comparison libcruft/lapack/slag2.f @ 7789:82be108cc558

First attempt at single precision tyeps * * * corrections to qrupdate single precision routines * * * prefer demotion to single over promotion to double * * * Add single precision support to log2 function * * * Trivial PROJECT file update * * * Cache optimized hermitian/transpose methods * * * Add tests for tranpose/hermitian and ChangeLog entry for new transpose code
author David Bateman <dbateman@free.fr>
date Sun, 27 Apr 2008 22:34:17 +0200
parents
children
comparison
equal deleted inserted replaced
7788:45f5faba05a2 7789:82be108cc558
1 SUBROUTINE SLAG2( A, LDA, B, LDB, SAFMIN, SCALE1, SCALE2, WR1,
2 $ WR2, WI )
3 *
4 * -- LAPACK auxiliary routine (version 3.1) --
5 * Univ. of Tennessee, Univ. of California Berkeley and NAG Ltd..
6 * November 2006
7 *
8 * .. Scalar Arguments ..
9 INTEGER LDA, LDB
10 REAL SAFMIN, SCALE1, SCALE2, WI, WR1, WR2
11 * ..
12 * .. Array Arguments ..
13 REAL A( LDA, * ), B( LDB, * )
14 * ..
15 *
16 * Purpose
17 * =======
18 *
19 * SLAG2 computes the eigenvalues of a 2 x 2 generalized eigenvalue
20 * problem A - w B, with scaling as necessary to avoid over-/underflow.
21 *
22 * The scaling factor "s" results in a modified eigenvalue equation
23 *
24 * s A - w B
25 *
26 * where s is a non-negative scaling factor chosen so that w, w B,
27 * and s A do not overflow and, if possible, do not underflow, either.
28 *
29 * Arguments
30 * =========
31 *
32 * A (input) REAL array, dimension (LDA, 2)
33 * On entry, the 2 x 2 matrix A. It is assumed that its 1-norm
34 * is less than 1/SAFMIN. Entries less than
35 * sqrt(SAFMIN)*norm(A) are subject to being treated as zero.
36 *
37 * LDA (input) INTEGER
38 * The leading dimension of the array A. LDA >= 2.
39 *
40 * B (input) REAL array, dimension (LDB, 2)
41 * On entry, the 2 x 2 upper triangular matrix B. It is
42 * assumed that the one-norm of B is less than 1/SAFMIN. The
43 * diagonals should be at least sqrt(SAFMIN) times the largest
44 * element of B (in absolute value); if a diagonal is smaller
45 * than that, then +/- sqrt(SAFMIN) will be used instead of
46 * that diagonal.
47 *
48 * LDB (input) INTEGER
49 * The leading dimension of the array B. LDB >= 2.
50 *
51 * SAFMIN (input) REAL
52 * The smallest positive number s.t. 1/SAFMIN does not
53 * overflow. (This should always be SLAMCH('S') -- it is an
54 * argument in order to avoid having to call SLAMCH frequently.)
55 *
56 * SCALE1 (output) REAL
57 * A scaling factor used to avoid over-/underflow in the
58 * eigenvalue equation which defines the first eigenvalue. If
59 * the eigenvalues are complex, then the eigenvalues are
60 * ( WR1 +/- WI i ) / SCALE1 (which may lie outside the
61 * exponent range of the machine), SCALE1=SCALE2, and SCALE1
62 * will always be positive. If the eigenvalues are real, then
63 * the first (real) eigenvalue is WR1 / SCALE1 , but this may
64 * overflow or underflow, and in fact, SCALE1 may be zero or
65 * less than the underflow threshhold if the exact eigenvalue
66 * is sufficiently large.
67 *
68 * SCALE2 (output) REAL
69 * A scaling factor used to avoid over-/underflow in the
70 * eigenvalue equation which defines the second eigenvalue. If
71 * the eigenvalues are complex, then SCALE2=SCALE1. If the
72 * eigenvalues are real, then the second (real) eigenvalue is
73 * WR2 / SCALE2 , but this may overflow or underflow, and in
74 * fact, SCALE2 may be zero or less than the underflow
75 * threshhold if the exact eigenvalue is sufficiently large.
76 *
77 * WR1 (output) REAL
78 * If the eigenvalue is real, then WR1 is SCALE1 times the
79 * eigenvalue closest to the (2,2) element of A B**(-1). If the
80 * eigenvalue is complex, then WR1=WR2 is SCALE1 times the real
81 * part of the eigenvalues.
82 *
83 * WR2 (output) REAL
84 * If the eigenvalue is real, then WR2 is SCALE2 times the
85 * other eigenvalue. If the eigenvalue is complex, then
86 * WR1=WR2 is SCALE1 times the real part of the eigenvalues.
87 *
88 * WI (output) REAL
89 * If the eigenvalue is real, then WI is zero. If the
90 * eigenvalue is complex, then WI is SCALE1 times the imaginary
91 * part of the eigenvalues. WI will always be non-negative.
92 *
93 * =====================================================================
94 *
95 * .. Parameters ..
96 REAL ZERO, ONE, TWO
97 PARAMETER ( ZERO = 0.0E+0, ONE = 1.0E+0, TWO = 2.0E+0 )
98 REAL HALF
99 PARAMETER ( HALF = ONE / TWO )
100 REAL FUZZY1
101 PARAMETER ( FUZZY1 = ONE+1.0E-5 )
102 * ..
103 * .. Local Scalars ..
104 REAL A11, A12, A21, A22, ABI22, ANORM, AS11, AS12,
105 $ AS22, ASCALE, B11, B12, B22, BINV11, BINV22,
106 $ BMIN, BNORM, BSCALE, BSIZE, C1, C2, C3, C4, C5,
107 $ DIFF, DISCR, PP, QQ, R, RTMAX, RTMIN, S1, S2,
108 $ SAFMAX, SHIFT, SS, SUM, WABS, WBIG, WDET,
109 $ WSCALE, WSIZE, WSMALL
110 * ..
111 * .. Intrinsic Functions ..
112 INTRINSIC ABS, MAX, MIN, SIGN, SQRT
113 * ..
114 * .. Executable Statements ..
115 *
116 RTMIN = SQRT( SAFMIN )
117 RTMAX = ONE / RTMIN
118 SAFMAX = ONE / SAFMIN
119 *
120 * Scale A
121 *
122 ANORM = MAX( ABS( A( 1, 1 ) )+ABS( A( 2, 1 ) ),
123 $ ABS( A( 1, 2 ) )+ABS( A( 2, 2 ) ), SAFMIN )
124 ASCALE = ONE / ANORM
125 A11 = ASCALE*A( 1, 1 )
126 A21 = ASCALE*A( 2, 1 )
127 A12 = ASCALE*A( 1, 2 )
128 A22 = ASCALE*A( 2, 2 )
129 *
130 * Perturb B if necessary to insure non-singularity
131 *
132 B11 = B( 1, 1 )
133 B12 = B( 1, 2 )
134 B22 = B( 2, 2 )
135 BMIN = RTMIN*MAX( ABS( B11 ), ABS( B12 ), ABS( B22 ), RTMIN )
136 IF( ABS( B11 ).LT.BMIN )
137 $ B11 = SIGN( BMIN, B11 )
138 IF( ABS( B22 ).LT.BMIN )
139 $ B22 = SIGN( BMIN, B22 )
140 *
141 * Scale B
142 *
143 BNORM = MAX( ABS( B11 ), ABS( B12 )+ABS( B22 ), SAFMIN )
144 BSIZE = MAX( ABS( B11 ), ABS( B22 ) )
145 BSCALE = ONE / BSIZE
146 B11 = B11*BSCALE
147 B12 = B12*BSCALE
148 B22 = B22*BSCALE
149 *
150 * Compute larger eigenvalue by method described by C. van Loan
151 *
152 * ( AS is A shifted by -SHIFT*B )
153 *
154 BINV11 = ONE / B11
155 BINV22 = ONE / B22
156 S1 = A11*BINV11
157 S2 = A22*BINV22
158 IF( ABS( S1 ).LE.ABS( S2 ) ) THEN
159 AS12 = A12 - S1*B12
160 AS22 = A22 - S1*B22
161 SS = A21*( BINV11*BINV22 )
162 ABI22 = AS22*BINV22 - SS*B12
163 PP = HALF*ABI22
164 SHIFT = S1
165 ELSE
166 AS12 = A12 - S2*B12
167 AS11 = A11 - S2*B11
168 SS = A21*( BINV11*BINV22 )
169 ABI22 = -SS*B12
170 PP = HALF*( AS11*BINV11+ABI22 )
171 SHIFT = S2
172 END IF
173 QQ = SS*AS12
174 IF( ABS( PP*RTMIN ).GE.ONE ) THEN
175 DISCR = ( RTMIN*PP )**2 + QQ*SAFMIN
176 R = SQRT( ABS( DISCR ) )*RTMAX
177 ELSE
178 IF( PP**2+ABS( QQ ).LE.SAFMIN ) THEN
179 DISCR = ( RTMAX*PP )**2 + QQ*SAFMAX
180 R = SQRT( ABS( DISCR ) )*RTMIN
181 ELSE
182 DISCR = PP**2 + QQ
183 R = SQRT( ABS( DISCR ) )
184 END IF
185 END IF
186 *
187 * Note: the test of R in the following IF is to cover the case when
188 * DISCR is small and negative and is flushed to zero during
189 * the calculation of R. On machines which have a consistent
190 * flush-to-zero threshhold and handle numbers above that
191 * threshhold correctly, it would not be necessary.
192 *
193 IF( DISCR.GE.ZERO .OR. R.EQ.ZERO ) THEN
194 SUM = PP + SIGN( R, PP )
195 DIFF = PP - SIGN( R, PP )
196 WBIG = SHIFT + SUM
197 *
198 * Compute smaller eigenvalue
199 *
200 WSMALL = SHIFT + DIFF
201 IF( HALF*ABS( WBIG ).GT.MAX( ABS( WSMALL ), SAFMIN ) ) THEN
202 WDET = ( A11*A22-A12*A21 )*( BINV11*BINV22 )
203 WSMALL = WDET / WBIG
204 END IF
205 *
206 * Choose (real) eigenvalue closest to 2,2 element of A*B**(-1)
207 * for WR1.
208 *
209 IF( PP.GT.ABI22 ) THEN
210 WR1 = MIN( WBIG, WSMALL )
211 WR2 = MAX( WBIG, WSMALL )
212 ELSE
213 WR1 = MAX( WBIG, WSMALL )
214 WR2 = MIN( WBIG, WSMALL )
215 END IF
216 WI = ZERO
217 ELSE
218 *
219 * Complex eigenvalues
220 *
221 WR1 = SHIFT + PP
222 WR2 = WR1
223 WI = R
224 END IF
225 *
226 * Further scaling to avoid underflow and overflow in computing
227 * SCALE1 and overflow in computing w*B.
228 *
229 * This scale factor (WSCALE) is bounded from above using C1 and C2,
230 * and from below using C3 and C4.
231 * C1 implements the condition s A must never overflow.
232 * C2 implements the condition w B must never overflow.
233 * C3, with C2,
234 * implement the condition that s A - w B must never overflow.
235 * C4 implements the condition s should not underflow.
236 * C5 implements the condition max(s,|w|) should be at least 2.
237 *
238 C1 = BSIZE*( SAFMIN*MAX( ONE, ASCALE ) )
239 C2 = SAFMIN*MAX( ONE, BNORM )
240 C3 = BSIZE*SAFMIN
241 IF( ASCALE.LE.ONE .AND. BSIZE.LE.ONE ) THEN
242 C4 = MIN( ONE, ( ASCALE / SAFMIN )*BSIZE )
243 ELSE
244 C4 = ONE
245 END IF
246 IF( ASCALE.LE.ONE .OR. BSIZE.LE.ONE ) THEN
247 C5 = MIN( ONE, ASCALE*BSIZE )
248 ELSE
249 C5 = ONE
250 END IF
251 *
252 * Scale first eigenvalue
253 *
254 WABS = ABS( WR1 ) + ABS( WI )
255 WSIZE = MAX( SAFMIN, C1, FUZZY1*( WABS*C2+C3 ),
256 $ MIN( C4, HALF*MAX( WABS, C5 ) ) )
257 IF( WSIZE.NE.ONE ) THEN
258 WSCALE = ONE / WSIZE
259 IF( WSIZE.GT.ONE ) THEN
260 SCALE1 = ( MAX( ASCALE, BSIZE )*WSCALE )*
261 $ MIN( ASCALE, BSIZE )
262 ELSE
263 SCALE1 = ( MIN( ASCALE, BSIZE )*WSCALE )*
264 $ MAX( ASCALE, BSIZE )
265 END IF
266 WR1 = WR1*WSCALE
267 IF( WI.NE.ZERO ) THEN
268 WI = WI*WSCALE
269 WR2 = WR1
270 SCALE2 = SCALE1
271 END IF
272 ELSE
273 SCALE1 = ASCALE*BSIZE
274 SCALE2 = SCALE1
275 END IF
276 *
277 * Scale second eigenvalue (if real)
278 *
279 IF( WI.EQ.ZERO ) THEN
280 WSIZE = MAX( SAFMIN, C1, FUZZY1*( ABS( WR2 )*C2+C3 ),
281 $ MIN( C4, HALF*MAX( ABS( WR2 ), C5 ) ) )
282 IF( WSIZE.NE.ONE ) THEN
283 WSCALE = ONE / WSIZE
284 IF( WSIZE.GT.ONE ) THEN
285 SCALE2 = ( MAX( ASCALE, BSIZE )*WSCALE )*
286 $ MIN( ASCALE, BSIZE )
287 ELSE
288 SCALE2 = ( MIN( ASCALE, BSIZE )*WSCALE )*
289 $ MAX( ASCALE, BSIZE )
290 END IF
291 WR2 = WR2*WSCALE
292 ELSE
293 SCALE2 = ASCALE*BSIZE
294 END IF
295 END IF
296 *
297 * End of SLAG2
298 *
299 RETURN
300 END