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_options_t options;
SuperLUStat_t stat;
SuperMatrix A;
ScalePermstruct_t ScalePermstruct;
LUstruct_t LUstruct;
gridinfo_t grid;
double *berr;
doublecomplex *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();
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. */
zreadhb_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", grid.nprow, 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);
/* Save a copy of the right-hand side. */
if ( !(b1 = doublecomplexMalloc_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. */
zallocateA_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. */
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. */
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. */
zCreate_CompCol_Matrix_dist(&A, m, n, nnz, a1, asub1, xa1,
SLU_NC, SLU_Z, SLU_GE);
/* Solve the linear system. */
pzgssvx_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");
zinf_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;
}
