void
zgstrs (trans_t trans, SuperMatrix *L, SuperMatrix *U,
int *perm_c, int *perm_r, SuperMatrix *B,
SuperLUStat_t *stat, int *info)
{
/*
* Purpose
* =======
*
* ZGSTRS solves a system of linear equations A*X=B or A'*X=B
* with A sparse and B dense, using the LU factorization computed by
* ZGSTRF.
*
* See supermatrix.h for the definition of 'SuperMatrix' structure.
*
* Arguments
* =========
*
* trans (input) trans_t
* Specifies the form of the system of equations:
* = NOTRANS: A * X = B (No transpose)
* = TRANS: A'* X = B (Transpose)
* = CONJ: A**H * X = B (Conjugate transpose)
*
* L (input) SuperMatrix*
* The factor L from the factorization Pr*A*Pc=L*U as computed by
* zgstrf(). Use compressed row subscripts storage for supernodes,
* i.e., L has types: Stype = SLU_SC, Dtype = SLU_Z, Mtype = SLU_TRLU.
*
* U (input) SuperMatrix*
* The factor U from the factorization Pr*A*Pc=L*U as computed by
* zgstrf(). Use column-wise storage scheme, i.e., U has types:
* Stype = SLU_NC, Dtype = SLU_Z, Mtype = SLU_TRU.
*
* perm_c (input) int*, dimension (L->ncol)
* Column permutation vector, which defines the
* permutation matrix Pc; perm_c[i] = j means column i of A is
* in position j in A*Pc.
*
* perm_r (input) int*, dimension (L->nrow)
* Row permutation vector, which defines the permutation matrix Pr;
* perm_r[i] = j means row i of A is in position j in Pr*A.
*
* B (input/output) SuperMatrix*
* B has types: Stype = SLU_DN, Dtype = SLU_Z, Mtype = SLU_GE.
* On entry, the right hand side matrix.
* On exit, the solution matrix if info = 0;
*
* stat (output) SuperLUStat_t*
* Record the statistics on runtime and floating-point operation count.
* See util.h for the definition of 'SuperLUStat_t'.
*
* info (output) int*
* = 0: successful exit
* < 0: if info = -i, the i-th argument had an illegal value
*
*/
#ifdef _CRAY
_fcd ftcs1, ftcs2, ftcs3, ftcs4;
#endif
int incx = 1, incy = 1;
#ifdef USE_VENDOR_BLAS
doublecomplex alpha = {1.0, 0.0}, beta = {1.0, 0.0};
doublecomplex *work_col;
#endif
doublecomplex temp_comp;
DNformat *Bstore;
doublecomplex *Bmat;
SCformat *Lstore;
NCformat *Ustore;
doublecomplex *Lval, *Uval;
int fsupc, nrow, nsupr, nsupc, luptr, istart, irow;
int i, j, k, iptr, jcol, n, ldb, nrhs;
doublecomplex *work, *rhs_work, *soln;
flops_t solve_ops;
void zprint_soln();
/* Test input parameters ... */
*info = 0;
Bstore = B->Store;
ldb = Bstore->lda;
nrhs = B->ncol;
if ( trans != NOTRANS && trans != TRANS && trans != CONJ ) *info = -1;
else if ( L->nrow != L->ncol || L->nrow < 0 ||
L->Stype != SLU_SC || L->Dtype != SLU_Z || L->Mtype != SLU_TRLU )
*info = -2;
else if ( U->nrow != U->ncol || U->nrow < 0 ||
U->Stype != SLU_NC || U->Dtype != SLU_Z || U->Mtype != SLU_TRU )
*info = -3;
else if ( ldb < SUPERLU_MAX(0, L->nrow) ||
B->Stype != SLU_DN || B->Dtype != SLU_Z || B->Mtype != SLU_GE )
*info = -6;
if ( *info ) {
i = -(*info);
xerbla_("zgstrs", &i);
return;
}
n = L->nrow;
work = doublecomplexCalloc(n * nrhs);
if ( !work ) ABORT("Malloc fails for local work[].");
soln = doublecomplexMalloc(n);
if ( !soln ) ABORT("Malloc fails for local soln[].");
Bmat = Bstore->nzval;
Lstore = L->Store;
Lval = Lstore->nzval;
Ustore = U->Store;
Uval = Ustore->nzval;
solve_ops = 0;
if ( trans == NOTRANS ) {
/* Permute right hand sides to form Pr*B */
for (i = 0; i < nrhs; i++) {
rhs_work = &Bmat[i*ldb];
for (k = 0; k < n; k++) soln[perm_r[k]] = rhs_work[k];
for (k = 0; k < n; k++) rhs_work[k] = soln[k];
}
/* Forward solve PLy=Pb. */
for (k = 0; k <= Lstore->nsuper; k++) {
fsupc = L_FST_SUPC(k);
istart = L_SUB_START(fsupc);
nsupr = L_SUB_START(fsupc+1) - istart;
nsupc = L_FST_SUPC(k+1) - fsupc;
nrow = nsupr - nsupc;
solve_ops += 4 * nsupc * (nsupc - 1) * nrhs;
solve_ops += 8 * nrow * nsupc * nrhs;
if ( nsupc == 1 ) {
for (j = 0; j < nrhs; j++) {
rhs_work = &Bmat[j*ldb];
luptr = L_NZ_START(fsupc);
for (iptr=istart+1; iptr < L_SUB_START(fsupc+1); iptr++){
irow = L_SUB(iptr);
++luptr;
zz_mult(&temp_comp, &rhs_work[fsupc], &Lval[luptr]);
z_sub(&rhs_work[irow], &rhs_work[irow], &temp_comp);
}
}
} else {
luptr = L_NZ_START(fsupc);
#ifdef USE_VENDOR_BLAS
#ifdef _CRAY
ftcs1 = _cptofcd("L", strlen("L"));
ftcs2 = _cptofcd("N", strlen("N"));
ftcs3 = _cptofcd("U", strlen("U"));
CTRSM( ftcs1, ftcs1, ftcs2, ftcs3, &nsupc, &nrhs, &alpha,
&Lval[luptr], &nsupr, &Bmat[fsupc], &ldb);
CGEMM( ftcs2, ftcs2, &nrow, &nrhs, &nsupc, &alpha,
&Lval[luptr+nsupc], &nsupr, &Bmat[fsupc], &ldb,
&beta, &work[0], &n );
#else
ztrsm_("L", "L", "N", "U", &nsupc, &nrhs, &alpha,
&Lval[luptr], &nsupr, &Bmat[fsupc], &ldb);
zgemm_( "N", "N", &nrow, &nrhs, &nsupc, &alpha,
&Lval[luptr+nsupc], &nsupr, &Bmat[fsupc], &ldb,
&beta, &work[0], &n );
#endif
for (j = 0; j < nrhs; j++) {
rhs_work = &Bmat[j*ldb];
work_col = &work[j*n];
iptr = istart + nsupc;
for (i = 0; i < nrow; i++) {
irow = L_SUB(iptr);
z_sub(&rhs_work[irow], &rhs_work[irow], &work_col[i]);
work_col[i].r = 0.0;
work_col[i].i = 0.0;
iptr++;
}
}
#else
for (j = 0; j < nrhs; j++) {
rhs_work = &Bmat[j*ldb];
zlsolve (nsupr, nsupc, &Lval[luptr], &rhs_work[fsupc]);
zmatvec (nsupr, nrow, nsupc, &Lval[luptr+nsupc],
&rhs_work[fsupc], &work[0] );
iptr = istart + nsupc;
for (i = 0; i < nrow; i++) {
irow = L_SUB(iptr);
z_sub(&rhs_work[irow], &rhs_work[irow], &work[i]);
work[i].r = 0.;
work[i].i = 0.;
iptr++;
}
}
#endif
} /* else ... */
} /* for L-solve */
#ifdef DEBUG
printf("After L-solve: y=\n");
zprint_soln(n, nrhs, Bmat);
#endif
/*
* Back solve Ux=y.
*/
for (k = Lstore->nsuper; k >= 0; k--) {
fsupc = L_FST_SUPC(k);
istart = L_SUB_START(fsupc);
nsupr = L_SUB_START(fsupc+1) - istart;
nsupc = L_FST_SUPC(k+1) - fsupc;
luptr = L_NZ_START(fsupc);
solve_ops += 4 * nsupc * (nsupc + 1) * nrhs;
if ( nsupc == 1 ) {
rhs_work = &Bmat[0];
for (j = 0; j < nrhs; j++) {
z_div(&rhs_work[fsupc], &rhs_work[fsupc], &Lval[luptr]);
rhs_work += ldb;
}
} else {
#ifdef USE_VENDOR_BLAS
#ifdef _CRAY
ftcs1 = _cptofcd("L", strlen("L"));
ftcs2 = _cptofcd("U", strlen("U"));
ftcs3 = _cptofcd("N", strlen("N"));
CTRSM( ftcs1, ftcs2, ftcs3, ftcs3, &nsupc, &nrhs, &alpha,
&Lval[luptr], &nsupr, &Bmat[fsupc], &ldb);
#else
ztrsm_("L", "U", "N", "N", &nsupc, &nrhs, &alpha,
&Lval[luptr], &nsupr, &Bmat[fsupc], &ldb);
#endif
#else
for (j = 0; j < nrhs; j++)
zusolve ( nsupr, nsupc, &Lval[luptr], &Bmat[fsupc+j*ldb] );
#endif
}
for (j = 0; j < nrhs; ++j) {
rhs_work = &Bmat[j*ldb];
for (jcol = fsupc; jcol < fsupc + nsupc; jcol++) {
solve_ops += 8*(U_NZ_START(jcol+1) - U_NZ_START(jcol));
for (i = U_NZ_START(jcol); i < U_NZ_START(jcol+1); i++ ){
irow = U_SUB(i);
zz_mult(&temp_comp, &rhs_work[jcol], &Uval[i]);
z_sub(&rhs_work[irow], &rhs_work[irow], &temp_comp);
}
}
}
} /* for U-solve */
#ifdef DEBUG
printf("After U-solve: x=\n");
zprint_soln(n, nrhs, Bmat);
#endif
/* Compute the final solution X := Pc*X. */
for (i = 0; i < nrhs; i++) {
rhs_work = &Bmat[i*ldb];
for (k = 0; k < n; k++) soln[k] = rhs_work[perm_c[k]];
for (k = 0; k < n; k++) rhs_work[k] = soln[k];
}
stat->ops[SOLVE] = solve_ops;
} else { /* Solve A'*X=B */
/* Permute right hand sides to form Pc'*B. */
for (i = 0; i < nrhs; i++) {
rhs_work = &Bmat[i*ldb];
for (k = 0; k < n; k++) soln[perm_c[k]] = rhs_work[k];
for (k = 0; k < n; k++) rhs_work[k] = soln[k];
}
stat->ops[SOLVE] = 0;
if (trans == TRANS) {
for (k = 0; k < nrhs; ++k) {
/* Multiply by inv(U'). */
sp_ztrsv("U", "T", "N", L, U, &Bmat[k*ldb], stat, info);
/* Multiply by inv(L'). */
sp_ztrsv("L", "T", "U", L, U, &Bmat[k*ldb], stat, info);
}
}
else {
for (k = 0; k < nrhs; ++k) {
/* Multiply by inv(U'). */
sp_ztrsv("U", "C", "N", L, U, &Bmat[k*ldb], stat, info);
/* Multiply by inv(L'). */
sp_ztrsv("L", "C", "U", L, U, &Bmat[k*ldb], stat, info);
}
}
/* Compute the final solution X := Pr'*X (=inv(Pr)*X) */
for (i = 0; i < nrhs; i++) {
rhs_work = &Bmat[i*ldb];
for (k = 0; k < n; k++) soln[k] = rhs_work[perm_r[k]];
for (k = 0; k < n; k++) rhs_work[k] = soln[k];
}
}
SUPERLU_FREE(work);
SUPERLU_FREE(soln);
}
void
zgstrs (trans_t trans, SuperMatrix *L, SuperMatrix *U,
int *perm_c, int *perm_r, SuperMatrix *B,
SuperLUStat_t *stat, int *info)
{
#ifdef _CRAY
_fcd ftcs1, ftcs2, ftcs3, ftcs4;
#endif
int incx = 1, incy = 1;
#ifdef USE_VENDOR_BLAS
doublecomplex alpha = {1.0, 0.0}, beta = {1.0, 0.0};
doublecomplex *work_col;
#endif
doublecomplex temp_comp;
DNformat *Bstore;
doublecomplex *Bmat;
SCformat *Lstore;
NCformat *Ustore;
doublecomplex *Lval, *Uval;
int fsupc, nrow, nsupr, nsupc, luptr, istart, irow;
int i, j, k, iptr, jcol, n, ldb, nrhs;
doublecomplex *work, *rhs_work, *soln;
flops_t solve_ops;
void zprint_soln();
/* Test input parameters ... */
*info = 0;
Bstore = B->Store;
ldb = Bstore->lda;
nrhs = B->ncol;
if ( trans != NOTRANS && trans != TRANS && trans != CONJ ) *info = -1;
else if ( L->nrow != L->ncol || L->nrow < 0 ||
L->Stype != SLU_SC || L->Dtype != SLU_Z || L->Mtype != SLU_TRLU )
*info = -2;
else if ( U->nrow != U->ncol || U->nrow < 0 ||
U->Stype != SLU_NC || U->Dtype != SLU_Z || U->Mtype != SLU_TRU )
*info = -3;
else if ( ldb < SUPERLU_MAX(0, L->nrow) ||
B->Stype != SLU_DN || B->Dtype != SLU_Z || B->Mtype != SLU_GE )
*info = -6;
if ( *info ) {
i = -(*info);
input_error("zgstrs", &i);
return;
}
n = L->nrow;
work = doublecomplexCalloc(n * nrhs);
if ( !work ) ABORT("Malloc fails for local work[].");
soln = doublecomplexMalloc(n);
if ( !soln ) ABORT("Malloc fails for local soln[].");
Bmat = Bstore->nzval;
Lstore = L->Store;
Lval = Lstore->nzval;
Ustore = U->Store;
Uval = Ustore->nzval;
solve_ops = 0;
if ( trans == NOTRANS ) {
/* Permute right hand sides to form Pr*B */
for (i = 0; i < nrhs; i++) {
rhs_work = &Bmat[i*ldb];
for (k = 0; k < n; k++) soln[perm_r[k]] = rhs_work[k];
for (k = 0; k < n; k++) rhs_work[k] = soln[k];
}
/* Forward solve PLy=Pb. */
for (k = 0; k <= Lstore->nsuper; k++) {
fsupc = L_FST_SUPC(k);
istart = L_SUB_START(fsupc);
nsupr = L_SUB_START(fsupc+1) - istart;
nsupc = L_FST_SUPC(k+1) - fsupc;
nrow = nsupr - nsupc;
solve_ops += 4 * nsupc * (nsupc - 1) * nrhs;
solve_ops += 8 * nrow * nsupc * nrhs;
if ( nsupc == 1 ) {
for (j = 0; j < nrhs; j++) {
rhs_work = &Bmat[j*ldb];
luptr = L_NZ_START(fsupc);
for (iptr=istart+1; iptr < L_SUB_START(fsupc+1); iptr++){
irow = L_SUB(iptr);
++luptr;
zz_mult(&temp_comp, &rhs_work[fsupc], &Lval[luptr]);
z_sub(&rhs_work[irow], &rhs_work[irow], &temp_comp);
}
}
} else {
luptr = L_NZ_START(fsupc);
#ifdef USE_VENDOR_BLAS
#ifdef _CRAY
ftcs1 = _cptofcd("L", strlen("L"));
ftcs2 = _cptofcd("N", strlen("N"));
ftcs3 = _cptofcd("U", strlen("U"));
CTRSM( ftcs1, ftcs1, ftcs2, ftcs3, &nsupc, &nrhs, &alpha,
&Lval[luptr], &nsupr, &Bmat[fsupc], &ldb);
CGEMM( ftcs2, ftcs2, &nrow, &nrhs, &nsupc, &alpha,
&Lval[luptr+nsupc], &nsupr, &Bmat[fsupc], &ldb,
&beta, &work[0], &n );
#else
ztrsm_("L", "L", "N", "U", &nsupc, &nrhs, &alpha,
&Lval[luptr], &nsupr, &Bmat[fsupc], &ldb);
zgemm_( "N", "N", &nrow, &nrhs, &nsupc, &alpha,
&Lval[luptr+nsupc], &nsupr, &Bmat[fsupc], &ldb,
&beta, &work[0], &n );
#endif
for (j = 0; j < nrhs; j++) {
rhs_work = &Bmat[j*ldb];
work_col = &work[j*n];
iptr = istart + nsupc;
for (i = 0; i < nrow; i++) {
irow = L_SUB(iptr);
z_sub(&rhs_work[irow], &rhs_work[irow], &work_col[i]);
work_col[i].r = 0.0;
work_col[i].i = 0.0;
iptr++;
}
}
#else
for (j = 0; j < nrhs; j++) {
rhs_work = &Bmat[j*ldb];
zlsolve (nsupr, nsupc, &Lval[luptr], &rhs_work[fsupc]);
zmatvec (nsupr, nrow, nsupc, &Lval[luptr+nsupc],
&rhs_work[fsupc], &work[0] );
iptr = istart + nsupc;
for (i = 0; i < nrow; i++) {
irow = L_SUB(iptr);
z_sub(&rhs_work[irow], &rhs_work[irow], &work[i]);
work[i].r = 0.;
work[i].i = 0.;
iptr++;
}
}
#endif
} /* else ... */
} /* for L-solve */
#ifdef DEBUG
printf("After L-solve: y=\n");
zprint_soln(n, nrhs, Bmat);
#endif
/*
* Back solve Ux=y.
*/
for (k = Lstore->nsuper; k >= 0; k--) {
fsupc = L_FST_SUPC(k);
istart = L_SUB_START(fsupc);
nsupr = L_SUB_START(fsupc+1) - istart;
nsupc = L_FST_SUPC(k+1) - fsupc;
luptr = L_NZ_START(fsupc);
solve_ops += 4 * nsupc * (nsupc + 1) * nrhs;
if ( nsupc == 1 ) {
rhs_work = &Bmat[0];
for (j = 0; j < nrhs; j++) {
z_div(&rhs_work[fsupc], &rhs_work[fsupc], &Lval[luptr]);
rhs_work += ldb;
}
} else {
#ifdef USE_VENDOR_BLAS
#ifdef _CRAY
ftcs1 = _cptofcd("L", strlen("L"));
ftcs2 = _cptofcd("U", strlen("U"));
ftcs3 = _cptofcd("N", strlen("N"));
CTRSM( ftcs1, ftcs2, ftcs3, ftcs3, &nsupc, &nrhs, &alpha,
&Lval[luptr], &nsupr, &Bmat[fsupc], &ldb);
#else
ztrsm_("L", "U", "N", "N", &nsupc, &nrhs, &alpha,
&Lval[luptr], &nsupr, &Bmat[fsupc], &ldb);
#endif
#else
for (j = 0; j < nrhs; j++)
zusolve ( nsupr, nsupc, &Lval[luptr], &Bmat[fsupc+j*ldb] );
#endif
}
for (j = 0; j < nrhs; ++j) {
rhs_work = &Bmat[j*ldb];
for (jcol = fsupc; jcol < fsupc + nsupc; jcol++) {
solve_ops += 8*(U_NZ_START(jcol+1) - U_NZ_START(jcol));
for (i = U_NZ_START(jcol); i < U_NZ_START(jcol+1); i++ ){
irow = U_SUB(i);
zz_mult(&temp_comp, &rhs_work[jcol], &Uval[i]);
z_sub(&rhs_work[irow], &rhs_work[irow], &temp_comp);
}
}
}
} /* for U-solve */
#ifdef DEBUG
printf("After U-solve: x=\n");
zprint_soln(n, nrhs, Bmat);
#endif
/* Compute the final solution X := Pc*X. */
for (i = 0; i < nrhs; i++) {
rhs_work = &Bmat[i*ldb];
for (k = 0; k < n; k++) soln[k] = rhs_work[perm_c[k]];
for (k = 0; k < n; k++) rhs_work[k] = soln[k];
}
stat->ops[SOLVE] = solve_ops;
} else { /* Solve A'*X=B or CONJ(A)*X=B */
/* Permute right hand sides to form Pc'*B. */
for (i = 0; i < nrhs; i++) {
rhs_work = &Bmat[i*ldb];
for (k = 0; k < n; k++) soln[perm_c[k]] = rhs_work[k];
for (k = 0; k < n; k++) rhs_work[k] = soln[k];
}
stat->ops[SOLVE] = 0;
if (trans == TRANS) {
for (k = 0; k < nrhs; ++k) {
/* Multiply by inv(U'). */
sp_ztrsv("U", "T", "N", L, U, &Bmat[k*ldb], stat, info);
/* Multiply by inv(L'). */
sp_ztrsv("L", "T", "U", L, U, &Bmat[k*ldb], stat, info);
}
} else { /* trans == CONJ */
for (k = 0; k < nrhs; ++k) {
/* Multiply by conj(inv(U')). */
sp_ztrsv("U", "C", "N", L, U, &Bmat[k*ldb], stat, info);
/* Multiply by conj(inv(L')). */
sp_ztrsv("L", "C", "U", L, U, &Bmat[k*ldb], stat, info);
}
}
/* Compute the final solution X := Pr'*X (=inv(Pr)*X) */
for (i = 0; i < nrhs; i++) {
rhs_work = &Bmat[i*ldb];
for (k = 0; k < n; k++) soln[k] = rhs_work[perm_r[k]];
for (k = 0; k < n; k++) rhs_work[k] = soln[k];
}
}
SUPERLU_FREE(work);
SUPERLU_FREE(soln);
}
void pzgstrf
/************************************************************************/
(
superlu_options_t_Distributed *options, int m, int n, double anorm,
LUstruct_t *LUstruct, gridinfo_t *grid, SuperLUStat_t *stat, int *info
)
/*
* Purpose
* =======
*
* pzgstrf performs the LU factorization in parallel.
*
* Arguments
* =========
*
* options (input) superlu_options_t_Distributed*
* The structure defines the input parameters to control
* how the LU decomposition will be performed.
* The following field should be defined:
* o ReplaceTinyPivot (yes_no_t)
* Specifies whether to replace the tiny diagonals by
* sqrt(epsilon)*norm(A) during LU factorization.
*
* m (input) int
* Number of rows in the matrix.
*
* n (input) int
* Number of columns in the matrix.
*
* anorm (input) double
* The norm of the original matrix A, or the scaled A if
* equilibration was done.
*
* LUstruct (input/output) LUstruct_t*
* The data structures to store the distributed L and U factors.
* The following fields should be defined:
*
* o Glu_persist (input) Glu_persist_t*
* Global data structure (xsup, supno) replicated on all processes,
* describing the supernode partition in the factored matrices
* L and U:
* xsup[s] is the leading column of the s-th supernode,
* supno[i] is the supernode number to which column i belongs.
*
* o Llu (input/output) LocalLU_t*
* The distributed data structures to store L and U factors.
* See superlu_zdefs.h for the definition of 'LocalLU_t'.
*
* grid (input) gridinfo_t*
* The 2D process mesh. It contains the MPI communicator, the number
* of process rows (NPROW), the number of process columns (NPCOL),
* and my process rank. It is an input argument to all the
* parallel routines.
* Grid can be initialized by subroutine SUPERLU_GRIDINIT.
* See superlu_zdefs.h for the definition of 'gridinfo_t'.
*
* stat (output) SuperLUStat_t*
* Record the statistics on runtime and floating-point operation count.
* See util.h for the definition of 'SuperLUStat_t'.
*
* info (output) int*
* = 0: successful exit
* < 0: if info = -i, the i-th argument had an illegal value
* > 0: if info = i, U(i,i) is exactly zero. The factorization has
* been completed, but the factor U is exactly singular,
* and division by zero will occur if it is used to solve a
* system of equations.
*
*/
{
#ifdef _CRAY
_fcd ftcs = _cptofcd("N", strlen("N"));
_fcd ftcs1 = _cptofcd("L", strlen("L"));
_fcd ftcs2 = _cptofcd("N", strlen("N"));
_fcd ftcs3 = _cptofcd("U", strlen("U"));
#endif
doublecomplex zero = {0.0, 0.0};
doublecomplex alpha = {1.0, 0.0}, beta = {0.0, 0.0};
int_t *xsup;
int_t *lsub, *lsub1, *usub, *Usub_buf,
*Lsub_buf_2[2]; /* Need 2 buffers to implement Irecv. */
doublecomplex *lusup, *lusup1, *uval, *Uval_buf,
*Lval_buf_2[2]; /* Need 2 buffers to implement Irecv. */
int_t fnz, i, ib, ijb, ilst, it, iukp, jb, jj, klst, knsupc,
lb, lib, ldv, ljb, lptr, lptr0, lptrj, luptr, luptr0, luptrj,
nlb, nub, nsupc, rel, rukp;
int_t Pc, Pr;
int iam, kcol, krow, mycol, myrow, pi, pj;
int j, k, lk, nsupers;
int nsupr, nbrow, segsize;
int msgcnt[4]; /* Count the size of the message xfer'd in each buffer:
* 0 : transferred in Lsub_buf[]
* 1 : transferred in Lval_buf[]
* 2 : transferred in Usub_buf[]
* 3 : transferred in Uval_buf[]
*/
int_t msg0, msg2;
int_t **Ufstnz_br_ptr, **Lrowind_bc_ptr;
doublecomplex **Unzval_br_ptr, **Lnzval_bc_ptr;
int_t *index;
doublecomplex *nzval;
int_t *iuip, *ruip;/* Pointers to U index/nzval; size ceil(NSUPERS/Pr). */
doublecomplex *ucol;
int_t *indirect;
doublecomplex *tempv, *tempv2d;
int_t iinfo;
int_t *ToRecv, *ToSendD, **ToSendR;
Glu_persist_t *Glu_persist = LUstruct->Glu_persist;
LocalLU_t *Llu = LUstruct->Llu;
superlu_scope_t *scp;
double s_eps, thresh;
doublecomplex *tempU2d, *tempu;
int full, ldt, ldu, lead_zero, ncols;
MPI_Request recv_req[4], *send_req;
MPI_Status status;
#if ( DEBUGlevel>=1 )
int_t num_copy=0, num_update=0;
#endif
#if ( PRNTlevel==3 )
int_t zero_msg = 0, total_msg = 0;
#endif
#if ( PROFlevel>=1 )
double t1, t2;
float msg_vol = 0, msg_cnt = 0;
int_t iword = sizeof(int_t), zword = sizeof(doublecomplex);
#endif
/* Test the input parameters. */
*info = 0;
if ( m < 0 ) *info = -2;
else if ( n < 0 ) *info = -3;
if ( *info ) {
pxerbla("pzgstrf", grid, -*info);
return;
}
/* Quick return if possible. */
if ( m == 0 || n == 0 ) return;
/*
* Initialization.
*/
iam = grid->iam;
Pc = grid->npcol;
Pr = grid->nprow;
myrow = MYROW( iam, grid );
mycol = MYCOL( iam, grid );
nsupers = Glu_persist->supno[n-1] + 1;
xsup = Glu_persist->xsup;
s_eps = slamch_("Epsilon");
thresh = s_eps * anorm / 256.;
#if ( DEBUGlevel>=1 )
CHECK_MALLOC(iam, "Enter pzgstrf()");
#endif
stat->ops[FACT] = 0.0;
if ( Pr*Pc > 1) {
i = Llu->bufmax[0];
if ( !(Llu->Lsub_buf_2[0] = intMalloc_dist(2 * i)) )
ABORT("Malloc fails for Lsub_buf.");
Llu->Lsub_buf_2[1] = Llu->Lsub_buf_2[0] + i;
i = Llu->bufmax[1];
if ( !(Llu->Lval_buf_2[0] = doublecomplexMalloc_dist(2 * i)) )
ABORT("Malloc fails for Lval_buf[].");
Llu->Lval_buf_2[1] = Llu->Lval_buf_2[0] + i;
if ( Llu->bufmax[2] != 0 )
if ( !(Llu->Usub_buf = intMalloc_dist(Llu->bufmax[2])) )
ABORT("Malloc fails for Usub_buf[].");
if ( Llu->bufmax[3] != 0 )
if ( !(Llu->Uval_buf = doublecomplexMalloc_dist(Llu->bufmax[3])) )
ABORT("Malloc fails for Uval_buf[].");
if ( !(send_req =
(MPI_Request *) SUPERLU_MALLOC(2*Pc*sizeof(MPI_Request))))
ABORT("Malloc fails for send_req[].");
}
if ( !(Llu->ujrow = doublecomplexMalloc_dist(sp_ienv_dist(3))) )
ABORT("Malloc fails for ujrow[].");
#if ( PRNTlevel>=1 )
if ( !iam ) {
printf(".. thresh = s_eps %e * anorm %e / 256. = %e\n", s_eps, anorm, thresh);
printf(".. Buffer size: Lsub %d\tLval %d\tUsub %d\tUval %d\tLDA %d\n",
Llu->bufmax[0], Llu->bufmax[1],
Llu->bufmax[2], Llu->bufmax[3], Llu->bufmax[4]);
}
#endif
Lsub_buf_2[0] = Llu->Lsub_buf_2[0];
Lsub_buf_2[1] = Llu->Lsub_buf_2[1];
Lval_buf_2[0] = Llu->Lval_buf_2[0];
Lval_buf_2[1] = Llu->Lval_buf_2[1];
Usub_buf = Llu->Usub_buf;
Uval_buf = Llu->Uval_buf;
Lrowind_bc_ptr = Llu->Lrowind_bc_ptr;
Lnzval_bc_ptr = Llu->Lnzval_bc_ptr;
Ufstnz_br_ptr = Llu->Ufstnz_br_ptr;
Unzval_br_ptr = Llu->Unzval_br_ptr;
ToRecv = Llu->ToRecv;
ToSendD = Llu->ToSendD;
ToSendR = Llu->ToSendR;
ldt = sp_ienv_dist(3); /* Size of maximum supernode */
if ( !(tempv2d = doublecomplexCalloc_dist(2*ldt*ldt)) )
ABORT("Calloc fails for tempv2d[].");
tempU2d = tempv2d + ldt*ldt;
if ( !(indirect = intMalloc_dist(ldt)) )
ABORT("Malloc fails for indirect[].");
k = CEILING( nsupers, Pr ); /* Number of local block rows */
if ( !(iuip = intMalloc_dist(k)) )
ABORT("Malloc fails for iuip[].");
if ( !(ruip = intMalloc_dist(k)) )
ABORT("Malloc fails for ruip[].");
/* ---------------------------------------------------------------
Handle the first block column separately to start the pipeline.
--------------------------------------------------------------- */
if ( mycol == 0 ) {
pzgstrf2(options, 0, thresh, Glu_persist, grid, Llu, stat, info);
scp = &grid->rscp; /* The scope of process row. */
/* Process column *kcol* multicasts numeric values of L(:,k)
to process rows. */
lsub = Lrowind_bc_ptr[0];
lusup = Lnzval_bc_ptr[0];
if ( lsub ) {
msgcnt[0] = lsub[1] + BC_HEADER + lsub[0]*LB_DESCRIPTOR;
msgcnt[1] = lsub[1] * SuperSize( 0 );
} else {
msgcnt[0] = msgcnt[1] = 0;
}
for (pj = 0; pj < Pc; ++pj) {
if ( ToSendR[0][pj] != EMPTY ) {
#if ( PROFlevel>=1 )
TIC(t1);
#endif
MPI_Isend( lsub, msgcnt[0], mpi_int_t, pj, 0, scp->comm,
&send_req[pj] );
MPI_Isend( lusup, msgcnt[1], SuperLU_MPI_DOUBLE_COMPLEX,
pj, 1, scp->comm, &send_req[pj+Pc] );
#if ( DEBUGlevel>=2 )
printf("(%d) Send L(:,%4d): lsub %4d, lusup %4d to Pc %2d\n",
iam, 0, msgcnt[0], msgcnt[1], pj);
#endif
#if ( PROFlevel>=1 )
TOC(t2, t1);
stat->utime[COMM] += t2;
msg_cnt += 2;
msg_vol += msgcnt[0]*iword + msgcnt[1]*zword;
#endif
}
} /* for pj ... */
} else { /* Post immediate receives. */
if ( ToRecv[0] >= 1 ) { /* Recv block column L(:,0). */
scp = &grid->rscp; /* The scope of process row. */
MPI_Irecv( Lsub_buf_2[0], Llu->bufmax[0], mpi_int_t, 0,
0, scp->comm, &recv_req[0] );
MPI_Irecv( Lval_buf_2[0], Llu->bufmax[1],
SuperLU_MPI_DOUBLE_COMPLEX, 0,
1, scp->comm, &recv_req[1] );
#if ( DEBUGlevel>=2 )
printf("(%d) Post Irecv L(:,%4d)\n", iam, 0);
#endif
}
} /* if mycol == 0 */
/* ------------------------------------------
MAIN LOOP: Loop through all block columns.
------------------------------------------ */
for (k = 0; k < nsupers; ++k) {
knsupc = SuperSize( k );
krow = PROW( k, grid );
kcol = PCOL( k, grid );
if ( mycol == kcol ) {
lk = LBj( k, grid ); /* Local block number. */
for (pj = 0; pj < Pc; ++pj) {
/* Wait for Isend to complete before using lsub/lusup. */
if ( ToSendR[lk][pj] != EMPTY ) {
MPI_Wait( &send_req[pj], &status );
MPI_Wait( &send_req[pj+Pc], &status );
}
}
lsub = Lrowind_bc_ptr[lk];
lusup = Lnzval_bc_ptr[lk];
} else {
if ( ToRecv[k] >= 1 ) { /* Recv block column L(:,k). */
scp = &grid->rscp; /* The scope of process row. */
#if ( PROFlevel>=1 )
TIC(t1);
#endif
/*probe_recv(iam, kcol, (4*k)%NTAGS, mpi_int_t, scp->comm,
Llu->bufmax[0]);*/
/*MPI_Recv( Lsub_buf, Llu->bufmax[0], mpi_int_t, kcol,
(4*k)%NTAGS, scp->comm, &status );*/
MPI_Wait( &recv_req[0], &status );
MPI_Get_count( &status, mpi_int_t, &msgcnt[0] );
/*probe_recv(iam, kcol, (4*k+1)%NTAGS, MPI_DOUBLE, scp->comm,
Llu->bufmax[1]);*/
/*MPI_Recv( Lval_buf, Llu->bufmax[1], SuperLU_MPI_DOUBLE_COMPLEX,
kcol, (4*k+1)%NTAGS, scp->comm, &status );*/
MPI_Wait( &recv_req[1], &status );
MPI_Get_count(&status, SuperLU_MPI_DOUBLE_COMPLEX, &msgcnt[1]);
#if ( PROFlevel>=1 )
TOC(t2, t1);
stat->utime[COMM] += t2;
#endif
#if ( DEBUGlevel>=2 )
printf("(%d) Recv L(:,%4d): lsub %4d, lusup %4d from Pc %2d\n",
iam, k, msgcnt[0], msgcnt[1], kcol);
#endif
lsub = Lsub_buf_2[k%2];
lusup = Lval_buf_2[k%2];
#if ( PRNTlevel==3 )
++total_msg;
if ( !msgcnt[0] ) ++zero_msg;
#endif
} else msgcnt[0] = 0;
} /* if mycol = Pc(k) */
scp = &grid->cscp; /* The scope of process column. */
if ( myrow == krow ) {
/* Parallel triangular solve across process row *krow* --
U(k,j) = L(k,k) \ A(k,j). */
#ifdef _CRAY
pzgstrs2(n, k, Glu_persist, grid, Llu, stat, ftcs1, ftcs2, ftcs3);
#else
pzgstrs2(n, k, Glu_persist, grid, Llu, stat);
#endif
/* Multicasts U(k,:) to process columns. */
lk = LBi( k, grid );
usub = Ufstnz_br_ptr[lk];
uval = Unzval_br_ptr[lk];
if ( usub ) {
msgcnt[2] = usub[2];
msgcnt[3] = usub[1];
} else {
msgcnt[2] = msgcnt[3] = 0;
}
if ( ToSendD[lk] == YES ) {
for (pi = 0; pi < Pr; ++pi)
if ( pi != myrow ) {
#if ( PROFlevel>=1 )
TIC(t1);
#endif
MPI_Send( usub, msgcnt[2], mpi_int_t, pi,
(4*k+2)%NTAGS, scp->comm);
MPI_Send( uval, msgcnt[3], SuperLU_MPI_DOUBLE_COMPLEX,
pi, (4*k+3)%NTAGS, scp->comm);
#if ( PROFlevel>=1 )
TOC(t2, t1);
stat->utime[COMM] += t2;
msg_cnt += 2;
msg_vol += msgcnt[2]*iword + msgcnt[3]*zword;
#endif
#if ( DEBUGlevel>=2 )
printf("(%d) Send U(%4d,:) to Pr %2d\n", iam, k, pi);
#endif
} /* if pi ... */
} /* if ToSendD ... */
} else { /* myrow != krow */
if ( ToRecv[k] == 2 ) { /* Recv block row U(k,:). */
#if ( PROFlevel>=1 )
TIC(t1);
#endif
/*probe_recv(iam, krow, (4*k+2)%NTAGS, mpi_int_t, scp->comm,
Llu->bufmax[2]);*/
MPI_Recv( Usub_buf, Llu->bufmax[2], mpi_int_t, krow,
(4*k+2)%NTAGS, scp->comm, &status );
MPI_Get_count( &status, mpi_int_t, &msgcnt[2] );
/*probe_recv(iam, krow, (4*k+3)%NTAGS,
SuperLU_MPI_DOUBLE_COMPLEX, scp->comm,
Llu->bufmax[3]);*/
MPI_Recv( Uval_buf, Llu->bufmax[3], SuperLU_MPI_DOUBLE_COMPLEX,
krow, (4*k+3)%NTAGS, scp->comm, &status );
MPI_Get_count(&status, SuperLU_MPI_DOUBLE_COMPLEX, &msgcnt[3]);
#if ( PROFlevel>=1 )
TOC(t2, t1);
stat->utime[COMM] += t2;
#endif
usub = Usub_buf;
uval = Uval_buf;
#if ( DEBUGlevel>=2 )
printf("(%d) Recv U(%4d,:) from Pr %2d\n", iam, k, krow);
#endif
#if ( PRNTlevel==3 )
++total_msg;
if ( !msgcnt[2] ) ++zero_msg;
#endif
} else msgcnt[2] = 0;
} /* if myrow == Pr(k) */
/*
* Parallel rank-k update; pair up blocks L(i,k) and U(k,j).
* for (j = k+1; k < N; ++k) {
* for (i = k+1; i < N; ++i)
* if ( myrow == PROW( i, grid ) && mycol == PCOL( j, grid )
* && L(i,k) != 0 && U(k,j) != 0 )
* A(i,j) = A(i,j) - L(i,k) * U(k,j);
*/
msg0 = msgcnt[0];
msg2 = msgcnt[2];
if ( msg0 && msg2 ) { /* L(:,k) and U(k,:) are not empty. */
nsupr = lsub[1]; /* LDA of lusup. */
if ( myrow == krow ) { /* Skip diagonal block L(k,k). */
lptr0 = BC_HEADER + LB_DESCRIPTOR + lsub[BC_HEADER+1];
luptr0 = knsupc;
nlb = lsub[0] - 1;
} else {
lptr0 = BC_HEADER;
luptr0 = 0;
nlb = lsub[0];
}
lptr = lptr0;
for (lb = 0; lb < nlb; ++lb) { /* Initialize block row pointers. */
ib = lsub[lptr];
lib = LBi( ib, grid );
iuip[lib] = BR_HEADER;
ruip[lib] = 0;
lptr += LB_DESCRIPTOR + lsub[lptr+1];
}
nub = usub[0]; /* Number of blocks in the block row U(k,:) */
iukp = BR_HEADER; /* Skip header; Pointer to index[] of U(k,:) */
rukp = 0; /* Pointer to nzval[] of U(k,:) */
klst = FstBlockC( k+1 );
/*
* Update the first block column A(:,k+1).
*/
jb = usub[iukp]; /* Global block number of block U(k,j). */
if ( jb == k+1 ) { /* First update (k+1)-th block. */
--nub;
lptr = lptr0;
luptr = luptr0;
ljb = LBj( jb, grid ); /* Local block number of U(k,j). */
nsupc = SuperSize( jb );
iukp += UB_DESCRIPTOR; /* Start fstnz of block U(k,j). */
/* Prepare to call DGEMM. */
jj = iukp;
while ( usub[jj] == klst ) ++jj;
ldu = klst - usub[jj++];
ncols = 1;
full = 1;
for (; jj < iukp+nsupc; ++jj) {
segsize = klst - usub[jj];
if ( segsize ) {
++ncols;
if ( segsize != ldu ) full = 0;
if ( segsize > ldu ) ldu = segsize;
}
}
#if ( DEBUGlevel>=1 )
++num_update;
#endif
if ( full ) {
tempu = &uval[rukp];
} else { /* Copy block U(k,j) into tempU2d. */
#if ( DEBUGlevel>=1 )
printf("(%d) full=%d,k=%d,jb=%d,ldu=%d,ncols=%d,nsupc=%d\n",
iam, full, k, jb, ldu, ncols, nsupc);
++num_copy;
#endif
tempu = tempU2d;
for (jj = iukp; jj < iukp+nsupc; ++jj) {
segsize = klst - usub[jj];
if ( segsize ) {
lead_zero = ldu - segsize;
for (i = 0; i < lead_zero; ++i) tempu[i] = zero;
tempu += lead_zero;
for (i = 0; i < segsize; ++i)
tempu[i] = uval[rukp+i];
rukp += segsize;
tempu += segsize;
}
}
tempu = tempU2d;
rukp -= usub[iukp - 1]; /* Return to start of U(k,j). */
} /* if full ... */
for (lb = 0; lb < nlb; ++lb) {
ib = lsub[lptr]; /* Row block L(i,k). */
nbrow = lsub[lptr+1]; /* Number of full rows. */
lptr += LB_DESCRIPTOR; /* Skip descriptor. */
tempv = tempv2d;
#ifdef _CRAY
CGEMM(ftcs, ftcs, &nbrow, &ncols, &ldu, &alpha,
&lusup[luptr+(knsupc-ldu)*nsupr], &nsupr,
tempu, &ldu, &beta, tempv, &ldt);
#else
zgemm_("N", "N", &nbrow, &ncols, &ldu, &alpha,
&lusup[luptr+(knsupc-ldu)*nsupr], &nsupr,
tempu, &ldu, &beta, tempv, &ldt);
#endif
stat->ops[FACT] += 8 * nbrow * ldu * ncols;
/* Now gather the result into the destination block. */
if ( ib < jb ) { /* A(i,j) is in U. */
ilst = FstBlockC( ib+1 );
lib = LBi( ib, grid );
index = Ufstnz_br_ptr[lib];
ijb = index[iuip[lib]];
while ( ijb < jb ) { /* Search for dest block. */
ruip[lib] += index[iuip[lib]+1];
iuip[lib] += UB_DESCRIPTOR + SuperSize( ijb );
ijb = index[iuip[lib]];
}
iuip[lib] += UB_DESCRIPTOR; /* Skip descriptor. */
tempv = tempv2d;
for (jj = 0; jj < nsupc; ++jj) {
segsize = klst - usub[iukp + jj];
fnz = index[iuip[lib]++];
if ( segsize ) { /* Nonzero segment in U(k.j). */
ucol = &Unzval_br_ptr[lib][ruip[lib]];
for (i = 0, it = 0; i < nbrow; ++i) {
rel = lsub[lptr + i] - fnz;
z_sub(&ucol[rel], &ucol[rel], &tempv[it]);
++it;
}
tempv += ldt;
}
ruip[lib] += ilst - fnz;
}
} else { /* A(i,j) is in L. */
index = Lrowind_bc_ptr[ljb];
ldv = index[1]; /* LDA of the dest lusup. */
lptrj = BC_HEADER;
luptrj = 0;
ijb = index[lptrj];
while ( ijb != ib ) { /* Search for dest block --
blocks are not ordered! */
luptrj += index[lptrj+1];
lptrj += LB_DESCRIPTOR + index[lptrj+1];
ijb = index[lptrj];
}
/*
* Build indirect table. This is needed because the
* indices are not sorted.
*/
fnz = FstBlockC( ib );
lptrj += LB_DESCRIPTOR;
for (i = 0; i < index[lptrj-1]; ++i) {
rel = index[lptrj + i] - fnz;
indirect[rel] = i;
}
nzval = Lnzval_bc_ptr[ljb] + luptrj;
tempv = tempv2d;
for (jj = 0; jj < nsupc; ++jj) {
segsize = klst - usub[iukp + jj];
if ( segsize ) {
/*#pragma _CRI cache_bypass nzval,tempv*/
for (it = 0, i = 0; i < nbrow; ++i) {
rel = lsub[lptr + i] - fnz;
z_sub(&nzval[indirect[rel]],
&nzval[indirect[rel]],
&tempv[it]);
++it;
}
tempv += ldt;
}
nzval += ldv;
}
} /* if ib < jb ... */
lptr += nbrow;
luptr += nbrow;
} /* for lb ... */
rukp += usub[iukp - 1]; /* Move to block U(k,j+1) */
iukp += nsupc;
} /* if jb == k+1 */
} /* if L(:,k) and U(k,:) not empty */
if ( k+1 < nsupers ) {
kcol = PCOL( k+1, grid );
if ( mycol == kcol ) {
/* Factor diagonal and subdiagonal blocks and test for exact
singularity. */
pzgstrf2(options, k+1, thresh, Glu_persist, grid, Llu, stat, info);
/* Process column *kcol+1* multicasts numeric values of L(:,k+1)
to process rows. */
lk = LBj( k+1, grid ); /* Local block number. */
lsub1 = Lrowind_bc_ptr[lk];
if ( lsub1 ) {
msgcnt[0] = lsub1[1] + BC_HEADER + lsub1[0]*LB_DESCRIPTOR;
msgcnt[1] = lsub1[1] * SuperSize( k+1 );
} else {
msgcnt[0] = 0;
msgcnt[1] = 0;
}
scp = &grid->rscp; /* The scope of process row. */
for (pj = 0; pj < Pc; ++pj) {
if ( ToSendR[lk][pj] != EMPTY ) {
lusup1 = Lnzval_bc_ptr[lk];
#if ( PROFlevel>=1 )
TIC(t1);
#endif
MPI_Isend( lsub1, msgcnt[0], mpi_int_t, pj,
(4*(k+1))%NTAGS, scp->comm, &send_req[pj] );
MPI_Isend( lusup1, msgcnt[1], SuperLU_MPI_DOUBLE_COMPLEX,
pj, (4*(k+1)+1)%NTAGS, scp->comm,
&send_req[pj+Pc] );
#if ( PROFlevel>=1 )
TOC(t2, t1);
stat->utime[COMM] += t2;
msg_cnt += 2;
msg_vol += msgcnt[0]*iword + msgcnt[1]*zword;
#endif
#if ( DEBUGlevel>=2 )
printf("(%d) Send L(:,%4d): lsub %4d, lusup %4d to Pc %2d\n",
iam, k+1, msgcnt[0], msgcnt[1], pj);
#endif
}
} /* for pj ... */
} else { /* Post Recv of block column L(:,k+1). */
if ( ToRecv[k+1] >= 1 ) {
scp = &grid->rscp; /* The scope of process row. */
MPI_Irecv(Lsub_buf_2[(k+1)%2], Llu->bufmax[0], mpi_int_t, kcol,
(4*(k+1))%NTAGS, scp->comm, &recv_req[0]);
MPI_Irecv(Lval_buf_2[(k+1)%2], Llu->bufmax[1],
SuperLU_MPI_DOUBLE_COMPLEX, kcol,
(4*(k+1)+1)%NTAGS, scp->comm, &recv_req[1]);
}
} /* if mycol == Pc(k+1) */
} /* if k+1 < nsupers */
if ( msg0 && msg2 ) { /* L(:,k) and U(k,:) are not empty. */
/*
* Update all other blocks using block row U(k,:)
*/
for (j = 0; j < nub; ++j) {
lptr = lptr0;
luptr = luptr0;
jb = usub[iukp]; /* Global block number of block U(k,j). */
ljb = LBj( jb, grid ); /* Local block number of U(k,j). */
nsupc = SuperSize( jb );
iukp += UB_DESCRIPTOR; /* Start fstnz of block U(k,j). */
/* Prepare to call DGEMM. */
jj = iukp;
while ( usub[jj] == klst ) ++jj;
ldu = klst - usub[jj++];
ncols = 1;
full = 1;
for (; jj < iukp+nsupc; ++jj) {
segsize = klst - usub[jj];
if ( segsize ) {
++ncols;
if ( segsize != ldu ) full = 0;
if ( segsize > ldu ) ldu = segsize;
}
}
#if ( DEBUGlevel>=2 )
printf("(%d) full=%d,k=%d,jb=%d,ldu=%d,ncols=%d,nsupc=%d\n",
iam, full, k, jb, ldu, ncols, nsupc);
++num_update;
#endif
if ( full ) {
tempu = &uval[rukp];
} else { /* Copy block U(k,j) into tempU2d. */
#if ( DEBUGlevel>=1 )
++num_copy;
#endif
tempu = tempU2d;
for (jj = iukp; jj < iukp+nsupc; ++jj) {
segsize = klst - usub[jj];
if ( segsize ) {
lead_zero = ldu - segsize;
for (i = 0; i < lead_zero; ++i) tempu[i] = zero;
tempu += lead_zero;
for (i = 0; i < segsize; ++i)
tempu[i] = uval[rukp+i];
rukp += segsize;
tempu += segsize;
}
}
tempu = tempU2d;
rukp -= usub[iukp - 1]; /* Return to start of U(k,j). */
} /* if full ... */
for (lb = 0; lb < nlb; ++lb) {
ib = lsub[lptr]; /* Row block L(i,k). */
nbrow = lsub[lptr+1]; /* Number of full rows. */
lptr += LB_DESCRIPTOR; /* Skip descriptor. */
tempv = tempv2d;
#ifdef _CRAY
CGEMM(ftcs, ftcs, &nbrow, &ncols, &ldu, &alpha,
&lusup[luptr+(knsupc-ldu)*nsupr], &nsupr,
tempu, &ldu, &beta, tempv, &ldt);
#else
zgemm_("N", "N", &nbrow, &ncols, &ldu, &alpha,
&lusup[luptr+(knsupc-ldu)*nsupr], &nsupr,
tempu, &ldu, &beta, tempv, &ldt);
#endif
stat->ops[FACT] += 8 * nbrow * ldu * ncols;
/* Now gather the result into the destination block. */
if ( ib < jb ) { /* A(i,j) is in U. */
ilst = FstBlockC( ib+1 );
lib = LBi( ib, grid );
index = Ufstnz_br_ptr[lib];
ijb = index[iuip[lib]];
while ( ijb < jb ) { /* Search for dest block. */
ruip[lib] += index[iuip[lib]+1];
iuip[lib] += UB_DESCRIPTOR + SuperSize( ijb );
ijb = index[iuip[lib]];
}
iuip[lib] += UB_DESCRIPTOR; /* Skip descriptor. */
tempv = tempv2d;
for (jj = 0; jj < nsupc; ++jj) {
segsize = klst - usub[iukp + jj];
fnz = index[iuip[lib]++];
if ( segsize ) { /* Nonzero segment in U(k.j). */
ucol = &Unzval_br_ptr[lib][ruip[lib]];
for (i = 0, it = 0; i < nbrow; ++i) {
rel = lsub[lptr + i] - fnz;
z_sub(&ucol[rel], &ucol[rel], &tempv[it]);
++it;
}
tempv += ldt;
}
ruip[lib] += ilst - fnz;
}
} else { /* A(i,j) is in L. */
index = Lrowind_bc_ptr[ljb];
ldv = index[1]; /* LDA of the dest lusup. */
lptrj = BC_HEADER;
luptrj = 0;
ijb = index[lptrj];
while ( ijb != ib ) { /* Search for dest block --
blocks are not ordered! */
luptrj += index[lptrj+1];
lptrj += LB_DESCRIPTOR + index[lptrj+1];
ijb = index[lptrj];
}
/*
* Build indirect table. This is needed because the
* indices are not sorted.
*/
fnz = FstBlockC( ib );
lptrj += LB_DESCRIPTOR;
for (i = 0; i < index[lptrj-1]; ++i) {
rel = index[lptrj + i] - fnz;
indirect[rel] = i;
}
nzval = Lnzval_bc_ptr[ljb] + luptrj;
tempv = tempv2d;
for (jj = 0; jj < nsupc; ++jj) {
segsize = klst - usub[iukp + jj];
if ( segsize ) {
/*#pragma _CRI cache_bypass nzval,tempv*/
for (it = 0, i = 0; i < nbrow; ++i) {
rel = lsub[lptr + i] - fnz;
z_sub(&nzval[indirect[rel]],
&nzval[indirect[rel]],
&tempv[it]);
++it;
}
tempv += ldt;
}
nzval += ldv;
}
} /* if ib < jb ... */
lptr += nbrow;
luptr += nbrow;
} /* for lb ... */
rukp += usub[iukp - 1]; /* Move to block U(k,j+1) */
iukp += nsupc;
} /* for j ... */
} /* if k L(:,k) and U(k,:) are not empty */
}
/* ------------------------------------------
END MAIN LOOP: for k = ...
------------------------------------------ */
if ( Pr*Pc > 1 ) {
SUPERLU_FREE(Lsub_buf_2[0]); /* also free Lsub_buf_2[1] */
SUPERLU_FREE(Lval_buf_2[0]); /* also free Lval_buf_2[1] */
if ( Llu->bufmax[2] != 0 ) SUPERLU_FREE(Usub_buf);
if ( Llu->bufmax[3] != 0 ) SUPERLU_FREE(Uval_buf);
SUPERLU_FREE(send_req);
}
SUPERLU_FREE(Llu->ujrow);
SUPERLU_FREE(tempv2d);
SUPERLU_FREE(indirect);
SUPERLU_FREE(iuip);
SUPERLU_FREE(ruip);
/* Prepare error message. */
if ( *info == 0 ) *info = n + 1;
#if ( PROFlevel>=1 )
TIC(t1);
#endif
MPI_Allreduce( info, &iinfo, 1, mpi_int_t, MPI_MIN, grid->comm );
#if ( PROFlevel>=1 )
TOC(t2, t1);
stat->utime[COMM] += t2;
{
float msg_vol_max, msg_vol_sum, msg_cnt_max, msg_cnt_sum;
MPI_Reduce( &msg_cnt, &msg_cnt_sum,
1, MPI_FLOAT, MPI_SUM, 0, grid->comm );
MPI_Reduce( &msg_cnt, &msg_cnt_max,
1, MPI_FLOAT, MPI_MAX, 0, grid->comm );
MPI_Reduce( &msg_vol, &msg_vol_sum,
1, MPI_FLOAT, MPI_SUM, 0, grid->comm );
MPI_Reduce( &msg_vol, &msg_vol_max,
1, MPI_FLOAT, MPI_MAX, 0, grid->comm );
if ( !iam ) {
printf("\tPZGSTRF comm stat:"
"\tAvg\tMax\t\tAvg\tMax\n"
"\t\t\tCount:\t%.0f\t%.0f\tVol(MB)\t%.2f\t%.2f\n",
msg_cnt_sum/Pr/Pc, msg_cnt_max,
msg_vol_sum/Pr/Pc*1e-6, msg_vol_max*1e-6);
}
}
#endif
if ( iinfo == n + 1 ) *info = 0;
else *info = iinfo;
#if ( PRNTlevel==3 )
MPI_Allreduce( &zero_msg, &iinfo, 1, mpi_int_t, MPI_SUM, grid->comm );
if ( !iam ) printf(".. # msg of zero size\t%d\n", iinfo);
MPI_Allreduce( &total_msg, &iinfo, 1, mpi_int_t, MPI_SUM, grid->comm );
if ( !iam ) printf(".. # total msg\t%d\n", iinfo);
#endif
#if ( PRNTlevel==2 )
for (i = 0; i < Pr * Pc; ++i) {
if ( iam == i ) {
PrintLblocks(iam, nsupers, grid, Glu_persist, Llu);
PrintUblocks(iam, nsupers, grid, Glu_persist, Llu);
printf("(%d)\n", iam);
PrintInt10("Recv", nsupers, Llu->ToRecv);
}
MPI_Barrier( grid->comm );
}
#endif
#if ( DEBUGlevel>=1 )
printf("(%d) num_copy=%d, num_update=%d\n", iam, num_copy, num_update);
CHECK_MALLOC(iam, "Exit pzgstrf()");
#endif
} /* PZGSTRF_AGLOBAL */
void zlsum_fmod
/************************************************************************/
(
doublecomplex *lsum, /* Sum of local modifications. */
doublecomplex *x, /* X array (local) */
doublecomplex *xk, /* X[k]. */
doublecomplex *rtemp, /* Result of full matrix-vector multiply. */
int nrhs, /* Number of right-hand sides. */
int knsupc, /* Size of supernode k. */
int_t k, /* The k-th component of X. */
int_t *fmod, /* Modification count for L-solve. */
int_t nlb, /* Number of L blocks. */
int_t lptr, /* Starting position in lsub[*]. */
int_t luptr, /* Starting position in lusup[*]. */
int_t *xsup,
gridinfo_t *grid,
LocalLU_t *Llu,
MPI_Request send_req[],
SuperLUStat_t *stat
)
{
doublecomplex alpha = {1.0, 0.0}, beta = {0.0, 0.0};
doublecomplex *lusup, *lusup1;
doublecomplex *dest;
int iam, iknsupc, myrow, nbrow, nsupr, nsupr1, p, pi;
int_t i, ii, ik, il, ikcol, irow, j, lb, lk, rel;
int_t *lsub, *lsub1, nlb1, lptr1, luptr1;
int_t *ilsum = Llu->ilsum; /* Starting position of each supernode in lsum. */
int_t *frecv = Llu->frecv;
int_t **fsendx_plist = Llu->fsendx_plist;
MPI_Status status;
int test_flag;
iam = grid->iam;
myrow = MYROW( iam, grid );
lk = LBj( k, grid ); /* Local block number, column-wise. */
lsub = Llu->Lrowind_bc_ptr[lk];
lusup = Llu->Lnzval_bc_ptr[lk];
nsupr = lsub[1];
for (lb = 0; lb < nlb; ++lb) {
ik = lsub[lptr]; /* Global block number, row-wise. */
nbrow = lsub[lptr+1];
#ifdef _CRAY
CGEMM( ftcs2, ftcs2, &nbrow, &nrhs, &knsupc,
&alpha, &lusup[luptr], &nsupr, xk,
&knsupc, &beta, rtemp, &nbrow );
#else
zgemm_( "N", "N", &nbrow, &nrhs, &knsupc,
&alpha, &lusup[luptr], &nsupr, xk,
&knsupc, &beta, rtemp, &nbrow );
#endif
stat->ops[SOLVE] += 8 * nbrow * nrhs * knsupc + 2 * nbrow * nrhs;
lk = LBi( ik, grid ); /* Local block number, row-wise. */
iknsupc = SuperSize( ik );
il = LSUM_BLK( lk );
dest = &lsum[il];
lptr += LB_DESCRIPTOR;
rel = xsup[ik]; /* Global row index of block ik. */
for (i = 0; i < nbrow; ++i) {
irow = lsub[lptr++] - rel; /* Relative row. */
RHS_ITERATE(j)
z_sub(&dest[irow + j*iknsupc],
&dest[irow + j*iknsupc],
&rtemp[i + j*nbrow]);
}
luptr += nbrow;
if ( (--fmod[lk])==0 ) { /* Local accumulation done. */
ikcol = PCOL( ik, grid );
p = PNUM( myrow, ikcol, grid );
if ( iam != p ) {
#ifdef ISEND_IRECV
#if 1
MPI_Test( &send_req[myrow], &test_flag, &status );
#else
if ( send_req[myrow] != MPI_REQUEST_NULL )
MPI_Wait( &send_req[myrow], &status );
#endif
MPI_Isend( &lsum[il - LSUM_H], iknsupc * nrhs + LSUM_H,
SuperLU_MPI_DOUBLE_COMPLEX, p, LSUM, grid->comm, &send_req[myrow] );
#else
#ifdef BSEND
MPI_Bsend( &lsum[il - LSUM_H], iknsupc * nrhs + LSUM_H,
SuperLU_MPI_DOUBLE_COMPLEX, p, LSUM, grid->comm );
#else
MPI_Send( &lsum[il - LSUM_H], iknsupc * nrhs + LSUM_H,
SuperLU_MPI_DOUBLE_COMPLEX, p, LSUM, grid->comm );
#endif
#endif
#if ( DEBUGlevel>=2 )
printf("(%2d) Sent LSUM[%2.0f], size %2d, to P %2d\n",
iam, lsum[il-LSUM_H], iknsupc*nrhs+LSUM_H, p);
#endif
} else { /* Diagonal process: X[i] += lsum[i]. */
ii = X_BLK( lk );
RHS_ITERATE(j)
for (i = 0; i < iknsupc; ++i)
z_add(&x[i + ii + j*iknsupc],
&x[i + ii + j*iknsupc],
&lsum[i + il + j*iknsupc]);
if ( frecv[lk]==0 ) { /* Becomes a leaf node. */
fmod[lk] = -1; /* Do not solve X[k] in the future. */
lk = LBj( ik, grid );/* Local block number, column-wise. */
lsub1 = Llu->Lrowind_bc_ptr[lk];
lusup1 = Llu->Lnzval_bc_ptr[lk];
nsupr1 = lsub1[1];
#ifdef _CRAY
CTRSM(ftcs1, ftcs1, ftcs2, ftcs3, &iknsupc, &nrhs, &alpha,
lusup1, &nsupr1, &x[ii], &iknsupc);
#else
ztrsm_("L", "L", "N", "U", &iknsupc, &nrhs, &alpha,
lusup1, &nsupr1, &x[ii], &iknsupc);
#endif
stat->ops[SOLVE] += 4 * iknsupc * (iknsupc - 1) * nrhs
+ 10 * knsupc * nrhs; /* complex division */
#if ( DEBUGlevel>=2 )
printf("(%2d) Solve X[%2d]\n", iam, ik);
#endif
/*
* Send Xk to process column Pc[k].
*/
for (p = 0; p < grid->nprow; ++p)
if ( fsendx_plist[lk][p] != EMPTY ) {
pi = PNUM( p, ikcol, grid );
#ifdef ISEND_IRECV
#if 1
MPI_Test( &send_req[p], &test_flag, &status );
#else
if ( send_req[p] != MPI_REQUEST_NULL )
MPI_Wait( &send_req[p], &status );
#endif
MPI_Isend( &x[ii - XK_H], iknsupc * nrhs + XK_H,
SuperLU_MPI_DOUBLE_COMPLEX, pi, Xk, grid->comm,
&send_req[p] );
#else
#ifdef BSEND
MPI_Bsend( &x[ii - XK_H], iknsupc * nrhs + XK_H,
SuperLU_MPI_DOUBLE_COMPLEX, pi, Xk, grid->comm );
#else
MPI_Send( &x[ii - XK_H], iknsupc * nrhs + XK_H,
SuperLU_MPI_DOUBLE_COMPLEX, pi, Xk, grid->comm );
#endif
#endif
#if ( DEBUGlevel>=2 )
printf("(%2d) Sent X[%2.0f] to P %2d\n",
iam, x[ii-XK_H], pi);
#endif
}
/*
* Perform local block modifications.
*/
nlb1 = lsub1[0] - 1;
lptr1 = BC_HEADER + LB_DESCRIPTOR + iknsupc;
luptr1 = iknsupc; /* Skip diagonal block L(I,I). */
zlsum_fmod(lsum, x, &x[ii], rtemp, nrhs, iknsupc, ik,
fmod, nlb1, lptr1, luptr1, xsup,
grid, Llu, send_req, stat);
#ifdef ISEND_IRECV
/* Wait for previous Isends to complete. */
for (p = 0; p < grid->nprow; ++p) {
if ( fsendx_plist[lk][p] != EMPTY )
/*MPI_Wait( &send_req[p], &status );*/
MPI_Test( &send_req[p], &test_flag, &status );
}
#endif
} /* if frecv[lk] == 0 */
} /* if iam == p */
} /* if fmod[lk] == 0 */
} /* for lb ... */
