PetscErrorCode MatDestroy_SuperLU_DIST(Mat A)
{
PetscErrorCode ierr;
Mat_SuperLU_DIST *lu = (Mat_SuperLU_DIST*)A->spptr;
PetscBool flg;
PetscFunctionBegin;
if (lu && lu->CleanUpSuperLU_Dist) {
/* Deallocate SuperLU_DIST storage */
if (lu->MatInputMode == GLOBAL) {
Destroy_CompCol_Matrix_dist(&lu->A_sup);
} else {
Destroy_CompRowLoc_Matrix_dist(&lu->A_sup);
if ( lu->options.SolveInitialized ) {
#if defined(PETSC_USE_COMPLEX)
zSolveFinalize(&lu->options, &lu->SOLVEstruct);
#else
dSolveFinalize(&lu->options, &lu->SOLVEstruct);
#endif
}
}
Destroy_LU(A->cmap->N, &lu->grid, &lu->LUstruct);
ScalePermstructFree(&lu->ScalePermstruct);
LUstructFree(&lu->LUstruct);
/* Release the SuperLU_DIST process grid. */
superlu_gridexit(&lu->grid);
ierr = MPI_Comm_free(&(lu->comm_superlu));CHKERRQ(ierr);
}
ierr = PetscFree(A->spptr);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQAIJ,&flg);CHKERRQ(ierr);
if (flg) {
ierr = MatDestroy_SeqAIJ(A);CHKERRQ(ierr);
} else {
ierr = MatDestroy_MPIAIJ(A);CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
int main(int argc, char *argv[])
{
superlu_dist_options_t options;
SuperLUStat_t stat;
SuperMatrix A;
ScalePermstruct_t ScalePermstruct;
LUstruct_t LUstruct;
gridinfo_t grid1, grid2;
double *berr;
double *a, *b, *xtrue;
int_t *asub, *xa;
int_t i, j, m, n, nnz;
int_t nprow, npcol, ldumap, p;
int_t usermap[6];
int iam, info, ldb, ldx, nprocs;
int nrhs = 1; /* Number of right-hand side. */
char trans[1];
char **cpp, c;
FILE *fp, *fopen();
/* prototypes */
extern void LUstructInit(const int_t, LUstruct_t *);
extern void LUstructFree(LUstruct_t *);
extern void Destroy_LU(int_t, gridinfo_t *, LUstruct_t *);
/* ------------------------------------------------------------
INITIALIZE MPI ENVIRONMENT.
------------------------------------------------------------*/
MPI_Init( &argc, &argv );
MPI_Comm_size( MPI_COMM_WORLD, &nprocs );
if ( nprocs < 10 ) {
fprintf(stderr, "Requires at least 10 processes\n");
exit(-1);
}
/* Parse command line argv[]. */
for (cpp = argv+1; *cpp; ++cpp) {
if ( **cpp == '-' ) {
c = *(*cpp+1);
++cpp;
switch (c) {
case 'h':
printf("Options:\n");
printf("\t-r <int>: process rows (default %d)\n", nprow);
printf("\t-c <int>: process columns (default %d)\n", npcol);
exit(0);
break;
case 'r': nprow = atoi(*cpp);
break;
case 'c': npcol = atoi(*cpp);
break;
}
} else { /* Last arg is considered a filename */
if ( !(fp = fopen(*cpp, "r")) ) {
ABORT("File does not exist");
}
break;
}
}
/* ------------------------------------------------------------
INITIALIZE THE SUPERLU PROCESS GRID 1.
------------------------------------------------------------*/
nprow = 2;
npcol = 3;
ldumap = 2;
p = 0; /* Grid 1 starts from process 0. */
for (i = 0; i < nprow; ++i)
for (j = 0; j < npcol; ++j) usermap[i+j*ldumap] = p++;
superlu_gridmap(MPI_COMM_WORLD, nprow, npcol, usermap, ldumap, &grid1);
/* ------------------------------------------------------------
INITIALIZE THE SUPERLU PROCESS GRID 2.
------------------------------------------------------------*/
nprow = 2;
npcol = 2;
ldumap = 2;
p = 6; /* Grid 2 starts from process 6. */
for (i = 0; i < nprow; ++i)
for (j = 0; j < npcol; ++j) usermap[i+j*ldumap] = p++;
superlu_gridmap(MPI_COMM_WORLD, nprow, npcol, usermap, ldumap, &grid2);
/* Bail out if I do not belong in any of the 2 grids. */
MPI_Comm_rank( MPI_COMM_WORLD, &iam );
if ( iam >= 10 ) goto out;
#if ( DEBUGlevel>=1 )
CHECK_MALLOC(iam, "Enter main()");
#endif
if ( iam >= 0 && iam < 6 ) { /* I am in grid 1. */
iam = grid1.iam; /* Get the logical number in the new grid. */
/* ------------------------------------------------------------
PROCESS 0 READS THE MATRIX A, AND THEN BROADCASTS IT TO ALL
THE OTHER PROCESSES.
------------------------------------------------------------*/
if ( !iam ) {
/* Read the matrix stored on disk in Harwell-Boeing format. */
dreadhb_dist(iam, fp, &m, &n, &nnz, &a, &asub, &xa);
printf("\tDimension\t%dx%d\t # nonzeros %d\n", m, n, nnz);
printf("\tProcess grid\t%d X %d\n", (int) grid1.nprow, (int) grid1.npcol);
/* Broadcast matrix A to the other PEs. */
MPI_Bcast( &m, 1, mpi_int_t, 0, grid1.comm );
MPI_Bcast( &n, 1, mpi_int_t, 0, grid1.comm );
MPI_Bcast( &nnz, 1, mpi_int_t, 0, grid1.comm );
MPI_Bcast( a, nnz, MPI_DOUBLE, 0, grid1.comm );
MPI_Bcast( asub, nnz, mpi_int_t, 0, grid1.comm );
MPI_Bcast( xa, n+1, mpi_int_t, 0, grid1.comm );
} else {
/* Receive matrix A from PE 0. */
MPI_Bcast( &m, 1, mpi_int_t, 0, grid1.comm );
MPI_Bcast( &n, 1, mpi_int_t, 0, grid1.comm );
MPI_Bcast( &nnz, 1, mpi_int_t, 0, grid1.comm );
/* Allocate storage for compressed column representation. */
dallocateA_dist(n, nnz, &a, &asub, &xa);
MPI_Bcast( a, nnz, MPI_DOUBLE, 0, grid1.comm );
MPI_Bcast( asub, nnz, mpi_int_t, 0, grid1.comm );
MPI_Bcast( xa, n+1, mpi_int_t, 0, grid1.comm );
}
/* Create compressed column matrix for A. */
dCreate_CompCol_Matrix_dist(&A, m, n, nnz, a, asub, xa,
SLU_NC, SLU_D, SLU_GE);
/* Generate the exact solution and compute the right-hand side. */
if (!(b=doubleMalloc_dist(m*nrhs))) ABORT("Malloc fails for b[]");
if (!(xtrue=doubleMalloc_dist(n*nrhs))) ABORT("Malloc fails for xtrue[]");
*trans = 'N';
ldx = n;
ldb = m;
dGenXtrue_dist(n, nrhs, xtrue, ldx);
dFillRHS_dist(trans, nrhs, xtrue, ldx, &A, b, ldb);
if ( !(berr = doubleMalloc_dist(nrhs)) )
ABORT("Malloc fails for berr[].");
/* ------------------------------------------------------------
NOW WE SOLVE THE LINEAR SYSTEM.
------------------------------------------------------------*/
/* Set the default input options:
options.Fact = DOFACT;
options.Equil = YES;
options.ColPerm = METIS_AT_PLUS_A;
options.RowPerm = LargeDiag;
options.ReplaceTinyPivot = YES;
options.Trans = NOTRANS;
options.IterRefine = DOUBLE;
options.SolveInitialized = NO;
options.RefineInitialized = NO;
options.PrintStat = YES;
*/
set_default_options_dist(&options);
if (!iam) {
print_sp_ienv_dist(&options);
print_options_dist(&options);
}
/* Initialize ScalePermstruct and LUstruct. */
ScalePermstructInit(m, n, &ScalePermstruct);
LUstructInit(n, &LUstruct);
/* Initialize the statistics variables. */
PStatInit(&stat);
/* Call the linear equation solver: factorize and solve. */
pdgssvx_ABglobal(&options, &A, &ScalePermstruct, b, ldb, nrhs, &grid1,
&LUstruct, berr, &stat, &info);
/* Check the accuracy of the solution. */
if ( !iam ) {
dinf_norm_error_dist(n, nrhs, b, ldb, xtrue, ldx, &grid1);
}
/* Print the statistics. */
PStatPrint(&options, &stat, &grid1);
/* ------------------------------------------------------------
DEALLOCATE STORAGE.
------------------------------------------------------------*/
PStatFree(&stat);
Destroy_CompCol_Matrix_dist(&A);
Destroy_LU(n, &grid1, &LUstruct);
ScalePermstructFree(&ScalePermstruct);
LUstructFree(&LUstruct);
SUPERLU_FREE(b);
SUPERLU_FREE(xtrue);
SUPERLU_FREE(berr);
} else { /* I am in grid 2. */
iam = grid2.iam; /* Get the logical number in the new grid. */
/* ------------------------------------------------------------
PROCESS 0 READS THE MATRIX A, AND THEN BROADCASTS IT TO ALL
THE OTHER PROCESSES.
------------------------------------------------------------*/
if ( !iam ) {
/* Read the matrix stored on disk in Harwell-Boeing format. */
dreadhb_dist(iam, fp, &m, &n, &nnz, &a, &asub, &xa);
printf("\tDimension\t%dx%d\t # nonzeros %d\n", m, n, nnz);
printf("\tProcess grid\t%d X %d\n", (int) grid2.nprow, (int) grid2.npcol);
/* Broadcast matrix A to the other PEs. */
MPI_Bcast( &m, 1, mpi_int_t, 0, grid2.comm );
MPI_Bcast( &n, 1, mpi_int_t, 0, grid2.comm );
MPI_Bcast( &nnz, 1, mpi_int_t, 0, grid2.comm );
MPI_Bcast( a, nnz, MPI_DOUBLE, 0, grid2.comm );
MPI_Bcast( asub, nnz, mpi_int_t, 0, grid2.comm );
MPI_Bcast( xa, n+1, mpi_int_t, 0, grid2.comm );
} else {
/* Receive matrix A from PE 0. */
MPI_Bcast( &m, 1, mpi_int_t, 0, grid2.comm );
MPI_Bcast( &n, 1, mpi_int_t, 0, grid2.comm );
MPI_Bcast( &nnz, 1, mpi_int_t, 0, grid2.comm );
/* Allocate storage for compressed column representation. */
dallocateA_dist(n, nnz, &a, &asub, &xa);
MPI_Bcast( a, nnz, MPI_DOUBLE, 0, grid2.comm );
MPI_Bcast( asub, nnz, mpi_int_t, 0, grid2.comm );
MPI_Bcast( xa, n+1, mpi_int_t, 0, grid2.comm );
}
/* Create compressed column matrix for A. */
dCreate_CompCol_Matrix_dist(&A, m, n, nnz, a, asub, xa,
SLU_NC, SLU_D, SLU_GE);
/* Generate the exact solution and compute the right-hand side. */
if (!(b=doubleMalloc_dist(m*nrhs))) ABORT("Malloc fails for b[]");
if (!(xtrue=doubleMalloc_dist(n*nrhs))) ABORT("Malloc fails for xtrue[]");
*trans = 'N';
ldx = n;
ldb = m;
dGenXtrue_dist(n, nrhs, xtrue, ldx);
dFillRHS_dist(trans, nrhs, xtrue, ldx, &A, b, ldb);
if ( !(berr = doubleMalloc_dist(nrhs)) )
ABORT("Malloc fails for berr[].");
/* ------------------------------------------------------------
NOW WE SOLVE THE LINEAR SYSTEM.
------------------------------------------------------------*/
/* Set the default input options:
options.Fact = DOFACT;
options.Equil = YES;
options.ColPerm = MMD_AT_PLUS_A;
options.RowPerm = LargeDiag;
options.ReplaceTinyPivot = YES;
options.Trans = NOTRANS;
options.IterRefine = DOUBLE;
options.SolveInitialized = NO;
options.RefineInitialized = NO;
options.PrintStat = YES;
*/
set_default_options_dist(&options);
/* Initialize ScalePermstruct and LUstruct. */
ScalePermstructInit(m, n, &ScalePermstruct);
LUstructInit(n, &LUstruct);
/* Initialize the statistics variables. */
PStatInit(&stat);
/* Call the linear equation solver: factorize and solve. */
pdgssvx_ABglobal(&options, &A, &ScalePermstruct, b, ldb, nrhs, &grid2,
&LUstruct, berr, &stat, &info);
/* Check the accuracy of the solution. */
if ( !iam ) {
dinf_norm_error_dist(n, nrhs, b, ldb, xtrue, ldx, &grid2);
}
/* Print the statistics. */
PStatPrint(&options, &stat, &grid2);
/* ------------------------------------------------------------
DEALLOCATE STORAGE.
------------------------------------------------------------*/
PStatFree(&stat);
Destroy_CompCol_Matrix_dist(&A);
Destroy_LU(n, &grid2, &LUstruct);
ScalePermstructFree(&ScalePermstruct);
LUstructFree(&LUstruct);
SUPERLU_FREE(b);
SUPERLU_FREE(xtrue);
SUPERLU_FREE(berr);
}
/* ------------------------------------------------------------
RELEASE THE SUPERLU PROCESS GRIDS.
------------------------------------------------------------*/
superlu_gridexit(&grid1);
superlu_gridexit(&grid2);
out:
/* ------------------------------------------------------------
TERMINATES THE MPI EXECUTION ENVIRONMENT.
------------------------------------------------------------*/
MPI_Finalize();
#if ( DEBUGlevel>=1 )
CHECK_MALLOC(iam, "Exit main()");
#endif
}
void
pdgssvx(superlu_options_t *options, SuperMatrix *A,
ScalePermstruct_t *ScalePermstruct,
double B[], int ldb, int nrhs, gridinfo_t *grid,
LUstruct_t *LUstruct, SOLVEstruct_t *SOLVEstruct, double *berr,
SuperLUStat_t *stat, int *info)
{
/*
* -- Distributed SuperLU routine (version 2.2) --
* Lawrence Berkeley National Lab, Univ. of California Berkeley.
* November 1, 2007
* Feburary 20, 2008
*
*
* Purpose
* =======
*
* PDGSSVX solves a system of linear equations A*X=B,
* by using Gaussian elimination with "static pivoting" to
* compute the LU factorization of A.
*
* Static pivoting is a technique that combines the numerical stability
* of partial pivoting with the scalability of Cholesky (no pivoting),
* to run accurately and efficiently on large numbers of processors.
* See our paper at http://www.nersc.gov/~xiaoye/SuperLU/ for a detailed
* description of the parallel algorithms.
*
* The input matrices A and B are distributed by block rows.
* Here is a graphical illustration (0-based indexing):
*
* A B
* 0 --------------- ------
* | | | |
* | | P0 | |
* | | | |
* --------------- ------
* - fst_row->| | | |
* | | | | |
* m_loc | | P1 | |
* | | | | |
* - | | | |
* --------------- ------
* | . | |. |
* | . | |. |
* | . | |. |
* --------------- ------
*
* where, fst_row is the row number of the first row,
* m_loc is the number of rows local to this processor
* These are defined in the 'SuperMatrix' structure, see supermatrix.h.
*
*
* Here are the options for using this code:
*
* 1. Independent of all the other options specified below, the
* user must supply
*
* - B, the matrix of right-hand sides, distributed by block rows,
* and its dimensions ldb (local) and nrhs (global)
* - grid, a structure describing the 2D processor mesh
* - options->IterRefine, which determines whether or not to
* improve the accuracy of the computed solution using
* iterative refinement
*
* On output, B is overwritten with the solution X.
*
* 2. Depending on options->Fact, the user has four options
* for solving A*X=B. The standard option is for factoring
* A "from scratch". (The other options, described below,
* are used when A is sufficiently similar to a previously
* solved problem to save time by reusing part or all of
* the previous factorization.)
*
* - options->Fact = DOFACT: A is factored "from scratch"
*
* In this case the user must also supply
*
* o A, the input matrix
*
* as well as the following options to determine what matrix to
* factorize.
*
* o options->Equil, to specify how to scale the rows and columns
* of A to "equilibrate" it (to try to reduce its
* condition number and so improve the
* accuracy of the computed solution)
*
* o options->RowPerm, to specify how to permute the rows of A
* (typically to control numerical stability)
*
* o options->ColPerm, to specify how to permute the columns of A
* (typically to control fill-in and enhance
* parallelism during factorization)
*
* o options->ReplaceTinyPivot, to specify how to deal with tiny
* pivots encountered during factorization
* (to control numerical stability)
*
* The outputs returned include
*
* o ScalePermstruct, modified to describe how the input matrix A
* was equilibrated and permuted:
* . ScalePermstruct->DiagScale, indicates whether the rows and/or
* columns of A were scaled
* . ScalePermstruct->R, array of row scale factors
* . ScalePermstruct->C, array of column scale factors
* . ScalePermstruct->perm_r, row permutation vector
* . ScalePermstruct->perm_c, column permutation vector
*
* (part of ScalePermstruct may also need to be supplied on input,
* depending on options->RowPerm and options->ColPerm as described
* later).
*
* o A, the input matrix A overwritten by the scaled and permuted
* matrix diag(R)*A*diag(C)*Pc^T, where
* Pc is the row permutation matrix determined by
* ScalePermstruct->perm_c
* diag(R) and diag(C) are diagonal scaling matrices determined
* by ScalePermstruct->DiagScale, ScalePermstruct->R and
* ScalePermstruct->C
*
* o LUstruct, which contains the L and U factorization of A1 where
*
* A1 = Pc*Pr*diag(R)*A*diag(C)*Pc^T = L*U
*
* (Note that A1 = Pc*Pr*Aout, where Aout is the matrix stored
* in A on output.)
*
* 3. The second value of options->Fact assumes that a matrix with the same
* sparsity pattern as A has already been factored:
*
* - options->Fact = SamePattern: A is factored, assuming that it has
* the same nonzero pattern as a previously factored matrix. In
* this case the algorithm saves time by reusing the previously
* computed column permutation vector stored in
* ScalePermstruct->perm_c and the "elimination tree" of A
* stored in LUstruct->etree
*
* In this case the user must still specify the following options
* as before:
*
* o options->Equil
* o options->RowPerm
* o options->ReplaceTinyPivot
*
* but not options->ColPerm, whose value is ignored. This is because the
* previous column permutation from ScalePermstruct->perm_c is used as
* input. The user must also supply
*
* o A, the input matrix
* o ScalePermstruct->perm_c, the column permutation
* o LUstruct->etree, the elimination tree
*
* The outputs returned include
*
* o A, the input matrix A overwritten by the scaled and permuted
* matrix as described above
* o ScalePermstruct, modified to describe how the input matrix A was
* equilibrated and row permuted
* o LUstruct, modified to contain the new L and U factors
*
* 4. The third value of options->Fact assumes that a matrix B with the same
* sparsity pattern as A has already been factored, and where the
* row permutation of B can be reused for A. This is useful when A and B
* have similar numerical values, so that the same row permutation
* will make both factorizations numerically stable. This lets us reuse
* all of the previously computed structure of L and U.
*
* - options->Fact = SamePattern_SameRowPerm: A is factored,
* assuming not only the same nonzero pattern as the previously
* factored matrix B, but reusing B's row permutation.
*
* In this case the user must still specify the following options
* as before:
*
* o options->Equil
* o options->ReplaceTinyPivot
*
* but not options->RowPerm or options->ColPerm, whose values are
* ignored. This is because the permutations from ScalePermstruct->perm_r
* and ScalePermstruct->perm_c are used as input.
*
* The user must also supply
*
* o A, the input matrix
* o ScalePermstruct->DiagScale, how the previous matrix was row
* and/or column scaled
* o ScalePermstruct->R, the row scalings of the previous matrix,
* if any
* o ScalePermstruct->C, the columns scalings of the previous matrix,
* if any
* o ScalePermstruct->perm_r, the row permutation of the previous
* matrix
* o ScalePermstruct->perm_c, the column permutation of the previous
* matrix
* o all of LUstruct, the previously computed information about
* L and U (the actual numerical values of L and U
* stored in LUstruct->Llu are ignored)
*
* The outputs returned include
*
* o A, the input matrix A overwritten by the scaled and permuted
* matrix as described above
* o ScalePermstruct, modified to describe how the input matrix A was
* equilibrated (thus ScalePermstruct->DiagScale,
* R and C may be modified)
* o LUstruct, modified to contain the new L and U factors
*
* 5. The fourth and last value of options->Fact assumes that A is
* identical to a matrix that has already been factored on a previous
* call, and reuses its entire LU factorization
*
* - options->Fact = Factored: A is identical to a previously
* factorized matrix, so the entire previous factorization
* can be reused.
*
* In this case all the other options mentioned above are ignored
* (options->Equil, options->RowPerm, options->ColPerm,
* options->ReplaceTinyPivot)
*
* The user must also supply
*
* o A, the unfactored matrix, only in the case that iterative
* refinment is to be done (specifically A must be the output
* A from the previous call, so that it has been scaled and permuted)
* o all of ScalePermstruct
* o all of LUstruct, including the actual numerical values of
* L and U
*
* all of which are unmodified on output.
*
* Arguments
* =========
*
* options (input) superlu_options_t* (global)
* The structure defines the input parameters to control
* how the LU decomposition will be performed.
* The following fields should be defined for this structure:
*
* o Fact (fact_t)
* Specifies whether or not the factored form of the matrix
* A is supplied on entry, and if not, how the matrix A should
* be factorized based on the previous history.
*
* = DOFACT: The matrix A will be factorized from scratch.
* Inputs: A
* options->Equil, RowPerm, ColPerm, ReplaceTinyPivot
* Outputs: modified A
* (possibly row and/or column scaled and/or
* permuted)
* all of ScalePermstruct
* all of LUstruct
*
* = SamePattern: the matrix A will be factorized assuming
* that a factorization of a matrix with the same sparsity
* pattern was performed prior to this one. Therefore, this
* factorization will reuse column permutation vector
* ScalePermstruct->perm_c and the elimination tree
* LUstruct->etree
* Inputs: A
* options->Equil, RowPerm, ReplaceTinyPivot
* ScalePermstruct->perm_c
* LUstruct->etree
* Outputs: modified A
* (possibly row and/or column scaled and/or
* permuted)
* rest of ScalePermstruct (DiagScale, R, C, perm_r)
* rest of LUstruct (GLU_persist, Llu)
*
* = SamePattern_SameRowPerm: the matrix A will be factorized
* assuming that a factorization of a matrix with the same
* sparsity pattern and similar numerical values was performed
* prior to this one. Therefore, this factorization will reuse
* both row and column scaling factors R and C, and the
* both row and column permutation vectors perm_r and perm_c,
* distributed data structure set up from the previous symbolic
* factorization.
* Inputs: A
* options->Equil, ReplaceTinyPivot
* all of ScalePermstruct
* all of LUstruct
* Outputs: modified A
* (possibly row and/or column scaled and/or
* permuted)
* modified LUstruct->Llu
* = FACTORED: the matrix A is already factored.
* Inputs: all of ScalePermstruct
* all of LUstruct
*
* o Equil (yes_no_t)
* Specifies whether to equilibrate the system.
* = NO: no equilibration.
* = YES: scaling factors are computed to equilibrate the system:
* diag(R)*A*diag(C)*inv(diag(C))*X = diag(R)*B.
* Whether or not the system will be equilibrated depends
* on the scaling of the matrix A, but if equilibration is
* used, A is overwritten by diag(R)*A*diag(C) and B by
* diag(R)*B.
*
* o RowPerm (rowperm_t)
* Specifies how to permute rows of the matrix A.
* = NATURAL: use the natural ordering.
* = LargeDiag: use the Duff/Koster algorithm to permute rows of
* the original matrix to make the diagonal large
* relative to the off-diagonal.
* = MY_PERMR: use the ordering given in ScalePermstruct->perm_r
* input by the user.
*
* o ColPerm (colperm_t)
* Specifies what type of column permutation to use to reduce fill.
* = NATURAL: natural ordering.
* = MMD_AT_PLUS_A: minimum degree ordering on structure of A'+A.
* = MMD_ATA: minimum degree ordering on structure of A'*A.
* = MY_PERMC: the ordering given in ScalePermstruct->perm_c.
*
* o ReplaceTinyPivot (yes_no_t)
* = NO: do not modify pivots
* = YES: replace tiny pivots by sqrt(epsilon)*norm(A) during
* LU factorization.
*
* o IterRefine (IterRefine_t)
* Specifies how to perform iterative refinement.
* = NO: no iterative refinement.
* = DOUBLE: accumulate residual in double precision.
* = EXTRA: accumulate residual in extra precision.
*
* NOTE: all options must be indentical on all processes when
* calling this routine.
*
* A (input/output) SuperMatrix* (local)
* On entry, matrix A in A*X=B, of dimension (A->nrow, A->ncol).
* The number of linear equations is A->nrow. The type of A must be:
* Stype = SLU_NR_loc; Dtype = SLU_D; Mtype = SLU_GE.
* That is, A is stored in distributed compressed row format.
* See supermatrix.h for the definition of 'SuperMatrix'.
* This routine only handles square A, however, the LU factorization
* routine PDGSTRF can factorize rectangular matrices.
* On exit, A may be overwtirren by diag(R)*A*diag(C)*Pc^T,
* depending on ScalePermstruct->DiagScale and options->ColPerm:
* if ScalePermstruct->DiagScale != NOEQUIL, A is overwritten by
* diag(R)*A*diag(C).
* if options->ColPerm != NATURAL, A is further overwritten by
* diag(R)*A*diag(C)*Pc^T.
* If all the above condition are true, the LU decomposition is
* performed on the matrix Pc*Pr*diag(R)*A*diag(C)*Pc^T.
*
* ScalePermstruct (input/output) ScalePermstruct_t* (global)
* The data structure to store the scaling and permutation vectors
* describing the transformations performed to the matrix A.
* It contains the following fields:
*
* o DiagScale (DiagScale_t)
* Specifies the form of equilibration that was done.
* = NOEQUIL: no equilibration.
* = ROW: row equilibration, i.e., A was premultiplied by
* diag(R).
* = COL: Column equilibration, i.e., A was postmultiplied
* by diag(C).
* = BOTH: both row and column equilibration, i.e., A was
* replaced by diag(R)*A*diag(C).
* If options->Fact = FACTORED or SamePattern_SameRowPerm,
* DiagScale is an input argument; otherwise it is an output
* argument.
*
* o perm_r (int*)
* Row permutation vector, which defines the permutation matrix Pr;
* perm_r[i] = j means row i of A is in position j in Pr*A.
* If options->RowPerm = MY_PERMR, or
* options->Fact = SamePattern_SameRowPerm, perm_r is an
* input argument; otherwise it is an output argument.
*
* o perm_c (int*)
* Column permutation vector, which defines the
* permutation matrix Pc; perm_c[i] = j means column i of A is
* in position j in A*Pc.
* If options->ColPerm = MY_PERMC or options->Fact = SamePattern
* or options->Fact = SamePattern_SameRowPerm, perm_c is an
* input argument; otherwise, it is an output argument.
* On exit, perm_c may be overwritten by the product of the input
* perm_c and a permutation that postorders the elimination tree
* of Pc*A'*A*Pc'; perm_c is not changed if the elimination tree
* is already in postorder.
*
* o R (double*) dimension (A->nrow)
* The row scale factors for A.
* If DiagScale = ROW or BOTH, A is multiplied on the left by
* diag(R).
* If DiagScale = NOEQUIL or COL, R is not defined.
* If options->Fact = FACTORED or SamePattern_SameRowPerm, R is
* an input argument; otherwise, R is an output argument.
*
* o C (double*) dimension (A->ncol)
* The column scale factors for A.
* If DiagScale = COL or BOTH, A is multiplied on the right by
* diag(C).
* If DiagScale = NOEQUIL or ROW, C is not defined.
* If options->Fact = FACTORED or SamePattern_SameRowPerm, C is
* an input argument; otherwise, C is an output argument.
*
* B (input/output) double* (local)
* On entry, the right-hand side matrix of dimension (m_loc, nrhs),
* where, m_loc is the number of rows stored locally on my
* process and is defined in the data structure of matrix A.
* On exit, the solution matrix if info = 0;
*
* ldb (input) int (local)
* The leading dimension of matrix B.
*
* nrhs (input) int (global)
* The number of right-hand sides.
* If nrhs = 0, only LU decomposition is performed, the forward
* and back substitutions are skipped.
*
* grid (input) gridinfo_t* (global)
* The 2D process mesh. It contains the MPI communicator, the number
* of process rows (NPROW), the number of process columns (NPCOL),
* and my process rank. It is an input argument to all the
* parallel routines.
* Grid can be initialized by subroutine SUPERLU_GRIDINIT.
* See superlu_ddefs.h for the definition of 'gridinfo_t'.
*
* LUstruct (input/output) LUstruct_t*
* The data structures to store the distributed L and U factors.
* It contains the following fields:
*
* o etree (int*) dimension (A->ncol) (global)
* Elimination tree of Pc*(A'+A)*Pc' or Pc*A'*A*Pc'.
* It is computed in sp_colorder() during the first factorization,
* and is reused in the subsequent factorizations of the matrices
* with the same nonzero pattern.
* On exit of sp_colorder(), the columns of A are permuted so that
* the etree is in a certain postorder. This postorder is reflected
* in ScalePermstruct->perm_c.
* NOTE:
* Etree is a vector of parent pointers for a forest whose vertices
* are the integers 0 to A->ncol-1; etree[root]==A->ncol.
*
* o Glu_persist (Glu_persist_t*) (global)
* Global data structure (xsup, supno) replicated on all processes,
* describing the supernode partition in the factored matrices
* L and U:
* xsup[s] is the leading column of the s-th supernode,
* supno[i] is the supernode number to which column i belongs.
*
* o Llu (LocalLU_t*) (local)
* The distributed data structures to store L and U factors.
* See superlu_ddefs.h for the definition of 'LocalLU_t'.
*
* SOLVEstruct (input/output) SOLVEstruct_t*
* The data structure to hold the communication pattern used
* in the phases of triangular solution and iterative refinement.
* This pattern should be intialized only once for repeated solutions.
* If options->SolveInitialized = YES, it is an input argument.
* If options->SolveInitialized = NO and nrhs != 0, it is an output
* argument. See superlu_ddefs.h for the definition of 'SOLVEstruct_t'.
*
* berr (output) double*, dimension (nrhs) (global)
* The componentwise relative backward error of each solution
* vector X(j) (i.e., the smallest relative change in
* any element of A or B that makes X(j) an exact solution).
*
* stat (output) SuperLUStat_t*
* Record the statistics on runtime and floating-point operation count.
* See util.h for the definition of 'SuperLUStat_t'.
*
* info (output) int*
* = 0: successful exit
* > 0: if info = i, and i is
* <= A->ncol: U(i,i) is exactly zero. The factorization has
* been completed, but the factor U is exactly singular,
* so the solution could not be computed.
* > A->ncol: number of bytes allocated when memory allocation
* failure occurred, plus A->ncol.
*
* See superlu_ddefs.h for the definitions of varioous data types.
*
*/
NRformat_loc *Astore;
SuperMatrix GA; /* Global A in NC format */
NCformat *GAstore;
double *a_GA;
SuperMatrix GAC; /* Global A in NCP format (add n end pointers) */
NCPformat *GACstore;
Glu_persist_t *Glu_persist = LUstruct->Glu_persist;
Glu_freeable_t *Glu_freeable;
/* The nonzero structures of L and U factors, which are
replicated on all processrs.
(lsub, xlsub) contains the compressed subscript of
supernodes in L.
(usub, xusub) contains the compressed subscript of
nonzero segments in U.
If options->Fact != SamePattern_SameRowPerm, they are
computed by SYMBFACT routine, and then used by PDDISTRIBUTE
routine. They will be freed after PDDISTRIBUTE routine.
If options->Fact == SamePattern_SameRowPerm, these
structures are not used. */
fact_t Fact;
double *a;
int_t *colptr, *rowind;
int_t *perm_r; /* row permutations from partial pivoting */
int_t *perm_c; /* column permutation vector */
int_t *etree; /* elimination tree */
int_t *rowptr, *colind; /* Local A in NR*/
int_t *rowind_loc, *colptr_loc;
int_t colequ, Equil, factored, job, notran, rowequ, need_value;
int_t i, iinfo, j, irow, m, n, nnz, permc_spec, dist_mem_use;
int_t nnz_loc, m_loc, fst_row, icol;
int iam;
int ldx; /* LDA for matrix X (local). */
char equed[1], norm[1];
double *C, *R, *C1, *R1, amax, anorm, colcnd, rowcnd;
double *X, *b_col, *b_work, *x_col;
double t;
static mem_usage_t num_mem_usage, symb_mem_usage;
#if ( PRNTlevel>= 2 )
double dmin, dsum, dprod;
#endif
int_t procs;
/* Structures needed for parallel symbolic factorization */
int_t *sizes, *fstVtxSep, parSymbFact;
int noDomains, nprocs_num;
MPI_Comm symb_comm; /* communicator for symbolic factorization */
int col, key; /* parameters for creating a new communicator */
Pslu_freeable_t Pslu_freeable;
float flinfo;
/* Initialization. */
m = A->nrow;
n = A->ncol;
Astore = (NRformat_loc *) A->Store;
nnz_loc = Astore->nnz_loc;
m_loc = Astore->m_loc;
fst_row = Astore->fst_row;
a = (double *) Astore->nzval;
rowptr = Astore->rowptr;
colind = Astore->colind;
sizes = NULL;
fstVtxSep = NULL;
symb_comm = MPI_COMM_NULL;
/* Test the input parameters. */
*info = 0;
Fact = options->Fact;
if ( Fact < 0 || Fact > FACTORED )
*info = -1;
else if ( options->RowPerm < 0 || options->RowPerm > MY_PERMR )
*info = -1;
else if ( options->ColPerm < 0 || options->ColPerm > MY_PERMC )
*info = -1;
else if ( options->IterRefine < 0 || options->IterRefine > EXTRA )
*info = -1;
else if ( options->IterRefine == EXTRA ) {
*info = -1;
fprintf(stderr, "Extra precise iterative refinement yet to support.");
} else if ( A->nrow != A->ncol || A->nrow < 0 || A->Stype != SLU_NR_loc
|| A->Dtype != SLU_D || A->Mtype != SLU_GE )
*info = -2;
else if ( ldb < m_loc )
*info = -5;
else if ( nrhs < 0 )
*info = -6;
if ( *info ) {
i = -(*info);
pxerbla("pdgssvx", grid, -*info);
return;
}
factored = (Fact == FACTORED);
Equil = (!factored && options->Equil == YES);
notran = (options->Trans == NOTRANS);
iam = grid->iam;
job = 5;
if ( factored || (Fact == SamePattern_SameRowPerm && Equil) ) {
rowequ = (ScalePermstruct->DiagScale == ROW) ||
(ScalePermstruct->DiagScale == BOTH);
colequ = (ScalePermstruct->DiagScale == COL) ||
(ScalePermstruct->DiagScale == BOTH);
} else rowequ = colequ = FALSE;
/* The following arrays are replicated on all processes. */
perm_r = ScalePermstruct->perm_r;
perm_c = ScalePermstruct->perm_c;
etree = LUstruct->etree;
R = ScalePermstruct->R;
C = ScalePermstruct->C;
/********/
#if ( DEBUGlevel>=1 )
CHECK_MALLOC(iam, "Enter pdgssvx()");
#endif
/* Not factored & ask for equilibration */
if ( Equil && Fact != SamePattern_SameRowPerm ) {
/* Allocate storage if not done so before. */
switch ( ScalePermstruct->DiagScale ) {
case NOEQUIL:
if ( !(R = (double *) doubleMalloc_dist(m)) )
ABORT("Malloc fails for R[].");
if ( !(C = (double *) doubleMalloc_dist(n)) )
ABORT("Malloc fails for C[].");
ScalePermstruct->R = R;
ScalePermstruct->C = C;
break;
case ROW:
if ( !(C = (double *) doubleMalloc_dist(n)) )
ABORT("Malloc fails for C[].");
ScalePermstruct->C = C;
break;
case COL:
if ( !(R = (double *) doubleMalloc_dist(m)) )
ABORT("Malloc fails for R[].");
ScalePermstruct->R = R;
break;
}
}
/* ------------------------------------------------------------
Diagonal scaling to equilibrate the matrix.
------------------------------------------------------------*/
if ( Equil ) {
#if ( DEBUGlevel>=1 )
CHECK_MALLOC(iam, "Enter equil");
#endif
t = SuperLU_timer_();
if ( Fact == SamePattern_SameRowPerm ) {
/* Reuse R and C. */
switch ( ScalePermstruct->DiagScale ) {
case NOEQUIL:
break;
case ROW:
irow = fst_row;
for (j = 0; j < m_loc; ++j) {
for (i = rowptr[j]; i < rowptr[j+1]; ++i) {
a[i] *= R[irow]; /* Scale rows. */
}
++irow;
}
break;
case COL:
for (j = 0; j < m_loc; ++j)
for (i = rowptr[j]; i < rowptr[j+1]; ++i){
icol = colind[i];
a[i] *= C[icol]; /* Scale columns. */
}
break;
case BOTH:
irow = fst_row;
for (j = 0; j < m_loc; ++j) {
for (i = rowptr[j]; i < rowptr[j+1]; ++i) {
icol = colind[i];
a[i] *= R[irow] * C[icol]; /* Scale rows and cols. */
}
++irow;
}
break;
}
} else { /* Compute R & C from scratch */
/* Compute the row and column scalings. */
pdgsequ(A, R, C, &rowcnd, &colcnd, &amax, &iinfo, grid);
/* Equilibrate matrix A if it is badly-scaled. */
pdlaqgs(A, R, C, rowcnd, colcnd, amax, equed);
if ( lsame_(equed, "R") ) {
ScalePermstruct->DiagScale = rowequ = ROW;
} else if ( lsame_(equed, "C") ) {
ScalePermstruct->DiagScale = colequ = COL;
} else if ( lsame_(equed, "B") ) {
ScalePermstruct->DiagScale = BOTH;
rowequ = ROW;
colequ = COL;
} else ScalePermstruct->DiagScale = NOEQUIL;
#if ( PRNTlevel>=1 )
if ( !iam ) {
printf(".. equilibrated? *equed = %c\n", *equed);
/*fflush(stdout);*/
}
#endif
} /* if Fact ... */
stat->utime[EQUIL] = SuperLU_timer_() - t;
#if ( DEBUGlevel>=1 )
CHECK_MALLOC(iam, "Exit equil");
#endif
} /* if Equil ... */
if ( !factored ) { /* Skip this if already factored. */
/*
* Gather A from the distributed compressed row format to
* global A in compressed column format.
* Numerical values are gathered only when a row permutation
* for large diagonal is sought after.
*/
if ( Fact != SamePattern_SameRowPerm ) {
need_value = (options->RowPerm == LargeDiag);
pdCompRow_loc_to_CompCol_global(need_value, A, grid, &GA);
GAstore = (NCformat *) GA.Store;
colptr = GAstore->colptr;
rowind = GAstore->rowind;
nnz = GAstore->nnz;
if ( need_value ) a_GA = (double *) GAstore->nzval;
else assert(GAstore->nzval == NULL);
}
/* ------------------------------------------------------------
Find the row permutation for A.
------------------------------------------------------------*/
if ( options->RowPerm != NO ) {
t = SuperLU_timer_();
if ( Fact != SamePattern_SameRowPerm ) {
if ( options->RowPerm == MY_PERMR ) { /* Use user's perm_r. */
/* Permute the global matrix GA for symbfact() */
for (i = 0; i < colptr[n]; ++i) {
irow = rowind[i];
rowind[i] = perm_r[irow];
}
} else { /* options->RowPerm == LargeDiag */
/* Get a new perm_r[] */
if ( job == 5 ) {
/* Allocate storage for scaling factors. */
if ( !(R1 = doubleMalloc_dist(m)) )
ABORT("SUPERLU_MALLOC fails for R1[]");
if ( !(C1 = doubleMalloc_dist(n)) )
ABORT("SUPERLU_MALLOC fails for C1[]");
}
if ( !iam ) {
/* Process 0 finds a row permutation */
dldperm(job, m, nnz, colptr, rowind, a_GA,
perm_r, R1, C1);
MPI_Bcast( perm_r, m, mpi_int_t, 0, grid->comm );
if ( job == 5 && Equil ) {
MPI_Bcast( R1, m, MPI_DOUBLE, 0, grid->comm );
MPI_Bcast( C1, n, MPI_DOUBLE, 0, grid->comm );
}
} else {
MPI_Bcast( perm_r, m, mpi_int_t, 0, grid->comm );
if ( job == 5 && Equil ) {
MPI_Bcast( R1, m, MPI_DOUBLE, 0, grid->comm );
MPI_Bcast( C1, n, MPI_DOUBLE, 0, grid->comm );
}
}
#if ( PRNTlevel>=2 )
dmin = dlamch_("Overflow");
dsum = 0.0;
dprod = 1.0;
#endif
if ( job == 5 ) {
if ( Equil ) {
for (i = 0; i < n; ++i) {
R1[i] = exp(R1[i]);
C1[i] = exp(C1[i]);
}
/* Scale the distributed matrix */
irow = fst_row;
for (j = 0; j < m_loc; ++j) {
for (i = rowptr[j]; i < rowptr[j+1]; ++i) {
icol = colind[i];
a[i] *= R1[irow] * C1[icol];
#if ( PRNTlevel>=2 )
if ( perm_r[irow] == icol ) { /* New diagonal */
if ( job == 2 || job == 3 )
dmin = SUPERLU_MIN(dmin, fabs(a[i]));
else if ( job == 4 )
dsum += fabs(a[i]);
else if ( job == 5 )
dprod *= fabs(a[i]);
}
#endif
}
++irow;
}
/* Multiply together the scaling factors. */
if ( rowequ ) for (i = 0; i < m; ++i) R[i] *= R1[i];
else for (i = 0; i < m; ++i) R[i] = R1[i];
if ( colequ ) for (i = 0; i < n; ++i) C[i] *= C1[i];
else for (i = 0; i < n; ++i) C[i] = C1[i];
ScalePermstruct->DiagScale = BOTH;
rowequ = colequ = 1;
} /* end Equil */
/* Now permute global A to prepare for symbfact() */
for (j = 0; j < n; ++j) {
for (i = colptr[j]; i < colptr[j+1]; ++i) {
irow = rowind[i];
rowind[i] = perm_r[irow];
}
}
SUPERLU_FREE (R1);
SUPERLU_FREE (C1);
} else { /* job = 2,3,4 */
for (j = 0; j < n; ++j) {
for (i = colptr[j]; i < colptr[j+1]; ++i) {
irow = rowind[i];
rowind[i] = perm_r[irow];
} /* end for i ... */
} /* end for j ... */
} /* end else job ... */
#if ( PRNTlevel>=2 )
if ( job == 2 || job == 3 ) {
if ( !iam ) printf("\tsmallest diagonal %e\n", dmin);
} else if ( job == 4 ) {
if ( !iam ) printf("\tsum of diagonal %e\n", dsum);
} else if ( job == 5 ) {
if ( !iam ) printf("\t product of diagonal %e\n", dprod);
}
#endif
} /* end if options->RowPerm ... */
t = SuperLU_timer_() - t;
stat->utime[ROWPERM] = t;
#if ( PRNTlevel>=1 )
if ( !iam ) printf(".. LDPERM job %d\t time: %.2f\n", job, t);
#endif
} /* end if Fact ... */
} else { /* options->RowPerm == NOROWPERM */
for (i = 0; i < m; ++i) perm_r[i] = i;
}
#if ( DEBUGlevel>=2 )
if ( !iam ) PrintInt10("perm_r", m, perm_r);
#endif
} /* end if (!factored) */
if ( !factored || options->IterRefine ) {
/* Compute norm(A), which will be used to adjust small diagonal. */
if ( notran ) *(unsigned char *)norm = '1';
else *(unsigned char *)norm = 'I';
anorm = pdlangs(norm, A, grid);
#if ( PRNTlevel>=1 )
if ( !iam ) printf(".. anorm %e\n", anorm);
#endif
}
/* ------------------------------------------------------------
Perform the LU factorization.
------------------------------------------------------------*/
if ( !factored ) {
t = SuperLU_timer_();
/*
* Get column permutation vector perm_c[], according to permc_spec:
* permc_spec = NATURAL: natural ordering
* permc_spec = MMD_AT_PLUS_A: minimum degree on structure of A'+A
* permc_spec = MMD_ATA: minimum degree on structure of A'*A
* permc_spec = METIS_AT_PLUS_A: METIS on structure of A'+A
* permc_spec = PARMETIS: parallel METIS on structure of A'+A
* permc_spec = MY_PERMC: the ordering already supplied in perm_c[]
*/
permc_spec = options->ColPerm;
parSymbFact = options->ParSymbFact;
#if ( PRNTlevel>=1 )
if ( parSymbFact && permc_spec != PARMETIS )
if ( !iam ) printf(".. Parallel symbolic factorization"
" only works wth ParMetis!\n");
#endif
if ( parSymbFact == YES || permc_spec == PARMETIS ) {
nprocs_num = grid->nprow * grid->npcol;
noDomains = (int) ( pow(2, ((int) LOG2( nprocs_num ))));
/* create a new communicator for the first noDomains processors in
grid->comm */
key = iam;
if (iam < noDomains) col = 0;
else col = MPI_UNDEFINED;
MPI_Comm_split (grid->comm, col, key, &symb_comm );
permc_spec = PARMETIS; /* only works with PARMETIS */
}
if ( permc_spec != MY_PERMC && Fact == DOFACT ) {
if ( permc_spec == PARMETIS ) {
/* Get column permutation vector in perm_c. *
* This routine takes as input the distributed input matrix A *
* and does not modify it. It also allocates memory for *
* sizes[] and fstVtxSep[] arrays, that contain information *
* on the separator tree computed by ParMETIS. */
flinfo = get_perm_c_parmetis(A, perm_r, perm_c, nprocs_num,
noDomains, &sizes, &fstVtxSep,
grid, &symb_comm);
if (flinfo > 0)
ABORT("ERROR in get perm_c parmetis.");
} else {
get_perm_c_dist(iam, permc_spec, &GA, perm_c);
}
}
stat->utime[COLPERM] = SuperLU_timer_() - t;
/* Compute the elimination tree of Pc*(A'+A)*Pc' or Pc*A'*A*Pc'
(a.k.a. column etree), depending on the choice of ColPerm.
Adjust perm_c[] to be consistent with a postorder of etree.
Permute columns of A to form A*Pc'. */
if ( Fact != SamePattern_SameRowPerm ) {
if ( parSymbFact == NO ) {
int_t *GACcolbeg, *GACcolend, *GACrowind;
sp_colorder(options, &GA, perm_c, etree, &GAC);
/* Form Pc*A*Pc' to preserve the diagonal of the matrix GAC. */
GACstore = (NCPformat *) GAC.Store;
GACcolbeg = GACstore->colbeg;
GACcolend = GACstore->colend;
GACrowind = GACstore->rowind;
for (j = 0; j < n; ++j) {
for (i = GACcolbeg[j]; i < GACcolend[j]; ++i) {
irow = GACrowind[i];
GACrowind[i] = perm_c[irow];
}
}
/* Perform a symbolic factorization on Pc*Pr*A*Pc' and set up
the nonzero data structures for L & U. */
#if ( PRNTlevel>=1 )
if ( !iam )
printf(".. symbfact(): relax %4d, maxsuper %4d, fill %4d\n",
sp_ienv_dist(2), sp_ienv_dist(3), sp_ienv_dist(6));
#endif
t = SuperLU_timer_();
if ( !(Glu_freeable = (Glu_freeable_t *)
SUPERLU_MALLOC(sizeof(Glu_freeable_t))) )
ABORT("Malloc fails for Glu_freeable.");
/* Every process does this. */
iinfo = symbfact(options, iam, &GAC, perm_c, etree,
Glu_persist, Glu_freeable);
stat->utime[SYMBFAC] = SuperLU_timer_() - t;
if ( iinfo < 0 ) { /* Successful return */
QuerySpace_dist(n, -iinfo, Glu_freeable, &symb_mem_usage);
#if ( PRNTlevel>=1 )
if ( !iam ) {
printf("\tNo of supers %ld\n", Glu_persist->supno[n-1]+1);
printf("\tSize of G(L) %ld\n", Glu_freeable->xlsub[n]);
printf("\tSize of G(U) %ld\n", Glu_freeable->xusub[n]);
printf("\tint %d, short %d, float %d, double %d\n",
sizeof(int_t), sizeof(short), sizeof(float),
sizeof(double));
printf("\tSYMBfact (MB):\tL\\U %.2f\ttotal %.2f\texpansions %d\n",
symb_mem_usage.for_lu*1e-6,
symb_mem_usage.total*1e-6,
symb_mem_usage.expansions);
}
#endif
} else {
if ( !iam ) {
fprintf(stderr,"symbfact() error returns %d\n",iinfo);
exit(-1);
}
}
} /* end if serial symbolic factorization */
else { /* parallel symbolic factorization */
t = SuperLU_timer_();
flinfo = symbfact_dist(nprocs_num, noDomains, A, perm_c, perm_r,
sizes, fstVtxSep, &Pslu_freeable,
&(grid->comm), &symb_comm,
&symb_mem_usage);
stat->utime[SYMBFAC] = SuperLU_timer_() - t;
if (flinfo > 0)
ABORT("Insufficient memory for parallel symbolic factorization.");
}
} /* end if Fact ... */
#if ( PRNTlevel>=1 )
if (!iam) printf("\tSYMBfact time: %.2f\n", stat->utime[SYMBFAC]);
#endif
if (sizes) SUPERLU_FREE (sizes);
if (fstVtxSep) SUPERLU_FREE (fstVtxSep);
if (symb_comm != MPI_COMM_NULL)
MPI_Comm_free (&symb_comm);
if (parSymbFact == NO || Fact == SamePattern_SameRowPerm) {
/* Apply column permutation to the original distributed A */
for (j = 0; j < nnz_loc; ++j) colind[j] = perm_c[colind[j]];
/* Distribute Pc*Pr*diag(R)*A*diag(C)*Pc' into L and U storage.
NOTE: the row permutation Pc*Pr is applied internally in the
distribution routine. */
t = SuperLU_timer_();
dist_mem_use = pddistribute(Fact, n, A, ScalePermstruct,
Glu_freeable, LUstruct, grid);
stat->utime[DIST] = SuperLU_timer_() - t;
/* Deallocate storage used in symbolic factorization. */
if ( Fact != SamePattern_SameRowPerm ) {
iinfo = symbfact_SubFree(Glu_freeable);
SUPERLU_FREE(Glu_freeable);
}
} else {
/* Distribute Pc*Pr*diag(R)*A*diag(C)*Pc' into L and U storage.
NOTE: the row permutation Pc*Pr is applied internally in the
distribution routine. */
/* Apply column permutation to the original distributed A */
for (j = 0; j < nnz_loc; ++j) colind[j] = perm_c[colind[j]];
t = SuperLU_timer_();
dist_mem_use = ddist_psymbtonum(Fact, n, A, ScalePermstruct,
&Pslu_freeable, LUstruct, grid);
if (dist_mem_use > 0)
ABORT ("Not enough memory available for dist_psymbtonum\n");
stat->utime[DIST] = SuperLU_timer_() - t;
}
#if ( PRNTlevel>=1 )
if (!iam) printf ("\tDISTRIBUTE time %8.2f\n", stat->utime[DIST]);
#endif
/* Perform numerical factorization in parallel. */
t = SuperLU_timer_();
pdgstrf(options, m, n, anorm, LUstruct, grid, stat, info);
stat->utime[FACT] = SuperLU_timer_() - t;
#if ( PRNTlevel>=1 )
{
int_t TinyPivots;
float for_lu, total, max, avg, temp;
dQuerySpace_dist(n, LUstruct, grid, &num_mem_usage);
MPI_Reduce( &num_mem_usage.for_lu, &for_lu,
1, MPI_FLOAT, MPI_SUM, 0, grid->comm );
MPI_Reduce( &num_mem_usage.total, &total,
1, MPI_FLOAT, MPI_SUM, 0, grid->comm );
temp = SUPERLU_MAX(symb_mem_usage.total,
symb_mem_usage.for_lu +
(float)dist_mem_use + num_mem_usage.for_lu);
if (parSymbFact == TRUE)
/* The memory used in the redistribution routine
includes the memory used for storing the symbolic
structure and the memory allocated for numerical
factorization */
temp = SUPERLU_MAX(symb_mem_usage.total,
(float)dist_mem_use);
temp = SUPERLU_MAX(temp, num_mem_usage.total);
MPI_Reduce( &temp, &max,
1, MPI_FLOAT, MPI_MAX, 0, grid->comm );
MPI_Reduce( &temp, &avg,
1, MPI_FLOAT, MPI_SUM, 0, grid->comm );
MPI_Allreduce( &stat->TinyPivots, &TinyPivots, 1, mpi_int_t,
MPI_SUM, grid->comm );
stat->TinyPivots = TinyPivots;
if ( !iam ) {
printf("\tNUMfact (MB) all PEs:\tL\\U\t%.2f\tall\t%.2f\n",
for_lu*1e-6, total*1e-6);
printf("\tAll space (MB):"
"\t\ttotal\t%.2f\tAvg\t%.2f\tMax\t%.2f\n",
avg*1e-6, avg/grid->nprow/grid->npcol*1e-6, max*1e-6);
printf("\tNumber of tiny pivots: %10d\n", stat->TinyPivots);
}
}
#endif
/* Destroy GA */
if ( Fact != SamePattern_SameRowPerm )
Destroy_CompCol_Matrix_dist(&GA);
} /* end if (!factored) */
/* ------------------------------------------------------------
Compute the solution matrix X.
------------------------------------------------------------*/
if ( nrhs ) {
if ( !(b_work = doubleMalloc_dist(n)) )
ABORT("Malloc fails for b_work[]");
/* ------------------------------------------------------------
Scale the right-hand side if equilibration was performed.
------------------------------------------------------------*/
if ( notran ) {
if ( rowequ ) {
b_col = B;
for (j = 0; j < nrhs; ++j) {
irow = fst_row;
for (i = 0; i < m_loc; ++i) {
b_col[i] *= R[irow];
++irow;
}
b_col += ldb;
}
}
} else if ( colequ ) {
b_col = B;
for (j = 0; j < nrhs; ++j) {
irow = fst_row;
for (i = 0; i < m_loc; ++i) {
b_col[i] *= C[irow];
++irow;
}
b_col += ldb;
}
}
/* Save a copy of the right-hand side. */
ldx = ldb;
if ( !(X = doubleMalloc_dist(((size_t)ldx) * nrhs)) )
ABORT("Malloc fails for X[]");
x_col = X; b_col = B;
for (j = 0; j < nrhs; ++j) {
for (i = 0; i < m_loc; ++i) x_col[i] = b_col[i];
x_col += ldx; b_col += ldb;
}
/* ------------------------------------------------------------
Solve the linear system.
------------------------------------------------------------*/
if ( options->SolveInitialized == NO ) {
dSolveInit(options, A, perm_r, perm_c, nrhs, LUstruct, grid,
SOLVEstruct);
}
pdgstrs(n, LUstruct, ScalePermstruct, grid, X, m_loc,
fst_row, ldb, nrhs, SOLVEstruct, stat, info);
/* ------------------------------------------------------------
Use iterative refinement to improve the computed solution and
compute error bounds and backward error estimates for it.
------------------------------------------------------------*/
if ( options->IterRefine ) {
/* Improve the solution by iterative refinement. */
int_t *it, *colind_gsmv = SOLVEstruct->A_colind_gsmv;
SOLVEstruct_t *SOLVEstruct1; /* Used by refinement. */
t = SuperLU_timer_();
if ( options->RefineInitialized == NO || Fact == DOFACT ) {
/* All these cases need to re-initialize gsmv structure */
if ( options->RefineInitialized )
pdgsmv_finalize(SOLVEstruct->gsmv_comm);
pdgsmv_init(A, SOLVEstruct->row_to_proc, grid,
SOLVEstruct->gsmv_comm);
/* Save a copy of the transformed local col indices
in colind_gsmv[]. */
if ( colind_gsmv ) SUPERLU_FREE(colind_gsmv);
if ( !(it = intMalloc_dist(nnz_loc)) )
ABORT("Malloc fails for colind_gsmv[]");
colind_gsmv = SOLVEstruct->A_colind_gsmv = it;
for (i = 0; i < nnz_loc; ++i) colind_gsmv[i] = colind[i];
options->RefineInitialized = YES;
} else if ( Fact == SamePattern ||
Fact == SamePattern_SameRowPerm ) {
double at;
int_t k, jcol, p;
/* Swap to beginning the part of A corresponding to the
local part of X, as was done in pdgsmv_init() */
for (i = 0; i < m_loc; ++i) { /* Loop through each row */
k = rowptr[i];
for (j = rowptr[i]; j < rowptr[i+1]; ++j) {
jcol = colind[j];
p = SOLVEstruct->row_to_proc[jcol];
if ( p == iam ) { /* Local */
at = a[k]; a[k] = a[j]; a[j] = at;
++k;
}
}
}
/* Re-use the local col indices of A obtained from the
previous call to pdgsmv_init() */
for (i = 0; i < nnz_loc; ++i) colind[i] = colind_gsmv[i];
}
if ( nrhs == 1 ) { /* Use the existing solve structure */
SOLVEstruct1 = SOLVEstruct;
} else { /* For nrhs > 1, since refinement is performed for RHS
one at a time, the communication structure for pdgstrs
is different than the solve with nrhs RHS.
So we use SOLVEstruct1 for the refinement step.
*/
if ( !(SOLVEstruct1 = (SOLVEstruct_t *)
SUPERLU_MALLOC(sizeof(SOLVEstruct_t))) )
ABORT("Malloc fails for SOLVEstruct1");
/* Copy the same stuff */
SOLVEstruct1->row_to_proc = SOLVEstruct->row_to_proc;
SOLVEstruct1->inv_perm_c = SOLVEstruct->inv_perm_c;
SOLVEstruct1->num_diag_procs = SOLVEstruct->num_diag_procs;
SOLVEstruct1->diag_procs = SOLVEstruct->diag_procs;
SOLVEstruct1->diag_len = SOLVEstruct->diag_len;
SOLVEstruct1->gsmv_comm = SOLVEstruct->gsmv_comm;
SOLVEstruct1->A_colind_gsmv = SOLVEstruct->A_colind_gsmv;
/* Initialize the *gstrs_comm for 1 RHS. */
if ( !(SOLVEstruct1->gstrs_comm = (pxgstrs_comm_t *)
SUPERLU_MALLOC(sizeof(pxgstrs_comm_t))) )
ABORT("Malloc fails for gstrs_comm[]");
pxgstrs_init(n, m_loc, 1, fst_row, perm_r, perm_c, grid,
Glu_persist, SOLVEstruct1);
}
pdgsrfs(n, A, anorm, LUstruct, ScalePermstruct, grid,
B, ldb, X, ldx, nrhs, SOLVEstruct1, berr, stat, info);
/* Deallocate the storage associated with SOLVEstruct1 */
if ( nrhs > 1 ) {
pxgstrs_finalize(SOLVEstruct1->gstrs_comm);
SUPERLU_FREE(SOLVEstruct1);
}
stat->utime[REFINE] = SuperLU_timer_() - t;
}
/* Permute the solution matrix B <= Pc'*X. */
pdPermute_Dense_Matrix(fst_row, m_loc, SOLVEstruct->row_to_proc,
SOLVEstruct->inv_perm_c,
X, ldx, B, ldb, nrhs, grid);
#if ( DEBUGlevel>=2 )
printf("\n (%d) .. After pdPermute_Dense_Matrix(): b =\n", iam);
for (i = 0; i < m_loc; ++i)
printf("\t(%d)\t%4d\t%.10f\n", iam, i+fst_row, B[i]);
#endif
/* Transform the solution matrix X to a solution of the original
system before the equilibration. */
if ( notran ) {
if ( colequ ) {
b_col = B;
for (j = 0; j < nrhs; ++j) {
irow = fst_row;
for (i = 0; i < m_loc; ++i) {
b_col[i] *= C[irow];
++irow;
}
b_col += ldb;
}
}
} else if ( rowequ ) {
b_col = B;
for (j = 0; j < nrhs; ++j) {
irow = fst_row;
for (i = 0; i < m_loc; ++i) {
b_col[i] *= R[irow];
++irow;
}
b_col += ldb;
}
}
SUPERLU_FREE(b_work);
SUPERLU_FREE(X);
} /* end if nrhs != 0 */
#if ( PRNTlevel>=1 )
if ( !iam ) printf(".. DiagScale = %d\n", ScalePermstruct->DiagScale);
#endif
/* Deallocate R and/or C if it was not used. */
if ( Equil && Fact != SamePattern_SameRowPerm ) {
switch ( ScalePermstruct->DiagScale ) {
case NOEQUIL:
SUPERLU_FREE(R);
SUPERLU_FREE(C);
break;
case ROW:
SUPERLU_FREE(C);
break;
case COL:
SUPERLU_FREE(R);
break;
}
}
if ( !factored && Fact != SamePattern_SameRowPerm && !parSymbFact)
Destroy_CompCol_Permuted_dist(&GAC);
#if ( DEBUGlevel>=1 )
CHECK_MALLOC(iam, "Exit pdgssvx()");
#endif
}
int main(int argc, char *argv[])
{
superlu_dist_options_t options;
SuperLUStat_t stat;
SuperMatrix A;
ScalePermstruct_t ScalePermstruct;
LUstruct_t LUstruct;
gridinfo_t grid;
double *berr;
doublecomplex *a, *b, *xtrue;
int_t *asub, *xa;
int_t m, n, nnz;
int_t nprow, npcol;
int iam, info, ldb, ldx, nrhs;
char trans[1];
char **cpp, c;
FILE *fp, *fopen();
extern int cpp_defs();
/* prototypes */
extern void LUstructInit(const int_t, LUstruct_t *);
extern void LUstructFree(LUstruct_t *);
extern void Destroy_LU(int_t, gridinfo_t *, LUstruct_t *);
nprow = 1; /* Default process rows. */
npcol = 1; /* Default process columns. */
nrhs = 1; /* Number of right-hand side. */
/* ------------------------------------------------------------
INITIALIZE MPI ENVIRONMENT.
------------------------------------------------------------*/
MPI_Init( &argc, &argv );
/* Parse command line argv[]. */
for (cpp = argv+1; *cpp; ++cpp) {
if ( **cpp == '-' ) {
c = *(*cpp+1);
++cpp;
switch (c) {
case 'h':
printf("Options:\n");
printf("\t-r <int>: process rows (default " IFMT ")\n", nprow);
printf("\t-c <int>: process columns (default " IFMT ")\n", npcol);
exit(0);
break;
case 'r': nprow = atoi(*cpp);
break;
case 'c': npcol = atoi(*cpp);
break;
}
} else { /* Last arg is considered a filename */
if ( !(fp = fopen(*cpp, "r")) ) {
ABORT("File does not exist");
}
break;
}
}
/* ------------------------------------------------------------
INITIALIZE THE SUPERLU PROCESS GRID.
------------------------------------------------------------*/
superlu_gridinit(MPI_COMM_WORLD, nprow, npcol, &grid);
/* Bail out if I do not belong in the grid. */
iam = grid.iam;
if ( iam >= nprow * npcol )
goto out;
#if ( DEBUGlevel>=1 )
CHECK_MALLOC(iam, "Enter main()");
#endif
/* ------------------------------------------------------------
PROCESS 0 READS THE MATRIX A, AND THEN BROADCASTS IT TO ALL
THE OTHER PROCESSES.
------------------------------------------------------------*/
if ( !iam ) {
/* Print the CPP definitions. */
cpp_defs();
/* Read the matrix stored on disk in Harwell-Boeing format. */
zreadhb_dist(iam, fp, &m, &n, &nnz, &a, &asub, &xa);
printf("Input matrix file: %s\n", *cpp);
printf("\tDimension\t" IFMT "x" IFMT "\t # nonzeros " IFMT "\n", m, n, nnz);
printf("\tProcess grid\t%d X %d\n", (int) grid.nprow, (int) grid.npcol);
/* Broadcast matrix A to the other PEs. */
MPI_Bcast( &m, 1, mpi_int_t, 0, grid.comm );
MPI_Bcast( &n, 1, mpi_int_t, 0, grid.comm );
MPI_Bcast( &nnz, 1, mpi_int_t, 0, grid.comm );
MPI_Bcast( a, nnz, SuperLU_MPI_DOUBLE_COMPLEX, 0, grid.comm );
MPI_Bcast( asub, nnz, mpi_int_t, 0, grid.comm );
MPI_Bcast( xa, n+1, mpi_int_t, 0, grid.comm );
} else {
/* Receive matrix A from PE 0. */
MPI_Bcast( &m, 1, mpi_int_t, 0, grid.comm );
MPI_Bcast( &n, 1, mpi_int_t, 0, grid.comm );
MPI_Bcast( &nnz, 1, mpi_int_t, 0, grid.comm );
/* Allocate storage for compressed column representation. */
zallocateA_dist(n, nnz, &a, &asub, &xa);
MPI_Bcast( a, nnz, SuperLU_MPI_DOUBLE_COMPLEX, 0, grid.comm );
MPI_Bcast( asub, nnz, mpi_int_t, 0, grid.comm );
MPI_Bcast( xa, n+1, mpi_int_t, 0, grid.comm );
}
/* Create compressed column matrix for A. */
zCreate_CompCol_Matrix_dist(&A, m, n, nnz, a, asub, xa,
SLU_NC, SLU_Z, SLU_GE);
/* Generate the exact solution and compute the right-hand side. */
if (!(b=doublecomplexMalloc_dist(m*nrhs))) ABORT("Malloc fails for b[]");
if (!(xtrue=doublecomplexMalloc_dist(n*nrhs))) ABORT("Malloc fails for xtrue[]");
*trans = 'N';
ldx = n;
ldb = m;
zGenXtrue_dist(n, nrhs, xtrue, ldx);
zFillRHS_dist(trans, nrhs, xtrue, ldx, &A, b, ldb);
if ( !(berr = doubleMalloc_dist(nrhs)) )
ABORT("Malloc fails for berr[].");
/* ------------------------------------------------------------
NOW WE SOLVE THE LINEAR SYSTEM.
------------------------------------------------------------*/
/* Set the default input options:
options.Fact = DOFACT;
options.Equil = YES;
options.ColPerm = METIS_AT_PLUS_A;
options.RowPerm = LargeDiag_MC64;
options.ReplaceTinyPivot = YES;
options.Trans = NOTRANS;
options.IterRefine = DOUBLE;
options.SolveInitialized = NO;
options.RefineInitialized = NO;
options.PrintStat = YES;
*/
set_default_options_dist(&options);
if (!iam) {
print_sp_ienv_dist(&options);
print_options_dist(&options);
}
/* Initialize ScalePermstruct and LUstruct. */
ScalePermstructInit(m, n, &ScalePermstruct);
LUstructInit(n, &LUstruct);
/* Initialize the statistics variables. */
PStatInit(&stat);
/* Call the linear equation solver. */
pzgssvx_ABglobal(&options, &A, &ScalePermstruct, b, ldb, nrhs, &grid,
&LUstruct, berr, &stat, &info);
/* Check the accuracy of the solution. */
if ( !iam ) {
zinf_norm_error_dist(n, nrhs, b, ldb, xtrue, ldx, &grid);
}
PStatPrint(&options, &stat, &grid); /* Print the statistics. */
/* ------------------------------------------------------------
DEALLOCATE STORAGE.
------------------------------------------------------------*/
PStatFree(&stat);
Destroy_CompCol_Matrix_dist(&A);
Destroy_LU(n, &grid, &LUstruct);
ScalePermstructFree(&ScalePermstruct);
LUstructFree(&LUstruct);
SUPERLU_FREE(b);
SUPERLU_FREE(xtrue);
SUPERLU_FREE(berr);
/* ------------------------------------------------------------
RELEASE THE SUPERLU PROCESS GRID.
------------------------------------------------------------*/
out:
superlu_gridexit(&grid);
/* ------------------------------------------------------------
TERMINATES THE MPI EXECUTION ENVIRONMENT.
------------------------------------------------------------*/
MPI_Finalize();
#if ( DEBUGlevel>=1 )
CHECK_MALLOC(iam, "Exit main()");
#endif
}
int main(int argc, char *argv[])
{
superlu_dist_options_t options;
SuperLUStat_t stat;
SuperMatrix A;
ScalePermstruct_t ScalePermstruct;
LUstruct_t LUstruct;
gridinfo_t grid;
double *berr;
double *a, *a1, *b, *b1, *xtrue;
int_t *asub, *asub1, *xa, *xa1;
int_t i, j, m, n, nnz;
int_t nprow, npcol;
int iam, info, ldb, ldx, nrhs;
char trans[1];
char **cpp, c;
FILE *fp, *fopen();
extern int cpp_defs();
/* prototypes */
extern void LUstructInit(const int_t, LUstruct_t *);
extern void LUstructFree(LUstruct_t *);
extern void Destroy_LU(int_t, gridinfo_t *, LUstruct_t *);
nprow = 1; /* Default process rows. */
npcol = 1; /* Default process columns. */
nrhs = 1; /* Number of right-hand side. */
/* ------------------------------------------------------------
INITIALIZE MPI ENVIRONMENT.
------------------------------------------------------------*/
MPI_Init( &argc, &argv );
/* Parse command line argv[]. */
for (cpp = argv+1; *cpp; ++cpp) {
if ( **cpp == '-' ) {
c = *(*cpp+1);
++cpp;
switch (c) {
case 'h':
printf("Options:\n");
printf("\t-r <int>: process rows (default %d)\n", nprow);
printf("\t-c <int>: process columns (default %d)\n", npcol);
exit(0);
break;
case 'r': nprow = atoi(*cpp);
break;
case 'c': npcol = atoi(*cpp);
break;
}
} else { /* Last arg is considered a filename */
if ( !(fp = fopen(*cpp, "r")) ) {
ABORT("File does not exist");
}
break;
}
}
/* ------------------------------------------------------------
INITIALIZE THE SUPERLU PROCESS GRID.
------------------------------------------------------------*/
superlu_gridinit(MPI_COMM_WORLD, nprow, npcol, &grid);
/* Bail out if I do not belong in the grid. */
iam = grid.iam;
if ( iam >= nprow * npcol )
goto out;
#if ( DEBUGlevel>=1 )
CHECK_MALLOC(iam, "Enter main()");
#endif
/* ------------------------------------------------------------
Process 0 reads the matrix A, and then broadcasts it to all
the other processes.
------------------------------------------------------------*/
if ( !iam ) {
/* Print the CPP definitions. */
cpp_defs();
/* Read the matrix stored on disk in Harwell-Boeing format. */
dreadhb_dist(iam, fp, &m, &n, &nnz, &a, &asub, &xa);
printf("Input matrix file: %s\n", *cpp);
printf("\tDimension\t%dx%d\t # nonzeros %d\n", m, n, nnz);
printf("\tProcess grid\t%d X %d\n", (int) grid.nprow, (int) grid.npcol);
/* Broadcast matrix A to the other PEs. */
MPI_Bcast( &m, 1, mpi_int_t, 0, grid.comm );
MPI_Bcast( &n, 1, mpi_int_t, 0, grid.comm );
MPI_Bcast( &nnz, 1, mpi_int_t, 0, grid.comm );
MPI_Bcast( a, nnz, MPI_DOUBLE, 0, grid.comm );
MPI_Bcast( asub, nnz, mpi_int_t, 0, grid.comm );
MPI_Bcast( xa, n+1, mpi_int_t, 0, grid.comm );
} else {
/* Receive matrix A from PE 0. */
MPI_Bcast( &m, 1, mpi_int_t, 0, grid.comm );
MPI_Bcast( &n, 1, mpi_int_t, 0, grid.comm );
MPI_Bcast( &nnz, 1, mpi_int_t, 0, grid.comm );
/* Allocate storage for compressed column representation. */
dallocateA_dist(n, nnz, &a, &asub, &xa);
MPI_Bcast( a, nnz, MPI_DOUBLE, 0, grid.comm );
MPI_Bcast( asub, nnz, mpi_int_t, 0, grid.comm );
MPI_Bcast( xa, n+1, mpi_int_t, 0, grid.comm );
}
/* Create compressed column matrix for A. */
dCreate_CompCol_Matrix_dist(&A, m, n, nnz, a, asub, xa,
SLU_NC, SLU_D, SLU_GE);
/* Generate the exact solution and compute the right-hand side. */
if (!(b=doubleMalloc_dist(m * nrhs))) ABORT("Malloc fails for b[]");
if (!(xtrue=doubleMalloc_dist(n*nrhs))) ABORT("Malloc fails for xtrue[]");
*trans = 'N';
ldx = n;
ldb = m;
dGenXtrue_dist(n, nrhs, xtrue, ldx);
dFillRHS_dist(trans, nrhs, xtrue, ldx, &A, b, ldb);
/* Save a copy of the right-hand side. */
if ( !(b1 = doubleMalloc_dist(m * nrhs)) ) ABORT("Malloc fails for b1[]");
for (j = 0; j < nrhs; ++j)
for (i = 0; i < m; ++i) b1[i+j*ldb] = b[i+j*ldb];
if ( !(berr = doubleMalloc_dist(nrhs)) )
ABORT("Malloc fails for berr[].");
/* Save a copy of the matrix A. */
dallocateA_dist(n, nnz, &a1, &asub1, &xa1);
for (i = 0; i < nnz; ++i) { a1[i] = a[i]; asub1[i] = asub[i]; }
for (i = 0; i < n+1; ++i) xa1[i] = xa[i];
/* ------------------------------------------------------------
WE SOLVE THE LINEAR SYSTEM FOR THE FIRST TIME.
------------------------------------------------------------*/
/* Set the default input options:
options.Fact = DOFACT;
options.Equil = YES;
options.ColPerm = METIS_AT_PLUS_A;
options.RowPerm = LargeDiag;
options.ReplaceTinyPivot = YES;
options.Trans = NOTRANS;
options.IterRefine = DOUBLE;
options.SolveInitialized = NO;
options.RefineInitialized = NO;
options.PrintStat = YES;
*/
set_default_options_dist(&options);
if (!iam) {
print_sp_ienv_dist(&options);
print_options_dist(&options);
}
/* Initialize ScalePermstruct and LUstruct. */
ScalePermstructInit(m, n, &ScalePermstruct);
LUstructInit(n, &LUstruct);
/* Initialize the statistics variables. */
PStatInit(&stat);
/* Call the linear equation solver: factorize and solve. */
pdgssvx_ABglobal(&options, &A, &ScalePermstruct, b, ldb, nrhs, &grid,
&LUstruct, berr, &stat, &info);
/* Check the accuracy of the solution. */
if ( !iam ) {
dinf_norm_error_dist(n, nrhs, b, ldb, xtrue, ldx, &grid);
}
PStatPrint(&options, &stat, &grid); /* Print the statistics. */
PStatFree(&stat);
Destroy_CompCol_Matrix_dist(&A); /* Deallocate storage of matrix A. */
Destroy_LU(n, &grid, &LUstruct); /* Deallocate storage associated with
the L and U matrices. */
SUPERLU_FREE(b); /* Free storage of right-hand side. */
/* ------------------------------------------------------------
NOW WE SOLVE ANOTHER LINEAR SYSTEM.
ONLY THE SPARSITY PATTERN OF MATRIX A IS THE SAME.
------------------------------------------------------------*/
options.Fact = SamePattern;
PStatInit(&stat); /* Initialize the statistics variables. */
/* Create compressed column matrix for A. */
dCreate_CompCol_Matrix_dist(&A, m, n, nnz, a1, asub1, xa1,
SLU_NC, SLU_D, SLU_GE);
/* Solve the linear system. */
pdgssvx_ABglobal(&options, &A, &ScalePermstruct, b1, ldb, nrhs, &grid,
&LUstruct, berr, &stat, &info);
/* Check the accuracy of the solution. */
if ( !iam ) {
printf("Solve the system with the same sparsity pattern.\n");
dinf_norm_error_dist(n, nrhs, b1, ldb, xtrue, ldx, &grid);
}
/* Print the statistics. */
PStatPrint(&options, &stat, &grid);
/* ------------------------------------------------------------
DEALLOCATE STORAGE.
------------------------------------------------------------*/
PStatFree(&stat);
Destroy_CompCol_Matrix_dist(&A); /* Deallocate storage of matrix A. */
Destroy_LU(n, &grid, &LUstruct); /* Deallocate storage associated with
the L and U matrices. */
ScalePermstructFree(&ScalePermstruct);
LUstructFree(&LUstruct); /* Deallocate the structure of L and U.*/
SUPERLU_FREE(b1); /* Free storage of right-hand side. */
SUPERLU_FREE(xtrue); /* Free storage of the exact solution. */
SUPERLU_FREE(berr);
/* ------------------------------------------------------------
RELEASE THE SUPERLU PROCESS GRID.
------------------------------------------------------------*/
out:
superlu_gridexit(&grid);
/* ------------------------------------------------------------
TERMINATES THE MPI EXECUTION ENVIRONMENT.
------------------------------------------------------------*/
MPI_Finalize();
#if ( DEBUGlevel>=1 )
CHECK_MALLOC(iam, "Exit main()");
#endif
}
PetscErrorCode MatLUFactorNumeric_SuperLU_DIST(Mat F,Mat A,const MatFactorInfo *info)
{
Mat *tseq,A_seq = NULL;
Mat_SeqAIJ *aa,*bb;
Mat_SuperLU_DIST *lu = (Mat_SuperLU_DIST*)(F)->spptr;
PetscErrorCode ierr;
PetscInt M=A->rmap->N,N=A->cmap->N,i,*ai,*aj,*bi,*bj,nz,rstart,*garray,
m=A->rmap->n, colA_start,j,jcol,jB,countA,countB,*bjj,*ajj;
int sinfo; /* SuperLU_Dist info flag is always an int even with long long indices */
PetscMPIInt size;
SuperLUStat_t stat;
double *berr=0;
IS isrow;
Mat F_diag=NULL;
#if defined(PETSC_USE_COMPLEX)
doublecomplex *av, *bv;
#else
double *av, *bv;
#endif
PetscFunctionBegin;
ierr = MPI_Comm_size(PetscObjectComm((PetscObject)A),&size);CHKERRQ(ierr);
if (lu->MatInputMode == GLOBAL) { /* global mat input */
if (size > 1) { /* convert mpi A to seq mat A */
ierr = ISCreateStride(PETSC_COMM_SELF,M,0,1,&isrow);CHKERRQ(ierr);
ierr = MatGetSubMatrices(A,1,&isrow,&isrow,MAT_INITIAL_MATRIX,&tseq);CHKERRQ(ierr);
ierr = ISDestroy(&isrow);CHKERRQ(ierr);
A_seq = *tseq;
ierr = PetscFree(tseq);CHKERRQ(ierr);
aa = (Mat_SeqAIJ*)A_seq->data;
} else {
PetscBool flg;
ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&flg);CHKERRQ(ierr);
if (flg) {
Mat_MPIAIJ *At = (Mat_MPIAIJ*)A->data;
A = At->A;
}
aa = (Mat_SeqAIJ*)A->data;
}
/* Convert Petsc NR matrix to SuperLU_DIST NC.
Note: memories of lu->val, col and row are allocated by CompRow_to_CompCol_dist()! */
if (lu->options.Fact != DOFACT) {/* successive numeric factorization, sparsity pattern is reused. */
PetscStackCall("SuperLU_DIST:Destroy_CompCol_Matrix_dist",Destroy_CompCol_Matrix_dist(&lu->A_sup));
if (lu->FactPattern == SamePattern_SameRowPerm) {
lu->options.Fact = SamePattern_SameRowPerm; /* matrix has similar numerical values */
} else { /* lu->FactPattern == SamePattern */
PetscStackCall("SuperLU_DIST:Destroy_LU",Destroy_LU(N, &lu->grid, &lu->LUstruct));
lu->options.Fact = SamePattern;
}
}
#if defined(PETSC_USE_COMPLEX)
PetscStackCall("SuperLU_DIST:zCompRow_to_CompCol_dist",zCompRow_to_CompCol_dist(M,N,aa->nz,(doublecomplex*)aa->a,(int_t*)aa->j,(int_t*)aa->i,&lu->val,&lu->col, &lu->row));
#else
PetscStackCall("SuperLU_DIST:dCompRow_to_CompCol_dist",dCompRow_to_CompCol_dist(M,N,aa->nz,aa->a,(int_t*)aa->j,(int_t*)aa->i,&lu->val, &lu->col, &lu->row));
#endif
/* Create compressed column matrix A_sup. */
#if defined(PETSC_USE_COMPLEX)
PetscStackCall("SuperLU_DIST:zCreate_CompCol_Matrix_dist",zCreate_CompCol_Matrix_dist(&lu->A_sup, M, N, aa->nz, lu->val, lu->col, lu->row, SLU_NC, SLU_Z, SLU_GE));
#else
PetscStackCall("SuperLU_DIST:dCreate_CompCol_Matrix_dist",dCreate_CompCol_Matrix_dist(&lu->A_sup, M, N, aa->nz, lu->val, lu->col, lu->row, SLU_NC, SLU_D, SLU_GE));
#endif
} else { /* distributed mat input */
Mat_MPIAIJ *mat = (Mat_MPIAIJ*)A->data;
aa=(Mat_SeqAIJ*)(mat->A)->data;
bb=(Mat_SeqAIJ*)(mat->B)->data;
ai=aa->i; aj=aa->j;
bi=bb->i; bj=bb->j;
#if defined(PETSC_USE_COMPLEX)
av=(doublecomplex*)aa->a;
bv=(doublecomplex*)bb->a;
#else
av=aa->a;
bv=bb->a;
#endif
rstart = A->rmap->rstart;
nz = aa->nz + bb->nz;
garray = mat->garray;
if (lu->options.Fact == DOFACT) { /* first numeric factorization */
#if defined(PETSC_USE_COMPLEX)
PetscStackCall("SuperLU_DIST:zallocateA_dist",zallocateA_dist(m, nz, &lu->val, &lu->col, &lu->row));
#else
PetscStackCall("SuperLU_DIST:dallocateA_dist",dallocateA_dist(m, nz, &lu->val, &lu->col, &lu->row));
#endif
} else { /* successive numeric factorization, sparsity pattern and perm_c are reused. */
/* Destroy_CompRowLoc_Matrix_dist(&lu->A_sup); */ /* this leads to crash! However, see SuperLU_DIST_2.5/EXAMPLE/pzdrive2.c */
if (lu->FactPattern == SamePattern_SameRowPerm) {
lu->options.Fact = SamePattern_SameRowPerm; /* matrix has similar numerical values */
} else {
PetscStackCall("SuperLU_DIST:Destroy_LU",Destroy_LU(N, &lu->grid, &lu->LUstruct)); /* Deallocate storage associated with the L and U matrices. */
lu->options.Fact = SamePattern;
}
}
nz = 0;
for (i=0; i<m; i++) {
lu->row[i] = nz;
countA = ai[i+1] - ai[i];
countB = bi[i+1] - bi[i];
ajj = aj + ai[i]; /* ptr to the beginning of this row */
bjj = bj + bi[i];
/* B part, smaller col index */
colA_start = rstart + ajj[0]; /* the smallest global col index of A */
jB = 0;
for (j=0; j<countB; j++) {
jcol = garray[bjj[j]];
if (jcol > colA_start) {
jB = j;
break;
}
lu->col[nz] = jcol;
lu->val[nz++] = *bv++;
if (j==countB-1) jB = countB;
}
/* A part */
for (j=0; j<countA; j++) {
lu->col[nz] = rstart + ajj[j];
lu->val[nz++] = *av++;
}
/* B part, larger col index */
for (j=jB; j<countB; j++) {
lu->col[nz] = garray[bjj[j]];
lu->val[nz++] = *bv++;
}
}
lu->row[m] = nz;
#if defined(PETSC_USE_COMPLEX)
PetscStackCall("SuperLU_DIST:zCreate_CompRowLoc_Matrix_dist",zCreate_CompRowLoc_Matrix_dist(&lu->A_sup, M, N, nz, m, rstart,lu->val, lu->col, lu->row, SLU_NR_loc, SLU_Z, SLU_GE));
#else
PetscStackCall("SuperLU_DIST:dCreate_CompRowLoc_Matrix_dist",dCreate_CompRowLoc_Matrix_dist(&lu->A_sup, M, N, nz, m, rstart,lu->val, lu->col, lu->row, SLU_NR_loc, SLU_D, SLU_GE));
#endif
}
/* Factor the matrix. */
PetscStackCall("SuperLU_DIST:PStatInit",PStatInit(&stat)); /* Initialize the statistics variables. */
if (lu->MatInputMode == GLOBAL) { /* global mat input */
#if defined(PETSC_USE_COMPLEX)
PetscStackCall("SuperLU_DIST:pzgssvx_ABglobal",pzgssvx_ABglobal(&lu->options, &lu->A_sup, &lu->ScalePermstruct, 0, M, 0,&lu->grid, &lu->LUstruct, berr, &stat, &sinfo));
#else
PetscStackCall("SuperLU_DIST:pdgssvx_ABglobal",pdgssvx_ABglobal(&lu->options, &lu->A_sup, &lu->ScalePermstruct, 0, M, 0,&lu->grid, &lu->LUstruct, berr, &stat, &sinfo));
#endif
} else { /* distributed mat input */
#if defined(PETSC_USE_COMPLEX)
PetscStackCall("SuperLU_DIST:pzgssvx",pzgssvx(&lu->options, &lu->A_sup, &lu->ScalePermstruct, 0, m, 0, &lu->grid,&lu->LUstruct, &lu->SOLVEstruct, berr, &stat, &sinfo));
if (sinfo) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"pzgssvx fails, info: %d\n",sinfo);
#else
PetscStackCall("SuperLU_DIST:pdgssvx",pdgssvx(&lu->options, &lu->A_sup, &lu->ScalePermstruct, 0, m, 0, &lu->grid,&lu->LUstruct, &lu->SOLVEstruct, berr, &stat, &sinfo));
if (sinfo) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"pdgssvx fails, info: %d\n",sinfo);
#endif
}
if (lu->MatInputMode == GLOBAL && size > 1) {
ierr = MatDestroy(&A_seq);CHKERRQ(ierr);
}
if (lu->options.PrintStat) {
PStatPrint(&lu->options, &stat, &lu->grid); /* Print the statistics. */
}
PStatFree(&stat);
if (size > 1) {
F_diag = ((Mat_MPIAIJ*)(F)->data)->A;
F_diag->assembled = PETSC_TRUE;
}
(F)->assembled = PETSC_TRUE;
(F)->preallocated = PETSC_TRUE;
lu->options.Fact = FACTORED; /* The factored form of A is supplied. Local option used by this func. only */
PetscFunctionReturn(0);
}
int zcreate_dist_matrix(SuperMatrix *A, int_t m, int_t n, int_t nnz,
doublecomplex *nzval_g, int_t *rowind_g, int_t *colptr_g,
gridinfo_t *grid)
{
SuperMatrix GA; /* global A */
int_t *rowind, *colptr; /* global */
doublecomplex *nzval; /* global */
doublecomplex *nzval_loc; /* local */
int_t *colind, *rowptr; /* local */
int_t m_loc, fst_row, nnz_loc;
int_t m_loc_fst; /* Record m_loc of the first p-1 processors,
when mod(m, p) is not zero. */
int_t iam, row, col, i, j, relpos;
char trans[1];
int_t *marker;
iam = grid->iam;
#if ( DEBUGlevel>=1 )
CHECK_MALLOC(iam, "Enter zcreate_dist_matrix()");
#endif
if ( !iam ) {
/* Allocate storage for compressed column representation. */
zallocateA_dist(n, nnz, &nzval, &rowind, &colptr);
/* Copy the global matrix. */
#if 0
/* and ADJUST to 0-based indexing
which is required by the C routines.*/
#endif
for(i=0; i<nnz; i++){
nzval[i]=nzval_g[i];
rowind[i]=rowind_g[i]; /* - 1;*/
}
for(i=0; i<n+1; i++)
colptr[i]=colptr_g[i]; /* - 1;*/
/* Broadcast matrix A to the other PEs. */
MPI_Bcast( &m, 1, mpi_int_t, 0, grid->comm );
MPI_Bcast( &n, 1, mpi_int_t, 0, grid->comm );
MPI_Bcast( &nnz, 1, mpi_int_t, 0, grid->comm );
MPI_Bcast( nzval, nnz, SuperLU_MPI_DOUBLE_COMPLEX, 0, grid->comm );
MPI_Bcast( rowind, nnz, mpi_int_t, 0, grid->comm );
MPI_Bcast( colptr, n+1, mpi_int_t, 0, grid->comm );
} else {
/* Receive matrix A from PE 0. */
MPI_Bcast( &m, 1, mpi_int_t, 0, grid->comm );
MPI_Bcast( &n, 1, mpi_int_t, 0, grid->comm );
MPI_Bcast( &nnz, 1, mpi_int_t, 0, grid->comm );
/* Allocate storage for compressed column representation. */
zallocateA_dist(n, nnz, &nzval, &rowind, &colptr);
MPI_Bcast( nzval, nnz, SuperLU_MPI_DOUBLE_COMPLEX, 0, grid->comm );
MPI_Bcast( rowind, nnz, mpi_int_t, 0, grid->comm );
MPI_Bcast( colptr, n+1, mpi_int_t, 0, grid->comm );
}
#if 0
nzval[0]=0.1;
#endif
/* Compute the number of rows to be distributed to local process */
m_loc = m / (grid->nprow * grid->npcol);
m_loc_fst = m_loc;
/* When m / procs is not an integer */
if ((m_loc * grid->nprow * grid->npcol) != m) {
m_loc = m_loc+1;
m_loc_fst = m_loc;
if (iam == (grid->nprow * grid->npcol - 1))
m_loc = m - m_loc_fst * (grid->nprow * grid->npcol - 1);
}
/* Create compressed column matrix for GA. */
zCreate_CompCol_Matrix_dist(&GA, m, n, nnz, nzval, rowind, colptr,
SLU_NC, SLU_Z, SLU_GE);
/*************************************************
* Change GA to a local A with NR_loc format *
*************************************************/
rowptr = (int_t *) intMalloc_dist(m_loc+1);
marker = (int_t *) intCalloc_dist(n);
/* Get counts of each row of GA */
for (i = 0; i < n; ++i)
for (j = colptr[i]; j < colptr[i+1]; ++j) ++marker[rowind[j]];
/* Set up row pointers */
rowptr[0] = 0;
fst_row = iam * m_loc_fst;
nnz_loc = 0;
for (j = 0; j < m_loc; ++j) {
row = fst_row + j;
rowptr[j+1] = rowptr[j] + marker[row];
marker[j] = rowptr[j];
}
nnz_loc = rowptr[m_loc];
nzval_loc = (doublecomplex *) doublecomplexMalloc_dist(nnz_loc);
colind = (int_t *) intMalloc_dist(nnz_loc);
/* Transfer the matrix into the compressed row storage */
for (i = 0; i < n; ++i) {
for (j = colptr[i]; j < colptr[i+1]; ++j) {
row = rowind[j];
if ( (row>=fst_row) && (row<fst_row+m_loc) ) {
row = row - fst_row;
relpos = marker[row];
colind[relpos] = i;
nzval_loc[relpos] = nzval[j];
++marker[row];
}
}
}
#if ( DEBUGlevel>=1 )
if ( !iam ) dPrint_CompCol_Matrix_dist(&GA);
#endif
/* Destroy GA */
Destroy_CompCol_Matrix_dist(&GA);
/******************************************************/
/* Change GA to a local A with NR_loc format */
/******************************************************/
/* Set up the local A in NR_loc format */
zCreate_CompRowLoc_Matrix_dist(A, m, n, nnz_loc, m_loc, fst_row,
nzval_loc, colind, rowptr,
SLU_NR_loc, SLU_Z, SLU_GE);
SUPERLU_FREE(marker);
#if ( DEBUGlevel>=1 )
printf("sizeof(NRforamt_loc) %d\n", sizeof(NRformat_loc));
CHECK_MALLOC(iam, "Exit dcreate_dist_matrix()");
#endif
return 0;
}