GLOBAL void UMF_assemble #else GLOBAL void UMF_assemble_fixq #endif ( NumericType *Numeric, WorkType *Work ) { /* ---------------------------------------------------------------------- */ /* local variables */ /* ---------------------------------------------------------------------- */ Int e, i, row, col, i2, nrows, ncols, f, tpi, extcdeg, extrdeg, rdeg0, cdeg0, son_list, next, nrows_to_assemble, ncols_to_assemble, ngetrows, j, j2, nrowsleft, /* number of rows remaining in S */ ncolsleft, /* number of columns remaining in S */ prior_Lson, prior_Uson, *E, *Cols, *Rows, *Wm, *Woo, *Row_tuples, *Row_degree, *Row_tlen, *Col_tuples, *Col_tlen ; Unit *Memory, *p ; Element *ep ; Tuple *tp, *tp1, *tp2, *tpend ; Entry *S, /* a pointer into the contribution block of a son */ *Fcblock, /* current contribution block */ *Fcol ; /* a column of FC */ Int *Frpos, *Fcpos, fnrows, /* number of rows in contribution block in F */ fncols ; /* number of columns in contribution block in F */ #if !defined (FIXQ) || !defined (NDEBUG) Int *Col_degree ; #endif #ifndef NDEBUG Int n_row, n_col ; n_row = Work->n_row ; n_col = Work->n_col ; DEBUG3 (("::Assemble SCANS 1-4\n")) ; UMF_dump_current_front (Numeric, Work, TRUE) ; #endif #if !defined (FIXQ) || !defined (NDEBUG) Col_degree = Numeric->Cperm ; /* not updated if FIXQ is true */ #endif /* ---------------------------------------------------------------------- */ /* get parameters */ /* ---------------------------------------------------------------------- */ fncols = Work->fncols ; fnrows = Work->fnrows ; Fcpos = Work->Fcpos ; Frpos = Work->Frpos ; Row_degree = Numeric->Rperm ; Row_tuples = Numeric->Uip ; Row_tlen = Numeric->Uilen ; Col_tuples = Numeric->Lip ; Col_tlen = Numeric->Lilen ; E = Work->E ; Memory = Numeric->Memory ; Wm = Work->Wm ; Woo = Work->Woo ; rdeg0 = Work->rdeg0 ; cdeg0 = Work->cdeg0 ; #ifndef NDEBUG DEBUG6 (("============================================\n")) ; DEBUG6 (("Degree update, assembly.\n")) ; DEBUG6 (("pivot row pattern: fncols="ID"\n", fncols)) ; for (j = 0 ; j < fncols ; j++) { col = Work->Fcols [j] ; DEBUG6 ((ID" ", col)) ; ASSERT (Fcpos [col] == j * Work->fnr_curr) ; ASSERT (NON_PIVOTAL_COL (col)) ; } ASSERT (Fcpos [Work->pivcol] >= 0) ; DEBUG6 (("pivcol: "ID" pos "ID" fnr_curr "ID" fncols "ID"\n", Work->pivcol, Fcpos [Work->pivcol], Work->fnr_curr, fncols)) ; ASSERT (Fcpos [Work->pivcol] < fncols * Work->fnr_curr) ; DEBUG6 (("\npivot col pattern: fnrows="ID"\n", fnrows)) ; for (i = 0 ; i < fnrows ; i++) { row = Work->Frows [i] ; DEBUG6 ((ID" ", row)) ; ASSERT (Frpos [row] == i) ; ASSERT (NON_PIVOTAL_ROW (row)) ; } DEBUG6 (("\n")) ; ASSERT (Frpos [Work->pivrow] >= 0) ; ASSERT (Frpos [Work->pivrow] < fnrows) ; ASSERT (Work->Flublock == (Entry *) (Numeric->Memory + E [0])) ; ASSERT (Work->Fcblock == Work->Flublock + Work->nb * (Work->nb + Work->fnr_curr + Work->fnc_curr)) ; #endif Fcblock = Work->Fcblock ; /* ---------------------------------------------------------------------- */ /* determine the largest actual frontal matrix size (for Info only) */ /* ---------------------------------------------------------------------- */ ASSERT (fnrows == Work->fnrows_new + 1) ; ASSERT (fncols == Work->fncols_new + 1) ; Numeric->maxnrows = MAX (Numeric->maxnrows, fnrows) ; Numeric->maxncols = MAX (Numeric->maxncols, fncols) ; /* this is safe from integer overflow, since the current frontal matrix * is already allocated. */ Numeric->maxfrsize = MAX (Numeric->maxfrsize, fnrows * fncols) ; /* ---------------------------------------------------------------------- */ /* assemble from prior elements into the current frontal matrix */ /* ---------------------------------------------------------------------- */ DEBUG2 (("New assemble start [prior_element:"ID"\n", Work->prior_element)) ; /* Currently no rows or columns are marked. No elements are scanned, */ /* that is, (ep->next == EMPTY) is true for all elements */ son_list = 0 ; /* start creating son_list [ */ /* ---------------------------------------------------------------------- */ /* determine if most recent element is Lson or Uson of current front */ /* ---------------------------------------------------------------------- */ if (!Work->do_extend) { prior_Uson = ( Work->pivcol_in_front && !Work->pivrow_in_front) ; prior_Lson = (!Work->pivcol_in_front && Work->pivrow_in_front) ; if (prior_Uson || prior_Lson) { e = Work->prior_element ; if (e != EMPTY) { ASSERT (E [e]) ; p = Memory + E [e] ; ep = (Element *) p ; ep->next = son_list ; son_list = e ; #ifndef NDEBUG DEBUG2 (("e "ID" is Prior son "ID" "ID"\n", e, prior_Uson, prior_Lson)) ; UMF_dump_element (Numeric, Work, e, FALSE) ; #endif ASSERT (E [e]) ; } } } Work->prior_element = EMPTY ; /* ---------------------------------------------------------------------- */ /* SCAN1-row: scan the element lists of each new row in the pivot col */ /* and compute the external column degree for each frontal */ /* ---------------------------------------------------------------------- */ for (i2 = Work->fscan_row ; i2 < fnrows ; i2++) { /* Get a row */ row = Work->NewRows [i2] ; if (row < 0) row = FLIP (row) ; ASSERT (row >= 0 && row < n_row) ; DEBUG6 (("SCAN1-row: "ID"\n", row)) ; #ifndef NDEBUG UMF_dump_rowcol (0, Numeric, Work, row, FALSE) ; #endif ASSERT (NON_PIVOTAL_ROW (row)) ; tpi = Row_tuples [row] ; if (!tpi) continue ; tp = (Tuple *) (Memory + tpi) ; tp1 = tp ; tp2 = tp ; tpend = tp + Row_tlen [row] ; for ( ; tp < tpend ; tp++) { e = tp->e ; ASSERT (e > 0 && e <= Work->nel) ; if (!E [e]) continue ; /* element already deallocated */ f = tp->f ; p = Memory + E [e] ; ep = (Element *) p ; p += UNITS (Element, 1) ; Rows = ((Int *) p) + ep->ncols ; if (Rows [f] == EMPTY) continue ; /* row already assembled */ ASSERT (row == Rows [f]) ; if (ep->cdeg < cdeg0) { /* first time seen in scan1-row */ ep->cdeg = ep->nrowsleft + cdeg0 ; DEBUG6 (("e "ID" First seen: cdeg: "ID" ", e, ep->cdeg-cdeg0)) ; ASSERT (ep->ncolsleft > 0 && ep->nrowsleft > 0) ; } ep->cdeg-- ; /* decrement external column degree */ DEBUG6 (("e "ID" New ext col deg: "ID"\n", e, ep->cdeg - cdeg0)) ; /* this element is not yet in the new son list */ if (ep->cdeg == cdeg0 && ep->next == EMPTY) { /* A new LUson or Uson has been found */ ep->next = son_list ; son_list = e ; } ASSERT (ep->cdeg >= cdeg0) ; *tp2++ = *tp ; /* leave the tuple in the list */ } Row_tlen [row] = tp2 - tp1 ; } /* ---------------------------------------------------------------------- */ /* SCAN1-col: scan the element lists of each new col in the pivot row */ /* and compute the external row degree for each frontal */ /* ---------------------------------------------------------------------- */ for (j2 = Work->fscan_col ; j2 < fncols ; j2++) { /* Get a column */ col = Work->NewCols [j2] ; if (col < 0) col = FLIP (col) ; ASSERT (col >= 0 && col < n_col) ; DEBUG6 (("SCAN 1-col: "ID"\n", col)) ; #ifndef NDEBUG UMF_dump_rowcol (1, Numeric, Work, col, FALSE) ; #endif ASSERT (NON_PIVOTAL_COL (col)) ; tpi = Col_tuples [col] ; if (!tpi) continue ; tp = (Tuple *) (Memory + tpi) ; tp1 = tp ; tp2 = tp ; tpend = tp + Col_tlen [col] ; for ( ; tp < tpend ; tp++) { e = tp->e ; ASSERT (e > 0 && e <= Work->nel) ; if (!E [e]) continue ; /* element already deallocated */ f = tp->f ; p = Memory + E [e] ; ep = (Element *) p ; p += UNITS (Element, 1) ; Cols = (Int *) p ; if (Cols [f] == EMPTY) continue ; /* column already assembled */ ASSERT (col == Cols [f]) ; if (ep->rdeg < rdeg0) { /* first time seen in scan1-col */ ep->rdeg = ep->ncolsleft + rdeg0 ; DEBUG6 (("e "ID" First seen: rdeg: "ID" ", e, ep->rdeg-rdeg0)) ; ASSERT (ep->ncolsleft > 0 && ep->nrowsleft > 0) ; } ep->rdeg-- ; /* decrement external row degree */ DEBUG6 (("e "ID" New ext row degree: "ID"\n", e, ep->rdeg-rdeg0)) ; if (ep->rdeg == rdeg0 && ep->next == EMPTY) { /* A new LUson or Lson has been found */ ep->next = son_list ; son_list = e ; } ASSERT (ep->rdeg >= rdeg0) ; *tp2++ = *tp ; /* leave the tuple in the list */ } Col_tlen [col] = tp2 - tp1 ; } /* ---------------------------------------------------------------------- */ /* assemble new sons via full scans */ /* ---------------------------------------------------------------------- */ next = EMPTY ; for (e = son_list ; e > 0 ; e = next) { ASSERT (e > 0 && e <= Work->nel && E [e]) ; p = Memory + E [e] ; DEBUG2 (("New son: "ID"\n", e)) ; #ifndef NDEBUG UMF_dump_element (Numeric, Work, e, FALSE) ; #endif GET_ELEMENT (ep, p, Cols, Rows, ncols, nrows, S) ; nrowsleft = ep->nrowsleft ; ncolsleft = ep->ncolsleft ; next = ep->next ; ep->next = EMPTY ; extrdeg = (ep->rdeg < rdeg0) ? ncolsleft : (ep->rdeg - rdeg0) ; extcdeg = (ep->cdeg < cdeg0) ? nrowsleft : (ep->cdeg - cdeg0) ; ncols_to_assemble = ncolsleft - extrdeg ; nrows_to_assemble = nrowsleft - extcdeg ; DEBUG2 (("extrdeg "ID" extcdeg "ID"\n", extrdeg, extcdeg)) ; if (extrdeg == 0 && extcdeg == 0) { /* -------------------------------------------------------------- */ /* this is an LUson - assemble an entire contribution block */ /* -------------------------------------------------------------- */ DEBUG6 (("LUson found: "ID"\n", e)) ; if (nrows == nrowsleft) { /* ---------------------------------------------------------- */ /* no rows assembled out of this LUson yet */ /* ---------------------------------------------------------- */ /* compute the compressed column offset vector*/ /* [ use Wm [0..nrows-1] for offsets */ #pragma ivdep for (i = 0 ; i < nrows ; i++) { row = Rows [i] ; Row_degree [row] -= ncolsleft ; Wm [i] = Frpos [row] ; } if (ncols == ncolsleft) { /* ------------------------------------------------------ */ /* no rows or cols assembled out of LUson yet */ /* ------------------------------------------------------ */ for (j = 0 ; j < ncols ; j++) { col = Cols [j] ; #ifndef FIXQ Col_degree [col] -= nrowsleft ; #endif Fcol = Fcblock + Fcpos [col] ; #pragma ivdep for (i = 0 ; i < nrows ; i++) { /* Fcol [Wm [i]] += S [i] ; */ ASSEMBLE (Fcol [Wm [i]], S [i]) ; } S += nrows ; } } else { /* ------------------------------------------------------ */ /* only cols have been assembled out of LUson */ /* ------------------------------------------------------ */ for (j = 0 ; j < ncols ; j++) { col = Cols [j] ; if (col >= 0) { #ifndef FIXQ Col_degree [col] -= nrowsleft ; #endif Fcol = Fcblock + Fcpos [col] ; #pragma ivdep for (i = 0 ; i < nrows ; i++) { /* Fcol [Wm [i]] += S [i] ; */ ASSEMBLE (Fcol [Wm [i]], S [i]) ; } } S += nrows ; } } /* ] done using Wm [0..nrows-1] for offsets */ } else { /* ---------------------------------------------------------- */ /* some rows have been assembled out of this LUson */ /* ---------------------------------------------------------- */ /* compute the compressed column offset vector*/ /* [ use Woo,Wm [0..nrowsleft-1] for offsets */ ngetrows = 0 ; for (i = 0 ; i < nrows ; i++) { row = Rows [i] ; if (row >= 0) { Row_degree [row] -= ncolsleft ; Woo [ngetrows] = i ; Wm [ngetrows++] = Frpos [row] ; } } ASSERT (ngetrows == nrowsleft) ; if (ncols == ncolsleft) { /* ------------------------------------------------------ */ /* only rows have been assembled out of this LUson */ /* ------------------------------------------------------ */ for (j = 0 ; j < ncols ; j++) { col = Cols [j] ; #ifndef FIXQ Col_degree [col] -= nrowsleft ; #endif Fcol = Fcblock + Fcpos [col] ; #pragma ivdep for (i = 0 ; i < nrowsleft ; i++) { /* Fcol [Wm [i]] += S [Woo [i]] ; */ ASSEMBLE (Fcol [Wm [i]], S [Woo [i]]) ; } S += nrows ; } } else { /* ------------------------------------------------------ */ /* both rows and columns have been assembled out of LUson */ /* ------------------------------------------------------ */ for (j = 0 ; j < ncols ; j++) { col = Cols [j] ; if (col >= 0) { #ifndef FIXQ Col_degree [col] -= nrowsleft ; #endif Fcol = Fcblock + Fcpos [col] ; #pragma ivdep for (i = 0 ; i < nrowsleft ; i++) { /* Fcol [Wm [i]] += S [Woo [i]] ; */ ASSEMBLE (Fcol [Wm [i]], S [Woo [i]]) ; } } S += nrows ; } } /* ] done using Woo,Wm [0..nrowsleft-1] */ } /* deallocate the element: remove from ordered list */ UMF_mem_free_tail_block (Numeric, E [e]) ; E [e] = 0 ; } else if (extcdeg == 0) { /* -------------------------------------------------------------- */ /* this is a Uson - assemble all possible columns */ /* -------------------------------------------------------------- */ DEBUG6 (("New USON: "ID"\n", e)) ; ASSERT (extrdeg > 0) ; DEBUG6 (("New uson "ID" cols to do "ID"\n", e, ncols_to_assemble)) ; if (ncols_to_assemble > 0) { Int skip = FALSE ; if (ncols_to_assemble * 16 < ncols && nrows == 1) { /* this is a tall and thin frontal matrix consisting of * only one column (most likely an original column). Do * not assemble it. It cannot be the pivot column, since * the pivot column element would be an LU son, not an Lson, * of the current frontal matrix. */ ASSERT (nrowsleft == 1) ; ASSERT (Rows [0] >= 0 && Rows [0] < Work->n_row) ; skip = TRUE ; Work->any_skip = TRUE ; } if (!skip) { if (nrows == nrowsleft) { /* -------------------------------------------------- */ /* no rows have been assembled out of this Uson yet */ /* -------------------------------------------------- */ /* compute the compressed column offset vector */ /* [ use Wm [0..nrows-1] for offsets */ #pragma ivdep for (i = 0 ; i < nrows ; i++) { row = Rows [i] ; ASSERT (row >= 0 && row < n_row) ; Row_degree [row] -= ncols_to_assemble ; Wm [i] = Frpos [row] ; } for (j = 0 ; j < ncols ; j++) { col = Cols [j] ; if ((col >= 0) && (Fcpos [col] >= 0)) { #ifndef FIXQ Col_degree [col] -= nrowsleft ; #endif Fcol = Fcblock + Fcpos [col] ; #pragma ivdep for (i = 0 ; i < nrows ; i++) { /* Fcol [Wm [i]] += S [i] ; */ ASSEMBLE (Fcol [Wm [i]], S [i]) ; } /* flag the column as assembled from Uson */ Cols [j] = EMPTY ; } S += nrows ; } /* ] done using Wm [0..nrows-1] for offsets */ } else { /* -------------------------------------------------- */ /* some rows have been assembled out of this Uson */ /* -------------------------------------------------- */ /* compute the compressed column offset vector*/ /* [ use Woo,Wm [0..nrows-1] for offsets */ ngetrows = 0 ; for (i = 0 ; i < nrows ; i++) { row = Rows [i] ; if (row >= 0) { Row_degree [row] -= ncols_to_assemble ; ASSERT (row < n_row && Frpos [row] >= 0) ; Woo [ngetrows] = i ; Wm [ngetrows++] = Frpos [row] ; } } ASSERT (ngetrows == nrowsleft) ; for (j = 0 ; j < ncols ; j++) { col = Cols [j] ; if ((col >= 0) && (Fcpos [col] >= 0)) { #ifndef FIXQ Col_degree [col] -= nrowsleft ; #endif Fcol = Fcblock + Fcpos [col] ; #pragma ivdep for (i = 0 ; i < nrowsleft ; i++) { /* Fcol [Wm [i]] += S [Woo [i]] ; */ ASSEMBLE (Fcol [Wm [i]], S [Woo [i]]) ; } /* flag the column as assembled from Uson */ Cols [j] = EMPTY ; } S += nrows ; } /* ] done using Woo,Wm */ } ep->ncolsleft = extrdeg ; } } } else { /* -------------------------------------------------------------- */ /* this is an Lson - assemble all possible rows */ /* -------------------------------------------------------------- */ DEBUG6 (("New LSON: "ID"\n", e)) ; ASSERT (extrdeg == 0 && extcdeg > 0) ; DEBUG6 (("New lson "ID" rows to do "ID"\n", e, nrows_to_assemble)) ; if (nrows_to_assemble > 0) { Int skip = FALSE ; if (nrows_to_assemble * 16 < nrows && ncols == 1) { /* this is a tall and thin frontal matrix consisting of * only one column (most likely an original column). Do * not assemble it. It cannot be the pivot column, since * the pivot column element would be an LU son, not an Lson, * of the current frontal matrix. */ ASSERT (ncolsleft == 1) ; ASSERT (Cols [0] >= 0 && Cols [0] < Work->n_col) ; Work->any_skip = TRUE ; skip = TRUE ; } if (!skip) { /* compute the compressed column offset vector */ /* [ use Woo,Wm [0..nrows-1] for offsets */ ngetrows = 0 ; for (i = 0 ; i < nrows ; i++) { row = Rows [i] ; if ((row >= 0) && (Frpos [row] >= 0)) { ASSERT (row < n_row) ; Row_degree [row] -= ncolsleft ; Woo [ngetrows] = i ; Wm [ngetrows++] = Frpos [row] ; /* flag the row as assembled from the Lson */ Rows [i] = EMPTY ; } } ASSERT (nrowsleft - ngetrows == extcdeg) ; ASSERT (ngetrows == nrows_to_assemble) ; if (ncols == ncolsleft) { /* -------------------------------------------------- */ /* no columns assembled out this Lson yet */ /* -------------------------------------------------- */ for (j = 0 ; j < ncols ; j++) { col = Cols [j] ; ASSERT (col >= 0 && col < n_col) ; #ifndef FIXQ Col_degree [col] -= nrows_to_assemble ; #endif Fcol = Fcblock + Fcpos [col] ; #pragma ivdep for (i = 0 ; i < nrows_to_assemble ; i++) { /* Fcol [Wm [i]] += S [Woo [i]] ; */ ASSEMBLE (Fcol [Wm [i]], S [Woo [i]]) ; } S += nrows ; } } else { /* -------------------------------------------------- */ /* some columns have been assembled out of this Lson */ /* -------------------------------------------------- */ for (j = 0 ; j < ncols ; j++) { col = Cols [j] ; ASSERT (col < n_col) ; if (col >= 0) { #ifndef FIXQ Col_degree [col] -= nrows_to_assemble ; #endif Fcol = Fcblock + Fcpos [col] ; #pragma ivdep for (i = 0 ; i < nrows_to_assemble ; i++) { /* Fcol [Wm [i]] += S [Woo [i]] ; */ ASSEMBLE (Fcol [Wm [i]], S [Woo [i]]) ; } } S += nrows ; } } /* ] done using Woo,Wm */ ep->nrowsleft = extcdeg ; } } } } /* Note that garbage collection, and build tuples */ /* both destroy the son list. */ /* ] son_list now empty */ /* ---------------------------------------------------------------------- */ /* If frontal matrix extended, assemble old L/Usons from new rows/cols */ /* ---------------------------------------------------------------------- */ /* ---------------------------------------------------------------------- */ /* SCAN2-row: assemble rows of old Lsons from the new rows */ /* ---------------------------------------------------------------------- */ #ifndef NDEBUG DEBUG7 (("Current frontal matrix: (prior to scan2-row)\n")) ; UMF_dump_current_front (Numeric, Work, TRUE) ; #endif /* rescan the pivot row */ if (Work->any_skip) { row_assemble (Work->pivrow, Numeric, Work) ; } if (Work->do_scan2row) { for (i2 = Work->fscan_row ; i2 < fnrows ; i2++) { /* Get a row */ row = Work->NewRows [i2] ; if (row < 0) row = FLIP (row) ; ASSERT (row >= 0 && row < n_row) ; if (!(row == Work->pivrow && Work->any_skip)) { /* assemble it */ row_assemble (row, Numeric, Work) ; } } } /* ---------------------------------------------------------------------- */ /* SCAN2-col: assemble columns of old Usons from the new columns */ /* ---------------------------------------------------------------------- */ #ifndef NDEBUG DEBUG7 (("Current frontal matrix: (prior to scan2-col)\n")) ; UMF_dump_current_front (Numeric, Work, TRUE) ; #endif /* rescan the pivot col */ if (Work->any_skip) { col_assemble (Work->pivcol, Numeric, Work) ; } if (Work->do_scan2col) { for (j2 = Work->fscan_col ; j2 < fncols ; j2++) { /* Get a column */ col = Work->NewCols [j2] ; if (col < 0) col = FLIP (col) ; ASSERT (col >= 0 && col < n_col) ; if (!(col == Work->pivcol && Work->any_skip)) { /* assemble it */ col_assemble (col, Numeric, Work) ; } } } /* ---------------------------------------------------------------------- */ /* done. the remainder of this routine is used only when in debug mode */ /* ---------------------------------------------------------------------- */ #ifndef NDEBUG /* ---------------------------------------------------------------------- */ /* when debugging: make sure the assembly did everything that it could */ /* ---------------------------------------------------------------------- */ DEBUG3 (("::Assemble done\n")) ; for (i2 = 0 ; i2 < fnrows ; i2++) { /* Get a row */ row = Work->Frows [i2] ; ASSERT (row >= 0 && row < n_row) ; DEBUG6 (("DEBUG SCAN 1: "ID"\n", row)) ; UMF_dump_rowcol (0, Numeric, Work, row, TRUE) ; ASSERT (NON_PIVOTAL_ROW (row)) ; tpi = Row_tuples [row] ; if (!tpi) continue ; tp = (Tuple *) (Memory + tpi) ; tpend = tp + Row_tlen [row] ; for ( ; tp < tpend ; tp++) { e = tp->e ; ASSERT (e > 0 && e <= Work->nel) ; if (!E [e]) continue ; /* element already deallocated */ f = tp->f ; p = Memory + E [e] ; ep = (Element *) p ; p += UNITS (Element, 1) ; Cols = (Int *) p ; Rows = ((Int *) p) + ep->ncols ; if (Rows [f] == EMPTY) continue ; /* row already assembled */ ASSERT (row == Rows [f]) ; extrdeg = (ep->rdeg < rdeg0) ? ep->ncolsleft : (ep->rdeg - rdeg0) ; extcdeg = (ep->cdeg < cdeg0) ? ep->nrowsleft : (ep->cdeg - cdeg0) ; DEBUG6 (( "e "ID" After assembly ext row deg: "ID" ext col degree "ID"\n", e, extrdeg, extcdeg)) ; if (Work->any_skip) { /* no Lsons in any row, except for very tall and thin ones */ ASSERT (extrdeg >= 0) ; if (extrdeg == 0) { /* this is an unassemble Lson */ ASSERT (ep->ncols == 1) ; ASSERT (ep->ncolsleft == 1) ; col = Cols [0] ; ASSERT (col != Work->pivcol) ; } } else { /* no Lsons in any row */ ASSERT (extrdeg > 0) ; /* Uson external row degree is = number of cols left */ ASSERT (IMPLIES (extcdeg == 0, extrdeg == ep->ncolsleft)) ; } } } /* ---------------------------------------------------------------------- */ for (j2 = 0 ; j2 < fncols ; j2++) { /* Get a column */ col = Work->Fcols [j2] ; ASSERT (col >= 0 && col < n_col) ; DEBUG6 (("DEBUG SCAN 2: "ID"\n", col)) ; #ifndef FIXQ UMF_dump_rowcol (1, Numeric, Work, col, TRUE) ; #else UMF_dump_rowcol (1, Numeric, Work, col, FALSE) ; #endif ASSERT (NON_PIVOTAL_COL (col)) ; tpi = Col_tuples [col] ; if (!tpi) continue ; tp = (Tuple *) (Memory + tpi) ; tpend = tp + Col_tlen [col] ; for ( ; tp < tpend ; tp++) { e = tp->e ; ASSERT (e > 0 && e <= Work->nel) ; if (!E [e]) continue ; /* element already deallocated */ f = tp->f ; p = Memory + E [e] ; ep = (Element *) p ; p += UNITS (Element, 1) ; Cols = (Int *) p ; Rows = ((Int *) p) + ep->ncols ; if (Cols [f] == EMPTY) continue ; /* column already assembled */ ASSERT (col == Cols [f]) ; extrdeg = (ep->rdeg < rdeg0) ? ep->ncolsleft : (ep->rdeg - rdeg0) ; extcdeg = (ep->cdeg < cdeg0) ? ep->nrowsleft : (ep->cdeg - cdeg0) ; DEBUG6 (("e "ID" After assembly ext col deg: "ID"\n", e, extcdeg)) ; if (Work->any_skip) { /* no Usons in any column, except for very tall and thin ones */ ASSERT (extcdeg >= 0) ; if (extcdeg == 0) { /* this is an unassemble Uson */ ASSERT (ep->nrows == 1) ; ASSERT (ep->nrowsleft == 1) ; row = Rows [0] ; ASSERT (row != Work->pivrow) ; } } else { /* no Usons in any column */ ASSERT (extcdeg > 0) ; /* Lson external column degree is = number of rows left */ ASSERT (IMPLIES (extrdeg == 0, extcdeg == ep->nrowsleft)) ; } } } #endif /* NDEBUG */ }
GLOBAL Int UMF_build_tuples ( NumericType *Numeric, WorkType *Work ) { /* ---------------------------------------------------------------------- */ /* local variables */ /* ---------------------------------------------------------------------- */ Int e, nrows, ncols, nel, *Rows, *Cols, row, col, n_row, n_col, *E, *Row_tuples, *Row_degree, *Row_tlen, *Col_tuples, *Col_degree, *Col_tlen, n1 ; Element *ep ; Unit *p ; Tuple tuple, *tp ; /* ---------------------------------------------------------------------- */ /* get parameters */ /* ---------------------------------------------------------------------- */ E = Work->E ; Col_degree = Numeric->Cperm ; /* for NON_PIVOTAL_COL macro */ Row_degree = Numeric->Rperm ; /* for NON_PIVOTAL_ROW macro */ Row_tuples = Numeric->Uip ; Row_tlen = Numeric->Uilen ; Col_tuples = Numeric->Lip ; Col_tlen = Numeric->Lilen ; n_row = Work->n_row ; n_col = Work->n_col ; nel = Work->nel ; n1 = Work->n1 ; DEBUG3 (("BUILD_TUPLES: n_row "ID" n_col "ID" nel "ID"\n", n_row, n_col, nel)) ; /* ---------------------------------------------------------------------- */ /* allocate space for the tuple lists */ /* ---------------------------------------------------------------------- */ /* Garbage collection and memory reallocation have already attempted to */ /* ensure that there is enough memory for all the tuple lists. If */ /* memory allocation fails here, then there is nothing more to be done. */ for (row = n1 ; row < n_row ; row++) { if (NON_PIVOTAL_ROW (row)) { Row_tuples [row] = UMF_mem_alloc_tail_block (Numeric, UNITS (Tuple, TUPLES (Row_tlen [row]))) ; if (!Row_tuples [row]) { /* :: out of memory for row tuples :: */ DEBUGm4 (("out of memory: build row tuples\n")) ; return (FALSE) ; /* out of memory for row tuples */ } Row_tlen [row] = 0 ; } } /* push on stack in reverse order, so column tuples are in the order */ /* that they will be deleted. */ for (col = n_col-1 ; col >= n1 ; col--) { if (NON_PIVOTAL_COL (col)) { Col_tuples [col] = UMF_mem_alloc_tail_block (Numeric, UNITS (Tuple, TUPLES (Col_tlen [col]))) ; if (!Col_tuples [col]) { /* :: out of memory for col tuples :: */ DEBUGm4 (("out of memory: build col tuples\n")) ; return (FALSE) ; /* out of memory for col tuples */ } Col_tlen [col] = 0 ; } } #ifndef NDEBUG UMF_dump_memory (Numeric) ; #endif /* ---------------------------------------------------------------------- */ /* create the tuple lists (exclude element 0) */ /* ---------------------------------------------------------------------- */ /* for all elements, in order of creation */ for (e = 1 ; e <= nel ; e++) { DEBUG9 (("Adding tuples for element: "ID" at "ID"\n", e, E [e])) ; ASSERT (E [e]) ; /* no external fragmentation */ p = Numeric->Memory + E [e] ; GET_ELEMENT_PATTERN (ep, p, Cols, Rows, ncols) ; nrows = ep->nrows ; ASSERT (e != 0) ; ASSERT (e == 0 || nrows == ep->nrowsleft) ; ASSERT (e == 0 || ncols == ep->ncolsleft) ; tuple.e = e ; for (tuple.f = 0 ; tuple.f < ncols ; tuple.f++) { col = Cols [tuple.f] ; ASSERT (col >= n1 && col < n_col) ; ASSERT (NON_PIVOTAL_COL (col)) ; ASSERT (Col_tuples [col]) ; tp = ((Tuple *) (Numeric->Memory + Col_tuples [col])) + Col_tlen [col]++ ; *tp = tuple ; #ifndef NDEBUG UMF_dump_rowcol (1, Numeric, Work, col, FALSE) ; #endif } for (tuple.f = 0 ; tuple.f < nrows ; tuple.f++) { row = Rows [tuple.f] ; ASSERT (row >= n1 && row < n_row) ; ASSERT (NON_PIVOTAL_COL (col)) ; ASSERT (Row_tuples [row]) ; tp = ((Tuple *) (Numeric->Memory + Row_tuples [row])) + Row_tlen [row]++ ; *tp = tuple ; #ifndef NDEBUG UMF_dump_rowcol (0, Numeric, Work, row, FALSE) ; #endif } } /* ---------------------------------------------------------------------- */ /* the tuple lists are now valid, and can be scanned */ /* ---------------------------------------------------------------------- */ #ifndef NDEBUG UMF_dump_memory (Numeric) ; UMF_dump_matrix (Numeric, Work, FALSE) ; #endif DEBUG3 (("BUILD_TUPLES: done\n")) ; return (TRUE) ; }
PRIVATE void row_assemble ( Int row, NumericType *Numeric, WorkType *Work ) { Entry *S, *Fcblock, *Frow ; Int tpi, e, *E, *Fcpos, *Frpos, *Row_degree, *Row_tuples, *Row_tlen, rdeg0, f, nrows, ncols, *Rows, *Cols, col, ncolsleft, j ; Tuple *tp, *tp1, *tp2, *tpend ; Unit *Memory, *p ; Element *ep ; #ifndef FIXQ Int *Col_degree ; Col_degree = Numeric->Cperm ; #endif Row_tuples = Numeric->Uip ; tpi = Row_tuples [row] ; if (!tpi) return ; Memory = Numeric->Memory ; E = Work->E ; Fcpos = Work->Fcpos ; Frpos = Work->Frpos ; Row_degree = Numeric->Rperm ; Row_tlen = Numeric->Uilen ; E = Work->E ; Memory = Numeric->Memory ; rdeg0 = Work->rdeg0 ; Fcblock = Work->Fcblock ; #ifndef NDEBUG DEBUG6 (("SCAN2-row: "ID"\n", row)) ; UMF_dump_rowcol (0, Numeric, Work, row, FALSE) ; #endif ASSERT (NON_PIVOTAL_ROW (row)) ; tp = (Tuple *) (Memory + tpi) ; tp1 = tp ; tp2 = tp ; tpend = tp + Row_tlen [row] ; for ( ; tp < tpend ; tp++) { e = tp->e ; ASSERT (e > 0 && e <= Work->nel) ; if (!E [e]) continue ; /* element already deallocated */ f = tp->f ; p = Memory + E [e] ; ep = (Element *) p ; p += UNITS (Element, 1) ; Cols = (Int *) p ; Rows = Cols + ep->ncols ; if (Rows [f] == EMPTY) continue ; /* row already assembled */ ASSERT (row == Rows [f] && row >= 0 && row < Work->n_row) ; if (ep->rdeg == rdeg0) { /* ------------------------------------------------------ */ /* this is an old Lson - assemble just one row */ /* ------------------------------------------------------ */ /* flag the row as assembled from the Lson */ Rows [f] = EMPTY ; nrows = ep->nrows ; ncols = ep->ncols ; p += UNITS (Int, ncols + nrows) ; S = ((Entry *) p) + f ; DEBUG6 (("Old LSON: "ID"\n", e)) ; #ifndef NDEBUG UMF_dump_element (Numeric, Work, e, FALSE) ; #endif ncolsleft = ep->ncolsleft ; Frow = Fcblock + Frpos [row] ; DEBUG6 (("LSON found (in scan2-row): "ID"\n", e)) ; Row_degree [row] -= ncolsleft ; if (ncols == ncolsleft) { /* -------------------------------------------------- */ /* no columns assembled out this Lson yet */ /* -------------------------------------------------- */ #pragma ivdep for (j = 0 ; j < ncols ; j++) { col = Cols [j] ; ASSERT (col >= 0 && col < Work->n_col) ; #ifndef FIXQ Col_degree [col] -- ; #endif /* Frow [Fcpos [col]] += *S ; */ ASSEMBLE (Frow [Fcpos [col]], *S) ; S += nrows ; } } else { /* -------------------------------------------------- */ /* some columns have been assembled out of this Lson */ /* -------------------------------------------------- */ #pragma ivdep for (j = 0 ; j < ncols ; j++) { col = Cols [j] ; if (col >= 0) { ASSERT (col < Work->n_col) ; #ifndef FIXQ Col_degree [col] -- ; #endif /* Frow [Fcpos [col]] += *S ; */ ASSEMBLE (Frow [Fcpos [col]], *S) ; } S += nrows ; } } ep->nrowsleft-- ; ASSERT (ep->nrowsleft > 0) ; } else { *tp2++ = *tp ; /* leave the tuple in the list */ } } Row_tlen [row] = tp2 - tp1 ; #ifndef NDEBUG DEBUG7 (("row assembled in scan2-row: "ID"\n", row)) ; UMF_dump_rowcol (0, Numeric, Work, row, FALSE) ; DEBUG7 (("Current frontal matrix: (scan 1b)\n")) ; UMF_dump_current_front (Numeric, Work, TRUE) ; #endif }
PRIVATE void col_assemble ( Int col, NumericType *Numeric, WorkType *Work ) { Entry *S, *Fcblock, *Fcol ; Int tpi, e, *E, *Fcpos, *Frpos, *Row_degree, *Col_tuples, *Col_tlen, cdeg0, f, nrows, ncols, *Rows, *Cols, row, nrowsleft, i ; Tuple *tp, *tp1, *tp2, *tpend ; Unit *Memory, *p ; Element *ep ; #if !defined (FIXQ) || !defined (NDEBUG) Int *Col_degree ; Col_degree = Numeric->Cperm ; #endif Col_tuples = Numeric->Lip ; tpi = Col_tuples [col] ; if (!tpi) return ; Memory = Numeric->Memory ; E = Work->E ; Fcpos = Work->Fcpos ; Frpos = Work->Frpos ; Row_degree = Numeric->Rperm ; Col_tlen = Numeric->Lilen ; E = Work->E ; Memory = Numeric->Memory ; cdeg0 = Work->cdeg0 ; Fcblock = Work->Fcblock ; DEBUG6 (("SCAN2-col: "ID"\n", col)) ; #ifndef NDEBUG UMF_dump_rowcol (1, Numeric, Work, col, FALSE) ; #endif ASSERT (NON_PIVOTAL_COL (col)) ; tp = (Tuple *) (Memory + tpi) ; tp1 = tp ; tp2 = tp ; tpend = tp + Col_tlen [col] ; for ( ; tp < tpend ; tp++) { e = tp->e ; ASSERT (e > 0 && e <= Work->nel) ; if (!E [e]) continue ; /* element already deallocated */ f = tp->f ; p = Memory + E [e] ; ep = (Element *) p ; p += UNITS (Element, 1) ; Cols = (Int *) p ; if (Cols [f] == EMPTY) continue ; /* col already assembled */ ASSERT (col == Cols [f] && col >= 0 && col < Work->n_col) ; if (ep->cdeg == cdeg0) { /* ------------------------------------------------------ */ /* this is an old Uson - assemble just one col */ /* ------------------------------------------------------ */ /* flag the col as assembled from the Uson */ Cols [f] = EMPTY ; nrows = ep->nrows ; ncols = ep->ncols ; Rows = Cols + ncols ; p += UNITS (Int, ncols + nrows) ; S = ((Entry *) p) + f * nrows ; DEBUG6 (("Old USON: "ID"\n", e)) ; #ifndef NDEBUG UMF_dump_element (Numeric, Work, e, FALSE) ; #endif nrowsleft = ep->nrowsleft ; Fcol = Fcblock + Fcpos [col] ; DEBUG6 (("USON found (in scan2-col): "ID"\n", e)) ; #ifndef FIXQ Col_degree [col] -= nrowsleft ; #endif if (nrows == nrowsleft) { /* -------------------------------------------------- */ /* no rows assembled out of this Uson yet */ /* -------------------------------------------------- */ #pragma ivdep for (i = 0 ; i < nrows ; i++) { row = Rows [i] ; ASSERT (row >= 0 && row < Work->n_row) ; Row_degree [row]-- ; /* Fcol [Frpos [row]] += S [i] ; */ ASSEMBLE (Fcol [Frpos [row]], S [i]) ; } } else { /* -------------------------------------------------- */ /* some rows have been assembled out of this Uson */ /* -------------------------------------------------- */ #pragma ivdep for (i = 0 ; i < nrows ; i++) { row = Rows [i] ; if (row >= 0) { ASSERT (row < Work->n_row) ; Row_degree [row]-- ; /* Fcol [Frpos [row]] += S [i] ; */ ASSEMBLE (Fcol [Frpos [row]], S [i]) ; } } } ep->ncolsleft-- ; ASSERT (ep->ncolsleft > 0) ; } else { *tp2++ = *tp ; /* leave the tuple in the list */ } } Col_tlen [col] = tp2 - tp1 ; #ifndef NDEBUG DEBUG7 (("Column assembled in scan2-col: "ID"\n", col)) ; UMF_dump_rowcol (1, Numeric, Work, col, FALSE) ; DEBUG7 (("Current frontal matrix: after scan2-col\n")) ; UMF_dump_current_front (Numeric, Work, TRUE) ; #endif }
GLOBAL Int UMF_row_search ( NumericType *Numeric, WorkType *Work, SymbolicType *Symbolic, Int cdeg0, /* length of column in Front */ Int cdeg1, /* length of column outside Front */ const Int Pattern [ ], /* pattern of column, Pattern [0..cdeg1 -1] */ const Int Pos [ ], /* Pos [Pattern [0..cdeg1 -1]] = 0..cdeg1 -1 */ Int pivrow [2], /* pivrow [IN] and pivrow [OUT] */ Int rdeg [2], /* rdeg [IN] and rdeg [OUT] */ Int W_i [ ], /* pattern of pivrow [IN], */ /* either Fcols or Woi */ Int W_o [ ], /* pattern of pivrow [OUT], */ /* either Wio or Woo */ Int prior_pivrow [2], /* the two other rows just scanned, if any */ const Entry Wxy [ ], /* numerical values Wxy [0..cdeg1-1], either Wx or Wy */ Int pivcol, /* the candidate column being searched */ Int freebie [ ] ) { /* ---------------------------------------------------------------------- */ /* local variables */ /* ---------------------------------------------------------------------- */ double maxval, toler, toler2, value, pivot [2] ; Int i, row, deg, col, *Frpos, fnrows, *E, j, ncols, *Cols, *Rows, e, f, Wrpflag, *Fcpos, fncols, tpi, max_rdeg, nans_in_col, was_offdiag, diag_row, prefer_diagonal, *Wrp, found, *Diagonal_map ; Tuple *tp, *tpend, *tp1, *tp2 ; Unit *Memory, *p ; Element *ep ; Int *Row_tuples, *Row_degree, *Row_tlen ; #ifndef NDEBUG Int *Col_degree ; DEBUG2 (("Row_search:\n")) ; for (i = 0 ; i < cdeg1 ; i++) { row = Pattern [i] ; DEBUG4 ((" row: "ID"\n", row)) ; ASSERT (row >= 0 && row < Numeric->n_row) ; ASSERT (i == Pos [row]) ; } /* If row is not in Pattern [0..cdeg1-1], then Pos [row] == EMPTY */ if (UMF_debug > 0 || Numeric->n_row < 1000) { Int cnt = cdeg1 ; DEBUG4 (("Scan all rows:\n")) ; for (row = 0 ; row < Numeric->n_row ; row++) { if (Pos [row] < 0) { cnt++ ; } else { DEBUG4 ((" row: "ID" pos "ID"\n", row, Pos [row])) ; } } ASSERT (cnt == Numeric->n_row) ; } Col_degree = Numeric->Cperm ; /* for NON_PIVOTAL_COL macro only */ ASSERT (pivcol >= 0 && pivcol < Work->n_col) ; ASSERT (NON_PIVOTAL_COL (pivcol)) ; #endif pivot [IN] = 0. ; pivot [OUT] = 0. ; /* ---------------------------------------------------------------------- */ /* get parameters */ /* ---------------------------------------------------------------------- */ Row_degree = Numeric->Rperm ; Row_tuples = Numeric->Uip ; Row_tlen = Numeric->Uilen ; Wrp = Work->Wrp ; Frpos = Work->Frpos ; E = Work->E ; Memory = Numeric->Memory ; fnrows = Work->fnrows ; prefer_diagonal = Symbolic->prefer_diagonal ; Diagonal_map = Work->Diagonal_map ; if (Diagonal_map) { diag_row = Diagonal_map [pivcol] ; was_offdiag = diag_row < 0 ; if (was_offdiag) { /* the "diagonal" entry in this column was permuted here by an * earlier pivot choice. The tighter off-diagonal tolerance will * be used instead of the symmetric tolerance. */ diag_row = FLIP (diag_row) ; } ASSERT (diag_row >= 0 && diag_row < Numeric->n_row) ; } else { diag_row = EMPTY ; /* unused */ was_offdiag = EMPTY ; /* unused */ } /* pivot row degree cannot exceed max_rdeg */ max_rdeg = Work->fncols_max ; /* ---------------------------------------------------------------------- */ /* scan pivot column for candidate rows */ /* ---------------------------------------------------------------------- */ maxval = 0.0 ; nans_in_col = FALSE ; for (i = 0 ; i < cdeg1 ; i++) { APPROX_ABS (value, Wxy [i]) ; if (SCALAR_IS_NAN (value)) { nans_in_col = TRUE ; maxval = value ; break ; } /* This test can now ignore the NaN case: */ maxval = MAX (maxval, value) ; } /* if maxval is zero, the matrix is numerically singular */ toler = Numeric->relpt * maxval ; toler2 = Numeric->relpt2 * maxval ; toler2 = was_offdiag ? toler : toler2 ; DEBUG5 (("Row_search begins [ maxval %g toler %g %g\n", maxval, toler, toler2)) ; if (SCALAR_IS_NAN (toler) || SCALAR_IS_NAN (toler2)) { nans_in_col = TRUE ; } if (!nans_in_col) { /* look for the diagonal entry, if it exists */ found = FALSE ; ASSERT (!SCALAR_IS_NAN (toler)) ; if (prefer_diagonal) { ASSERT (diag_row != EMPTY) ; i = Pos [diag_row] ; if (i >= 0) { double a ; ASSERT (i < cdeg1) ; ASSERT (diag_row == Pattern [i]) ; APPROX_ABS (a, Wxy [i]) ; ASSERT (!SCALAR_IS_NAN (a)) ; ASSERT (!SCALAR_IS_NAN (toler2)) ; if (SCALAR_IS_NONZERO (a) && a >= toler2) { /* found it! */ DEBUG3 (("Symmetric pivot: "ID" "ID"\n", pivcol, diag_row)); found = TRUE ; if (Frpos [diag_row] >= 0 && Frpos [diag_row] < fnrows) { pivrow [IN] = diag_row ; pivrow [OUT] = EMPTY ; } else { pivrow [IN] = EMPTY ; pivrow [OUT] = diag_row ; } } } } /* either no diagonal found, or we didn't look for it */ if (!found) { if (cdeg0 > 0) { /* this is a column in the front */ for (i = 0 ; i < cdeg0 ; i++) { double a ; APPROX_ABS (a, Wxy [i]) ; ASSERT (!SCALAR_IS_NAN (a)) ; ASSERT (!SCALAR_IS_NAN (toler)) ; if (SCALAR_IS_NONZERO (a) && a >= toler) { row = Pattern [i] ; deg = Row_degree [row] ; #ifndef NDEBUG DEBUG6 ((ID" candidate row "ID" deg "ID" absval %g\n", i, row, deg, a)) ; UMF_dump_rowcol (0, Numeric, Work, row, TRUE) ; #endif ASSERT (Frpos [row] >= 0 && Frpos [row] < fnrows) ; ASSERT (Frpos [row] == i) ; /* row is in the current front */ DEBUG4 ((" in front\n")) ; if (deg < rdeg [IN] /* break ties by picking the largest entry: */ || (deg == rdeg [IN] && a > pivot [IN]) /* break ties by picking the diagonal entry: */ /* || (deg == rdeg [IN] && row == diag_row) */ ) { /* best row in front, so far */ pivrow [IN] = row ; rdeg [IN] = deg ; pivot [IN] = a ; } } } for ( ; i < cdeg1 ; i++) { double a ; APPROX_ABS (a, Wxy [i]) ; ASSERT (!SCALAR_IS_NAN (a)) ; ASSERT (!SCALAR_IS_NAN (toler)) ; if (SCALAR_IS_NONZERO (a) && a >= toler) { row = Pattern [i] ; deg = Row_degree [row] ; #ifndef NDEBUG DEBUG6 ((ID" candidate row "ID" deg "ID" absval %g\n", i, row, deg, a)) ; UMF_dump_rowcol (0, Numeric, Work, row, TRUE) ; #endif ASSERT (Frpos [row] == i) ; /* row is not in the current front */ DEBUG4 ((" NOT in front\n")) ; if (deg < rdeg [OUT] /* break ties by picking the largest entry: */ || (deg == rdeg [OUT] && a > pivot [OUT]) /* break ties by picking the diagonal entry: */ /* || (deg == rdeg [OUT] && row == diag_row) */ ) { /* best row not in front, so far */ pivrow [OUT] = row ; rdeg [OUT] = deg ; pivot [OUT] = a ; } } } } else { /* this column is not in the front */ for (i = 0 ; i < cdeg1 ; i++) { double a ; APPROX_ABS (a, Wxy [i]) ; ASSERT (!SCALAR_IS_NAN (a)) ; ASSERT (!SCALAR_IS_NAN (toler)) ; if (SCALAR_IS_NONZERO (a) && a >= toler) { row = Pattern [i] ; deg = Row_degree [row] ; #ifndef NDEBUG DEBUG6 ((ID" candidate row "ID" deg "ID" absval %g\n", i, row, deg, a)) ; UMF_dump_rowcol (0, Numeric, Work, row, TRUE) ; #endif if (Frpos [row] >= 0 && Frpos [row] < fnrows) { /* row is in the current front */ DEBUG4 ((" in front\n")) ; if (deg < rdeg [IN] /* break ties by picking the largest entry: */ || (deg == rdeg [IN] && a > pivot [IN]) /* break ties by picking the diagonal entry: */ /* || (deg == rdeg [IN] && row == diag_row) */ ) { /* best row in front, so far */ pivrow [IN] = row ; rdeg [IN] = deg ; pivot [IN] = a ; } } else { /* row is not in the current front */ DEBUG4 ((" NOT in front\n")) ; if (deg < rdeg [OUT] /* break ties by picking the largest entry: */ || (deg == rdeg[OUT] && a > pivot [OUT]) /* break ties by picking the diagonal entry: */ /* || (deg == rdeg[OUT] && row == diag_row) */ ) { /* best row not in front, so far */ pivrow [OUT] = row ; rdeg [OUT] = deg ; pivot [OUT] = a ; } } } } } } } /* ---------------------------------------------------------------------- */ /* NaN handling */ /* ---------------------------------------------------------------------- */ /* if cdeg1 > 0 then we must have found a pivot row ... unless NaN's */ /* exist. Try with no numerical tests if no pivot found. */ if (cdeg1 > 0 && pivrow [IN] == EMPTY && pivrow [OUT] == EMPTY) { /* cleanup for the NaN case */ DEBUG0 (("Found a NaN in pivot column!\n")) ; /* grab the first entry in the pivot column, ignoring degree, */ /* numerical stability, and symmetric preference */ row = Pattern [0] ; deg = Row_degree [row] ; if (Frpos [row] >= 0 && Frpos [row] < fnrows) { /* row is in the current front */ DEBUG4 ((" in front\n")) ; pivrow [IN] = row ; rdeg [IN] = deg ; } else { /* row is not in the current front */ DEBUG4 ((" NOT in front\n")) ; pivrow [OUT] = row ; rdeg [OUT] = deg ; } /* We are now guaranteed to have a pivot, no matter how broken */ /* (non-IEEE compliant) the underlying numerical operators are. */ /* This is particularly a problem for Microsoft compilers (they do */ /* not handle NaN's properly). Now try to find a sparser pivot, if */ /* possible. */ for (i = 1 ; i < cdeg1 ; i++) { row = Pattern [i] ; deg = Row_degree [row] ; if (Frpos [row] >= 0 && Frpos [row] < fnrows) { /* row is in the current front */ DEBUG4 ((" in front\n")) ; if (deg < rdeg [IN] || (deg == rdeg [IN] && row == diag_row)) { /* best row in front, so far */ pivrow [IN] = row ; rdeg [IN] = deg ; } } else { /* row is not in the current front */ DEBUG4 ((" NOT in front\n")) ; if (deg < rdeg [OUT] || (deg == rdeg [OUT] && row == diag_row)) { /* best row not in front, so far */ pivrow [OUT] = row ; rdeg [OUT] = deg ; } } } } /* We found a pivot if there are entries (even zero ones) in pivot col */ ASSERT (IMPLIES (cdeg1 > 0, pivrow[IN] != EMPTY || pivrow[OUT] != EMPTY)) ; /* If there are no entries in the pivot column, then no pivot is found */ ASSERT (IMPLIES (cdeg1 == 0, pivrow[IN] == EMPTY && pivrow[OUT] == EMPTY)) ; /* ---------------------------------------------------------------------- */ /* check for singular matrix */ /* ---------------------------------------------------------------------- */ if (cdeg1 == 0) { if (fnrows > 0) { /* Get the pivrow [OUT][IN] from the current front. The frontal matrix looks like this: pivcol[OUT] | v x x x x 0 <- so grab this row as the pivrow [OUT][IN]. x x x x 0 x x x x 0 0 0 0 0 0 The current frontal matrix has some rows in it. The degree of the pivcol[OUT] is zero. The column is empty, and the current front does not contribute to it. */ pivrow [IN] = Work->Frows [0] ; DEBUGm4 (("Got zero pivrow[OUT][IN] "ID" from current front\n", pivrow [IN])) ; } else { /* Get a pivot row from the row-merge tree, use as pivrow [OUT][OUT]. pivrow [IN] remains EMPTY. This can only happen if the current front is 0-by-0. */ Int *Front_leftmostdesc, *Front_1strow, *Front_new1strow, row1, row2, fleftmost, nfr, n_row, frontid ; ASSERT (Work->fncols == 0) ; Front_leftmostdesc = Symbolic->Front_leftmostdesc ; Front_1strow = Symbolic->Front_1strow ; Front_new1strow = Work->Front_new1strow ; nfr = Symbolic->nfr ; n_row = Numeric->n_row ; frontid = Work->frontid ; DEBUGm4 (("Note: pivcol: "ID" is empty front "ID"\n", pivcol, frontid)) ; #ifndef NDEBUG DEBUG1 (("Calling dump rowmerge\n")) ; UMF_dump_rowmerge (Numeric, Symbolic, Work) ; #endif /* Row-merge set is the non-pivotal rows in the range */ /* Front_new1strow [Front_leftmostdesc [frontid]] to */ /* Front_1strow [frontid+1] - 1. */ /* If this is empty, then use the empty rows, in the range */ /* Front_new1strow [nfr] to n_row-1. */ /* If this too is empty, then pivrow [OUT] will be empty. */ /* In both cases, update Front_new1strow [...]. */ fleftmost = Front_leftmostdesc [frontid] ; row1 = Front_new1strow [fleftmost] ; row2 = Front_1strow [frontid+1] - 1 ; DEBUG1 (("Leftmost: "ID" Rows ["ID" to "ID"] srch ["ID" to "ID"]\n", fleftmost, Front_1strow [frontid], row2, row1, row2)) ; /* look in the range row1 ... row2 */ for (row = row1 ; row <= row2 ; row++) { DEBUG3 ((" Row: "ID"\n", row)) ; if (NON_PIVOTAL_ROW (row)) { /* found it */ DEBUG3 ((" Row: "ID" found\n", row)) ; ASSERT (Frpos [row] == EMPTY) ; pivrow [OUT] = row ; DEBUGm4 (("got row merge pivrow %d\n", pivrow [OUT])) ; break ; } } Front_new1strow [fleftmost] = row ; if (pivrow [OUT] == EMPTY) { /* not found, look in empty row set in "dummy" front */ row1 = Front_new1strow [nfr] ; row2 = n_row-1 ; DEBUG3 (("Empty: "ID" Rows ["ID" to "ID"] srch["ID" to "ID"]\n", nfr, Front_1strow [nfr], row2, row1, row2)) ; /* look in the range row1 ... row2 */ for (row = row1 ; row <= row2 ; row++) { DEBUG3 ((" Empty Row: "ID"\n", row)) ; if (NON_PIVOTAL_ROW (row)) { /* found it */ DEBUG3 ((" Empty Row: "ID" found\n", row)) ; ASSERT (Frpos [row] == EMPTY) ; pivrow [OUT] = row ; DEBUGm4 (("got dummy row pivrow %d\n", pivrow [OUT])) ; break ; } } Front_new1strow [nfr] = row ; } if (pivrow [OUT] == EMPTY) { /* Row-merge set is empty. We can just discard */ /* the candidate pivot column. */ DEBUG0 (("Note: row-merge set empty\n")) ; DEBUGm4 (("got no pivrow \n")) ; return (UMFPACK_WARNING_singular_matrix) ; } } } /* ---------------------------------------------------------------------- */ /* construct the candidate row in the front, if any */ /* ---------------------------------------------------------------------- */ #ifndef NDEBUG /* check Wrp */ ASSERT (Work->Wrpflag > 0) ; if (UMF_debug > 0 || Work->n_col < 1000) { for (i = 0 ; i < Work->n_col ; i++) { ASSERT (Wrp [i] < Work->Wrpflag) ; } } #endif #ifndef NDEBUG DEBUG4 (("pivrow [IN]: "ID"\n", pivrow [IN])) ; UMF_dump_rowcol (0, Numeric, Work, pivrow [IN], TRUE) ; #endif if (pivrow [IN] != EMPTY) { /* the row merge candidate row is not pivrow [IN] */ freebie [IN] = (pivrow [IN] == prior_pivrow [IN]) && (cdeg1 > 0) ; ASSERT (cdeg1 >= 0) ; if (!freebie [IN]) { /* include current front in the degree of this row */ Fcpos = Work->Fcpos ; fncols = Work->fncols ; Wrpflag = Work->Wrpflag ; /* -------------------------------------------------------------- */ /* construct the pattern of the IN row */ /* -------------------------------------------------------------- */ #ifndef NDEBUG /* check Fcols */ DEBUG5 (("ROW ASSEMBLE: rdeg "ID"\nREDUCE ROW "ID"\n", fncols, pivrow [IN])) ; for (j = 0 ; j < fncols ; j++) { col = Work->Fcols [j] ; ASSERT (col >= 0 && col < Work->n_col) ; ASSERT (Fcpos [col] >= 0) ; } if (UMF_debug > 0 || Work->n_col < 1000) { Int cnt = fncols ; for (col = 0 ; col < Work->n_col ; col++) { if (Fcpos [col] < 0) cnt++ ; } ASSERT (cnt == Work->n_col) ; } #endif rdeg [IN] = fncols ; ASSERT (pivrow [IN] >= 0 && pivrow [IN] < Work->n_row) ; ASSERT (NON_PIVOTAL_ROW (pivrow [IN])) ; /* add the pivot column itself */ ASSERT (Wrp [pivcol] != Wrpflag) ; if (Fcpos [pivcol] < 0) { DEBUG3 (("Adding pivot col to pivrow [IN] pattern\n")) ; if (rdeg [IN] >= max_rdeg) { /* :: pattern change (in) :: */ return (UMFPACK_ERROR_different_pattern) ; } Wrp [pivcol] = Wrpflag ; W_i [rdeg [IN]++] = pivcol ; } tpi = Row_tuples [pivrow [IN]] ; if (tpi) { tp = (Tuple *) (Memory + tpi) ; tp1 = tp ; tp2 = tp ; tpend = tp + Row_tlen [pivrow [IN]] ; for ( ; tp < tpend ; tp++) { e = tp->e ; ASSERT (e > 0 && e <= Work->nel) ; if (!E [e]) { continue ; /* element already deallocated */ } f = tp->f ; p = Memory + E [e] ; ep = (Element *) p ; p += UNITS (Element, 1) ; Cols = (Int *) p ; ncols = ep->ncols ; Rows = Cols + ncols ; if (Rows [f] == EMPTY) { continue ; /* row already assembled */ } ASSERT (pivrow [IN] == Rows [f]) ; for (j = 0 ; j < ncols ; j++) { col = Cols [j] ; ASSERT (col >= EMPTY && col < Work->n_col) ; if ((col >= 0) && (Wrp [col] != Wrpflag) && Fcpos [col] <0) { ASSERT (NON_PIVOTAL_COL (col)) ; if (rdeg [IN] >= max_rdeg) { /* :: pattern change (rdeg in failure) :: */ DEBUGm4 (("rdeg [IN] >= max_rdeg failure\n")) ; return (UMFPACK_ERROR_different_pattern) ; } Wrp [col] = Wrpflag ; W_i [rdeg [IN]++] = col ; } } *tp2++ = *tp ; /* leave the tuple in the list */ } Row_tlen [pivrow [IN]] = tp2 - tp1 ; } #ifndef NDEBUG DEBUG4 (("Reduced IN row:\n")) ; for (j = 0 ; j < fncols ; j++) { DEBUG6 ((" "ID" "ID" "ID"\n", j, Work->Fcols [j], Fcpos [Work->Fcols [j]])) ; ASSERT (Fcpos [Work->Fcols [j]] >= 0) ; } for (j = fncols ; j < rdeg [IN] ; j++) { DEBUG6 ((" "ID" "ID" "ID"\n", j, W_i [j], Wrp [W_i [j]])); ASSERT (W_i [j] >= 0 && W_i [j] < Work->n_col) ; ASSERT (Wrp [W_i [j]] == Wrpflag) ; } /* mark the end of the pattern in case we scan it by mistake */ /* Note that this means W_i must be of size >= fncols_max + 1 */ W_i [rdeg [IN]] = EMPTY ; #endif /* rdeg [IN] is now the exact degree of the IN row */ /* clear Work->Wrp. */ Work->Wrpflag++ ; /* All Wrp [0..n_col] is now < Wrpflag */ } } #ifndef NDEBUG /* check Wrp */ if (UMF_debug > 0 || Work->n_col < 1000) { for (i = 0 ; i < Work->n_col ; i++) { ASSERT (Wrp [i] < Work->Wrpflag) ; } } #endif /* ---------------------------------------------------------------------- */ /* construct the candidate row not in the front, if any */ /* ---------------------------------------------------------------------- */ #ifndef NDEBUG DEBUG4 (("pivrow [OUT]: "ID"\n", pivrow [OUT])) ; UMF_dump_rowcol (0, Numeric, Work, pivrow [OUT], TRUE) ; #endif /* If this is a candidate row from the row merge set, force it to be */ /* scanned (ignore prior_pivrow [OUT]). */ if (pivrow [OUT] != EMPTY) { freebie [OUT] = (pivrow [OUT] == prior_pivrow [OUT]) && cdeg1 > 0 ; ASSERT (cdeg1 >= 0) ; if (!freebie [OUT]) { Wrpflag = Work->Wrpflag ; /* -------------------------------------------------------------- */ /* construct the pattern of the row */ /* -------------------------------------------------------------- */ rdeg [OUT] = 0 ; ASSERT (pivrow [OUT] >= 0 && pivrow [OUT] < Work->n_row) ; ASSERT (NON_PIVOTAL_ROW (pivrow [OUT])) ; /* add the pivot column itself */ ASSERT (Wrp [pivcol] != Wrpflag) ; DEBUG3 (("Adding pivot col to pivrow [OUT] pattern\n")) ; if (rdeg [OUT] >= max_rdeg) { /* :: pattern change (out) :: */ return (UMFPACK_ERROR_different_pattern) ; } Wrp [pivcol] = Wrpflag ; W_o [rdeg [OUT]++] = pivcol ; tpi = Row_tuples [pivrow [OUT]] ; if (tpi) { tp = (Tuple *) (Memory + tpi) ; tp1 = tp ; tp2 = tp ; tpend = tp + Row_tlen [pivrow [OUT]] ; for ( ; tp < tpend ; tp++) { e = tp->e ; ASSERT (e > 0 && e <= Work->nel) ; if (!E [e]) { continue ; /* element already deallocated */ } f = tp->f ; p = Memory + E [e] ; ep = (Element *) p ; p += UNITS (Element, 1) ; Cols = (Int *) p ; ncols = ep->ncols ; Rows = Cols + ncols ; if (Rows [f] == EMPTY) { continue ; /* row already assembled */ } ASSERT (pivrow [OUT] == Rows [f]) ; for (j = 0 ; j < ncols ; j++) { col = Cols [j] ; ASSERT (col >= EMPTY && col < Work->n_col) ; if ((col >= 0) && (Wrp [col] != Wrpflag)) { ASSERT (NON_PIVOTAL_COL (col)) ; if (rdeg [OUT] >= max_rdeg) { /* :: pattern change (rdeg out failure) :: */ DEBUGm4 (("rdeg [OUT] failure\n")) ; return (UMFPACK_ERROR_different_pattern) ; } Wrp [col] = Wrpflag ; W_o [rdeg [OUT]++] = col ; } } *tp2++ = *tp ; /* leave the tuple in the list */ } Row_tlen [pivrow [OUT]] = tp2 - tp1 ; } #ifndef NDEBUG DEBUG4 (("Reduced row OUT:\n")) ; for (j = 0 ; j < rdeg [OUT] ; j++) { DEBUG6 ((" "ID" "ID" "ID"\n", j, W_o [j], Wrp [W_o [j]])) ; ASSERT (W_o [j] >= 0 && W_o [j] < Work->n_col) ; ASSERT (Wrp [W_o [j]] == Wrpflag) ; } /* mark the end of the pattern in case we scan it by mistake */ /* Note that this means W_o must be of size >= fncols_max + 1 */ W_o [rdeg [OUT]] = EMPTY ; #endif /* rdeg [OUT] is now the exact degree of the row */ /* clear Work->Wrp. */ Work->Wrpflag++ ; /* All Wrp [0..n] is now < Wrpflag */ } } DEBUG5 (("Row_search end ] \n")) ; #ifndef NDEBUG /* check Wrp */ if (UMF_debug > 0 || Work->n_col < 1000) { for (i = 0 ; i < Work->n_col ; i++) { ASSERT (Wrp [i] < Work->Wrpflag) ; } } #endif return (UMFPACK_OK) ; }
GLOBAL Int UMF_start_front /* returns TRUE if successful, FALSE otherwise */ ( Int chain, NumericType *Numeric, WorkType *Work, SymbolicType *Symbolic ) { Int fnrows_max, fncols_max, fnr2, fnc2, fsize, fcurr_size, maxfrsize, overflow, nb, f, cdeg ; double maxbytes ; nb = Symbolic->nb ; fnrows_max = Symbolic->Chain_maxrows [chain] ; fncols_max = Symbolic->Chain_maxcols [chain] ; DEBUGm2 (("Start Front for chain "ID". fnrows_max "ID" fncols_max "ID"\n", chain, fnrows_max, fncols_max)) ; Work->fnrows_max = fnrows_max ; Work->fncols_max = fncols_max ; Work->any_skip = FALSE ; maxbytes = sizeof (Entry) * (double) (fnrows_max + nb) * (double) (fncols_max + nb) ; fcurr_size = Work->fcurr_size ; if (Symbolic->prefer_diagonal) { /* Get a rough upper bound on the degree of the first pivot column in * this front. Note that Col_degree is not maintained if diagonal * pivoting is preferred. For most matrices, the first pivot column * of the first frontal matrix of a new chain has only one tuple in * it anyway, so this bound is exact in that case. */ Int col, tpi, e, *E, *Col_tuples, *Col_tlen, *Cols ; Tuple *tp, *tpend ; Unit *Memory, *p ; Element *ep ; E = Work->E ; Memory = Numeric->Memory ; Col_tuples = Numeric->Lip ; Col_tlen = Numeric->Lilen ; col = Work->nextcand ; tpi = Col_tuples [col] ; tp = (Tuple *) Memory + tpi ; tpend = tp + Col_tlen [col] ; cdeg = 0 ; DEBUGm3 (("\n=============== start front: col "ID" tlen "ID"\n", col, Col_tlen [col])) ; for ( ; tp < tpend ; tp++) { DEBUG1 (("Tuple ("ID","ID")\n", tp->e, tp->f)) ; e = tp->e ; if (!E [e]) continue ; f = tp->f ; p = Memory + E [e] ; ep = (Element *) p ; p += UNITS (Element, 1) ; Cols = (Int *) p ; if (Cols [f] == EMPTY) continue ; DEBUG1 ((" nrowsleft "ID"\n", ep->nrowsleft)) ; cdeg += ep->nrowsleft ; } #ifndef NDEBUG DEBUGm3 (("start front cdeg: "ID" col "ID"\n", cdeg, col)) ; UMF_dump_rowcol (1, Numeric, Work, col, FALSE) ; #endif /* cdeg is now the rough upper bound on the degree of the next pivot * column. */ /* If AMD was called, we know the maximum number of nonzeros in any * column of L. Use this as an upper bound for cdeg, but add 2 to * account for a small amount of off-diagonal pivoting. */ if (Symbolic->amd_dmax > 0) { cdeg = MIN (cdeg, Symbolic->amd_dmax) ; } /* Increase it to account for larger columns later on. * Also ensure that it's larger than zero. */ cdeg += 2 ; /* cdeg cannot be larger than fnrows_max */ cdeg = MIN (cdeg, fnrows_max) ; } else { /* don't do the above cdeg computation */ cdeg = 0 ; } DEBUGm2 (("fnrows max "ID" fncols_max "ID"\n", fnrows_max, fncols_max)) ; /* the current frontal matrix is empty */ ASSERT (Work->fnrows == 0 && Work->fncols == 0 && Work->fnpiv == 0) ; /* maximum row dimension is always odd, to avoid bad cache effects */ ASSERT (fnrows_max >= 0) ; ASSERT (fnrows_max % 2 == 1) ; /* ---------------------------------------------------------------------- * allocate working array for current frontal matrix: * minimum size: 1-by-1 * maximum size: fnrows_max-by-fncols_max * desired size: * * if Numeric->front_alloc_init >= 0: * * for unsymmetric matrices: * Numeric->front_alloc_init * (fnrows_max-by-fncols_max) * * for symmetric matrices (diagonal pivoting preference, actually): * Numeric->front_alloc_init * (fnrows_max-by-fncols_max), or * cdeg*cdeg, whichever is smaller. * * if Numeric->front_alloc_init < 0: * allocate a front of size -Numeric->front_alloc_init. * * Allocate the whole thing if it's small (less than 2*nb^2). Make sure the * leading dimension of the frontal matrix is odd. * * Also allocate the nb-by-nb LU block, the dr-by-nb L block, and the * nb-by-dc U block. * ---------------------------------------------------------------------- */ /* get the maximum front size, avoiding integer overflow */ overflow = INT_OVERFLOW (maxbytes) ; if (overflow) { /* :: int overflow, max front size :: */ maxfrsize = Int_MAX / sizeof (Entry) ; } else { maxfrsize = (fnrows_max + nb) * (fncols_max + nb) ; } ASSERT (!INT_OVERFLOW ((double) maxfrsize * sizeof (Entry))) ; if (Numeric->front_alloc_init < 0) { /* allocate a front of -Numeric->front_alloc_init entries */ fsize = -Numeric->front_alloc_init ; fsize = MAX (1, fsize) ; } else { if (INT_OVERFLOW (Numeric->front_alloc_init * maxbytes)) { /* :: int overflow, requested front size :: */ fsize = Int_MAX / sizeof (Entry) ; } else { fsize = Numeric->front_alloc_init * maxfrsize ; } if (cdeg > 0) { /* diagonal pivoting is in use. cdeg was computed above */ Int fsize2 ; /* add the L and U blocks */ cdeg += nb ; if (INT_OVERFLOW (((double) cdeg * (double) cdeg) * sizeof (Entry))) { /* :: int overflow, symmetric front size :: */ fsize2 = Int_MAX / sizeof (Entry) ; } else { fsize2 = MAX (cdeg * cdeg, fcurr_size) ; } fsize = MIN (fsize, fsize2) ; } } fsize = MAX (fsize, 2*nb*nb) ; /* fsize and maxfrsize are now safe from integer overflow. They both * include the size of the pivot blocks. */ ASSERT (!INT_OVERFLOW ((double) fsize * sizeof (Entry))) ; Work->fnrows_new = 0 ; Work->fncols_new = 0 ; /* desired size is fnr2-by-fnc2 (includes L and U blocks): */ DEBUGm2 ((" fsize "ID" fcurr_size "ID"\n", fsize, fcurr_size)) ; DEBUGm2 ((" maxfrsize "ID" fnr_curr "ID" fnc_curr "ID"\n", maxfrsize, Work->fnr_curr, Work->fnc_curr)) ; if (fsize >= maxfrsize && !overflow) { /* max working array is small, allocate all of it */ fnr2 = fnrows_max + nb ; fnc2 = fncols_max + nb ; fsize = maxfrsize ; DEBUGm1 ((" sufficient for ("ID"+"ID")-by-("ID"+"ID")\n", fnrows_max, nb, fncols_max, nb)) ; } else { /* allocate a smaller working array */ if (fnrows_max <= fncols_max) { fnr2 = (Int) sqrt ((double) fsize) ; /* make sure fnr2 is odd */ fnr2 = MAX (fnr2, 1) ; if (fnr2 % 2 == 0) fnr2++ ; fnr2 = MIN (fnr2, fnrows_max + nb) ; fnc2 = fsize / fnr2 ; } else { fnc2 = (Int) sqrt ((double) fsize) ; fnc2 = MIN (fnc2, fncols_max + nb) ; fnr2 = fsize / fnc2 ; /* make sure fnr2 is odd */ fnr2 = MAX (fnr2, 1) ; if (fnr2 % 2 == 0) { fnr2++ ; fnc2 = fsize / fnr2 ; } } DEBUGm1 ((" smaller "ID"-by-"ID"\n", fnr2, fnc2)) ; } fnr2 = MIN (fnr2, fnrows_max + nb) ; fnc2 = MIN (fnc2, fncols_max + nb) ; ASSERT (fnr2 % 2 == 1) ; ASSERT (fnr2 * fnc2 <= fsize) ; fnr2 -= nb ; fnc2 -= nb ; ASSERT (fnr2 >= 0) ; ASSERT (fnc2 >= 0) ; if (fsize > fcurr_size) { DEBUGm1 ((" Grow front \n")) ; Work->do_grow = TRUE ; if (!UMF_grow_front (Numeric, fnr2, fnc2, Work, -1)) { /* since the minimum front size is 1-by-1, it would be nearly * impossible to run out of memory here. */ DEBUGm4 (("out of memory: start front\n")) ; return (FALSE) ; } } else { /* use the existing front */ DEBUGm1 ((" existing front ok\n")) ; Work->fnr_curr = fnr2 ; Work->fnc_curr = fnc2 ; Work->Flblock = Work->Flublock + nb * nb ; Work->Fublock = Work->Flblock + nb * fnr2 ; Work->Fcblock = Work->Fublock + nb * fnc2 ; } ASSERT (Work->Flblock == Work->Flublock + Work->nb*Work->nb) ; ASSERT (Work->Fublock == Work->Flblock + Work->fnr_curr*Work->nb) ; ASSERT (Work->Fcblock == Work->Fublock + Work->nb*Work->fnc_curr) ; return (TRUE) ; }
GLOBAL Int UMF_create_element ( NumericType *Numeric, WorkType *Work, SymbolicType *Symbolic ) { /* ---------------------------------------------------------------------- */ /* local variables */ /* ---------------------------------------------------------------------- */ Int j, col, row, *Fcols, *Frows, fnrows, fncols, *Cols, len, needunits, t1, t2, size, e, i, *E, *Fcpos, *Frpos, *Rows, eloc, fnr_curr, f, got_memory, *Row_tuples, *Row_degree, *Row_tlen, *Col_tuples, max_mark, *Col_degree, *Col_tlen, nn, n_row, n_col, r2, c2, do_Fcpos ; Entry *C, *Fcol ; Element *ep ; Unit *p, *Memory ; Tuple *tp, *tp1, *tp2, tuple, *tpend ; #ifndef NDEBUG DEBUG2 (("FRONTAL WRAPUP\n")) ; UMF_dump_current_front (Numeric, Work, TRUE) ; #endif /* ---------------------------------------------------------------------- */ /* get parameters */ /* ---------------------------------------------------------------------- */ ASSERT (Work->fnpiv == 0) ; ASSERT (Work->fnzeros == 0) ; Row_degree = Numeric->Rperm ; Row_tuples = Numeric->Uip ; Row_tlen = Numeric->Uilen ; Col_degree = Numeric->Cperm ; Col_tuples = Numeric->Lip ; Col_tlen = Numeric->Lilen ; n_row = Work->n_row ; n_col = Work->n_col ; nn = MAX (n_row, n_col) ; Fcols = Work->Fcols ; Frows = Work->Frows ; Fcpos = Work->Fcpos ; Frpos = Work->Frpos ; Memory = Numeric->Memory ; fncols = Work->fncols ; fnrows = Work->fnrows ; tp = (Tuple *) NULL ; tp1 = (Tuple *) NULL ; tp2 = (Tuple *) NULL ; /* ---------------------------------------------------------------------- */ /* add the current frontal matrix to the degrees of each column */ /* ---------------------------------------------------------------------- */ if (!Symbolic->fixQ) { /* but only if the column ordering is not fixed */ #pragma ivdep for (j = 0 ; j < fncols ; j++) { /* add the current frontal matrix to the degree */ ASSERT (Fcols [j] >= 0 && Fcols [j] < n_col) ; Col_degree [Fcols [j]] += fnrows ; } } /* ---------------------------------------------------------------------- */ /* add the current frontal matrix to the degrees of each row */ /* ---------------------------------------------------------------------- */ #pragma ivdep for (i = 0 ; i < fnrows ; i++) { /* add the current frontal matrix to the degree */ ASSERT (Frows [i] >= 0 && Frows [i] < n_row) ; Row_degree [Frows [i]] += fncols ; } /* ---------------------------------------------------------------------- */ /* Reset the external degree counters */ /* ---------------------------------------------------------------------- */ E = Work->E ; max_mark = MAX_MARK (nn) ; if (!Work->pivcol_in_front) { /* clear the external column degrees. no more Usons of current front */ Work->cdeg0 += (nn + 1) ; if (Work->cdeg0 >= max_mark) { /* guard against integer overflow. This is very rare */ DEBUG1 (("Integer overflow, cdeg\n")) ; Work->cdeg0 = 1 ; #pragma ivdep for (e = 1 ; e <= Work->nel ; e++) { if (E [e]) { ep = (Element *) (Memory + E [e]) ; ep->cdeg = 0 ; } } } } if (!Work->pivrow_in_front) { /* clear the external row degrees. no more Lsons of current front */ Work->rdeg0 += (nn + 1) ; if (Work->rdeg0 >= max_mark) { /* guard against integer overflow. This is very rare */ DEBUG1 (("Integer overflow, rdeg\n")) ; Work->rdeg0 = 1 ; #pragma ivdep for (e = 1 ; e <= Work->nel ; e++) { if (E [e]) { ep = (Element *) (Memory + E [e]) ; ep->rdeg = 0 ; } } } } /* ---------------------------------------------------------------------- */ /* clear row/col offsets */ /* ---------------------------------------------------------------------- */ if (!Work->pivrow_in_front) { #pragma ivdep for (j = 0 ; j < fncols ; j++) { Fcpos [Fcols [j]] = EMPTY ; } } if (!Work->pivcol_in_front) { #pragma ivdep for (i = 0 ; i < fnrows ; i++) { Frpos [Frows [i]] = EMPTY ; } } if (fncols <= 0 || fnrows <= 0) { /* no element to create */ DEBUG2 (("Element evaporation\n")) ; Work->prior_element = EMPTY ; return (TRUE) ; } /* ---------------------------------------------------------------------- */ /* create element for later assembly */ /* ---------------------------------------------------------------------- */ #ifndef NDEBUG UMF_allocfail = FALSE ; if (UMF_gprob > 0) { double rrr = ((double) (rand ( ))) / (((double) RAND_MAX) + 1) ; DEBUG4 (("Check random %e %e\n", rrr, UMF_gprob)) ; UMF_allocfail = rrr < UMF_gprob ; if (UMF_allocfail) DEBUGm2 (("Random garbage collection (create)\n")); } #endif needunits = 0 ; got_memory = FALSE ; eloc = UMF_mem_alloc_element (Numeric, fnrows, fncols, &Rows, &Cols, &C, &needunits, &ep) ; /* if UMF_get_memory needs to be called */ if (Work->do_grow) { /* full compaction of current frontal matrix, since UMF_grow_front will * be called next anyway. */ r2 = fnrows ; c2 = fncols ; do_Fcpos = FALSE ; } else { /* partial compaction. */ r2 = MAX (fnrows, Work->fnrows_new + 1) ; c2 = MAX (fncols, Work->fncols_new + 1) ; /* recompute Fcpos if pivot row is in the front */ do_Fcpos = Work->pivrow_in_front ; } if (!eloc) { /* Do garbage collection, realloc, and try again. */ /* Compact the current front if it needs to grow anyway. */ /* Note that there are no pivot rows or columns in the current front */ DEBUGm3 (("get_memory from umf_create_element, 1\n")) ; if (!UMF_get_memory (Numeric, Work, needunits, r2, c2, do_Fcpos)) { /* :: out of memory in umf_create_element (1) :: */ DEBUGm4 (("out of memory: create element (1)\n")) ; return (FALSE) ; /* out of memory */ } got_memory = TRUE ; Memory = Numeric->Memory ; eloc = UMF_mem_alloc_element (Numeric, fnrows, fncols, &Rows, &Cols, &C, &needunits, &ep) ; ASSERT (eloc >= 0) ; if (!eloc) { /* :: out of memory in umf_create_element (2) :: */ DEBUGm4 (("out of memory: create element (2)\n")) ; return (FALSE) ; /* out of memory */ } } e = ++(Work->nel) ; /* get the name of this new frontal matrix */ Work->prior_element = e ; DEBUG8 (("wrapup e "ID" nel "ID"\n", e, Work->nel)) ; ASSERT (e > 0 && e < Work->elen) ; ASSERT (E [e] == 0) ; E [e] = eloc ; if (Work->pivcol_in_front) { /* the new element is a Uson of the next frontal matrix */ ep->cdeg = Work->cdeg0 ; } if (Work->pivrow_in_front) { /* the new element is an Lson of the next frontal matrix */ ep->rdeg = Work->rdeg0 ; } /* ---------------------------------------------------------------------- */ /* copy frontal matrix into the new element */ /* ---------------------------------------------------------------------- */ #pragma ivdep for (i = 0 ; i < fnrows ; i++) { Rows [i] = Frows [i] ; } #pragma ivdep for (i = 0 ; i < fncols ; i++) { Cols [i] = Fcols [i] ; } Fcol = Work->Fcblock ; DEBUG0 (("copy front "ID" by "ID"\n", fnrows, fncols)) ; fnr_curr = Work->fnr_curr ; ASSERT (fnr_curr >= 0 && fnr_curr % 2 == 1) ; for (j = 0 ; j < fncols ; j++) { copy_column (fnrows, Fcol, C) ; Fcol += fnr_curr ; C += fnrows ; } DEBUG8 (("element copied\n")) ; /* ---------------------------------------------------------------------- */ /* add tuples for the new element */ /* ---------------------------------------------------------------------- */ tuple.e = e ; if (got_memory) { /* ------------------------------------------------------------------ */ /* UMF_get_memory ensures enough space exists for each new tuple */ /* ------------------------------------------------------------------ */ /* place (e,f) in the element list of each column */ for (tuple.f = 0 ; tuple.f < fncols ; tuple.f++) { col = Fcols [tuple.f] ; ASSERT (col >= 0 && col < n_col) ; ASSERT (NON_PIVOTAL_COL (col)) ; ASSERT (Col_tuples [col]) ; tp = ((Tuple *) (Memory + Col_tuples [col])) + Col_tlen [col]++ ; *tp = tuple ; } /* ------------------------------------------------------------------ */ /* place (e,f) in the element list of each row */ for (tuple.f = 0 ; tuple.f < fnrows ; tuple.f++) { row = Frows [tuple.f] ; ASSERT (row >= 0 && row < n_row) ; ASSERT (NON_PIVOTAL_ROW (row)) ; ASSERT (Row_tuples [row]) ; tp = ((Tuple *) (Memory + Row_tuples [row])) + Row_tlen [row]++ ; *tp = tuple ; } } else { /* ------------------------------------------------------------------ */ /* place (e,f) in the element list of each column */ /* ------------------------------------------------------------------ */ /* might not have enough space for each tuple */ for (tuple.f = 0 ; tuple.f < fncols ; tuple.f++) { col = Fcols [tuple.f] ; ASSERT (col >= 0 && col < n_col) ; ASSERT (NON_PIVOTAL_COL (col)) ; t1 = Col_tuples [col] ; DEBUG1 (("Placing on col:"ID" , tuples at "ID"\n", col, Col_tuples [col])) ; size = 0 ; len = 0 ; if (t1) { p = Memory + t1 ; tp = (Tuple *) p ; size = GET_BLOCK_SIZE (p) ; len = Col_tlen [col] ; tp2 = tp + len ; } needunits = UNITS (Tuple, len + 1) ; DEBUG1 (("len: "ID" size: "ID" needunits: "ID"\n", len, size, needunits)); if (needunits > size && t1) { /* prune the tuples */ tp1 = tp ; tp2 = tp ; tpend = tp + len ; for ( ; tp < tpend ; tp++) { e = tp->e ; ASSERT (e > 0 && e <= Work->nel) ; if (!E [e]) continue ; /* element already deallocated */ f = tp->f ; p = Memory + E [e] ; ep = (Element *) p ; p += UNITS (Element, 1) ; Cols = (Int *) p ; ; if (Cols [f] == EMPTY) continue ; /* already assembled */ ASSERT (col == Cols [f]) ; *tp2++ = *tp ; /* leave the tuple in the list */ } len = tp2 - tp1 ; Col_tlen [col] = len ; needunits = UNITS (Tuple, len + 1) ; } if (needunits > size) { /* no room exists - reallocate elsewhere */ DEBUG1 (("REALLOCATE Col: "ID", size "ID" to "ID"\n", col, size, 2*needunits)) ; #ifndef NDEBUG UMF_allocfail = FALSE ; if (UMF_gprob > 0) /* a double relop, but ignore NaN case */ { double rrr = ((double) (rand ( ))) / (((double) RAND_MAX) + 1) ; DEBUG1 (("Check random %e %e\n", rrr, UMF_gprob)) ; UMF_allocfail = rrr < UMF_gprob ; if (UMF_allocfail) DEBUGm2 (("Random gar. (col tuple)\n")) ; } #endif needunits = MIN (2*needunits, (Int) UNITS (Tuple, nn)) ; t2 = UMF_mem_alloc_tail_block (Numeric, needunits) ; if (!t2) { /* :: get memory in umf_create_element (1) :: */ /* get memory, reconstruct all tuple lists, and return */ /* Compact the current front if it needs to grow anyway. */ /* Note: no pivot rows or columns in the current front */ DEBUGm4 (("get_memory from umf_create_element, 1\n")) ; return (UMF_get_memory (Numeric, Work, 0, r2, c2,do_Fcpos)); } Col_tuples [col] = t2 ; tp2 = (Tuple *) (Memory + t2) ; if (t1) { for (i = 0 ; i < len ; i++) { *tp2++ = *tp1++ ; } UMF_mem_free_tail_block (Numeric, t1) ; } } /* place the new (e,f) tuple in the element list of the column */ Col_tlen [col]++ ; *tp2 = tuple ; } /* ------------------------------------------------------------------ */ /* place (e,f) in the element list of each row */ /* ------------------------------------------------------------------ */ for (tuple.f = 0 ; tuple.f < fnrows ; tuple.f++) { row = Frows [tuple.f] ; ASSERT (row >= 0 && row < n_row) ; ASSERT (NON_PIVOTAL_ROW (row)) ; t1 = Row_tuples [row] ; DEBUG1 (("Placing on row:"ID" , tuples at "ID"\n", row, Row_tuples [row])) ; size = 0 ; len = 0 ; if (t1) { p = Memory + t1 ; tp = (Tuple *) p ; size = GET_BLOCK_SIZE (p) ; len = Row_tlen [row] ; tp2 = tp + len ; } needunits = UNITS (Tuple, len + 1) ; DEBUG1 (("len: "ID" size: "ID" needunits: "ID"\n", len, size, needunits)) ; if (needunits > size && t1) { /* prune the tuples */ tp1 = tp ; tp2 = tp ; tpend = tp + len ; for ( ; tp < tpend ; tp++) { e = tp->e ; ASSERT (e > 0 && e <= Work->nel) ; if (!E [e]) { continue ; /* element already deallocated */ } f = tp->f ; p = Memory + E [e] ; ep = (Element *) p ; p += UNITS (Element, 1) ; Cols = (Int *) p ; Rows = Cols + (ep->ncols) ; if (Rows [f] == EMPTY) continue ; /* already assembled */ ASSERT (row == Rows [f]) ; *tp2++ = *tp ; /* leave the tuple in the list */ } len = tp2 - tp1 ; Row_tlen [row] = len ; needunits = UNITS (Tuple, len + 1) ; } if (needunits > size) { /* no room exists - reallocate elsewhere */ DEBUG1 (("REALLOCATE Row: "ID", size "ID" to "ID"\n", row, size, 2*needunits)) ; #ifndef NDEBUG UMF_allocfail = FALSE ; if (UMF_gprob > 0) /* a double relop, but ignore NaN case */ { double rrr = ((double) (rand ( ))) / (((double) RAND_MAX) + 1) ; DEBUG1 (("Check random %e %e\n", rrr, UMF_gprob)) ; UMF_allocfail = rrr < UMF_gprob ; if (UMF_allocfail) DEBUGm2 (("Random gar. (row tuple)\n")) ; } #endif needunits = MIN (2*needunits, (Int) UNITS (Tuple, nn)) ; t2 = UMF_mem_alloc_tail_block (Numeric, needunits) ; if (!t2) { /* :: get memory in umf_create_element (2) :: */ /* get memory, reconstruct all tuple lists, and return */ /* Compact the current front if it needs to grow anyway. */ /* Note: no pivot rows or columns in the current front */ DEBUGm4 (("get_memory from umf_create_element, 2\n")) ; return (UMF_get_memory (Numeric, Work, 0, r2, c2,do_Fcpos)); } Row_tuples [row] = t2 ; tp2 = (Tuple *) (Memory + t2) ; if (t1) { for (i = 0 ; i < len ; i++) { *tp2++ = *tp1++ ; } UMF_mem_free_tail_block (Numeric, t1) ; } } /* place the new (e,f) tuple in the element list of the row */ Row_tlen [row]++ ; *tp2 = tuple ; } } /* ---------------------------------------------------------------------- */ #ifndef NDEBUG DEBUG1 (("Done extending\nFINAL: element row pattern: len="ID"\n", fncols)); for (j = 0 ; j < fncols ; j++) DEBUG1 ((""ID"\n", Fcols [j])) ; DEBUG1 (("FINAL: element col pattern: len="ID"\n", fnrows)) ; for (j = 0 ; j < fnrows ; j++) DEBUG1 ((""ID"\n", Frows [j])) ; for (j = 0 ; j < fncols ; j++) { col = Fcols [j] ; ASSERT (col >= 0 && col < n_col) ; UMF_dump_rowcol (1, Numeric, Work, col, !Symbolic->fixQ) ; } for (j = 0 ; j < fnrows ; j++) { row = Frows [j] ; ASSERT (row >= 0 && row < n_row) ; UMF_dump_rowcol (0, Numeric, Work, row, TRUE) ; } if (n_row < 1000 && n_col < 1000) { UMF_dump_memory (Numeric) ; } DEBUG1 (("New element, after filling with stuff: "ID"\n", e)) ; UMF_dump_element (Numeric, Work, e, TRUE) ; if (nn < 1000) { DEBUG4 (("Matrix dump, after New element: "ID"\n", e)) ; UMF_dump_matrix (Numeric, Work, TRUE) ; } DEBUG3 (("FRONTAL WRAPUP DONE\n")) ; #endif return (TRUE) ; }