int main(int argc, char *argv[])
{
char equed[1];
yes_no_t equil;
trans_t trans;
SuperMatrix A, L, U;
SuperMatrix B, X;
NCformat *Astore;
NCformat *Ustore;
SCformat *Lstore;
float *a;
int *asub, *xa;
int *perm_r; /* row permutations from partial pivoting */
int *perm_c; /* column permutation vector */
int *etree;
void *work;
int info, lwork, nrhs, ldx;
int i, m, n, nnz;
float *rhsb, *rhsx, *xact;
float *R, *C;
float *ferr, *berr;
float u, rpg, rcond;
mem_usage_t mem_usage;
superlu_options_t options;
SuperLUStat_t stat;
extern void parse_command_line();
#if ( DEBUGlevel>=1 )
CHECK_MALLOC("Enter main()");
#endif
/* Defaults */
lwork = 0;
nrhs = 1;
equil = YES;
u = 1.0;
trans = NOTRANS;
/* Set the default input options:
options.Fact = DOFACT;
options.Equil = YES;
options.ColPerm = COLAMD;
options.DiagPivotThresh = 1.0;
options.Trans = NOTRANS;
options.IterRefine = NOREFINE;
options.SymmetricMode = NO;
options.PivotGrowth = NO;
options.ConditionNumber = NO;
options.PrintStat = YES;
*/
set_default_options(&options);
/* Can use command line input to modify the defaults. */
parse_command_line(argc, argv, &lwork, &u, &equil, &trans);
options.Equil = equil;
options.DiagPivotThresh = u;
options.Trans = trans;
/* Add more functionalities that the defaults. */
options.PivotGrowth = YES; /* Compute reciprocal pivot growth */
options.ConditionNumber = YES;/* Compute reciprocal condition number */
options.IterRefine = SLU_SINGLE; /* Perform single-precision refinement */
if ( lwork > 0 ) {
work = SUPERLU_MALLOC(lwork);
if ( !work ) {
ABORT("SLINSOLX: cannot allocate work[]");
}
}
/* Read matrix A from a file in Harwell-Boeing format.*/
sreadhb(&m, &n, &nnz, &a, &asub, &xa);
sCreate_CompCol_Matrix(&A, m, n, nnz, a, asub, xa, SLU_NC, SLU_S, SLU_GE);
Astore = A.Store;
printf("Dimension %dx%d; # nonzeros %d\n", A.nrow, A.ncol, Astore->nnz);
if ( !(rhsb = floatMalloc(m * nrhs)) ) ABORT("Malloc fails for rhsb[].");
if ( !(rhsx = floatMalloc(m * nrhs)) ) ABORT("Malloc fails for rhsx[].");
sCreate_Dense_Matrix(&B, m, nrhs, rhsb, m, SLU_DN, SLU_S, SLU_GE);
sCreate_Dense_Matrix(&X, m, nrhs, rhsx, m, SLU_DN, SLU_S, SLU_GE);
xact = floatMalloc(n * nrhs);
ldx = n;
sGenXtrue(n, nrhs, xact, ldx);
sFillRHS(trans, nrhs, xact, ldx, &A, &B);
if ( !(etree = intMalloc(n)) ) ABORT("Malloc fails for etree[].");
if ( !(perm_r = intMalloc(m)) ) ABORT("Malloc fails for perm_r[].");
if ( !(perm_c = intMalloc(n)) ) ABORT("Malloc fails for perm_c[].");
if ( !(R = (float *) SUPERLU_MALLOC(A.nrow * sizeof(float))) )
ABORT("SUPERLU_MALLOC fails for R[].");
if ( !(C = (float *) SUPERLU_MALLOC(A.ncol * sizeof(float))) )
ABORT("SUPERLU_MALLOC fails for C[].");
if ( !(ferr = (float *) SUPERLU_MALLOC(nrhs * sizeof(float))) )
ABORT("SUPERLU_MALLOC fails for ferr[].");
if ( !(berr = (float *) SUPERLU_MALLOC(nrhs * sizeof(float))) )
ABORT("SUPERLU_MALLOC fails for berr[].");
/* Initialize the statistics variables. */
StatInit(&stat);
/* Solve the system and compute the condition number
and error bounds using dgssvx. */
sgssvx(&options, &A, perm_c, perm_r, etree, equed, R, C,
&L, &U, work, lwork, &B, &X, &rpg, &rcond, ferr, berr,
&mem_usage, &stat, &info);
printf("sgssvx(): info %d\n", info);
if ( info == 0 || info == n+1 ) {
/* This is how you could access the solution matrix. */
float *sol = (float*) ((DNformat*) X.Store)->nzval;
if ( options.PivotGrowth == YES )
printf("Recip. pivot growth = %e\n", rpg);
if ( options.ConditionNumber == YES )
printf("Recip. condition number = %e\n", rcond);
if ( options.IterRefine != NOREFINE ) {
printf("Iterative Refinement:\n");
printf("%8s%8s%16s%16s\n", "rhs", "Steps", "FERR", "BERR");
for (i = 0; i < nrhs; ++i)
printf("%8d%8d%16e%16e\n", i+1, stat.RefineSteps, ferr[i], berr[i]);
}
Lstore = (SCformat *) L.Store;
Ustore = (NCformat *) U.Store;
printf("No of nonzeros in factor L = %d\n", Lstore->nnz);
printf("No of nonzeros in factor U = %d\n", Ustore->nnz);
printf("No of nonzeros in L+U = %d\n", Lstore->nnz + Ustore->nnz - n);
printf("FILL ratio = %.1f\n", (float)(Lstore->nnz + Ustore->nnz - n)/nnz);
printf("L\\U MB %.3f\ttotal MB needed %.3f\n",
mem_usage.for_lu/1e6, mem_usage.total_needed/1e6);
fflush(stdout);
} else if ( info > 0 && lwork == -1 ) {
printf("** Estimated memory: %d bytes\n", info - n);
}
if ( options.PrintStat ) StatPrint(&stat);
StatFree(&stat);
SUPERLU_FREE (rhsb);
SUPERLU_FREE (rhsx);
SUPERLU_FREE (xact);
SUPERLU_FREE (etree);
SUPERLU_FREE (perm_r);
SUPERLU_FREE (perm_c);
SUPERLU_FREE (R);
SUPERLU_FREE (C);
SUPERLU_FREE (ferr);
SUPERLU_FREE (berr);
Destroy_CompCol_Matrix(&A);
Destroy_SuperMatrix_Store(&B);
Destroy_SuperMatrix_Store(&X);
if ( lwork == 0 ) {
Destroy_SuperNode_Matrix(&L);
Destroy_CompCol_Matrix(&U);
} else if ( lwork > 0 ) {
SUPERLU_FREE(work);
}
#if ( DEBUGlevel>=1 )
CHECK_MALLOC("Exit main()");
#endif
}
int main(int argc, char *argv[])
{
void smatvec_mult(float alpha, float x[], float beta, float y[]);
void spsolve(int n, float x[], float y[]);
extern int sfgmr( int n,
void (*matvec_mult)(float, float [], float, float []),
void (*psolve)(int n, float [], float[]),
float *rhs, float *sol, double tol, int restrt, int *itmax,
FILE *fits);
extern int sfill_diag(int n, NCformat *Astore);
char equed[1] = {'B'};
yes_no_t equil;
trans_t trans;
SuperMatrix A, L, U;
SuperMatrix B, X;
NCformat *Astore;
NCformat *Ustore;
SCformat *Lstore;
GlobalLU_t Glu; /* facilitate multiple factorizations with
SamePattern_SameRowPerm */
float *a;
int *asub, *xa;
int *etree;
int *perm_c; /* column permutation vector */
int *perm_r; /* row permutations from partial pivoting */
int nrhs, ldx, lwork, info, m, n, nnz;
float *rhsb, *rhsx, *xact;
float *work = NULL;
float *R, *C;
float u, rpg, rcond;
float zero = 0.0;
float one = 1.0;
mem_usage_t mem_usage;
superlu_options_t options;
SuperLUStat_t stat;
FILE *fp = stdin;
int restrt, iter, maxit, i;
double resid;
float *x, *b;
#ifdef DEBUG
extern int num_drop_L, num_drop_U;
#endif
#if ( DEBUGlevel>=1 )
CHECK_MALLOC("Enter main()");
#endif
/* Defaults */
lwork = 0;
nrhs = 1;
trans = NOTRANS;
/* Set the default input options:
options.Fact = DOFACT;
options.Equil = YES;
options.ColPerm = COLAMD;
options.DiagPivotThresh = 0.1; //different from complete LU
options.Trans = NOTRANS;
options.IterRefine = NOREFINE;
options.SymmetricMode = NO;
options.PivotGrowth = NO;
options.ConditionNumber = NO;
options.PrintStat = YES;
options.RowPerm = LargeDiag;
options.ILU_DropTol = 1e-4;
options.ILU_FillTol = 1e-2;
options.ILU_FillFactor = 10.0;
options.ILU_DropRule = DROP_BASIC | DROP_AREA;
options.ILU_Norm = INF_NORM;
options.ILU_MILU = SILU;
*/
ilu_set_default_options(&options);
/* Modify the defaults. */
options.PivotGrowth = YES; /* Compute reciprocal pivot growth */
options.ConditionNumber = YES;/* Compute reciprocal condition number */
if ( lwork > 0 ) {
work = SUPERLU_MALLOC(lwork);
if ( !work ) ABORT("Malloc fails for work[].");
}
/* Read matrix A from a file in Harwell-Boeing format.*/
if (argc < 2)
{
printf("Usage:\n%s [OPTION] < [INPUT] > [OUTPUT]\nOPTION:\n"
"-h -hb:\n\t[INPUT] is a Harwell-Boeing format matrix.\n"
"-r -rb:\n\t[INPUT] is a Rutherford-Boeing format matrix.\n"
"-t -triplet:\n\t[INPUT] is a triplet format matrix.\n",
argv[0]);
return 0;
}
else
{
switch (argv[1][1])
{
case 'H':
case 'h':
printf("Input a Harwell-Boeing format matrix:\n");
sreadhb(fp, &m, &n, &nnz, &a, &asub, &xa);
break;
case 'R':
case 'r':
printf("Input a Rutherford-Boeing format matrix:\n");
sreadrb(&m, &n, &nnz, &a, &asub, &xa);
break;
case 'T':
case 't':
printf("Input a triplet format matrix:\n");
sreadtriple(&m, &n, &nnz, &a, &asub, &xa);
break;
default:
printf("Unrecognized format.\n");
return 0;
}
}
sCreate_CompCol_Matrix(&A, m, n, nnz, a, asub, xa,
SLU_NC, SLU_S, SLU_GE);
Astore = A.Store;
sfill_diag(n, Astore);
printf("Dimension %dx%d; # nonzeros %d\n", A.nrow, A.ncol, Astore->nnz);
fflush(stdout);
/* Generate the right-hand side */
if ( !(rhsb = floatMalloc(m * nrhs)) ) ABORT("Malloc fails for rhsb[].");
if ( !(rhsx = floatMalloc(m * nrhs)) ) ABORT("Malloc fails for rhsx[].");
sCreate_Dense_Matrix(&B, m, nrhs, rhsb, m, SLU_DN, SLU_S, SLU_GE);
sCreate_Dense_Matrix(&X, m, nrhs, rhsx, m, SLU_DN, SLU_S, SLU_GE);
xact = floatMalloc(n * nrhs);
ldx = n;
sGenXtrue(n, nrhs, xact, ldx);
sFillRHS(trans, nrhs, xact, ldx, &A, &B);
if ( !(etree = intMalloc(n)) ) ABORT("Malloc fails for etree[].");
if ( !(perm_r = intMalloc(m)) ) ABORT("Malloc fails for perm_r[].");
if ( !(perm_c = intMalloc(n)) ) ABORT("Malloc fails for perm_c[].");
if ( !(R = (float *) SUPERLU_MALLOC(A.nrow * sizeof(float))) )
ABORT("SUPERLU_MALLOC fails for R[].");
if ( !(C = (float *) SUPERLU_MALLOC(A.ncol * sizeof(float))) )
ABORT("SUPERLU_MALLOC fails for C[].");
info = 0;
#ifdef DEBUG
num_drop_L = 0;
num_drop_U = 0;
#endif
/* Initialize the statistics variables. */
StatInit(&stat);
/* Compute the incomplete factorization and compute the condition number
and pivot growth using dgsisx. */
B.ncol = 0; /* not to perform triangular solution */
sgsisx(&options, &A, perm_c, perm_r, etree, equed, R, C, &L, &U, work,
lwork, &B, &X, &rpg, &rcond, &Glu, &mem_usage, &stat, &info);
/* Set RHS for GMRES. */
if (!(b = floatMalloc(m))) ABORT("Malloc fails for b[].");
if (*equed == 'R' || *equed == 'B') {
for (i = 0; i < n; ++i) b[i] = rhsb[i] * R[i];
} else {
for (i = 0; i < m; i++) b[i] = rhsb[i];
}
printf("sgsisx(): info %d, equed %c\n", info, equed[0]);
if (info > 0 || rcond < 1e-8 || rpg > 1e8)
printf("WARNING: This preconditioner might be unstable.\n");
if ( info == 0 || info == n+1 ) {
if ( options.PivotGrowth == YES )
printf("Recip. pivot growth = %e\n", rpg);
if ( options.ConditionNumber == YES )
printf("Recip. condition number = %e\n", rcond);
} else if ( info > 0 && lwork == -1 ) {
printf("** Estimated memory: %d bytes\n", info - n);
}
Lstore = (SCformat *) L.Store;
Ustore = (NCformat *) U.Store;
printf("n(A) = %d, nnz(A) = %d\n", n, Astore->nnz);
printf("No of nonzeros in factor L = %d\n", Lstore->nnz);
printf("No of nonzeros in factor U = %d\n", Ustore->nnz);
printf("No of nonzeros in L+U = %d\n", Lstore->nnz + Ustore->nnz - n);
printf("Fill ratio: nnz(F)/nnz(A) = %.3f\n",
((double)(Lstore->nnz) + (double)(Ustore->nnz) - (double)n)
/ (double)Astore->nnz);
printf("L\\U MB %.3f\ttotal MB needed %.3f\n",
mem_usage.for_lu/1e6, mem_usage.total_needed/1e6);
fflush(stdout);
/* Set the global variables. */
GLOBAL_A = &A;
GLOBAL_L = &L;
GLOBAL_U = &U;
GLOBAL_STAT = &stat;
GLOBAL_PERM_C = perm_c;
GLOBAL_PERM_R = perm_r;
GLOBAL_OPTIONS = &options;
GLOBAL_R = R;
GLOBAL_C = C;
GLOBAL_MEM_USAGE = &mem_usage;
/* Set the options to do solve-only. */
options.Fact = FACTORED;
options.PivotGrowth = NO;
options.ConditionNumber = NO;
/* Set the variables used by GMRES. */
restrt = SUPERLU_MIN(n / 3 + 1, 50);
maxit = 1000;
iter = maxit;
resid = 1e-8;
if (!(x = floatMalloc(n))) ABORT("Malloc fails for x[].");
if (info <= n + 1)
{
int i_1 = 1;
double maxferr = 0.0, nrmA, nrmB, res, t;
float temp;
extern float snrm2_(int *, float [], int *);
extern void saxpy_(int *, float *, float [], int *, float [], int *);
/* Initial guess */
for (i = 0; i < n; i++) x[i] = zero;
t = SuperLU_timer_();
/* Call GMRES */
sfgmr(n, smatvec_mult, spsolve, b, x, resid, restrt, &iter, stdout);
t = SuperLU_timer_() - t;
/* Output the result. */
nrmA = snrm2_(&(Astore->nnz), (float *)((DNformat *)A.Store)->nzval,
&i_1);
nrmB = snrm2_(&m, b, &i_1);
sp_sgemv("N", -1.0, &A, x, 1, 1.0, b, 1);
res = snrm2_(&m, b, &i_1);
resid = res / nrmB;
printf("||A||_F = %.1e, ||B||_2 = %.1e, ||B-A*X||_2 = %.1e, "
"relres = %.1e\n", nrmA, nrmB, res, resid);
if (iter >= maxit)
{
if (resid >= 1.0) iter = -180;
else if (resid > 1e-8) iter = -111;
}
printf("iteration: %d\nresidual: %.1e\nGMRES time: %.2f seconds.\n",
iter, resid, t);
/* Scale the solution back if equilibration was performed. */
if (*equed == 'C' || *equed == 'B')
for (i = 0; i < n; i++) x[i] *= C[i];
for (i = 0; i < m; i++) {
maxferr = SUPERLU_MAX(maxferr, fabs(x[i] - xact[i]));
}
printf("||X-X_true||_oo = %.1e\n", maxferr);
}
#ifdef DEBUG
printf("%d entries in L and %d entries in U dropped.\n",
num_drop_L, num_drop_U);
#endif
fflush(stdout);
if ( options.PrintStat ) StatPrint(&stat);
StatFree(&stat);
SUPERLU_FREE (rhsb);
SUPERLU_FREE (rhsx);
SUPERLU_FREE (xact);
SUPERLU_FREE (etree);
SUPERLU_FREE (perm_r);
SUPERLU_FREE (perm_c);
SUPERLU_FREE (R);
SUPERLU_FREE (C);
Destroy_CompCol_Matrix(&A);
Destroy_SuperMatrix_Store(&B);
Destroy_SuperMatrix_Store(&X);
if ( lwork >= 0 ) {
Destroy_SuperNode_Matrix(&L);
Destroy_CompCol_Matrix(&U);
}
SUPERLU_FREE(b);
SUPERLU_FREE(x);
#if ( DEBUGlevel>=1 )
CHECK_MALLOC("Exit main()");
#endif
return 0;
}
int main(int argc, char *argv[])
{
/*
* Purpose
* =======
*
* The driver program CLINSOLX2.
*
* This example illustrates how to use CGSSVX to solve systems repeatedly
* with the same sparsity pattern of matrix A.
* In this case, the column permutation vector perm_c is computed once.
* The following data structures will be reused in the subsequent call to
* CGSSVX: perm_c, etree
*
*/
char equed[1];
yes_no_t equil;
trans_t trans;
SuperMatrix A, A1, L, U;
SuperMatrix B, B1, X;
NCformat *Astore;
NCformat *Ustore;
SCformat *Lstore;
complex *a, *a1;
int *asub, *xa, *asub1, *xa1;
int *perm_r; /* row permutations from partial pivoting */
int *perm_c; /* column permutation vector */
int *etree;
void *work;
int info, lwork, nrhs, ldx;
int i, j, m, n, nnz;
complex *rhsb, *rhsb1, *rhsx, *xact;
float *R, *C;
float *ferr, *berr;
float u, rpg, rcond;
mem_usage_t mem_usage;
superlu_options_t options;
SuperLUStat_t stat;
extern void parse_command_line();
#if ( DEBUGlevel>=1 )
CHECK_MALLOC("Enter main()");
#endif
/* Defaults */
lwork = 0;
nrhs = 1;
equil = YES;
u = 1.0;
trans = NOTRANS;
/* Set the default input options:
options.Fact = DOFACT;
options.Equil = YES;
options.ColPerm = COLAMD;
options.DiagPivotThresh = 1.0;
options.Trans = NOTRANS;
options.IterRefine = NOREFINE;
options.SymmetricMode = NO;
options.PivotGrowth = NO;
options.ConditionNumber = NO;
options.PrintStat = YES;
*/
set_default_options(&options);
/* Can use command line input to modify the defaults. */
parse_command_line(argc, argv, &lwork, &u, &equil, &trans);
options.Equil = equil;
options.DiagPivotThresh = u;
options.Trans = trans;
if ( lwork > 0 ) {
work = SUPERLU_MALLOC(lwork);
if ( !work ) {
ABORT("DLINSOLX: cannot allocate work[]");
}
}
/* Read matrix A from a file in Harwell-Boeing format.*/
creadhb(&m, &n, &nnz, &a, &asub, &xa);
if ( !(a1 = complexMalloc(nnz)) ) ABORT("Malloc fails for a1[].");
if ( !(asub1 = intMalloc(nnz)) ) ABORT("Malloc fails for asub1[].");
if ( !(xa1 = intMalloc(n+1)) ) ABORT("Malloc fails for 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];
cCreate_CompCol_Matrix(&A, m, n, nnz, a, asub, xa, SLU_NC, SLU_C, SLU_GE);
Astore = A.Store;
printf("Dimension %dx%d; # nonzeros %d\n", A.nrow, A.ncol, Astore->nnz);
if ( !(rhsb = complexMalloc(m * nrhs)) ) ABORT("Malloc fails for rhsb[].");
if ( !(rhsb1 = complexMalloc(m * nrhs)) ) ABORT("Malloc fails for rhsb1[].");
if ( !(rhsx = complexMalloc(m * nrhs)) ) ABORT("Malloc fails for rhsx[].");
cCreate_Dense_Matrix(&B, m, nrhs, rhsb, m, SLU_DN, SLU_C, SLU_GE);
cCreate_Dense_Matrix(&X, m, nrhs, rhsx, m, SLU_DN, SLU_C, SLU_GE);
xact = complexMalloc(n * nrhs);
ldx = n;
cGenXtrue(n, nrhs, xact, ldx);
cFillRHS(trans, nrhs, xact, ldx, &A, &B);
for (j = 0; j < nrhs; ++j)
for (i = 0; i < m; ++i) rhsb1[i+j*m] = rhsb[i+j*m];
if ( !(perm_c = intMalloc(n)) ) ABORT("Malloc fails for perm_c[].");
if ( !(perm_r = intMalloc(m)) ) ABORT("Malloc fails for perm_r[].");
if ( !(etree = intMalloc(n)) ) ABORT("Malloc fails for etree[].");
if ( !(R = (float *) SUPERLU_MALLOC(A.nrow * sizeof(float))) )
ABORT("SUPERLU_MALLOC fails for R[].");
if ( !(C = (float *) SUPERLU_MALLOC(A.ncol * sizeof(float))) )
ABORT("SUPERLU_MALLOC fails for C[].");
if ( !(ferr = (float *) SUPERLU_MALLOC(nrhs * sizeof(float))) )
ABORT("SUPERLU_MALLOC fails for ferr[].");
if ( !(berr = (float *) SUPERLU_MALLOC(nrhs * sizeof(float))) )
ABORT("SUPERLU_MALLOC fails for berr[].");
/* Initialize the statistics variables. */
StatInit(&stat);
/* ------------------------------------------------------------
WE SOLVE THE LINEAR SYSTEM FOR THE FIRST TIME: AX = B
------------------------------------------------------------*/
cgssvx(&options, &A, perm_c, perm_r, etree, equed, R, C,
&L, &U, work, lwork, &B, &X, &rpg, &rcond, ferr, berr,
&mem_usage, &stat, &info);
printf("First system: cgssvx() returns info %d\n", info);
if ( info == 0 || info == n+1 ) {
/* This is how you could access the solution matrix. */
complex *sol = (complex*) ((DNformat*) X.Store)->nzval;
if ( options.PivotGrowth ) printf("Recip. pivot growth = %e\n", rpg);
if ( options.ConditionNumber )
printf("Recip. condition number = %e\n", rcond);
Lstore = (SCformat *) L.Store;
Ustore = (NCformat *) U.Store;
printf("No of nonzeros in factor L = %d\n", Lstore->nnz);
printf("No of nonzeros in factor U = %d\n", Ustore->nnz);
printf("No of nonzeros in L+U = %d\n", Lstore->nnz + Ustore->nnz - n);
printf("FILL ratio = %.1f\n", (float)(Lstore->nnz + Ustore->nnz - n)/nnz);
printf("L\\U MB %.3f\ttotal MB needed %.3f\n",
mem_usage.for_lu/1e6, mem_usage.total_needed/1e6);
if ( options.IterRefine ) {
printf("Iterative Refinement:\n");
printf("%8s%8s%16s%16s\n", "rhs", "Steps", "FERR", "BERR");
for (i = 0; i < nrhs; ++i)
printf("%8d%8d%16e%16e\n", i+1, stat.RefineSteps, ferr[i], berr[i]);
}
fflush(stdout);
} else if ( info > 0 && lwork == -1 ) {
printf("** Estimated memory: %d bytes\n", info - n);
}
if ( options.PrintStat ) StatPrint(&stat);
StatFree(&stat);
Destroy_CompCol_Matrix(&A);
Destroy_Dense_Matrix(&B);
if ( lwork >= 0 ) { /* Deallocate storage associated with L and U. */
Destroy_SuperNode_Matrix(&L);
Destroy_CompCol_Matrix(&U);
}
/* ------------------------------------------------------------
NOW WE SOLVE ANOTHER LINEAR SYSTEM: A1*X = B1
ONLY THE SPARSITY PATTERN OF A1 IS THE SAME AS THAT OF A.
------------------------------------------------------------*/
options.Fact = SamePattern;
StatInit(&stat); /* Initialize the statistics variables. */
cCreate_CompCol_Matrix(&A1, m, n, nnz, a1, asub1, xa1,
SLU_NC, SLU_C, SLU_GE);
cCreate_Dense_Matrix(&B1, m, nrhs, rhsb1, m, SLU_DN, SLU_C, SLU_GE);
cgssvx(&options, &A1, perm_c, perm_r, etree, equed, R, C,
&L, &U, work, lwork, &B1, &X, &rpg, &rcond, ferr, berr,
&mem_usage, &stat, &info);
printf("\nSecond system: cgssvx() returns info %d\n", info);
if ( info == 0 || info == n+1 ) {
/* This is how you could access the solution matrix. */
complex *sol = (complex*) ((DNformat*) X.Store)->nzval;
if ( options.PivotGrowth ) printf("Recip. pivot growth = %e\n", rpg);
if ( options.ConditionNumber )
printf("Recip. condition number = %e\n", rcond);
Lstore = (SCformat *) L.Store;
Ustore = (NCformat *) U.Store;
printf("No of nonzeros in factor L = %d\n", Lstore->nnz);
printf("No of nonzeros in factor U = %d\n", Ustore->nnz);
printf("No of nonzeros in L+U = %d\n", Lstore->nnz + Ustore->nnz - n);
printf("L\\U MB %.3f\ttotal MB needed %.3f\n",
mem_usage.for_lu/1e6, mem_usage.total_needed/1e6);
if ( options.IterRefine ) {
printf("Iterative Refinement:\n");
printf("%8s%8s%16s%16s\n", "rhs", "Steps", "FERR", "BERR");
for (i = 0; i < nrhs; ++i)
printf("%8d%8d%16e%16e\n", i+1, stat.RefineSteps, ferr[i], berr[i]);
}
fflush(stdout);
} else if ( info > 0 && lwork == -1 ) {
printf("** Estimated memory: %d bytes\n", info - n);
}
if ( options.PrintStat ) StatPrint(&stat);
StatFree(&stat);
SUPERLU_FREE (xact);
SUPERLU_FREE (etree);
SUPERLU_FREE (perm_r);
SUPERLU_FREE (perm_c);
SUPERLU_FREE (R);
SUPERLU_FREE (C);
SUPERLU_FREE (ferr);
SUPERLU_FREE (berr);
Destroy_CompCol_Matrix(&A1);
Destroy_Dense_Matrix(&B1);
Destroy_Dense_Matrix(&X);
if ( lwork == 0 ) {
Destroy_SuperNode_Matrix(&L);
Destroy_CompCol_Matrix(&U);
} else if ( lwork > 0 ) {
SUPERLU_FREE(work);
}
#if ( DEBUGlevel>=1 )
CHECK_MALLOC("Exit main()");
#endif
}
bool SuperLUSolver::Solve(SparseMatrixType& rA, VectorType& rX, VectorType& rB)
{
//std::cout << "matrix size in solver: " << rA.size1() << std::endl;
//std::cout << "RHS size in solver SLU: " << rB.size() << std::endl;
// typedef ublas::compressed_matrix<double, ublas::row_major, 0,
// ublas::unbounded_array<int>, ublas::unbounded_array<double> > cm_t;
//make a copy of the RHS
VectorType rC = rB;
superlu_options_t options;
SuperLUStat_t stat;
/* Set the default input options:
options.Fact = DOFACT;
options.Equil = YES;
options.ColPerm = COLAMD;
options.DiagPivotThresh = 1.0;
options.Trans = NOTRANS;
options.IterRefine = NOREFINE;
options.SymmetricMode = NO;
options.PivotGrowth = NO;
options.ConditionNumber = NO;
options.PrintStat = YES;
*/
set_default_options(&options);
options.IterRefine = SLU_DOUBLE;
// options.ColPerm = MMD_AT_PLUS_A;
//Fill the SuperLU matrices
SuperMatrix Aslu, B, L, U;
//create a copy of the matrix
int *index1_vector = new (std::nothrow) int[rA.index1_data().size()];
int *index2_vector = new (std::nothrow) int[rA.index2_data().size()];
// double *values_vector = new (std::nothrow) double[rA.value_data().size()];
for( int unsigned i = 0; i < rA.index1_data().size(); i++ )
index1_vector[i] = (int)rA.index1_data()[i];
for( unsigned int i = 0; i < rA.index2_data().size(); i++ )
index2_vector[i] = (int)rA.index2_data()[i];
/* for( unsigned int i = 0; i < rA.value_data().size(); i++ )
values_vector[i] = (double)rA.value_data()[i];*/
//create a copy of the rhs vector (it will be overwritten with the solution)
/* double *b_vector = new (std::nothrow) double[rB.size()];
for( unsigned int i = 0; i < rB.size(); i++ )
b_vector[i] = rB[i];*/
/*
dCreate_CompCol_Matrix (&Aslu, rA.size1(), rA.size2(),
rA.nnz(),
values_vector,
index2_vector,
index1_vector,
SLU_NR, SLU_D, SLU_GE
);*/
//works also with dCreate_CompCol_Matrix
dCreate_CompRow_Matrix (&Aslu, rA.size1(), rA.size2(),
rA.nnz(),
rA.value_data().begin(),
index2_vector, //can not avoid a copy as ublas uses unsigned int internally
index1_vector, //can not avoid a copy as ublas uses unsigned int internally
SLU_NR, SLU_D, SLU_GE
);
dCreate_Dense_Matrix (&B, rB.size(), 1,&rB[0],rB.size(),SLU_DN, SLU_D, SLU_GE);
//allocate memory for permutation arrays
int* perm_c;
int* perm_r;
if ( !(perm_c = intMalloc(rA.size1())) ) ABORT("Malloc fails for perm_c[].");
if ( !(perm_r = intMalloc(rA.size2())) ) ABORT("Malloc fails for perm_r[].");
//initialize container for statistical data
StatInit(&stat);
//call solver routine
int info;
dgssv(&options, &Aslu, perm_c, perm_r, &L, &U, &B, &stat, &info);
//print output
if (options.PrintStat) {
StatPrint(&stat);
}
//resubstitution of results
#pragma omp parallel for
for(int i=0; i<static_cast<int>(rB.size()); i++ )
rX[i] = rB[i]; // B(i,0);
//recover the RHS
rB=rC;
//deallocate memory used
StatFree(&stat);
SUPERLU_FREE (perm_r);
SUPERLU_FREE (perm_c);
Destroy_SuperMatrix_Store(&Aslu); //note that by using the "store" function we will take care of deallocation ourselves
Destroy_SuperMatrix_Store(&B);
Destroy_SuperNode_Matrix(&L);
Destroy_CompCol_Matrix(&U);
delete [] index1_vector;
delete [] index2_vector;
// delete [] b_vector;
//CHECK WITH VALGRIND IF THIS IS NEEDED ...or if it is done by the lines above
//deallocate tempory storage used for the matrix
// if(b_vector!=NULL) delete [] index1_vector;
// // if(b_vector!=NULL) delete [] index2_vector;
// if(b_vector!=NULL) delete [] values_vector;
// if(b_vector!=NULL) delete [] b_vector;
return true;
}
main(int argc, char *argv[])
{
/*
* Purpose
* =======
*
* The driver program ZLINSOLX1.
*
* This example illustrates how to use ZGSSVX to solve systems with the same
* A but different right-hand side.
* In this case, we factorize A only once in the first call to DGSSVX,
* and reuse the following data structures in the subsequent call to ZGSSVX:
* perm_c, perm_r, R, C, L, U.
*
*/
char equed[1];
yes_no_t equil;
trans_t trans;
SuperMatrix A, L, U;
SuperMatrix B, X;
NCformat *Astore;
NCformat *Ustore;
SCformat *Lstore;
doublecomplex *a;
int *asub, *xa;
int *perm_c; /* column permutation vector */
int *perm_r; /* row permutations from partial pivoting */
int *etree;
void *work;
int info, lwork, nrhs, ldx;
int i, m, n, nnz;
doublecomplex *rhsb, *rhsx, *xact;
double *R, *C;
double *ferr, *berr;
double u, rpg, rcond;
mem_usage_t mem_usage;
superlu_options_t options;
SuperLUStat_t stat;
extern void parse_command_line();
#if ( DEBUGlevel>=1 )
CHECK_MALLOC("Enter main()");
#endif
/* Defaults */
lwork = 0;
nrhs = 1;
equil = YES;
u = 1.0;
trans = NOTRANS;
/* Set the default values for options argument:
options.Fact = DOFACT;
options.Equil = YES;
options.ColPerm = COLAMD;
options.DiagPivotThresh = 1.0;
options.Trans = NOTRANS;
options.IterRefine = NOREFINE;
options.SymmetricMode = NO;
options.PivotGrowth = NO;
options.ConditionNumber = NO;
options.PrintStat = YES;
*/
set_default_options(&options);
/* Can use command line input to modify the defaults. */
parse_command_line(argc, argv, &lwork, &u, &equil, &trans);
options.Equil = equil;
options.DiagPivotThresh = u;
options.Trans = trans;
if ( lwork > 0 ) {
work = SUPERLU_MALLOC(lwork);
if ( !work ) {
ABORT("ZLINSOLX: cannot allocate work[]");
}
}
/* Read matrix A from a file in Harwell-Boeing format.*/
zreadhb(&m, &n, &nnz, &a, &asub, &xa);
zCreate_CompCol_Matrix(&A, m, n, nnz, a, asub, xa, SLU_NC, SLU_Z, SLU_GE);
Astore = A.Store;
printf("Dimension %dx%d; # nonzeros %d\n", A.nrow, A.ncol, Astore->nnz);
if ( !(rhsb = doublecomplexMalloc(m * nrhs)) ) ABORT("Malloc fails for rhsb[].");
if ( !(rhsx = doublecomplexMalloc(m * nrhs)) ) ABORT("Malloc fails for rhsx[].");
zCreate_Dense_Matrix(&B, m, nrhs, rhsb, m, SLU_DN, SLU_Z, SLU_GE);
zCreate_Dense_Matrix(&X, m, nrhs, rhsx, m, SLU_DN, SLU_Z, SLU_GE);
xact = doublecomplexMalloc(n * nrhs);
ldx = n;
zGenXtrue(n, nrhs, xact, ldx);
zFillRHS(trans, nrhs, xact, ldx, &A, &B);
if ( !(etree = intMalloc(n)) ) ABORT("Malloc fails for etree[].");
if ( !(perm_r = intMalloc(m)) ) ABORT("Malloc fails for perm_r[].");
if ( !(perm_c = intMalloc(n)) ) ABORT("Malloc fails for perm_c[].");
if ( !(R = (double *) SUPERLU_MALLOC(A.nrow * sizeof(double))) )
ABORT("SUPERLU_MALLOC fails for R[].");
if ( !(C = (double *) SUPERLU_MALLOC(A.ncol * sizeof(double))) )
ABORT("SUPERLU_MALLOC fails for C[].");
if ( !(ferr = (double *) SUPERLU_MALLOC(nrhs * sizeof(double))) )
ABORT("SUPERLU_MALLOC fails for ferr[].");
if ( !(berr = (double *) SUPERLU_MALLOC(nrhs * sizeof(double))) )
ABORT("SUPERLU_MALLOC fails for berr[].");
/* Initialize the statistics variables. */
StatInit(&stat);
/* ONLY PERFORM THE LU DECOMPOSITION */
B.ncol = 0; /* Indicate not to solve the system */
zgssvx(&options, &A, perm_c, perm_r, etree, equed, R, C,
&L, &U, work, lwork, &B, &X, &rpg, &rcond, ferr, berr,
&mem_usage, &stat, &info);
printf("LU factorization: zgssvx() returns info %d\n", info);
if ( info == 0 || info == n+1 ) {
if ( options.PivotGrowth ) printf("Recip. pivot growth = %e\n", rpg);
if ( options.ConditionNumber )
printf("Recip. condition number = %e\n", rcond);
Lstore = (SCformat *) L.Store;
Ustore = (NCformat *) U.Store;
printf("No of nonzeros in factor L = %d\n", Lstore->nnz);
printf("No of nonzeros in factor U = %d\n", Ustore->nnz);
printf("No of nonzeros in L+U = %d\n", Lstore->nnz + Ustore->nnz - n);
printf("FILL ratio = %.1f\n", (float)(Lstore->nnz + Ustore->nnz - n)/nnz);
printf("L\\U MB %.3f\ttotal MB needed %.3f\n",
mem_usage.for_lu/1e6, mem_usage.total_needed/1e6);
fflush(stdout);
} else if ( info > 0 && lwork == -1 ) {
printf("** Estimated memory: %d bytes\n", info - n);
}
if ( options.PrintStat ) StatPrint(&stat);
StatFree(&stat);
/* ------------------------------------------------------------
NOW WE SOLVE THE LINEAR SYSTEM USING THE FACTORED FORM OF A.
------------------------------------------------------------*/
options.Fact = FACTORED; /* Indicate the factored form of A is supplied. */
B.ncol = nrhs; /* Set the number of right-hand side */
/* Initialize the statistics variables. */
StatInit(&stat);
zgssvx(&options, &A, perm_c, perm_r, etree, equed, R, C,
&L, &U, work, lwork, &B, &X, &rpg, &rcond, ferr, berr,
&mem_usage, &stat, &info);
printf("Triangular solve: zgssvx() returns info %d\n", info);
if ( info == 0 || info == n+1 ) {
/* This is how you could access the solution matrix. */
doublecomplex *sol = (doublecomplex*) ((DNformat*) X.Store)->nzval;
if ( options.IterRefine ) {
printf("Iterative Refinement:\n");
printf("%8s%8s%16s%16s\n", "rhs", "Steps", "FERR", "BERR");
for (i = 0; i < nrhs; ++i)
printf("%8d%8d%16e%16e\n", i+1, stat.RefineSteps, ferr[i], berr[i]);
}
fflush(stdout);
} else if ( info > 0 && lwork == -1 ) {
printf("** Estimated memory: %d bytes\n", info - n);
}
if ( options.PrintStat ) StatPrint(&stat);
StatFree(&stat);
SUPERLU_FREE (rhsb);
SUPERLU_FREE (rhsx);
SUPERLU_FREE (xact);
SUPERLU_FREE (etree);
SUPERLU_FREE (perm_r);
SUPERLU_FREE (perm_c);
SUPERLU_FREE (R);
SUPERLU_FREE (C);
SUPERLU_FREE (ferr);
SUPERLU_FREE (berr);
Destroy_CompCol_Matrix(&A);
Destroy_SuperMatrix_Store(&B);
Destroy_SuperMatrix_Store(&X);
if ( lwork == 0 ) {
Destroy_SuperNode_Matrix(&L);
Destroy_CompCol_Matrix(&U);
} else if ( lwork > 0 ) {
SUPERLU_FREE(work);
}
#if ( DEBUGlevel>=1 )
CHECK_MALLOC("Exit main()");
#endif
}
main(int argc, char *argv[])
{
SuperMatrix A;
NCformat *Astore;
double *a;
int *asub, *xa;
int *perm_c; /* column permutation vector */
int *perm_r; /* row permutations from partial pivoting */
SuperMatrix L; /* factor L */
SCformat *Lstore;
SuperMatrix U; /* factor U */
NCformat *Ustore;
SuperMatrix B;
int nrhs, ldx, info, m, n, nnz;
double *xact, *rhs;
mem_usage_t mem_usage;
superlu_options_t options;
SuperLUStat_t stat;
#if ( DEBUGlevel>=1 )
CHECK_MALLOC("Enter main()");
#endif
/* Set the default input options:
options.Fact = DOFACT;
options.Equil = YES;
options.ColPerm = COLAMD;
options.DiagPivotThresh = 1.0;
options.Trans = NOTRANS;
options.IterRefine = NOREFINE;
options.SymmetricMode = NO;
options.PivotGrowth = NO;
options.ConditionNumber = NO;
options.PrintStat = YES;
*/
set_default_options(&options);
/* Now we modify the default options to use the symmetric mode. */
options.SymmetricMode = YES;
options.ColPerm = MMD_AT_PLUS_A;
options.DiagPivotThresh = 0.001;
#if 1
/* Read matrix A from a file in Harwell-Boeing format.*/
if (argc < 2)
{
printf("Usage:\n%s [OPTION] < [INPUT] > [OUTPUT]\nOPTION:\n"
"-h -hb:\n\t[INPUT] is a Harwell-Boeing format matrix.\n"
"-r -rb:\n\t[INPUT] is a Rutherford-Boeing format matrix.\n"
"-t -triplet:\n\t[INPUT] is a triplet format matrix.\n",
argv[0]);
return 0;
}
else
{
switch (argv[1][1])
{
case 'H':
case 'h':
printf("Input a Harwell-Boeing format matrix:\n");
dreadhb(&m, &n, &nnz, &a, &asub, &xa);
break;
case 'R':
case 'r':
printf("Input a Rutherford-Boeing format matrix:\n");
dreadrb(&m, &n, &nnz, &a, &asub, &xa);
break;
case 'T':
case 't':
printf("Input a triplet format matrix:\n");
dreadtriple(&m, &n, &nnz, &a, &asub, &xa);
break;
default:
printf("Unrecognized format.\n");
return 0;
}
}
#else
/* Read the matrix in Harwell-Boeing format. */
dreadhb(&m, &n, &nnz, &a, &asub, &xa);
#endif
dCreate_CompCol_Matrix(&A, m, n, nnz, a, asub, xa, SLU_NC, SLU_D, SLU_GE);
Astore = A.Store;
printf("Dimension %dx%d; # nonzeros %d\n", A.nrow, A.ncol, Astore->nnz);
nrhs = 1;
if ( !(rhs = doubleMalloc(m * nrhs)) ) ABORT("Malloc fails for rhs[].");
dCreate_Dense_Matrix(&B, m, nrhs, rhs, m, SLU_DN, SLU_D, SLU_GE);
xact = doubleMalloc(n * nrhs);
ldx = n;
dGenXtrue(n, nrhs, xact, ldx);
dFillRHS(options.Trans, nrhs, xact, ldx, &A, &B);
if ( !(perm_c = intMalloc(n)) ) ABORT("Malloc fails for perm_c[].");
if ( !(perm_r = intMalloc(m)) ) ABORT("Malloc fails for perm_r[].");
/* Initialize the statistics variables. */
StatInit(&stat);
dgssv(&options, &A, perm_c, perm_r, &L, &U, &B, &stat, &info);
if ( info == 0 ) {
/* This is how you could access the solution matrix. */
double *sol = (double*) ((DNformat*) B.Store)->nzval;
/* Compute the infinity norm of the error. */
dinf_norm_error(nrhs, &B, xact);
Lstore = (SCformat *) L.Store;
Ustore = (NCformat *) U.Store;
printf("No of nonzeros in factor L = %d\n", Lstore->nnz);
printf("No of nonzeros in factor U = %d\n", Ustore->nnz);
printf("No of nonzeros in L+U = %d\n", Lstore->nnz + Ustore->nnz - n);
printf("FILL ratio = %.1f\n", (float)(Lstore->nnz + Ustore->nnz - n)/nnz);
dQuerySpace(&L, &U, &mem_usage);
printf("L\\U MB %.3f\ttotal MB needed %.3f\n",
mem_usage.for_lu/1e6, mem_usage.total_needed/1e6);
} else {
printf("dgssv() error returns INFO= %d\n", info);
if ( info <= n ) { /* factorization completes */
dQuerySpace(&L, &U, &mem_usage);
printf("L\\U MB %.3f\ttotal MB needed %.3f\n",
mem_usage.for_lu/1e6, mem_usage.total_needed/1e6);
}
}
if ( options.PrintStat ) StatPrint(&stat);
StatFree(&stat);
SUPERLU_FREE (rhs);
SUPERLU_FREE (xact);
SUPERLU_FREE (perm_r);
SUPERLU_FREE (perm_c);
Destroy_CompCol_Matrix(&A);
Destroy_SuperMatrix_Store(&B);
Destroy_SuperNode_Matrix(&L);
Destroy_CompCol_Matrix(&U);
#if ( DEBUGlevel>=1 )
CHECK_MALLOC("Exit main()");
#endif
}
int main ( int argc, char *argv[] )
/**********************************************************************/
/*
Purpose:
SUPER_LU_S3 solves a sparse system read from a file using SGSSVX.
Discussion:
The sparse matrix is stored in a file using the Harwell-Boeing
sparse matrix format. The file should be assigned to the standard
input of this program. For instance, if the matrix is stored
in the file "g10_rua.txt", the execution command might be:
super_lu_s3 < g10_rua.txt
Modified:
25 April 2004
Reference:
James Demmel, John Gilbert, Xiaoye Li,
SuperLU Users's Guide,
Sections 1 and 2.
Local parameters:
SuperMatrix L, the computed L factor.
int *perm_c, the column permutation vector.
int *perm_r, the row permutations from partial pivoting.
SuperMatrix U, the computed U factor.
*/
{
SuperMatrix A;
NCformat *Astore;
float *a;
int *asub;
SuperMatrix B;
float *berr;
float *C;
char equed[1];
yes_no_t equil;
int *etree;
float *ferr;
int i;
int info;
SuperMatrix L;
int ldx;
SCformat *Lstore;
int lwork;
int m;
mem_usage_t mem_usage;
int n;
int nnz;
int nrhs;
superlu_options_t options;
int *perm_c;
int *perm_r;
float *R;
float rcond;
float *rhsb;
float *rhsx;
float rpg;
float *sol;
SuperLUStat_t stat;
trans_t trans;
SuperMatrix U;
float u;
NCformat *Ustore;
void *work;
SuperMatrix X;
int *xa;
float *xact;
/*
Say hello.
*/
printf ( "\n" );
printf ( "SUPER_LU_S3:\n" );
printf ( " Read a sparse matrix A from standard input,\n");
printf ( " stored in Harwell-Boeing Sparse Matrix format.\n" );
printf ( "\n" );
printf ( " Solve a linear system A * X = B using SGSSVX.\n" );
/*
Defaults
*/
lwork = 0;
nrhs = 1;
equil = YES;
u = 1.0;
trans = NOTRANS;
/*
Set the default input options:
options.Fact = DOFACT;
options.Equil = YES;
options.ColPerm = COLAMD;
options.DiagPivotThresh = 1.0;
options.Trans = NOTRANS;
options.IterRefine = NOREFINE;
options.SymmetricMode = NO;
options.PivotGrowth = NO;
options.ConditionNumber = NO;
options.PrintStat = YES;
*/
set_default_options ( &options );
/*
Can use command line input to modify the defaults.
*/
parse_command_line ( argc, argv, &lwork, &u, &equil, &trans );
options.Equil = equil;
options.DiagPivotThresh = u;
options.Trans = trans;
printf ( "\n" );
printf ( " Length of work array LWORK = %d\n", lwork );
printf ( " Equilibration option EQUIL = %d\n", equil );
printf ( " Diagonal pivot threshhold value U = %f\n", u );
printf ( " Tranpose option TRANS = %d\n", trans );
/*
Add more functionalities that the defaults.
Compute reciprocal pivot growth
*/
options.PivotGrowth = YES;
/*
Compute reciprocal condition number
*/
options.ConditionNumber = YES;
/*
Perform single-precision refinement
*/
options.IterRefine = SINGLE;
if ( 0 < lwork )
{
work = SUPERLU_MALLOC(lwork);
if ( !work )
{
ABORT ( "SUPERLU_MALLOC cannot allocate work[]" );
}
}
/*
Read matrix A from a file in Harwell-Boeing format.
*/
sreadhb ( &m, &n, &nnz, &a, &asub, &xa );
/*
Create storage for a compressed column matrix.
*/
sCreate_CompCol_Matrix ( &A, m, n, nnz, a, asub, xa, SLU_NC, SLU_S, SLU_GE );
Astore = A.Store;
printf ( "\n" );
printf ( " Dimension %dx%d; # nonzeros %d\n", A.nrow, A.ncol, Astore->nnz );
rhsb = floatMalloc ( m * nrhs );
if ( !rhsb )
{
ABORT ( "Malloc fails for rhsb[]." );
}
rhsx = floatMalloc ( m * nrhs );
if ( !rhsx )
{
ABORT ( "Malloc fails for rhsx[]." );
}
sCreate_Dense_Matrix ( &B, m, nrhs, rhsb, m, SLU_DN, SLU_S, SLU_GE );
sCreate_Dense_Matrix ( &X, m, nrhs, rhsx, m, SLU_DN, SLU_S, SLU_GE );
xact = floatMalloc ( n * nrhs );
if ( !xact )
{
ABORT ( "SUPERLU_MALLOC cannot allocate xact[]" );
}
ldx = n;
sGenXtrue ( n, nrhs, xact, ldx );
sFillRHS ( trans, nrhs, xact, ldx, &A, &B );
etree = intMalloc ( n );
if ( !etree )
{
ABORT ( "Malloc fails for etree[]." );
}
perm_c = intMalloc ( n );
if ( !perm_c )
{
ABORT ( "Malloc fails for perm_c[]." );
}
perm_r = intMalloc ( m );
if ( !perm_r )
{
ABORT ( "Malloc fails for perm_r[]." );
}
R = (float *) SUPERLU_MALLOC ( A.nrow * sizeof(float) );
if ( !R )
{
ABORT ( "SUPERLU_MALLOC fails for R[]." );
}
C = (float *) SUPERLU_MALLOC ( A.ncol * sizeof(float) );
if ( !C )
{
ABORT ( "SUPERLU_MALLOC fails for C[]." );
}
ferr = (float *) SUPERLU_MALLOC ( nrhs * sizeof(float) );
if ( !ferr )
{
ABORT ( "SUPERLU_MALLOC fails for ferr[]." );
}
berr = (float *) SUPERLU_MALLOC ( nrhs * sizeof(float) );
if ( !berr )
{
ABORT ( "SUPERLU_MALLOC fails for berr[]." );
}
/*
Initialize the statistics variables.
*/
StatInit(&stat);
/*
Solve the system and compute the condition number and error bounds using SGSSVX.
*/
sgssvx ( &options, &A, perm_c, perm_r, etree, equed, R, C,
&L, &U, work, lwork, &B, &X, &rpg, &rcond, ferr, berr,
&mem_usage, &stat, &info );
printf ( "\n" );
printf ( " SGSSVX returns INFO = %d\n", info );
if ( info == 0 || info == n+1 )
{
sol = (float*) ((DNformat*) X.Store)->nzval;
if ( options.PivotGrowth == YES )
{
printf ( "\n" );
printf ( " Reciprocal pivot growth = %e\n", rpg);
}
if ( options.ConditionNumber == YES )
{
printf ( "\n" );
printf ( " Reciprocal condition number = %e\n", rcond);
}
if ( options.IterRefine != NOREFINE )
{
printf ( "\n" );
printf ( " Iterative Refinement:\n");
printf ( "%8s%8s%16s%16s\n", "rhs", "Steps", "FERR", "BERR");
for ( i = 0; i < nrhs; i++ )
{
printf ( "%8d%8d%16e%16e\n", i+1, stat.RefineSteps, ferr[i], berr[i]);
}
}
Lstore = (SCformat *) L.Store;
Ustore = (NCformat *) U.Store;
printf ( "\n" );
printf ( " Number of nonzeros in factor L = %d\n", Lstore->nnz );
printf ( " Number of nonzeros in factor U = %d\n", Ustore->nnz );
printf ( " Number of nonzeros in L+U = %d\n",
Lstore->nnz + Ustore->nnz - n );
printf ( "\n" );
printf ( " L\\U MB %.3f\ttotal MB needed %.3f\texpansions %d\n",
mem_usage.for_lu/1e6, mem_usage.total_needed/1e6,
mem_usage.expansions );
fflush ( stdout );
}
else if ( info > 0 && lwork == -1 )
{
printf ( "\n" );
printf ( " Estimated memory: %d bytes\n", info - n );
}
if ( options.PrintStat )
{
StatPrint ( &stat );
}
StatFree ( &stat );
SUPERLU_FREE ( rhsb );
SUPERLU_FREE ( rhsx );
SUPERLU_FREE ( xact );
SUPERLU_FREE ( etree );
SUPERLU_FREE ( perm_r );
SUPERLU_FREE ( perm_c );
SUPERLU_FREE ( R );
SUPERLU_FREE ( C );
SUPERLU_FREE ( ferr );
SUPERLU_FREE ( berr );
Destroy_CompCol_Matrix ( &A );
Destroy_SuperMatrix_Store ( &B );
Destroy_SuperMatrix_Store ( &X );
if ( 0 <= lwork )
{
Destroy_SuperNode_Matrix ( &L );
Destroy_CompCol_Matrix ( &U );
}
/*
Say goodbye.
*/
printf ( "\n" );
printf ( "SUPER_LU_S3:\n" );
printf ( " Normal end of execution.\n");
return 0;
}
main(int argc, char *argv[])
{
SuperMatrix A;
NCformat *Astore;
doublecomplex *a;
int *asub, *xa;
int *perm_c; /* column permutation vector */
int *perm_r; /* row permutations from partial pivoting */
SuperMatrix L; /* factor L */
SCformat *Lstore;
SuperMatrix U; /* factor U */
NCformat *Ustore;
SuperMatrix B;
int nrhs, ldx, info, m, n, nnz;
doublecomplex *xact, *rhs;
mem_usage_t mem_usage;
superlu_options_t options;
SuperLUStat_t stat;
#if ( DEBUGlevel>=1 )
CHECK_MALLOC("Enter main()");
#endif
/* Set the default input options:
options.Fact = DOFACT;
options.Equil = YES;
options.ColPerm = COLAMD;
options.DiagPivotThresh = 1.0;
options.Trans = NOTRANS;
options.IterRefine = NOREFINE;
options.SymmetricMode = NO;
options.PivotGrowth = NO;
options.ConditionNumber = NO;
options.PrintStat = YES;
*/
set_default_options(&options);
/* Now we modify the default options to use the symmetric mode. */
options.SymmetricMode = YES;
options.ColPerm = MMD_AT_PLUS_A;
options.DiagPivotThresh = 0.001;
/* Read the matrix in Harwell-Boeing format. */
zreadhb(&m, &n, &nnz, &a, &asub, &xa);
zCreate_CompCol_Matrix(&A, m, n, nnz, a, asub, xa, SLU_NC, SLU_Z, SLU_GE);
Astore = A.Store;
printf("Dimension %dx%d; # nonzeros %d\n", A.nrow, A.ncol, Astore->nnz);
nrhs = 1;
if ( !(rhs = doublecomplexMalloc(m * nrhs)) ) ABORT("Malloc fails for rhs[].");
zCreate_Dense_Matrix(&B, m, nrhs, rhs, m, SLU_DN, SLU_Z, SLU_GE);
xact = doublecomplexMalloc(n * nrhs);
ldx = n;
zGenXtrue(n, nrhs, xact, ldx);
zFillRHS(options.Trans, nrhs, xact, ldx, &A, &B);
if ( !(perm_c = intMalloc(n)) ) ABORT("Malloc fails for perm_c[].");
if ( !(perm_r = intMalloc(m)) ) ABORT("Malloc fails for perm_r[].");
/* Initialize the statistics variables. */
StatInit(&stat);
zgssv(&options, &A, perm_c, perm_r, &L, &U, &B, &stat, &info);
if ( info == 0 ) {
/* This is how you could access the solution matrix. */
doublecomplex *sol = (doublecomplex*) ((DNformat*) B.Store)->nzval;
/* Compute the infinity norm of the error. */
zinf_norm_error(nrhs, &B, xact);
Lstore = (SCformat *) L.Store;
Ustore = (NCformat *) U.Store;
printf("No of nonzeros in factor L = %d\n", Lstore->nnz);
printf("No of nonzeros in factor U = %d\n", Ustore->nnz);
printf("No of nonzeros in L+U = %d\n", Lstore->nnz + Ustore->nnz - n);
zQuerySpace(&L, &U, &mem_usage);
printf("L\\U MB %.3f\ttotal MB needed %.3f\texpansions %d\n",
mem_usage.for_lu/1e6, mem_usage.total_needed/1e6,
mem_usage.expansions);
} else {
printf("zgssv() error returns INFO= %d\n", info);
if ( info <= n ) { /* factorization completes */
zQuerySpace(&L, &U, &mem_usage);
printf("L\\U MB %.3f\ttotal MB needed %.3f\texpansions %d\n",
mem_usage.for_lu/1e6, mem_usage.total_needed/1e6,
mem_usage.expansions);
}
}
if ( options.PrintStat ) StatPrint(&stat);
StatFree(&stat);
SUPERLU_FREE (rhs);
SUPERLU_FREE (xact);
SUPERLU_FREE (perm_r);
SUPERLU_FREE (perm_c);
Destroy_CompCol_Matrix(&A);
Destroy_SuperMatrix_Store(&B);
Destroy_SuperNode_Matrix(&L);
Destroy_CompCol_Matrix(&U);
#if ( DEBUGlevel>=1 )
CHECK_MALLOC("Exit main()");
#endif
}
PetscErrorCode MatSolve_SuperLU_Private(Mat A,Vec b,Vec x)
{
Mat_SuperLU *lu = (Mat_SuperLU*)A->spptr;
PetscScalar *barray,*xarray;
PetscErrorCode ierr;
PetscInt info,i,n=x->map->n;
PetscReal ferr,berr;
PetscFunctionBegin;
if ( lu->lwork == -1 ) {
PetscFunctionReturn(0);
}
lu->B.ncol = 1; /* Set the number of right-hand side */
if (lu->options.Equil && !lu->rhs_dup){
/* superlu overwrites b when Equil is used, thus create rhs_dup to keep user's b unchanged */
ierr = PetscMalloc(n*sizeof(PetscScalar),&lu->rhs_dup);CHKERRQ(ierr);
}
if (lu->options.Equil){
/* Copy b into rsh_dup */
ierr = VecGetArray(b,&barray);CHKERRQ(ierr);
ierr = PetscMemcpy(lu->rhs_dup,barray,n*sizeof(PetscScalar));CHKERRQ(ierr);
ierr = VecRestoreArray(b,&barray);CHKERRQ(ierr);
barray = lu->rhs_dup;
} else {
ierr = VecGetArray(b,&barray);CHKERRQ(ierr);
}
ierr = VecGetArray(x,&xarray);CHKERRQ(ierr);
#if defined(PETSC_USE_COMPLEX)
((DNformat*)lu->B.Store)->nzval = (doublecomplex*)barray;
((DNformat*)lu->X.Store)->nzval = (doublecomplex*)xarray;
#else
((DNformat*)lu->B.Store)->nzval = barray;
((DNformat*)lu->X.Store)->nzval = xarray;
#endif
lu->options.Fact = FACTORED; /* Indicate the factored form of A is supplied. */
if (A->factortype == MAT_FACTOR_LU){
#if defined(PETSC_USE_COMPLEX)
zgssvx(&lu->options, &lu->A, lu->perm_c, lu->perm_r, lu->etree, lu->equed, lu->R, lu->C,
&lu->L, &lu->U, lu->work, lu->lwork, &lu->B, &lu->X, &lu->rpg, &lu->rcond, &ferr, &berr,
&lu->mem_usage, &lu->stat, &info);
#else
dgssvx(&lu->options, &lu->A, lu->perm_c, lu->perm_r, lu->etree, lu->equed, lu->R, lu->C,
&lu->L, &lu->U, lu->work, lu->lwork, &lu->B, &lu->X, &lu->rpg, &lu->rcond, &ferr, &berr,
&lu->mem_usage, &lu->stat, &info);
#endif
} else if (A->factortype == MAT_FACTOR_ILU){
#if defined(PETSC_USE_COMPLEX)
zgsisx(&lu->options, &lu->A, lu->perm_c, lu->perm_r, lu->etree, lu->equed, lu->R, lu->C,
&lu->L, &lu->U, lu->work, lu->lwork, &lu->B, &lu->X, &lu->rpg, &lu->rcond,
&lu->mem_usage, &lu->stat, &info);
#else
dgsisx(&lu->options, &lu->A, lu->perm_c, lu->perm_r, lu->etree, lu->equed, lu->R, lu->C,
&lu->L, &lu->U, lu->work, lu->lwork, &lu->B, &lu->X, &lu->rpg, &lu->rcond,
&lu->mem_usage, &lu->stat, &info);
#endif
} else {
SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Factor type not supported");
}
if (!lu->options.Equil){
ierr = VecRestoreArray(b,&barray);CHKERRQ(ierr);
}
ierr = VecRestoreArray(x,&xarray);CHKERRQ(ierr);
if ( !info || info == lu->A.ncol+1 ) {
if ( lu->options.IterRefine ) {
ierr = PetscPrintf(PETSC_COMM_SELF,"Iterative Refinement:\n");
ierr = PetscPrintf(PETSC_COMM_SELF," %8s%8s%16s%16s\n", "rhs", "Steps", "FERR", "BERR");
for (i = 0; i < 1; ++i)
ierr = PetscPrintf(PETSC_COMM_SELF," %8d%8d%16e%16e\n", i+1, lu->stat.RefineSteps, ferr, berr);
}
} else if ( info > 0 ){
if ( lu->lwork == -1 ) {
ierr = PetscPrintf(PETSC_COMM_SELF," ** Estimated memory: %D bytes\n", info - lu->A.ncol);
} else {
ierr = PetscPrintf(PETSC_COMM_SELF," Warning: gssvx() returns info %D\n",info);
}
} else if (info < 0){
SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB, "info = %D, the %D-th argument in gssvx() had an illegal value", info,-info);
}
if ( lu->options.PrintStat ) {
ierr = PetscPrintf(PETSC_COMM_SELF,"MatSolve__SuperLU():\n");
StatPrint(&lu->stat);
}
PetscFunctionReturn(0);
}
SolveSuperLU (const MatriceMorse<R> &AA, int strategy, double ttgv, double epsilon,
double pivot, double pivot_sym, string ¶m_char, KN<long> pperm_r,
KN<long> pperm_c):
eps(epsilon), epsr(0),
tgv(ttgv),
etree(0), string_option(param_char), perm_r(pperm_r), perm_c(pperm_c),
RR(0), CC(0),
tol_pivot_sym(pivot_sym), tol_pivot(pivot) {
SuperMatrix B, X;
SuperLUStat_t stat;
void *work = 0;
int info, lwork = 0/*, nrhs = 1*/;
int i;
double ferr[1];
double berr[1];
double rpg, rcond;
R *bb;
R *xx;
A.Store = 0;
B.Store = 0;
X.Store = 0;
L.Store = 0;
U.Store = 0;
int status;
n = AA.n;
m = AA.m;
nnz = AA.nbcoef;
arow = AA.a;
asubrow = AA.cl;
xarow = AA.lg;
/* FreeFem++ use Morse Format */
// FFCS - "this->" required by g++ 4.7
this->CompRow_to_CompCol(m, n, nnz, arow, asubrow, xarow,
&a, &asub, &xa);
/* Defaults */
lwork = 0;
// nrhs = 0;
/* Set the default values for options argument:
* options.Fact = DOFACT;
* options.Equil = YES;
* options.ColPerm = COLAMD;
* options.DiagPivotThresh = 1.0;
* options.Trans = NOTRANS;
* options.IterRefine = NOREFINE;
* options.SymmetricMode = NO;
* options.PivotGrowth = NO;
* options.ConditionNumber = NO;
* options.PrintStat = YES;
*/
set_default_options(&options);
printf(".. default options:\n");
printf("\tFact\t %8d\n", options.Fact);
printf("\tEquil\t %8d\n", options.Equil);
printf("\tColPerm\t %8d\n", options.ColPerm);
printf("\tDiagPivotThresh %8.4f\n", options.DiagPivotThresh);
printf("\tTrans\t %8d\n", options.Trans);
printf("\tIterRefine\t%4d\n", options.IterRefine);
printf("\tSymmetricMode\t%4d\n", options.SymmetricMode);
printf("\tPivotGrowth\t%4d\n", options.PivotGrowth);
printf("\tConditionNumber\t%4d\n", options.ConditionNumber);
printf("..\n");
if (!string_option.empty()) {read_options_freefem(string_option, &options);}
printf(".. options:\n");
printf("\tFact\t %8d\n", options.Fact);
printf("\tEquil\t %8d\n", options.Equil);
printf("\tColPerm\t %8d\n", options.ColPerm);
printf("\tDiagPivotThresh %8.4f\n", options.DiagPivotThresh);
printf("\tTrans\t %8d\n", options.Trans);
printf("\tIterRefine\t%4d\n", options.IterRefine);
printf("\tSymmetricMode\t%4d\n", options.SymmetricMode);
printf("\tPivotGrowth\t%4d\n", options.PivotGrowth);
printf("\tConditionNumber\t%4d\n", options.ConditionNumber);
printf("..\n");
Dtype_t R_SLU = SuperLUDriver<R>::R_SLU_T();
// FFCS - "this->" required by g++ 4.7
this->Create_CompCol_Matrix(&A, m, n, nnz, a, asub, xa, SLU_NC, R_SLU, SLU_GE);
this->Create_Dense_Matrix(&B, m, 0, (R *)0, m, SLU_DN, R_SLU, SLU_GE);
this->Create_Dense_Matrix(&X, m, 0, (R *)0, m, SLU_DN, R_SLU, SLU_GE);
if (etree.size() == 0) {etree.resize(n);}
if (perm_r.size() == 0) {perm_r.resize(n);}
if (perm_c.size() == 0) {perm_c.resize(n);}
if (!(RR = new double[n])) {
ABORT("SUPERLU_MALLOC fails for R[].");
}
for (int ii = 0; ii < n; ii++) {
RR[ii] = 1.;
}
if (!(CC = new double[m])) {
ABORT("SUPERLU_MALLOC fails for C[].");
}
for (int ii = 0; ii < n; ii++) {
CC[ii] = 1.;
}
ferr[0] = 0;
berr[0] = 0;
/* Initialize the statistics variables. */
StatInit(&stat);
/* ONLY PERFORM THE LU DECOMPOSITION */
B.ncol = 0; /* Indicate not to solve the system */
SuperLUDriver<R>::gssvx(&options, &A, perm_c, perm_r, etree, equed, RR, CC,
&L, &U, work, lwork, &B, &X, &rpg, &rcond, ferr, berr, &Glu,
&mem_usage, &stat, &info);
if (verbosity > 2) {
printf("LU factorization: dgssvx() returns info %d\n", info);
}
if (verbosity > 3) {
if (info == 0 || info == n + 1) {
if (options.PivotGrowth) {printf("Recip. pivot growth = %e\n", rpg);}
if (options.ConditionNumber) {
printf("Recip. condition number = %e\n", rcond);
}
Lstore = (SCformat *)L.Store;
Ustore = (NCformat *)U.Store;
printf("No of nonzeros in factor L = %d\n", Lstore->nnz);
printf("No of nonzeros in factor U = %d\n", Ustore->nnz);
printf("No of nonzeros in L+U = %d\n", Lstore->nnz + Ustore->nnz - n);
printf("L\\U MB %.3f\ttotal MB needed %.3f\texpansions %d\n",
mem_usage.for_lu / 1e6, mem_usage.total_needed / 1e6,
stat.expansions
);
fflush(stdout);
} else if (info > 0 && lwork == -1) {
printf("** Estimated memory: %d bytes\n", info - n);
}
}
if (verbosity > 5) {StatPrint(&stat);}
StatFree(&stat);
if (B.Store) {Destroy_SuperMatrix_Store(&B);}
if (X.Store) {Destroy_SuperMatrix_Store(&X);}
options.Fact = FACTORED;/* Indicate the factored form of A is supplied. */
}
PetscErrorCode MatLUFactorNumeric_SuperLU(Mat F,Mat A,const MatFactorInfo *info)
{
Mat_SuperLU *lu = (Mat_SuperLU*)F->spptr;
Mat_SeqAIJ *aa;
PetscErrorCode ierr;
PetscInt sinfo;
PetscReal ferr, berr;
NCformat *Ustore;
SCformat *Lstore;
PetscFunctionBegin;
if (lu->flg == SAME_NONZERO_PATTERN){ /* successing numerical factorization */
lu->options.Fact = SamePattern;
/* Ref: ~SuperLU_3.0/EXAMPLE/dlinsolx2.c */
Destroy_SuperMatrix_Store(&lu->A);
if (lu->options.Equil){
ierr = MatCopy_SeqAIJ(A,lu->A_dup,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
}
if ( lu->lwork >= 0 ) {
Destroy_SuperNode_Matrix(&lu->L);
Destroy_CompCol_Matrix(&lu->U);
lu->options.Fact = SamePattern;
}
}
/* Create the SuperMatrix for lu->A=A^T:
Since SuperLU likes column-oriented matrices,we pass it the transpose,
and then solve A^T X = B in MatSolve(). */
if (lu->options.Equil){
aa = (Mat_SeqAIJ*)(lu->A_dup)->data;
} else {
aa = (Mat_SeqAIJ*)(A)->data;
}
#if defined(PETSC_USE_COMPLEX)
zCreate_CompCol_Matrix(&lu->A,A->cmap->n,A->rmap->n,aa->nz,(doublecomplex*)aa->a,aa->j,aa->i,
SLU_NC,SLU_Z,SLU_GE);
#else
dCreate_CompCol_Matrix(&lu->A,A->cmap->n,A->rmap->n,aa->nz,aa->a,aa->j,aa->i,
SLU_NC,SLU_D,SLU_GE);
#endif
/* Numerical factorization */
lu->B.ncol = 0; /* Indicate not to solve the system */
if (F->factortype == MAT_FACTOR_LU){
#if defined(PETSC_USE_COMPLEX)
zgssvx(&lu->options, &lu->A, lu->perm_c, lu->perm_r, lu->etree, lu->equed, lu->R, lu->C,
&lu->L, &lu->U, lu->work, lu->lwork, &lu->B, &lu->X, &lu->rpg, &lu->rcond, &ferr, &berr,
&lu->mem_usage, &lu->stat, &sinfo);
#else
dgssvx(&lu->options, &lu->A, lu->perm_c, lu->perm_r, lu->etree, lu->equed, lu->R, lu->C,
&lu->L, &lu->U, lu->work, lu->lwork, &lu->B, &lu->X, &lu->rpg, &lu->rcond, &ferr, &berr,
&lu->mem_usage, &lu->stat, &sinfo);
#endif
} else if (F->factortype == MAT_FACTOR_ILU){
/* Compute the incomplete factorization, condition number and pivot growth */
#if defined(PETSC_USE_COMPLEX)
zgsisx(&lu->options, &lu->A, lu->perm_c, lu->perm_r,lu->etree, lu->equed, lu->R, lu->C,
&lu->L, &lu->U, lu->work, lu->lwork, &lu->B, &lu->X, &lu->rpg, &lu->rcond,
&lu->mem_usage, &lu->stat, &sinfo);
#else
dgsisx(&lu->options, &lu->A, lu->perm_c, lu->perm_r, lu->etree, lu->equed, lu->R, lu->C,
&lu->L, &lu->U, lu->work, lu->lwork, &lu->B, &lu->X, &lu->rpg, &lu->rcond,
&lu->mem_usage, &lu->stat, &sinfo);
#endif
} else {
SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Factor type not supported");
}
if ( !sinfo || sinfo == lu->A.ncol+1 ) {
if ( lu->options.PivotGrowth )
ierr = PetscPrintf(PETSC_COMM_SELF," Recip. pivot growth = %e\n", lu->rpg);
if ( lu->options.ConditionNumber )
ierr = PetscPrintf(PETSC_COMM_SELF," Recip. condition number = %e\n", lu->rcond);
} else if ( sinfo > 0 ){
if ( lu->lwork == -1 ) {
ierr = PetscPrintf(PETSC_COMM_SELF," ** Estimated memory: %D bytes\n", sinfo - lu->A.ncol);
} else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot in row %D",sinfo);
} else { /* sinfo < 0 */
SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB, "info = %D, the %D-th argument in gssvx() had an illegal value", sinfo,-sinfo);
}
if ( lu->options.PrintStat ) {
ierr = PetscPrintf(PETSC_COMM_SELF,"MatLUFactorNumeric_SuperLU():\n");
StatPrint(&lu->stat);
Lstore = (SCformat *) lu->L.Store;
Ustore = (NCformat *) lu->U.Store;
ierr = PetscPrintf(PETSC_COMM_SELF," No of nonzeros in factor L = %D\n", Lstore->nnz);
ierr = PetscPrintf(PETSC_COMM_SELF," No of nonzeros in factor U = %D\n", Ustore->nnz);
ierr = PetscPrintf(PETSC_COMM_SELF," No of nonzeros in L+U = %D\n", Lstore->nnz + Ustore->nnz - lu->A.ncol);
ierr = PetscPrintf(PETSC_COMM_SELF," L\\U MB %.3f\ttotal MB needed %.3f\n",
lu->mem_usage.for_lu/1e6, lu->mem_usage.total_needed/1e6);
}
lu->flg = SAME_NONZERO_PATTERN;
F->ops->solve = MatSolve_SuperLU;
F->ops->solvetranspose = MatSolveTranspose_SuperLU;
F->ops->matsolve = MatMatSolve_SuperLU;
PetscFunctionReturn(0);
}
bool SparseMatrix::solveSLU (Vector& B)
{
int ierr = ncol+1;
if (!factored) this->optimiseSLU();
#ifdef HAS_SUPERLU_MT
if (!slu) {
// Create a new SuperLU matrix
slu = new SuperLUdata;
slu->perm_c = new int[ncol];
slu->perm_r = new int[nrow];
dCreate_CompCol_Matrix(&slu->A, nrow, ncol, this->size(),
&A.front(), &JA.front(), &IA.front(),
SLU_NC, SLU_D, SLU_GE);
}
else {
Destroy_SuperMatrix_Store(&slu->A);
Destroy_SuperNode_Matrix(&slu->L);
Destroy_CompCol_Matrix(&slu->U);
dCreate_CompCol_Matrix(&slu->A, nrow, ncol, this->size(),
&A.front(), &JA.front(), &IA.front(),
SLU_NC, SLU_D, SLU_GE);
}
// Get column permutation vector perm_c[], according to permc_spec:
// permc_spec = 0: natural ordering
// permc_spec = 1: minimum degree ordering on structure of A'*A
// permc_spec = 2: minimum degree ordering on structure of A'+A
// permc_spec = 3: approximate minimum degree for unsymmetric matrices
int permc_spec = 1;
get_perm_c(permc_spec, &slu->A, slu->perm_c);
// Create right-hand-side/solution vector(s)
size_t nrhs = B.size() / nrow;
SuperMatrix Bmat;
dCreate_Dense_Matrix(&Bmat, nrow, nrhs, B.ptr(), nrow,
SLU_DN, SLU_D, SLU_GE);
// Invoke the simple driver
pdgssv(numThreads, &slu->A, slu->perm_c, slu->perm_r,
&slu->L, &slu->U, &Bmat, &ierr);
if (ierr > 0)
std::cerr <<"SuperLU_MT Failure "<< ierr << std::endl;
Destroy_SuperMatrix_Store(&Bmat);
#elif defined(HAS_SUPERLU)
if (!slu) {
// Create a new SuperLU matrix
slu = new SuperLUdata(1);
slu->perm_c = new int[ncol];
slu->perm_r = new int[nrow];
dCreate_CompCol_Matrix(&slu->A, nrow, ncol, this->size(),
&A.front(), &JA.front(), &IA.front(),
SLU_NC, SLU_D, SLU_GE);
}
else if (factored)
slu->opts->Fact = FACTORED; // Re-use previous factorization
else {
Destroy_SuperMatrix_Store(&slu->A);
Destroy_SuperNode_Matrix(&slu->L);
Destroy_CompCol_Matrix(&slu->U);
dCreate_CompCol_Matrix(&slu->A, nrow, ncol, this->size(),
&A.front(), &JA.front(), &IA.front(),
SLU_NC, SLU_D, SLU_GE);
}
// Create right-hand-side/solution vector(s)
size_t nrhs = B.size() / nrow;
SuperMatrix Bmat;
dCreate_Dense_Matrix(&Bmat, nrow, nrhs, B.ptr(), nrow,
SLU_DN, SLU_D, SLU_GE);
SuperLUStat_t stat;
StatInit(&stat);
// Invoke the simple driver
dgssv(slu->opts, &slu->A, slu->perm_c, slu->perm_r,
&slu->L, &slu->U, &Bmat, &stat, &ierr);
if (ierr > 0)
std::cerr <<"SuperLU Failure "<< ierr << std::endl;
else
factored = true;
if (printSLUstat)
StatPrint(&stat);
StatFree(&stat);
Destroy_SuperMatrix_Store(&Bmat);
#else
std::cerr <<"SparseMatrix::solve: SuperLU solver not available"<< std::endl;
#endif
return ierr == 0;
}
bool SparseMatrix::solveSLUx (Vector& B, Real* rcond)
{
int ierr = ncol+1;
if (!factored) this->optimiseSLU();
#ifdef HAS_SUPERLU_MT
if (!slu) {
// Create a new SuperLU matrix
slu = new SuperLUdata(numThreads);
slu->equed = NOEQUIL;
slu->perm_c = new int[ncol];
slu->perm_r = new int[nrow];
slu->C = new Real[ncol];
slu->R = new Real[nrow];
slu->opts->etree = new int[ncol];
slu->opts->colcnt_h = new int[ncol];
slu->opts->part_super_h = new int[ncol];
memset(slu->opts->colcnt_h, 0, ncol*sizeof(int));
memset(slu->opts->part_super_h, 0, ncol*sizeof(int));
memset(slu->opts->etree, 0, ncol*sizeof(int));
dCreate_CompCol_Matrix(&slu->A, nrow, ncol, this->size(),
&A.front(), &JA.front(), &IA.front(),
SLU_NC, SLU_D, SLU_GE);
// Get column permutation vector perm_c[], according to permc_spec:
// permc_spec = 0: natural ordering
// permc_spec = 1: minimum degree ordering on structure of A'*A
// permc_spec = 2: minimum degree ordering on structure of A'+A
// permc_spec = 3: approximate minimum degree for unsymmetric matrices
int permc_spec = 1;
get_perm_c(permc_spec, &slu->A, slu->perm_c);
}
else if (factored)
slu->opts->fact = FACTORED; // Re-use previous factorization
else
slu->opts->refact = YES; // Re-use previous ordering
// Create right-hand-side and solution vector(s)
Vector X(B.size());
SuperMatrix Bmat, Xmat;
const size_t nrhs = B.size() / nrow;
dCreate_Dense_Matrix(&Bmat, nrow, nrhs, B.ptr(), nrow,
SLU_DN, SLU_D, SLU_GE);
dCreate_Dense_Matrix(&Xmat, nrow, nrhs, X.ptr(), nrow,
SLU_DN, SLU_D, SLU_GE);
Real ferr[nrhs], berr[nrhs];
superlu_memusage_t mem_usage;
// Invoke the expert driver
pdgssvx(numThreads, slu->opts, &slu->A, slu->perm_c, slu->perm_r,
&slu->equed, slu->R, slu->C, &slu->L, &slu->U, &Bmat, &Xmat,
&slu->rpg, &slu->rcond, ferr, berr, &mem_usage, &ierr);
B.swap(X);
if (ierr > 0)
std::cerr <<"SuperLU_MT Failure "<< ierr << std::endl;
else if (!factored)
{
factored = true;
if (rcond)
*rcond = slu->rcond;
}
Destroy_SuperMatrix_Store(&Bmat);
Destroy_SuperMatrix_Store(&Xmat);
#elif defined(HAS_SUPERLU)
if (!slu) {
// Create a new SuperLU matrix
slu = new SuperLUdata(1);
slu->perm_c = new int[ncol];
slu->perm_r = new int[nrow];
slu->etree = new int[ncol];
slu->C = new Real[ncol];
slu->R = new Real[nrow];
dCreate_CompCol_Matrix(&slu->A, nrow, ncol, this->size(),
&A.front(), &JA.front(), &IA.front(),
SLU_NC, SLU_D, SLU_GE);
}
else if (factored)
slu->opts->Fact = FACTORED; // Re-use previous factorization
else {
Destroy_SuperMatrix_Store(&slu->A);
Destroy_SuperNode_Matrix(&slu->L);
Destroy_CompCol_Matrix(&slu->U);
dCreate_CompCol_Matrix(&slu->A, nrow, ncol, this->size(),
&A.front(), &JA.front(), &IA.front(),
SLU_NC, SLU_D, SLU_GE);
}
// Create right-hand-side vector and solution vector
Vector X(B.size());
SuperMatrix Bmat, Xmat;
const size_t nrhs = B.size() / nrow;
dCreate_Dense_Matrix(&Bmat, nrow, nrhs, B.ptr(), nrow,
SLU_DN, SLU_D, SLU_GE);
dCreate_Dense_Matrix(&Xmat, nrow, nrhs, X.ptr(), nrow,
SLU_DN, SLU_D, SLU_GE);
slu->opts->ConditionNumber = printSLUstat || rcond ? YES : NO;
slu->opts->PivotGrowth = printSLUstat ? YES : NO;
void* work = 0;
int lwork = 0;
Real ferr[nrhs], berr[nrhs];
mem_usage_t mem_usage;
SuperLUStat_t stat;
StatInit(&stat);
// Invoke the expert driver
#if SUPERLU_VERSION == 5
GlobalLU_t Glu;
dgssvx(slu->opts, &slu->A, slu->perm_c, slu->perm_r, slu->etree, slu->equed,
slu->R, slu->C, &slu->L, &slu->U, work, lwork, &Bmat, &Xmat,
&slu->rpg, &slu->rcond, ferr, berr, &Glu, &mem_usage, &stat, &ierr);
#else
dgssvx(slu->opts, &slu->A, slu->perm_c, slu->perm_r, slu->etree, slu->equed,
slu->R, slu->C, &slu->L, &slu->U, work, lwork, &Bmat, &Xmat,
&slu->rpg, &slu->rcond, ferr, berr, &mem_usage, &stat, &ierr);
#endif
B.swap(X);
if (ierr > 0)
std::cerr <<"SuperLU Failure "<< ierr << std::endl;
else if (!factored)
{
factored = true;
if (rcond)
*rcond = slu->rcond;
}
if (printSLUstat)
{
StatPrint(&stat);
IFEM::cout <<"Reciprocal condition number = "<< slu->rcond
<<"\nReciprocal pivot growth = "<< slu->rpg << std::endl;
}
StatFree(&stat);
Destroy_SuperMatrix_Store(&Bmat);
Destroy_SuperMatrix_Store(&Xmat);
#else
std::cerr <<"SparseMatrix::solve: SuperLU solver not available"<< std::endl;
#endif
return ierr == 0;
}
void RKWidget::step ( const size_t nStep )
{
DNformat *Bstore;
DNformat *Xstore;
double temp;
if(unstable) return;
samset = false;
if( dirty ) {
if(aexist) {
Destroy_CompCol_Matrix(&A);
if ( lwork == 0 ) {
Destroy_SuperNode_Matrix(&L);
Destroy_CompCol_Matrix(&Up);
} else if ( lwork > 0 ) {
SUPERLU_FREE(work);
}
// these may be freed in dgssvx or Destroy_CompCol_Matrix I think
aexist = false;
}
a = new double[nvar*N_];
xa = new int[N_+1];
asub = new int[nvar*N_];
updateCoef(method);
if(nblock == 1) {
// load with coef[] ???
fillA();
} else if (nup+ndn == 0) {
// solve seperate independent blocks???
} else {
// load block system
blockFillA();
}
dCreate_CompCol_Matrix(&A, N_, N_, nnz, a, asub, xa, SLU_NC, SLU_D, SLU_GE);
aexist = true;
/* Initialize the statistics variables. */
StatInit(&stat);
dPrint_CompCol_Matrix("A matrix", &A);
options.Fact = DOFACT;
//options.ColPerm=NATURAL;
options.ColPerm=COLAMD;
options.PivotGrowth = NO;
options.ConditionNumber = NO;
/* ONLY PERFORM THE LU DECOMPOSITION */
B.ncol = 0; /* Indicate not to solve the system */
dgssvx(&options, &A, perm_c, perm_r, etree, equed, R, C,
&L, &Up, work, lwork, &B, &X, &rpg, &rcond, ferr, berr,
&mem_usage, &stat, &info);
//dPrint_CompCol_Matrix("A matrix", &A);
printf("LU factorization: dgssvx() returns info %d\n", info);
if ( info == 0 || info == N_+1 ) {
if ( options.PivotGrowth ) printf("Recip. pivot growth = %e\n", rpg);
if ( options.ConditionNumber )
printf("Recip. condition number = %e\n", rcond);
printf("L\\U_ MB %.3f\ttotal MB needed %.3f\n",
mem_usage.for_lu/1e6, mem_usage.total_needed/1e6);
fflush(stdout);
options.Fact = FACTORED; /* Indicate the factored form of A is supplied. */
B.ncol = 1;
dirty = false;
} else if ( info > 0 && lwork == -1 ) {
printf("** Estimated memory: %d bytes\n", info - n);
}
if ( options.PrintStat ) StatPrint(&stat);
StatFree(&stat);
}
for ( size_t n = 0; n < nStep; n++ ) {
for( int ns = 0; ns < nStage; ns++ ) {
///set B matrix
Bstore= (DNformat *)B.Store;
rhsb=(double*)Bstore->nzval;
if (nup+ndn == 0) {
// solve seperate independent blocks???
std::cout << "Discontinuous Galerkin Not Implimented\n";
return;
} else {
fillB(ns);
}
///solve factored system
StatInit(&stat);
options.Fact = FACTORED;
dgssvx(&options, &A, perm_c, perm_r, etree, equed, R, C,
&L, &Up, work, lwork, &B, &X, &rpg, &rcond, ferr, berr,
&mem_usage, &stat, &info);
//if( n == 1 ) printf("Triangular solve: dgssvx() returns info %d\n", info);
///update U_
if ( info == 0 || info == N_+1 ) {
/* This is how you could access the solution matrix. */
Xstore = (DNformat *)X.Store;
rhsx = (double*)(Xstore->nzval);
for ( size_t i = 0; i < N_ ; i++ ) {
if(rhsx[i]> 1e16) unstable = true;
b_k[i+N_*ns]=rhsx[i];
}
} else {
std::cout << "ERROR: Matrix Solution Failed info = " << info << std::endl;
}
StatFree(&stat);
}
for(int j=0 ; j<nStage; j++) {
temp=b_b[j];
if( temp != 0 ) {
for(size_t i = 0; i < N_ ; i++ ) {
U_[i] += temp*b_k[i+j*N_];
}
}
}
cStep++;
totCFL += CFL;
}
}
int main ( int argc, char *argv[] )
/**********************************************************************/
/*
Purpose:
SUPER_LU_S2 solves a symmetric sparse system read from a file.
Discussion:
The sparse matrix is stored in a file using the Harwell-Boeing
sparse matrix format. The file should be assigned to the standard
input of this program. For instance, if the matrix is stored
in the file "g10_rua.txt", the execution command might be:
super_lu_s2 < g10_rua.txt
Modified:
25 April 2004
Reference:
James Demmel, John Gilbert, Xiaoye Li,
SuperLU Users's Guide,
Sections 1 and 2.
Local parameters:
SuperMatrix L, the computed L factor.
int *perm_c, the column permutation vector.
int *perm_r, the row permutations from partial pivoting.
SuperMatrix U, the computed U factor.
*/
{
SuperMatrix A;
NCformat *Astore;
float *a;
int *asub;
SuperMatrix B;
int info;
SuperMatrix L;
int ldx;
SCformat *Lstore;
int m;
mem_usage_t mem_usage;
int n;
int nnz;
int nrhs;
superlu_options_t options;
int *perm_c;
int *perm_r;
float *rhs;
float *sol;
SuperLUStat_t stat;
SuperMatrix U;
NCformat *Ustore;
int *xa;
float *xact;
/*
Say hello.
*/
printf ( "\n" );
printf ( "SUPER_LU_S2:\n" );
printf ( " Read a symmetric sparse matrix A from standard input,\n");
printf ( " stored in Harwell-Boeing Sparse Matrix format.\n" );
printf ( "\n" );
printf ( " Solve a linear system A * X = B.\n" );
/*
Set the default input options:
options.Fact = DOFACT;
options.Equil = YES;
options.ColPerm = COLAMD;
options.DiagPivotThresh = 1.0;
options.Trans = NOTRANS;
options.IterRefine = NOREFINE;
options.SymmetricMode = NO;
options.PivotGrowth = NO;
options.ConditionNumber = NO;
options.PrintStat = YES;
*/
set_default_options ( &options );
/*
Now we modify the default options to use the symmetric mode.
*/
options.SymmetricMode = YES;
options.ColPerm = MMD_AT_PLUS_A;
options.DiagPivotThresh = 0.001;
/*
Read the matrix in Harwell-Boeing format.
*/
sreadhb ( &m, &n, &nnz, &a, &asub, &xa );
/*
Create storage for a compressed column matrix.
*/
sCreate_CompCol_Matrix ( &A, m, n, nnz, a, asub, xa, SLU_NC, SLU_S, SLU_GE );
Astore = A.Store;
printf ( "\n" );
printf ( " Dimension %dx%d; # nonzeros %d\n", A.nrow, A.ncol, Astore->nnz );
/*
Set up the right hand side.
*/
nrhs = 1;
rhs = floatMalloc ( m * nrhs );
if ( !rhs )
{
ABORT ( " Malloc fails for rhs[]." );
}
sCreate_Dense_Matrix ( &B, m, nrhs, rhs, m, SLU_DN, SLU_S, SLU_GE );
xact = floatMalloc ( n * nrhs );
if ( !xact )
{
ABORT ( " Malloc fails for rhs[]." );
}
ldx = n;
sGenXtrue ( n, nrhs, xact, ldx );
sFillRHS ( options.Trans, nrhs, xact, ldx, &A, &B );
perm_c = intMalloc ( n );
if ( !perm_c )
{
ABORT ( "Malloc fails for perm_c[]." );
}
perm_r = intMalloc ( m );
if ( !perm_r )
{
ABORT ( "Malloc fails for perm_r[]." );
}
/*
Initialize the statistics variables.
*/
StatInit ( &stat );
/*
Call SGSSV to factor the matrix and solve the linear system.
*/
sgssv ( &options, &A, perm_c, perm_r, &L, &U, &B, &stat, &info );
if ( info == 0 )
{
/*
To conveniently access the solution matrix, you need to get a pointer to it.
*/
sol = (float*) ((DNformat*) B.Store)->nzval;
/*
Compute the infinity norm of the error.
*/
sinf_norm_error ( nrhs, &B, xact );
Lstore = (SCformat *) L.Store;
Ustore = (NCformat *) U.Store;
printf ( "\n" );
printf ( " Number of nonzeros in factor L = %d\n", Lstore->nnz );
printf ( " Number of nonzeros in factor U = %d\n", Ustore->nnz );
printf ( " Number of nonzeros in L+U = %d\n",
Lstore->nnz + Ustore->nnz - n );
sQuerySpace ( &L, &U, &mem_usage );
printf ( "\n" );
printf ( " L\\U MB %.3f\ttotal MB needed %.3f\texpansions %d\n",
mem_usage.for_lu/1e6, mem_usage.total_needed/1e6,
mem_usage.expansions);
}
else
{
printf ( "\n" );
printf ( " SGSSV error returns INFO= %d\n", info );
if ( info <= n )
{
sQuerySpace ( &L, &U, &mem_usage );
printf ( "\n" );
printf (" L\\U MB %.3f\ttotal MB needed %.3f\texpansions %d\n",
mem_usage.for_lu/1e6, mem_usage.total_needed/1e6,
mem_usage.expansions );
}
}
if ( options.PrintStat )
{
StatPrint ( &stat );
}
StatFree ( &stat );
/*
Free the memory.
*/
SUPERLU_FREE ( rhs );
SUPERLU_FREE ( xact );
SUPERLU_FREE ( perm_r );
SUPERLU_FREE ( perm_c );
Destroy_CompCol_Matrix ( &A );
Destroy_SuperMatrix_Store ( &B );
Destroy_SuperNode_Matrix ( &L );
Destroy_CompCol_Matrix ( &U );
/*
Say goodbye.
*/
printf ( "\n" );
printf ( "SUPER_LU_S2:\n" );
printf ( " Normal end of execution.\n");
return 0;
}
