/* ========================================================================== */ /* === UMF_2by2 ============================================================= */ /* ========================================================================== */ /* -------------------------------------------------------------------------- */ /* UMFPACK Copyright (c) Timothy A. Davis, CISE, */ /* Univ. of Florida. All Rights Reserved. See ../Doc/License for License. */ /* web: http://www.cise.ufl.edu/research/sparse/umfpack */ /* -------------------------------------------------------------------------- */ /* Not user-callable. Computes a row permutation P so that A (P,:) has a * mostly zero-free diagonal, with large entries on the diagonal. It does this * by swapping pairs of rows. Once a row is swapped it is not swapped again. * This is a "cheap" assignment, not a complete max. transversal or * bi-partite matching. It is only a partial matching. For most matrices * for which this algorithm is used, however, the matching is complete (in * UMFPACK this algorithm is used for matrices with roughly symmetric pattern, * and these matrices typically have a mostly-zero-free diagonal to begin with. * This algorithm is not meant to be used on arbitrary unsymmetric matrices * (for those matrices, UMFPACK uses its unsymmetric strategy and does not * use this algorithm). * * Even if incomplete, the matching is usually good enough for UMFPACK's * symmetric strategy, which can easily pivot off the diagonal during numerical * factorization if it finds a weak diagonal entry. * * The algorithms works as follows. First, row scaling factors are computed, * and weak diagonal entries are found. A weak entry is a value A(k,k) whose * absolute value is < tol * max (abs (A (:,k))). For each weak diagonal k in * increasing order of degree in A+A', the algorithm finds an index j such * that A (k,j) and A (j,k) are "large" (greater than or equal to tol times * the largest magnitude in their columns). Row j must also not have already * been swapped. Rows j and k are then swapped. If we come to a diagonal k * that has already been swapped, then it is not modified. This case occurs * for "oxo" pivots: * * k j * k o x * j x o * * which are swapped once to obtain * * k j * j x o * k o x * * These two rows are then not modified any further (A (j,j) was weak, but * after one swap the permuted the jth diagonal entry is strong. * * This algorithm only works on square matrices (real, complex, or pattern- * only). The numerical values are optional. If not present, each entry is * treated as numerically acceptable (tol is ignored), and the algorithm * operates by just using the pattern, not the values. Each column of the * input matrix A must be sorted, with no duplicate entries. The matrix A * can be optionally scaled prior to the numerical test. The matrix A (:,P) * has the same diagonal entries as A (:,P), except in different order. So * the output permutation P can also be used to swap the columns of A. */ #include "umf_internal.h" #ifndef NDEBUG #include "umf_is_permutation.h" #endif /* x is "weak" if it is less than ctol. If x or ctol are NaN, then define * x as not "weak". This is a rather arbitrary choice, made to simplify the * computation. On all but a PC with Microsoft C/C++, this test becomes * ((x) - ctol < 0). */ #define WEAK(x,ctol) (SCALAR_IS_LTZERO ((x)-(ctol))) /* For flag value in Next [col] */ #define IS_WEAK -2 /* ========================================================================== */ /* === two_by_two =========================================================== */ /* ========================================================================== */ PRIVATE Int two_by_two /* returns # unmatched weak diagonals */ ( /* input, not modified */ Int n2, /* C is n2-by-n2 */ Int Cp [ ], /* size n2+1, column pointers for C */ Int Ci [ ], /* size snz = Cp [n2], row indices for C */ Int Degree [ ], /* Degree [i] = degree of row i of C+C' */ /* input, not defined on output */ Int Next [ ], /* Next [k] == IS_WEAK if k is a weak diagonal */ Int Ri [ ], /* Ri [i] is the length of row i in C */ /* output, not defined on input */ Int P [ ], /* workspace, not defined on input or output */ Int Rp [ ], Int Head [ ] ) { Int deg, newcol, row, col, p, p2, unmatched, k, j, j2, j_best, best, jdiff, jdiff_best, jdeg, jdeg_best, cp, cp1, cp2, rp, rp1, rp2, maxdeg, mindeg ; /* ---------------------------------------------------------------------- */ /* place weak diagonals in the degree lists */ /* ---------------------------------------------------------------------- */ for (deg = 0 ; deg < n2 ; deg++) { Head [deg] = EMPTY ; } maxdeg = 0 ; mindeg = Int_MAX ; for (newcol = n2-1 ; newcol >= 0 ; newcol--) { if (Next [newcol] == IS_WEAK) { /* add this column to the list of weak nodes */ DEBUGm1 ((" newcol "ID" has a weak diagonal deg "ID"\n", newcol, deg)) ; deg = Degree [newcol] ; ASSERT (deg >= 0 && deg < n2) ; Next [newcol] = Head [deg] ; Head [deg] = newcol ; maxdeg = MAX (maxdeg, deg) ; mindeg = MIN (mindeg, deg) ; } } /* ---------------------------------------------------------------------- */ /* construct R = C' (C = strong entries in pruned submatrix) */ /* ---------------------------------------------------------------------- */ /* Ri [0..n2-1] is the length of each row of R */ /* use P as temporary pointer into the row form of R [ */ Rp [0] = 0 ; for (row = 0 ; row < n2 ; row++) { Rp [row+1] = Rp [row] + Ri [row] ; P [row] = Rp [row] ; } /* Ri no longer needed for row counts */ /* all entries in C are strong */ for (col = 0 ; col < n2 ; col++) { p2 = Cp [col+1] ; for (p = Cp [col] ; p < p2 ; p++) { /* place the column index in row = Ci [p] */ Ri [P [Ci [p]]++] = col ; } } /* contents of P no longer needed ] */ #ifndef NDEBUG DEBUG0 (("==================R: row form of strong entries in A:\n")) ; UMF_dump_col_matrix ((double *) NULL, #ifdef COMPLEX (double *) NULL, #endif Ri, Rp, n2, n2, Rp [n2]) ; #endif ASSERT (AMD_valid (n2, n2, Rp, Ri) == AMD_OK) ; /* ---------------------------------------------------------------------- */ /* for each weak diagonal, find a pair of strong off-diagonal entries */ /* ---------------------------------------------------------------------- */ for (row = 0 ; row < n2 ; row++) { P [row] = EMPTY ; } unmatched = 0 ; best = EMPTY ; jdiff = EMPTY ; jdeg = EMPTY ; for (deg = mindeg ; deg <= maxdeg ; deg++) { /* find the next weak diagonal of lowest degree */ DEBUGm2 (("---------------------------------- Deg: "ID"\n", deg)) ; for (k = Head [deg] ; k != EMPTY ; k = Next [k]) { DEBUGm2 (("k: "ID"\n", k)) ; if (P [k] == EMPTY) { /* C (k,k) is a weak diagonal entry. Find an index j != k such * that C (j,k) and C (k,j) are both strong, and also such * that Degree [j] is minimized. In case of a tie, pick * the smallest index j. C and R contain the pattern of * strong entries only. * * Note that row k of R and column k of C are both sorted. */ DEBUGm4 (("===== Weak diagonal k = "ID"\n", k)) ; DEBUG1 (("Column k of C:\n")) ; for (p = Cp [k] ; p < Cp [k+1] ; p++) { DEBUG1 ((" "ID": deg "ID"\n", Ci [p], Degree [Ci [p]])); } DEBUG1 (("Row k of R (strong entries only):\n")) ; for (p = Rp [k] ; p < Rp [k+1] ; p++) { DEBUG1 ((" "ID": deg "ID"\n", Ri [p], Degree [Ri [p]])); } /* no (C (k,j), C (j,k)) pair exists yet */ j_best = EMPTY ; jdiff_best = Int_MAX ; jdeg_best = Int_MAX ; /* pointers into column k (including values) */ cp1 = Cp [k] ; cp2 = Cp [k+1] ; cp = cp1 ; /* pointers into row k (strong entries only, no values) */ rp1 = Rp [k] ; rp2 = Rp [k+1] ; rp = rp1 ; /* while entries searched in column k and row k */ while (TRUE) { if (cp >= cp2) { /* no more entries in this column */ break ; } /* get C (j,k), which is strong */ j = Ci [cp] ; if (rp >= rp2) { /* no more entries in this column */ break ; } /* get R (k,j2), which is strong */ j2 = Ri [rp] ; if (j < j2) { /* C (j,k) is strong, but R (k,j) is not strong */ cp++ ; continue ; } if (j2 < j) { /* C (k,j2) is strong, but R (j2,k) is not strong */ rp++ ; continue ; } /* j == j2: C (j,k) is strong and R (k,j) is strong */ best = FALSE ; if (P [j] == EMPTY) { /* j has not yet been matched */ jdeg = Degree [j] ; jdiff = SCALAR_ABS (k-j) ; DEBUG1 (("Try candidate j "ID" deg "ID" diff "ID "\n", j, jdeg, jdiff)) ; if (j_best == EMPTY) { /* this is the first candidate seen */ DEBUG1 ((" first\n")) ; best = TRUE ; } else { if (jdeg < jdeg_best) { /* the degree of j is best seen so far. */ DEBUG1 ((" least degree\n")) ; best = TRUE ; } else if (jdeg == jdeg_best) { /* degree of j and j_best are the same */ /* tie break by nearest node number */ if (jdiff < jdiff_best) { DEBUG1 ((" tie degree, closer\n")) ; best = TRUE ; } else if (jdiff == jdiff_best) { /* |j-k| = |j_best-k|. For any given k * and j_best there is only one other j * than can be just as close as j_best. * Tie break by picking the smaller of * j and j_best */ DEBUG1 ((" tie degree, as close\n")); best = j < j_best ; } } else { /* j has higher degree than best so far */ best = FALSE ; } } } if (best) { /* j is best match for k */ /* found a strong pair, A (j,k) and A (k,j) */ DEBUG1 ((" --- Found pair k: "ID" j: " ID " jdeg: "ID" jdiff: "ID"\n", k, j, jdeg, jdiff)) ; ASSERT (jdiff != EMPTY) ; ASSERT (jdeg != EMPTY) ; j_best = j ; jdeg_best = jdeg ; jdiff_best = jdiff ; } /* get the next entries in column k and row k */ cp++ ; rp++ ; } /* save the pair (j,k), if we found one */ if (j_best != EMPTY) { j = j_best ; DEBUGm4 ((" --- best pair j: "ID" for k: "ID"\n", j, k)) ; P [k] = j ; P [j] = k ; } else { /* no match was found for k */ unmatched++ ; } } } } /* ---------------------------------------------------------------------- */ /* finalize the row permutation, P */ /* ---------------------------------------------------------------------- */ for (k = 0 ; k < n2 ; k++) { if (P [k] == EMPTY) { P [k] = k ; } } ASSERT (UMF_is_permutation (P, Rp, n2, n2)) ; return (unmatched) ; } /* ========================================================================== */ /* === UMF_2by2 ============================================================= */ /* ========================================================================== */ GLOBAL void UMF_2by2 ( /* input, not modified: */ Int n, /* A is n-by-n */ const Int Ap [ ], /* size n+1 */ const Int Ai [ ], /* size nz = Ap [n] */ const double Ax [ ], /* size nz if present */ #ifdef COMPLEX const double Az [ ], /* size nz if present */ #endif double tol, /* tolerance for determining whether or not an * entry is numerically acceptable. If tol <= 0 * then all numerical values ignored. */ Int scale, /* scaling to perform (none, sum, or max) */ Int Cperm1 [ ], /* singleton permutations */ #ifndef NDEBUG Int Rperm1 [ ], /* not needed, since Rperm1 = Cperm1 for submatrix S */ #endif Int InvRperm1 [ ], /* inverse of Rperm1 */ Int n1, /* number of singletons */ Int nempty, /* number of empty rows/cols */ /* input, contents undefined on output: */ Int Degree [ ], /* Degree [j] is the number of off-diagonal * entries in row/column j of S+S', where * where S = A (Cperm1 [n1..], Rperm1 [n1..]). * Note that S is not used, nor formed. */ /* output: */ Int P [ ], /* P [k] = i means original row i is kth row in S(P,:) * where S = A (Cperm1 [n1..], Rperm1 [n1..]) */ Int *p_nweak, Int *p_unmatched, /* workspace (not defined on input or output): */ Int Ri [ ], /* of size >= max (nz, n) */ Int Rp [ ], /* of size n+1 */ double Rs [ ], /* of size n if present. Rs = sum (abs (A),2) or * max (abs (A),2), the sum or max of each row. Unused * if scale is equal to UMFPACK_SCALE_NONE. */ Int Head [ ], /* of size n. Head pointers for bucket sort */ Int Next [ ], /* of size n. Next pointers for bucket sort */ Int Ci [ ], /* size nz */ Int Cp [ ] /* size n+1 */ ) { /* ---------------------------------------------------------------------- */ /* local variables */ /* ---------------------------------------------------------------------- */ Entry aij ; double cmax, value, rs, ctol, dvalue ; Int k, p, row, col, do_values, do_sum, do_max, do_scale, nweak, weak, p1, p2, dfound, unmatched, n2, oldrow, newrow, oldcol, newcol, pp ; #ifdef COMPLEX Int split = SPLIT (Az) ; #endif #ifndef NRECIPROCAL Int do_recip = FALSE ; #endif #ifndef NDEBUG /* UMF_debug += 99 ; */ DEBUGm3 (("\n ==================================UMF_2by2: tol %g\n", tol)) ; ASSERT (AMD_valid (n, n, Ap, Ai) == AMD_OK) ; for (k = n1 ; k < n - nempty ; k++) { ASSERT (Cperm1 [k] == Rperm1 [k]) ; } #endif /* ---------------------------------------------------------------------- */ /* determine scaling options */ /* ---------------------------------------------------------------------- */ /* use the values, but only if they are present */ /* ignore the values if tol <= 0 */ do_values = (tol > 0) && (Ax != (double *) NULL) ; if (do_values && (Rs != (double *) NULL)) { do_sum = (scale == UMFPACK_SCALE_SUM) ; do_max = (scale == UMFPACK_SCALE_MAX) ; } else { /* no scaling */ do_sum = FALSE ; do_max = FALSE ; } do_scale = do_max || do_sum ; DEBUGm3 (("do_values "ID" do_sum "ID" do_max "ID" do_scale "ID"\n", do_values, do_sum, do_max, do_scale)) ; /* ---------------------------------------------------------------------- */ /* compute the row scaling, if requested */ /* ---------------------------------------------------------------------- */ /* see also umf_kernel_init */ if (do_scale) { #ifndef NRECIPROCAL double rsmin ; #endif for (row = 0 ; row < n ; row++) { Rs [row] = 0.0 ; } for (col = 0 ; col < n ; col++) { p2 = Ap [col+1] ; for (p = Ap [col] ; p < p2 ; p++) { row = Ai [p] ; ASSIGN (aij, Ax, Az, p, split) ; APPROX_ABS (value, aij) ; rs = Rs [row] ; if (!SCALAR_IS_NAN (rs)) { if (SCALAR_IS_NAN (value)) { /* if any entry in a row is NaN, then the scale factor * for the row is NaN. It will be set to 1 later. */ Rs [row] = value ; } else if (do_max) { Rs [row] = MAX (rs, value) ; } else { Rs [row] += value ; } } } } #ifndef NRECIPROCAL rsmin = Rs [0] ; if (SCALAR_IS_ZERO (rsmin) || SCALAR_IS_NAN (rsmin)) { rsmin = 1.0 ; } #endif for (row = 0 ; row < n ; row++) { /* do not scale an empty row, or a row with a NaN */ rs = Rs [row] ; if (SCALAR_IS_ZERO (rs) || SCALAR_IS_NAN (rs)) { Rs [row] = 1.0 ; } #ifndef NRECIPROCAL rsmin = MIN (rsmin, Rs [row]) ; #endif } #ifndef NRECIPROCAL /* multiply by the reciprocal if Rs is not too small */ do_recip = (rsmin >= RECIPROCAL_TOLERANCE) ; if (do_recip) { /* invert the scale factors */ for (row = 0 ; row < n ; row++) { Rs [row] = 1.0 / Rs [row] ; } } #endif } /* ---------------------------------------------------------------------- */ /* compute the max in each column and find diagonal */ /* ---------------------------------------------------------------------- */ nweak = 0 ; #ifndef NDEBUG for (k = 0 ; k < n ; k++) { ASSERT (Rperm1 [k] >= 0 && Rperm1 [k] < n) ; ASSERT (InvRperm1 [Rperm1 [k]] == k) ; } #endif n2 = n - n1 - nempty ; /* use Ri to count the number of strong entries in each row */ for (row = 0 ; row < n2 ; row++) { Ri [row] = 0 ; } pp = 0 ; ctol = 0 ; dvalue = 1 ; /* construct C = pruned submatrix, strong values only, column form */ for (k = n1 ; k < n - nempty ; k++) { oldcol = Cperm1 [k] ; newcol = k - n1 ; Next [newcol] = EMPTY ; DEBUGm1 (("Column "ID" newcol "ID" oldcol "ID"\n", k, newcol, oldcol)) ; Cp [newcol] = pp ; dfound = FALSE ; p1 = Ap [oldcol] ; p2 = Ap [oldcol+1] ; if (do_values) { cmax = 0 ; dvalue = 0 ; if (!do_scale) { /* no scaling */ for (p = p1 ; p < p2 ; p++) { oldrow = Ai [p] ; ASSERT (oldrow >= 0 && oldrow < n) ; newrow = InvRperm1 [oldrow] - n1 ; ASSERT (newrow >= -n1 && newrow < n2) ; if (newrow < 0) continue ; ASSIGN (aij, Ax, Az, p, split) ; APPROX_ABS (value, aij) ; /* if either cmax or value is NaN, define cmax as NaN */ if (!SCALAR_IS_NAN (cmax)) { if (SCALAR_IS_NAN (value)) { cmax = value ; } else { cmax = MAX (cmax, value) ; } } if (oldrow == oldcol) { /* we found the diagonal entry in this column */ dvalue = value ; dfound = TRUE ; ASSERT (newrow == newcol) ; } } } #ifndef NRECIPROCAL else if (do_recip) { /* multiply by the reciprocal */ for (p = p1 ; p < p2 ; p++) { oldrow = Ai [p] ; ASSERT (oldrow >= 0 && oldrow < n) ; newrow = InvRperm1 [oldrow] - n1 ; ASSERT (newrow >= -n1 && newrow < n2) ; if (newrow < 0) continue ; ASSIGN (aij, Ax, Az, p, split) ; APPROX_ABS (value, aij) ; value *= Rs [oldrow] ; /* if either cmax or value is NaN, define cmax as NaN */ if (!SCALAR_IS_NAN (cmax)) { if (SCALAR_IS_NAN (value)) { cmax = value ; } else { cmax = MAX (cmax, value) ; } } if (oldrow == oldcol) { /* we found the diagonal entry in this column */ dvalue = value ; dfound = TRUE ; ASSERT (newrow == newcol) ; } } } #endif else { /* divide instead */ for (p = p1 ; p < p2 ; p++) { oldrow = Ai [p] ; ASSERT (oldrow >= 0 && oldrow < n) ; newrow = InvRperm1 [oldrow] - n1 ; ASSERT (newrow >= -n1 && newrow < n2) ; if (newrow < 0) continue ; ASSIGN (aij, Ax, Az, p, split) ; APPROX_ABS (value, aij) ; value /= Rs [oldrow] ; /* if either cmax or value is NaN, define cmax as NaN */ if (!SCALAR_IS_NAN (cmax)) { if (SCALAR_IS_NAN (value)) { cmax = value ; } else { cmax = MAX (cmax, value) ; } } if (oldrow == oldcol) { /* we found the diagonal entry in this column */ dvalue = value ; dfound = TRUE ; ASSERT (newrow == newcol) ; } } } ctol = tol * cmax ; DEBUGm1 ((" cmax col "ID" %g ctol %g\n", oldcol, cmax, ctol)) ; } else { for (p = p1 ; p < p2 ; p++) { oldrow = Ai [p] ; ASSERT (oldrow >= 0 && oldrow < n) ; newrow = InvRperm1 [oldrow] - n1 ; ASSERT (newrow >= -n1 && newrow < n2) ; if (newrow < 0) continue ; Ci [pp++] = newrow ; if (oldrow == oldcol) { /* we found the diagonal entry in this column */ ASSERT (newrow == newcol) ; dfound = TRUE ; } /* count the entries in each column */ Ri [newrow]++ ; } } /* ------------------------------------------------------------------ */ /* flag the weak diagonals */ /* ------------------------------------------------------------------ */ if (!dfound) { /* no diagonal entry present */ weak = TRUE ; } else { /* diagonal entry is present, check its value */ weak = (do_values) ? WEAK (dvalue, ctol) : FALSE ; } if (weak) { /* flag this column as weak */ DEBUG0 (("Weak!\n")) ; Next [newcol] = IS_WEAK ; nweak++ ; } /* ------------------------------------------------------------------ */ /* count entries in each row that are not numerically weak */ /* ------------------------------------------------------------------ */ if (do_values) { if (!do_scale) { /* no scaling */ for (p = p1 ; p < p2 ; p++) { oldrow = Ai [p] ; newrow = InvRperm1 [oldrow] - n1 ; if (newrow < 0) continue ; ASSIGN (aij, Ax, Az, p, split) ; APPROX_ABS (value, aij) ; weak = WEAK (value, ctol) ; if (!weak) { DEBUG0 ((" strong: row "ID": %g\n", oldrow, value)) ; Ci [pp++] = newrow ; Ri [newrow]++ ; } } } #ifndef NRECIPROCAL else if (do_recip) { /* multiply by the reciprocal */ for (p = p1 ; p < p2 ; p++) { oldrow = Ai [p] ; newrow = InvRperm1 [oldrow] - n1 ; if (newrow < 0) continue ; ASSIGN (aij, Ax, Az, p, split) ; APPROX_ABS (value, aij) ; value *= Rs [oldrow] ; weak = WEAK (value, ctol) ; if (!weak) { DEBUG0 ((" strong: row "ID": %g\n", oldrow, value)) ; Ci [pp++] = newrow ; Ri [newrow]++ ; } } } #endif else { /* divide instead */ for (p = p1 ; p < p2 ; p++) { oldrow = Ai [p] ; newrow = InvRperm1 [oldrow] - n1 ; if (newrow < 0) continue ; ASSIGN (aij, Ax, Az, p, split) ; APPROX_ABS (value, aij) ; value /= Rs [oldrow] ; weak = WEAK (value, ctol) ; if (!weak) { DEBUG0 ((" strong: row "ID": %g\n", oldrow, value)) ; Ci [pp++] = newrow ; Ri [newrow]++ ; } } } } } Cp [n2] = pp ; ASSERT (AMD_valid (n2, n2, Cp, Ci) == AMD_OK) ; if (nweak == 0) { /* nothing to do, quick return */ DEBUGm2 (("\n =============================UMF_2by2: quick return\n")) ; for (k = 0 ; k < n ; k++) { P [k] = k ; } *p_nweak = 0 ; *p_unmatched = 0 ; return ; } #ifndef NDEBUG for (k = 0 ; k < n2 ; k++) { P [k] = EMPTY ; } for (k = 0 ; k < n2 ; k++) { ASSERT (Degree [k] >= 0 && Degree [k] < n2) ; } #endif /* ---------------------------------------------------------------------- */ /* find the 2-by-2 permutation */ /* ---------------------------------------------------------------------- */ /* The matrix S is now mapped to the index range 0 to n2-1. We have * S = A (Rperm [n1 .. n-nempty-1], Cperm [n1 .. n-nempty-1]), and then * C = pattern of strong entries in S. A weak diagonal k in S is marked * with Next [k] = IS_WEAK. */ unmatched = two_by_two (n2, Cp, Ci, Degree, Next, Ri, P, Rp, Head) ; /* ---------------------------------------------------------------------- */ *p_nweak = nweak ; *p_unmatched = unmatched ; #ifndef NDEBUG DEBUGm4 (("UMF_2by2: weak "ID" unmatched "ID"\n", nweak, unmatched)) ; for (row = 0 ; row < n ; row++) { DEBUGm2 (("P ["ID"] = "ID"\n", row, P [row])) ; } DEBUGm2 (("\n =============================UMF_2by2: done\n\n")) ; #endif }