/*
Function: hypre_CSRMatrixEliminateRowsCols
Eliminate rows and columns of A, store eliminated values in Ae.
If 'diag' is nonzero, the eliminated diagonal of A is set to identity.
If 'col_remap' is not NULL it specifies renumbering of columns of Ae.
*/
void hypre_CSRMatrixEliminateRowsCols(hypre_CSRMatrix *A,
hypre_CSRMatrix *Ae,
HYPRE_Int nrows, HYPRE_Int *rows,
HYPRE_Int ncols, HYPRE_Int *cols,
int diag, HYPRE_Int* col_remap)
{
HYPRE_Int i, j, k, col;
HYPRE_Int A_beg, Ae_beg, A_end;
HYPRE_Real a;
HYPRE_Int *A_i = hypre_CSRMatrixI(A);
HYPRE_Int *A_j = hypre_CSRMatrixJ(A);
HYPRE_Real *A_data = hypre_CSRMatrixData(A);
HYPRE_Int A_rows = hypre_CSRMatrixNumRows(A);
HYPRE_Int *Ae_i = hypre_CSRMatrixI(Ae);
HYPRE_Int *Ae_j = hypre_CSRMatrixJ(Ae);
HYPRE_Real *Ae_data = hypre_CSRMatrixData(Ae);
for (i = 0; i < A_rows; i++)
{
A_beg = A_i[i];
A_end = A_i[i+1];
Ae_beg = Ae_i[i];
if (hypre_BinarySearch(rows, i, nrows) >= 0)
{
/* eliminate row */
for (j = A_beg, k = Ae_beg; j < A_end; j++, k++)
{
col = A_j[j];
Ae_j[k] = col_remap ? col_remap[col] : col;
a = (diag && col == i) ? 1.0 : 0.0;
Ae_data[k] = A_data[j] - a;
A_data[j] = a;
}
}
else
{
/* eliminate columns */
for (j = A_beg, k = Ae_beg; j < A_end; j++)
{
col = A_j[j];
if (hypre_BinarySearch(cols, col, ncols) >= 0)
{
Ae_j[k] = col_remap ? col_remap[col] : col;
Ae_data[k] = A_data[j];
A_data[j] = 0.0;
k++;
}
}
}
}
}
/*
Function: hypre_CSRMatrixElimCreate
Prepare the Ae matrix: count nnz, initialize I, allocate J and data.
*/
void hypre_CSRMatrixElimCreate(hypre_CSRMatrix *A,
hypre_CSRMatrix *Ae,
HYPRE_Int nrows, HYPRE_Int *rows,
HYPRE_Int ncols, HYPRE_Int *cols,
HYPRE_Int *col_mark)
{
HYPRE_Int i, j, col;
HYPRE_Int A_beg, A_end;
HYPRE_Int *A_i = hypre_CSRMatrixI(A);
HYPRE_Int *A_j = hypre_CSRMatrixJ(A);
HYPRE_Int A_rows = hypre_CSRMatrixNumRows(A);
hypre_CSRMatrixI(Ae) = hypre_TAlloc(HYPRE_Int, A_rows+1);
HYPRE_Int *Ae_i = hypre_CSRMatrixI(Ae);
HYPRE_Int nnz = 0;
for (i = 0; i < A_rows; i++)
{
Ae_i[i] = nnz;
A_beg = A_i[i];
A_end = A_i[i+1];
if (hypre_BinarySearch(rows, i, nrows) >= 0)
{
/* full row */
nnz += A_end - A_beg;
if (col_mark)
{
for (j = A_beg; j < A_end; j++)
{
col_mark[A_j[j]] = 1;
}
}
}
else
{
/* count columns */
for (j = A_beg; j < A_end; j++)
{
col = A_j[j];
if (hypre_BinarySearch(cols, col, ncols) >= 0)
{
nnz++;
if (col_mark) {
col_mark[col] = 1;
}
}
}
}
}
Ae_i[A_rows] = nnz;
hypre_CSRMatrixJ(Ae) = hypre_TAlloc(HYPRE_Int, nnz);
hypre_CSRMatrixData(Ae) = hypre_TAlloc(HYPRE_Real, nnz);
hypre_CSRMatrixNumNonzeros(Ae) = nnz;
}
/* coarse (marked >=0) rows of P copied from C Both matrices have the same sizes. */
void hypre_ParCSRMatrixCopy_C( hypre_ParCSRMatrix * P,
hypre_ParCSRMatrix * C, HYPRE_Int * CF_marker )
{
hypre_CSRMatrix *C_diag = hypre_ParCSRMatrixDiag(C);
hypre_CSRMatrix *C_offd = hypre_ParCSRMatrixOffd(C);
hypre_CSRMatrix *P_diag = hypre_ParCSRMatrixDiag(P);
hypre_CSRMatrix *P_offd = hypre_ParCSRMatrixOffd(P);
double *C_diag_data = hypre_CSRMatrixData(C_diag);
HYPRE_Int *C_diag_i = hypre_CSRMatrixI(C_diag);
double *C_offd_data = hypre_CSRMatrixData(C_offd);
HYPRE_Int *C_offd_i = hypre_CSRMatrixI(C_offd);
double *P_diag_data = hypre_CSRMatrixData(P_diag);
double *P_offd_data = hypre_CSRMatrixData(P_offd);
HYPRE_Int num_rows_diag_C = hypre_CSRMatrixNumRows(C_diag);
HYPRE_Int num_rows_offd_C = hypre_CSRMatrixNumRows(C_offd);
HYPRE_Int num_cols_offd_C = hypre_CSRMatrixNumCols(C_offd);
HYPRE_Int i1, m;
for ( i1= 0; i1 < num_rows_diag_C; i1++ )
{
if ( CF_marker[i1] >= 0 ) /* Coarse rows only */
{
for ( m=C_diag_i[i1]; m<C_diag_i[i1+1]; ++m )
{
P_diag_data[m] = C_diag_data[m];
}
}
}
if ( num_cols_offd_C )
for ( i1= 0; i1 < num_rows_offd_C; i1++ )
{
if ( CF_marker[i1] >= 0 ) /* Coarse rows only */
{
for ( m=C_offd_i[i1]; m<C_offd_i[i1+1]; ++m )
{
P_offd_data[m] = C_offd_data[m];
}
}
}
}
/* fine (marked <0 ) rows of Pnew set to 0 */
void hypre_ParCSRMatrixZero_F( hypre_ParCSRMatrix * P, HYPRE_Int * CF_marker )
{
hypre_CSRMatrix *P_diag = hypre_ParCSRMatrixDiag(P);
hypre_CSRMatrix *P_offd = hypre_ParCSRMatrixOffd(P);
double *P_diag_data = hypre_CSRMatrixData(P_diag);
HYPRE_Int *P_diag_i = hypre_CSRMatrixI(P_diag);
double *P_offd_data = hypre_CSRMatrixData(P_offd);
HYPRE_Int *P_offd_i = hypre_CSRMatrixI(P_offd);
HYPRE_Int num_rows_diag_P = hypre_CSRMatrixNumRows(P_diag);
HYPRE_Int num_rows_offd_P = hypre_CSRMatrixNumRows(P_offd);
HYPRE_Int num_cols_offd_P = hypre_CSRMatrixNumCols(P_offd);
HYPRE_Int i1, m;
for ( i1= 0; i1 < num_rows_diag_P; i1++ )
{
if ( CF_marker[i1] < 0 ) /* Fine rows only */
{
for ( m=P_diag_i[i1]; m<P_diag_i[i1+1]; ++m )
{
P_diag_data[m] = 0;
}
}
}
if ( num_cols_offd_P )
for ( i1= 0; i1 < num_rows_offd_P; i1++ )
{
if ( CF_marker[i1] < 0 ) /* Fine rows only */
{
for ( m=P_offd_i[i1]; m<P_offd_i[i1+1]; ++m )
{
P_offd_data[m] = 0;
}
}
}
}
/*
Function: hypre_CSRMatrixEliminateOffdRowsAXB
Eliminate (zero) the given list of rows of A.
*/
void hypre_CSRMatrixEliminateOffdRowsAXB(hypre_CSRMatrix *A,
HYPRE_Int nrows_to_eliminate,
HYPRE_Int *rows_to_eliminate)
{
HYPRE_Int *Ai = hypre_CSRMatrixI(A);
HYPRE_Real *Adata = hypre_CSRMatrixData(A);
HYPRE_Int i, j;
HYPRE_Int irow, ibeg, iend;
for (i = 0; i < nrows_to_eliminate; i++)
{
irow = rows_to_eliminate[i];
ibeg = Ai[irow];
iend = Ai[irow+1];
for (j = ibeg; j < iend; j++)
{
Adata[j] = 0.0;
}
}
}
HYPRE_Int
HYPRE_ParCSRDiagScale( HYPRE_Solver solver,
HYPRE_ParCSRMatrix HA,
HYPRE_ParVector Hy,
HYPRE_ParVector Hx )
{
hypre_ParCSRMatrix *A = (hypre_ParCSRMatrix *) HA;
hypre_ParVector *y = (hypre_ParVector *) Hy;
hypre_ParVector *x = (hypre_ParVector *) Hx;
double *x_data = hypre_VectorData(hypre_ParVectorLocalVector(x));
double *y_data = hypre_VectorData(hypre_ParVectorLocalVector(y));
double *A_data = hypre_CSRMatrixData(hypre_ParCSRMatrixDiag(A));
HYPRE_Int *A_i = hypre_CSRMatrixI(hypre_ParCSRMatrixDiag(A));
HYPRE_Int local_size = hypre_VectorSize(hypre_ParVectorLocalVector(x));
HYPRE_Int i, ierr = 0;
for (i=0; i < local_size; i++)
{
x_data[i] = y_data[i]/A_data[A_i[i]];
}
return ierr;
}
int
hypre_CSRMatrixMatvec( double alpha,
hypre_CSRMatrix *A,
hypre_Vector *x,
double beta,
hypre_Vector *y )
{
double *A_data = hypre_CSRMatrixData(A);
int *A_i = hypre_CSRMatrixI(A);
int *A_j = hypre_CSRMatrixJ(A);
int num_rows = hypre_CSRMatrixNumRows(A);
int num_cols = hypre_CSRMatrixNumCols(A);
int *A_rownnz = hypre_CSRMatrixRownnz(A);
int num_rownnz = hypre_CSRMatrixNumRownnz(A);
double *x_data = hypre_VectorData(x);
double *y_data = hypre_VectorData(y);
int x_size = hypre_VectorSize(x);
int y_size = hypre_VectorSize(y);
int num_vectors = hypre_VectorNumVectors(x);
int idxstride_y = hypre_VectorIndexStride(y);
int vecstride_y = hypre_VectorVectorStride(y);
int idxstride_x = hypre_VectorIndexStride(x);
int vecstride_x = hypre_VectorVectorStride(x);
double temp, tempx;
int i, j, jj;
int m;
double xpar=0.7;
int ierr = 0;
/*---------------------------------------------------------------------
* Check for size compatibility. Matvec returns ierr = 1 if
* length of X doesn't equal the number of columns of A,
* ierr = 2 if the length of Y doesn't equal the number of rows
* of A, and ierr = 3 if both are true.
*
* Because temporary vectors are often used in Matvec, none of
* these conditions terminates processing, and the ierr flag
* is informational only.
*--------------------------------------------------------------------*/
hypre_assert( num_vectors == hypre_VectorNumVectors(y) );
if (num_cols != x_size)
ierr = 1;
if (num_rows != y_size)
ierr = 2;
if (num_cols != x_size && num_rows != y_size)
ierr = 3;
/*-----------------------------------------------------------------------
* Do (alpha == 0.0) computation - RDF: USE MACHINE EPS
*-----------------------------------------------------------------------*/
if (alpha == 0.0)
{
for (i = 0; i < num_rows*num_vectors; i++)
y_data[i] *= beta;
return ierr;
}
/*-----------------------------------------------------------------------
* y = (beta/alpha)*y
*-----------------------------------------------------------------------*/
temp = beta / alpha;
if (temp != 1.0)
{
if (temp == 0.0)
{
for (i = 0; i < num_rows*num_vectors; i++)
y_data[i] = 0.0;
}
else
{
for (i = 0; i < num_rows*num_vectors; i++)
y_data[i] *= temp;
}
}
/*-----------------------------------------------------------------
* y += A*x
*-----------------------------------------------------------------*/
/* use rownnz pointer to do the A*x multiplication when num_rownnz is smaller than num_rows */
if (num_rownnz < xpar*(num_rows))
{
for (i = 0; i < num_rownnz; i++)
{
m = A_rownnz[i];
/*
* for (jj = A_i[m]; jj < A_i[m+1]; jj++)
* {
* j = A_j[jj];
* y_data[m] += A_data[jj] * x_data[j];
* } */
if ( num_vectors==1 )
{
tempx = y_data[m];
for (jj = A_i[m]; jj < A_i[m+1]; jj++)
tempx += A_data[jj] * x_data[A_j[jj]];
y_data[m] = tempx;
}
else
for ( j=0; j<num_vectors; ++j )
{
tempx = y_data[ j*vecstride_y + m*idxstride_y ];
for (jj = A_i[m]; jj < A_i[m+1]; jj++)
tempx += A_data[jj] * x_data[ j*vecstride_x + A_j[jj]*idxstride_x ];
y_data[ j*vecstride_y + m*idxstride_y] = tempx;
}
}
}
else
{
#pragma omp parallel for private(i,jj,temp) schedule(static)
for (i = 0; i < num_rows; i++)
{
if ( num_vectors==1 )
{
temp = y_data[i];
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
temp += A_data[jj] * x_data[A_j[jj]];
y_data[i] = temp;
}
else
for ( j=0; j<num_vectors; ++j )
{
temp = y_data[ j*vecstride_y + i*idxstride_y ];
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
temp += A_data[jj] * x_data[ j*vecstride_x + A_j[jj]*idxstride_x ];
}
y_data[ j*vecstride_y + i*idxstride_y ] = temp;
}
}
}
/*-----------------------------------------------------------------
* y = alpha*y
*-----------------------------------------------------------------*/
if (alpha != 1.0)
{
for (i = 0; i < num_rows*num_vectors; i++)
y_data[i] *= alpha;
}
return ierr;
}
void hypre_BoomerAMGJacobiInterp_1( hypre_ParCSRMatrix * A,
hypre_ParCSRMatrix ** P,
hypre_ParCSRMatrix * S,
HYPRE_Int * CF_marker, HYPRE_Int level,
HYPRE_Real truncation_threshold,
HYPRE_Real truncation_threshold_minus,
HYPRE_Int * dof_func, HYPRE_Int * dof_func_offd,
HYPRE_Real weight_AF)
/* One step of Jacobi interpolation:
A is the linear system.
P is an interpolation matrix, input and output
CF_marker identifies coarse and fine points
If we imagine P and A as split into coarse and fine submatrices,
[ AFF AFC ] [ AF ] [ IFC ]
A = [ ] = [ ] , P = [ ]
[ ACF ACC ] [ AC ] [ ICC ]
(note that ICC is an identity matrix, applied to coarse points only)
then this function computes
IFCnew = IFCold - DFF(-1) * ( AFF*IFCold + AFC )
= IFCold - DFF(-1) * AF * Pold)
where DFF is the diagonal of AFF, (-1) represents the inverse, and
where "old" denotes a value on entry to this function, "new" a returned value.
*/
{
hypre_ParCSRMatrix * Pnew;
hypre_ParCSRMatrix * C;
hypre_CSRMatrix *P_diag = hypre_ParCSRMatrixDiag(*P);
hypre_CSRMatrix *P_offd = hypre_ParCSRMatrixOffd(*P);
HYPRE_Real *P_diag_data = hypre_CSRMatrixData(P_diag);
HYPRE_Int *P_diag_i = hypre_CSRMatrixI(P_diag);
HYPRE_Int *P_diag_j = hypre_CSRMatrixJ(P_diag);
HYPRE_Real *P_offd_data = hypre_CSRMatrixData(P_offd);
HYPRE_Int *P_offd_i = hypre_CSRMatrixI(P_offd);
hypre_CSRMatrix *C_diag;
hypre_CSRMatrix *C_offd;
hypre_CSRMatrix *Pnew_diag;
hypre_CSRMatrix *Pnew_offd;
HYPRE_Int num_rows_diag_P = hypre_CSRMatrixNumRows(P_diag);
HYPRE_Int i;
HYPRE_Int Jnochanges=0, Jchanges, Pnew_num_nonzeros;
HYPRE_Int CF_coarse=0;
HYPRE_Int * J_marker = hypre_CTAlloc( HYPRE_Int, num_rows_diag_P );
HYPRE_Int nc, ncmax, ncmin, nc1;
HYPRE_Int num_procs, my_id;
MPI_Comm comm = hypre_ParCSRMatrixComm( A );
#ifdef HYPRE_JACINT_PRINT_ROW_SUMS
HYPRE_Int m, nmav, npav;
HYPRE_Real PIi, PIimax, PIimin, PIimav, PIipav, randthresh;
HYPRE_Real eps = 1.0e-17;
#endif
#ifdef HYPRE_JACINT_PRINT_MATRICES
char filename[80];
HYPRE_Int i_dummy, j_dummy;
HYPRE_Int *base_i_ptr = &i_dummy;
HYPRE_Int *base_j_ptr = &j_dummy;
#endif
#ifdef HYPRE_JACINT_PRINT_SOME_ROWS
HYPRE_Int sample_rows[50], n_sample_rows=0, isamp;
#endif
hypre_MPI_Comm_size(comm, &num_procs);
hypre_MPI_Comm_rank(comm,&my_id);
for ( i=0; i<num_rows_diag_P; ++i )
{
J_marker[i] = CF_marker[i];
if (CF_marker[i]>=0) ++CF_coarse;
}
#ifdef HYPRE_JACINT_PRINT_DIAGNOSTICS
hypre_printf("%i %i Jacobi_Interp_1, P has %i+%i=%i nonzeros, local sum %e\n", my_id, level,
hypre_CSRMatrixNumNonzeros(P_diag), hypre_CSRMatrixNumNonzeros(P_offd),
hypre_CSRMatrixNumNonzeros(P_diag)+hypre_CSRMatrixNumNonzeros(P_offd),
hypre_ParCSRMatrixLocalSumElts(*P) );
#endif
/* row sum computations, for output */
#ifdef HYPRE_JACINT_PRINT_ROW_SUMS
PIimax=-1.0e12, PIimin=1.0e12, PIimav=0, PIipav=0;
nmav=0, npav=0;
for ( i=0; i<num_rows_diag_P; ++i )
{
PIi = 0; /* i-th value of P*1, i.e. sum of row i of P */
for ( m=P_diag_i[i]; m<P_diag_i[i+1]; ++m )
PIi += P_diag_data[m];
for ( m=P_offd_i[i]; m<P_offd_i[i+1]; ++m )
PIi += P_offd_data[m];
if (CF_marker[i]<0)
{
PIimax = hypre_max( PIimax, PIi );
PIimin = hypre_min( PIimin, PIi );
if (PIi<=1-eps) { PIimav+=PIi; ++nmav; };
if (PIi>=1+eps) { PIipav+=PIi; ++npav; };
}
}
if ( nmav>0 ) PIimav = PIimav/nmav;
if ( npav>0 ) PIipav = PIipav/npav;
hypre_printf("%i %i P in max,min row sums %e %e\n", my_id, level, PIimax, PIimin );
#endif
ncmax=0; ncmin=num_rows_diag_P; nc1=0;
for ( i=0; i<num_rows_diag_P; ++i )
if (CF_marker[i]<0)
{
nc = P_diag_i[i+1] - P_diag_i[i];
if (nc<=1)
{
++nc1;
}
ncmax = hypre_max( nc, ncmax );
ncmin = hypre_min( nc, ncmin );
}
#if 0
/* a very agressive reduction in how much the Jacobi step does: */
for ( i=0; i<num_rows_diag_P; ++i )
if (CF_marker[i]<0)
{
nc = P_diag_i[i+1] - P_diag_i[i];
if (nc>ncmin+1)
/*if ( nc > ncmin + 0.5*(ncmax-ncmin) )*/
{
J_marker[i] = 1;
++Jnochanges;
}
}
#endif
Jchanges = num_rows_diag_P - Jnochanges - CF_coarse;
#ifdef HYPRE_JACINT_PRINT_SOME_ROWS
hypre_printf("some rows to be changed: ");
randthresh = 15/(HYPRE_Real)Jchanges;
for ( i=0; i<num_rows_diag_P; ++i )
{
if ( J_marker[i]<0 )
{
if ( ((HYPRE_Real)rand())/RAND_MAX < randthresh )
{
hypre_printf( "%i: ", i );
for ( m=P_diag_i[i]; m<P_diag_i[i+1]; ++m )
hypre_printf( " %i %f, ", P_diag_j[m], P_diag_data[m] );
hypre_printf("; ");
sample_rows[n_sample_rows] = i;
++n_sample_rows;
}
}
}
hypre_printf("\n");
#endif
#ifdef HYPRE_JACINT_PRINT_DIAGNOSTICS
hypre_printf("%i %i P has %i rows, %i changeable, %i don't change-good, %i coarse\n",
my_id, level, num_rows_diag_P, Jchanges, Jnochanges, CF_coarse );
hypre_printf("%i %i min,max diag cols per row: %i, %i; no.rows w.<=1 col: %i\n", my_id, level, ncmin, ncmax, nc1 );
#endif
#ifdef HYPRE_JACINT_PRINT_MATRICES
if ( num_rows_diag_P <= HYPRE_MAX_PRINTABLE_MATRIX )
{
hypre_sprintf( filename, "Ain%i", level );
hypre_ParCSRMatrixPrintIJ( A,0,0,filename);
hypre_sprintf( filename, "Sin%i", level );
hypre_ParCSRMatrixPrintIJ( S,0,0,filename);
hypre_sprintf( filename, "Pin%i", level );
hypre_ParCSRMatrixPrintIJ( *P,0,0,filename);
}
#endif
C = hypre_ParMatmul_FC( A, *P, J_marker, dof_func, dof_func_offd );
/* hypre_parMatmul_FC creates and returns C, a variation of the
matrix product A*P in which only the "Fine"-designated rows have
been computed. (all columns are Coarse because all columns of P
are). "Fine" is defined solely by the marker array, and for
example could be a proper subset of the fine points of a
multigrid hierarchy.
As a matrix, C is the size of A*P. But only the marked rows have
been computed.
*/
#ifdef HYPRE_JACINT_PRINT_MATRICES
hypre_sprintf( filename, "C%i", level );
if ( num_rows_diag_P <= HYPRE_MAX_PRINTABLE_MATRIX ) hypre_ParCSRMatrixPrintIJ( C,0,0,filename);
#endif
C_diag = hypre_ParCSRMatrixDiag(C);
C_offd = hypre_ParCSRMatrixOffd(C);
#ifdef HYPRE_JACINT_PRINT_DIAGNOSTICS
hypre_printf("%i %i Jacobi_Interp_1 after matmul, C has %i+%i=%i nonzeros, local sum %e\n",
my_id, level, hypre_CSRMatrixNumNonzeros(C_diag),
hypre_CSRMatrixNumNonzeros(C_offd),
hypre_CSRMatrixNumNonzeros(C_diag)+hypre_CSRMatrixNumNonzeros(C_offd),
hypre_ParCSRMatrixLocalSumElts(C) );
#endif
hypre_ParMatScaleDiagInv_F( C, A, weight_AF, J_marker );
/* hypre_ParMatScaleDiagInv scales of its first argument by premultiplying with
a submatrix of the inverse of the diagonal of its second argument.
The marker array determines which diagonal elements are used. The marker
array should select exactly the right number of diagonal elements (the number
of rows of AP_FC).
*/
#ifdef HYPRE_JACINT_PRINT_MATRICES
hypre_sprintf( filename, "Cout%i", level );
if ( num_rows_diag_P <= HYPRE_MAX_PRINTABLE_MATRIX ) hypre_ParCSRMatrixPrintIJ( C,0,0,filename);
#endif
Pnew = hypre_ParMatMinus_F( *P, C, J_marker );
/* hypre_ParMatMinus_F subtracts rows of its second argument from selected rows
of its first argument. The marker array determines which rows of the first
argument are affected, and they should exactly correspond to all the rows
of the second argument.
*/
Pnew_diag = hypre_ParCSRMatrixDiag(Pnew);
Pnew_offd = hypre_ParCSRMatrixOffd(Pnew);
Pnew_num_nonzeros = hypre_CSRMatrixNumNonzeros(Pnew_diag)+hypre_CSRMatrixNumNonzeros(Pnew_offd);
#ifdef HYPRE_JACINT_PRINT_DIAGNOSTICS
hypre_printf("%i %i Jacobi_Interp_1 after MatMinus, Pnew has %i+%i=%i nonzeros, local sum %e\n",
my_id, level, hypre_CSRMatrixNumNonzeros(Pnew_diag),
hypre_CSRMatrixNumNonzeros(Pnew_offd), Pnew_num_nonzeros,
hypre_ParCSRMatrixLocalSumElts(Pnew) );
#endif
/* Transfer ownership of col_starts from P to Pnew ... */
if ( hypre_ParCSRMatrixColStarts(*P) &&
hypre_ParCSRMatrixColStarts(*P)==hypre_ParCSRMatrixColStarts(Pnew) )
{
if ( hypre_ParCSRMatrixOwnsColStarts(*P) && !hypre_ParCSRMatrixOwnsColStarts(Pnew) )
{
hypre_ParCSRMatrixSetColStartsOwner(*P,0);
hypre_ParCSRMatrixSetColStartsOwner(Pnew,1);
}
}
hypre_ParCSRMatrixDestroy( C );
hypre_ParCSRMatrixDestroy( *P );
/* Note that I'm truncating all the fine rows, not just the J-marked ones. */
#if 0
if ( Pnew_num_nonzeros < 10000 ) /* a fixed number like this makes it no.procs.-depdendent */
{ /* ad-hoc attempt to reduce zero-matrix problems seen in testing..*/
truncation_threshold = 1.0e-6 * truncation_threshold;
truncation_threshold_minus = 1.0e-6 * truncation_threshold_minus;
}
#endif
hypre_BoomerAMGTruncateInterp( Pnew, truncation_threshold,
truncation_threshold_minus, CF_marker );
hypre_MatvecCommPkgCreate ( Pnew );
*P = Pnew;
P_diag = hypre_ParCSRMatrixDiag(*P);
P_offd = hypre_ParCSRMatrixOffd(*P);
P_diag_data = hypre_CSRMatrixData(P_diag);
P_diag_i = hypre_CSRMatrixI(P_diag);
P_diag_j = hypre_CSRMatrixJ(P_diag);
P_offd_data = hypre_CSRMatrixData(P_offd);
P_offd_i = hypre_CSRMatrixI(P_offd);
/* row sum computations, for output */
#ifdef HYPRE_JACINT_PRINT_ROW_SUMS
PIimax=-1.0e12, PIimin=1.0e12, PIimav=0, PIipav=0;
nmav=0, npav=0;
for ( i=0; i<num_rows_diag_P; ++i )
{
PIi = 0; /* i-th value of P*1, i.e. sum of row i of P */
for ( m=P_diag_i[i]; m<P_diag_i[i+1]; ++m )
PIi += P_diag_data[m];
for ( m=P_offd_i[i]; m<P_offd_i[i+1]; ++m )
PIi += P_offd_data[m];
if (CF_marker[i]<0)
{
PIimax = hypre_max( PIimax, PIi );
PIimin = hypre_min( PIimin, PIi );
if (PIi<=1-eps) { PIimav+=PIi; ++nmav; };
if (PIi>=1+eps) { PIipav+=PIi; ++npav; };
}
}
if ( nmav>0 ) PIimav = PIimav/nmav;
if ( npav>0 ) PIipav = PIipav/npav;
hypre_printf("%i %i P out max,min row sums %e %e\n", my_id, level, PIimax, PIimin );
#endif
#ifdef HYPRE_JACINT_PRINT_SOME_ROWS
hypre_printf("some changed rows: ");
for ( isamp=0; isamp<n_sample_rows; ++isamp )
{
i = sample_rows[isamp];
hypre_printf( "%i: ", i );
for ( m=P_diag_i[i]; m<P_diag_i[i+1]; ++m )
hypre_printf( " %i %f, ", P_diag_j[m], P_diag_data[m] );
hypre_printf("; ");
}
hypre_printf("\n");
#endif
ncmax=0; ncmin=num_rows_diag_P; nc1=0;
for ( i=0; i<num_rows_diag_P; ++i )
if (CF_marker[i]<0)
{
nc = P_diag_i[i+1] - P_diag_i[i];
if (nc<=1) ++nc1;
ncmax = hypre_max( nc, ncmax );
ncmin = hypre_min( nc, ncmin );
}
#ifdef HYPRE_JACINT_PRINT_DIAGNOSTICS
hypre_printf("%i %i P has %i rows, %i changeable, %i too good, %i coarse\n",
my_id, level, num_rows_diag_P, num_rows_diag_P-Jnochanges-CF_coarse, Jnochanges, CF_coarse );
hypre_printf("%i %i min,max diag cols per row: %i, %i; no.rows w.<=1 col: %i\n", my_id, level, ncmin, ncmax, nc1 );
hypre_printf("%i %i Jacobi_Interp_1 after truncation (%e), Pnew has %i+%i=%i nonzeros, local sum %e\n",
my_id, level, truncation_threshold,
hypre_CSRMatrixNumNonzeros(Pnew_diag), hypre_CSRMatrixNumNonzeros(Pnew_offd),
hypre_CSRMatrixNumNonzeros(Pnew_diag)+hypre_CSRMatrixNumNonzeros(Pnew_offd),
hypre_ParCSRMatrixLocalSumElts(Pnew) );
#endif
/* Programming Notes:
1. Judging by around line 299 of par_interp.c, they typical use of CF_marker
is that CF_marker>=0 means Coarse, CF_marker<0 means Fine.
*/
#ifdef HYPRE_JACINT_PRINT_MATRICES
hypre_sprintf( filename, "Pout%i", level );
if ( num_rows_diag_P <= HYPRE_MAX_PRINTABLE_MATRIX ) hypre_ParCSRMatrixPrintIJ( *P,0,0,filename);
#endif
hypre_TFree( J_marker );
}
HYPRE_Int hypre_BoomerAMGRelaxT( hypre_ParCSRMatrix *A,
hypre_ParVector *f,
HYPRE_Int *cf_marker,
HYPRE_Int relax_type,
HYPRE_Int relax_points,
HYPRE_Real relax_weight,
hypre_ParVector *u,
hypre_ParVector *Vtemp )
{
hypre_CSRMatrix *A_diag = hypre_ParCSRMatrixDiag(A);
HYPRE_Real *A_diag_data = hypre_CSRMatrixData(A_diag);
HYPRE_Int *A_diag_i = hypre_CSRMatrixI(A_diag);
HYPRE_Int n_global= hypre_ParCSRMatrixGlobalNumRows(A);
HYPRE_Int n = hypre_CSRMatrixNumRows(A_diag);
HYPRE_Int first_index = hypre_ParVectorFirstIndex(u);
hypre_Vector *u_local = hypre_ParVectorLocalVector(u);
HYPRE_Real *u_data = hypre_VectorData(u_local);
hypre_Vector *Vtemp_local = hypre_ParVectorLocalVector(Vtemp);
HYPRE_Real *Vtemp_data = hypre_VectorData(Vtemp_local);
hypre_CSRMatrix *A_CSR;
HYPRE_Int *A_CSR_i;
HYPRE_Int *A_CSR_j;
HYPRE_Real *A_CSR_data;
hypre_Vector *f_vector;
HYPRE_Real *f_vector_data;
HYPRE_Int i;
HYPRE_Int jj;
HYPRE_Int column;
HYPRE_Int relax_error = 0;
HYPRE_Real *A_mat;
HYPRE_Real *b_vec;
HYPRE_Real zero = 0.0;
/*-----------------------------------------------------------------------
* Switch statement to direct control based on relax_type:
* relax_type = 7 -> Jacobi (uses ParMatvec)
* relax_type = 9 -> Direct Solve
*-----------------------------------------------------------------------*/
switch (relax_type)
{
case 7: /* Jacobi (uses ParMatvec) */
{
/*-----------------------------------------------------------------
* Copy f into temporary vector.
*-----------------------------------------------------------------*/
hypre_ParVectorCopy(f,Vtemp);
/*-----------------------------------------------------------------
* Perform MatvecT Vtemp=f-A^Tu
*-----------------------------------------------------------------*/
hypre_ParCSRMatrixMatvecT(-1.0,A, u, 1.0, Vtemp);
for (i = 0; i < n; i++)
{
/*-----------------------------------------------------------
* If diagonal is nonzero, relax point i; otherwise, skip it.
*-----------------------------------------------------------*/
if (A_diag_data[A_diag_i[i]] != zero)
{
u_data[i] += relax_weight * Vtemp_data[i]
/ A_diag_data[A_diag_i[i]];
}
}
}
break;
case 9: /* Direct solve: use gaussian elimination */
{
/*-----------------------------------------------------------------
* Generate CSR matrix from ParCSRMatrix A
*-----------------------------------------------------------------*/
if (n)
{
A_CSR = hypre_ParCSRMatrixToCSRMatrixAll(A);
f_vector = hypre_ParVectorToVectorAll(f);
A_CSR_i = hypre_CSRMatrixI(A_CSR);
A_CSR_j = hypre_CSRMatrixJ(A_CSR);
A_CSR_data = hypre_CSRMatrixData(A_CSR);
f_vector_data = hypre_VectorData(f_vector);
A_mat = hypre_CTAlloc(HYPRE_Real, n_global*n_global);
b_vec = hypre_CTAlloc(HYPRE_Real, n_global);
/*---------------------------------------------------------------
* Load transpose of CSR matrix into A_mat.
*---------------------------------------------------------------*/
for (i = 0; i < n_global; i++)
{
for (jj = A_CSR_i[i]; jj < A_CSR_i[i+1]; jj++)
{
column = A_CSR_j[jj];
A_mat[column*n_global+i] = A_CSR_data[jj];
}
b_vec[i] = f_vector_data[i];
}
relax_error = gselim(A_mat,b_vec,n_global);
for (i = 0; i < n; i++)
{
u_data[i] = b_vec[first_index+i];
}
hypre_TFree(A_mat);
hypre_TFree(b_vec);
hypre_CSRMatrixDestroy(A_CSR);
A_CSR = NULL;
hypre_SeqVectorDestroy(f_vector);
f_vector = NULL;
}
}
break;
}
return(relax_error);
}
/*
Assume that we are given a fine and coarse topology and the
coarse degrees of freedom (DOFs) have been chosen. Assume also,
that the global interpolation matrix dof_DOF has a prescribed
nonzero pattern. Then, the fine degrees of freedom can be split
into 4 groups (here "i" stands for "interior"):
NODEidof - dofs which are interpolated only from the DOF
in one coarse vertex
EDGEidof - dofs which are interpolated only from the DOFs
in one coarse edge
FACEidof - dofs which are interpolated only from the DOFs
in one coarse face
ELEMidof - dofs which are interpolated only from the DOFs
in one coarse element
The interpolation operator dof_DOF can be build in 4 steps, by
consequently filling-in the rows corresponding to the above groups.
The code below uses harmonic extension to extend the interpolation
from one group to the next.
*/
HYPRE_Int hypre_ND1AMGeInterpolation (hypre_ParCSRMatrix * Aee,
hypre_ParCSRMatrix * ELEM_idof,
hypre_ParCSRMatrix * FACE_idof,
hypre_ParCSRMatrix * EDGE_idof,
hypre_ParCSRMatrix * ELEM_FACE,
hypre_ParCSRMatrix * ELEM_EDGE,
HYPRE_Int num_OffProcRows,
hypre_MaxwellOffProcRow ** OffProcRows,
hypre_IJMatrix * IJ_dof_DOF)
{
HYPRE_Int ierr = 0;
HYPRE_Int i, j, k;
HYPRE_Int *offproc_rnums, *swap;
hypre_ParCSRMatrix * dof_DOF = hypre_IJMatrixObject(IJ_dof_DOF);
hypre_ParCSRMatrix * ELEM_DOF = ELEM_EDGE;
hypre_ParCSRMatrix * ELEM_FACEidof;
hypre_ParCSRMatrix * ELEM_EDGEidof;
hypre_CSRMatrix *A, *P;
HYPRE_Int numELEM = hypre_CSRMatrixNumRows(hypre_ParCSRMatrixDiag(ELEM_EDGE));
HYPRE_Int getrow_ierr;
HYPRE_Int three_dimensional_problem;
MPI_Comm comm= hypre_ParCSRMatrixComm(Aee);
HYPRE_Int myproc;
hypre_MPI_Comm_rank(comm, &myproc);
#if 0
hypre_IJMatrix * ij_dof_DOF = hypre_CTAlloc(hypre_IJMatrix, 1);
/* Convert dof_DOF to IJ matrix, so we can use AddToValues */
hypre_IJMatrixComm(ij_dof_DOF) = hypre_ParCSRMatrixComm(dof_DOF);
hypre_IJMatrixRowPartitioning(ij_dof_DOF) =
hypre_ParCSRMatrixRowStarts(dof_DOF);
hypre_IJMatrixColPartitioning(ij_dof_DOF) =
hypre_ParCSRMatrixColStarts(dof_DOF);
hypre_IJMatrixObject(ij_dof_DOF) = dof_DOF;
hypre_IJMatrixAssembleFlag(ij_dof_DOF) = 1;
#endif
/* sort the offproc rows to get quicker comparison for later */
if (num_OffProcRows)
{
offproc_rnums= hypre_TAlloc(HYPRE_Int, num_OffProcRows);
swap = hypre_TAlloc(HYPRE_Int, num_OffProcRows);
for (i= 0; i< num_OffProcRows; i++)
{
offproc_rnums[i]=(OffProcRows[i] -> row);
swap[i] = i;
}
}
if (num_OffProcRows > 1)
{
hypre_qsort2i(offproc_rnums, swap, 0, num_OffProcRows-1);
}
if (FACE_idof == EDGE_idof)
three_dimensional_problem = 0;
else
three_dimensional_problem = 1;
/* ELEM_FACEidof = ELEM_FACE x FACE_idof */
if (three_dimensional_problem)
ELEM_FACEidof = hypre_ParMatmul(ELEM_FACE, FACE_idof);
/* ELEM_EDGEidof = ELEM_EDGE x EDGE_idof */
ELEM_EDGEidof = hypre_ParMatmul(ELEM_EDGE, EDGE_idof);
/* Loop over local coarse elements */
k = hypre_ParCSRMatrixFirstRowIndex(ELEM_EDGE);
for (i = 0; i < numELEM; i++, k++)
{
HYPRE_Int size1, size2;
HYPRE_Int *col_ind0, *col_ind1, *col_ind2;
HYPRE_Int num_DOF, *DOF0, *DOF;
HYPRE_Int num_idof, *idof0, *idof;
HYPRE_Int num_bdof, *bdof;
double *boolean_data;
/* Determine the coarse DOFs */
hypre_ParCSRMatrixGetRow (ELEM_DOF, k, &num_DOF, &DOF0, &boolean_data);
DOF= hypre_TAlloc(HYPRE_Int, num_DOF);
for (j= 0; j< num_DOF; j++)
{
DOF[j]= DOF0[j];
}
hypre_ParCSRMatrixRestoreRow (ELEM_DOF, k, &num_DOF, &DOF0, &boolean_data);
qsort0(DOF,0,num_DOF-1);
/* Find the fine dofs interior for the current coarse element */
hypre_ParCSRMatrixGetRow (ELEM_idof, k, &num_idof, &idof0, &boolean_data);
idof= hypre_TAlloc(HYPRE_Int, num_idof);
for (j= 0; j< num_idof; j++)
{
idof[j]= idof0[j];
}
hypre_ParCSRMatrixRestoreRow (ELEM_idof, k, &num_idof, &idof0, &boolean_data);
/* Sort the interior dofs according to their global number */
qsort0(idof,0,num_idof-1);
/* Find the fine dofs on the boundary of the current coarse element */
if (three_dimensional_problem)
{
hypre_ParCSRMatrixGetRow (ELEM_FACEidof, k, &size1, &col_ind0, &boolean_data);
col_ind1= hypre_TAlloc(HYPRE_Int, size1);
for (j= 0; j< size1; j++)
{
col_ind1[j]= col_ind0[j];
}
hypre_ParCSRMatrixRestoreRow (ELEM_FACEidof, k, &size1, &col_ind0, &boolean_data);
}
else
size1 = 0;
hypre_ParCSRMatrixGetRow (ELEM_EDGEidof, k, &size2, &col_ind0, &boolean_data);
col_ind2= hypre_TAlloc(HYPRE_Int, size2);
for (j= 0; j< size2; j++)
{
col_ind2[j]= col_ind0[j];
}
hypre_ParCSRMatrixRestoreRow (ELEM_EDGEidof, k, &size2, &col_ind0, &boolean_data);
/* Merge and sort the boundary dofs according to their global number */
num_bdof = size1 + size2;
bdof = hypre_CTAlloc(HYPRE_Int, num_bdof);
if (three_dimensional_problem)
memcpy(bdof, col_ind1, size1*sizeof(HYPRE_Int));
memcpy(bdof+size1, col_ind2, size2*sizeof(HYPRE_Int));
qsort0(bdof,0,num_bdof-1);
/* A = extract_rows(Aee, idof) */
A = hypre_CSRMatrixCreate (num_idof, num_idof + num_bdof,
num_idof * (num_idof + num_bdof));
hypre_CSRMatrixInitialize(A);
{
HYPRE_Int *I = hypre_CSRMatrixI(A);
HYPRE_Int *J = hypre_CSRMatrixJ(A);
double *data = hypre_CSRMatrixData(A);
HYPRE_Int *tmp_J;
double *tmp_data;
I[0] = 0;
for (j = 0; j < num_idof; j++)
{
getrow_ierr= hypre_ParCSRMatrixGetRow (Aee, idof[j], &I[j+1], &tmp_J, &tmp_data);
if (getrow_ierr <0)
hypre_printf("getrow Aee off proc[%d] = \n",myproc);
memcpy(J, tmp_J, I[j+1]*sizeof(HYPRE_Int));
memcpy(data, tmp_data, I[j+1]*sizeof(double));
J+= I[j+1];
data+= I[j+1];
hypre_ParCSRMatrixRestoreRow (Aee, idof[j], &I[j+1], &tmp_J, &tmp_data);
I[j+1] += I[j];
}
}
/* P = extract_rows(dof_DOF, idof+bdof) */
P = hypre_CSRMatrixCreate (num_idof + num_bdof, num_DOF,
(num_idof + num_bdof) * num_DOF);
hypre_CSRMatrixInitialize(P);
{
HYPRE_Int *I = hypre_CSRMatrixI(P);
HYPRE_Int *J = hypre_CSRMatrixJ(P);
double *data = hypre_CSRMatrixData(P);
HYPRE_Int m;
HYPRE_Int *tmp_J;
double *tmp_data;
I[0] = 0;
for (j = 0; j < num_idof; j++)
{
getrow_ierr= hypre_ParCSRMatrixGetRow (dof_DOF, idof[j], &I[j+1], &tmp_J, &tmp_data);
if (getrow_ierr >= 0)
{
memcpy(J, tmp_J, I[j+1]*sizeof(HYPRE_Int));
memcpy(data, tmp_data, I[j+1]*sizeof(double));
J+= I[j+1];
data+= I[j+1];
hypre_ParCSRMatrixRestoreRow (dof_DOF, idof[j], &I[j+1], &tmp_J, &tmp_data);
I[j+1] += I[j];
}
else /* row offproc */
{
hypre_ParCSRMatrixRestoreRow (dof_DOF, idof[j], &I[j+1], &tmp_J, &tmp_data);
/* search for OffProcRows */
m= 0;
while (m < num_OffProcRows)
{
if (offproc_rnums[m] == idof[j])
{
break;
}
else
{
m++;
}
}
I[j+1]= (OffProcRows[swap[m]] -> ncols);
tmp_J = (OffProcRows[swap[m]] -> cols);
tmp_data= (OffProcRows[swap[m]] -> data);
memcpy(J, tmp_J, I[j+1]*sizeof(HYPRE_Int));
memcpy(data, tmp_data, I[j+1]*sizeof(double));
J+= I[j+1];
data+= I[j+1];
I[j+1] += I[j];
}
}
for ( ; j < num_idof + num_bdof; j++)
{
getrow_ierr= hypre_ParCSRMatrixGetRow (dof_DOF, bdof[j-num_idof], &I[j+1], &tmp_J, &tmp_data);
if (getrow_ierr >= 0)
{
memcpy(J, tmp_J, I[j+1]*sizeof(HYPRE_Int));
memcpy(data, tmp_data, I[j+1]*sizeof(double));
J+= I[j+1];
data+= I[j+1];
hypre_ParCSRMatrixRestoreRow (dof_DOF, bdof[j-num_idof], &I[j+1], &tmp_J, &tmp_data);
I[j+1] += I[j];
}
else /* row offproc */
{
hypre_ParCSRMatrixRestoreRow (dof_DOF, bdof[j-num_idof], &I[j+1], &tmp_J, &tmp_data);
/* search for OffProcRows */
m= 0;
while (m < num_OffProcRows)
{
if (offproc_rnums[m] == bdof[j-num_idof])
{
break;
}
else
{
m++;
}
}
if (m>= num_OffProcRows)hypre_printf("here the mistake\n");
I[j+1]= (OffProcRows[swap[m]] -> ncols);
tmp_J = (OffProcRows[swap[m]] -> cols);
tmp_data= (OffProcRows[swap[m]] -> data);
memcpy(J, tmp_J, I[j+1]*sizeof(HYPRE_Int));
memcpy(data, tmp_data, I[j+1]*sizeof(double));
J+= I[j+1];
data+= I[j+1];
I[j+1] += I[j];
}
}
}
/* Pi = Aii^{-1} Aib Pb */
hypre_HarmonicExtension (A, P, num_DOF, DOF,
num_idof, idof, num_bdof, bdof);
/* Insert Pi in dof_DOF */
{
HYPRE_Int * ncols = hypre_CTAlloc(HYPRE_Int, num_idof);
for (j = 0; j < num_idof; j++)
ncols[j] = num_DOF;
hypre_IJMatrixAddToValuesParCSR (IJ_dof_DOF,
num_idof, ncols, idof,
hypre_CSRMatrixJ(P),
hypre_CSRMatrixData(P));
hypre_TFree(ncols);
}
hypre_TFree(DOF);
hypre_TFree(idof);
if (three_dimensional_problem)
{
hypre_TFree(col_ind1);
}
hypre_TFree(col_ind2);
hypre_TFree(bdof);
hypre_CSRMatrixDestroy(A);
hypre_CSRMatrixDestroy(P);
}
#if 0
hypre_TFree(ij_dof_DOF);
#endif
if (three_dimensional_problem)
hypre_ParCSRMatrixDestroy(ELEM_FACEidof);
hypre_ParCSRMatrixDestroy(ELEM_EDGEidof);
if (num_OffProcRows)
{
hypre_TFree(offproc_rnums);
hypre_TFree(swap);
}
return ierr;
}
HYPRE_Int AmgCGCChoose (hypre_CSRMatrix *G,HYPRE_Int *vertexrange,HYPRE_Int mpisize,HYPRE_Int **coarse)
/* chooses one grid for every processor
* ============================================================
* G : the connectivity graph
* map : the parallel layout
* mpisize : number of procs
* coarse : the chosen coarse grids
* ===========================================================*/
{
HYPRE_Int i,j,jj,p,choice,*processor,ierr=0;
HYPRE_Int measure,new_measure;
/* MPI_Comm comm = hypre_ParCSRMatrixComm(G); */
/* hypre_ParCSRCommPkg *comm_pkg = hypre_ParCSRMatrixCommPkg (G); */
/* hypre_ParCSRCommHandle *comm_handle; */
HYPRE_Real *G_data = hypre_CSRMatrixData (G);
HYPRE_Real max;
HYPRE_Int *G_i = hypre_CSRMatrixI(G);
HYPRE_Int *G_j = hypre_CSRMatrixJ(G);
hypre_CSRMatrix *H,*HT;
HYPRE_Int *H_i,*H_j,*HT_i,*HT_j;
HYPRE_Int jG,jH;
HYPRE_Int num_vertices = hypre_CSRMatrixNumRows (G);
HYPRE_Int *measure_array;
HYPRE_Int *lists,*where;
hypre_LinkList LoL_head = NULL;
hypre_LinkList LoL_tail = NULL;
processor = hypre_CTAlloc (HYPRE_Int,num_vertices);
*coarse = hypre_CTAlloc (HYPRE_Int,mpisize);
memset (*coarse,0,sizeof(HYPRE_Int)*mpisize);
measure_array = hypre_CTAlloc (HYPRE_Int,num_vertices);
lists = hypre_CTAlloc (HYPRE_Int,num_vertices);
where = hypre_CTAlloc (HYPRE_Int,num_vertices);
/* for (p=0;p<mpisize;p++) hypre_printf ("%d: %d-%d\n",p,range[p]+1,range[p+1]); */
/******************************************************************
* determine heavy edges
******************************************************************/
jG = G_i[num_vertices];
H = hypre_CSRMatrixCreate (num_vertices,num_vertices,jG);
H_i = hypre_CTAlloc (HYPRE_Int,num_vertices+1);
H_j = hypre_CTAlloc (HYPRE_Int,jG);
hypre_CSRMatrixI(H) = H_i;
hypre_CSRMatrixJ(H) = H_j;
for (i=0,p=0;i<num_vertices;i++) {
while (vertexrange[p+1]<=i) p++;
processor[i]=p;
}
H_i[0]=0;
for (i=0,jj=0;i<num_vertices;i++) {
#if 0
hypre_printf ("neighbors of grid %d:",i);
#endif
H_i[i+1]=H_i[i];
for (j=G_i[i],choice=-1,max=0;j<G_i[i+1];j++) {
#if 0
if (G_data[j]>=0.0)
hypre_printf ("G[%d,%d]=0. G_j(j)=%d, G_data(j)=%f.\n",i,G_j[j],j,G_data[j]);
#endif
/* G_data is always negative, so this test is sufficient */
if (choice==-1 || G_data[j]>max) {
choice = G_j[j];
max = G_data[j];
}
if (j==G_i[i+1]-1 || processor[G_j[j+1]] > processor[choice]) {
/* we are done for this processor boundary */
H_j[jj++]=choice;
H_i[i+1]++;
#if 0
hypre_printf (" %d",choice);
#endif
choice = -1; max=0;
}
}
#if 0
hypre_printf("\n");
#endif
}
/******************************************************************
* compute H^T, the transpose of H
******************************************************************/
jH = H_i[num_vertices];
HT = hypre_CSRMatrixCreate (num_vertices,num_vertices,jH);
HT_i = hypre_CTAlloc (HYPRE_Int,num_vertices+1);
HT_j = hypre_CTAlloc (HYPRE_Int,jH);
hypre_CSRMatrixI(HT) = HT_i;
hypre_CSRMatrixJ(HT) = HT_j;
for (i=0; i <= num_vertices; i++)
HT_i[i] = 0;
for (i=0; i < jH; i++) {
HT_i[H_j[i]+1]++;
}
for (i=0; i < num_vertices; i++) {
HT_i[i+1] += HT_i[i];
}
for (i=0; i < num_vertices; i++) {
for (j=H_i[i]; j < H_i[i+1]; j++) {
HYPRE_Int myindex = H_j[j];
HT_j[HT_i[myindex]] = i;
HT_i[myindex]++;
}
}
for (i = num_vertices; i > 0; i--) {
HT_i[i] = HT_i[i-1];
}
HT_i[0] = 0;
/*****************************************************************
* set initial vertex weights
*****************************************************************/
for (i=0;i<num_vertices;i++) {
measure_array[i] = H_i[i+1] - H_i[i] + HT_i[i+1] - HT_i[i];
enter_on_lists (&LoL_head,&LoL_tail,measure_array[i],i,lists,where);
}
/******************************************************************
* apply CGC iteration
******************************************************************/
while (LoL_head && measure_array[LoL_head->head]) {
choice = LoL_head->head;
measure = measure_array[choice];
#if 0
hypre_printf ("Choice: %d, measure %d, processor %d\n",choice, measure,processor[choice]);
fflush(stdout);
#endif
(*coarse)[processor[choice]] = choice+1; /* add one because coarsegrid indexing starts with 1, not 0 */
/* new maximal weight */
new_measure = measure+1;
for (i=vertexrange[processor[choice]];i<vertexrange[processor[choice]+1];i++) {
/* set weights for all remaining vertices on this processor to zero */
measure = measure_array[i];
remove_point (&LoL_head,&LoL_tail,measure,i,lists,where);
measure_array[i]=0;
}
for (j=H_i[choice];j<H_i[choice+1];j++){
jj = H_j[j];
/* if no vertex is chosen on this proc, set weights of all heavily coupled vertices to max1 */
if (!(*coarse)[processor[jj]]) {
measure = measure_array[jj];
remove_point (&LoL_head,&LoL_tail,measure,jj,lists,where);
enter_on_lists (&LoL_head,&LoL_tail,new_measure,jj,lists,where);
measure_array[jj]=new_measure;
}
}
for (j=HT_i[choice];j<HT_i[choice+1];j++) {
jj = HT_j[j];
/* if no vertex is chosen on this proc, set weights of all heavily coupled vertices to max1 */
if (!(*coarse)[processor[jj]]) {
measure = measure_array[jj];
remove_point (&LoL_head,&LoL_tail,measure,jj,lists,where);
enter_on_lists (&LoL_head,&LoL_tail,new_measure,jj,lists,where);
measure_array[jj]=new_measure;
}
}
}
/* remove remaining list elements, if they exist. They all should have measure 0 */
while (LoL_head) {
i = LoL_head->head;
measure = measure_array[i];
#if 0
hypre_assert (measure==0);
#endif
remove_point (&LoL_head,&LoL_tail,measure,i,lists,where);
}
for (p=0;p<mpisize;p++)
/* if the algorithm has not determined a coarse vertex for this proc, simply take the last one
Do not take the first one, it might by empty! */
if (!(*coarse)[p]) {
(*coarse)[p] = vertexrange[p+1];
/* hypre_printf ("choice for processor %d: %d\n",p,range[p]+1); */
}
/********************************************
* clean up
********************************************/
hypre_CSRMatrixDestroy (H);
hypre_CSRMatrixDestroy (HT);
hypre_TFree (processor);
hypre_TFree (measure_array);
hypre_TFree (lists);
hypre_TFree (where);
return(ierr);
}
HYPRE_Int
hypre_BoomerAMGCreateScalarCFS(hypre_ParCSRMatrix *SN,
HYPRE_Int *CFN_marker,
HYPRE_Int *col_offd_SN_to_AN,
HYPRE_Int num_functions,
HYPRE_Int nodal,
HYPRE_Int data,
HYPRE_Int **dof_func_ptr,
HYPRE_Int **CF_marker_ptr,
HYPRE_Int **col_offd_S_to_A_ptr,
hypre_ParCSRMatrix **S_ptr)
{
MPI_Comm comm = hypre_ParCSRMatrixComm(SN);
hypre_ParCSRMatrix *S;
hypre_CSRMatrix *S_diag;
HYPRE_Int *S_diag_i;
HYPRE_Int *S_diag_j;
double *S_diag_data;
hypre_CSRMatrix *S_offd;
HYPRE_Int *S_offd_i;
HYPRE_Int *S_offd_j;
double *S_offd_data;
HYPRE_Int *row_starts_S;
HYPRE_Int *col_starts_S;
HYPRE_Int *row_starts_SN = hypre_ParCSRMatrixRowStarts(SN);
HYPRE_Int *col_starts_SN = hypre_ParCSRMatrixColStarts(SN);
hypre_CSRMatrix *SN_diag = hypre_ParCSRMatrixDiag(SN);
HYPRE_Int *SN_diag_i = hypre_CSRMatrixI(SN_diag);
HYPRE_Int *SN_diag_j = hypre_CSRMatrixJ(SN_diag);
double *SN_diag_data;
hypre_CSRMatrix *SN_offd = hypre_ParCSRMatrixOffd(SN);
HYPRE_Int *SN_offd_i = hypre_CSRMatrixI(SN_offd);
HYPRE_Int *SN_offd_j = hypre_CSRMatrixJ(SN_offd);
double *SN_offd_data;
HYPRE_Int *CF_marker;
HYPRE_Int *col_map_offd_SN = hypre_ParCSRMatrixColMapOffd(SN);
HYPRE_Int *col_map_offd_S;
HYPRE_Int *dof_func;
HYPRE_Int num_nodes = hypre_CSRMatrixNumRows(SN_diag);
HYPRE_Int num_variables;
hypre_ParCSRCommPkg *comm_pkg = hypre_ParCSRMatrixCommPkg(SN);
HYPRE_Int num_sends;
HYPRE_Int num_recvs;
HYPRE_Int *send_procs;
HYPRE_Int *send_map_starts;
HYPRE_Int *send_map_elmts;
HYPRE_Int *recv_procs;
HYPRE_Int *recv_vec_starts;
hypre_ParCSRCommPkg *comm_pkg_S;
HYPRE_Int *send_procs_S;
HYPRE_Int *send_map_starts_S;
HYPRE_Int *send_map_elmts_S;
HYPRE_Int *recv_procs_S;
HYPRE_Int *recv_vec_starts_S;
HYPRE_Int *col_offd_S_to_A = NULL;
HYPRE_Int num_coarse_nodes;
HYPRE_Int i,j,k,k1,jj,cnt;
HYPRE_Int row, start, end;
HYPRE_Int num_procs;
HYPRE_Int num_cols_offd_SN = hypre_CSRMatrixNumCols(SN_offd);
HYPRE_Int num_cols_offd_S;
HYPRE_Int SN_num_nonzeros_diag;
HYPRE_Int SN_num_nonzeros_offd;
HYPRE_Int S_num_nonzeros_diag;
HYPRE_Int S_num_nonzeros_offd;
HYPRE_Int global_num_vars;
HYPRE_Int global_num_cols;
HYPRE_Int global_num_nodes;
HYPRE_Int ierr = 0;
hypre_MPI_Comm_size(comm, &num_procs);
num_variables = num_functions*num_nodes;
CF_marker = hypre_CTAlloc(HYPRE_Int, num_variables);
if (nodal < 0)
{
cnt = 0;
num_coarse_nodes = 0;
for (i=0; i < num_nodes; i++)
{
if (CFN_marker[i] == 1) num_coarse_nodes++;
for (j=0; j < num_functions; j++)
CF_marker[cnt++] = CFN_marker[i];
}
dof_func = hypre_CTAlloc(HYPRE_Int,num_coarse_nodes*num_functions);
cnt = 0;
for (i=0; i < num_nodes; i++)
{
if (CFN_marker[i] == 1)
{
for (k=0; k < num_functions; k++)
dof_func[cnt++] = k;
}
}
*dof_func_ptr = dof_func;
}
else
{
cnt = 0;
for (i=0; i < num_nodes; i++)
for (j=0; j < num_functions; j++)
CF_marker[cnt++] = CFN_marker[i];
}
*CF_marker_ptr = CF_marker;
#ifdef HYPRE_NO_GLOBAL_PARTITION
row_starts_S = hypre_CTAlloc(HYPRE_Int,2);
for (i=0; i < 2; i++)
row_starts_S[i] = num_functions*row_starts_SN[i];
if (row_starts_SN != col_starts_SN)
{
col_starts_S = hypre_CTAlloc(HYPRE_Int,2);
for (i=0; i < 2; i++)
col_starts_S[i] = num_functions*col_starts_SN[i];
}
else
{
col_starts_S = row_starts_S;
}
#else
row_starts_S = hypre_CTAlloc(HYPRE_Int,num_procs+1);
for (i=0; i < num_procs+1; i++)
row_starts_S[i] = num_functions*row_starts_SN[i];
if (row_starts_SN != col_starts_SN)
{
col_starts_S = hypre_CTAlloc(HYPRE_Int,num_procs+1);
for (i=0; i < num_procs+1; i++)
col_starts_S[i] = num_functions*col_starts_SN[i];
}
else
{
col_starts_S = row_starts_S;
}
#endif
SN_num_nonzeros_diag = SN_diag_i[num_nodes];
SN_num_nonzeros_offd = SN_offd_i[num_nodes];
global_num_nodes = hypre_ParCSRMatrixGlobalNumRows(SN);
global_num_cols = hypre_ParCSRMatrixGlobalNumCols(SN)*num_functions;
global_num_vars = global_num_nodes*num_functions;
S_num_nonzeros_diag = num_functions*SN_num_nonzeros_diag;
S_num_nonzeros_offd = num_functions*SN_num_nonzeros_offd;
num_cols_offd_S = num_functions*num_cols_offd_SN;
S = hypre_ParCSRMatrixCreate(comm, global_num_vars, global_num_cols,
row_starts_S, col_starts_S, num_cols_offd_S,
S_num_nonzeros_diag, S_num_nonzeros_offd);
S_diag = hypre_ParCSRMatrixDiag(S);
S_offd = hypre_ParCSRMatrixOffd(S);
S_diag_i = hypre_CTAlloc(HYPRE_Int, num_variables+1);
S_offd_i = hypre_CTAlloc(HYPRE_Int, num_variables+1);
S_diag_j = hypre_CTAlloc(HYPRE_Int, S_num_nonzeros_diag);
hypre_CSRMatrixI(S_diag) = S_diag_i;
hypre_CSRMatrixJ(S_diag) = S_diag_j;
if (data)
{
SN_diag_data = hypre_CSRMatrixData(SN_diag);
S_diag_data = hypre_CTAlloc(double, S_num_nonzeros_diag);
hypre_CSRMatrixData(S_diag) = S_diag_data;
if (num_cols_offd_S)
{
SN_offd_data = hypre_CSRMatrixData(SN_offd);
S_offd_data = hypre_CTAlloc(double, S_num_nonzeros_offd);
hypre_CSRMatrixData(S_offd) = S_offd_data;
}
}
hypre_CSRMatrixI(S_offd) = S_offd_i;
if (comm_pkg)
{
comm_pkg_S = hypre_CTAlloc(hypre_ParCSRCommPkg,1);
hypre_ParCSRCommPkgComm(comm_pkg_S) = comm;
num_sends = hypre_ParCSRCommPkgNumSends(comm_pkg);
hypre_ParCSRCommPkgNumSends(comm_pkg_S) = num_sends;
num_recvs = hypre_ParCSRCommPkgNumRecvs(comm_pkg);
hypre_ParCSRCommPkgNumRecvs(comm_pkg_S) = num_recvs;
send_procs = hypre_ParCSRCommPkgSendProcs(comm_pkg);
send_map_starts = hypre_ParCSRCommPkgSendMapStarts(comm_pkg);
send_map_elmts = hypre_ParCSRCommPkgSendMapElmts(comm_pkg);
recv_procs = hypre_ParCSRCommPkgRecvProcs(comm_pkg);
recv_vec_starts = hypre_ParCSRCommPkgRecvVecStarts(comm_pkg);
send_procs_S = NULL;
send_map_elmts_S = NULL;
if (num_sends)
{
send_procs_S = hypre_CTAlloc(HYPRE_Int,num_sends);
send_map_elmts_S = hypre_CTAlloc(HYPRE_Int,
num_functions*send_map_starts[num_sends]);
}
send_map_starts_S = hypre_CTAlloc(HYPRE_Int,num_sends+1);
recv_vec_starts_S = hypre_CTAlloc(HYPRE_Int,num_recvs+1);
recv_procs_S = NULL;
if (num_recvs) recv_procs_S = hypre_CTAlloc(HYPRE_Int,num_recvs);
send_map_starts_S[0] = 0;
for (i=0; i < num_sends; i++)
{
send_procs_S[i] = send_procs[i];
send_map_starts_S[i+1] = num_functions*send_map_starts[i+1];
}
recv_vec_starts_S[0] = 0;
for (i=0; i < num_recvs; i++)
{
recv_procs_S[i] = recv_procs[i];
recv_vec_starts_S[i+1] = num_functions*recv_vec_starts[i+1];
}
cnt = 0;
for (i=0; i < send_map_starts[num_sends]; i++)
{
k1 = num_functions*send_map_elmts[i];
for (j=0; j < num_functions; j++)
{
send_map_elmts_S[cnt++] = k1+j;
}
}
hypre_ParCSRCommPkgSendProcs(comm_pkg_S) = send_procs_S;
hypre_ParCSRCommPkgSendMapStarts(comm_pkg_S) = send_map_starts_S;
hypre_ParCSRCommPkgSendMapElmts(comm_pkg_S) = send_map_elmts_S;
hypre_ParCSRCommPkgRecvProcs(comm_pkg_S) = recv_procs_S;
hypre_ParCSRCommPkgRecvVecStarts(comm_pkg_S) = recv_vec_starts_S;
hypre_ParCSRMatrixCommPkg(S) = comm_pkg_S;
}
if (num_cols_offd_S)
{
S_offd_j = hypre_CTAlloc(HYPRE_Int, S_num_nonzeros_offd);
hypre_CSRMatrixJ(S_offd) = S_offd_j;
col_map_offd_S = hypre_CTAlloc(HYPRE_Int, num_cols_offd_S);
cnt = 0;
for (i=0; i < num_cols_offd_SN; i++)
{
k1 = col_map_offd_SN[i]*num_functions;
for (j=0; j < num_functions; j++)
col_map_offd_S[cnt++] = k1+j;
}
hypre_ParCSRMatrixColMapOffd(S) = col_map_offd_S;
}
if (col_offd_SN_to_AN)
{
col_offd_S_to_A = hypre_CTAlloc(HYPRE_Int, num_cols_offd_S);
cnt = 0;
for (i=0; i < num_cols_offd_SN; i++)
{
k1 = col_offd_SN_to_AN[i]*num_functions;
for (j=0; j < num_functions; j++)
col_offd_S_to_A[cnt++] = k1+j;
}
*col_offd_S_to_A_ptr = col_offd_S_to_A;
}
cnt = 0;
row = 0;
for (i=0; i < num_nodes; i++)
{
row++;
start = cnt;
for (j=SN_diag_i[i]; j < SN_diag_i[i+1]; j++)
{
jj = SN_diag_j[j];
if (data) S_diag_data[cnt] = SN_diag_data[j];
S_diag_j[cnt++] = jj*num_functions;
}
end = cnt;
S_diag_i[row] = cnt;
for (k1=1; k1 < num_functions; k1++)
{
row++;
for (k=start; k < end; k++)
{
if (data) S_diag_data[cnt] = S_diag_data[k];
S_diag_j[cnt++] = S_diag_j[k]+k1;
}
S_diag_i[row] = cnt;
}
}
cnt = 0;
row = 0;
for (i=0; i < num_nodes; i++)
{
row++;
start = cnt;
for (j=SN_offd_i[i]; j < SN_offd_i[i+1]; j++)
{
jj = SN_offd_j[j];
if (data) S_offd_data[cnt] = SN_offd_data[j];
S_offd_j[cnt++] = jj*num_functions;
}
end = cnt;
S_offd_i[row] = cnt;
for (k1=1; k1 < num_functions; k1++)
{
row++;
for (k=start; k < end; k++)
{
if (data) S_offd_data[cnt] = S_offd_data[k];
S_offd_j[cnt++] = S_offd_j[k]+k1;
}
S_offd_i[row] = cnt;
}
}
*S_ptr = S;
return (ierr);
}
HYPRE_Int main( HYPRE_Int argc, char *argv[] )
{
hypre_ParCSRMatrix *par_matrix, *g_matrix, **submatrices;
hypre_CSRMatrix *A_diag, *A_offd;
hypre_CSRBlockMatrix *diag;
hypre_CSRBlockMatrix *offd;
hypre_ParCSRBlockMatrix *par_blk_matrix, *par_blk_matrixT, *rap_matrix;
hypre_Vector *x_local;
hypre_Vector *y_local;
hypre_ParVector *x;
hypre_ParVector *y;
HYPRE_Solver gmres_solver, precon;
HYPRE_Int *diag_i, *diag_j, *offd_i, *offd_j;
HYPRE_Int *diag_i2, *diag_j2, *offd_i2, *offd_j2;
double *diag_d, *diag_d2, *offd_d, *offd_d2;
HYPRE_Int mypid, local_size, nprocs;
HYPRE_Int global_num_rows, global_num_cols, num_cols_offd;
HYPRE_Int num_nonzeros_diag, num_nonzeros_offd, *colMap;
HYPRE_Int ii, jj, kk, row, col, nnz, *indices, *colMap2;
double *data, ddata, *y_data;
HYPRE_Int *row_starts, *col_starts, *rstarts, *cstarts;
HYPRE_Int *row_starts2, *col_starts2;
HYPRE_Int block_size=2, bnnz=4, *index_set;
FILE *fp;
/* --------------------------------------------- */
/* Initialize MPI */
/* --------------------------------------------- */
hypre_MPI_Init(&argc, &argv);
hypre_MPI_Comm_rank(hypre_MPI_COMM_WORLD, &mypid);
hypre_MPI_Comm_size(hypre_MPI_COMM_WORLD, &nprocs);
/* build and fetch matrix */
MyBuildParLaplacian9pt((HYPRE_ParCSRMatrix *) &par_matrix);
global_num_rows = hypre_ParCSRMatrixGlobalNumRows(par_matrix);
global_num_cols = hypre_ParCSRMatrixGlobalNumCols(par_matrix);
row_starts = hypre_ParCSRMatrixRowStarts(par_matrix);
col_starts = hypre_ParCSRMatrixColStarts(par_matrix);
A_diag = hypre_ParCSRMatrixDiag(par_matrix);
A_offd = hypre_ParCSRMatrixOffd(par_matrix);
num_cols_offd = hypre_CSRMatrixNumCols(A_offd);
num_nonzeros_diag = hypre_CSRMatrixNumNonzeros(A_diag);
num_nonzeros_offd = hypre_CSRMatrixNumNonzeros(A_offd);
/* --------------------------------------------- */
/* build vector and apply matvec */
/* --------------------------------------------- */
x = hypre_ParVectorCreate(hypre_MPI_COMM_WORLD,global_num_cols,col_starts);
hypre_ParVectorSetPartitioningOwner(x,0);
hypre_ParVectorInitialize(x);
x_local = hypre_ParVectorLocalVector(x);
data = hypre_VectorData(x_local);
local_size = col_starts[mypid+1] - col_starts[mypid];
for (ii = 0; ii < local_size; ii++) data[ii] = 1.0;
y = hypre_ParVectorCreate(hypre_MPI_COMM_WORLD,global_num_rows,row_starts);
hypre_ParVectorSetPartitioningOwner(y,0);
hypre_ParVectorInitialize(y);
hypre_ParCSRMatrixMatvec (1.0, par_matrix, x, 0.0, y);
ddata = hypre_ParVectorInnerProd(y, y);
if (mypid == 0) hypre_printf("y inner product = %e\n", ddata);
hypre_ParVectorDestroy(x);
hypre_ParVectorDestroy(y);
/* --------------------------------------------- */
/* build block matrix */
/* --------------------------------------------- */
rstarts = hypre_CTAlloc(HYPRE_Int, nprocs+1);
for (ii = 0; ii <= nprocs; ii++) rstarts[ii] = row_starts[ii];
cstarts = hypre_CTAlloc(HYPRE_Int, nprocs+1);
for (ii = 0; ii <= nprocs; ii++) cstarts[ii] = col_starts[ii];
par_blk_matrix = hypre_ParCSRBlockMatrixCreate(hypre_MPI_COMM_WORLD,block_size,
global_num_rows, global_num_cols, rstarts,
cstarts, num_cols_offd, num_nonzeros_diag,
num_nonzeros_offd);
colMap = hypre_ParCSRMatrixColMapOffd(par_matrix);
if (num_cols_offd > 0) colMap2 = hypre_CTAlloc(HYPRE_Int, num_cols_offd);
else colMap2 = NULL;
for (ii = 0; ii < num_cols_offd; ii++) colMap2[ii] = colMap[ii];
hypre_ParCSRBlockMatrixColMapOffd(par_blk_matrix) = colMap2;
diag_i = hypre_CSRMatrixI(hypre_ParCSRMatrixDiag(par_matrix));
diag_j = hypre_CSRMatrixJ(hypre_ParCSRMatrixDiag(par_matrix));
diag_d = hypre_CSRMatrixData(hypre_ParCSRMatrixDiag(par_matrix));
diag = hypre_ParCSRBlockMatrixDiag(par_blk_matrix);
diag_i2 = hypre_CTAlloc(HYPRE_Int, local_size+1);
diag_j2 = hypre_CTAlloc(HYPRE_Int, num_nonzeros_diag);
diag_d2 = hypre_CTAlloc(double, num_nonzeros_diag*bnnz);
for (ii = 0; ii <= local_size; ii++) diag_i2[ii] = diag_i[ii];
for (ii = 0; ii < num_nonzeros_diag; ii++) diag_j2[ii] = diag_j[ii];
hypre_CSRBlockMatrixI(diag) = diag_i2;
hypre_CSRBlockMatrixJ(diag) = diag_j2;
for (ii = 0; ii < num_nonzeros_diag; ii++)
{
for (jj = 0; jj < block_size; jj++)
for (kk = 0; kk < block_size; kk++)
{
if (jj <= kk)
diag_d2[ii*bnnz+jj*block_size+kk] = diag_d[ii];
else
diag_d2[ii*bnnz+jj*block_size+kk] = 0.0;
}
}
hypre_CSRBlockMatrixData(diag) = diag_d2;
offd_i = hypre_CSRMatrixI(hypre_ParCSRMatrixOffd(par_matrix));
offd_j = hypre_CSRMatrixJ(hypre_ParCSRMatrixOffd(par_matrix));
offd_d = hypre_CSRMatrixData(hypre_ParCSRMatrixOffd(par_matrix));
offd = hypre_ParCSRBlockMatrixOffd(par_blk_matrix);
offd_i2 = hypre_CTAlloc(HYPRE_Int, local_size+1);
for (ii = 0; ii <= local_size; ii++) offd_i2[ii] = offd_i[ii];
hypre_CSRBlockMatrixI(offd) = offd_i2;
if (num_cols_offd)
{
offd_j2 = hypre_CTAlloc(HYPRE_Int, num_nonzeros_offd);
for (ii = 0; ii < num_nonzeros_offd; ii++) offd_j2[ii] = offd_j[ii];
hypre_CSRBlockMatrixJ(offd) = offd_j2;
offd_d2 = hypre_CTAlloc(double, num_nonzeros_offd*bnnz);
for (ii = 0; ii < num_nonzeros_offd; ii++)
{
for (jj = 0; jj < block_size; jj++)
for (kk = 0; kk < block_size; kk++)
{
if (jj <= kk)
offd_d2[ii*bnnz+jj*block_size+kk] = offd_d[ii];
else
offd_d2[ii*bnnz+jj*block_size+kk] = 0.0;
}
}
hypre_CSRBlockMatrixData(offd) = offd_d2;
}
else
{
/* Delete any matrix entry C(i,j) for which the corresponding entry P(i,j) doesn't exist -
but only for "fine" rows C(i)<0
This is done as a purely local computation - C and P must have the same data distribution
(among processors).
*/
void hypre_ParCSRMatrixDropEntries( hypre_ParCSRMatrix * C,
hypre_ParCSRMatrix * P, HYPRE_Int * CF_marker )
{
hypre_CSRMatrix *C_diag = hypre_ParCSRMatrixDiag(C);
hypre_CSRMatrix *C_offd = hypre_ParCSRMatrixOffd(C);
double *C_diag_data = hypre_CSRMatrixData(C_diag);
HYPRE_Int *C_diag_i = hypre_CSRMatrixI(C_diag);
HYPRE_Int *C_diag_j = hypre_CSRMatrixJ(C_diag);
double *C_offd_data = hypre_CSRMatrixData(C_offd);
HYPRE_Int *C_offd_i = hypre_CSRMatrixI(C_offd);
HYPRE_Int *C_offd_j = hypre_CSRMatrixJ(C_offd);
hypre_CSRMatrix *P_diag = hypre_ParCSRMatrixDiag(P);
hypre_CSRMatrix *P_offd = hypre_ParCSRMatrixOffd(P);
HYPRE_Int *P_diag_i = hypre_CSRMatrixI(P_diag);
HYPRE_Int *P_diag_j = hypre_CSRMatrixJ(P_diag);
HYPRE_Int *P_offd_i = hypre_CSRMatrixI(P_offd);
HYPRE_Int *P_offd_j = hypre_CSRMatrixJ(P_offd);
HYPRE_Int *new_C_diag_i;
HYPRE_Int *new_C_offd_i;
HYPRE_Int num_rows_diag_C = hypre_CSRMatrixNumRows(C_diag);
HYPRE_Int num_rows_offd_C = hypre_CSRMatrixNumCols(C_offd);
HYPRE_Int num_nonzeros_diag = hypre_CSRMatrixNumNonzeros(C_diag);
HYPRE_Int num_nonzeros_offd = hypre_CSRMatrixNumNonzeros(C_offd);
double vmax = 0.0;
double vmin = 0.0;
double v, old_sum, new_sum, scale;
HYPRE_Int i1, m, m1d, m1o, jC, mP, keep;
/* Repack the i,j,and data arrays of C so as to discard those elements for which
there is no corresponding element in P.
Elements of Coarse rows (CF_marker>=0) are always kept.
The arrays are not re-allocated, so there will generally be unused space
at the ends of the arrays. */
new_C_diag_i = hypre_CTAlloc( HYPRE_Int, num_rows_diag_C+1 );
new_C_offd_i = hypre_CTAlloc( HYPRE_Int, num_rows_offd_C+1 );
m1d = C_diag_i[0];
m1o = C_offd_i[0];
for ( i1 = 0; i1 < num_rows_diag_C; i1++ )
{
old_sum = 0;
new_sum = 0;
for ( m=C_diag_i[i1]; m<C_diag_i[i1+1]; ++m )
{
v = C_diag_data[m];
jC = C_diag_j[m];
old_sum += v;
/* Do we know anything about the order of P_diag_j? It would be better
not to search through it all here. If we know nothing, some ordering or
index scheme will be needed for efficiency (worth doing iff this function
gets called at all ) (may2006: this function is no longer called) */
keep=0;
for ( mP=P_diag_i[i1]; mP<P_diag_i[i1+1]; ++mP )
{
if ( jC==P_diag_j[m] )
{
keep=1;
break;
}
}
if ( CF_marker[i1]>=0 || keep==1 )
{ /* keep v in C */
new_sum += v;
C_diag_j[m1d] = C_diag_j[m];
C_diag_data[m1d] = C_diag_data[m];
++m1d;
}
else
{ /* discard v */
--num_nonzeros_diag;
}
}
for ( m=C_offd_i[i1]; m<C_offd_i[i1+1]; ++m )
{
v = C_offd_data[m];
jC = C_diag_j[m];
old_sum += v;
keep=0;
for ( mP=P_offd_i[i1]; mP<P_offd_i[i1+1]; ++mP )
{
if ( jC==P_offd_j[m] )
{
keep=1;
break;
}
}
if ( CF_marker[i1]>=0 || v>=vmax || v<=vmin )
{ /* keep v in C */
new_sum += v;
C_offd_j[m1o] = C_offd_j[m];
C_offd_data[m1o] = C_offd_data[m];
++m1o;
}
else
{ /* discard v */
--num_nonzeros_offd;
}
}
new_C_diag_i[i1+1] = m1d;
if ( i1<num_rows_offd_C ) new_C_offd_i[i1+1] = m1o;
/* rescale to keep row sum the same */
if (new_sum!=0) scale = old_sum/new_sum; else scale = 1.0;
for ( m=new_C_diag_i[i1]; m<new_C_diag_i[i1+1]; ++m )
C_diag_data[m] *= scale;
if ( i1<num_rows_offd_C ) /* this test fails when there is no offd block */
for ( m=new_C_offd_i[i1]; m<new_C_offd_i[i1+1]; ++m )
C_offd_data[m] *= scale;
}
for ( i1 = 1; i1 <= num_rows_diag_C; i1++ )
{
C_diag_i[i1] = new_C_diag_i[i1];
if ( i1<num_rows_offd_C ) C_offd_i[i1] = new_C_offd_i[i1];
}
hypre_TFree( new_C_diag_i );
if ( num_rows_offd_C>0 ) hypre_TFree( new_C_offd_i );
hypre_CSRMatrixNumNonzeros(C_diag) = num_nonzeros_diag;
hypre_CSRMatrixNumNonzeros(C_offd) = num_nonzeros_offd;
/* SetNumNonzeros, SetDNumNonzeros are global, need hypre_MPI_Allreduce.
I suspect, but don't know, that other parts of hypre do not assume that
the correct values have been set.
hypre_ParCSRMatrixSetNumNonzeros( C );
hypre_ParCSRMatrixSetDNumNonzeros( C );*/
hypre_ParCSRMatrixNumNonzeros( C ) = 0;
hypre_ParCSRMatrixDNumNonzeros( C ) = 0.0;
}
void hypre_ParMatScaleDiagInv_F(
hypre_ParCSRMatrix * C, hypre_ParCSRMatrix * A, double weight, HYPRE_Int * CF_marker )
/* hypre_ParMatScaleDiagInv scales certain rows of its first
* argument by premultiplying with a submatrix of the inverse of
* the diagonal of its second argument; and _also_ multiplying by the scalar
* third argument.
* The marker array determines rows are changed and which diagonal elements
* are used.
*/
{
/*
If A=(Aij),C=(Cik), i&j in Fine+Coarse, k in Coarse, we want
new Cik = (1/aii)*Cik, for Fine i only, all k.
Unlike a matmul, this computation is purely local, only the diag
blocks are involved.
*/
hypre_CSRMatrix *A_diag = hypre_ParCSRMatrixDiag(A);
hypre_CSRMatrix *C_diag = hypre_ParCSRMatrixDiag(C);
hypre_CSRMatrix *C_offd = hypre_ParCSRMatrixOffd(C);
double *A_diag_data = hypre_CSRMatrixData(A_diag);
HYPRE_Int *A_diag_i = hypre_CSRMatrixI(A_diag);
HYPRE_Int *A_diag_j = hypre_CSRMatrixJ(A_diag);
double *C_diag_data = hypre_CSRMatrixData(C_diag);
double *C_offd_data = hypre_CSRMatrixData(C_offd);
HYPRE_Int *C_diag_i = hypre_CSRMatrixI(C_diag);
HYPRE_Int *C_offd_i = hypre_CSRMatrixI(C_offd);
HYPRE_Int num_rows_diag_C = hypre_CSRMatrixNumRows(C_diag);
HYPRE_Int num_cols_offd_C = hypre_CSRMatrixNumCols(C_offd);
HYPRE_Int i1, i2;
HYPRE_Int jj2, jj3;
double a_entry;
/*-----------------------------------------------------------------------
* Loop over C_diag rows.
*-----------------------------------------------------------------------*/
for (i1 = 0; i1 < num_rows_diag_C; i1++)
{
if ( CF_marker[i1] < 0 ) /* Fine data only */
{
/*-----------------------------------------------------------------
* Loop over A_diag data
*-----------------------------------------------------------------*/
for (jj2 = A_diag_i[i1]; jj2 < A_diag_i[i1+1]; jj2++)
{
i2 = A_diag_j[jj2];
if ( i1==i2 ) /* diagonal of A only */
{
a_entry = A_diag_data[jj2] * weight;
/*-----------------------------------------------------------
* Loop over entries in current row of C_diag.
*-----------------------------------------------------------*/
for (jj3 = C_diag_i[i2]; jj3 < C_diag_i[i2+1]; jj3++)
{
C_diag_data[jj3] = C_diag_data[jj3] / a_entry;
}
/*-----------------------------------------------------------
* Loop over entries in current row of C_offd.
*-----------------------------------------------------------*/
if ( num_cols_offd_C )
{
for (jj3 = C_offd_i[i2]; jj3 < C_offd_i[i2+1]; jj3++)
{
C_offd_data[jj3] = C_offd_data[jj3] / a_entry;
}
}
}
}
}
}
}
int
hypre_CSRMatrixMatvec_FF( double alpha,
hypre_CSRMatrix *A,
hypre_Vector *x,
double beta,
hypre_Vector *y,
int *CF_marker_x,
int *CF_marker_y,
int fpt )
{
double *A_data = hypre_CSRMatrixData(A);
int *A_i = hypre_CSRMatrixI(A);
int *A_j = hypre_CSRMatrixJ(A);
int num_rows = hypre_CSRMatrixNumRows(A);
int num_cols = hypre_CSRMatrixNumCols(A);
double *x_data = hypre_VectorData(x);
double *y_data = hypre_VectorData(y);
int x_size = hypre_VectorSize(x);
int y_size = hypre_VectorSize(y);
double temp;
int i, jj;
int ierr = 0;
/*---------------------------------------------------------------------
* Check for size compatibility. Matvec returns ierr = 1 if
* length of X doesn't equal the number of columns of A,
* ierr = 2 if the length of Y doesn't equal the number of rows
* of A, and ierr = 3 if both are true.
*
* Because temporary vectors are often used in Matvec, none of
* these conditions terminates processing, and the ierr flag
* is informational only.
*--------------------------------------------------------------------*/
if (num_cols != x_size)
ierr = 1;
if (num_rows != y_size)
ierr = 2;
if (num_cols != x_size && num_rows != y_size)
ierr = 3;
/*-----------------------------------------------------------------------
* Do (alpha == 0.0) computation - RDF: USE MACHINE EPS
*-----------------------------------------------------------------------*/
if (alpha == 0.0)
{
for (i = 0; i < num_rows; i++)
if (CF_marker_x[i] == fpt) y_data[i] *= beta;
return ierr;
}
/*-----------------------------------------------------------------------
* y = (beta/alpha)*y
*-----------------------------------------------------------------------*/
temp = beta / alpha;
if (temp != 1.0)
{
if (temp == 0.0)
{
for (i = 0; i < num_rows; i++)
if (CF_marker_x[i] == fpt) y_data[i] = 0.0;
}
else
{
for (i = 0; i < num_rows; i++)
if (CF_marker_x[i] == fpt) y_data[i] *= temp;
}
}
/*-----------------------------------------------------------------
* y += A*x
*-----------------------------------------------------------------*/
for (i = 0; i < num_rows; i++)
{
if (CF_marker_x[i] == fpt)
{
temp = y_data[i];
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
if (CF_marker_y[A_j[jj]] == fpt) temp += A_data[jj] * x_data[A_j[jj]];
y_data[i] = temp;
}
}
/*-----------------------------------------------------------------
* y = alpha*y
*-----------------------------------------------------------------*/
if (alpha != 1.0)
{
for (i = 0; i < num_rows; i++)
if (CF_marker_x[i] == fpt) y_data[i] *= alpha;
}
return ierr;
}
int
hypre_CSRMatrixMatvecT( double alpha,
hypre_CSRMatrix *A,
hypre_Vector *x,
double beta,
hypre_Vector *y )
{
double *A_data = hypre_CSRMatrixData(A);
int *A_i = hypre_CSRMatrixI(A);
int *A_j = hypre_CSRMatrixJ(A);
int num_rows = hypre_CSRMatrixNumRows(A);
int num_cols = hypre_CSRMatrixNumCols(A);
double *x_data = hypre_VectorData(x);
double *y_data = hypre_VectorData(y);
int x_size = hypre_VectorSize(x);
int y_size = hypre_VectorSize(y);
int num_vectors = hypre_VectorNumVectors(x);
int idxstride_y = hypre_VectorIndexStride(y);
int vecstride_y = hypre_VectorVectorStride(y);
int idxstride_x = hypre_VectorIndexStride(x);
int vecstride_x = hypre_VectorVectorStride(x);
double temp;
int i, i1, j, jv, jj, ns, ne, size, rest;
int num_threads;
int ierr = 0;
/*---------------------------------------------------------------------
* Check for size compatibility. MatvecT returns ierr = 1 if
* length of X doesn't equal the number of rows of A,
* ierr = 2 if the length of Y doesn't equal the number of
* columns of A, and ierr = 3 if both are true.
*
* Because temporary vectors are often used in MatvecT, none of
* these conditions terminates processing, and the ierr flag
* is informational only.
*--------------------------------------------------------------------*/
hypre_assert( num_vectors == hypre_VectorNumVectors(y) );
if (num_rows != x_size)
ierr = 1;
if (num_cols != y_size)
ierr = 2;
if (num_rows != x_size && num_cols != y_size)
ierr = 3;
/*-----------------------------------------------------------------------
* Do (alpha == 0.0) computation - RDF: USE MACHINE EPS
*-----------------------------------------------------------------------*/
if (alpha == 0.0)
{
for (i = 0; i < num_cols*num_vectors; i++)
y_data[i] *= beta;
return ierr;
}
/*-----------------------------------------------------------------------
* y = (beta/alpha)*y
*-----------------------------------------------------------------------*/
temp = beta / alpha;
if (temp != 1.0)
{
if (temp == 0.0)
{
for (i = 0; i < num_cols*num_vectors; i++)
y_data[i] = 0.0;
}
else
{
for (i = 0; i < num_cols*num_vectors; i++)
y_data[i] *= temp;
}
}
/*-----------------------------------------------------------------
* y += A^T*x
*-----------------------------------------------------------------*/
num_threads = hypre_NumThreads();
if (num_threads > 1)
{
for (i1 = 0; i1 < num_threads; i1++)
{
size = num_cols/num_threads;
rest = num_cols - size*num_threads;
if (i1 < rest)
{
ns = i1*size+i1-1;
ne = (i1+1)*size+i1+1;
}
else
{
ns = i1*size+rest-1;
ne = (i1+1)*size+rest;
}
if ( num_vectors==1 )
{
for (i = 0; i < num_rows; i++)
{
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
j = A_j[jj];
if (j > ns && j < ne)
y_data[j] += A_data[jj] * x_data[i];
}
}
}
else
{
for (i = 0; i < num_rows; i++)
{
for ( jv=0; jv<num_vectors; ++jv )
{
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
j = A_j[jj];
if (j > ns && j < ne)
y_data[ j*idxstride_y + jv*vecstride_y ] +=
A_data[jj] * x_data[ i*idxstride_x + jv*vecstride_x];
}
}
}
}
}
}
else
{
for (i = 0; i < num_rows; i++)
{
if ( num_vectors==1 )
{
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
j = A_j[jj];
y_data[j] += A_data[jj] * x_data[i];
}
}
else
{
for ( jv=0; jv<num_vectors; ++jv )
{
for (jj = A_i[i]; jj < A_i[i+1]; jj++)
{
j = A_j[jj];
y_data[ j*idxstride_y + jv*vecstride_y ] +=
A_data[jj] * x_data[ i*idxstride_x + jv*vecstride_x ];
}
}
}
}
}
/*-----------------------------------------------------------------
* y = alpha*y
*-----------------------------------------------------------------*/
if (alpha != 1.0)
{
for (i = 0; i < num_cols*num_vectors; i++)
y_data[i] *= alpha;
}
return ierr;
}
/*****************************************************************************
*
* Routine for constructing graph domain_dof with minimal overlap
* and computing the respective matrix inverses to be
* used in an overlapping Schwarz procedure (like smoother
* in AMG);
*
*****************************************************************************/
HYPRE_Int
hypre_AMGCreateDomainDof(hypre_CSRMatrix *A,
HYPRE_Int **i_domain_dof_pointer,
HYPRE_Int **j_domain_dof_pointer,
HYPRE_Real **domain_matrixinverse_pointer,
HYPRE_Int *num_domains_pointer)
{
HYPRE_Int *i_domain_dof, *j_domain_dof;
HYPRE_Real *domain_matrixinverse;
HYPRE_Int num_domains;
HYPRE_Int *i_dof_dof = hypre_CSRMatrixI(A);
HYPRE_Int *j_dof_dof = hypre_CSRMatrixJ(A);
HYPRE_Real *a_dof_dof = hypre_CSRMatrixData(A);
HYPRE_Int num_dofs = hypre_CSRMatrixNumRows(A);
/* HYPRE_Int *i_dof_to_accept_weight; */
HYPRE_Int *i_dof_to_prefer_weight,
*w_dof_dof, *i_dof_weight;
HYPRE_Int *i_dof_to_aggregate, *i_aggregate_dof, *j_aggregate_dof;
HYPRE_Int *i_dof_index;
HYPRE_Int ierr = 0;
HYPRE_Int i,j,k, l_loc, i_loc, j_loc;
HYPRE_Int i_dof;
HYPRE_Int *i_local_to_global;
HYPRE_Int *i_global_to_local;
HYPRE_Int local_dof_counter, max_local_dof_counter=0;
HYPRE_Int domain_dof_counter = 0, domain_matrixinverse_counter = 0;
HYPRE_Real *AE, *XE;
/* PCG arrays: --------------------------------------------------- */
/* HYPRE_Real *x, *rhs, *v, *w, *d, *aux;
HYPRE_Int max_iter; */
/* --------------------------------------------------------------------- */
/*=======================================================================*/
/* create artificial domains by agglomeration; */
/*=======================================================================*/
hypre_printf("----------- create artificials domain by agglomeration; ======\n");
i_dof_to_prefer_weight = (HYPRE_Int *) malloc(num_dofs * sizeof(HYPRE_Int));
w_dof_dof = (HYPRE_Int *) malloc(i_dof_dof[num_dofs] * sizeof(HYPRE_Int));
for (i=0; i < num_dofs; i++)
i_dof_to_prefer_weight[i] = 0;
for (i=0; i<num_dofs; i++)
for (j=i_dof_dof[i]; j< i_dof_dof[i+1]; j++)
{
if (j_dof_dof[j] == i)
w_dof_dof[j]=0;
else
w_dof_dof[j]=1;
}
hypre_printf("end computing weights for agglomeration procedure: --------\n");
i_dof_weight = (HYPRE_Int *) malloc(num_dofs * sizeof(HYPRE_Int));
i_aggregate_dof = (HYPRE_Int *) malloc(num_dofs * sizeof(HYPRE_Int));
j_aggregate_dof= (HYPRE_Int *) malloc(num_dofs * sizeof(HYPRE_Int));
ierr = hypre_AMGeAgglomerate(i_aggregate_dof, j_aggregate_dof,
i_dof_dof, j_dof_dof, w_dof_dof,
i_dof_dof, j_dof_dof,
i_dof_dof, j_dof_dof,
i_dof_to_prefer_weight,
i_dof_weight,
num_dofs, num_dofs,
&num_domains);
hypre_printf("num_dofs: %d, num_domains: %d\n", num_dofs, num_domains);
i_dof_to_aggregate = (HYPRE_Int *) malloc(num_dofs * sizeof(HYPRE_Int));
for (i=0; i < num_domains; i++)
for (j=i_aggregate_dof[i]; j < i_aggregate_dof[i+1]; j++)
i_dof_to_aggregate[j_aggregate_dof[j]] = i;
/*
hypre_printf("========================================================\n");
hypre_printf("== artificial non--overlapping domains (aggregates): ===\n");
hypre_printf("========================================================\n");
for (i=0; i < num_domains; i++)
{
hypre_printf("\n aggregate %d:\n", i);
for (j=i_aggregate_dof[i]; j < i_aggregate_dof[i+1]; j++)
hypre_printf("%d, ", j_aggregate_dof[j]);
hypre_printf("\n");
}
*/
free(i_dof_to_prefer_weight);
free(i_dof_weight);
free(w_dof_dof);
/* make domains from aggregates: *********************************/
i_domain_dof = (HYPRE_Int *) malloc((num_domains+1) * sizeof(HYPRE_Int));
i_dof_index = (HYPRE_Int *) malloc(num_dofs * sizeof(HYPRE_Int));
for (i=0; i < num_dofs; i++)
i_dof_index[i] = -1;
domain_dof_counter=0;
for (i=0; i < num_domains; i++)
{
i_domain_dof[i] = domain_dof_counter;
for (j=i_aggregate_dof[i]; j < i_aggregate_dof[i+1]; j++)
for (k=i_dof_dof[j_aggregate_dof[j]];
k<i_dof_dof[j_aggregate_dof[j]+1]; k++)
if (i_dof_to_aggregate[j_dof_dof[k]] >= i
&& i_dof_index[j_dof_dof[k]]==-1)
{
i_dof_index[j_dof_dof[k]]++;
domain_dof_counter++;
}
for (j=i_aggregate_dof[i]; j < i_aggregate_dof[i+1]; j++)
for (k=i_dof_dof[j_aggregate_dof[j]];
k<i_dof_dof[j_aggregate_dof[j]+1]; k++)
i_dof_index[j_dof_dof[k]]=-1;
}
i_domain_dof[num_domains] = domain_dof_counter;
j_domain_dof = (HYPRE_Int *) malloc(domain_dof_counter * sizeof(HYPRE_Int));
domain_dof_counter=0;
for (i=0; i < num_domains; i++)
{
for (j=i_aggregate_dof[i]; j < i_aggregate_dof[i+1]; j++)
for (k=i_dof_dof[j_aggregate_dof[j]];
k<i_dof_dof[j_aggregate_dof[j]+1]; k++)
if (i_dof_to_aggregate[j_dof_dof[k]] >= i
&& i_dof_index[j_dof_dof[k]]==-1)
{
i_dof_index[j_dof_dof[k]]++;
j_domain_dof[domain_dof_counter] = j_dof_dof[k];
domain_dof_counter++;
}
for (j=i_aggregate_dof[i]; j < i_aggregate_dof[i+1]; j++)
for (k=i_dof_dof[j_aggregate_dof[j]];
k<i_dof_dof[j_aggregate_dof[j]+1]; k++)
i_dof_index[j_dof_dof[k]]=-1;
}
free(i_aggregate_dof);
free(j_aggregate_dof);
free(i_dof_to_aggregate);
/*
i_domain_dof = i_aggregate_dof;
j_domain_dof = j_aggregate_dof;
*/
hypre_printf("END domain_dof computations: =================================\n");
domain_matrixinverse_counter = 0;
local_dof_counter = 0;
for (i=0; i < num_domains; i++)
{
local_dof_counter = i_domain_dof[i+1]-i_domain_dof[i];
domain_matrixinverse_counter+= local_dof_counter * local_dof_counter;
if (local_dof_counter > max_local_dof_counter)
max_local_dof_counter = local_dof_counter;
}
domain_matrixinverse = hypre_CTAlloc(HYPRE_Real, domain_matrixinverse_counter);
i_local_to_global = hypre_CTAlloc(HYPRE_Int, max_local_dof_counter);
AE = hypre_CTAlloc(HYPRE_Real, max_local_dof_counter *
max_local_dof_counter);
XE = hypre_CTAlloc(HYPRE_Real, max_local_dof_counter *
max_local_dof_counter);
/* i_dof_index = (HYPRE_Int *) malloc(num_dofs * sizeof(HYPRE_Int)); */
i_global_to_local = i_dof_index;
for (i=0; i < num_dofs; i++)
i_global_to_local[i] = -1;
domain_matrixinverse_counter = 0;
for (i=0; i < num_domains; i++)
{
local_dof_counter = 0;
for (j=i_domain_dof[i]; j < i_domain_dof[i+1]; j++)
{
i_global_to_local[j_domain_dof[j]] = local_dof_counter;
i_local_to_global[local_dof_counter] = j_domain_dof[j];
local_dof_counter++;
}
/* get local matrix in AE: ======================================== */
for (i_loc=0; i_loc < local_dof_counter; i_loc++)
for (j_loc=0; j_loc < local_dof_counter; j_loc++)
AE[i_loc + j_loc * local_dof_counter] = 0.e0;
for (i_loc=0; i_loc < local_dof_counter; i_loc++)
{
i_dof = i_local_to_global[i_loc];
for (j=i_dof_dof[i_dof]; j < i_dof_dof[i_dof+1]; j++)
{
j_loc = i_global_to_local[j_dof_dof[j]];
if (j_loc >=0)
AE[i_loc + j_loc * local_dof_counter] = a_dof_dof[j];
}
}
/* get block for Schwarz smoother: ============================= */
ierr = matinv(XE, AE, local_dof_counter);
/* hypre_printf("ierr_AE_inv: %d\n", ierr); */
for (i_loc=0; i_loc < local_dof_counter; i_loc++)
for (j_loc=0; j_loc < local_dof_counter; j_loc++)
domain_matrixinverse[domain_matrixinverse_counter
+ i_loc + j_loc * local_dof_counter]
= XE[i_loc + j_loc * local_dof_counter];
domain_matrixinverse_counter+=local_dof_counter*local_dof_counter;
for (l_loc=0; l_loc < local_dof_counter; l_loc++)
i_global_to_local[i_local_to_global[l_loc]] = -1;
}
hypre_TFree(i_local_to_global);
hypre_TFree(AE);
hypre_TFree(XE);
hypre_TFree(i_dof_index);
*i_domain_dof_pointer = i_domain_dof;
*j_domain_dof_pointer = j_domain_dof;
*num_domains_pointer = num_domains;
*domain_matrixinverse_pointer = domain_matrixinverse;
/*
x = hypre_CTAlloc(HYPRE_Real, num_dofs);
rhs = hypre_CTAlloc(HYPRE_Real, num_dofs);
v = hypre_CTAlloc(HYPRE_Real, num_dofs);
w = hypre_CTAlloc(HYPRE_Real, num_dofs);
d = hypre_CTAlloc(HYPRE_Real, num_dofs);
aux = hypre_CTAlloc(HYPRE_Real, num_dofs);
for (i=0; i < num_dofs; i++)
x[i] = 0.e0;
for (i=0; i < num_dofs; i++)
rhs[i] = rand();
max_iter = 1000;
hypre_printf("\nenter SchwarzPCG: =======================================\n");
ierr = hypre_Schwarzpcg(x, rhs,
a_dof_dof,
i_dof_dof, j_dof_dof,
i_domain_dof, j_domain_dof,
domain_matrixinverse,
num_domains,
v, w, d, aux,
max_iter,
num_dofs);
hypre_printf("\n\n=======================================================\n");
hypre_printf(" END test PCG solve: \n");
hypre_printf("===========================================================\n");
hypre_TFree(x);
hypre_TFree(rhs);
hypre_TFree(aux);
hypre_TFree(v);
hypre_TFree(w);
hypre_TFree(d);
hypre_TFree(i_domain_dof);
hypre_TFree(j_domain_dof);
hypre_TFree(domain_matrixinverse);
*/
return ierr;
}
HYPRE_Int hypre_SchwarzSolve(hypre_CSRMatrix *A,
hypre_Vector *rhs_vector,
HYPRE_Int num_domains,
HYPRE_Int *i_domain_dof,
HYPRE_Int *j_domain_dof,
HYPRE_Real *domain_matrixinverse,
hypre_Vector *x_vector,
hypre_Vector *aux_vector)
{
HYPRE_Int ierr = 0;
/* HYPRE_Int num_dofs; */
HYPRE_Int *i_dof_dof;
HYPRE_Int *j_dof_dof;
HYPRE_Real *a_dof_dof;
HYPRE_Real *x;
HYPRE_Real *rhs;
HYPRE_Real *aux;
HYPRE_Int i,j,k, j_loc, k_loc;
HYPRE_Int matrix_size, matrix_size_counter = 0;
/* initiate: ----------------------------------------------- */
/* num_dofs = hypre_CSRMatrixNumRows(A); */
i_dof_dof = hypre_CSRMatrixI(A);
j_dof_dof = hypre_CSRMatrixJ(A);
a_dof_dof = hypre_CSRMatrixData(A);
x = hypre_VectorData(x_vector);
rhs = hypre_VectorData(rhs_vector);
aux = hypre_VectorData(aux_vector);
/* for (i=0; i < num_dofs; i++)
x[i] = 0.e0; */
/* forward solve: ----------------------------------------------- */
matrix_size_counter = 0;
for (i=0; i < num_domains; i++)
{
matrix_size = i_domain_dof[i+1] - i_domain_dof[i];
/* compute residual: ---------------------------------------- */
for (j=i_domain_dof[i]; j < i_domain_dof[i+1]; j++)
{
aux[j_domain_dof[j]] = rhs[j_domain_dof[j]];
for (k=i_dof_dof[j_domain_dof[j]];
k<i_dof_dof[j_domain_dof[j]+1]; k++)
aux[j_domain_dof[j]] -= a_dof_dof[k] * x[j_dof_dof[k]];
}
/* solve for correction: ------------------------------------- */
for (j=i_domain_dof[i]; j < i_domain_dof[i+1]; j++)
{
j_loc = j-i_domain_dof[i];
for (k=i_domain_dof[i]; k < i_domain_dof[i+1]; k++)
{
k_loc = k-i_domain_dof[i];
x[j_domain_dof[j]]+=
domain_matrixinverse[matrix_size_counter
+ j_loc + k_loc * matrix_size]
* aux[j_domain_dof[k]];
}
}
matrix_size_counter += matrix_size * matrix_size;
}
/* backward solve: ------------------------------------------------ */
for (i=num_domains-1; i > -1; i--)
{
matrix_size = i_domain_dof[i+1] - i_domain_dof[i];
matrix_size_counter -= matrix_size * matrix_size;
/* compute residual: ---------------------------------------- */
for (j=i_domain_dof[i]; j < i_domain_dof[i+1]; j++)
{
aux[j_domain_dof[j]] = rhs[j_domain_dof[j]];
for (k=i_dof_dof[j_domain_dof[j]];
k<i_dof_dof[j_domain_dof[j]+1]; k++)
aux[j_domain_dof[j]] -= a_dof_dof[k] * x[j_dof_dof[k]];
}
/* solve for correction: ------------------------------------- */
for (j=i_domain_dof[i]; j < i_domain_dof[i+1]; j++)
{
j_loc = j-i_domain_dof[i];
for (k=i_domain_dof[i]; k < i_domain_dof[i+1]; k++)
{
k_loc = k-i_domain_dof[i];
x[j_domain_dof[j]]+=
domain_matrixinverse[matrix_size_counter
+ j_loc + k_loc * matrix_size]
* aux[j_domain_dof[k]];
}
}
}
return ierr;
}
HYPRE_Int
hypre_AMGNodalSchwarzSmoother( hypre_CSRMatrix *A,
HYPRE_Int *dof_func,
HYPRE_Int num_functions,
HYPRE_Int option,
HYPRE_Int **i_domain_dof_pointer,
HYPRE_Int **j_domain_dof_pointer,
HYPRE_Real **domain_matrixinverse_pointer,
HYPRE_Int *num_domains_pointer)
{
/* option = 0: nodal symGS;
1: next to nodal symGS (overlapping Schwarz) */
HYPRE_Int *i_domain_dof, *j_domain_dof;
HYPRE_Real *domain_matrixinverse;
HYPRE_Int num_domains;
HYPRE_Int *i_dof_node, *j_dof_node;
HYPRE_Int *i_node_dof, *j_node_dof;
HYPRE_Int *i_node_dof_dof, *j_node_dof_dof;
HYPRE_Int *i_node_node, *j_node_node;
HYPRE_Int num_nodes;
HYPRE_Int *i_dof_dof = hypre_CSRMatrixI(A);
HYPRE_Int *j_dof_dof = hypre_CSRMatrixJ(A);
HYPRE_Real *a_dof_dof = hypre_CSRMatrixData(A);
HYPRE_Int num_dofs = hypre_CSRMatrixNumRows(A);
HYPRE_Int ierr = 0;
HYPRE_Int i,j,k, l_loc, i_loc, j_loc;
HYPRE_Int i_dof, j_dof;
HYPRE_Int *i_local_to_global;
HYPRE_Int *i_global_to_local;
HYPRE_Int *i_int;
HYPRE_Int *i_int_to_local;
HYPRE_Int int_dof_counter, local_dof_counter, max_local_dof_counter=0;
HYPRE_Int domain_dof_counter = 0, domain_matrixinverse_counter = 0;
HYPRE_Real *AE, *XE;
/* PCG arrays: ---------------------------------------------------
HYPRE_Real *x, *rhs, *v, *w, *d, *aux;
HYPRE_Int max_iter;
------------------------------------------------------------------ */
/* build dof_node graph: ----------------------------------------- */
num_nodes = num_dofs / num_functions;
hypre_printf("\nnum_nodes: %d, num_dofs: %d = %d x %d\n", num_nodes, num_dofs,
num_nodes, num_functions);
i_dof_node = hypre_CTAlloc(HYPRE_Int, num_dofs+1);
j_dof_node = hypre_CTAlloc(HYPRE_Int, num_dofs);
for (i=0; i < num_dofs+1; i++)
i_dof_node[i] = i;
for (j = 0; j < num_nodes; j++)
for (k = 0; k < num_functions; k++)
j_dof_node[j*num_functions+k] = j;
/* build node_dof graph: ----------------------------------------- */
ierr = transpose_matrix_create(&i_node_dof, &j_node_dof,
i_dof_node, j_dof_node,
num_dofs, num_nodes);
/* build node_node graph: ----------------------------------------- */
ierr = matrix_matrix_product(&i_node_dof_dof,
&j_node_dof_dof,
i_node_dof, j_node_dof,
i_dof_dof, j_dof_dof,
num_nodes, num_dofs, num_dofs);
ierr = matrix_matrix_product(&i_node_node,
&j_node_node,
i_node_dof_dof,
j_node_dof_dof,
i_dof_node, j_dof_node,
num_nodes, num_dofs, num_nodes);
hypre_TFree(i_node_dof_dof);
hypre_TFree(j_node_dof_dof);
/* compute for each node the local information: -------------------- */
i_global_to_local = i_dof_node;
for (i_dof =0; i_dof < num_dofs; i_dof++)
i_global_to_local[i_dof] = -1;
domain_matrixinverse_counter = 0;
domain_dof_counter = 0;
for (i=0; i < num_nodes; i++)
{
local_dof_counter = 0;
for (j=i_node_node[i]; j < i_node_node[i+1]; j++)
for (k=i_node_dof[j_node_node[j]];
k<i_node_dof[j_node_node[j]+1]; k++)
{
j_dof = j_node_dof[k];
if (i_global_to_local[j_dof] < 0)
{
i_global_to_local[j_dof] = local_dof_counter;
local_dof_counter++;
}
}
domain_matrixinverse_counter += local_dof_counter*local_dof_counter;
domain_dof_counter += local_dof_counter;
if (local_dof_counter > max_local_dof_counter)
max_local_dof_counter = local_dof_counter;
for (j=i_node_node[i]; j < i_node_node[i+1]; j++)
for (k=i_node_dof[j_node_node[j]];
k<i_node_dof[j_node_node[j]+1]; k++)
{
j_dof = j_node_dof[k];
i_global_to_local[j_dof] = -1;
}
}
num_domains = num_nodes;
i_domain_dof = hypre_CTAlloc(HYPRE_Int, num_domains+1);
if (option == 1)
j_domain_dof = hypre_CTAlloc(HYPRE_Int, domain_dof_counter);
else
j_domain_dof = hypre_CTAlloc(HYPRE_Int, num_dofs);
if (option == 1)
domain_matrixinverse = hypre_CTAlloc(HYPRE_Real, domain_matrixinverse_counter);
else
domain_matrixinverse = hypre_CTAlloc(HYPRE_Real, num_dofs * num_functions);
i_local_to_global = hypre_CTAlloc(HYPRE_Int, max_local_dof_counter);
AE = hypre_CTAlloc(HYPRE_Real, max_local_dof_counter *
max_local_dof_counter);
XE = hypre_CTAlloc(HYPRE_Real, max_local_dof_counter *
max_local_dof_counter);
i_int_to_local = hypre_CTAlloc(HYPRE_Int, max_local_dof_counter);
i_int = hypre_CTAlloc(HYPRE_Int, max_local_dof_counter);
for (l_loc=0; l_loc < max_local_dof_counter; l_loc++)
i_int[l_loc] = -1;
domain_dof_counter = 0;
domain_matrixinverse_counter = 0;
for (i=0; i < num_nodes; i++)
{
i_domain_dof[i] = domain_dof_counter;
local_dof_counter = 0;
for (j=i_node_node[i]; j < i_node_node[i+1]; j++)
for (k=i_node_dof[j_node_node[j]];
k<i_node_dof[j_node_node[j]+1]; k++)
{
j_dof = j_node_dof[k];
if (i_global_to_local[j_dof] < 0)
{
i_global_to_local[j_dof] = local_dof_counter;
i_local_to_global[local_dof_counter] = j_dof;
local_dof_counter++;
}
}
for (j=i_node_dof[i]; j < i_node_dof[i+1]; j++)
for (k=i_dof_dof[j_node_dof[j]]; k < i_dof_dof[j_node_dof[j]+1]; k++)
if (i_global_to_local[j_dof_dof[k]] < 0)
hypre_printf("WRONG local indexing: ====================== \n");
int_dof_counter = 0;
for (k=i_node_dof[i]; k < i_node_dof[i+1]; k++)
{
i_dof = j_node_dof[k];
i_loc = i_global_to_local[i_dof];
i_int[i_loc] = int_dof_counter;
i_int_to_local[int_dof_counter] = i_loc;
int_dof_counter++;
}
/* get local matrix AE: ======================================== */
if (option == 1)
{
for (i_loc=0; i_loc < local_dof_counter; i_loc++)
for (j_loc=0; j_loc < local_dof_counter; j_loc++)
AE[i_loc + j_loc * local_dof_counter] = 0.e0;
for (i_loc=0; i_loc < local_dof_counter; i_loc++)
{
i_dof = i_local_to_global[i_loc];
for (j=i_dof_dof[i_dof]; j < i_dof_dof[i_dof+1]; j++)
{
j_loc = i_global_to_local[j_dof_dof[j]];
if (j_loc >=0)
AE[i_loc + j_loc * local_dof_counter] = a_dof_dof[j];
}
}
/* get block for Schwarz smoother: ============================= */
ierr = matinv(XE, AE, local_dof_counter);
/* hypre_printf("ierr_AE_inv: %d\n", ierr); */
}
if (option == 1)
for (i_loc=0; i_loc < local_dof_counter; i_loc++)
j_domain_dof[domain_dof_counter+i_loc]
= i_local_to_global[i_loc];
if (option == 1)
for (i_loc=0; i_loc < local_dof_counter; i_loc++)
for (j_loc=0; j_loc < local_dof_counter; j_loc++)
domain_matrixinverse[domain_matrixinverse_counter
+ i_loc + j_loc * local_dof_counter]
= XE[i_loc + j_loc * local_dof_counter];
if (option == 0)
{
for (i_loc=0; i_loc < int_dof_counter; i_loc++)
for (j_loc=0; j_loc < int_dof_counter; j_loc++)
AE[i_loc + j_loc * int_dof_counter] = 0.e0;
for (l_loc=0; l_loc < int_dof_counter; l_loc++)
{
i_loc = i_int_to_local[l_loc];
i_dof = i_local_to_global[i_loc];
for (j=i_dof_dof[i_dof]; j < i_dof_dof[i_dof+1]; j++)
{
j_loc = i_global_to_local[j_dof_dof[j]];
if (j_loc >=0)
if (i_int[j_loc] >=0)
AE[i_loc + i_int[j_loc] * int_dof_counter]
= a_dof_dof[j];
}
}
ierr = matinv(XE, AE, int_dof_counter);
for (i_loc=0; i_loc < int_dof_counter; i_loc++)
{
j_domain_dof[domain_dof_counter + i_loc] =
i_local_to_global[i_int_to_local[i_loc]];
for (j_loc=0; j_loc < int_dof_counter; j_loc++)
domain_matrixinverse[domain_matrixinverse_counter
+ i_loc + j_loc * int_dof_counter]
= XE[i_loc + j_loc * int_dof_counter];
}
domain_dof_counter+=int_dof_counter;
domain_matrixinverse_counter+=int_dof_counter*int_dof_counter;
}
else
{
domain_dof_counter+=local_dof_counter;
domain_matrixinverse_counter+=local_dof_counter*local_dof_counter;
}
for (l_loc=0; l_loc < local_dof_counter; l_loc++)
{
i_int[l_loc] = -1;
i_global_to_local[i_local_to_global[l_loc]] = -1;
}
}
i_domain_dof[num_nodes] = domain_dof_counter;
hypre_TFree(i_dof_node);
hypre_TFree(j_dof_node);
hypre_TFree(i_node_dof);
hypre_TFree(j_node_dof);
hypre_TFree(i_node_node);
hypre_TFree(j_node_node);
hypre_TFree(i_int);
hypre_TFree(i_int_to_local);
hypre_TFree(i_local_to_global);
hypre_TFree(AE);
hypre_TFree(XE);
*i_domain_dof_pointer = i_domain_dof;
*j_domain_dof_pointer = j_domain_dof;
*num_domains_pointer = num_domains;
*domain_matrixinverse_pointer = domain_matrixinverse;
/* hypre_printf("exit *Schwarz*: ===============================\n\n"); */
/* -----------------------------------------------------------------
x = hypre_CTAlloc(HYPRE_Real, num_dofs);
rhs = hypre_CTAlloc(HYPRE_Real, num_dofs);
v = hypre_CTAlloc(HYPRE_Real, num_dofs);
w = hypre_CTAlloc(HYPRE_Real, num_dofs);
d = hypre_CTAlloc(HYPRE_Real, num_dofs);
aux = hypre_CTAlloc(HYPRE_Real, num_dofs);
for (i=0; i < num_dofs; i++)
x[i] = 0.e0;
for (i=0; i < num_dofs; i++)
rhs[i] = rand();
max_iter = 1000;
hypre_printf("\nenter SchwarzPCG: =======================================\n");
ierr = hypre_Schwarzpcg(x, rhs,
a_dof_dof,
i_dof_dof, j_dof_dof,
i_domain_dof, j_domain_dof,
domain_matrixinverse,
num_domains,
v, w, d, aux,
max_iter,
num_dofs);
hypre_printf("\n\n=======================================================\n");
hypre_printf(" END test PCG solve: \n");
hypre_printf("===========================================================\n");
hypre_TFree(x);
hypre_TFree(rhs);
hypre_TFree(aux);
hypre_TFree(v);
hypre_TFree(w);
hypre_TFree(d);
----------------------------------------------------------------------- */
return ierr;
}
hypre_ParCSRMatrix * hypre_ParMatmul_FC(
hypre_ParCSRMatrix * A, hypre_ParCSRMatrix * P, HYPRE_Int * CF_marker,
HYPRE_Int * dof_func, HYPRE_Int * dof_func_offd )
/* hypre_parMatmul_FC creates and returns the "Fine"-designated rows of the
matrix product A*P. A's size is (nC+nF)*(nC+nF), P's size is (nC+nF)*nC
where nC is the number of coarse rows/columns, nF the number of fine
rows/columns. The size of C=A*P is (nC+nF)*nC, even though not all rows
of C are actually computed. If we were to construct a matrix consisting
only of the computed rows of C, its size would be nF*nC.
"Fine" is defined solely by the marker array, and for example could be
a proper subset of the fine points of a multigrid hierarchy.
*/
{
/* To compute a submatrix of C containing only the computed data, i.e.
only "Fine" rows, we would have to do a lot of computational work,
with a lot of communication. The communication is because such a
matrix would need global information that depends on which rows are
"Fine".
*/
MPI_Comm comm = hypre_ParCSRMatrixComm(A);
hypre_CSRMatrix *A_diag = hypre_ParCSRMatrixDiag(A);
double *A_diag_data = hypre_CSRMatrixData(A_diag);
HYPRE_Int *A_diag_i = hypre_CSRMatrixI(A_diag);
HYPRE_Int *A_diag_j = hypre_CSRMatrixJ(A_diag);
hypre_CSRMatrix *A_offd = hypre_ParCSRMatrixOffd(A);
double *A_offd_data = hypre_CSRMatrixData(A_offd);
HYPRE_Int *A_offd_i = hypre_CSRMatrixI(A_offd);
HYPRE_Int *A_offd_j = hypre_CSRMatrixJ(A_offd);
HYPRE_Int *row_starts_A = hypre_ParCSRMatrixRowStarts(A);
HYPRE_Int num_rows_diag_A = hypre_CSRMatrixNumRows(A_diag);
HYPRE_Int num_cols_diag_A = hypre_CSRMatrixNumCols(A_diag);
HYPRE_Int num_cols_offd_A = hypre_CSRMatrixNumCols(A_offd);
hypre_CSRMatrix *P_diag = hypre_ParCSRMatrixDiag(P);
double *P_diag_data = hypre_CSRMatrixData(P_diag);
HYPRE_Int *P_diag_i = hypre_CSRMatrixI(P_diag);
HYPRE_Int *P_diag_j = hypre_CSRMatrixJ(P_diag);
hypre_CSRMatrix *P_offd = hypre_ParCSRMatrixOffd(P);
HYPRE_Int *col_map_offd_P = hypre_ParCSRMatrixColMapOffd(P);
double *P_offd_data = hypre_CSRMatrixData(P_offd);
HYPRE_Int *P_offd_i = hypre_CSRMatrixI(P_offd);
HYPRE_Int *P_offd_j = hypre_CSRMatrixJ(P_offd);
HYPRE_Int first_col_diag_P = hypre_ParCSRMatrixFirstColDiag(P);
HYPRE_Int last_col_diag_P;
HYPRE_Int *col_starts_P = hypre_ParCSRMatrixColStarts(P);
HYPRE_Int num_rows_diag_P = hypre_CSRMatrixNumRows(P_diag);
HYPRE_Int num_cols_diag_P = hypre_CSRMatrixNumCols(P_diag);
HYPRE_Int num_cols_offd_P = hypre_CSRMatrixNumCols(P_offd);
hypre_ParCSRMatrix *C;
HYPRE_Int *col_map_offd_C;
HYPRE_Int *map_P_to_C;
hypre_CSRMatrix *C_diag;
double *C_diag_data;
HYPRE_Int *C_diag_i;
HYPRE_Int *C_diag_j;
hypre_CSRMatrix *C_offd;
double *C_offd_data=NULL;
HYPRE_Int *C_offd_i=NULL;
HYPRE_Int *C_offd_j=NULL;
HYPRE_Int C_diag_size;
HYPRE_Int C_offd_size;
HYPRE_Int num_cols_offd_C = 0;
hypre_CSRMatrix *Ps_ext;
double *Ps_ext_data;
HYPRE_Int *Ps_ext_i;
HYPRE_Int *Ps_ext_j;
double *P_ext_diag_data;
HYPRE_Int *P_ext_diag_i;
HYPRE_Int *P_ext_diag_j;
HYPRE_Int P_ext_diag_size;
double *P_ext_offd_data;
HYPRE_Int *P_ext_offd_i;
HYPRE_Int *P_ext_offd_j;
HYPRE_Int P_ext_offd_size;
HYPRE_Int *P_marker;
HYPRE_Int *temp;
HYPRE_Int i, j;
HYPRE_Int i1, i2, i3;
HYPRE_Int jj2, jj3;
HYPRE_Int jj_count_diag, jj_count_offd;
HYPRE_Int jj_row_begin_diag, jj_row_begin_offd;
HYPRE_Int start_indexing = 0; /* start indexing for C_data at 0 */
HYPRE_Int n_rows_A_global, n_cols_A_global;
HYPRE_Int n_rows_P_global, n_cols_P_global;
HYPRE_Int allsquare = 0;
HYPRE_Int cnt, cnt_offd, cnt_diag;
HYPRE_Int num_procs;
HYPRE_Int value;
double a_entry;
double a_b_product;
n_rows_A_global = hypre_ParCSRMatrixGlobalNumRows(A);
n_cols_A_global = hypre_ParCSRMatrixGlobalNumCols(A);
n_rows_P_global = hypre_ParCSRMatrixGlobalNumRows(P);
n_cols_P_global = hypre_ParCSRMatrixGlobalNumCols(P);
if (n_cols_A_global != n_rows_P_global || num_cols_diag_A != num_rows_diag_P)
{
hypre_printf(" Error! Incompatible matrix dimensions!\n");
return NULL;
}
/* if (num_rows_A==num_cols_P) allsquare = 1; */
/*-----------------------------------------------------------------------
* Extract P_ext, i.e. portion of P that is stored on neighbor procs
* and needed locally for matrix matrix product
*-----------------------------------------------------------------------*/
hypre_MPI_Comm_size(comm, &num_procs);
if (num_procs > 1)
{
/*---------------------------------------------------------------------
* If there exists no CommPkg for A, a CommPkg is generated using
* equally load balanced partitionings within
* hypre_ParCSRMatrixExtractBExt
*--------------------------------------------------------------------*/
Ps_ext = hypre_ParCSRMatrixExtractBExt(P,A,1);
Ps_ext_data = hypre_CSRMatrixData(Ps_ext);
Ps_ext_i = hypre_CSRMatrixI(Ps_ext);
Ps_ext_j = hypre_CSRMatrixJ(Ps_ext);
}
P_ext_diag_i = hypre_CTAlloc(HYPRE_Int, num_cols_offd_A+1);
P_ext_offd_i = hypre_CTAlloc(HYPRE_Int, num_cols_offd_A+1);
P_ext_diag_size = 0;
P_ext_offd_size = 0;
last_col_diag_P = first_col_diag_P + num_cols_diag_P -1;
for (i=0; i < num_cols_offd_A; i++)
{
for (j=Ps_ext_i[i]; j < Ps_ext_i[i+1]; j++)
if (Ps_ext_j[j] < first_col_diag_P || Ps_ext_j[j] > last_col_diag_P)
P_ext_offd_size++;
else
P_ext_diag_size++;
P_ext_diag_i[i+1] = P_ext_diag_size;
P_ext_offd_i[i+1] = P_ext_offd_size;
}
if (P_ext_diag_size)
{
P_ext_diag_j = hypre_CTAlloc(HYPRE_Int, P_ext_diag_size);
P_ext_diag_data = hypre_CTAlloc(double, P_ext_diag_size);
}
if (P_ext_offd_size)
{
P_ext_offd_j = hypre_CTAlloc(HYPRE_Int, P_ext_offd_size);
P_ext_offd_data = hypre_CTAlloc(double, P_ext_offd_size);
}
cnt_offd = 0;
cnt_diag = 0;
for (i=0; i < num_cols_offd_A; i++)
{
for (j=Ps_ext_i[i]; j < Ps_ext_i[i+1]; j++)
if (Ps_ext_j[j] < first_col_diag_P || Ps_ext_j[j] > last_col_diag_P)
{
P_ext_offd_j[cnt_offd] = Ps_ext_j[j];
P_ext_offd_data[cnt_offd++] = Ps_ext_data[j];
}
else
{
P_ext_diag_j[cnt_diag] = Ps_ext_j[j] - first_col_diag_P;
P_ext_diag_data[cnt_diag++] = Ps_ext_data[j];
}
}
if (num_procs > 1)
{
hypre_CSRMatrixDestroy(Ps_ext);
Ps_ext = NULL;
}
cnt = 0;
if (P_ext_offd_size || num_cols_offd_P)
{
temp = hypre_CTAlloc(HYPRE_Int, P_ext_offd_size+num_cols_offd_P);
for (i=0; i < P_ext_offd_size; i++)
temp[i] = P_ext_offd_j[i];
cnt = P_ext_offd_size;
for (i=0; i < num_cols_offd_P; i++)
temp[cnt++] = col_map_offd_P[i];
}
if (cnt)
{
qsort0(temp, 0, cnt-1);
num_cols_offd_C = 1;
value = temp[0];
for (i=1; i < cnt; i++)
{
if (temp[i] > value)
{
value = temp[i];
temp[num_cols_offd_C++] = value;
}
}
}
if (num_cols_offd_C)
col_map_offd_C = hypre_CTAlloc(HYPRE_Int,num_cols_offd_C);
for (i=0; i < num_cols_offd_C; i++)
col_map_offd_C[i] = temp[i];
if (P_ext_offd_size || num_cols_offd_P)
hypre_TFree(temp);
for (i=0 ; i < P_ext_offd_size; i++)
P_ext_offd_j[i] = hypre_BinarySearch(col_map_offd_C,
P_ext_offd_j[i],
num_cols_offd_C);
if (num_cols_offd_P)
{
map_P_to_C = hypre_CTAlloc(HYPRE_Int,num_cols_offd_P);
cnt = 0;
for (i=0; i < num_cols_offd_C; i++)
if (col_map_offd_C[i] == col_map_offd_P[cnt])
{
map_P_to_C[cnt++] = i;
if (cnt == num_cols_offd_P) break;
}
}
/*-----------------------------------------------------------------------
* Allocate marker array.
*-----------------------------------------------------------------------*/
P_marker = hypre_CTAlloc(HYPRE_Int, num_cols_diag_P+num_cols_offd_C);
/*-----------------------------------------------------------------------
* Initialize some stuff.
*-----------------------------------------------------------------------*/
for (i1 = 0; i1 < num_cols_diag_P+num_cols_offd_C; i1++)
{
P_marker[i1] = -1;
}
/* no changes for the marked version above this point */
/* This function call is the first pass: */
hypre_ParMatmul_RowSizes_Marked(
&C_diag_i, &C_offd_i, &P_marker,
A_diag_i, A_diag_j, A_offd_i, A_offd_j,
P_diag_i, P_diag_j, P_offd_i, P_offd_j,
P_ext_diag_i, P_ext_diag_j, P_ext_offd_i, P_ext_offd_j,
map_P_to_C,
&C_diag_size, &C_offd_size,
num_rows_diag_A, num_cols_offd_A, allsquare,
num_cols_diag_P, num_cols_offd_P,
num_cols_offd_C, CF_marker, dof_func, dof_func_offd
);
/* The above call of hypre_ParMatmul_RowSizes_Marked computed
two scalars: C_diag_size, C_offd_size,
and two arrays: C_diag_i, C_offd_i
( P_marker is also computed, but only used internally )
*/
/*-----------------------------------------------------------------------
* Allocate C_diag_data and C_diag_j arrays.
* Allocate C_offd_data and C_offd_j arrays.
*-----------------------------------------------------------------------*/
last_col_diag_P = first_col_diag_P + num_cols_diag_P - 1;
C_diag_data = hypre_CTAlloc(double, C_diag_size);
C_diag_j = hypre_CTAlloc(HYPRE_Int, C_diag_size);
if (C_offd_size)
{
C_offd_data = hypre_CTAlloc(double, C_offd_size);
C_offd_j = hypre_CTAlloc(HYPRE_Int, C_offd_size);
}
/*-----------------------------------------------------------------------
* Second Pass: Fill in C_diag_data and C_diag_j.
* Second Pass: Fill in C_offd_data and C_offd_j.
*-----------------------------------------------------------------------*/
/*-----------------------------------------------------------------------
* Initialize some stuff.
*-----------------------------------------------------------------------*/
jj_count_diag = start_indexing;
jj_count_offd = start_indexing;
for (i1 = 0; i1 < num_cols_diag_P+num_cols_offd_C; i1++)
{
P_marker[i1] = -1;
}
/*-----------------------------------------------------------------------
* Loop over interior c-points.
*-----------------------------------------------------------------------*/
for (i1 = 0; i1 < num_rows_diag_A; i1++)
{
if ( CF_marker[i1] < 0 ) /* i1 is a fine row */
/* ... This and the coarse row code are the only parts between first pass
and near the end where
hypre_ParMatmul_FC is different from the regular hypre_ParMatmul */
{
/*--------------------------------------------------------------------
* Create diagonal entry, C_{i1,i1}
*--------------------------------------------------------------------*/
jj_row_begin_diag = jj_count_diag;
jj_row_begin_offd = jj_count_offd;
/*-----------------------------------------------------------------
* Loop over entries in row i1 of A_offd.
*-----------------------------------------------------------------*/
if (num_cols_offd_A)
{
for (jj2 = A_offd_i[i1]; jj2 < A_offd_i[i1+1]; jj2++)
{
i2 = A_offd_j[jj2];
if( dof_func==NULL || dof_func[i1] == dof_func_offd[i2] )
{ /* interpolate only like "functions" */
a_entry = A_offd_data[jj2];
/*-----------------------------------------------------------
* Loop over entries in row i2 of P_ext.
*-----------------------------------------------------------*/
for (jj3 = P_ext_offd_i[i2]; jj3 < P_ext_offd_i[i2+1]; jj3++)
{
i3 = num_cols_diag_P+P_ext_offd_j[jj3];
a_b_product = a_entry * P_ext_offd_data[jj3];
/*--------------------------------------------------------
* Check P_marker to see that C_{i1,i3} has not already
* been accounted for. If it has not, create a new entry.
* If it has, add new contribution.
*--------------------------------------------------------*/
if (P_marker[i3] < jj_row_begin_offd)
{
P_marker[i3] = jj_count_offd;
C_offd_data[jj_count_offd] = a_b_product;
C_offd_j[jj_count_offd] = i3-num_cols_diag_P;
jj_count_offd++;
}
else
C_offd_data[P_marker[i3]] += a_b_product;
}
for (jj3 = P_ext_diag_i[i2]; jj3 < P_ext_diag_i[i2+1]; jj3++)
{
i3 = P_ext_diag_j[jj3];
a_b_product = a_entry * P_ext_diag_data[jj3];
if (P_marker[i3] < jj_row_begin_diag)
{
P_marker[i3] = jj_count_diag;
C_diag_data[jj_count_diag] = a_b_product;
C_diag_j[jj_count_diag] = i3;
jj_count_diag++;
}
else
C_diag_data[P_marker[i3]] += a_b_product;
}
}
else
{ /* Interpolation mat should be 0 where i1 and i2 correspond to
different "functions". As we haven't created an entry for
C(i1,i2), nothing needs to be done. */
}
}
}
/*-----------------------------------------------------------------
* Loop over entries in row i1 of A_diag.
*-----------------------------------------------------------------*/
for (jj2 = A_diag_i[i1]; jj2 < A_diag_i[i1+1]; jj2++)
{
i2 = A_diag_j[jj2];
if( dof_func==NULL || dof_func[i1] == dof_func[i2] )
{ /* interpolate only like "functions" */
a_entry = A_diag_data[jj2];
/*-----------------------------------------------------------
* Loop over entries in row i2 of P_diag.
*-----------------------------------------------------------*/
for (jj3 = P_diag_i[i2]; jj3 < P_diag_i[i2+1]; jj3++)
{
i3 = P_diag_j[jj3];
a_b_product = a_entry * P_diag_data[jj3];
/*--------------------------------------------------------
* Check P_marker to see that C_{i1,i3} has not already
* been accounted for. If it has not, create a new entry.
* If it has, add new contribution.
*--------------------------------------------------------*/
if (P_marker[i3] < jj_row_begin_diag)
{
P_marker[i3] = jj_count_diag;
C_diag_data[jj_count_diag] = a_b_product;
C_diag_j[jj_count_diag] = i3;
jj_count_diag++;
}
else
{
C_diag_data[P_marker[i3]] += a_b_product;
}
}
if (num_cols_offd_P)
{
for (jj3 = P_offd_i[i2]; jj3 < P_offd_i[i2+1]; jj3++)
{
i3 = num_cols_diag_P+map_P_to_C[P_offd_j[jj3]];
a_b_product = a_entry * P_offd_data[jj3];
/*--------------------------------------------------------
* Check P_marker to see that C_{i1,i3} has not already
* been accounted for. If it has not, create a new entry.
* If it has, add new contribution.
*--------------------------------------------------------*/
if (P_marker[i3] < jj_row_begin_offd)
{
P_marker[i3] = jj_count_offd;
C_offd_data[jj_count_offd] = a_b_product;
C_offd_j[jj_count_offd] = i3-num_cols_diag_P;
jj_count_offd++;
}
else
{
C_offd_data[P_marker[i3]] += a_b_product;
}
}
}
}
else
{ /* Interpolation mat should be 0 where i1 and i2 correspond to
different "functions". As we haven't created an entry for
C(i1,i2), nothing needs to be done. */
}
}
}
else /* i1 is a coarse row.*/
/* Copy P coarse-row values to C. This is useful if C is meant to
become a replacement for P */
{
if (num_cols_offd_P)
{
for (jj2 = P_offd_i[i1]; jj2 < P_offd_i[i1+1]; jj2++)
{
C_offd_j[jj_count_offd] = P_offd_j[jj_count_offd];
C_offd_data[jj_count_offd] = P_offd_data[jj_count_offd];
++jj_count_offd;
}
}
for (jj2 = P_diag_i[i1]; jj2 < P_diag_i[i1+1]; jj2++)
{
C_diag_j[jj_count_diag] = P_diag_j[jj2];
C_diag_data[jj_count_diag] = P_diag_data[jj2];
++jj_count_diag;
}
}
}
C = hypre_ParCSRMatrixCreate(
comm, n_rows_A_global, n_cols_P_global,
row_starts_A, col_starts_P, num_cols_offd_C, C_diag_size, C_offd_size );
/* Note that C does not own the partitionings */
hypre_ParCSRMatrixSetRowStartsOwner(C,0);
hypre_ParCSRMatrixSetColStartsOwner(C,0);
C_diag = hypre_ParCSRMatrixDiag(C);
hypre_CSRMatrixData(C_diag) = C_diag_data;
hypre_CSRMatrixI(C_diag) = C_diag_i;
hypre_CSRMatrixJ(C_diag) = C_diag_j;
C_offd = hypre_ParCSRMatrixOffd(C);
hypre_CSRMatrixI(C_offd) = C_offd_i;
hypre_ParCSRMatrixOffd(C) = C_offd;
if (num_cols_offd_C)
{
hypre_CSRMatrixData(C_offd) = C_offd_data;
hypre_CSRMatrixJ(C_offd) = C_offd_j;
hypre_ParCSRMatrixColMapOffd(C) = col_map_offd_C;
}
/*-----------------------------------------------------------------------
* Free various arrays
*-----------------------------------------------------------------------*/
hypre_TFree(P_marker);
hypre_TFree(P_ext_diag_i);
if (P_ext_diag_size)
{
hypre_TFree(P_ext_diag_j);
hypre_TFree(P_ext_diag_data);
}
hypre_TFree(P_ext_offd_i);
if (P_ext_offd_size)
{
hypre_TFree(P_ext_offd_j);
hypre_TFree(P_ext_offd_data);
}
if (num_cols_offd_P) hypre_TFree(map_P_to_C);
return C;
}
HYPRE_Int hypre_seqAMGSetup( hypre_ParAMGData *amg_data,
HYPRE_Int p_level,
HYPRE_Int coarse_threshold)
{
/* Par Data Structure variables */
hypre_ParCSRMatrix **Par_A_array = hypre_ParAMGDataAArray(amg_data);
MPI_Comm comm = hypre_ParCSRMatrixComm(Par_A_array[0]);
MPI_Comm new_comm, seq_comm;
hypre_ParCSRMatrix *A_seq = NULL;
hypre_CSRMatrix *A_seq_diag;
hypre_CSRMatrix *A_seq_offd;
hypre_ParVector *F_seq = NULL;
hypre_ParVector *U_seq = NULL;
hypre_ParCSRMatrix *A;
HYPRE_Int **dof_func_array;
HYPRE_Int num_procs, my_id;
HYPRE_Int not_finished_coarsening;
HYPRE_Int level;
HYPRE_Solver coarse_solver;
/* misc */
dof_func_array = hypre_ParAMGDataDofFuncArray(amg_data);
/*MPI Stuff */
hypre_MPI_Comm_size(comm, &num_procs);
hypre_MPI_Comm_rank(comm,&my_id);
/*initial */
level = p_level;
not_finished_coarsening = 1;
/* convert A at this level to sequential */
A = Par_A_array[level];
{
double *A_seq_data = NULL;
HYPRE_Int *A_seq_i = NULL;
HYPRE_Int *A_seq_offd_i = NULL;
HYPRE_Int *A_seq_j = NULL;
double *A_tmp_data = NULL;
HYPRE_Int *A_tmp_i = NULL;
HYPRE_Int *A_tmp_j = NULL;
HYPRE_Int *info, *displs, *displs2;
HYPRE_Int i, j, size, num_nonzeros, total_nnz, cnt;
hypre_CSRMatrix *A_diag = hypre_ParCSRMatrixDiag(A);
hypre_CSRMatrix *A_offd = hypre_ParCSRMatrixOffd(A);
HYPRE_Int *col_map_offd = hypre_ParCSRMatrixColMapOffd(A);
HYPRE_Int *A_diag_i = hypre_CSRMatrixI(A_diag);
HYPRE_Int *A_offd_i = hypre_CSRMatrixI(A_offd);
HYPRE_Int *A_diag_j = hypre_CSRMatrixJ(A_diag);
HYPRE_Int *A_offd_j = hypre_CSRMatrixJ(A_offd);
double *A_diag_data = hypre_CSRMatrixData(A_diag);
double *A_offd_data = hypre_CSRMatrixData(A_offd);
HYPRE_Int num_rows = hypre_CSRMatrixNumRows(A_diag);
HYPRE_Int first_row_index = hypre_ParCSRMatrixFirstRowIndex(A);
hypre_MPI_Group orig_group, new_group;
HYPRE_Int *ranks, new_num_procs, *row_starts;
info = hypre_CTAlloc(HYPRE_Int, num_procs);
hypre_MPI_Allgather(&num_rows, 1, HYPRE_MPI_INT, info, 1, HYPRE_MPI_INT, comm);
ranks = hypre_CTAlloc(HYPRE_Int, num_procs);
new_num_procs = 0;
for (i=0; i < num_procs; i++)
if (info[i])
{
ranks[new_num_procs] = i;
info[new_num_procs++] = info[i];
}
MPI_Comm_group(comm, &orig_group);
hypre_MPI_Group_incl(orig_group, new_num_procs, ranks, &new_group);
MPI_Comm_create(comm, new_group, &new_comm);
hypre_MPI_Group_free(&new_group);
hypre_MPI_Group_free(&orig_group);
if (num_rows)
{
/* alloc space in seq data structure only for participating procs*/
HYPRE_BoomerAMGCreate(&coarse_solver);
HYPRE_BoomerAMGSetMaxRowSum(coarse_solver,
hypre_ParAMGDataMaxRowSum(amg_data));
HYPRE_BoomerAMGSetStrongThreshold(coarse_solver,
hypre_ParAMGDataStrongThreshold(amg_data));
HYPRE_BoomerAMGSetCoarsenType(coarse_solver,
hypre_ParAMGDataCoarsenType(amg_data));
HYPRE_BoomerAMGSetInterpType(coarse_solver,
hypre_ParAMGDataInterpType(amg_data));
HYPRE_BoomerAMGSetTruncFactor(coarse_solver,
hypre_ParAMGDataTruncFactor(amg_data));
HYPRE_BoomerAMGSetPMaxElmts(coarse_solver,
hypre_ParAMGDataPMaxElmts(amg_data));
if (hypre_ParAMGDataUserRelaxType(amg_data) > -1)
HYPRE_BoomerAMGSetRelaxType(coarse_solver,
hypre_ParAMGDataUserRelaxType(amg_data));
HYPRE_BoomerAMGSetRelaxOrder(coarse_solver,
hypre_ParAMGDataRelaxOrder(amg_data));
HYPRE_BoomerAMGSetRelaxWt(coarse_solver,
hypre_ParAMGDataUserRelaxWeight(amg_data));
if (hypre_ParAMGDataUserNumSweeps(amg_data) > -1)
HYPRE_BoomerAMGSetNumSweeps(coarse_solver,
hypre_ParAMGDataUserNumSweeps(amg_data));
HYPRE_BoomerAMGSetNumFunctions(coarse_solver,
hypre_ParAMGDataNumFunctions(amg_data));
HYPRE_BoomerAMGSetMaxIter(coarse_solver, 1);
HYPRE_BoomerAMGSetTol(coarse_solver, 0);
/* Create CSR Matrix, will be Diag part of new matrix */
A_tmp_i = hypre_CTAlloc(HYPRE_Int, num_rows+1);
A_tmp_i[0] = 0;
for (i=1; i < num_rows+1; i++)
A_tmp_i[i] = A_diag_i[i]-A_diag_i[i-1]+A_offd_i[i]-A_offd_i[i-1];
num_nonzeros = A_offd_i[num_rows]+A_diag_i[num_rows];
A_tmp_j = hypre_CTAlloc(HYPRE_Int, num_nonzeros);
A_tmp_data = hypre_CTAlloc(double, num_nonzeros);
cnt = 0;
for (i=0; i < num_rows; i++)
{
for (j=A_diag_i[i]; j < A_diag_i[i+1]; j++)
{
A_tmp_j[cnt] = A_diag_j[j]+first_row_index;
A_tmp_data[cnt++] = A_diag_data[j];
}
for (j=A_offd_i[i]; j < A_offd_i[i+1]; j++)
{
A_tmp_j[cnt] = col_map_offd[A_offd_j[j]];
A_tmp_data[cnt++] = A_offd_data[j];
}
}
displs = hypre_CTAlloc(HYPRE_Int, new_num_procs+1);
displs[0] = 0;
for (i=1; i < new_num_procs+1; i++)
displs[i] = displs[i-1]+info[i-1];
size = displs[new_num_procs];
A_seq_i = hypre_CTAlloc(HYPRE_Int, size+1);
A_seq_offd_i = hypre_CTAlloc(HYPRE_Int, size+1);
hypre_MPI_Allgatherv ( &A_tmp_i[1], num_rows, HYPRE_MPI_INT, &A_seq_i[1], info,
displs, HYPRE_MPI_INT, new_comm );
displs2 = hypre_CTAlloc(HYPRE_Int, new_num_procs+1);
A_seq_i[0] = 0;
displs2[0] = 0;
for (j=1; j < displs[1]; j++)
A_seq_i[j] = A_seq_i[j]+A_seq_i[j-1];
for (i=1; i < new_num_procs; i++)
{
for (j=displs[i]; j < displs[i+1]; j++)
{
A_seq_i[j] = A_seq_i[j]+A_seq_i[j-1];
}
}
A_seq_i[size] = A_seq_i[size]+A_seq_i[size-1];
displs2[new_num_procs] = A_seq_i[size];
for (i=1; i < new_num_procs+1; i++)
{
displs2[i] = A_seq_i[displs[i]];
info[i-1] = displs2[i] - displs2[i-1];
}
total_nnz = displs2[new_num_procs];
A_seq_j = hypre_CTAlloc(HYPRE_Int, total_nnz);
A_seq_data = hypre_CTAlloc(double, total_nnz);
hypre_MPI_Allgatherv ( A_tmp_j, num_nonzeros, HYPRE_MPI_INT,
A_seq_j, info, displs2,
HYPRE_MPI_INT, new_comm );
hypre_MPI_Allgatherv ( A_tmp_data, num_nonzeros, hypre_MPI_DOUBLE,
A_seq_data, info, displs2,
hypre_MPI_DOUBLE, new_comm );
hypre_TFree(displs);
hypre_TFree(displs2);
hypre_TFree(A_tmp_i);
hypre_TFree(A_tmp_j);
hypre_TFree(A_tmp_data);
row_starts = hypre_CTAlloc(HYPRE_Int,2);
row_starts[0] = 0;
row_starts[1] = size;
/* Create 1 proc communicator */
seq_comm = hypre_MPI_COMM_SELF;
A_seq = hypre_ParCSRMatrixCreate(seq_comm,size,size,
row_starts, row_starts,
0,total_nnz,0);
A_seq_diag = hypre_ParCSRMatrixDiag(A_seq);
A_seq_offd = hypre_ParCSRMatrixOffd(A_seq);
hypre_CSRMatrixData(A_seq_diag) = A_seq_data;
hypre_CSRMatrixI(A_seq_diag) = A_seq_i;
hypre_CSRMatrixJ(A_seq_diag) = A_seq_j;
hypre_CSRMatrixI(A_seq_offd) = A_seq_offd_i;
F_seq = hypre_ParVectorCreate(seq_comm, size, row_starts);
U_seq = hypre_ParVectorCreate(seq_comm, size, row_starts);
hypre_ParVectorOwnsPartitioning(F_seq) = 0;
hypre_ParVectorOwnsPartitioning(U_seq) = 0;
hypre_ParVectorInitialize(F_seq);
hypre_ParVectorInitialize(U_seq);
hypre_BoomerAMGSetup(coarse_solver,A_seq,F_seq,U_seq);
hypre_ParAMGDataCoarseSolver(amg_data) = coarse_solver;
hypre_ParAMGDataACoarse(amg_data) = A_seq;
hypre_ParAMGDataFCoarse(amg_data) = F_seq;
hypre_ParAMGDataUCoarse(amg_data) = U_seq;
hypre_ParAMGDataNewComm(amg_data) = new_comm;
}
hypre_TFree(info);
hypre_TFree(ranks);
}
return 0;
}
hypre_ParCSRMatrix * hypre_ParMatMinus_F(
hypre_ParCSRMatrix * P, hypre_ParCSRMatrix * C, HYPRE_Int * CF_marker )
/* hypre_ParMatMinus_F subtracts selected rows of its second argument
from selected rows of its first argument. The marker array
determines which rows are affected - those for which CF_marker<0.
The result is returned as a new matrix.
*/
{
/*
If P=(Pik),C=(Cik), i in Fine+Coarse, k in Coarse, we want
new Pik = Pik - Cik, for Fine i only, all k.
This computation is purely local.
*/
/* This is _not_ a general-purpose matrix subtraction function.
This is written for an interpolation problem where it is known that C(i,k)
exists whenever P(i,k) does (because C=A*P where A has nonzero diagonal elements).
*/
hypre_ParCSRMatrix *Pnew;
hypre_CSRMatrix *P_diag = hypre_ParCSRMatrixDiag(P);
hypre_CSRMatrix *P_offd = hypre_ParCSRMatrixOffd(P);
hypre_CSRMatrix *C_diag = hypre_ParCSRMatrixDiag(C);
hypre_CSRMatrix *C_offd = hypre_ParCSRMatrixOffd(C);
hypre_CSRMatrix *Pnew_diag;
hypre_CSRMatrix *Pnew_offd;
double *P_diag_data = hypre_CSRMatrixData(P_diag);
HYPRE_Int *P_diag_i = hypre_CSRMatrixI(P_diag);
HYPRE_Int *P_diag_j = hypre_CSRMatrixJ(P_diag);
double *P_offd_data = hypre_CSRMatrixData(P_offd);
HYPRE_Int *P_offd_i = hypre_CSRMatrixI(P_offd);
HYPRE_Int *P_offd_j = hypre_CSRMatrixJ(P_offd);
HYPRE_Int *P_col_map_offd = hypre_ParCSRMatrixColMapOffd( P );
double *C_diag_data = hypre_CSRMatrixData(C_diag);
HYPRE_Int *C_diag_i = hypre_CSRMatrixI(C_diag);
HYPRE_Int *C_diag_j = hypre_CSRMatrixJ(C_diag);
double *C_offd_data = hypre_CSRMatrixData(C_offd);
HYPRE_Int *C_offd_i = hypre_CSRMatrixI(C_offd);
HYPRE_Int *C_offd_j = hypre_CSRMatrixJ(C_offd);
HYPRE_Int *C_col_map_offd = hypre_ParCSRMatrixColMapOffd( C );
HYPRE_Int *Pnew_diag_i;
HYPRE_Int *Pnew_diag_j;
double *Pnew_diag_data;
HYPRE_Int *Pnew_offd_i;
HYPRE_Int *Pnew_offd_j;
double *Pnew_offd_data;
HYPRE_Int *Pnew_j2m;
HYPRE_Int *Pnew_col_map_offd;
HYPRE_Int num_rows_diag_C = hypre_CSRMatrixNumRows(C_diag);
/* HYPRE_Int num_rows_offd_C = hypre_CSRMatrixNumRows(C_offd); */
HYPRE_Int num_cols_offd_C = hypre_CSRMatrixNumCols(C_offd);
HYPRE_Int num_cols_offd_P = hypre_CSRMatrixNumCols(P_offd);
HYPRE_Int num_cols_offd_Pnew, num_rows_offd_Pnew;
HYPRE_Int i1, jmin, jmax, jrange, jrangem1;
HYPRE_Int j, m, mc, mp, jc, jp, jP, jC, jg, jCg, jPg;
double dc, dp;
/* Pnew = hypre_ParCSRMatrixCompleteClone( C );*/
Pnew = hypre_ParCSRMatrixUnion( C, P );
;
hypre_ParCSRMatrixZero_F( Pnew, CF_marker ); /* fine rows of Pnew set to 0 */
hypre_ParCSRMatrixCopy_C( Pnew, C, CF_marker ); /* coarse rows of Pnew copied from C (or P) */
/* ...Zero_F may not be needed depending on how Pnew is made */
Pnew_diag = hypre_ParCSRMatrixDiag(Pnew);
Pnew_offd = hypre_ParCSRMatrixOffd(Pnew);
Pnew_diag_i = hypre_CSRMatrixI(Pnew_diag);
Pnew_diag_j = hypre_CSRMatrixJ(Pnew_diag);
Pnew_offd_i = hypre_CSRMatrixI(Pnew_offd);
Pnew_offd_j = hypre_CSRMatrixJ(Pnew_offd);
Pnew_diag_data = hypre_CSRMatrixData(Pnew_diag);
Pnew_offd_data = hypre_CSRMatrixData(Pnew_offd);
Pnew_col_map_offd = hypre_ParCSRMatrixColMapOffd( Pnew );
num_rows_offd_Pnew = hypre_CSRMatrixNumRows(Pnew_offd);
num_cols_offd_Pnew = hypre_CSRMatrixNumCols(Pnew_offd);
/* Find the j-ranges, needed to allocate a "reverse lookup" array. */
/* This is the max j - min j over P and Pnew (which here is a copy of C).
Only the diag block is considered. */
/* For scalability reasons (jrange can get big) this won't work for the offd block.
Also, indexing is more complicated in the offd block (c.f. col_map_offd).
It's not clear, though whether the "quadratic" algorithm I'm using for the offd
block is really any slower than the more complicated "linear" algorithm here. */
jrange = 0;
jrangem1=-1;
for ( i1 = 0; i1 < num_rows_diag_C; i1++ )
{
if ( CF_marker[i1]<0 && hypre_CSRMatrixNumNonzeros(Pnew_diag)>0 ) /* only Fine rows matter */
{
jmin = Pnew_diag_j[ Pnew_diag_i[i1] ];
jmax = Pnew_diag_j[ Pnew_diag_i[i1+1]-1 ];
jrangem1 = jmax-jmin;
jrange = hypre_max(jrange,jrangem1+1);
/* If columns (of a given row) were in increasing order, the above would be sufficient.
If not, the following would be necessary (and sufficient) */
jmin = Pnew_diag_j[ Pnew_diag_i[i1] ];
jmax = Pnew_diag_j[ Pnew_diag_i[i1] ];
for ( m=Pnew_diag_i[i1]+1; m<Pnew_diag_i[i1+1]; ++m )
{
j = Pnew_diag_j[m];
jmin = hypre_min( jmin, j );
jmax = hypre_max( jmax, j );
}
for ( m=P_diag_i[i1]; m<P_diag_i[i1+1]; ++m )
{
j = P_diag_j[m];
jmin = hypre_min( jmin, j );
jmax = hypre_max( jmax, j );
}
jrangem1 = jmax-jmin;
jrange = hypre_max(jrange,jrangem1+1);
}
}
/*-----------------------------------------------------------------------
* Loop over Pnew_diag rows. Construct a temporary reverse array:
* If j is a column number, Pnew_j2m[j] is the array index for j, i.e.
* Pnew_diag_j[ Pnew_j2m[j] ] = j
*-----------------------------------------------------------------------*/
Pnew_j2m = hypre_CTAlloc( HYPRE_Int, jrange );
for ( i1 = 0; i1 < num_rows_diag_C; i1++ )
{
if ( CF_marker[i1]<0 && hypre_CSRMatrixNumNonzeros(Pnew_diag)>0 ) /* Fine data only */
{
/* just needed for an assertion below... */
for ( j=0; j<jrange; ++j ) Pnew_j2m[j] = -1;
jmin = Pnew_diag_j[ Pnew_diag_i[i1] ];
/* If columns (of a given row) were in increasing order, the above line would be sufficient.
If not, the following loop would have to be added (or store the jmin computed above )*/
for ( m=Pnew_diag_i[i1]+1; m<Pnew_diag_i[i1+1]; ++m )
{
j = Pnew_diag_j[m];
jmin = hypre_min( jmin, j );
}
for ( m=P_diag_i[i1]; m<P_diag_i[i1+1]; ++m )
{
j = P_diag_j[m];
jmin = hypre_min( jmin, j );
}
for ( m = Pnew_diag_i[i1]; m<Pnew_diag_i[i1+1]; ++m )
{
j = Pnew_diag_j[m];
hypre_assert( j-jmin>=0 );
hypre_assert( j-jmin<jrange );
Pnew_j2m[ j-jmin ] = m;
}
/*-----------------------------------------------------------------------
* Loop over C_diag data for the current row.
* Subtract each C data entry from the corresponding Pnew entry.
*-----------------------------------------------------------------------*/
for ( mc=C_diag_i[i1]; mc<C_diag_i[i1+1]; ++mc )
{
jc = C_diag_j[mc];
dc = C_diag_data[mc];
m = Pnew_j2m[jc-jmin];
hypre_assert( m>=0 );
Pnew_diag_data[m] -= dc;
}
/*-----------------------------------------------------------------------
* Loop over P_diag data for the current row.
* Add each P data entry from the corresponding Pnew entry.
*-----------------------------------------------------------------------*/
for ( mp=P_diag_i[i1]; mp<P_diag_i[i1+1]; ++mp )
{
jp = P_diag_j[mp];
dp = P_diag_data[mp];
m = Pnew_j2m[jp-jmin];
hypre_assert( m>=0 );
Pnew_diag_data[m] += dp;
}
}
}
/*-----------------------------------------------------------------------
* Repeat for the offd block.
*-----------------------------------------------------------------------*/
for ( i1 = 0; i1 < num_rows_offd_Pnew; i1++ )
{
if ( CF_marker[i1]<0 && hypre_CSRMatrixNumNonzeros(Pnew_offd)>0 ) /* Fine data only */
{
if ( num_cols_offd_Pnew )
{
/* This is a simple quadratic algorithm. If necessary I may try
to implement the ideas used on the diag block later. */
for ( m = Pnew_offd_i[i1]; m<Pnew_offd_i[i1+1]; ++m )
{
j = Pnew_offd_j[m];
jg = Pnew_col_map_offd[j];
Pnew_offd_data[m] = 0;
if ( num_cols_offd_C )
for ( mc=C_offd_i[i1]; mc<C_offd_i[i1+1]; ++mc )
{
jC = C_offd_j[mc];
jCg = C_col_map_offd[jC];
if ( jCg==jg ) Pnew_offd_data[m] -= C_offd_data[mc];
}
if ( num_cols_offd_P )
for ( mp=P_offd_i[i1]; mp<P_offd_i[i1+1]; ++mp )
{
jP = P_offd_j[mp];
jPg = P_col_map_offd[jP];
if ( jPg==jg ) Pnew_offd_data[m] += P_offd_data[mp];
}
}
}
}
}
hypre_TFree(Pnew_j2m);
return Pnew;
}
int
hypre_BoomerAMGSetupStats( void *amg_vdata,
hypre_ParCSRMatrix *A )
{
MPI_Comm comm = hypre_ParCSRMatrixComm(A);
hypre_ParAMGData *amg_data = (hypre_ParAMGData*)amg_vdata;
/*hypre_SeqAMGData *seq_data = hypre_ParAMGDataSeqData(amg_data);*/
/* Data Structure variables */
hypre_ParCSRMatrix **A_array;
hypre_ParCSRMatrix **P_array;
hypre_CSRMatrix *A_diag;
double *A_diag_data;
int *A_diag_i;
hypre_CSRMatrix *A_offd;
double *A_offd_data;
int *A_offd_i;
hypre_CSRMatrix *P_diag;
double *P_diag_data;
int *P_diag_i;
hypre_CSRMatrix *P_offd;
double *P_offd_data;
int *P_offd_i;
int numrows;
HYPRE_BigInt *row_starts;
int num_levels;
int coarsen_type;
int interp_type;
int measure_type;
double global_nonzeros;
double *send_buff;
double *gather_buff;
/* Local variables */
int level;
int j;
HYPRE_BigInt fine_size;
int min_entries;
int max_entries;
int num_procs,my_id, num_threads;
double min_rowsum;
double max_rowsum;
double sparse;
int i;
HYPRE_BigInt coarse_size;
int entries;
double avg_entries;
double rowsum;
double min_weight;
double max_weight;
int global_min_e;
int global_max_e;
double global_min_rsum;
double global_max_rsum;
double global_min_wt;
double global_max_wt;
double *num_coeffs;
double *num_variables;
double total_variables;
double operat_cmplxty;
double grid_cmplxty;
/* amg solve params */
int max_iter;
int cycle_type;
int *num_grid_sweeps;
int *grid_relax_type;
int relax_order;
int **grid_relax_points;
double *relax_weight;
double *omega;
double tol;
int one = 1;
int minus_one = -1;
int zero = 0;
int smooth_type;
int smooth_num_levels;
int agg_num_levels;
/*int seq_cg = 0;*/
/*if (seq_data)
seq_cg = 1;*/
MPI_Comm_size(comm, &num_procs);
MPI_Comm_rank(comm,&my_id);
num_threads = hypre_NumThreads();
if (my_id == 0)
printf("\nNumber of MPI processes: %d , Number of OpenMP threads: %d\n", num_procs, num_threads);
A_array = hypre_ParAMGDataAArray(amg_data);
P_array = hypre_ParAMGDataPArray(amg_data);
num_levels = hypre_ParAMGDataNumLevels(amg_data);
coarsen_type = hypre_ParAMGDataCoarsenType(amg_data);
interp_type = hypre_ParAMGDataInterpType(amg_data);
measure_type = hypre_ParAMGDataMeasureType(amg_data);
smooth_type = hypre_ParAMGDataSmoothType(amg_data);
smooth_num_levels = hypre_ParAMGDataSmoothNumLevels(amg_data);
agg_num_levels = hypre_ParAMGDataAggNumLevels(amg_data);
/*----------------------------------------------------------
* Get the amg_data data
*----------------------------------------------------------*/
num_levels = hypre_ParAMGDataNumLevels(amg_data);
max_iter = hypre_ParAMGDataMaxIter(amg_data);
cycle_type = hypre_ParAMGDataCycleType(amg_data);
num_grid_sweeps = hypre_ParAMGDataNumGridSweeps(amg_data);
grid_relax_type = hypre_ParAMGDataGridRelaxType(amg_data);
grid_relax_points = hypre_ParAMGDataGridRelaxPoints(amg_data);
relax_weight = hypre_ParAMGDataRelaxWeight(amg_data);
relax_order = hypre_ParAMGDataRelaxOrder(amg_data);
omega = hypre_ParAMGDataOmega(amg_data);
tol = hypre_ParAMGDataTol(amg_data);
/*block_mode = hypre_ParAMGDataBlockMode(amg_data);*/
send_buff = hypre_CTAlloc(double, 6);
#ifdef HYPRE_NO_GLOBAL_PARTITION
gather_buff = hypre_CTAlloc(double,6);
#else
gather_buff = hypre_CTAlloc(double,6*num_procs);
#endif
if (my_id==0)
{
printf("\nBoomerAMG SETUP PARAMETERS:\n\n");
printf(" Max levels = %d\n",hypre_ParAMGDataMaxLevels(amg_data));
printf(" Num levels = %d\n\n",num_levels);
printf(" Strength Threshold = %f\n",
hypre_ParAMGDataStrongThreshold(amg_data));
printf(" Interpolation Truncation Factor = %f\n",
hypre_ParAMGDataTruncFactor(amg_data));
printf(" Maximum Row Sum Threshold for Dependency Weakening = %f\n\n",
hypre_ParAMGDataMaxRowSum(amg_data));
if (coarsen_type == 0)
{
printf(" Coarsening Type = Cleary-Luby-Jones-Plassman\n");
}
else if (abs(coarsen_type) == 1)
{
printf(" Coarsening Type = Ruge\n");
}
else if (abs(coarsen_type) == 2)
{
printf(" Coarsening Type = Ruge2B\n");
}
else if (abs(coarsen_type) == 3)
{
printf(" Coarsening Type = Ruge3\n");
}
else if (abs(coarsen_type) == 4)
{
printf(" Coarsening Type = Ruge 3c \n");
}
else if (abs(coarsen_type) == 5)
{
printf(" Coarsening Type = Ruge relax special points \n");
}
else if (abs(coarsen_type) == 6)
{
printf(" Coarsening Type = Falgout-CLJP \n");
}
else if (abs(coarsen_type) == 8)
{
printf(" Coarsening Type = PMIS \n");
}
else if (abs(coarsen_type) == 10)
{
printf(" Coarsening Type = HMIS \n");
}
else if (abs(coarsen_type) == 11)
{
printf(" Coarsening Type = Ruge 1st pass only \n");
}
else if (abs(coarsen_type) == 9)
{
printf(" Coarsening Type = PMIS fixed random \n");
}
else if (abs(coarsen_type) == 7)
{
printf(" Coarsening Type = CLJP, fixed random \n");
}
if (coarsen_type > 0)
{
printf(" Hybrid Coarsening (switch to CLJP when coarsening slows)\n");
}
if (coarsen_type)
printf(" measures are determined %s\n\n",
(measure_type ? "globally" : "locally"));
if (agg_num_levels)
printf(" no. of levels of aggressive coarsening: %d\n\n", agg_num_levels);
#ifdef HYPRE_NO_GLOBAL_PARTITION
printf( "\n No global partition option chosen.\n\n");
#endif
if (interp_type == 0)
{
printf(" Interpolation = modified classical interpolation\n");
}
else if (interp_type == 1)
{
printf(" Interpolation = LS interpolation \n");
}
else if (interp_type == 2)
{
printf(" Interpolation = modified classical interpolation for hyperbolic PDEs\n");
}
else if (interp_type == 3)
{
printf(" Interpolation = direct interpolation with separation of weights\n");
}
else if (interp_type == 4)
{
printf(" Interpolation = multipass interpolation\n");
}
else if (interp_type == 5)
{
printf(" Interpolation = multipass interpolation with separation of weights\n");
}
else if (interp_type == 6)
{
printf(" Interpolation = extended+i interpolation\n");
}
else if (interp_type == 7)
{
printf(" Interpolation = extended+i interpolation (only when needed)\n");
}
else if (interp_type == 8)
{
printf(" Interpolation = standard interpolation\n");
}
else if (interp_type == 9)
{
printf(" Interpolation = standard interpolation with separation of weights\n");
}
else if (interp_type == 12)
{
printf(" FF interpolation \n");
}
else if (interp_type == 13)
{
printf(" FF1 interpolation \n");
}
{
printf( "\nOperator Matrix Information:\n\n");
}
#if HYPRE_LONG_LONG
printf(" nonzero entries p");
printf("er row row sums\n");
printf("lev rows entries sparse min max ");
printf("avg min max\n");
printf("=======================================");
printf("==================================\n");
#else
printf(" nonzero entries p");
printf("er row row sums\n");
printf("lev rows entries sparse min max ");
printf("avg min max\n");
printf("=======================================");
printf("============================\n");
#endif
}
/*-----------------------------------------------------
* Enter Statistics Loop
*-----------------------------------------------------*/
num_coeffs = hypre_CTAlloc(double,num_levels);
num_variables = hypre_CTAlloc(double,num_levels);
for (level = 0; level < num_levels; level++)
{
{
A_diag = hypre_ParCSRMatrixDiag(A_array[level]);
A_diag_data = hypre_CSRMatrixData(A_diag);
A_diag_i = hypre_CSRMatrixI(A_diag);
A_offd = hypre_ParCSRMatrixOffd(A_array[level]);
A_offd_data = hypre_CSRMatrixData(A_offd);
A_offd_i = hypre_CSRMatrixI(A_offd);
row_starts = hypre_ParCSRMatrixRowStarts(A_array[level]);
fine_size = hypre_ParCSRMatrixGlobalNumRows(A_array[level]);
global_nonzeros = hypre_ParCSRMatrixDNumNonzeros(A_array[level]);
num_coeffs[level] = global_nonzeros;
num_variables[level] = (double) fine_size;
sparse = global_nonzeros /((double) fine_size * (double) fine_size);
min_entries = 0;
max_entries = 0;
min_rowsum = 0.0;
max_rowsum = 0.0;
if (hypre_CSRMatrixNumRows(A_diag))
{
min_entries = (A_diag_i[1]-A_diag_i[0])+(A_offd_i[1]-A_offd_i[0]);
for (j = A_diag_i[0]; j < A_diag_i[1]; j++)
min_rowsum += A_diag_data[j];
for (j = A_offd_i[0]; j < A_offd_i[1]; j++)
min_rowsum += A_offd_data[j];
max_rowsum = min_rowsum;
for (j = 0; j < hypre_CSRMatrixNumRows(A_diag); j++)
{
entries = (A_diag_i[j+1]-A_diag_i[j])+(A_offd_i[j+1]-A_offd_i[j]);
min_entries = hypre_min(entries, min_entries);
max_entries = hypre_max(entries, max_entries);
rowsum = 0.0;
for (i = A_diag_i[j]; i < A_diag_i[j+1]; i++)
rowsum += A_diag_data[i];
for (i = A_offd_i[j]; i < A_offd_i[j+1]; i++)
rowsum += A_offd_data[i];
min_rowsum = hypre_min(rowsum, min_rowsum);
max_rowsum = hypre_max(rowsum, max_rowsum);
}
}
avg_entries = global_nonzeros / ((double) fine_size);
}
#ifdef HYPRE_NO_GLOBAL_PARTITION
numrows = (int)(row_starts[1]-row_starts[0]);
if (!numrows) /* if we don't have any rows, then don't have this count toward
min row sum or min num entries */
{
min_entries = 1000000;
min_rowsum = 1.0e7;
}
send_buff[0] = - (double) min_entries;
send_buff[1] = (double) max_entries;
send_buff[2] = - min_rowsum;
send_buff[3] = max_rowsum;
MPI_Reduce(send_buff, gather_buff, 4, MPI_DOUBLE, MPI_MAX, 0, comm);
if (my_id ==0)
{
global_min_e = - gather_buff[0];
global_max_e = gather_buff[1];
global_min_rsum = - gather_buff[2];
global_max_rsum = gather_buff[3];
#ifdef HYPRE_LONG_LONG
printf( "%2d %12lld %8.0f %0.3f %4d %4d",
level, fine_size, global_nonzeros, sparse, global_min_e,
global_max_e);
#else
printf( "%2d %7d %8.0f %0.3f %4d %4d",
level, fine_size, global_nonzeros, sparse, global_min_e,
global_max_e);
#endif
printf(" %4.1f %10.3e %10.3e\n", avg_entries,
global_min_rsum, global_max_rsum);
}
#else
send_buff[0] = (double) min_entries;
send_buff[1] = (double) max_entries;
send_buff[2] = min_rowsum;
send_buff[3] = max_rowsum;
MPI_Gather(send_buff,4,MPI_DOUBLE,gather_buff,4,MPI_DOUBLE,0,comm);
if (my_id == 0)
{
global_min_e = 1000000;
global_max_e = 0;
global_min_rsum = 1.0e7;
global_max_rsum = 0.0;
for (j = 0; j < num_procs; j++)
{
numrows = row_starts[j+1]-row_starts[j];
if (numrows)
{
global_min_e = hypre_min(global_min_e, (int) gather_buff[j*4]);
global_min_rsum = hypre_min(global_min_rsum, gather_buff[j*4 +2]);
}
global_max_e = hypre_max(global_max_e, (int) gather_buff[j*4 +1]);
global_max_rsum = hypre_max(global_max_rsum, gather_buff[j*4 +3]);
}
#ifdef HYPRE_LONG_LONG
printf( "%2d %12lld %8.0f %0.3f %4d %4d",
level, fine_size, global_nonzeros, sparse, global_min_e,
global_max_e);
#else
printf( "%2d %7d %8.0f %0.3f %4d %4d",
level, fine_size, global_nonzeros, sparse, global_min_e,
global_max_e);
#endif
printf(" %4.1f %10.3e %10.3e\n", avg_entries,
global_min_rsum, global_max_rsum);
}
#endif
}
if (my_id == 0)
{
{
printf( "\n\nInterpolation Matrix Information:\n\n");
}
#if HYPRE_LONG_LONG
printf(" entries/row min max");
printf(" row sums\n");
printf("lev rows x cols min max ");
printf(" weight weight min max \n");
printf("=======================================");
printf("======================================\n");
#else
printf(" entries/row min max");
printf(" row sums\n");
printf("lev rows cols min max ");
printf(" weight weight min max \n");
printf("=======================================");
printf("==========================\n");
#endif
}
/*-----------------------------------------------------
* Enter Statistics Loop
*-----------------------------------------------------*/
for (level = 0; level < num_levels-1; level++)
{
{
P_diag = hypre_ParCSRMatrixDiag(P_array[level]);
P_diag_data = hypre_CSRMatrixData(P_diag);
P_diag_i = hypre_CSRMatrixI(P_diag);
P_offd = hypre_ParCSRMatrixOffd(P_array[level]);
P_offd_data = hypre_CSRMatrixData(P_offd);
P_offd_i = hypre_CSRMatrixI(P_offd);
row_starts = hypre_ParCSRMatrixRowStarts(P_array[level]);
fine_size = hypre_ParCSRMatrixGlobalNumRows(P_array[level]);
coarse_size = hypre_ParCSRMatrixGlobalNumCols(P_array[level]);
global_nonzeros = hypre_ParCSRMatrixNumNonzeros(P_array[level]);
min_weight = 1.0;
max_weight = 0.0;
max_rowsum = 0.0;
min_rowsum = 0.0;
min_entries = 0;
max_entries = 0;
if (hypre_CSRMatrixNumRows(P_diag))
{
if (hypre_CSRMatrixNumCols(P_diag)) min_weight = P_diag_data[0];
for (j = P_diag_i[0]; j < P_diag_i[1]; j++)
{
min_weight = hypre_min(min_weight, P_diag_data[j]);
if (P_diag_data[j] != 1.0)
max_weight = hypre_max(max_weight, P_diag_data[j]);
min_rowsum += P_diag_data[j];
}
for (j = P_offd_i[0]; j < P_offd_i[1]; j++)
{
min_weight = hypre_min(min_weight, P_offd_data[j]);
if (P_offd_data[j] != 1.0)
max_weight = hypre_max(max_weight, P_offd_data[j]);
min_rowsum += P_offd_data[j];
}
max_rowsum = min_rowsum;
min_entries = (P_diag_i[1]-P_diag_i[0])+(P_offd_i[1]-P_offd_i[0]);
max_entries = 0;
for (j = 0; j < hypre_CSRMatrixNumRows(P_diag); j++)
{
entries = (P_diag_i[j+1]-P_diag_i[j])+(P_offd_i[j+1]-P_offd_i[j]);
min_entries = hypre_min(entries, min_entries);
max_entries = hypre_max(entries, max_entries);
rowsum = 0.0;
for (i = P_diag_i[j]; i < P_diag_i[j+1]; i++)
{
min_weight = hypre_min(min_weight, P_diag_data[i]);
if (P_diag_data[i] != 1.0)
max_weight = hypre_max(max_weight, P_diag_data[i]);
rowsum += P_diag_data[i];
}
for (i = P_offd_i[j]; i < P_offd_i[j+1]; i++)
{
min_weight = hypre_min(min_weight, P_offd_data[i]);
if (P_offd_data[i] != 1.0)
max_weight = hypre_max(max_weight, P_offd_data[i]);
rowsum += P_offd_data[i];
}
min_rowsum = hypre_min(rowsum, min_rowsum);
max_rowsum = hypre_max(rowsum, max_rowsum);
}
}
avg_entries = ((double) global_nonzeros) / ((double) fine_size);
}
#ifdef HYPRE_NO_GLOBAL_PARTITION
numrows = (int)(row_starts[1]-row_starts[0]);
if (!numrows) /* if we don't have any rows, then don't have this count toward
min row sum or min num entries */
{
min_entries = 1000000;
min_rowsum = 1.0e7;
min_weight = 1.0e7;
}
send_buff[0] = - (double) min_entries;
send_buff[1] = (double) max_entries;
send_buff[2] = - min_rowsum;
send_buff[3] = max_rowsum;
send_buff[4] = - min_weight;
send_buff[5] = max_weight;
MPI_Reduce(send_buff, gather_buff, 6, MPI_DOUBLE, MPI_MAX, 0, comm);
if (my_id == 0)
{
global_min_e = - gather_buff[0];
global_max_e = gather_buff[1];
global_min_rsum = -gather_buff[2];
global_max_rsum = gather_buff[3];
global_min_wt = -gather_buff[4];
global_max_wt = gather_buff[5];
#ifdef HYPRE_LONG_LONG
printf( "%2d %12lld x %-12lld %3d %3d",
level, fine_size, coarse_size, global_min_e, global_max_e);
#else
printf( "%2d %5d x %-5d %3d %3d",
level, fine_size, coarse_size, global_min_e, global_max_e);
#endif
printf(" %10.3e %9.3e %9.3e %9.3e\n",
global_min_wt, global_max_wt,
global_min_rsum, global_max_rsum);
}
#else
send_buff[0] = (double) min_entries;
send_buff[1] = (double) max_entries;
send_buff[2] = min_rowsum;
send_buff[3] = max_rowsum;
send_buff[4] = min_weight;
send_buff[5] = max_weight;
MPI_Gather(send_buff,6,MPI_DOUBLE,gather_buff,6,MPI_DOUBLE,0,comm);
if (my_id == 0)
{
global_min_e = 1000000;
global_max_e = 0;
global_min_rsum = 1.0e7;
global_max_rsum = 0.0;
global_min_wt = 1.0e7;
global_max_wt = 0.0;
for (j = 0; j < num_procs; j++)
{
numrows = row_starts[j+1] - row_starts[j];
if (numrows)
{
global_min_e = hypre_min(global_min_e, (int) gather_buff[j*6]);
global_min_rsum = hypre_min(global_min_rsum, gather_buff[j*6+2]);
global_min_wt = hypre_min(global_min_wt, gather_buff[j*6+4]);
}
global_max_e = hypre_max(global_max_e, (int) gather_buff[j*6+1]);
global_max_rsum = hypre_max(global_max_rsum, gather_buff[j*6+3]);
global_max_wt = hypre_max(global_max_wt, gather_buff[j*6+5]);
}
#ifdef HYPRE_LONG_LONG
printf( "%2d %12lld x %-12lld %3d %3d",
level, fine_size, coarse_size, global_min_e, global_max_e);
#else
printf( "%2d %5d x %-5d %3d %3d",
level, fine_size, coarse_size, global_min_e, global_max_e);
#endif
printf(" %10.3e %9.3e %9.3e %9.3e\n",
global_min_wt, global_max_wt,
global_min_rsum, global_max_rsum);
}
#endif
}
total_variables = 0;
operat_cmplxty = 0;
for (j=0;j<hypre_ParAMGDataNumLevels(amg_data);j++)
{
operat_cmplxty += num_coeffs[j] / num_coeffs[0];
total_variables += num_variables[j];
}
if (num_variables[0] != 0)
grid_cmplxty = total_variables / num_variables[0];
if (my_id == 0 )
{
printf("\n\n Complexity: grid = %f\n",grid_cmplxty);
printf(" operator = %f\n",operat_cmplxty);
}
if (my_id == 0) printf("\n\n");
if (my_id == 0)
{
printf("\n\nBoomerAMG SOLVER PARAMETERS:\n\n");
printf( " Maximum number of cycles: %d \n",max_iter);
printf( " Stopping Tolerance: %e \n",tol);
printf( " Cycle type (1 = V, 2 = W, etc.): %d\n\n", cycle_type);
printf( " Relaxation Parameters:\n");
printf( " Visiting Grid: down up coarse\n");
printf( " Number of partial sweeps: %4d %2d %4d \n",
num_grid_sweeps[1],
num_grid_sweeps[2],num_grid_sweeps[3]);
printf( " Type 0=Jac, 3=hGS, 6=hSGS, 9=GE: %4d %2d %4d \n",
grid_relax_type[1],
grid_relax_type[2],grid_relax_type[3]);
#if 1 /* TO DO: may not want this to print if CG in the coarse grid */
printf( " Point types, partial sweeps (1=C, -1=F):\n");
if (grid_relax_points)
{
printf( " Pre-CG relaxation (down):");
for (j = 0; j < num_grid_sweeps[1]; j++)
printf(" %2d", grid_relax_points[1][j]);
printf( "\n");
printf( " Post-CG relaxation (up):");
for (j = 0; j < num_grid_sweeps[2]; j++)
printf(" %2d", grid_relax_points[2][j]);
printf( "\n");
printf( " Coarsest grid:");
for (j = 0; j < num_grid_sweeps[3]; j++)
printf(" %2d", grid_relax_points[3][j]);
printf( "\n\n");
}
else if (relax_order == 1)
{
printf( " Pre-CG relaxation (down):");
for (j = 0; j < num_grid_sweeps[1]; j++)
printf(" %2d %2d", one, minus_one);
printf( "\n");
printf( " Post-CG relaxation (up):");
for (j = 0; j < num_grid_sweeps[2]; j++)
printf(" %2d %2d", minus_one, one);
printf( "\n");
printf( " Coarsest grid:");
for (j = 0; j < num_grid_sweeps[3]; j++)
printf(" %2d", zero);
printf( "\n\n");
}
else
{
printf( " Pre-CG relaxation (down):");
for (j = 0; j < num_grid_sweeps[1]; j++)
printf(" %2d", zero);
printf( "\n");
printf( " Post-CG relaxation (up):");
for (j = 0; j < num_grid_sweeps[2]; j++)
printf(" %2d", zero);
printf( "\n");
printf( " Coarsest grid:");
for (j = 0; j < num_grid_sweeps[3]; j++)
printf(" %2d", zero);
printf( "\n\n");
}
#endif
if (smooth_type == 6)
for (j=0; j < smooth_num_levels; j++)
printf( " Schwarz Relaxation Weight %f level %d\n",
hypre_ParAMGDataSchwarzRlxWeight(amg_data),j);
for (j=0; j < num_levels; j++)
if (relax_weight[j] != 1)
printf( " Relaxation Weight %f level %d\n",relax_weight[j],j);
for (j=0; j < num_levels; j++)
if (omega[j] != 1)
printf( " Outer relaxation weight %f level %d\n",omega[j],j);
}
/*if (seq_cg)
{
hypre_seqAMGSetupStats(amg_data,num_coeffs[0],num_variables[0],
operat_cmplxty, grid_cmplxty );
}*/
hypre_TFree(num_coeffs);
hypre_TFree(num_variables);
hypre_TFree(send_buff);
hypre_TFree(gather_buff);
return(0);
}
HYPRE_ParCSRMatrix
GenerateRotate7pt( MPI_Comm comm,
HYPRE_Int nx,
HYPRE_Int ny,
HYPRE_Int P,
HYPRE_Int Q,
HYPRE_Int p,
HYPRE_Int q,
HYPRE_Real alpha,
HYPRE_Real eps )
{
hypre_ParCSRMatrix *A;
hypre_CSRMatrix *diag;
hypre_CSRMatrix *offd;
HYPRE_Int *diag_i;
HYPRE_Int *diag_j;
HYPRE_Real *diag_data;
HYPRE_Int *offd_i;
HYPRE_Int *offd_j;
HYPRE_Real *offd_data;
HYPRE_Real *value;
HYPRE_Real ac, bc, cc, s, c, pi, x;
HYPRE_Int *global_part;
HYPRE_Int ix, iy;
HYPRE_Int cnt, o_cnt;
HYPRE_Int local_num_rows;
HYPRE_Int *col_map_offd;
HYPRE_Int *work;
HYPRE_Int row_index;
HYPRE_Int i,j;
HYPRE_Int nx_local, ny_local;
HYPRE_Int nx_size, ny_size;
HYPRE_Int num_cols_offd;
HYPRE_Int grid_size;
HYPRE_Int *nx_part;
HYPRE_Int *ny_part;
HYPRE_Int num_procs, my_id;
HYPRE_Int P_busy, Q_busy;
hypre_MPI_Comm_size(comm,&num_procs);
hypre_MPI_Comm_rank(comm,&my_id);
grid_size = nx*ny;
value = hypre_CTAlloc(HYPRE_Real,4);
pi = 4.0*atan(1.0);
x = pi*alpha/180.0;
s = sin(x);
c = cos(x);
ac = -(c*c + eps*s*s);
bc = 2.0*(1.0 - eps)*s*c;
cc = -(s*s + eps*c*c);
value[0] = -2*(2*ac+bc+2*cc);
value[1] = 2*ac+bc;
value[2] = bc+2*cc;
value[3] = -bc;
hypre_GeneratePartitioning(nx,P,&nx_part);
hypre_GeneratePartitioning(ny,Q,&ny_part);
global_part = hypre_CTAlloc(HYPRE_Int,P*Q+1);
global_part[0] = 0;
cnt = 1;
for (iy = 0; iy < Q; iy++)
{
ny_size = ny_part[iy+1]-ny_part[iy];
for (ix = 0; ix < P; ix++)
{
nx_size = nx_part[ix+1] - nx_part[ix];
global_part[cnt] = global_part[cnt-1];
global_part[cnt++] += nx_size*ny_size;
}
}
nx_local = nx_part[p+1] - nx_part[p];
ny_local = ny_part[q+1] - ny_part[q];
my_id = q*P + p;
num_procs = P*Q;
local_num_rows = nx_local*ny_local;
diag_i = hypre_CTAlloc(HYPRE_Int, local_num_rows+1);
offd_i = hypre_CTAlloc(HYPRE_Int, local_num_rows+1);
P_busy = hypre_min(nx,P);
Q_busy = hypre_min(ny,Q);
num_cols_offd = 0;
if (p) num_cols_offd += ny_local;
if (p < P_busy-1) num_cols_offd += ny_local;
if (q) num_cols_offd += nx_local;
if (q < Q_busy-1) num_cols_offd += nx_local;
if (p && q) num_cols_offd++;
if (p && q < Q_busy-1 ) num_cols_offd++;
if (p < P_busy-1 && q ) num_cols_offd++;
if (p < P_busy-1 && q < Q_busy-1 ) num_cols_offd++;
if (!local_num_rows) num_cols_offd = 0;
col_map_offd = hypre_CTAlloc(HYPRE_Int, num_cols_offd);
cnt = 0;
o_cnt = 0;
diag_i[0] = 0;
offd_i[0] = 0;
for (iy = ny_part[q]; iy < ny_part[q+1]; iy++)
{
for (ix = nx_part[p]; ix < nx_part[p+1]; ix++)
{
cnt++;
o_cnt++;
diag_i[cnt] = diag_i[cnt-1];
offd_i[o_cnt] = offd_i[o_cnt-1];
diag_i[cnt]++;
if (iy > ny_part[q])
{
diag_i[cnt]++;
if (ix > nx_part[p])
{
diag_i[cnt]++;
}
else
{
if (ix)
offd_i[o_cnt]++;
}
}
else
{
if (iy)
{
offd_i[o_cnt]++;
if (ix > nx_part[p])
{
offd_i[o_cnt]++;
}
else if (ix)
{
offd_i[o_cnt]++;
}
}
}
if (ix > nx_part[p])
diag_i[cnt]++;
else
{
if (ix)
{
offd_i[o_cnt]++;
}
}
if (ix+1 < nx_part[p+1])
diag_i[cnt]++;
else
{
if (ix+1 < nx)
{
offd_i[o_cnt]++;
}
}
if (iy+1 < ny_part[q+1])
{
diag_i[cnt]++;
if (ix < nx_part[p+1]-1)
{
diag_i[cnt]++;
}
else
{
if (ix+1 < nx)
offd_i[o_cnt]++;
}
}
else
{
if (iy+1 < ny)
{
offd_i[o_cnt]++;
if (ix < nx_part[p+1]-1)
{
offd_i[o_cnt]++;
}
else if (ix < nx-1)
{
offd_i[o_cnt]++;
}
}
}
}
}
diag_j = hypre_CTAlloc(HYPRE_Int, diag_i[local_num_rows]);
diag_data = hypre_CTAlloc(HYPRE_Real, diag_i[local_num_rows]);
if (num_procs > 1)
{
offd_j = hypre_CTAlloc(HYPRE_Int, offd_i[local_num_rows]);
offd_data = hypre_CTAlloc(HYPRE_Real, offd_i[local_num_rows]);
}
row_index = 0;
cnt = 0;
o_cnt = 0;
for (iy = ny_part[q]; iy < ny_part[q+1]; iy++)
{
for (ix = nx_part[p]; ix < nx_part[p+1]; ix++)
{
diag_j[cnt] = row_index;
diag_data[cnt++] = value[0];
if (iy > ny_part[q])
{
if (ix > nx_part[p])
{
diag_j[cnt] = row_index-nx_local-1 ;
diag_data[cnt++] = value[3];
}
else
{
if (ix)
{
offd_j[o_cnt] = hypre_map2(ix-1,iy-1,p-1,q,P,Q,
nx_part,ny_part,global_part);
offd_data[o_cnt++] = value[3];
}
}
diag_j[cnt] = row_index-nx_local;
diag_data[cnt++] = value[2];
}
else
{
if (iy)
{
if (ix > nx_part[p])
{
offd_j[o_cnt] = hypre_map2(ix-1,iy-1,p,q-1,P,Q,
nx_part,ny_part,global_part);
offd_data[o_cnt++] = value[3];
}
else if (ix)
{
offd_j[o_cnt] = hypre_map2(ix-1,iy-1,p-1,q-1,P,Q,
nx_part,ny_part,global_part);
offd_data[o_cnt++] = value[3];
}
offd_j[o_cnt] = hypre_map2(ix,iy-1,p,q-1,P,Q,
nx_part,ny_part,global_part);
offd_data[o_cnt++] = value[2];
}
}
if (ix > nx_part[p])
{
diag_j[cnt] = row_index-1;
diag_data[cnt++] = value[1];
}
else
{
if (ix)
{
offd_j[o_cnt] = hypre_map2(ix-1,iy,p-1,q,P,Q,
nx_part,ny_part,global_part);
offd_data[o_cnt++] = value[1];
}
}
if (ix+1 < nx_part[p+1])
{
diag_j[cnt] = row_index+1;
diag_data[cnt++] = value[1];
}
else
{
if (ix+1 < nx)
{
offd_j[o_cnt] = hypre_map2(ix+1,iy,p+1,q,P,Q,
nx_part,ny_part,global_part);
offd_data[o_cnt++] = value[1];
}
}
if (iy+1 < ny_part[q+1])
{
diag_j[cnt] = row_index+nx_local;
diag_data[cnt++] = value[2];
if (ix < nx_part[p+1]-1)
{
diag_j[cnt] = row_index+nx_local+1 ;
diag_data[cnt++] = value[3];
}
else
{
if (ix+1 < nx)
{
offd_j[o_cnt] = hypre_map2(ix+1,iy+1,p+1,q,P,Q,
nx_part,ny_part,global_part);
offd_data[o_cnt++] = value[3];
}
}
}
else
{
if (iy+1 < ny)
{
offd_j[o_cnt] = hypre_map2(ix,iy+1,p,q+1,P,Q,
nx_part,ny_part,global_part);
offd_data[o_cnt++] = value[2];
if (ix < nx_part[p+1]-1)
{
offd_j[o_cnt] = hypre_map2(ix+1,iy+1,p,q+1,P,Q,
nx_part,ny_part,global_part);
offd_data[o_cnt++] = value[3];
}
else if (ix < nx-1)
{
offd_j[o_cnt] = hypre_map2(ix+1,iy+1,p+1,q+1,P,Q,
nx_part,ny_part,global_part);
offd_data[o_cnt++] = value[3];
}
}
}
row_index++;
}
}
if (num_procs > 1)
{
work = hypre_CTAlloc(HYPRE_Int,o_cnt);
for (i=0; i < o_cnt; i++)
work[i] = offd_j[i];
qsort0(work, 0, o_cnt-1);
col_map_offd[0] = work[0];
cnt = 0;
for (i=0; i < o_cnt; i++)
{
if (work[i] > col_map_offd[cnt])
{
cnt++;
col_map_offd[cnt] = work[i];
}
}
for (i=0; i < o_cnt; i++)
{
for (j=0; j < num_cols_offd; j++)
{
if (offd_j[i] == col_map_offd[j])
{
offd_j[i] = j;
break;
}
}
}
hypre_TFree(work);
}
A = hypre_ParCSRMatrixCreate(comm, grid_size, grid_size,
global_part, global_part, num_cols_offd,
diag_i[local_num_rows],
offd_i[local_num_rows]);
hypre_ParCSRMatrixColMapOffd(A) = col_map_offd;
diag = hypre_ParCSRMatrixDiag(A);
hypre_CSRMatrixI(diag) = diag_i;
hypre_CSRMatrixJ(diag) = diag_j;
hypre_CSRMatrixData(diag) = diag_data;
offd = hypre_ParCSRMatrixOffd(A);
hypre_CSRMatrixI(offd) = offd_i;
if (num_cols_offd)
{
hypre_CSRMatrixJ(offd) = offd_j;
hypre_CSRMatrixData(offd) = offd_data;
}
hypre_TFree(nx_part);
hypre_TFree(ny_part);
hypre_TFree(value);
return (HYPRE_ParCSRMatrix) A;
}
int HYPRE_ParCSR_SuperLUSetup(HYPRE_Solver solver, HYPRE_ParCSRMatrix A_csr,
HYPRE_ParVector b, HYPRE_ParVector x )
{
#ifdef HAVE_SUPERLU
int startRow, endRow, nrows, *partition, *AdiagI, *AdiagJ, nnz;
int irow, colNum, index, *cscI, *cscJ, jcol, *colLengs;
int *etree, permcSpec, lwork, panelSize, relax, info;
double *AdiagA, *cscA, diagPivotThresh, dropTol;
char refact[1];
hypre_CSRMatrix *Adiag;
HYPRE_SuperLU *sluPtr;
SuperMatrix sluAmat, auxAmat;
superlu_options_t slu_options;
SuperLUStat_t slu_stat;
/* ---------------------------------------------------------------- */
/* get matrix information */
/* ---------------------------------------------------------------- */
sluPtr = (HYPRE_SuperLU *) solver;
assert ( sluPtr != NULL );
HYPRE_ParCSRMatrixGetRowPartitioning( A_csr, &partition );
startRow = partition[0];
endRow = partition[1] - 1;
nrows = endRow - startRow + 1;
free( partition );
if ( startRow != 0 )
{
printf("HYPRE_ParCSR_SuperLUSetup ERROR - start row != 0.\n");
return -1;
}
/* ---------------------------------------------------------------- */
/* get hypre matrix */
/* ---------------------------------------------------------------- */
Adiag = hypre_ParCSRMatrixDiag((hypre_ParCSRMatrix *) A_csr);
AdiagI = hypre_CSRMatrixI(Adiag);
AdiagJ = hypre_CSRMatrixJ(Adiag);
AdiagA = hypre_CSRMatrixData(Adiag);
nnz = AdiagI[nrows];
/* ---------------------------------------------------------------- */
/* convert the csr matrix into csc matrix */
/* ---------------------------------------------------------------- */
colLengs = (int *) malloc(nrows * sizeof(int));
for ( irow = 0; irow < nrows; irow++ ) colLengs[irow] = 0;
for ( irow = 0; irow < nrows; irow++ )
for ( jcol = AdiagI[irow]; jcol < AdiagI[irow+1]; jcol++ )
colLengs[AdiagJ[jcol]]++;
cscJ = (int *) malloc( (nrows+1) * sizeof(int) );
cscI = (int *) malloc( nnz * sizeof(int) );
cscA = (double *) malloc( nnz * sizeof(double) );
cscJ[0] = 0;
nnz = 0;
for ( jcol = 1; jcol <= nrows; jcol++ )
{
nnz += colLengs[jcol-1];
cscJ[jcol] = nnz;
}
for ( irow = 0; irow < nrows; irow++ )
{
for ( jcol = AdiagI[irow]; jcol < AdiagI[irow+1]; jcol++ )
{
colNum = AdiagJ[jcol];
index = cscJ[colNum]++;
cscI[index] = irow;
cscA[index] = AdiagA[jcol];
}
}
cscJ[0] = 0;
nnz = 0;
for ( jcol = 1; jcol <= nrows; jcol++ )
{
nnz += colLengs[jcol-1];
cscJ[jcol] = nnz;
}
free(colLengs);
/* ---------------------------------------------------------------- */
/* create SuperMatrix */
/* ---------------------------------------------------------------- */
dCreate_CompCol_Matrix(&sluAmat,nrows,nrows,cscJ[nrows],cscA,cscI,
cscJ, SLU_NC, SLU_D, SLU_GE);
etree = (int *) malloc(nrows * sizeof(int));
sluPtr->permC_ = (int *) malloc(nrows * sizeof(int));
sluPtr->permR_ = (int *) malloc(nrows * sizeof(int));
permcSpec = 0;
get_perm_c(permcSpec, &sluAmat, sluPtr->permC_);
slu_options.Fact = DOFACT;
slu_options.SymmetricMode = NO;
sp_preorder(&slu_options, &sluAmat, sluPtr->permC_, etree, &auxAmat);
diagPivotThresh = 1.0;
dropTol = 0.0;
panelSize = sp_ienv(1);
relax = sp_ienv(2);
StatInit(&slu_stat);
lwork = 0;
slu_options.ColPerm = MY_PERMC;
slu_options.DiagPivotThresh = diagPivotThresh;
dgstrf(&slu_options, &auxAmat, dropTol, relax, panelSize,
etree, NULL, lwork, sluPtr->permC_, sluPtr->permR_,
&(sluPtr->SLU_Lmat), &(sluPtr->SLU_Umat), &slu_stat, &info);
Destroy_CompCol_Permuted(&auxAmat);
Destroy_CompCol_Matrix(&sluAmat);
free(etree);
sluPtr->factorized_ = 1;
StatFree(&slu_stat);
return 0;
#else
printf("HYPRE_ParCSR_SuperLUSetup ERROR - SuperLU not enabled.\n");
*solver = (HYPRE_Solver) NULL;
return -1;
#endif
}
HYPRE_Int
hypre_BoomerAMGCreateNodalA(hypre_ParCSRMatrix *A,
HYPRE_Int num_functions,
HYPRE_Int *dof_func,
HYPRE_Int option,
HYPRE_Int diag_option,
hypre_ParCSRMatrix **AN_ptr)
{
MPI_Comm comm = hypre_ParCSRMatrixComm(A);
hypre_CSRMatrix *A_diag = hypre_ParCSRMatrixDiag(A);
HYPRE_Int *A_diag_i = hypre_CSRMatrixI(A_diag);
double *A_diag_data = hypre_CSRMatrixData(A_diag);
hypre_CSRMatrix *A_offd = hypre_ParCSRMatrixOffd(A);
HYPRE_Int *A_offd_i = hypre_CSRMatrixI(A_offd);
double *A_offd_data = hypre_CSRMatrixData(A_offd);
HYPRE_Int *A_diag_j = hypre_CSRMatrixJ(A_diag);
HYPRE_Int *A_offd_j = hypre_CSRMatrixJ(A_offd);
HYPRE_Int *row_starts = hypre_ParCSRMatrixRowStarts(A);
HYPRE_Int *col_map_offd = hypre_ParCSRMatrixColMapOffd(A);
HYPRE_Int num_variables = hypre_CSRMatrixNumRows(A_diag);
HYPRE_Int num_nonzeros_offd = 0;
HYPRE_Int num_cols_offd = 0;
hypre_ParCSRMatrix *AN;
hypre_CSRMatrix *AN_diag;
HYPRE_Int *AN_diag_i;
HYPRE_Int *AN_diag_j;
double *AN_diag_data;
hypre_CSRMatrix *AN_offd;
HYPRE_Int *AN_offd_i;
HYPRE_Int *AN_offd_j;
double *AN_offd_data;
HYPRE_Int *col_map_offd_AN;
HYPRE_Int *new_col_map_offd;
HYPRE_Int *row_starts_AN;
HYPRE_Int AN_num_nonzeros_diag = 0;
HYPRE_Int AN_num_nonzeros_offd = 0;
HYPRE_Int num_cols_offd_AN;
HYPRE_Int new_num_cols_offd;
hypre_ParCSRCommPkg *comm_pkg = hypre_ParCSRMatrixCommPkg(A);
HYPRE_Int num_sends;
HYPRE_Int num_recvs;
HYPRE_Int *send_procs;
HYPRE_Int *send_map_starts;
HYPRE_Int *send_map_elmts;
HYPRE_Int *new_send_map_elmts;
HYPRE_Int *recv_procs;
HYPRE_Int *recv_vec_starts;
hypre_ParCSRCommPkg *comm_pkg_AN;
HYPRE_Int *send_procs_AN;
HYPRE_Int *send_map_starts_AN;
HYPRE_Int *send_map_elmts_AN;
HYPRE_Int *recv_procs_AN;
HYPRE_Int *recv_vec_starts_AN;
HYPRE_Int i, j, k, k_map;
HYPRE_Int ierr = 0;
HYPRE_Int index, row;
HYPRE_Int start_index;
HYPRE_Int num_procs;
HYPRE_Int node, cnt;
HYPRE_Int mode;
HYPRE_Int new_send_elmts_size;
HYPRE_Int global_num_nodes;
HYPRE_Int num_nodes;
HYPRE_Int num_fun2;
HYPRE_Int *map_to_node;
HYPRE_Int *map_to_map;
HYPRE_Int *counter;
double sum;
double *data;
hypre_MPI_Comm_size(comm,&num_procs);
if (!comm_pkg)
{
hypre_MatvecCommPkgCreate(A);
comm_pkg = hypre_ParCSRMatrixCommPkg(A);
}
mode = fabs(option);
comm_pkg_AN = NULL;
col_map_offd_AN = NULL;
#ifdef HYPRE_NO_GLOBAL_PARTITION
row_starts_AN = hypre_CTAlloc(HYPRE_Int, 2);
for (i=0; i < 2; i++)
{
row_starts_AN[i] = row_starts[i]/num_functions;
if (row_starts_AN[i]*num_functions < row_starts[i])
{
hypre_printf("nodes not properly aligned or incomplete info!\n");
return (87);
}
}
global_num_nodes = hypre_ParCSRMatrixGlobalNumRows(A)/num_functions;
#else
row_starts_AN = hypre_CTAlloc(HYPRE_Int, num_procs+1);
for (i=0; i < num_procs+1; i++)
{
row_starts_AN[i] = row_starts[i]/num_functions;
if (row_starts_AN[i]*num_functions < row_starts[i])
{
hypre_printf("nodes not properly aligned or incomplete info!\n");
return (87);
}
}
global_num_nodes = row_starts_AN[num_procs];
#endif
num_nodes = num_variables/num_functions;
num_fun2 = num_functions*num_functions;
map_to_node = hypre_CTAlloc(HYPRE_Int, num_variables);
AN_diag_i = hypre_CTAlloc(HYPRE_Int, num_nodes+1);
counter = hypre_CTAlloc(HYPRE_Int, num_nodes);
for (i=0; i < num_variables; i++)
map_to_node[i] = i/num_functions;
for (i=0; i < num_nodes; i++)
counter[i] = -1;
AN_num_nonzeros_diag = 0;
row = 0;
for (i=0; i < num_nodes; i++)
{
AN_diag_i[i] = AN_num_nonzeros_diag;
for (j=0; j < num_functions; j++)
{
for (k=A_diag_i[row]; k < A_diag_i[row+1]; k++)
{
k_map = map_to_node[A_diag_j[k]];
if (counter[k_map] < i)
{
counter[k_map] = i;
AN_num_nonzeros_diag++;
}
}
row++;
}
}
AN_diag_i[num_nodes] = AN_num_nonzeros_diag;
AN_diag_j = hypre_CTAlloc(HYPRE_Int, AN_num_nonzeros_diag);
AN_diag_data = hypre_CTAlloc(double, AN_num_nonzeros_diag);
AN_diag = hypre_CSRMatrixCreate(num_nodes,num_nodes,AN_num_nonzeros_diag);
hypre_CSRMatrixI(AN_diag) = AN_diag_i;
hypre_CSRMatrixJ(AN_diag) = AN_diag_j;
hypre_CSRMatrixData(AN_diag) = AN_diag_data;
for (i=0; i < num_nodes; i++)
counter[i] = -1;
index = 0;
start_index = 0;
row = 0;
switch (mode)
{
case 1: /* frobenius norm */
{
for (i=0; i < num_nodes; i++)
{
for (j=0; j < num_functions; j++)
{
for (k=A_diag_i[row]; k < A_diag_i[row+1]; k++)
{
k_map = map_to_node[A_diag_j[k]];
if (counter[k_map] < start_index)
{
counter[k_map] = index;
AN_diag_j[index] = k_map;
AN_diag_data[index] = A_diag_data[k]*A_diag_data[k];
index++;
}
else
{
AN_diag_data[counter[k_map]] +=
A_diag_data[k]*A_diag_data[k];
}
}
row++;
}
start_index = index;
}
for (i=0; i < AN_num_nonzeros_diag; i++)
AN_diag_data[i] = sqrt(AN_diag_data[i]);
}
break;
case 2: /* sum of abs. value of all elements in each block */
{
for (i=0; i < num_nodes; i++)
{
for (j=0; j < num_functions; j++)
{
for (k=A_diag_i[row]; k < A_diag_i[row+1]; k++)
{
k_map = map_to_node[A_diag_j[k]];
if (counter[k_map] < start_index)
{
counter[k_map] = index;
AN_diag_j[index] = k_map;
AN_diag_data[index] = fabs(A_diag_data[k]);
index++;
}
else
{
AN_diag_data[counter[k_map]] += fabs(A_diag_data[k]);
}
}
row++;
}
start_index = index;
}
for (i=0; i < AN_num_nonzeros_diag; i++)
AN_diag_data[i] /= num_fun2;
}
break;
case 3: /* largest element of each block (sets true value - not abs. value) */
{
for (i=0; i < num_nodes; i++)
{
for (j=0; j < num_functions; j++)
{
for (k=A_diag_i[row]; k < A_diag_i[row+1]; k++)
{
k_map = map_to_node[A_diag_j[k]];
if (counter[k_map] < start_index)
{
counter[k_map] = index;
AN_diag_j[index] = k_map;
AN_diag_data[index] = A_diag_data[k];
index++;
}
else
{
if (fabs(A_diag_data[k]) >
fabs(AN_diag_data[counter[k_map]]))
AN_diag_data[counter[k_map]] = A_diag_data[k];
}
}
row++;
}
start_index = index;
}
}
break;
case 4: /* inf. norm (row-sum) */
{
data = hypre_CTAlloc(double, AN_num_nonzeros_diag*num_functions);
for (i=0; i < num_nodes; i++)
{
for (j=0; j < num_functions; j++)
{
for (k=A_diag_i[row]; k < A_diag_i[row+1]; k++)
{
k_map = map_to_node[A_diag_j[k]];
if (counter[k_map] < start_index)
{
counter[k_map] = index;
AN_diag_j[index] = k_map;
data[index*num_functions + j] = fabs(A_diag_data[k]);
index++;
}
else
{
data[(counter[k_map])*num_functions + j] += fabs(A_diag_data[k]);
}
}
row++;
}
start_index = index;
}
for (i=0; i < AN_num_nonzeros_diag; i++)
{
AN_diag_data[i] = data[i*num_functions];
for (j=1; j< num_functions; j++)
{
AN_diag_data[i] = hypre_max( AN_diag_data[i],data[i*num_functions+j]);
}
}
hypre_TFree(data);
}
break;
case 6: /* sum of all elements in each block */
{
for (i=0; i < num_nodes; i++)
{
for (j=0; j < num_functions; j++)
{
for (k=A_diag_i[row]; k < A_diag_i[row+1]; k++)
{
k_map = map_to_node[A_diag_j[k]];
if (counter[k_map] < start_index)
{
counter[k_map] = index;
AN_diag_j[index] = k_map;
AN_diag_data[index] = (A_diag_data[k]);
index++;
}
else
{
AN_diag_data[counter[k_map]] += (A_diag_data[k]);
}
}
row++;
}
start_index = index;
}
}
break;
}
if (diag_option ==1 )
{
/* make the diag entry the negative of the sum of off-diag entries (DO MORE BELOW) */
for (i=0; i < num_nodes; i++)
{
index = AN_diag_i[i];
sum = 0.0;
for (k = AN_diag_i[i]+1; k < AN_diag_i[i+1]; k++)
{
sum += AN_diag_data[k];
}
AN_diag_data[index] = -sum;
}
}
else if (diag_option == 2)
{
/* make all diagonal entries negative */
/* the diagonal is the first element listed in each row - */
for (i=0; i < num_nodes; i++)
{
index = AN_diag_i[i];
AN_diag_data[index] = - AN_diag_data[index];
}
}
num_nonzeros_offd = A_offd_i[num_variables];
AN_offd_i = hypre_CTAlloc(HYPRE_Int, num_nodes+1);
num_cols_offd_AN = 0;
if (comm_pkg)
{
comm_pkg_AN = hypre_CTAlloc(hypre_ParCSRCommPkg,1);
hypre_ParCSRCommPkgComm(comm_pkg_AN) = comm;
num_sends = hypre_ParCSRCommPkgNumSends(comm_pkg);
hypre_ParCSRCommPkgNumSends(comm_pkg_AN) = num_sends;
num_recvs = hypre_ParCSRCommPkgNumRecvs(comm_pkg);
hypre_ParCSRCommPkgNumRecvs(comm_pkg_AN) = num_recvs;
send_procs = hypre_ParCSRCommPkgSendProcs(comm_pkg);
send_map_starts = hypre_ParCSRCommPkgSendMapStarts(comm_pkg);
send_map_elmts = hypre_ParCSRCommPkgSendMapElmts(comm_pkg);
recv_procs = hypre_ParCSRCommPkgRecvProcs(comm_pkg);
recv_vec_starts = hypre_ParCSRCommPkgRecvVecStarts(comm_pkg);
send_procs_AN = NULL;
send_map_elmts_AN = NULL;
if (num_sends)
{
send_procs_AN = hypre_CTAlloc(HYPRE_Int,num_sends);
send_map_elmts_AN = hypre_CTAlloc(HYPRE_Int,send_map_starts[num_sends]);
}
send_map_starts_AN = hypre_CTAlloc(HYPRE_Int,num_sends+1);
recv_vec_starts_AN = hypre_CTAlloc(HYPRE_Int,num_recvs+1);
recv_procs_AN = NULL;
if (num_recvs) recv_procs_AN = hypre_CTAlloc(HYPRE_Int,num_recvs);
for (i=0; i < num_sends; i++)
send_procs_AN[i] = send_procs[i];
for (i=0; i < num_recvs; i++)
recv_procs_AN[i] = recv_procs[i];
send_map_starts_AN[0] = 0;
cnt = 0;
for (i=0; i < num_sends; i++)
{
k_map = send_map_starts[i];
if (send_map_starts[i+1]-k_map)
send_map_elmts_AN[cnt++] = send_map_elmts[k_map]/num_functions;
for (j=send_map_starts[i]+1; j < send_map_starts[i+1]; j++)
{
node = send_map_elmts[j]/num_functions;
if (node > send_map_elmts_AN[cnt-1])
send_map_elmts_AN[cnt++] = node;
}
send_map_starts_AN[i+1] = cnt;
}
hypre_ParCSRCommPkgSendProcs(comm_pkg_AN) = send_procs_AN;
hypre_ParCSRCommPkgSendMapStarts(comm_pkg_AN) = send_map_starts_AN;
hypre_ParCSRCommPkgSendMapElmts(comm_pkg_AN) = send_map_elmts_AN;
hypre_ParCSRCommPkgRecvProcs(comm_pkg_AN) = recv_procs_AN;
hypre_ParCSRCommPkgRecvVecStarts(comm_pkg_AN) = recv_vec_starts_AN;
}
num_cols_offd = hypre_CSRMatrixNumCols(A_offd);
if (num_cols_offd)
{
if (num_cols_offd > num_variables)
{
hypre_TFree(map_to_node);
map_to_node = hypre_CTAlloc(HYPRE_Int,num_cols_offd);
}
num_cols_offd_AN = 1;
map_to_node[0] = col_map_offd[0]/num_functions;
for (i=1; i < num_cols_offd; i++)
{
map_to_node[i] = col_map_offd[i]/num_functions;
if (map_to_node[i] > map_to_node[i-1]) num_cols_offd_AN++;
}
if (num_cols_offd_AN > num_nodes)
{
hypre_TFree(counter);
counter = hypre_CTAlloc(HYPRE_Int,num_cols_offd_AN);
}
map_to_map = NULL;
col_map_offd_AN = NULL;
map_to_map = hypre_CTAlloc(HYPRE_Int, num_cols_offd);
col_map_offd_AN = hypre_CTAlloc(HYPRE_Int,num_cols_offd_AN);
col_map_offd_AN[0] = map_to_node[0];
recv_vec_starts_AN[0] = 0;
cnt = 1;
for (i=0; i < num_recvs; i++)
{
for (j=recv_vec_starts[i]; j < recv_vec_starts[i+1]; j++)
{
node = map_to_node[j];
if (node > col_map_offd_AN[cnt-1])
{
col_map_offd_AN[cnt++] = node;
}
map_to_map[j] = cnt-1;
}
recv_vec_starts_AN[i+1] = cnt;
}
for (i=0; i < num_cols_offd_AN; i++)
counter[i] = -1;
AN_num_nonzeros_offd = 0;
row = 0;
for (i=0; i < num_nodes; i++)
{
AN_offd_i[i] = AN_num_nonzeros_offd;
for (j=0; j < num_functions; j++)
{
for (k=A_offd_i[row]; k < A_offd_i[row+1]; k++)
{
k_map = map_to_map[A_offd_j[k]];
if (counter[k_map] < i)
{
counter[k_map] = i;
AN_num_nonzeros_offd++;
}
}
row++;
}
}
AN_offd_i[num_nodes] = AN_num_nonzeros_offd;
}
AN_offd = hypre_CSRMatrixCreate(num_nodes,num_cols_offd_AN,
AN_num_nonzeros_offd);
hypre_CSRMatrixI(AN_offd) = AN_offd_i;
if (AN_num_nonzeros_offd)
{
AN_offd_j = hypre_CTAlloc(HYPRE_Int, AN_num_nonzeros_offd);
AN_offd_data = hypre_CTAlloc(double, AN_num_nonzeros_offd);
hypre_CSRMatrixJ(AN_offd) = AN_offd_j;
hypre_CSRMatrixData(AN_offd) = AN_offd_data;
for (i=0; i < num_cols_offd_AN; i++)
counter[i] = -1;
index = 0;
row = 0;
AN_offd_i[0] = 0;
start_index = 0;
switch (mode)
{
case 1: /* frobenius norm */
{
for (i=0; i < num_nodes; i++)
{
for (j=0; j < num_functions; j++)
{
for (k=A_offd_i[row]; k < A_offd_i[row+1]; k++)
{
k_map = map_to_map[A_offd_j[k]];
if (counter[k_map] < start_index)
{
counter[k_map] = index;
AN_offd_j[index] = k_map;
AN_offd_data[index] = A_offd_data[k]*A_offd_data[k];
index++;
}
else
{
AN_offd_data[counter[k_map]] +=
A_offd_data[k]*A_offd_data[k];
}
}
row++;
}
start_index = index;
}
for (i=0; i < AN_num_nonzeros_offd; i++)
AN_offd_data[i] = sqrt(AN_offd_data[i]);
}
break;
case 2: /* sum of abs. value of all elements in block */
{
for (i=0; i < num_nodes; i++)
{
for (j=0; j < num_functions; j++)
{
for (k=A_offd_i[row]; k < A_offd_i[row+1]; k++)
{
k_map = map_to_map[A_offd_j[k]];
if (counter[k_map] < start_index)
{
counter[k_map] = index;
AN_offd_j[index] = k_map;
AN_offd_data[index] = fabs(A_offd_data[k]);
index++;
}
else
{
AN_offd_data[counter[k_map]] += fabs(A_offd_data[k]);
}
}
row++;
}
start_index = index;
}
for (i=0; i < AN_num_nonzeros_offd; i++)
AN_offd_data[i] /= num_fun2;
}
break;
case 3: /* largest element in each block (not abs. value ) */
{
for (i=0; i < num_nodes; i++)
{
for (j=0; j < num_functions; j++)
{
for (k=A_offd_i[row]; k < A_offd_i[row+1]; k++)
{
k_map = map_to_map[A_offd_j[k]];
if (counter[k_map] < start_index)
{
counter[k_map] = index;
AN_offd_j[index] = k_map;
AN_offd_data[index] = A_offd_data[k];
index++;
}
else
{
if (fabs(A_offd_data[k]) >
fabs(AN_offd_data[counter[k_map]]))
AN_offd_data[counter[k_map]] = A_offd_data[k];
}
}
row++;
}
start_index = index;
}
}
break;
case 4: /* inf. norm (row-sum) */
{
data = hypre_CTAlloc(double, AN_num_nonzeros_offd*num_functions);
for (i=0; i < num_nodes; i++)
{
for (j=0; j < num_functions; j++)
{
for (k=A_offd_i[row]; k < A_offd_i[row+1]; k++)
{
k_map = map_to_map[A_offd_j[k]];
if (counter[k_map] < start_index)
{
counter[k_map] = index;
AN_offd_j[index] = k_map;
data[index*num_functions + j] = fabs(A_offd_data[k]);
index++;
}
else
{
data[(counter[k_map])*num_functions + j] += fabs(A_offd_data[k]);
}
}
row++;
}
start_index = index;
}
for (i=0; i < AN_num_nonzeros_offd; i++)
{
AN_offd_data[i] = data[i*num_functions];
for (j=1; j< num_functions; j++)
{
AN_offd_data[i] = hypre_max( AN_offd_data[i],data[i*num_functions+j]);
}
}
hypre_TFree(data);
}
break;
case 6: /* sum of value of all elements in block */
{
for (i=0; i < num_nodes; i++)
{
for (j=0; j < num_functions; j++)
{
for (k=A_offd_i[row]; k < A_offd_i[row+1]; k++)
{
k_map = map_to_map[A_offd_j[k]];
if (counter[k_map] < start_index)
{
counter[k_map] = index;
AN_offd_j[index] = k_map;
AN_offd_data[index] = (A_offd_data[k]);
index++;
}
else
{
AN_offd_data[counter[k_map]] += (A_offd_data[k]);
}
}
row++;
}
start_index = index;
}
}
break;
}
hypre_TFree(map_to_map);
}
if (diag_option ==1 )
{
/* make the diag entry the negative of the sum of off-diag entries (here we are adding the
off_diag contribution)*/
/* the diagonal is the first element listed in each row of AN_diag_data - */
for (i=0; i < num_nodes; i++)
{
sum = 0.0;
for (k = AN_offd_i[i]; k < AN_offd_i[i+1]; k++)
{
sum += AN_offd_data[k];
}
index = AN_diag_i[i];/* location of diag entry in data */
AN_diag_data[index] -= sum; /* subtract from current value */
}
}
AN = hypre_ParCSRMatrixCreate(comm, global_num_nodes, global_num_nodes,
row_starts_AN, row_starts_AN, num_cols_offd_AN,
AN_num_nonzeros_diag, AN_num_nonzeros_offd);
/* we already created the diag and offd matrices - so we don't need the ones
created above */
hypre_CSRMatrixDestroy(hypre_ParCSRMatrixDiag(AN));
hypre_CSRMatrixDestroy(hypre_ParCSRMatrixOffd(AN));
hypre_ParCSRMatrixDiag(AN) = AN_diag;
hypre_ParCSRMatrixOffd(AN) = AN_offd;
hypre_ParCSRMatrixColMapOffd(AN) = col_map_offd_AN;
hypre_ParCSRMatrixCommPkg(AN) = comm_pkg_AN;
new_num_cols_offd = num_functions*num_cols_offd_AN;
if (new_num_cols_offd > num_cols_offd)
{
new_col_map_offd = hypre_CTAlloc(HYPRE_Int, new_num_cols_offd);
cnt = 0;
for (i=0; i < num_cols_offd_AN; i++)
{
for (j=0; j < num_functions; j++)
{
new_col_map_offd[cnt++] = num_functions*col_map_offd_AN[i]+j;
}
}
cnt = 0;
for (i=0; i < num_cols_offd; i++)
{
while (col_map_offd[i] > new_col_map_offd[cnt])
cnt++;
col_map_offd[i] = cnt++;
}
for (i=0; i < num_recvs+1; i++)
{
recv_vec_starts[i] = num_functions*recv_vec_starts_AN[i];
}
for (i=0; i < num_nonzeros_offd; i++)
{
j = A_offd_j[i];
A_offd_j[i] = col_map_offd[j];
}
hypre_ParCSRMatrixColMapOffd(A) = new_col_map_offd;
hypre_CSRMatrixNumCols(A_offd) = new_num_cols_offd;
hypre_TFree(col_map_offd);
}
hypre_TFree(map_to_node);
new_send_elmts_size = send_map_starts_AN[num_sends]*num_functions;
if (new_send_elmts_size > send_map_starts[num_sends])
{
new_send_map_elmts = hypre_CTAlloc(HYPRE_Int,new_send_elmts_size);
cnt = 0;
send_map_starts[0] = 0;
for (i=0; i < num_sends; i++)
{
send_map_starts[i+1] = send_map_starts_AN[i+1]*num_functions;
for (j=send_map_starts_AN[i]; j < send_map_starts_AN[i+1]; j++)
{
for (k=0; k < num_functions; k++)
new_send_map_elmts[cnt++] = send_map_elmts_AN[j]*num_functions+k;
}
}
hypre_TFree(send_map_elmts);
hypre_ParCSRCommPkgSendMapElmts(comm_pkg) = new_send_map_elmts;
}
*AN_ptr = AN;
hypre_TFree(counter);
return (ierr);
}
HYPRE_Int
hypre_InexactPartitionOfUnityInterpolation
(hypre_CSRMatrix **P_pointer,
HYPRE_Int *i_dof_dof,
HYPRE_Int *j_dof_dof,
HYPRE_Real *a_dof_dof,
HYPRE_Real *unit_vector,
HYPRE_Int *i_domain_dof,
HYPRE_Int *j_domain_dof,
HYPRE_Int num_domains, /* == num-coarsedofs */
HYPRE_Int num_dofs)
{
HYPRE_Int ierr = 0;
HYPRE_Int i,j,k;
HYPRE_Int ind = 1;
HYPRE_Int nu, nu_max = 1;
HYPRE_Real eps = 1.e-24;
HYPRE_Int max_iter = 1000;
HYPRE_Int iter;
HYPRE_Real delta0, delta_old, delta, alpha, tau, beta;
HYPRE_Real aux, diag;
HYPRE_Real *P_t_coeff;
hypre_CSRMatrix *P_t, *P;
HYPRE_Real *x,*r,*d,*g,*h;
HYPRE_Real *row_sum;
HYPRE_Int *i_global_to_local;
HYPRE_Int local_dof_counter;
HYPRE_Real *diag_dof_dof;
/* ------------------------------------------------------------------
domain_dof relation should satisfy the following property:
num_domains == num_coarsedofs;
each domain contains only one coarse dof;
------------------------------------------------------------------ */
i_global_to_local = hypre_CTAlloc(HYPRE_Int, num_dofs);
for (i=0; i < num_dofs; i++)
i_global_to_local[i] = -1;
local_dof_counter = 0;
for (i=0; i < num_domains; i++)
if (local_dof_counter < i_domain_dof[i+1]-i_domain_dof[i])
local_dof_counter = i_domain_dof[i+1]-i_domain_dof[i];
/* solve T x = unit_vector; --------------------------------------- */
/* cg loop: ------------------------------------------------------- */
hypre_printf("\n---------------------- num_domains: %d, nnz: %d;\n",
num_domains, i_domain_dof[num_domains]);
x = hypre_CTAlloc(HYPRE_Real, num_dofs);
d = hypre_CTAlloc(HYPRE_Real, num_dofs);
g = hypre_CTAlloc(HYPRE_Real, num_dofs);
r = hypre_CTAlloc(HYPRE_Real, num_dofs);
h = hypre_CTAlloc(HYPRE_Real, local_dof_counter);
diag_dof_dof = hypre_CTAlloc(HYPRE_Real, i_dof_dof[num_dofs]);
for (i=0; i<num_dofs; i++)
for (j=i_dof_dof[i]; j<i_dof_dof[i+1]; j++)
if (i!=j_dof_dof[j])
diag_dof_dof[j] = 0.e0;
else
diag_dof_dof[j] = a_dof_dof[j];
delta0 = 0.e0;
for (i=0; i < num_dofs; i++)
{
x[i] = 0.e0;
r[i] = unit_vector[i];
delta0+=r[i]*r[i];
}
/* compute initial iterate:
ierr =
compute_sum_A_i_action(x,
r,
i_domain_dof,
j_domain_dof,
i_dof_dof,
j_dof_dof,
a_dof_dof,
i_global_to_local,
num_domains,
num_dofs);
------------------------------------- */
/* matrix vector product: g < -- T x; ------------------------------ */
ierr=
compute_sym_GS_T_action(g,
x,
h,
i_domain_dof,
j_domain_dof,
nu_max,
i_dof_dof,
j_dof_dof,
a_dof_dof,
i_global_to_local,
num_domains,
num_dofs);
delta = 0;
for (i=0; i < num_dofs; i++)
{
r[i] -= g[i];
delta+=r[i]*r[i];
}
if (delta < eps * delta0)
goto end_cg;
ierr=
compute_sym_GS_T_action(g,
unit_vector,
h,
i_domain_dof,
j_domain_dof,
1,
i_dof_dof,
j_dof_dof,
diag_dof_dof,
i_global_to_local,
num_domains,
num_dofs);
/*
ierr =
compute_sum_A_i_action(d,
r,
i_domain_dof,
j_domain_dof,
i_dof_dof,
j_dof_dof,
a_dof_dof,
i_global_to_local,
num_domains,
num_dofs);
*/
for (i=0; i < num_dofs; i++)
d[i]=r[i]/g[i];
/* d contains precondtitioned residual: ------------------------ */
delta = 0.e0;
for (i=0; i < num_dofs; i++)
delta+=d[i]*r[i];
delta0 = delta;
eps = 1.e-12;
iter = 0;
loop:
/* matrix vector product: -------------------------------------- */
ierr=
compute_sym_GS_T_action(g,
d,
h,
i_domain_dof,
j_domain_dof,
nu_max,
i_dof_dof,
j_dof_dof,
a_dof_dof,
i_global_to_local,
num_domains,
num_dofs);
tau = 0.e0;
for (i=0; i < num_dofs; i++)
tau += d[i]*g[i];
alpha = delta/tau;
for (i=0; i < num_dofs; i++)
{
x[i] += alpha * d[i];
r[i] -= alpha * g[i];
}
iter++;
delta_old = delta;
/*
ierr =
compute_sum_A_i_action(g,
r,
i_domain_dof,
j_domain_dof,
i_dof_dof,
j_dof_dof,
a_dof_dof,
i_global_to_local,
num_domains,
num_dofs);
*/
ierr=
compute_sym_GS_T_action(g,
unit_vector,
h,
i_domain_dof,
j_domain_dof,
1,
i_dof_dof,
j_dof_dof,
diag_dof_dof,
i_global_to_local,
num_domains,
num_dofs);
for (i=0; i < num_dofs; i++)
g[i] = r[i]/g[i];
delta = 0.e0;
for (i=0; i < num_dofs; i++)
delta += g[i] * r[i];
hypre_printf("\n---------------------- iter: %d, delta: %le\n",
iter, delta);
if (delta < eps * delta0 || iter > max_iter)
goto end_cg;
beta = delta/delta_old;
for (i=0; i < num_dofs; i++)
d[i] = g[i] + beta * d[i];
goto loop;
end_cg:
hypre_printf("\n END CG in partition of unity interpolation; num_iters: %d\n",
iter);
hypre_TFree(r);
hypre_TFree(g);
hypre_TFree(d);
/* ith column of P is T_i x; ----------------------------------- */
P_t_coeff = hypre_CTAlloc(HYPRE_Real, i_domain_dof[num_domains]);
for (i=0; i < num_domains; i++)
{
for (j=i_domain_dof[i]; j < i_domain_dof[i+1]; j++)
{
i_global_to_local[j_domain_dof[j]] = j-i_domain_dof[i];
h[j-i_domain_dof[i]] = 0.e0;
}
nu = 0;
loop_nu:
for (j=i_domain_dof[i]; j < i_domain_dof[i+1]; j++)
{
aux = x[j_domain_dof[j]];
for (k=i_dof_dof[j_domain_dof[j]];
k<i_dof_dof[j_domain_dof[j]+1]; k++)
if (i_global_to_local[j_dof_dof[k]] > -1)
{
/* this is a_{i_loc, j_loc} --------------------------------- */
if (j_dof_dof[k] != j_domain_dof[j])
{
aux -= a_dof_dof[k] * h[i_global_to_local[j_dof_dof[k]]];
}
else
{
diag = a_dof_dof[k];
}
}
h[i_global_to_local[j_domain_dof[j]]] = aux/diag;
}
for (j=i_domain_dof[i+1]-1; j >= i_domain_dof[i]; j--)
{
aux = x[j_domain_dof[j]];
for (k =i_dof_dof[j_domain_dof[j]+1]-1;
k>=i_dof_dof[j_domain_dof[j]]; k--)
if (i_global_to_local[j_dof_dof[k]] > -1)
{
/* this is a_{i_loc, j_loc} --------------------------------- */
if (j_dof_dof[k] != j_domain_dof[j])
{
aux -= a_dof_dof[k] * h[i_global_to_local[j_dof_dof[k]]];
}
else
{
diag = a_dof_dof[k];
}
}
h[i_global_to_local[j_domain_dof[j]]] = aux/diag;
}
nu++;
if (nu < nu_max)
goto loop_nu;
for (j=i_domain_dof[i]; j < i_domain_dof[i+1]; j++)
{
P_t_coeff[j]= h[i_global_to_local[j_domain_dof[j]]];
i_global_to_local[j_domain_dof[j]] = -1;
}
}
hypre_TFree(diag_dof_dof);
hypre_TFree(x);
hypre_TFree(h);
hypre_TFree(i_global_to_local);
P_t = hypre_CSRMatrixCreate(num_domains, num_dofs,
i_domain_dof[num_domains]);
hypre_CSRMatrixData(P_t) = P_t_coeff;
hypre_CSRMatrixI(P_t) = i_domain_dof;
hypre_CSRMatrixJ(P_t) = j_domain_dof;
row_sum = hypre_CTAlloc(HYPRE_Real, num_dofs);
for (i=0; i < num_dofs; i++)
row_sum[i] = 0.e0;
for (i=0; i < num_domains; i++)
for (j=i_domain_dof[i]; j < i_domain_dof[i+1]; j++)
row_sum[j_domain_dof[j]]+=P_t_coeff[j];
delta = 0.e0;
for (i=0; i < num_dofs; i++)
delta+= (row_sum[i] - 1.e0)*(row_sum[i] - 1.e0);
hypre_printf("\n unit row_sum deviation in seq_PU_interpolation: %le\n",
sqrt(delta/num_dofs));
hypre_TFree(row_sum);
ind = 1;
ierr =
hypre_CSRMatrixTranspose(P_t, &P, ind);
*P_pointer = P;
hypre_CSRMatrixI(P_t) = NULL;
hypre_CSRMatrixJ(P_t) = NULL;
hypre_CSRMatrixDestroy(P_t);
return ierr;
}
HYPRE_Int
hypre_AMGSetupStats( void *amg_vdata )
{
hypre_AMGData *amg_data = amg_vdata;
/* Data Structure variables */
hypre_CSRMatrix **A_array;
hypre_CSRMatrix **P_array;
HYPRE_Int num_levels;
HYPRE_Int num_nonzeros;
/* HYPRE_Int amg_ioutdat;
char *log_file_name;
*/
/* Local variables */
HYPRE_Int *A_i;
double *A_data;
HYPRE_Int *P_i;
double *P_data;
HYPRE_Int level;
HYPRE_Int i,j;
HYPRE_Int fine_size;
HYPRE_Int coarse_size;
HYPRE_Int entries;
HYPRE_Int total_entries;
HYPRE_Int min_entries;
HYPRE_Int max_entries;
double avg_entries;
double rowsum;
double min_rowsum;
double max_rowsum;
double sparse;
double min_weight;
double max_weight;
double op_complxty=0;
double grid_complxty=0;
double num_nz0;
double num_var0;
A_array = hypre_AMGDataAArray(amg_data);
P_array = hypre_AMGDataPArray(amg_data);
num_levels = hypre_AMGDataNumLevels(amg_data);
/* amg_ioutdat = hypre_AMGDataIOutDat(amg_data);
log_file_name = hypre_AMGDataLogFileName(amg_data);
*/
hypre_printf("\n AMG SETUP PARAMETERS:\n\n");
hypre_printf(" Strength threshold = %f\n",hypre_AMGDataStrongThreshold(amg_data));
hypre_printf(" Max levels = %d\n",hypre_AMGDataMaxLevels(amg_data));
hypre_printf(" Num levels = %d\n\n",num_levels);
hypre_printf( "\nOperator Matrix Information:\n\n");
hypre_printf(" nonzero entries p");
hypre_printf("er row row sums\n");
hypre_printf("lev rows entries sparse min max ");
hypre_printf("avg min max\n");
hypre_printf("=======================================");
hypre_printf("==========================\n");
/*-----------------------------------------------------
* Enter Statistics Loop
*-----------------------------------------------------*/
num_var0 = (double) hypre_CSRMatrixNumRows(A_array[0]);
num_nz0 = (double) hypre_CSRMatrixNumNonzeros(A_array[0]);
for (level = 0; level < num_levels; level++)
{
A_i = hypre_CSRMatrixI(A_array[level]);
A_data = hypre_CSRMatrixData(A_array[level]);
fine_size = hypre_CSRMatrixNumRows(A_array[level]);
num_nonzeros = hypre_CSRMatrixNumNonzeros(A_array[level]);
sparse = num_nonzeros /((double) fine_size * (double) fine_size);
op_complxty += ((double)num_nonzeros/num_nz0);
grid_complxty += ((double)fine_size/num_var0);
min_entries = A_i[1]-A_i[0];
max_entries = 0;
total_entries = 0;
min_rowsum = 0.0;
max_rowsum = 0.0;
for (j = A_i[0]; j < A_i[1]; j++)
min_rowsum += A_data[j];
max_rowsum = min_rowsum;
for (j = 0; j < fine_size; j++)
{
entries = A_i[j+1] - A_i[j];
min_entries = hypre_min(entries, min_entries);
max_entries = hypre_max(entries, max_entries);
total_entries += entries;
rowsum = 0.0;
for (i = A_i[j]; i < A_i[j+1]; i++)
rowsum += A_data[i];
min_rowsum = hypre_min(rowsum, min_rowsum);
max_rowsum = hypre_max(rowsum, max_rowsum);
}
avg_entries = ((double) total_entries) / ((double) fine_size);
hypre_printf( "%2d %5d %7d %0.3f %3d %3d",
level, fine_size, num_nonzeros, sparse, min_entries,
max_entries);
hypre_printf(" %4.1f %10.3e %10.3e\n", avg_entries,
min_rowsum, max_rowsum);
}
hypre_printf( "\n\nInterpolation Matrix Information:\n\n");
hypre_printf(" entries/row min max");
hypre_printf(" row sums\n");
hypre_printf("lev rows cols min max ");
hypre_printf(" weight weight min max \n");
hypre_printf("=======================================");
hypre_printf("==========================\n");
/*-----------------------------------------------------
* Enter Statistics Loop
*-----------------------------------------------------*/
for (level = 0; level < num_levels-1; level++)
{
P_i = hypre_CSRMatrixI(P_array[level]);
P_data = hypre_CSRMatrixData(P_array[level]);
fine_size = hypre_CSRMatrixNumRows(P_array[level]);
coarse_size = hypre_CSRMatrixNumCols(P_array[level]);
num_nonzeros = hypre_CSRMatrixNumNonzeros(P_array[level]);
min_entries = P_i[1]-P_i[0];
max_entries = 0;
total_entries = 0;
min_rowsum = 0.0;
max_rowsum = 0.0;
min_weight = P_data[0];
max_weight = 0.0;
for (j = P_i[0]; j < P_i[1]; j++)
min_rowsum += P_data[j];
max_rowsum = min_rowsum;
for (j = 0; j < num_nonzeros; j++)
{
if (P_data[j] != 1.0)
{
min_weight = hypre_min(min_weight,P_data[j]);
max_weight = hypre_max(max_weight,P_data[j]);
}
}
for (j = 0; j < fine_size; j++)
{
entries = P_i[j+1] - P_i[j];
min_entries = hypre_min(entries, min_entries);
max_entries = hypre_max(entries, max_entries);
total_entries += entries;
rowsum = 0.0;
for (i = P_i[j]; i < P_i[j+1]; i++)
rowsum += P_data[i];
min_rowsum = hypre_min(rowsum, min_rowsum);
max_rowsum = hypre_max(rowsum, max_rowsum);
}
hypre_printf( "%2d %5d x %-5d %3d %3d",
level, fine_size, coarse_size, min_entries, max_entries);
hypre_printf(" %5.3e %5.3e %5.3e %5.3e\n",
min_weight, max_weight, min_rowsum, max_rowsum);
}
hypre_printf("\n Operator Complexity: %8.3f\n", op_complxty);
hypre_printf(" Grid Complexity: %8.3f\n", grid_complxty);
hypre_WriteSolverParams(amg_data);
return(0);
}
HYPRE_Int hypre_HarmonicExtension (hypre_CSRMatrix *A,
hypre_CSRMatrix *P,
HYPRE_Int num_DOF, HYPRE_Int *DOF,
HYPRE_Int num_idof, HYPRE_Int *idof,
HYPRE_Int num_bdof, HYPRE_Int *bdof)
{
HYPRE_Int ierr = 0;
HYPRE_Int i, j, k, l, m;
double factor;
HYPRE_Int *IA = hypre_CSRMatrixI(A);
HYPRE_Int *JA = hypre_CSRMatrixJ(A);
double *dataA = hypre_CSRMatrixData(A);
HYPRE_Int *IP = hypre_CSRMatrixI(P);
HYPRE_Int *JP = hypre_CSRMatrixJ(P);
double *dataP = hypre_CSRMatrixData(P);
double * Aii = hypre_CTAlloc(double, num_idof*num_idof);
double * Pi = hypre_CTAlloc(double, num_idof*num_DOF);
/* Loop over the rows of A */
for (i = 0; i < num_idof; i++)
for (j = IA[i]; j < IA[i+1]; j++)
{
/* Global to local*/
k = hypre_BinarySearch(idof,JA[j], num_idof);
/* If a column is a bdof, compute its participation in Pi = Aib x Pb */
if (k == -1)
{
k = hypre_BinarySearch(bdof,JA[j], num_bdof);
if (k > -1)
{
for (l = IP[k+num_idof]; l < IP[k+num_idof+1]; l++)
{
m = hypre_BinarySearch(DOF,JP[l], num_DOF);
if (m > -1)
{
m+=i*num_DOF;
/* Pi[i*num_DOF+m] += dataA[j] * dataP[l];*/
Pi[m] += dataA[j] * dataP[l];
}
}
}
}
/* If a column is an idof, put it in Aii */
else
Aii[i*num_idof+k] = dataA[j];
}
/* Perform Gaussian elimination in [Aii, Pi] */
for (j = 0; j < num_idof-1; j++)
if (Aii[j*num_idof+j] != 0.0)
for (i = j+1; i < num_idof; i++)
if (Aii[i*num_idof+j] != 0.0)
{
factor = Aii[i*num_idof+j]/Aii[j*num_idof+j];
for (m = j+1; m < num_idof; m++)
Aii[i*num_idof+m] -= factor * Aii[j*num_idof+m];
for (m = 0; m < num_DOF; m++)
Pi[i*num_DOF+m] -= factor * Pi[j*num_DOF+m];
}
/* Back Substitution */
for (i = num_idof-1; i >= 0; i--)
{
for (j = i+1; j < num_idof; j++)
if (Aii[i*num_idof+j] != 0.0)
for (m = 0; m < num_DOF; m++)
Pi[i*num_DOF+m] -= Aii[i*num_idof+j] * Pi[j*num_DOF+m];
for (m = 0; m < num_DOF; m++)
Pi[i*num_DOF+m] /= Aii[i*num_idof+i];
}
/* Put -Pi back in P. We assume that each idof depends on _all_ DOFs */
for (i = 0; i < num_idof; i++, JP += num_DOF, dataP += num_DOF)
for (j = 0; j < num_DOF; j++)
{
JP[j] = DOF[j];
dataP[j] = -Pi[i*num_DOF+j];
}
hypre_TFree(Aii);
hypre_TFree(Pi);
return ierr;
}