LAPACK 3.3.0
|
00001 INTEGER FUNCTION SLANEG( N, D, LLD, SIGMA, PIVMIN, R ) 00002 IMPLICIT NONE 00003 * 00004 * -- LAPACK auxiliary routine (version 3.2.2) -- 00005 * -- LAPACK is a software package provided by Univ. of Tennessee, -- 00006 * -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..-- 00007 * June 2010 00008 * 00009 * .. Scalar Arguments .. 00010 INTEGER N, R 00011 REAL PIVMIN, SIGMA 00012 * .. 00013 * .. Array Arguments .. 00014 REAL D( * ), LLD( * ) 00015 * .. 00016 * 00017 * Purpose 00018 * ======= 00019 * 00020 * SLANEG computes the Sturm count, the number of negative pivots 00021 * encountered while factoring tridiagonal T - sigma I = L D L^T. 00022 * This implementation works directly on the factors without forming 00023 * the tridiagonal matrix T. The Sturm count is also the number of 00024 * eigenvalues of T less than sigma. 00025 * 00026 * This routine is called from SLARRB. 00027 * 00028 * The current routine does not use the PIVMIN parameter but rather 00029 * requires IEEE-754 propagation of Infinities and NaNs. This 00030 * routine also has no input range restrictions but does require 00031 * default exception handling such that x/0 produces Inf when x is 00032 * non-zero, and Inf/Inf produces NaN. For more information, see: 00033 * 00034 * Marques, Riedy, and Voemel, "Benefits of IEEE-754 Features in 00035 * Modern Symmetric Tridiagonal Eigensolvers," SIAM Journal on 00036 * Scientific Computing, v28, n5, 2006. DOI 10.1137/050641624 00037 * (Tech report version in LAWN 172 with the same title.) 00038 * 00039 * Arguments 00040 * ========= 00041 * 00042 * N (input) INTEGER 00043 * The order of the matrix. 00044 * 00045 * D (input) REAL array, dimension (N) 00046 * The N diagonal elements of the diagonal matrix D. 00047 * 00048 * LLD (input) REAL array, dimension (N-1) 00049 * The (N-1) elements L(i)*L(i)*D(i). 00050 * 00051 * SIGMA (input) REAL 00052 * Shift amount in T - sigma I = L D L^T. 00053 * 00054 * PIVMIN (input) REAL 00055 * The minimum pivot in the Sturm sequence. May be used 00056 * when zero pivots are encountered on non-IEEE-754 00057 * architectures. 00058 * 00059 * R (input) INTEGER 00060 * The twist index for the twisted factorization that is used 00061 * for the negcount. 00062 * 00063 * Further Details 00064 * =============== 00065 * 00066 * Based on contributions by 00067 * Osni Marques, LBNL/NERSC, USA 00068 * Christof Voemel, University of California, Berkeley, USA 00069 * Jason Riedy, University of California, Berkeley, USA 00070 * 00071 * ===================================================================== 00072 * 00073 * .. Parameters .. 00074 REAL ZERO, ONE 00075 PARAMETER ( ZERO = 0.0E0, ONE = 1.0E0 ) 00076 * Some architectures propagate Infinities and NaNs very slowly, so 00077 * the code computes counts in BLKLEN chunks. Then a NaN can 00078 * propagate at most BLKLEN columns before being detected. This is 00079 * not a general tuning parameter; it needs only to be just large 00080 * enough that the overhead is tiny in common cases. 00081 INTEGER BLKLEN 00082 PARAMETER ( BLKLEN = 128 ) 00083 * .. 00084 * .. Local Scalars .. 00085 INTEGER BJ, J, NEG1, NEG2, NEGCNT 00086 REAL BSAV, DMINUS, DPLUS, GAMMA, P, T, TMP 00087 LOGICAL SAWNAN 00088 * .. 00089 * .. Intrinsic Functions .. 00090 INTRINSIC MIN, MAX 00091 * .. 00092 * .. External Functions .. 00093 LOGICAL SISNAN 00094 EXTERNAL SISNAN 00095 * .. 00096 * .. Executable Statements .. 00097 00098 NEGCNT = 0 00099 00100 * I) upper part: L D L^T - SIGMA I = L+ D+ L+^T 00101 T = -SIGMA 00102 DO 210 BJ = 1, R-1, BLKLEN 00103 NEG1 = 0 00104 BSAV = T 00105 DO 21 J = BJ, MIN(BJ+BLKLEN-1, R-1) 00106 DPLUS = D( J ) + T 00107 IF( DPLUS.LT.ZERO ) NEG1 = NEG1 + 1 00108 TMP = T / DPLUS 00109 T = TMP * LLD( J ) - SIGMA 00110 21 CONTINUE 00111 SAWNAN = SISNAN( T ) 00112 * Run a slower version of the above loop if a NaN is detected. 00113 * A NaN should occur only with a zero pivot after an infinite 00114 * pivot. In that case, substituting 1 for T/DPLUS is the 00115 * correct limit. 00116 IF( SAWNAN ) THEN 00117 NEG1 = 0 00118 T = BSAV 00119 DO 22 J = BJ, MIN(BJ+BLKLEN-1, R-1) 00120 DPLUS = D( J ) + T 00121 IF( DPLUS.LT.ZERO ) NEG1 = NEG1 + 1 00122 TMP = T / DPLUS 00123 IF (SISNAN(TMP)) TMP = ONE 00124 T = TMP * LLD(J) - SIGMA 00125 22 CONTINUE 00126 END IF 00127 NEGCNT = NEGCNT + NEG1 00128 210 CONTINUE 00129 * 00130 * II) lower part: L D L^T - SIGMA I = U- D- U-^T 00131 P = D( N ) - SIGMA 00132 DO 230 BJ = N-1, R, -BLKLEN 00133 NEG2 = 0 00134 BSAV = P 00135 DO 23 J = BJ, MAX(BJ-BLKLEN+1, R), -1 00136 DMINUS = LLD( J ) + P 00137 IF( DMINUS.LT.ZERO ) NEG2 = NEG2 + 1 00138 TMP = P / DMINUS 00139 P = TMP * D( J ) - SIGMA 00140 23 CONTINUE 00141 SAWNAN = SISNAN( P ) 00142 * As above, run a slower version that substitutes 1 for Inf/Inf. 00143 * 00144 IF( SAWNAN ) THEN 00145 NEG2 = 0 00146 P = BSAV 00147 DO 24 J = BJ, MAX(BJ-BLKLEN+1, R), -1 00148 DMINUS = LLD( J ) + P 00149 IF( DMINUS.LT.ZERO ) NEG2 = NEG2 + 1 00150 TMP = P / DMINUS 00151 IF (SISNAN(TMP)) TMP = ONE 00152 P = TMP * D(J) - SIGMA 00153 24 CONTINUE 00154 END IF 00155 NEGCNT = NEGCNT + NEG2 00156 230 CONTINUE 00157 * 00158 * III) Twist index 00159 * T was shifted by SIGMA initially. 00160 GAMMA = (T + SIGMA) + P 00161 IF( GAMMA.LT.ZERO ) NEGCNT = NEGCNT+1 00162 00163 SLANEG = NEGCNT 00164 END