Annotation of rpl/lapack/lapack/zlahef.f, revision 1.20
1.14 bertrand 1: *> \brief \b ZLAHEF computes a partial factorization of a complex Hermitian indefinite matrix using the Bunch-Kaufman diagonal pivoting method (blocked algorithm, calling Level 3 BLAS).
1.9 bertrand 2: *
3: * =========== DOCUMENTATION ===========
4: *
1.17 bertrand 5: * Online html documentation available at
6: * http://www.netlib.org/lapack/explore-html/
1.9 bertrand 7: *
8: *> \htmlonly
1.17 bertrand 9: *> Download ZLAHEF + dependencies
10: *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/zlahef.f">
11: *> [TGZ]</a>
12: *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/zlahef.f">
13: *> [ZIP]</a>
14: *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/zlahef.f">
1.9 bertrand 15: *> [TXT]</a>
1.17 bertrand 16: *> \endhtmlonly
1.9 bertrand 17: *
18: * Definition:
19: * ===========
20: *
21: * SUBROUTINE ZLAHEF( UPLO, N, NB, KB, A, LDA, IPIV, W, LDW, INFO )
1.17 bertrand 22: *
1.9 bertrand 23: * .. Scalar Arguments ..
24: * CHARACTER UPLO
25: * INTEGER INFO, KB, LDA, LDW, N, NB
26: * ..
27: * .. Array Arguments ..
28: * INTEGER IPIV( * )
29: * COMPLEX*16 A( LDA, * ), W( LDW, * )
30: * ..
1.17 bertrand 31: *
1.9 bertrand 32: *
33: *> \par Purpose:
34: * =============
35: *>
36: *> \verbatim
37: *>
38: *> ZLAHEF computes a partial factorization of a complex Hermitian
39: *> matrix A using the Bunch-Kaufman diagonal pivoting method. The
40: *> partial factorization has the form:
41: *>
42: *> A = ( I U12 ) ( A11 0 ) ( I 0 ) if UPLO = 'U', or:
43: *> ( 0 U22 ) ( 0 D ) ( U12**H U22**H )
44: *>
45: *> A = ( L11 0 ) ( D 0 ) ( L11**H L21**H ) if UPLO = 'L'
46: *> ( L21 I ) ( 0 A22 ) ( 0 I )
47: *>
48: *> where the order of D is at most NB. The actual order is returned in
49: *> the argument KB, and is either NB or NB-1, or N if N <= NB.
50: *> Note that U**H denotes the conjugate transpose of U.
51: *>
52: *> ZLAHEF is an auxiliary routine called by ZHETRF. It uses blocked code
53: *> (calling Level 3 BLAS) to update the submatrix A11 (if UPLO = 'U') or
54: *> A22 (if UPLO = 'L').
55: *> \endverbatim
56: *
57: * Arguments:
58: * ==========
59: *
60: *> \param[in] UPLO
61: *> \verbatim
62: *> UPLO is CHARACTER*1
63: *> Specifies whether the upper or lower triangular part of the
64: *> Hermitian matrix A is stored:
65: *> = 'U': Upper triangular
66: *> = 'L': Lower triangular
67: *> \endverbatim
68: *>
69: *> \param[in] N
70: *> \verbatim
71: *> N is INTEGER
72: *> The order of the matrix A. N >= 0.
73: *> \endverbatim
74: *>
75: *> \param[in] NB
76: *> \verbatim
77: *> NB is INTEGER
78: *> The maximum number of columns of the matrix A that should be
79: *> factored. NB should be at least 2 to allow for 2-by-2 pivot
80: *> blocks.
81: *> \endverbatim
82: *>
83: *> \param[out] KB
84: *> \verbatim
85: *> KB is INTEGER
86: *> The number of columns of A that were actually factored.
87: *> KB is either NB-1 or NB, or N if N <= NB.
88: *> \endverbatim
89: *>
90: *> \param[in,out] A
91: *> \verbatim
92: *> A is COMPLEX*16 array, dimension (LDA,N)
93: *> On entry, the Hermitian matrix A. If UPLO = 'U', the leading
94: *> n-by-n upper triangular part of A contains the upper
95: *> triangular part of the matrix A, and the strictly lower
96: *> triangular part of A is not referenced. If UPLO = 'L', the
97: *> leading n-by-n lower triangular part of A contains the lower
98: *> triangular part of the matrix A, and the strictly upper
99: *> triangular part of A is not referenced.
100: *> On exit, A contains details of the partial factorization.
101: *> \endverbatim
102: *>
103: *> \param[in] LDA
104: *> \verbatim
105: *> LDA is INTEGER
106: *> The leading dimension of the array A. LDA >= max(1,N).
107: *> \endverbatim
108: *>
109: *> \param[out] IPIV
110: *> \verbatim
111: *> IPIV is INTEGER array, dimension (N)
112: *> Details of the interchanges and the block structure of D.
113: *>
1.14 bertrand 114: *> If UPLO = 'U':
115: *> Only the last KB elements of IPIV are set.
116: *>
117: *> If IPIV(k) > 0, then rows and columns k and IPIV(k) were
118: *> interchanged and D(k,k) is a 1-by-1 diagonal block.
119: *>
120: *> If IPIV(k) = IPIV(k-1) < 0, then rows and columns
121: *> k-1 and -IPIV(k) were interchanged and D(k-1:k,k-1:k)
122: *> is a 2-by-2 diagonal block.
123: *>
124: *> If UPLO = 'L':
125: *> Only the first KB elements of IPIV are set.
126: *>
127: *> If IPIV(k) > 0, then rows and columns k and IPIV(k) were
128: *> interchanged and D(k,k) is a 1-by-1 diagonal block.
129: *>
130: *> If IPIV(k) = IPIV(k+1) < 0, then rows and columns
131: *> k+1 and -IPIV(k) were interchanged and D(k:k+1,k:k+1)
132: *> is a 2-by-2 diagonal block.
1.9 bertrand 133: *> \endverbatim
134: *>
135: *> \param[out] W
136: *> \verbatim
137: *> W is COMPLEX*16 array, dimension (LDW,NB)
138: *> \endverbatim
139: *>
140: *> \param[in] LDW
141: *> \verbatim
142: *> LDW is INTEGER
143: *> The leading dimension of the array W. LDW >= max(1,N).
144: *> \endverbatim
145: *>
146: *> \param[out] INFO
147: *> \verbatim
148: *> INFO is INTEGER
149: *> = 0: successful exit
150: *> > 0: if INFO = k, D(k,k) is exactly zero. The factorization
151: *> has been completed, but the block diagonal matrix D is
152: *> exactly singular.
153: *> \endverbatim
154: *
155: * Authors:
156: * ========
157: *
1.17 bertrand 158: *> \author Univ. of Tennessee
159: *> \author Univ. of California Berkeley
160: *> \author Univ. of Colorado Denver
161: *> \author NAG Ltd.
1.9 bertrand 162: *
163: *> \ingroup complex16HEcomputational
164: *
1.14 bertrand 165: *> \par Contributors:
166: * ==================
167: *>
168: *> \verbatim
169: *>
1.17 bertrand 170: *> December 2016, Igor Kozachenko,
1.14 bertrand 171: *> Computer Science Division,
172: *> University of California, Berkeley
173: *> \endverbatim
174: *
1.9 bertrand 175: * =====================================================================
1.1 bertrand 176: SUBROUTINE ZLAHEF( UPLO, N, NB, KB, A, LDA, IPIV, W, LDW, INFO )
177: *
1.20 ! bertrand 178: * -- LAPACK computational routine --
1.1 bertrand 179: * -- LAPACK is a software package provided by Univ. of Tennessee, --
180: * -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
181: *
182: * .. Scalar Arguments ..
183: CHARACTER UPLO
184: INTEGER INFO, KB, LDA, LDW, N, NB
185: * ..
186: * .. Array Arguments ..
187: INTEGER IPIV( * )
188: COMPLEX*16 A( LDA, * ), W( LDW, * )
189: * ..
190: *
191: * =====================================================================
192: *
193: * .. Parameters ..
194: DOUBLE PRECISION ZERO, ONE
195: PARAMETER ( ZERO = 0.0D+0, ONE = 1.0D+0 )
196: COMPLEX*16 CONE
197: PARAMETER ( CONE = ( 1.0D+0, 0.0D+0 ) )
198: DOUBLE PRECISION EIGHT, SEVTEN
199: PARAMETER ( EIGHT = 8.0D+0, SEVTEN = 17.0D+0 )
200: * ..
201: * .. Local Scalars ..
202: INTEGER IMAX, J, JB, JJ, JMAX, JP, K, KK, KKW, KP,
203: $ KSTEP, KW
204: DOUBLE PRECISION ABSAKK, ALPHA, COLMAX, R1, ROWMAX, T
205: COMPLEX*16 D11, D21, D22, Z
206: * ..
207: * .. External Functions ..
208: LOGICAL LSAME
209: INTEGER IZAMAX
210: EXTERNAL LSAME, IZAMAX
211: * ..
212: * .. External Subroutines ..
213: EXTERNAL ZCOPY, ZDSCAL, ZGEMM, ZGEMV, ZLACGV, ZSWAP
214: * ..
215: * .. Intrinsic Functions ..
216: INTRINSIC ABS, DBLE, DCONJG, DIMAG, MAX, MIN, SQRT
217: * ..
218: * .. Statement Functions ..
219: DOUBLE PRECISION CABS1
220: * ..
221: * .. Statement Function definitions ..
222: CABS1( Z ) = ABS( DBLE( Z ) ) + ABS( DIMAG( Z ) )
223: * ..
224: * .. Executable Statements ..
225: *
226: INFO = 0
227: *
228: * Initialize ALPHA for use in choosing pivot block size.
229: *
230: ALPHA = ( ONE+SQRT( SEVTEN ) ) / EIGHT
231: *
232: IF( LSAME( UPLO, 'U' ) ) THEN
233: *
234: * Factorize the trailing columns of A using the upper triangle
235: * of A and working backwards, and compute the matrix W = U12*D
236: * for use in updating A11 (note that conjg(W) is actually stored)
237: *
238: * K is the main loop index, decreasing from N in steps of 1 or 2
239: *
240: * KW is the column of W which corresponds to column K of A
241: *
242: K = N
243: 10 CONTINUE
244: KW = NB + K - N
245: *
246: * Exit from loop
247: *
248: IF( ( K.LE.N-NB+1 .AND. NB.LT.N ) .OR. K.LT.1 )
249: $ GO TO 30
250: *
1.14 bertrand 251: KSTEP = 1
252: *
1.1 bertrand 253: * Copy column K of A to column KW of W and update it
254: *
255: CALL ZCOPY( K-1, A( 1, K ), 1, W( 1, KW ), 1 )
256: W( K, KW ) = DBLE( A( K, K ) )
257: IF( K.LT.N ) THEN
258: CALL ZGEMV( 'No transpose', K, N-K, -CONE, A( 1, K+1 ), LDA,
259: $ W( K, KW+1 ), LDW, CONE, W( 1, KW ), 1 )
260: W( K, KW ) = DBLE( W( K, KW ) )
261: END IF
262: *
263: * Determine rows and columns to be interchanged and whether
264: * a 1-by-1 or 2-by-2 pivot block will be used
265: *
266: ABSAKK = ABS( DBLE( W( K, KW ) ) )
267: *
268: * IMAX is the row-index of the largest off-diagonal element in
1.14 bertrand 269: * column K, and COLMAX is its absolute value.
270: * Determine both COLMAX and IMAX.
1.1 bertrand 271: *
272: IF( K.GT.1 ) THEN
273: IMAX = IZAMAX( K-1, W( 1, KW ), 1 )
274: COLMAX = CABS1( W( IMAX, KW ) )
275: ELSE
276: COLMAX = ZERO
277: END IF
278: *
279: IF( MAX( ABSAKK, COLMAX ).EQ.ZERO ) THEN
280: *
1.14 bertrand 281: * Column K is zero or underflow: set INFO and continue
1.1 bertrand 282: *
283: IF( INFO.EQ.0 )
284: $ INFO = K
285: KP = K
286: A( K, K ) = DBLE( A( K, K ) )
287: ELSE
1.14 bertrand 288: *
289: * ============================================================
290: *
291: * BEGIN pivot search
292: *
293: * Case(1)
1.1 bertrand 294: IF( ABSAKK.GE.ALPHA*COLMAX ) THEN
295: *
296: * no interchange, use 1-by-1 pivot block
297: *
298: KP = K
299: ELSE
300: *
1.14 bertrand 301: * BEGIN pivot search along IMAX row
302: *
303: *
1.1 bertrand 304: * Copy column IMAX to column KW-1 of W and update it
305: *
306: CALL ZCOPY( IMAX-1, A( 1, IMAX ), 1, W( 1, KW-1 ), 1 )
307: W( IMAX, KW-1 ) = DBLE( A( IMAX, IMAX ) )
308: CALL ZCOPY( K-IMAX, A( IMAX, IMAX+1 ), LDA,
309: $ W( IMAX+1, KW-1 ), 1 )
310: CALL ZLACGV( K-IMAX, W( IMAX+1, KW-1 ), 1 )
311: IF( K.LT.N ) THEN
312: CALL ZGEMV( 'No transpose', K, N-K, -CONE,
313: $ A( 1, K+1 ), LDA, W( IMAX, KW+1 ), LDW,
314: $ CONE, W( 1, KW-1 ), 1 )
315: W( IMAX, KW-1 ) = DBLE( W( IMAX, KW-1 ) )
316: END IF
317: *
318: * JMAX is the column-index of the largest off-diagonal
1.14 bertrand 319: * element in row IMAX, and ROWMAX is its absolute value.
320: * Determine only ROWMAX.
1.1 bertrand 321: *
322: JMAX = IMAX + IZAMAX( K-IMAX, W( IMAX+1, KW-1 ), 1 )
323: ROWMAX = CABS1( W( JMAX, KW-1 ) )
324: IF( IMAX.GT.1 ) THEN
325: JMAX = IZAMAX( IMAX-1, W( 1, KW-1 ), 1 )
326: ROWMAX = MAX( ROWMAX, CABS1( W( JMAX, KW-1 ) ) )
327: END IF
328: *
1.14 bertrand 329: * Case(2)
1.1 bertrand 330: IF( ABSAKK.GE.ALPHA*COLMAX*( COLMAX / ROWMAX ) ) THEN
331: *
332: * no interchange, use 1-by-1 pivot block
333: *
334: KP = K
1.14 bertrand 335: *
336: * Case(3)
1.1 bertrand 337: ELSE IF( ABS( DBLE( W( IMAX, KW-1 ) ) ).GE.ALPHA*ROWMAX )
338: $ THEN
339: *
340: * interchange rows and columns K and IMAX, use 1-by-1
341: * pivot block
342: *
343: KP = IMAX
344: *
1.14 bertrand 345: * copy column KW-1 of W to column KW of W
1.1 bertrand 346: *
347: CALL ZCOPY( K, W( 1, KW-1 ), 1, W( 1, KW ), 1 )
1.14 bertrand 348: *
349: * Case(4)
1.1 bertrand 350: ELSE
351: *
352: * interchange rows and columns K-1 and IMAX, use 2-by-2
353: * pivot block
354: *
355: KP = IMAX
356: KSTEP = 2
357: END IF
1.14 bertrand 358: *
359: *
360: * END pivot search along IMAX row
361: *
1.1 bertrand 362: END IF
363: *
1.14 bertrand 364: * END pivot search
365: *
366: * ============================================================
367: *
368: * KK is the column of A where pivoting step stopped
369: *
1.1 bertrand 370: KK = K - KSTEP + 1
1.14 bertrand 371: *
372: * KKW is the column of W which corresponds to column KK of A
373: *
1.1 bertrand 374: KKW = NB + KK - N
375: *
1.14 bertrand 376: * Interchange rows and columns KP and KK.
377: * Updated column KP is already stored in column KKW of W.
1.1 bertrand 378: *
379: IF( KP.NE.KK ) THEN
380: *
1.14 bertrand 381: * Copy non-updated column KK to column KP of submatrix A
382: * at step K. No need to copy element into column K
383: * (or K and K-1 for 2-by-2 pivot) of A, since these columns
384: * will be later overwritten.
1.1 bertrand 385: *
386: A( KP, KP ) = DBLE( A( KK, KK ) )
387: CALL ZCOPY( KK-1-KP, A( KP+1, KK ), 1, A( KP, KP+1 ),
388: $ LDA )
389: CALL ZLACGV( KK-1-KP, A( KP, KP+1 ), LDA )
1.14 bertrand 390: IF( KP.GT.1 )
391: $ CALL ZCOPY( KP-1, A( 1, KK ), 1, A( 1, KP ), 1 )
1.1 bertrand 392: *
1.14 bertrand 393: * Interchange rows KK and KP in last K+1 to N columns of A
394: * (columns K (or K and K-1 for 2-by-2 pivot) of A will be
395: * later overwritten). Interchange rows KK and KP
396: * in last KKW to NB columns of W.
1.1 bertrand 397: *
1.14 bertrand 398: IF( K.LT.N )
399: $ CALL ZSWAP( N-K, A( KK, K+1 ), LDA, A( KP, K+1 ),
1.1 bertrand 400: $ LDA )
401: CALL ZSWAP( N-KK+1, W( KK, KKW ), LDW, W( KP, KKW ),
402: $ LDW )
403: END IF
404: *
405: IF( KSTEP.EQ.1 ) THEN
406: *
1.14 bertrand 407: * 1-by-1 pivot block D(k): column kw of W now holds
1.1 bertrand 408: *
1.14 bertrand 409: * W(kw) = U(k)*D(k),
1.1 bertrand 410: *
411: * where U(k) is the k-th column of U
412: *
1.14 bertrand 413: * (1) Store subdiag. elements of column U(k)
414: * and 1-by-1 block D(k) in column k of A.
415: * (NOTE: Diagonal element U(k,k) is a UNIT element
416: * and not stored)
417: * A(k,k) := D(k,k) = W(k,kw)
418: * A(1:k-1,k) := U(1:k-1,k) = W(1:k-1,kw)/D(k,k)
419: *
420: * (NOTE: No need to use for Hermitian matrix
421: * A( K, K ) = DBLE( W( K, K) ) to separately copy diagonal
422: * element D(k,k) from W (potentially saves only one load))
423: CALL ZCOPY( K, W( 1, KW ), 1, A( 1, K ), 1 )
424: IF( K.GT.1 ) THEN
425: *
426: * (NOTE: No need to check if A(k,k) is NOT ZERO,
427: * since that was ensured earlier in pivot search:
428: * case A(k,k) = 0 falls into 2x2 pivot case(4))
1.1 bertrand 429: *
1.14 bertrand 430: R1 = ONE / DBLE( A( K, K ) )
431: CALL ZDSCAL( K-1, R1, A( 1, K ), 1 )
432: *
433: * (2) Conjugate column W(kw)
1.1 bertrand 434: *
1.14 bertrand 435: CALL ZLACGV( K-1, W( 1, KW ), 1 )
436: END IF
1.1 bertrand 437: *
438: ELSE
439: *
1.14 bertrand 440: * 2-by-2 pivot block D(k): columns kw and kw-1 of W now hold
1.1 bertrand 441: *
1.14 bertrand 442: * ( W(kw-1) W(kw) ) = ( U(k-1) U(k) )*D(k)
1.1 bertrand 443: *
444: * where U(k) and U(k-1) are the k-th and (k-1)-th columns
445: * of U
446: *
1.14 bertrand 447: * (1) Store U(1:k-2,k-1) and U(1:k-2,k) and 2-by-2
448: * block D(k-1:k,k-1:k) in columns k-1 and k of A.
449: * (NOTE: 2-by-2 diagonal block U(k-1:k,k-1:k) is a UNIT
450: * block and not stored)
451: * A(k-1:k,k-1:k) := D(k-1:k,k-1:k) = W(k-1:k,kw-1:kw)
452: * A(1:k-2,k-1:k) := U(1:k-2,k:k-1:k) =
453: * = W(1:k-2,kw-1:kw) * ( D(k-1:k,k-1:k)**(-1) )
454: *
1.1 bertrand 455: IF( K.GT.2 ) THEN
456: *
1.14 bertrand 457: * Factor out the columns of the inverse of 2-by-2 pivot
458: * block D, so that each column contains 1, to reduce the
459: * number of FLOPS when we multiply panel
460: * ( W(kw-1) W(kw) ) by this inverse, i.e. by D**(-1).
461: *
462: * D**(-1) = ( d11 cj(d21) )**(-1) =
463: * ( d21 d22 )
464: *
465: * = 1/(d11*d22-|d21|**2) * ( ( d22) (-cj(d21) ) ) =
466: * ( (-d21) ( d11 ) )
467: *
468: * = 1/(|d21|**2) * 1/((d11/cj(d21))*(d22/d21)-1) *
469: *
470: * * ( d21*( d22/d21 ) conj(d21)*( - 1 ) ) =
471: * ( ( -1 ) ( d11/conj(d21) ) )
472: *
473: * = 1/(|d21|**2) * 1/(D22*D11-1) *
474: *
475: * * ( d21*( D11 ) conj(d21)*( -1 ) ) =
476: * ( ( -1 ) ( D22 ) )
477: *
478: * = (1/|d21|**2) * T * ( d21*( D11 ) conj(d21)*( -1 ) ) =
479: * ( ( -1 ) ( D22 ) )
480: *
481: * = ( (T/conj(d21))*( D11 ) (T/d21)*( -1 ) ) =
482: * ( ( -1 ) ( D22 ) )
483: *
484: * = ( conj(D21)*( D11 ) D21*( -1 ) )
485: * ( ( -1 ) ( D22 ) ),
486: *
487: * where D11 = d22/d21,
488: * D22 = d11/conj(d21),
489: * D21 = T/d21,
490: * T = 1/(D22*D11-1).
491: *
492: * (NOTE: No need to check for division by ZERO,
493: * since that was ensured earlier in pivot search:
494: * (a) d21 != 0, since in 2x2 pivot case(4)
495: * |d21| should be larger than |d11| and |d22|;
496: * (b) (D22*D11 - 1) != 0, since from (a),
497: * both |D11| < 1, |D22| < 1, hence |D22*D11| << 1.)
1.1 bertrand 498: *
499: D21 = W( K-1, KW )
500: D11 = W( K, KW ) / DCONJG( D21 )
501: D22 = W( K-1, KW-1 ) / D21
502: T = ONE / ( DBLE( D11*D22 )-ONE )
503: D21 = T / D21
1.14 bertrand 504: *
505: * Update elements in columns A(k-1) and A(k) as
506: * dot products of rows of ( W(kw-1) W(kw) ) and columns
507: * of D**(-1)
508: *
1.1 bertrand 509: DO 20 J = 1, K - 2
510: A( J, K-1 ) = D21*( D11*W( J, KW-1 )-W( J, KW ) )
511: A( J, K ) = DCONJG( D21 )*
512: $ ( D22*W( J, KW )-W( J, KW-1 ) )
513: 20 CONTINUE
514: END IF
515: *
516: * Copy D(k) to A
517: *
518: A( K-1, K-1 ) = W( K-1, KW-1 )
519: A( K-1, K ) = W( K-1, KW )
520: A( K, K ) = W( K, KW )
521: *
1.14 bertrand 522: * (2) Conjugate columns W(kw) and W(kw-1)
1.1 bertrand 523: *
524: CALL ZLACGV( K-1, W( 1, KW ), 1 )
525: CALL ZLACGV( K-2, W( 1, KW-1 ), 1 )
1.14 bertrand 526: *
1.1 bertrand 527: END IF
1.14 bertrand 528: *
1.1 bertrand 529: END IF
530: *
531: * Store details of the interchanges in IPIV
532: *
533: IF( KSTEP.EQ.1 ) THEN
534: IPIV( K ) = KP
535: ELSE
536: IPIV( K ) = -KP
537: IPIV( K-1 ) = -KP
538: END IF
539: *
540: * Decrease K and return to the start of the main loop
541: *
542: K = K - KSTEP
543: GO TO 10
544: *
545: 30 CONTINUE
546: *
547: * Update the upper triangle of A11 (= A(1:k,1:k)) as
548: *
1.8 bertrand 549: * A11 := A11 - U12*D*U12**H = A11 - U12*W**H
1.1 bertrand 550: *
551: * computing blocks of NB columns at a time (note that conjg(W) is
552: * actually stored)
553: *
554: DO 50 J = ( ( K-1 ) / NB )*NB + 1, 1, -NB
555: JB = MIN( NB, K-J+1 )
556: *
557: * Update the upper triangle of the diagonal block
558: *
559: DO 40 JJ = J, J + JB - 1
560: A( JJ, JJ ) = DBLE( A( JJ, JJ ) )
561: CALL ZGEMV( 'No transpose', JJ-J+1, N-K, -CONE,
562: $ A( J, K+1 ), LDA, W( JJ, KW+1 ), LDW, CONE,
563: $ A( J, JJ ), 1 )
564: A( JJ, JJ ) = DBLE( A( JJ, JJ ) )
565: 40 CONTINUE
566: *
567: * Update the rectangular superdiagonal block
568: *
569: CALL ZGEMM( 'No transpose', 'Transpose', J-1, JB, N-K,
570: $ -CONE, A( 1, K+1 ), LDA, W( J, KW+1 ), LDW,
571: $ CONE, A( 1, J ), LDA )
572: 50 CONTINUE
573: *
574: * Put U12 in standard form by partially undoing the interchanges
1.14 bertrand 575: * in columns k+1:n looping backwards from k+1 to n
1.1 bertrand 576: *
577: J = K + 1
578: 60 CONTINUE
1.14 bertrand 579: *
580: * Undo the interchanges (if any) of rows JJ and JP at each
581: * step J
582: *
583: * (Here, J is a diagonal index)
584: JJ = J
585: JP = IPIV( J )
586: IF( JP.LT.0 ) THEN
587: JP = -JP
588: * (Here, J is a diagonal index)
589: J = J + 1
590: END IF
591: * (NOTE: Here, J is used to determine row length. Length N-J+1
592: * of the rows to swap back doesn't include diagonal element)
1.1 bertrand 593: J = J + 1
1.14 bertrand 594: IF( JP.NE.JJ .AND. J.LE.N )
595: $ CALL ZSWAP( N-J+1, A( JP, J ), LDA, A( JJ, J ), LDA )
596: IF( J.LT.N )
1.1 bertrand 597: $ GO TO 60
598: *
599: * Set KB to the number of columns factorized
600: *
601: KB = N - K
602: *
603: ELSE
604: *
605: * Factorize the leading columns of A using the lower triangle
606: * of A and working forwards, and compute the matrix W = L21*D
607: * for use in updating A22 (note that conjg(W) is actually stored)
608: *
609: * K is the main loop index, increasing from 1 in steps of 1 or 2
610: *
611: K = 1
612: 70 CONTINUE
613: *
614: * Exit from loop
615: *
616: IF( ( K.GE.NB .AND. NB.LT.N ) .OR. K.GT.N )
617: $ GO TO 90
618: *
1.14 bertrand 619: KSTEP = 1
620: *
1.1 bertrand 621: * Copy column K of A to column K of W and update it
622: *
623: W( K, K ) = DBLE( A( K, K ) )
624: IF( K.LT.N )
625: $ CALL ZCOPY( N-K, A( K+1, K ), 1, W( K+1, K ), 1 )
626: CALL ZGEMV( 'No transpose', N-K+1, K-1, -CONE, A( K, 1 ), LDA,
627: $ W( K, 1 ), LDW, CONE, W( K, K ), 1 )
628: W( K, K ) = DBLE( W( K, K ) )
629: *
630: * Determine rows and columns to be interchanged and whether
631: * a 1-by-1 or 2-by-2 pivot block will be used
632: *
633: ABSAKK = ABS( DBLE( W( K, K ) ) )
634: *
635: * IMAX is the row-index of the largest off-diagonal element in
1.14 bertrand 636: * column K, and COLMAX is its absolute value.
637: * Determine both COLMAX and IMAX.
1.1 bertrand 638: *
639: IF( K.LT.N ) THEN
640: IMAX = K + IZAMAX( N-K, W( K+1, K ), 1 )
641: COLMAX = CABS1( W( IMAX, K ) )
642: ELSE
643: COLMAX = ZERO
644: END IF
645: *
646: IF( MAX( ABSAKK, COLMAX ).EQ.ZERO ) THEN
647: *
1.14 bertrand 648: * Column K is zero or underflow: set INFO and continue
1.1 bertrand 649: *
650: IF( INFO.EQ.0 )
651: $ INFO = K
652: KP = K
653: A( K, K ) = DBLE( A( K, K ) )
654: ELSE
1.14 bertrand 655: *
656: * ============================================================
657: *
658: * BEGIN pivot search
659: *
660: * Case(1)
1.1 bertrand 661: IF( ABSAKK.GE.ALPHA*COLMAX ) THEN
662: *
663: * no interchange, use 1-by-1 pivot block
664: *
665: KP = K
666: ELSE
667: *
1.14 bertrand 668: * BEGIN pivot search along IMAX row
669: *
670: *
1.1 bertrand 671: * Copy column IMAX to column K+1 of W and update it
672: *
673: CALL ZCOPY( IMAX-K, A( IMAX, K ), LDA, W( K, K+1 ), 1 )
674: CALL ZLACGV( IMAX-K, W( K, K+1 ), 1 )
675: W( IMAX, K+1 ) = DBLE( A( IMAX, IMAX ) )
676: IF( IMAX.LT.N )
677: $ CALL ZCOPY( N-IMAX, A( IMAX+1, IMAX ), 1,
678: $ W( IMAX+1, K+1 ), 1 )
679: CALL ZGEMV( 'No transpose', N-K+1, K-1, -CONE, A( K, 1 ),
680: $ LDA, W( IMAX, 1 ), LDW, CONE, W( K, K+1 ),
681: $ 1 )
682: W( IMAX, K+1 ) = DBLE( W( IMAX, K+1 ) )
683: *
684: * JMAX is the column-index of the largest off-diagonal
1.14 bertrand 685: * element in row IMAX, and ROWMAX is its absolute value.
686: * Determine only ROWMAX.
1.1 bertrand 687: *
688: JMAX = K - 1 + IZAMAX( IMAX-K, W( K, K+1 ), 1 )
689: ROWMAX = CABS1( W( JMAX, K+1 ) )
690: IF( IMAX.LT.N ) THEN
691: JMAX = IMAX + IZAMAX( N-IMAX, W( IMAX+1, K+1 ), 1 )
692: ROWMAX = MAX( ROWMAX, CABS1( W( JMAX, K+1 ) ) )
693: END IF
694: *
1.14 bertrand 695: * Case(2)
1.1 bertrand 696: IF( ABSAKK.GE.ALPHA*COLMAX*( COLMAX / ROWMAX ) ) THEN
697: *
698: * no interchange, use 1-by-1 pivot block
699: *
700: KP = K
1.14 bertrand 701: *
702: * Case(3)
1.1 bertrand 703: ELSE IF( ABS( DBLE( W( IMAX, K+1 ) ) ).GE.ALPHA*ROWMAX )
704: $ THEN
705: *
706: * interchange rows and columns K and IMAX, use 1-by-1
707: * pivot block
708: *
709: KP = IMAX
710: *
1.14 bertrand 711: * copy column K+1 of W to column K of W
1.1 bertrand 712: *
713: CALL ZCOPY( N-K+1, W( K, K+1 ), 1, W( K, K ), 1 )
1.14 bertrand 714: *
715: * Case(4)
1.1 bertrand 716: ELSE
717: *
718: * interchange rows and columns K+1 and IMAX, use 2-by-2
719: * pivot block
720: *
721: KP = IMAX
722: KSTEP = 2
723: END IF
1.14 bertrand 724: *
725: *
726: * END pivot search along IMAX row
727: *
1.1 bertrand 728: END IF
729: *
1.14 bertrand 730: * END pivot search
731: *
732: * ============================================================
733: *
734: * KK is the column of A where pivoting step stopped
735: *
1.1 bertrand 736: KK = K + KSTEP - 1
737: *
1.14 bertrand 738: * Interchange rows and columns KP and KK.
739: * Updated column KP is already stored in column KK of W.
1.1 bertrand 740: *
741: IF( KP.NE.KK ) THEN
742: *
1.14 bertrand 743: * Copy non-updated column KK to column KP of submatrix A
744: * at step K. No need to copy element into column K
745: * (or K and K+1 for 2-by-2 pivot) of A, since these columns
746: * will be later overwritten.
1.1 bertrand 747: *
748: A( KP, KP ) = DBLE( A( KK, KK ) )
749: CALL ZCOPY( KP-KK-1, A( KK+1, KK ), 1, A( KP, KK+1 ),
750: $ LDA )
751: CALL ZLACGV( KP-KK-1, A( KP, KK+1 ), LDA )
752: IF( KP.LT.N )
753: $ CALL ZCOPY( N-KP, A( KP+1, KK ), 1, A( KP+1, KP ), 1 )
754: *
1.14 bertrand 755: * Interchange rows KK and KP in first K-1 columns of A
756: * (columns K (or K and K+1 for 2-by-2 pivot) of A will be
757: * later overwritten). Interchange rows KK and KP
758: * in first KK columns of W.
1.1 bertrand 759: *
1.14 bertrand 760: IF( K.GT.1 )
761: $ CALL ZSWAP( K-1, A( KK, 1 ), LDA, A( KP, 1 ), LDA )
1.1 bertrand 762: CALL ZSWAP( KK, W( KK, 1 ), LDW, W( KP, 1 ), LDW )
763: END IF
764: *
765: IF( KSTEP.EQ.1 ) THEN
766: *
767: * 1-by-1 pivot block D(k): column k of W now holds
768: *
1.14 bertrand 769: * W(k) = L(k)*D(k),
1.1 bertrand 770: *
771: * where L(k) is the k-th column of L
772: *
1.14 bertrand 773: * (1) Store subdiag. elements of column L(k)
774: * and 1-by-1 block D(k) in column k of A.
775: * (NOTE: Diagonal element L(k,k) is a UNIT element
776: * and not stored)
777: * A(k,k) := D(k,k) = W(k,k)
778: * A(k+1:N,k) := L(k+1:N,k) = W(k+1:N,k)/D(k,k)
779: *
780: * (NOTE: No need to use for Hermitian matrix
781: * A( K, K ) = DBLE( W( K, K) ) to separately copy diagonal
782: * element D(k,k) from W (potentially saves only one load))
1.1 bertrand 783: CALL ZCOPY( N-K+1, W( K, K ), 1, A( K, K ), 1 )
784: IF( K.LT.N ) THEN
1.14 bertrand 785: *
786: * (NOTE: No need to check if A(k,k) is NOT ZERO,
787: * since that was ensured earlier in pivot search:
788: * case A(k,k) = 0 falls into 2x2 pivot case(4))
789: *
1.1 bertrand 790: R1 = ONE / DBLE( A( K, K ) )
791: CALL ZDSCAL( N-K, R1, A( K+1, K ), 1 )
792: *
1.14 bertrand 793: * (2) Conjugate column W(k)
1.1 bertrand 794: *
795: CALL ZLACGV( N-K, W( K+1, K ), 1 )
796: END IF
1.14 bertrand 797: *
1.1 bertrand 798: ELSE
799: *
800: * 2-by-2 pivot block D(k): columns k and k+1 of W now hold
801: *
802: * ( W(k) W(k+1) ) = ( L(k) L(k+1) )*D(k)
803: *
804: * where L(k) and L(k+1) are the k-th and (k+1)-th columns
805: * of L
806: *
1.14 bertrand 807: * (1) Store L(k+2:N,k) and L(k+2:N,k+1) and 2-by-2
808: * block D(k:k+1,k:k+1) in columns k and k+1 of A.
809: * (NOTE: 2-by-2 diagonal block L(k:k+1,k:k+1) is a UNIT
810: * block and not stored)
811: * A(k:k+1,k:k+1) := D(k:k+1,k:k+1) = W(k:k+1,k:k+1)
812: * A(k+2:N,k:k+1) := L(k+2:N,k:k+1) =
813: * = W(k+2:N,k:k+1) * ( D(k:k+1,k:k+1)**(-1) )
814: *
1.1 bertrand 815: IF( K.LT.N-1 ) THEN
816: *
1.14 bertrand 817: * Factor out the columns of the inverse of 2-by-2 pivot
818: * block D, so that each column contains 1, to reduce the
819: * number of FLOPS when we multiply panel
820: * ( W(kw-1) W(kw) ) by this inverse, i.e. by D**(-1).
821: *
822: * D**(-1) = ( d11 cj(d21) )**(-1) =
823: * ( d21 d22 )
824: *
825: * = 1/(d11*d22-|d21|**2) * ( ( d22) (-cj(d21) ) ) =
826: * ( (-d21) ( d11 ) )
827: *
828: * = 1/(|d21|**2) * 1/((d11/cj(d21))*(d22/d21)-1) *
829: *
830: * * ( d21*( d22/d21 ) conj(d21)*( - 1 ) ) =
831: * ( ( -1 ) ( d11/conj(d21) ) )
832: *
833: * = 1/(|d21|**2) * 1/(D22*D11-1) *
834: *
835: * * ( d21*( D11 ) conj(d21)*( -1 ) ) =
836: * ( ( -1 ) ( D22 ) )
837: *
838: * = (1/|d21|**2) * T * ( d21*( D11 ) conj(d21)*( -1 ) ) =
839: * ( ( -1 ) ( D22 ) )
840: *
841: * = ( (T/conj(d21))*( D11 ) (T/d21)*( -1 ) ) =
842: * ( ( -1 ) ( D22 ) )
843: *
844: * = ( conj(D21)*( D11 ) D21*( -1 ) )
845: * ( ( -1 ) ( D22 ) ),
846: *
847: * where D11 = d22/d21,
848: * D22 = d11/conj(d21),
849: * D21 = T/d21,
850: * T = 1/(D22*D11-1).
851: *
852: * (NOTE: No need to check for division by ZERO,
853: * since that was ensured earlier in pivot search:
854: * (a) d21 != 0, since in 2x2 pivot case(4)
855: * |d21| should be larger than |d11| and |d22|;
856: * (b) (D22*D11 - 1) != 0, since from (a),
857: * both |D11| < 1, |D22| < 1, hence |D22*D11| << 1.)
1.1 bertrand 858: *
859: D21 = W( K+1, K )
860: D11 = W( K+1, K+1 ) / D21
861: D22 = W( K, K ) / DCONJG( D21 )
862: T = ONE / ( DBLE( D11*D22 )-ONE )
863: D21 = T / D21
1.14 bertrand 864: *
865: * Update elements in columns A(k) and A(k+1) as
866: * dot products of rows of ( W(k) W(k+1) ) and columns
867: * of D**(-1)
868: *
1.1 bertrand 869: DO 80 J = K + 2, N
870: A( J, K ) = DCONJG( D21 )*
871: $ ( D11*W( J, K )-W( J, K+1 ) )
872: A( J, K+1 ) = D21*( D22*W( J, K+1 )-W( J, K ) )
873: 80 CONTINUE
874: END IF
875: *
876: * Copy D(k) to A
877: *
878: A( K, K ) = W( K, K )
879: A( K+1, K ) = W( K+1, K )
880: A( K+1, K+1 ) = W( K+1, K+1 )
881: *
1.14 bertrand 882: * (2) Conjugate columns W(k) and W(k+1)
1.1 bertrand 883: *
884: CALL ZLACGV( N-K, W( K+1, K ), 1 )
885: CALL ZLACGV( N-K-1, W( K+2, K+1 ), 1 )
1.14 bertrand 886: *
1.1 bertrand 887: END IF
1.14 bertrand 888: *
1.1 bertrand 889: END IF
890: *
891: * Store details of the interchanges in IPIV
892: *
893: IF( KSTEP.EQ.1 ) THEN
894: IPIV( K ) = KP
895: ELSE
896: IPIV( K ) = -KP
897: IPIV( K+1 ) = -KP
898: END IF
899: *
900: * Increase K and return to the start of the main loop
901: *
902: K = K + KSTEP
903: GO TO 70
904: *
905: 90 CONTINUE
906: *
907: * Update the lower triangle of A22 (= A(k:n,k:n)) as
908: *
1.8 bertrand 909: * A22 := A22 - L21*D*L21**H = A22 - L21*W**H
1.1 bertrand 910: *
911: * computing blocks of NB columns at a time (note that conjg(W) is
912: * actually stored)
913: *
914: DO 110 J = K, N, NB
915: JB = MIN( NB, N-J+1 )
916: *
917: * Update the lower triangle of the diagonal block
918: *
919: DO 100 JJ = J, J + JB - 1
920: A( JJ, JJ ) = DBLE( A( JJ, JJ ) )
921: CALL ZGEMV( 'No transpose', J+JB-JJ, K-1, -CONE,
922: $ A( JJ, 1 ), LDA, W( JJ, 1 ), LDW, CONE,
923: $ A( JJ, JJ ), 1 )
924: A( JJ, JJ ) = DBLE( A( JJ, JJ ) )
925: 100 CONTINUE
926: *
927: * Update the rectangular subdiagonal block
928: *
929: IF( J+JB.LE.N )
930: $ CALL ZGEMM( 'No transpose', 'Transpose', N-J-JB+1, JB,
931: $ K-1, -CONE, A( J+JB, 1 ), LDA, W( J, 1 ),
932: $ LDW, CONE, A( J+JB, J ), LDA )
933: 110 CONTINUE
934: *
935: * Put L21 in standard form by partially undoing the interchanges
1.14 bertrand 936: * of rows in columns 1:k-1 looping backwards from k-1 to 1
1.1 bertrand 937: *
938: J = K - 1
939: 120 CONTINUE
1.14 bertrand 940: *
941: * Undo the interchanges (if any) of rows JJ and JP at each
942: * step J
943: *
944: * (Here, J is a diagonal index)
945: JJ = J
946: JP = IPIV( J )
947: IF( JP.LT.0 ) THEN
948: JP = -JP
949: * (Here, J is a diagonal index)
950: J = J - 1
951: END IF
952: * (NOTE: Here, J is used to determine row length. Length J
953: * of the rows to swap back doesn't include diagonal element)
1.1 bertrand 954: J = J - 1
1.14 bertrand 955: IF( JP.NE.JJ .AND. J.GE.1 )
956: $ CALL ZSWAP( J, A( JP, 1 ), LDA, A( JJ, 1 ), LDA )
957: IF( J.GT.1 )
1.1 bertrand 958: $ GO TO 120
959: *
960: * Set KB to the number of columns factorized
961: *
962: KB = K - 1
963: *
964: END IF
965: RETURN
966: *
967: * End of ZLAHEF
968: *
969: END
CVSweb interface <joel.bertrand@systella.fr>
![[BACK]](/cvsweb/icons/back.gif){kind=icon}
Return to zlahef.f CVS log ![[TXT]](/cvsweb/icons/text.gif){kind=icon}
![[DIR]](/cvsweb/icons/dir.gif){kind=icon}
Up to [local] / rpl / lapack / lapack