffMat->matrix;
G = stokesSLE->gStiffMat->matrix;
if(bsscr->K2){
Stg_MatDestroy(&bsscr->K2);
bsscr->K2 = PETSC_NULL;
}
switch (bsscr->k2type) {
case (K2_DGMGD):
{/* Should only do this one if scaling is turned off. */
Vec D=0;
Vec MD = 0;
Mat GKG = 0;
Mat KG = 0;
Mat Mscale = 0;
//AugLagStokes_SLE * stokesSLE = (AugLagStokes_SLE*)bsscr->st_sle;
if (Solver->mStiffMat){
MatMatMult(K, G, MAT_INITIAL_MATRIX, PETSC_DEFAULT , &KG);
MatTransposeMatMult(G, KG, MAT_INITIAL_MATRIX, PETSC_DEFAULT ,&GKG);
MatGetVecs( GKG, PETSC_NULL, &MD );
MatGetDiagonal( GKG, MD );
M = Solver->mStiffMat->matrix;
MatDuplicate(M, MAT_COPY_VALUES, &Mscale );
MatDiagonalScale(Mscale, NULL, MD );
VecMin(MD, PETSC_NULL, &minD);
VecMax(MD, PETSC_NULL, &maxD);
VecScale(MD, 1.0/maxD);
VecMin(MD, PETSC_NULL, &minD);
VecMax(MD, PETSC_NULL, &maxD);
bsscr_GMiGt(&K2,K,G,Mscale); /* K2 created */
Stg_VecDestroy(&D);
Stg_VecDestroy(&MD);
Stg_MatDestroy(&GKG);
Stg_MatDestroy(&KG);
PetscPrintf( PETSC_COMM_WORLD, "\n\n----- K2_DGMGD ------");
PetscPrintf( PETSC_COMM_WORLD, " min %f, max %f \n\n", minD, maxD);
}
else{
PetscPrintf( PETSC_COMM_WORLD,"The Augmented Lagrangian Method DGMGD was specified but the SLE has no mStiffMat on it.\n");
PetscPrintf( PETSC_COMM_WORLD,"You need to use the AugLagStokes_SLE class.\n"); }
}
break;
case (K2_GMG):
{
//AugLagStokes_SLE * stokesSLE = (AugLagStokes_SLE*)bsscr->st_sle;
if (Solver->mStiffMat){
M = Solver->mStiffMat->matrix;
bsscr_GMiGt(&K2,K,G,M); /* K2 created */
PetscPrintf( PETSC_COMM_WORLD, "\n\n----- K2_GMG ------\n\n");
}else{
PetscPrintf( PETSC_COMM_WORLD,"The Augmented Lagrangian Method GMG was specified but the SLE has no mStiffMat on it.\n");
PetscPrintf( PETSC_COMM_WORLD,"You need to use the AugLagStokes_SLE class.\n");
}
}
break;
case (K2_GG):
{
bsscr_GGt(&K2,K,G); /* K2 created */
PetscPrintf( PETSC_COMM_WORLD, "\n\n----- K2_GG ------\n\n");
}
break;
case (K2_SLE):
{
//AugLagStokes_SLE * stokesSLE = (AugLagStokes_SLE*)bsscr->st_sle;
if (Solver->mStiffMat){
K2 = Solver->mStiffMat->matrix;
if(Solver->jForceVec){ f2 = Solver->jForceVec->vector; }
PetscPrintf( PETSC_COMM_WORLD, "\n\n----- K2_SLE ------\n\n");
}else{
PetscPrintf( PETSC_COMM_WORLD,"The Augmented Lagrangian Method SLE was specified but the Solver has no Matrix on it.\n");
PetscPrintf( PETSC_COMM_WORLD,"You need to set the K2 matrix on the StokesBlockKSPInterface class.\n");
}
}
break;
default:
PetscPrintf( PETSC_COMM_WORLD, "\n\n----- NO K2 ------\n\n");
}
bsscr->f2 = f2;
bsscr->K2 = K2;
PetscFunctionReturn(0);
}
int main(int argc,char **args)
{
Mat A,C,Bdense,Cdense;
PetscErrorCode ierr;
PetscViewer fd; /* viewer */
char file[PETSC_MAX_PATH_LEN]; /* input file name */
PetscBool flg,viewmats=PETSC_FALSE;
PetscMPIInt rank,size;
PetscReal fill=1.0;
PetscInt m,n,i,j,BN=10,rstart,rend,*rows,*cols;
PetscScalar *Barray,*Carray,rval,*array;
Vec x,y;
PetscRandom rand;
PetscInitialize(&argc,&args,(char*)0,help);
ierr = MPI_Comm_rank(PETSC_COMM_WORLD,&rank);CHKERRQ(ierr);
/* Determine file from which we read the matrix A */
ierr = PetscOptionsGetString(NULL,"-f",file,PETSC_MAX_PATH_LEN,&flg);CHKERRQ(ierr);
if (!flg) SETERRQ(PETSC_COMM_WORLD,1,"Must indicate binary file with the -f option");
/* Load matrix A */
ierr = PetscViewerBinaryOpen(PETSC_COMM_WORLD,file,FILE_MODE_READ,&fd);CHKERRQ(ierr);
ierr = MatCreate(PETSC_COMM_WORLD,&A);CHKERRQ(ierr);
ierr = MatLoad(A,fd);CHKERRQ(ierr);
ierr = PetscViewerDestroy(&fd);CHKERRQ(ierr);
/* Print (for testing only) */
ierr = PetscOptionsHasName(NULL, "-view_mats", &viewmats);CHKERRQ(ierr);
if (viewmats) {
if (!rank) printf("A_aij:\n");
ierr = MatView(A,0);CHKERRQ(ierr);
}
/* Test MatTransposeMatMult_aij_aij() */
ierr = MatTransposeMatMult(A,A,MAT_INITIAL_MATRIX,fill,&C);CHKERRQ(ierr);
if (viewmats) {
if (!rank) printf("\nC = A_aij^T * A_aij:\n");
ierr = MatView(C,0);CHKERRQ(ierr);
}
ierr = MatDestroy(&C);CHKERRQ(ierr);
ierr = MatGetLocalSize(A,&m,&n);CHKERRQ(ierr);
/* create a dense matrix Bdense */
ierr = MatCreate(PETSC_COMM_WORLD,&Bdense);CHKERRQ(ierr);
ierr = MatSetSizes(Bdense,m,PETSC_DECIDE,PETSC_DECIDE,BN);CHKERRQ(ierr);
ierr = MatSetType(Bdense,MATDENSE);CHKERRQ(ierr);
ierr = MatSetFromOptions(Bdense);CHKERRQ(ierr);
ierr = MatSetUp(Bdense);CHKERRQ(ierr);
ierr = MatGetOwnershipRange(Bdense,&rstart,&rend);CHKERRQ(ierr);
//printf("[%d] rstart/end %d %d; local size %d %d\n",rank,rstart,rend,m,n);
ierr = PetscMalloc3(m,PetscInt,&rows,BN,PetscInt,&cols,m*BN,PetscScalar,&array);CHKERRQ(ierr);
for (i=0; i<m; i++) rows[i] = rstart + i;
ierr = PetscRandomCreate(PETSC_COMM_WORLD,&rand);CHKERRQ(ierr);
ierr = PetscRandomSetFromOptions(rand);CHKERRQ(ierr);
for (j=0; j<BN; j++) {
cols[j] = j;
for (i=0; i<m; i++) {
ierr = PetscRandomGetValue(rand,&rval);CHKERRQ(ierr);
array[m*j+i] = rval;
}
}
ierr = MatSetValues(Bdense,m,rows,BN,cols,array,INSERT_VALUES);CHKERRQ(ierr);
ierr = MatAssemblyBegin(Bdense,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(Bdense,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = PetscRandomDestroy(&rand);CHKERRQ(ierr);
ierr = PetscFree3(rows,cols,array);CHKERRQ(ierr);
if (viewmats) {
if (!rank) printf("\nBdense:\n");
ierr = MatView(Bdense,0);CHKERRQ(ierr);
}
/* Test MatTransposeMatMult_aij_dense() */
ierr = MatTransposeMatMult(A,Bdense,MAT_INITIAL_MATRIX,fill,&C);CHKERRQ(ierr);
ierr = MatTransposeMatMult(A,Bdense,MAT_REUSE_MATRIX,fill,&C);CHKERRQ(ierr);
if (viewmats) {
if (!rank) printf("\nC=A^T*Bdense:\n");
ierr = MatView(C,0);CHKERRQ(ierr);
}
/* Check accuracy */
ierr = MatCreate(PETSC_COMM_WORLD,&Cdense);CHKERRQ(ierr);
ierr = MatSetSizes(Cdense,n,PETSC_DECIDE,PETSC_DECIDE,BN);CHKERRQ(ierr);
ierr = MatSetType(Cdense,MATDENSE);CHKERRQ(ierr);
ierr = MatSetFromOptions(Cdense);CHKERRQ(ierr);
ierr = MatSetUp(Cdense);CHKERRQ(ierr);
ierr = MatAssemblyBegin(Cdense,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(Cdense,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MPI_Comm_size(PETSC_COMM_WORLD,&size);CHKERRQ(ierr);
if (size == 1) {
ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,m,NULL,&x);CHKERRQ(ierr);
ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,n,NULL,&y);CHKERRQ(ierr);
} else {
ierr = VecCreateMPIWithArray(PETSC_COMM_WORLD,1,m,PETSC_DECIDE,NULL,&x);CHKERRQ(ierr);
ierr = VecCreateMPIWithArray(PETSC_COMM_WORLD,1,n,PETSC_DECIDE,NULL,&y);CHKERRQ(ierr);
}
/* Cdense[:,j] = A^T * Bdense[:,j] */
ierr = MatDenseGetArray(Bdense,&Barray);CHKERRQ(ierr);
ierr = MatDenseGetArray(Cdense,&Carray);CHKERRQ(ierr);
for (j=0; j<BN; j++) {
ierr = VecPlaceArray(x,Barray);CHKERRQ(ierr);
ierr = VecPlaceArray(y,Carray);CHKERRQ(ierr);
ierr = MatMultTranspose(A,x,y);CHKERRQ(ierr);
ierr = VecResetArray(x);CHKERRQ(ierr);
ierr = VecResetArray(y);CHKERRQ(ierr);
Barray += m;
Carray += n;
}
ierr = MatDenseRestoreArray(Bdense,&Barray);CHKERRQ(ierr);
ierr = MatDenseRestoreArray(Cdense,&Carray);CHKERRQ(ierr);
if (viewmats) {
if (!rank) printf("\nCdense:\n");
ierr = MatView(Cdense,0);CHKERRQ(ierr);
}
ierr = MatEqual(C,Cdense,&flg);CHKERRQ(ierr);
if (!flg) {
if (!rank) printf(" C != Cdense\n");
}
/* Free data structures */
ierr = MatDestroy(&A);CHKERRQ(ierr);
ierr = MatDestroy(&C);CHKERRQ(ierr);
ierr = MatDestroy(&Bdense);CHKERRQ(ierr);
ierr = MatDestroy(&Cdense);CHKERRQ(ierr);
ierr = VecDestroy(&x);CHKERRQ(ierr);
ierr = VecDestroy(&y);CHKERRQ(ierr);
ierr = PetscFinalize();
return 0;
}
static void PETScMatvecGenColumnMajor(void *x, PRIMME_INT ldx, void *y, PRIMME_INT ldy,
int blockSize, int trans, Mat matrix, MPI_Comm comm) {
PetscInt m, n, mLocal, nLocal;
PetscErrorCode ierr;
Mat X, Y, X0, Y0;
int xcompact, ycompact;
if (blockSize == 1) {
PETScMatvecGenNoBlock(x, ldx, y, ldy, blockSize, trans, matrix, comm);
return;
}
assert(sizeof(PetscScalar) == sizeof(SCALAR));
ierr = MatGetSize(matrix, &m, &n); CHKERRABORT(comm, ierr);
ierr = MatGetLocalSize(matrix, &mLocal, &nLocal); CHKERRABORT(comm, ierr);
if (trans == 0) {
ierr = MatCreateDense(comm,nLocal,PETSC_DECIDE,n,blockSize,x,&X);CHKERRABORT(comm, ierr);
ierr = MatCreateDense(comm,mLocal,PETSC_DECIDE,m,blockSize,y,&Y);CHKERRABORT(comm, ierr);
xcompact = nLocal == ldx;
ycompact = mLocal == ldy;
}
else {
ierr = MatCreateDense(comm,mLocal,PETSC_DECIDE,m,blockSize,x,&X);CHKERRABORT(comm, ierr);
ierr = MatCreateDense(comm,nLocal,PETSC_DECIDE,n,blockSize,y,&Y);CHKERRABORT(comm, ierr);
xcompact = mLocal == ldx;
ycompact = nLocal == ldy;
}
ierr = MatDenseGetLocalMatrix(X, &X0);CHKERRABORT(comm, ierr);
ierr = MatSeqDenseSetLDA(X0, (PetscInt)ldx);CHKERRABORT(comm, ierr);
ierr = MatDenseGetLocalMatrix(Y, &Y0);CHKERRABORT(comm, ierr);
ierr = MatSeqDenseSetLDA(Y0, (PetscInt)ldy);CHKERRABORT(comm, ierr);
/* MatMatMult doesn't support X to be non-contiguous */
if (xcompact) {
X0 = X;
}
else {
ierr = MatDuplicate(X, MAT_COPY_VALUES, &X0);CHKERRABORT(comm, ierr);
}
if (trans == 0) {
if (ycompact) {
ierr = MatMatMult(matrix, X0, MAT_REUSE_MATRIX, PETSC_DEFAULT, &Y); CHKERRABORT(comm, ierr);
}
else {
ierr = MatMatMult(matrix, X0, MAT_INITIAL_MATRIX, PETSC_DEFAULT, &Y0); CHKERRABORT(comm, ierr);
ierr = MatCopy(Y0, Y, SAME_NONZERO_PATTERN); CHKERRABORT(comm, ierr);
ierr = MatDestroy(&Y0); CHKERRABORT(comm, ierr);
}
}
else {
/* A^H*X is not implemented in PETSc, do instead (A^T*X^c)^c */
ierr = MatConjugate(X0); CHKERRABORT(comm, ierr);
if (ycompact) {
ierr = MatTransposeMatMult(matrix, X0, MAT_REUSE_MATRIX, PETSC_DEFAULT, &Y); CHKERRABORT(comm, ierr);
}
else {
ierr = MatTransposeMatMult(matrix, X0, MAT_INITIAL_MATRIX, PETSC_DEFAULT, &Y0); CHKERRABORT(comm, ierr);
ierr = MatCopy(Y0, Y, SAME_NONZERO_PATTERN); CHKERRABORT(comm, ierr);
ierr = MatDestroy(&Y0); CHKERRABORT(comm, ierr);
}
ierr = MatConjugate(Y); CHKERRABORT(comm, ierr);
if (xcompact) {
ierr = MatConjugate(X0); CHKERRABORT(comm, ierr);
}
}
if (!xcompact) {
ierr = MatDestroy(&X0);CHKERRABORT(comm, ierr);
}
ierr = MatDestroy(&X);CHKERRABORT(comm, ierr);
ierr = MatDestroy(&Y);CHKERRABORT(comm, ierr);
}
