/* Here is a driver inspired by A. Sheffer's "cow flattener". */
static NLboolean __nlFactorize_SUPERLU(__NLContext *context, NLint *permutation) {
/* OpenNL Context */
__NLSparseMatrix* M = (context->least_squares)? &context->MtM: &context->M;
NLuint n = context->n;
NLuint nnz = __nlSparseMatrixNNZ(M); /* number of non-zero coeffs */
/* Compressed Row Storage matrix representation */
NLint *xa = __NL_NEW_ARRAY(NLint, n+1);
NLfloat *rhs = __NL_NEW_ARRAY(NLfloat, n);
NLfloat *a = __NL_NEW_ARRAY(NLfloat, nnz);
NLint *asub = __NL_NEW_ARRAY(NLint, nnz);
NLint *etree = __NL_NEW_ARRAY(NLint, n);
/* SuperLU variables */
SuperMatrix At, AtP;
NLint info, panel_size, relax;
superlu_options_t options;
/* Temporary variables */
NLuint i, jj, count;
__nl_assert(!(M->storage & __NL_SYMMETRIC));
__nl_assert(M->storage & __NL_ROWS);
__nl_assert(M->m == M->n);
/* Convert M to compressed column format */
for(i=0, count=0; i<n; i++) {
__NLRowColumn *Ri = M->row + i;
xa[i] = count;
for(jj=0; jj<Ri->size; jj++, count++) {
a[count] = Ri->coeff[jj].value;
asub[count] = Ri->coeff[jj].index;
}
}
xa[n] = nnz;
/* Free M, don't need it anymore at this point */
__nlSparseMatrixClear(M);
/* Create superlu A matrix transposed */
sCreate_CompCol_Matrix(
&At, n, n, nnz, a, asub, xa,
SLU_NC, /* Colum wise, no supernode */
SLU_S, /* floats */
SLU_GE /* general storage */
);
/* Set superlu options */
set_default_options(&options);
options.ColPerm = MY_PERMC;
options.Fact = DOFACT;
StatInit(&(context->slu.stat));
panel_size = sp_ienv(1); /* sp_ienv give us the defaults */
relax = sp_ienv(2);
/* Compute permutation and permuted matrix */
context->slu.perm_r = __NL_NEW_ARRAY(NLint, n);
context->slu.perm_c = __NL_NEW_ARRAY(NLint, n);
if ((permutation == NULL) || (*permutation == -1)) {
get_perm_c(3, &At, context->slu.perm_c);
if (permutation)
memcpy(permutation, context->slu.perm_c, sizeof(NLint)*n);
}
else
memcpy(context->slu.perm_c, permutation, sizeof(NLint)*n);
sp_preorder(&options, &At, context->slu.perm_c, etree, &AtP);
/* Decompose into L and U */
sgstrf(&options, &AtP, relax, panel_size,
etree, NULL, 0, context->slu.perm_c, context->slu.perm_r,
&(context->slu.L), &(context->slu.U), &(context->slu.stat), &info);
/* Cleanup */
Destroy_SuperMatrix_Store(&At);
Destroy_CompCol_Permuted(&AtP);
__NL_DELETE_ARRAY(etree);
__NL_DELETE_ARRAY(xa);
__NL_DELETE_ARRAY(rhs);
__NL_DELETE_ARRAY(a);
__NL_DELETE_ARRAY(asub);
context->slu.alloc_slu = NL_TRUE;
return (info == 0);
}
NLboolean nlVariableIsLocked(NLuint index) {
__nl_assert(__nlCurrentContext->state != __NL_STATE_INITIAL);
__nl_parano_range_assert(index, 0, __nlCurrentContext->nb_variables - 1);
return __nlCurrentContext->variable[index].locked;
}
static void __nlCheckState(NLenum state) {
__nl_assert(__nlCurrentContext->state == state);
}
NLfloat nlGetVariable(NLuint rhsindex, NLuint index) {
__nl_assert(__nlCurrentContext->state != __NL_STATE_INITIAL);
__nl_parano_range_assert(index, 0, __nlCurrentContext->nb_variables - 1);
return __nlCurrentContext->variable[index].value[rhsindex];
}
/* Here is a driver inspired by A. Sheffer's "cow flattener". */
static NLboolean __nlSolve_SUPERLU( NLboolean do_perm) {
/* OpenNL Context */
__NLSparseMatrix* M = &(__nlCurrentContext->M);
NLfloat* b = __nlCurrentContext->b;
NLfloat* x = __nlCurrentContext->x;
/* Compressed Row Storage matrix representation */
NLuint n = __nlCurrentContext->n;
NLuint nnz = __nlSparseMatrixNNZ(M); /* Number of Non-Zero coeffs */
NLint* xa = __NL_NEW_ARRAY(NLint, n+1);
NLfloat* rhs = __NL_NEW_ARRAY(NLfloat, n);
NLfloat* a = __NL_NEW_ARRAY(NLfloat, nnz);
NLint* asub = __NL_NEW_ARRAY(NLint, nnz);
/* Permutation vector */
NLint* perm_r = __NL_NEW_ARRAY(NLint, n);
NLint* perm = __NL_NEW_ARRAY(NLint, n);
/* SuperLU variables */
SuperMatrix A, B; /* System */
SuperMatrix L, U; /* Inverse of A */
NLint info; /* status code */
DNformat *vals = NULL; /* access to result */
float *rvals = NULL; /* access to result */
/* SuperLU options and stats */
superlu_options_t options;
SuperLUStat_t stat;
/* Temporary variables */
__NLRowColumn* Ri = NULL;
NLuint i,jj,count;
__nl_assert(!(M->storage & __NL_SYMMETRIC));
__nl_assert(M->storage & __NL_ROWS);
__nl_assert(M->m == M->n);
/*
* Step 1: convert matrix M into SuperLU compressed column
* representation.
* -------------------------------------------------------
*/
count = 0;
for(i=0; i<n; i++) {
Ri = &(M->row[i]);
xa[i] = count;
for(jj=0; jj<Ri->size; jj++) {
a[count] = Ri->coeff[jj].value;
asub[count] = Ri->coeff[jj].index;
count++;
}
}
xa[n] = nnz;
/* Save memory for SuperLU */
__nlSparseMatrixClear(M);
/*
* Rem: symmetric storage does not seem to work with
* SuperLU ... (->deactivated in main SLS::Solver driver)
*/
sCreate_CompCol_Matrix(
&A, n, n, nnz, a, asub, xa,
SLU_NR, /* Row_wise, no supernode */
SLU_S, /* floats */
SLU_GE /* general storage */
);
/* Step 2: create vector */
sCreate_Dense_Matrix(
&B, n, 1, b, n,
SLU_DN, /* Fortran-type column-wise storage */
SLU_S, /* floats */
SLU_GE /* general */
);
/* Step 3: get permutation matrix
* ------------------------------
* com_perm: 0 -> no re-ordering
* 1 -> re-ordering for A^t.A
* 2 -> re-ordering for A^t+A
* 3 -> approximate minimum degree ordering
*/
get_perm_c(do_perm ? 3 : 0, &A, perm);
/* Step 4: call SuperLU main routine
* ---------------------------------
*/
set_default_options(&options);
options.ColPerm = MY_PERMC;
StatInit(&stat);
sgssv(&options, &A, perm, perm_r, &L, &U, &B, &stat, &info);
/* Step 5: get the solution
* ------------------------
* Fortran-type column-wise storage
*/
vals = (DNformat*)B.Store;
rvals = (float*)(vals->nzval);
if(info == 0) {
for(i = 0; i < n; i++){
x[i] = rvals[i];
}
}
/* Step 6: cleanup
* ---------------
*/
/*
* For these two ones, only the "store" structure
* needs to be deallocated (the arrays have been allocated
* by us).
*/
Destroy_SuperMatrix_Store(&A);
Destroy_SuperMatrix_Store(&B);
/*
* These ones need to be fully deallocated (they have been
* allocated by SuperLU).
*/
Destroy_SuperNode_Matrix(&L);
Destroy_CompCol_Matrix(&U);
StatFree(&stat);
__NL_DELETE_ARRAY(xa);
__NL_DELETE_ARRAY(rhs);
__NL_DELETE_ARRAY(a);
__NL_DELETE_ARRAY(asub);
__NL_DELETE_ARRAY(perm_r);
__NL_DELETE_ARRAY(perm);
return (info == 0);
}