1: SUBROUTINE ZGESDD( JOBZ, M, N, A, LDA, S, U, LDU, VT, LDVT, WORK,
2: $ LWORK, RWORK, IWORK, INFO )
3: *
4: * -- LAPACK driver routine (version 3.2) --
5: * -- LAPACK is a software package provided by Univ. of Tennessee, --
6: * -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
7: * November 2006
8: * 8-15-00: Improve consistency of WS calculations (eca)
9: *
10: * .. Scalar Arguments ..
11: CHARACTER JOBZ
12: INTEGER INFO, LDA, LDU, LDVT, LWORK, M, N
13: * ..
14: * .. Array Arguments ..
15: INTEGER IWORK( * )
16: DOUBLE PRECISION RWORK( * ), S( * )
17: COMPLEX*16 A( LDA, * ), U( LDU, * ), VT( LDVT, * ),
18: $ WORK( * )
19: * ..
20: *
21: * Purpose
22: * =======
23: *
24: * ZGESDD computes the singular value decomposition (SVD) of a complex
25: * M-by-N matrix A, optionally computing the left and/or right singular
26: * vectors, by using divide-and-conquer method. The SVD is written
27: *
28: * A = U * SIGMA * conjugate-transpose(V)
29: *
30: * where SIGMA is an M-by-N matrix which is zero except for its
31: * min(m,n) diagonal elements, U is an M-by-M unitary matrix, and
32: * V is an N-by-N unitary matrix. The diagonal elements of SIGMA
33: * are the singular values of A; they are real and non-negative, and
34: * are returned in descending order. The first min(m,n) columns of
35: * U and V are the left and right singular vectors of A.
36: *
37: * Note that the routine returns VT = V**H, not V.
38: *
39: * The divide and conquer algorithm makes very mild assumptions about
40: * floating point arithmetic. It will work on machines with a guard
41: * digit in add/subtract, or on those binary machines without guard
42: * digits which subtract like the Cray X-MP, Cray Y-MP, Cray C-90, or
43: * Cray-2. It could conceivably fail on hexadecimal or decimal machines
44: * without guard digits, but we know of none.
45: *
46: * Arguments
47: * =========
48: *
49: * JOBZ (input) CHARACTER*1
50: * Specifies options for computing all or part of the matrix U:
51: * = 'A': all M columns of U and all N rows of V**H are
52: * returned in the arrays U and VT;
53: * = 'S': the first min(M,N) columns of U and the first
54: * min(M,N) rows of V**H are returned in the arrays U
55: * and VT;
56: * = 'O': If M >= N, the first N columns of U are overwritten
57: * in the array A and all rows of V**H are returned in
58: * the array VT;
59: * otherwise, all columns of U are returned in the
60: * array U and the first M rows of V**H are overwritten
61: * in the array A;
62: * = 'N': no columns of U or rows of V**H are computed.
63: *
64: * M (input) INTEGER
65: * The number of rows of the input matrix A. M >= 0.
66: *
67: * N (input) INTEGER
68: * The number of columns of the input matrix A. N >= 0.
69: *
70: * A (input/output) COMPLEX*16 array, dimension (LDA,N)
71: * On entry, the M-by-N matrix A.
72: * On exit,
73: * if JOBZ = 'O', A is overwritten with the first N columns
74: * of U (the left singular vectors, stored
75: * columnwise) if M >= N;
76: * A is overwritten with the first M rows
77: * of V**H (the right singular vectors, stored
78: * rowwise) otherwise.
79: * if JOBZ .ne. 'O', the contents of A are destroyed.
80: *
81: * LDA (input) INTEGER
82: * The leading dimension of the array A. LDA >= max(1,M).
83: *
84: * S (output) DOUBLE PRECISION array, dimension (min(M,N))
85: * The singular values of A, sorted so that S(i) >= S(i+1).
86: *
87: * U (output) COMPLEX*16 array, dimension (LDU,UCOL)
88: * UCOL = M if JOBZ = 'A' or JOBZ = 'O' and M < N;
89: * UCOL = min(M,N) if JOBZ = 'S'.
90: * If JOBZ = 'A' or JOBZ = 'O' and M < N, U contains the M-by-M
91: * unitary matrix U;
92: * if JOBZ = 'S', U contains the first min(M,N) columns of U
93: * (the left singular vectors, stored columnwise);
94: * if JOBZ = 'O' and M >= N, or JOBZ = 'N', U is not referenced.
95: *
96: * LDU (input) INTEGER
97: * The leading dimension of the array U. LDU >= 1; if
98: * JOBZ = 'S' or 'A' or JOBZ = 'O' and M < N, LDU >= M.
99: *
100: * VT (output) COMPLEX*16 array, dimension (LDVT,N)
101: * If JOBZ = 'A' or JOBZ = 'O' and M >= N, VT contains the
102: * N-by-N unitary matrix V**H;
103: * if JOBZ = 'S', VT contains the first min(M,N) rows of
104: * V**H (the right singular vectors, stored rowwise);
105: * if JOBZ = 'O' and M < N, or JOBZ = 'N', VT is not referenced.
106: *
107: * LDVT (input) INTEGER
108: * The leading dimension of the array VT. LDVT >= 1; if
109: * JOBZ = 'A' or JOBZ = 'O' and M >= N, LDVT >= N;
110: * if JOBZ = 'S', LDVT >= min(M,N).
111: *
112: * WORK (workspace/output) COMPLEX*16 array, dimension (MAX(1,LWORK))
113: * On exit, if INFO = 0, WORK(1) returns the optimal LWORK.
114: *
115: * LWORK (input) INTEGER
116: * The dimension of the array WORK. LWORK >= 1.
117: * if JOBZ = 'N', LWORK >= 2*min(M,N)+max(M,N).
118: * if JOBZ = 'O',
119: * LWORK >= 2*min(M,N)*min(M,N)+2*min(M,N)+max(M,N).
120: * if JOBZ = 'S' or 'A',
121: * LWORK >= min(M,N)*min(M,N)+2*min(M,N)+max(M,N).
122: * For good performance, LWORK should generally be larger.
123: *
124: * If LWORK = -1, a workspace query is assumed. The optimal
125: * size for the WORK array is calculated and stored in WORK(1),
126: * and no other work except argument checking is performed.
127: *
128: * RWORK (workspace) DOUBLE PRECISION array, dimension (MAX(1,LRWORK))
129: * If JOBZ = 'N', LRWORK >= 5*min(M,N).
130: * Otherwise, LRWORK >= 5*min(M,N)*min(M,N) + 7*min(M,N)
131: *
132: * IWORK (workspace) INTEGER array, dimension (8*min(M,N))
133: *
134: * INFO (output) INTEGER
135: * = 0: successful exit.
136: * < 0: if INFO = -i, the i-th argument had an illegal value.
137: * > 0: The updating process of DBDSDC did not converge.
138: *
139: * Further Details
140: * ===============
141: *
142: * Based on contributions by
143: * Ming Gu and Huan Ren, Computer Science Division, University of
144: * California at Berkeley, USA
145: *
146: * =====================================================================
147: *
148: * .. Parameters ..
149: INTEGER LQUERV
150: PARAMETER ( LQUERV = -1 )
151: COMPLEX*16 CZERO, CONE
152: PARAMETER ( CZERO = ( 0.0D+0, 0.0D+0 ),
153: $ CONE = ( 1.0D+0, 0.0D+0 ) )
154: DOUBLE PRECISION ZERO, ONE
155: PARAMETER ( ZERO = 0.0D+0, ONE = 1.0D+0 )
156: * ..
157: * .. Local Scalars ..
158: LOGICAL WNTQA, WNTQAS, WNTQN, WNTQO, WNTQS
159: INTEGER BLK, CHUNK, I, IE, IERR, IL, IR, IRU, IRVT,
160: $ ISCL, ITAU, ITAUP, ITAUQ, IU, IVT, LDWKVT,
161: $ LDWRKL, LDWRKR, LDWRKU, MAXWRK, MINMN, MINWRK,
162: $ MNTHR1, MNTHR2, NRWORK, NWORK, WRKBL
163: DOUBLE PRECISION ANRM, BIGNUM, EPS, SMLNUM
164: * ..
165: * .. Local Arrays ..
166: INTEGER IDUM( 1 )
167: DOUBLE PRECISION DUM( 1 )
168: * ..
169: * .. External Subroutines ..
170: EXTERNAL DBDSDC, DLASCL, XERBLA, ZGEBRD, ZGELQF, ZGEMM,
171: $ ZGEQRF, ZLACP2, ZLACPY, ZLACRM, ZLARCM, ZLASCL,
172: $ ZLASET, ZUNGBR, ZUNGLQ, ZUNGQR, ZUNMBR
173: * ..
174: * .. External Functions ..
175: LOGICAL LSAME
176: INTEGER ILAENV
177: DOUBLE PRECISION DLAMCH, ZLANGE
178: EXTERNAL LSAME, ILAENV, DLAMCH, ZLANGE
179: * ..
180: * .. Intrinsic Functions ..
181: INTRINSIC INT, MAX, MIN, SQRT
182: * ..
183: * .. Executable Statements ..
184: *
185: * Test the input arguments
186: *
187: INFO = 0
188: MINMN = MIN( M, N )
189: MNTHR1 = INT( MINMN*17.0D0 / 9.0D0 )
190: MNTHR2 = INT( MINMN*5.0D0 / 3.0D0 )
191: WNTQA = LSAME( JOBZ, 'A' )
192: WNTQS = LSAME( JOBZ, 'S' )
193: WNTQAS = WNTQA .OR. WNTQS
194: WNTQO = LSAME( JOBZ, 'O' )
195: WNTQN = LSAME( JOBZ, 'N' )
196: MINWRK = 1
197: MAXWRK = 1
198: *
199: IF( .NOT.( WNTQA .OR. WNTQS .OR. WNTQO .OR. WNTQN ) ) THEN
200: INFO = -1
201: ELSE IF( M.LT.0 ) THEN
202: INFO = -2
203: ELSE IF( N.LT.0 ) THEN
204: INFO = -3
205: ELSE IF( LDA.LT.MAX( 1, M ) ) THEN
206: INFO = -5
207: ELSE IF( LDU.LT.1 .OR. ( WNTQAS .AND. LDU.LT.M ) .OR.
208: $ ( WNTQO .AND. M.LT.N .AND. LDU.LT.M ) ) THEN
209: INFO = -8
210: ELSE IF( LDVT.LT.1 .OR. ( WNTQA .AND. LDVT.LT.N ) .OR.
211: $ ( WNTQS .AND. LDVT.LT.MINMN ) .OR.
212: $ ( WNTQO .AND. M.GE.N .AND. LDVT.LT.N ) ) THEN
213: INFO = -10
214: END IF
215: *
216: * Compute workspace
217: * (Note: Comments in the code beginning "Workspace:" describe the
218: * minimal amount of workspace needed at that point in the code,
219: * as well as the preferred amount for good performance.
220: * CWorkspace refers to complex workspace, and RWorkspace to
221: * real workspace. NB refers to the optimal block size for the
222: * immediately following subroutine, as returned by ILAENV.)
223: *
224: IF( INFO.EQ.0 .AND. M.GT.0 .AND. N.GT.0 ) THEN
225: IF( M.GE.N ) THEN
226: *
227: * There is no complex work space needed for bidiagonal SVD
228: * The real work space needed for bidiagonal SVD is BDSPAC
229: * for computing singular values and singular vectors; BDSPAN
230: * for computing singular values only.
231: * BDSPAC = 5*N*N + 7*N
232: * BDSPAN = MAX(7*N+4, 3*N+2+SMLSIZ*(SMLSIZ+8))
233: *
234: IF( M.GE.MNTHR1 ) THEN
235: IF( WNTQN ) THEN
236: *
237: * Path 1 (M much larger than N, JOBZ='N')
238: *
239: MAXWRK = N + N*ILAENV( 1, 'ZGEQRF', ' ', M, N, -1,
240: $ -1 )
241: MAXWRK = MAX( MAXWRK, 2*N+2*N*
242: $ ILAENV( 1, 'ZGEBRD', ' ', N, N, -1, -1 ) )
243: MINWRK = 3*N
244: ELSE IF( WNTQO ) THEN
245: *
246: * Path 2 (M much larger than N, JOBZ='O')
247: *
248: WRKBL = N + N*ILAENV( 1, 'ZGEQRF', ' ', M, N, -1, -1 )
249: WRKBL = MAX( WRKBL, N+N*ILAENV( 1, 'ZUNGQR', ' ', M,
250: $ N, N, -1 ) )
251: WRKBL = MAX( WRKBL, 2*N+2*N*
252: $ ILAENV( 1, 'ZGEBRD', ' ', N, N, -1, -1 ) )
253: WRKBL = MAX( WRKBL, 2*N+N*
254: $ ILAENV( 1, 'ZUNMBR', 'QLN', N, N, N, -1 ) )
255: WRKBL = MAX( WRKBL, 2*N+N*
256: $ ILAENV( 1, 'ZUNMBR', 'PRC', N, N, N, -1 ) )
257: MAXWRK = M*N + N*N + WRKBL
258: MINWRK = 2*N*N + 3*N
259: ELSE IF( WNTQS ) THEN
260: *
261: * Path 3 (M much larger than N, JOBZ='S')
262: *
263: WRKBL = N + N*ILAENV( 1, 'ZGEQRF', ' ', M, N, -1, -1 )
264: WRKBL = MAX( WRKBL, N+N*ILAENV( 1, 'ZUNGQR', ' ', M,
265: $ N, N, -1 ) )
266: WRKBL = MAX( WRKBL, 2*N+2*N*
267: $ ILAENV( 1, 'ZGEBRD', ' ', N, N, -1, -1 ) )
268: WRKBL = MAX( WRKBL, 2*N+N*
269: $ ILAENV( 1, 'ZUNMBR', 'QLN', N, N, N, -1 ) )
270: WRKBL = MAX( WRKBL, 2*N+N*
271: $ ILAENV( 1, 'ZUNMBR', 'PRC', N, N, N, -1 ) )
272: MAXWRK = N*N + WRKBL
273: MINWRK = N*N + 3*N
274: ELSE IF( WNTQA ) THEN
275: *
276: * Path 4 (M much larger than N, JOBZ='A')
277: *
278: WRKBL = N + N*ILAENV( 1, 'ZGEQRF', ' ', M, N, -1, -1 )
279: WRKBL = MAX( WRKBL, N+M*ILAENV( 1, 'ZUNGQR', ' ', M,
280: $ M, N, -1 ) )
281: WRKBL = MAX( WRKBL, 2*N+2*N*
282: $ ILAENV( 1, 'ZGEBRD', ' ', N, N, -1, -1 ) )
283: WRKBL = MAX( WRKBL, 2*N+N*
284: $ ILAENV( 1, 'ZUNMBR', 'QLN', N, N, N, -1 ) )
285: WRKBL = MAX( WRKBL, 2*N+N*
286: $ ILAENV( 1, 'ZUNMBR', 'PRC', N, N, N, -1 ) )
287: MAXWRK = N*N + WRKBL
288: MINWRK = N*N + 2*N + M
289: END IF
290: ELSE IF( M.GE.MNTHR2 ) THEN
291: *
292: * Path 5 (M much larger than N, but not as much as MNTHR1)
293: *
294: MAXWRK = 2*N + ( M+N )*ILAENV( 1, 'ZGEBRD', ' ', M, N,
295: $ -1, -1 )
296: MINWRK = 2*N + M
297: IF( WNTQO ) THEN
298: MAXWRK = MAX( MAXWRK, 2*N+N*
299: $ ILAENV( 1, 'ZUNGBR', 'P', N, N, N, -1 ) )
300: MAXWRK = MAX( MAXWRK, 2*N+N*
301: $ ILAENV( 1, 'ZUNGBR', 'Q', M, N, N, -1 ) )
302: MAXWRK = MAXWRK + M*N
303: MINWRK = MINWRK + N*N
304: ELSE IF( WNTQS ) THEN
305: MAXWRK = MAX( MAXWRK, 2*N+N*
306: $ ILAENV( 1, 'ZUNGBR', 'P', N, N, N, -1 ) )
307: MAXWRK = MAX( MAXWRK, 2*N+N*
308: $ ILAENV( 1, 'ZUNGBR', 'Q', M, N, N, -1 ) )
309: ELSE IF( WNTQA ) THEN
310: MAXWRK = MAX( MAXWRK, 2*N+N*
311: $ ILAENV( 1, 'ZUNGBR', 'P', N, N, N, -1 ) )
312: MAXWRK = MAX( MAXWRK, 2*N+M*
313: $ ILAENV( 1, 'ZUNGBR', 'Q', M, M, N, -1 ) )
314: END IF
315: ELSE
316: *
317: * Path 6 (M at least N, but not much larger)
318: *
319: MAXWRK = 2*N + ( M+N )*ILAENV( 1, 'ZGEBRD', ' ', M, N,
320: $ -1, -1 )
321: MINWRK = 2*N + M
322: IF( WNTQO ) THEN
323: MAXWRK = MAX( MAXWRK, 2*N+N*
324: $ ILAENV( 1, 'ZUNMBR', 'PRC', N, N, N, -1 ) )
325: MAXWRK = MAX( MAXWRK, 2*N+N*
326: $ ILAENV( 1, 'ZUNMBR', 'QLN', M, N, N, -1 ) )
327: MAXWRK = MAXWRK + M*N
328: MINWRK = MINWRK + N*N
329: ELSE IF( WNTQS ) THEN
330: MAXWRK = MAX( MAXWRK, 2*N+N*
331: $ ILAENV( 1, 'ZUNMBR', 'PRC', N, N, N, -1 ) )
332: MAXWRK = MAX( MAXWRK, 2*N+N*
333: $ ILAENV( 1, 'ZUNMBR', 'QLN', M, N, N, -1 ) )
334: ELSE IF( WNTQA ) THEN
335: MAXWRK = MAX( MAXWRK, 2*N+N*
336: $ ILAENV( 1, 'ZUNGBR', 'PRC', N, N, N, -1 ) )
337: MAXWRK = MAX( MAXWRK, 2*N+M*
338: $ ILAENV( 1, 'ZUNGBR', 'QLN', M, M, N, -1 ) )
339: END IF
340: END IF
341: ELSE
342: *
343: * There is no complex work space needed for bidiagonal SVD
344: * The real work space needed for bidiagonal SVD is BDSPAC
345: * for computing singular values and singular vectors; BDSPAN
346: * for computing singular values only.
347: * BDSPAC = 5*M*M + 7*M
348: * BDSPAN = MAX(7*M+4, 3*M+2+SMLSIZ*(SMLSIZ+8))
349: *
350: IF( N.GE.MNTHR1 ) THEN
351: IF( WNTQN ) THEN
352: *
353: * Path 1t (N much larger than M, JOBZ='N')
354: *
355: MAXWRK = M + M*ILAENV( 1, 'ZGELQF', ' ', M, N, -1,
356: $ -1 )
357: MAXWRK = MAX( MAXWRK, 2*M+2*M*
358: $ ILAENV( 1, 'ZGEBRD', ' ', M, M, -1, -1 ) )
359: MINWRK = 3*M
360: ELSE IF( WNTQO ) THEN
361: *
362: * Path 2t (N much larger than M, JOBZ='O')
363: *
364: WRKBL = M + M*ILAENV( 1, 'ZGELQF', ' ', M, N, -1, -1 )
365: WRKBL = MAX( WRKBL, M+M*ILAENV( 1, 'ZUNGLQ', ' ', M,
366: $ N, M, -1 ) )
367: WRKBL = MAX( WRKBL, 2*M+2*M*
368: $ ILAENV( 1, 'ZGEBRD', ' ', M, M, -1, -1 ) )
369: WRKBL = MAX( WRKBL, 2*M+M*
370: $ ILAENV( 1, 'ZUNMBR', 'PRC', M, M, M, -1 ) )
371: WRKBL = MAX( WRKBL, 2*M+M*
372: $ ILAENV( 1, 'ZUNMBR', 'QLN', M, M, M, -1 ) )
373: MAXWRK = M*N + M*M + WRKBL
374: MINWRK = 2*M*M + 3*M
375: ELSE IF( WNTQS ) THEN
376: *
377: * Path 3t (N much larger than M, JOBZ='S')
378: *
379: WRKBL = M + M*ILAENV( 1, 'ZGELQF', ' ', M, N, -1, -1 )
380: WRKBL = MAX( WRKBL, M+M*ILAENV( 1, 'ZUNGLQ', ' ', M,
381: $ N, M, -1 ) )
382: WRKBL = MAX( WRKBL, 2*M+2*M*
383: $ ILAENV( 1, 'ZGEBRD', ' ', M, M, -1, -1 ) )
384: WRKBL = MAX( WRKBL, 2*M+M*
385: $ ILAENV( 1, 'ZUNMBR', 'PRC', M, M, M, -1 ) )
386: WRKBL = MAX( WRKBL, 2*M+M*
387: $ ILAENV( 1, 'ZUNMBR', 'QLN', M, M, M, -1 ) )
388: MAXWRK = M*M + WRKBL
389: MINWRK = M*M + 3*M
390: ELSE IF( WNTQA ) THEN
391: *
392: * Path 4t (N much larger than M, JOBZ='A')
393: *
394: WRKBL = M + M*ILAENV( 1, 'ZGELQF', ' ', M, N, -1, -1 )
395: WRKBL = MAX( WRKBL, M+N*ILAENV( 1, 'ZUNGLQ', ' ', N,
396: $ N, M, -1 ) )
397: WRKBL = MAX( WRKBL, 2*M+2*M*
398: $ ILAENV( 1, 'ZGEBRD', ' ', M, M, -1, -1 ) )
399: WRKBL = MAX( WRKBL, 2*M+M*
400: $ ILAENV( 1, 'ZUNMBR', 'PRC', M, M, M, -1 ) )
401: WRKBL = MAX( WRKBL, 2*M+M*
402: $ ILAENV( 1, 'ZUNMBR', 'QLN', M, M, M, -1 ) )
403: MAXWRK = M*M + WRKBL
404: MINWRK = M*M + 2*M + N
405: END IF
406: ELSE IF( N.GE.MNTHR2 ) THEN
407: *
408: * Path 5t (N much larger than M, but not as much as MNTHR1)
409: *
410: MAXWRK = 2*M + ( M+N )*ILAENV( 1, 'ZGEBRD', ' ', M, N,
411: $ -1, -1 )
412: MINWRK = 2*M + N
413: IF( WNTQO ) THEN
414: MAXWRK = MAX( MAXWRK, 2*M+M*
415: $ ILAENV( 1, 'ZUNGBR', 'P', M, N, M, -1 ) )
416: MAXWRK = MAX( MAXWRK, 2*M+M*
417: $ ILAENV( 1, 'ZUNGBR', 'Q', M, M, N, -1 ) )
418: MAXWRK = MAXWRK + M*N
419: MINWRK = MINWRK + M*M
420: ELSE IF( WNTQS ) THEN
421: MAXWRK = MAX( MAXWRK, 2*M+M*
422: $ ILAENV( 1, 'ZUNGBR', 'P', M, N, M, -1 ) )
423: MAXWRK = MAX( MAXWRK, 2*M+M*
424: $ ILAENV( 1, 'ZUNGBR', 'Q', M, M, N, -1 ) )
425: ELSE IF( WNTQA ) THEN
426: MAXWRK = MAX( MAXWRK, 2*M+N*
427: $ ILAENV( 1, 'ZUNGBR', 'P', N, N, M, -1 ) )
428: MAXWRK = MAX( MAXWRK, 2*M+M*
429: $ ILAENV( 1, 'ZUNGBR', 'Q', M, M, N, -1 ) )
430: END IF
431: ELSE
432: *
433: * Path 6t (N greater than M, but not much larger)
434: *
435: MAXWRK = 2*M + ( M+N )*ILAENV( 1, 'ZGEBRD', ' ', M, N,
436: $ -1, -1 )
437: MINWRK = 2*M + N
438: IF( WNTQO ) THEN
439: MAXWRK = MAX( MAXWRK, 2*M+M*
440: $ ILAENV( 1, 'ZUNMBR', 'PRC', M, N, M, -1 ) )
441: MAXWRK = MAX( MAXWRK, 2*M+M*
442: $ ILAENV( 1, 'ZUNMBR', 'QLN', M, M, N, -1 ) )
443: MAXWRK = MAXWRK + M*N
444: MINWRK = MINWRK + M*M
445: ELSE IF( WNTQS ) THEN
446: MAXWRK = MAX( MAXWRK, 2*M+M*
447: $ ILAENV( 1, 'ZUNGBR', 'PRC', M, N, M, -1 ) )
448: MAXWRK = MAX( MAXWRK, 2*M+M*
449: $ ILAENV( 1, 'ZUNGBR', 'QLN', M, M, N, -1 ) )
450: ELSE IF( WNTQA ) THEN
451: MAXWRK = MAX( MAXWRK, 2*M+N*
452: $ ILAENV( 1, 'ZUNGBR', 'PRC', N, N, M, -1 ) )
453: MAXWRK = MAX( MAXWRK, 2*M+M*
454: $ ILAENV( 1, 'ZUNGBR', 'QLN', M, M, N, -1 ) )
455: END IF
456: END IF
457: END IF
458: MAXWRK = MAX( MAXWRK, MINWRK )
459: END IF
460: IF( INFO.EQ.0 ) THEN
461: WORK( 1 ) = MAXWRK
462: IF( LWORK.LT.MINWRK .AND. LWORK.NE.LQUERV )
463: $ INFO = -13
464: END IF
465: *
466: * Quick returns
467: *
468: IF( INFO.NE.0 ) THEN
469: CALL XERBLA( 'ZGESDD', -INFO )
470: RETURN
471: END IF
472: IF( LWORK.EQ.LQUERV )
473: $ RETURN
474: IF( M.EQ.0 .OR. N.EQ.0 ) THEN
475: RETURN
476: END IF
477: *
478: * Get machine constants
479: *
480: EPS = DLAMCH( 'P' )
481: SMLNUM = SQRT( DLAMCH( 'S' ) ) / EPS
482: BIGNUM = ONE / SMLNUM
483: *
484: * Scale A if max element outside range [SMLNUM,BIGNUM]
485: *
486: ANRM = ZLANGE( 'M', M, N, A, LDA, DUM )
487: ISCL = 0
488: IF( ANRM.GT.ZERO .AND. ANRM.LT.SMLNUM ) THEN
489: ISCL = 1
490: CALL ZLASCL( 'G', 0, 0, ANRM, SMLNUM, M, N, A, LDA, IERR )
491: ELSE IF( ANRM.GT.BIGNUM ) THEN
492: ISCL = 1
493: CALL ZLASCL( 'G', 0, 0, ANRM, BIGNUM, M, N, A, LDA, IERR )
494: END IF
495: *
496: IF( M.GE.N ) THEN
497: *
498: * A has at least as many rows as columns. If A has sufficiently
499: * more rows than columns, first reduce using the QR
500: * decomposition (if sufficient workspace available)
501: *
502: IF( M.GE.MNTHR1 ) THEN
503: *
504: IF( WNTQN ) THEN
505: *
506: * Path 1 (M much larger than N, JOBZ='N')
507: * No singular vectors to be computed
508: *
509: ITAU = 1
510: NWORK = ITAU + N
511: *
512: * Compute A=Q*R
513: * (CWorkspace: need 2*N, prefer N+N*NB)
514: * (RWorkspace: need 0)
515: *
516: CALL ZGEQRF( M, N, A, LDA, WORK( ITAU ), WORK( NWORK ),
517: $ LWORK-NWORK+1, IERR )
518: *
519: * Zero out below R
520: *
521: CALL ZLASET( 'L', N-1, N-1, CZERO, CZERO, A( 2, 1 ),
522: $ LDA )
523: IE = 1
524: ITAUQ = 1
525: ITAUP = ITAUQ + N
526: NWORK = ITAUP + N
527: *
528: * Bidiagonalize R in A
529: * (CWorkspace: need 3*N, prefer 2*N+2*N*NB)
530: * (RWorkspace: need N)
531: *
532: CALL ZGEBRD( N, N, A, LDA, S, RWORK( IE ), WORK( ITAUQ ),
533: $ WORK( ITAUP ), WORK( NWORK ), LWORK-NWORK+1,
534: $ IERR )
535: NRWORK = IE + N
536: *
537: * Perform bidiagonal SVD, compute singular values only
538: * (CWorkspace: 0)
539: * (RWorkspace: need BDSPAN)
540: *
541: CALL DBDSDC( 'U', 'N', N, S, RWORK( IE ), DUM, 1, DUM, 1,
542: $ DUM, IDUM, RWORK( NRWORK ), IWORK, INFO )
543: *
544: ELSE IF( WNTQO ) THEN
545: *
546: * Path 2 (M much larger than N, JOBZ='O')
547: * N left singular vectors to be overwritten on A and
548: * N right singular vectors to be computed in VT
549: *
550: IU = 1
551: *
552: * WORK(IU) is N by N
553: *
554: LDWRKU = N
555: IR = IU + LDWRKU*N
556: IF( LWORK.GE.M*N+N*N+3*N ) THEN
557: *
558: * WORK(IR) is M by N
559: *
560: LDWRKR = M
561: ELSE
562: LDWRKR = ( LWORK-N*N-3*N ) / N
563: END IF
564: ITAU = IR + LDWRKR*N
565: NWORK = ITAU + N
566: *
567: * Compute A=Q*R
568: * (CWorkspace: need N*N+2*N, prefer M*N+N+N*NB)
569: * (RWorkspace: 0)
570: *
571: CALL ZGEQRF( M, N, A, LDA, WORK( ITAU ), WORK( NWORK ),
572: $ LWORK-NWORK+1, IERR )
573: *
574: * Copy R to WORK( IR ), zeroing out below it
575: *
576: CALL ZLACPY( 'U', N, N, A, LDA, WORK( IR ), LDWRKR )
577: CALL ZLASET( 'L', N-1, N-1, CZERO, CZERO, WORK( IR+1 ),
578: $ LDWRKR )
579: *
580: * Generate Q in A
581: * (CWorkspace: need 2*N, prefer N+N*NB)
582: * (RWorkspace: 0)
583: *
584: CALL ZUNGQR( M, N, N, A, LDA, WORK( ITAU ),
585: $ WORK( NWORK ), LWORK-NWORK+1, IERR )
586: IE = 1
587: ITAUQ = ITAU
588: ITAUP = ITAUQ + N
589: NWORK = ITAUP + N
590: *
591: * Bidiagonalize R in WORK(IR)
592: * (CWorkspace: need N*N+3*N, prefer M*N+2*N+2*N*NB)
593: * (RWorkspace: need N)
594: *
595: CALL ZGEBRD( N, N, WORK( IR ), LDWRKR, S, RWORK( IE ),
596: $ WORK( ITAUQ ), WORK( ITAUP ), WORK( NWORK ),
597: $ LWORK-NWORK+1, IERR )
598: *
599: * Perform bidiagonal SVD, computing left singular vectors
600: * of R in WORK(IRU) and computing right singular vectors
601: * of R in WORK(IRVT)
602: * (CWorkspace: need 0)
603: * (RWorkspace: need BDSPAC)
604: *
605: IRU = IE + N
606: IRVT = IRU + N*N
607: NRWORK = IRVT + N*N
608: CALL DBDSDC( 'U', 'I', N, S, RWORK( IE ), RWORK( IRU ),
609: $ N, RWORK( IRVT ), N, DUM, IDUM,
610: $ RWORK( NRWORK ), IWORK, INFO )
611: *
612: * Copy real matrix RWORK(IRU) to complex matrix WORK(IU)
613: * Overwrite WORK(IU) by the left singular vectors of R
614: * (CWorkspace: need 2*N*N+3*N, prefer M*N+N*N+2*N+N*NB)
615: * (RWorkspace: 0)
616: *
617: CALL ZLACP2( 'F', N, N, RWORK( IRU ), N, WORK( IU ),
618: $ LDWRKU )
619: CALL ZUNMBR( 'Q', 'L', 'N', N, N, N, WORK( IR ), LDWRKR,
620: $ WORK( ITAUQ ), WORK( IU ), LDWRKU,
621: $ WORK( NWORK ), LWORK-NWORK+1, IERR )
622: *
623: * Copy real matrix RWORK(IRVT) to complex matrix VT
624: * Overwrite VT by the right singular vectors of R
625: * (CWorkspace: need N*N+3*N, prefer M*N+2*N+N*NB)
626: * (RWorkspace: 0)
627: *
628: CALL ZLACP2( 'F', N, N, RWORK( IRVT ), N, VT, LDVT )
629: CALL ZUNMBR( 'P', 'R', 'C', N, N, N, WORK( IR ), LDWRKR,
630: $ WORK( ITAUP ), VT, LDVT, WORK( NWORK ),
631: $ LWORK-NWORK+1, IERR )
632: *
633: * Multiply Q in A by left singular vectors of R in
634: * WORK(IU), storing result in WORK(IR) and copying to A
635: * (CWorkspace: need 2*N*N, prefer N*N+M*N)
636: * (RWorkspace: 0)
637: *
638: DO 10 I = 1, M, LDWRKR
639: CHUNK = MIN( M-I+1, LDWRKR )
640: CALL ZGEMM( 'N', 'N', CHUNK, N, N, CONE, A( I, 1 ),
641: $ LDA, WORK( IU ), LDWRKU, CZERO,
642: $ WORK( IR ), LDWRKR )
643: CALL ZLACPY( 'F', CHUNK, N, WORK( IR ), LDWRKR,
644: $ A( I, 1 ), LDA )
645: 10 CONTINUE
646: *
647: ELSE IF( WNTQS ) THEN
648: *
649: * Path 3 (M much larger than N, JOBZ='S')
650: * N left singular vectors to be computed in U and
651: * N right singular vectors to be computed in VT
652: *
653: IR = 1
654: *
655: * WORK(IR) is N by N
656: *
657: LDWRKR = N
658: ITAU = IR + LDWRKR*N
659: NWORK = ITAU + N
660: *
661: * Compute A=Q*R
662: * (CWorkspace: need N*N+2*N, prefer N*N+N+N*NB)
663: * (RWorkspace: 0)
664: *
665: CALL ZGEQRF( M, N, A, LDA, WORK( ITAU ), WORK( NWORK ),
666: $ LWORK-NWORK+1, IERR )
667: *
668: * Copy R to WORK(IR), zeroing out below it
669: *
670: CALL ZLACPY( 'U', N, N, A, LDA, WORK( IR ), LDWRKR )
671: CALL ZLASET( 'L', N-1, N-1, CZERO, CZERO, WORK( IR+1 ),
672: $ LDWRKR )
673: *
674: * Generate Q in A
675: * (CWorkspace: need 2*N, prefer N+N*NB)
676: * (RWorkspace: 0)
677: *
678: CALL ZUNGQR( M, N, N, A, LDA, WORK( ITAU ),
679: $ WORK( NWORK ), LWORK-NWORK+1, IERR )
680: IE = 1
681: ITAUQ = ITAU
682: ITAUP = ITAUQ + N
683: NWORK = ITAUP + N
684: *
685: * Bidiagonalize R in WORK(IR)
686: * (CWorkspace: need N*N+3*N, prefer N*N+2*N+2*N*NB)
687: * (RWorkspace: need N)
688: *
689: CALL ZGEBRD( N, N, WORK( IR ), LDWRKR, S, RWORK( IE ),
690: $ WORK( ITAUQ ), WORK( ITAUP ), WORK( NWORK ),
691: $ LWORK-NWORK+1, IERR )
692: *
693: * Perform bidiagonal SVD, computing left singular vectors
694: * of bidiagonal matrix in RWORK(IRU) and computing right
695: * singular vectors of bidiagonal matrix in RWORK(IRVT)
696: * (CWorkspace: need 0)
697: * (RWorkspace: need BDSPAC)
698: *
699: IRU = IE + N
700: IRVT = IRU + N*N
701: NRWORK = IRVT + N*N
702: CALL DBDSDC( 'U', 'I', N, S, RWORK( IE ), RWORK( IRU ),
703: $ N, RWORK( IRVT ), N, DUM, IDUM,
704: $ RWORK( NRWORK ), IWORK, INFO )
705: *
706: * Copy real matrix RWORK(IRU) to complex matrix U
707: * Overwrite U by left singular vectors of R
708: * (CWorkspace: need N*N+3*N, prefer N*N+2*N+N*NB)
709: * (RWorkspace: 0)
710: *
711: CALL ZLACP2( 'F', N, N, RWORK( IRU ), N, U, LDU )
712: CALL ZUNMBR( 'Q', 'L', 'N', N, N, N, WORK( IR ), LDWRKR,
713: $ WORK( ITAUQ ), U, LDU, WORK( NWORK ),
714: $ LWORK-NWORK+1, IERR )
715: *
716: * Copy real matrix RWORK(IRVT) to complex matrix VT
717: * Overwrite VT by right singular vectors of R
718: * (CWorkspace: need N*N+3*N, prefer N*N+2*N+N*NB)
719: * (RWorkspace: 0)
720: *
721: CALL ZLACP2( 'F', N, N, RWORK( IRVT ), N, VT, LDVT )
722: CALL ZUNMBR( 'P', 'R', 'C', N, N, N, WORK( IR ), LDWRKR,
723: $ WORK( ITAUP ), VT, LDVT, WORK( NWORK ),
724: $ LWORK-NWORK+1, IERR )
725: *
726: * Multiply Q in A by left singular vectors of R in
727: * WORK(IR), storing result in U
728: * (CWorkspace: need N*N)
729: * (RWorkspace: 0)
730: *
731: CALL ZLACPY( 'F', N, N, U, LDU, WORK( IR ), LDWRKR )
732: CALL ZGEMM( 'N', 'N', M, N, N, CONE, A, LDA, WORK( IR ),
733: $ LDWRKR, CZERO, U, LDU )
734: *
735: ELSE IF( WNTQA ) THEN
736: *
737: * Path 4 (M much larger than N, JOBZ='A')
738: * M left singular vectors to be computed in U and
739: * N right singular vectors to be computed in VT
740: *
741: IU = 1
742: *
743: * WORK(IU) is N by N
744: *
745: LDWRKU = N
746: ITAU = IU + LDWRKU*N
747: NWORK = ITAU + N
748: *
749: * Compute A=Q*R, copying result to U
750: * (CWorkspace: need 2*N, prefer N+N*NB)
751: * (RWorkspace: 0)
752: *
753: CALL ZGEQRF( M, N, A, LDA, WORK( ITAU ), WORK( NWORK ),
754: $ LWORK-NWORK+1, IERR )
755: CALL ZLACPY( 'L', M, N, A, LDA, U, LDU )
756: *
757: * Generate Q in U
758: * (CWorkspace: need N+M, prefer N+M*NB)
759: * (RWorkspace: 0)
760: *
761: CALL ZUNGQR( M, M, N, U, LDU, WORK( ITAU ),
762: $ WORK( NWORK ), LWORK-NWORK+1, IERR )
763: *
764: * Produce R in A, zeroing out below it
765: *
766: CALL ZLASET( 'L', N-1, N-1, CZERO, CZERO, A( 2, 1 ),
767: $ LDA )
768: IE = 1
769: ITAUQ = ITAU
770: ITAUP = ITAUQ + N
771: NWORK = ITAUP + N
772: *
773: * Bidiagonalize R in A
774: * (CWorkspace: need 3*N, prefer 2*N+2*N*NB)
775: * (RWorkspace: need N)
776: *
777: CALL ZGEBRD( N, N, A, LDA, S, RWORK( IE ), WORK( ITAUQ ),
778: $ WORK( ITAUP ), WORK( NWORK ), LWORK-NWORK+1,
779: $ IERR )
780: IRU = IE + N
781: IRVT = IRU + N*N
782: NRWORK = IRVT + N*N
783: *
784: * Perform bidiagonal SVD, computing left singular vectors
785: * of bidiagonal matrix in RWORK(IRU) and computing right
786: * singular vectors of bidiagonal matrix in RWORK(IRVT)
787: * (CWorkspace: need 0)
788: * (RWorkspace: need BDSPAC)
789: *
790: CALL DBDSDC( 'U', 'I', N, S, RWORK( IE ), RWORK( IRU ),
791: $ N, RWORK( IRVT ), N, DUM, IDUM,
792: $ RWORK( NRWORK ), IWORK, INFO )
793: *
794: * Copy real matrix RWORK(IRU) to complex matrix WORK(IU)
795: * Overwrite WORK(IU) by left singular vectors of R
796: * (CWorkspace: need N*N+3*N, prefer N*N+2*N+N*NB)
797: * (RWorkspace: 0)
798: *
799: CALL ZLACP2( 'F', N, N, RWORK( IRU ), N, WORK( IU ),
800: $ LDWRKU )
801: CALL ZUNMBR( 'Q', 'L', 'N', N, N, N, A, LDA,
802: $ WORK( ITAUQ ), WORK( IU ), LDWRKU,
803: $ WORK( NWORK ), LWORK-NWORK+1, IERR )
804: *
805: * Copy real matrix RWORK(IRVT) to complex matrix VT
806: * Overwrite VT by right singular vectors of R
807: * (CWorkspace: need 3*N, prefer 2*N+N*NB)
808: * (RWorkspace: 0)
809: *
810: CALL ZLACP2( 'F', N, N, RWORK( IRVT ), N, VT, LDVT )
811: CALL ZUNMBR( 'P', 'R', 'C', N, N, N, A, LDA,
812: $ WORK( ITAUP ), VT, LDVT, WORK( NWORK ),
813: $ LWORK-NWORK+1, IERR )
814: *
815: * Multiply Q in U by left singular vectors of R in
816: * WORK(IU), storing result in A
817: * (CWorkspace: need N*N)
818: * (RWorkspace: 0)
819: *
820: CALL ZGEMM( 'N', 'N', M, N, N, CONE, U, LDU, WORK( IU ),
821: $ LDWRKU, CZERO, A, LDA )
822: *
823: * Copy left singular vectors of A from A to U
824: *
825: CALL ZLACPY( 'F', M, N, A, LDA, U, LDU )
826: *
827: END IF
828: *
829: ELSE IF( M.GE.MNTHR2 ) THEN
830: *
831: * MNTHR2 <= M < MNTHR1
832: *
833: * Path 5 (M much larger than N, but not as much as MNTHR1)
834: * Reduce to bidiagonal form without QR decomposition, use
835: * ZUNGBR and matrix multiplication to compute singular vectors
836: *
837: IE = 1
838: NRWORK = IE + N
839: ITAUQ = 1
840: ITAUP = ITAUQ + N
841: NWORK = ITAUP + N
842: *
843: * Bidiagonalize A
844: * (CWorkspace: need 2*N+M, prefer 2*N+(M+N)*NB)
845: * (RWorkspace: need N)
846: *
847: CALL ZGEBRD( M, N, A, LDA, S, RWORK( IE ), WORK( ITAUQ ),
848: $ WORK( ITAUP ), WORK( NWORK ), LWORK-NWORK+1,
849: $ IERR )
850: IF( WNTQN ) THEN
851: *
852: * Compute singular values only
853: * (Cworkspace: 0)
854: * (Rworkspace: need BDSPAN)
855: *
856: CALL DBDSDC( 'U', 'N', N, S, RWORK( IE ), DUM, 1, DUM, 1,
857: $ DUM, IDUM, RWORK( NRWORK ), IWORK, INFO )
858: ELSE IF( WNTQO ) THEN
859: IU = NWORK
860: IRU = NRWORK
861: IRVT = IRU + N*N
862: NRWORK = IRVT + N*N
863: *
864: * Copy A to VT, generate P**H
865: * (Cworkspace: need 2*N, prefer N+N*NB)
866: * (Rworkspace: 0)
867: *
868: CALL ZLACPY( 'U', N, N, A, LDA, VT, LDVT )
869: CALL ZUNGBR( 'P', N, N, N, VT, LDVT, WORK( ITAUP ),
870: $ WORK( NWORK ), LWORK-NWORK+1, IERR )
871: *
872: * Generate Q in A
873: * (CWorkspace: need 2*N, prefer N+N*NB)
874: * (RWorkspace: 0)
875: *
876: CALL ZUNGBR( 'Q', M, N, N, A, LDA, WORK( ITAUQ ),
877: $ WORK( NWORK ), LWORK-NWORK+1, IERR )
878: *
879: IF( LWORK.GE.M*N+3*N ) THEN
880: *
881: * WORK( IU ) is M by N
882: *
883: LDWRKU = M
884: ELSE
885: *
886: * WORK(IU) is LDWRKU by N
887: *
888: LDWRKU = ( LWORK-3*N ) / N
889: END IF
890: NWORK = IU + LDWRKU*N
891: *
892: * Perform bidiagonal SVD, computing left singular vectors
893: * of bidiagonal matrix in RWORK(IRU) and computing right
894: * singular vectors of bidiagonal matrix in RWORK(IRVT)
895: * (CWorkspace: need 0)
896: * (RWorkspace: need BDSPAC)
897: *
898: CALL DBDSDC( 'U', 'I', N, S, RWORK( IE ), RWORK( IRU ),
899: $ N, RWORK( IRVT ), N, DUM, IDUM,
900: $ RWORK( NRWORK ), IWORK, INFO )
901: *
902: * Multiply real matrix RWORK(IRVT) by P**H in VT,
903: * storing the result in WORK(IU), copying to VT
904: * (Cworkspace: need 0)
905: * (Rworkspace: need 3*N*N)
906: *
907: CALL ZLARCM( N, N, RWORK( IRVT ), N, VT, LDVT,
908: $ WORK( IU ), LDWRKU, RWORK( NRWORK ) )
909: CALL ZLACPY( 'F', N, N, WORK( IU ), LDWRKU, VT, LDVT )
910: *
911: * Multiply Q in A by real matrix RWORK(IRU), storing the
912: * result in WORK(IU), copying to A
913: * (CWorkspace: need N*N, prefer M*N)
914: * (Rworkspace: need 3*N*N, prefer N*N+2*M*N)
915: *
916: NRWORK = IRVT
917: DO 20 I = 1, M, LDWRKU
918: CHUNK = MIN( M-I+1, LDWRKU )
919: CALL ZLACRM( CHUNK, N, A( I, 1 ), LDA, RWORK( IRU ),
920: $ N, WORK( IU ), LDWRKU, RWORK( NRWORK ) )
921: CALL ZLACPY( 'F', CHUNK, N, WORK( IU ), LDWRKU,
922: $ A( I, 1 ), LDA )
923: 20 CONTINUE
924: *
925: ELSE IF( WNTQS ) THEN
926: *
927: * Copy A to VT, generate P**H
928: * (Cworkspace: need 2*N, prefer N+N*NB)
929: * (Rworkspace: 0)
930: *
931: CALL ZLACPY( 'U', N, N, A, LDA, VT, LDVT )
932: CALL ZUNGBR( 'P', N, N, N, VT, LDVT, WORK( ITAUP ),
933: $ WORK( NWORK ), LWORK-NWORK+1, IERR )
934: *
935: * Copy A to U, generate Q
936: * (Cworkspace: need 2*N, prefer N+N*NB)
937: * (Rworkspace: 0)
938: *
939: CALL ZLACPY( 'L', M, N, A, LDA, U, LDU )
940: CALL ZUNGBR( 'Q', M, N, N, U, LDU, WORK( ITAUQ ),
941: $ WORK( NWORK ), LWORK-NWORK+1, IERR )
942: *
943: * Perform bidiagonal SVD, computing left singular vectors
944: * of bidiagonal matrix in RWORK(IRU) and computing right
945: * singular vectors of bidiagonal matrix in RWORK(IRVT)
946: * (CWorkspace: need 0)
947: * (RWorkspace: need BDSPAC)
948: *
949: IRU = NRWORK
950: IRVT = IRU + N*N
951: NRWORK = IRVT + N*N
952: CALL DBDSDC( 'U', 'I', N, S, RWORK( IE ), RWORK( IRU ),
953: $ N, RWORK( IRVT ), N, DUM, IDUM,
954: $ RWORK( NRWORK ), IWORK, INFO )
955: *
956: * Multiply real matrix RWORK(IRVT) by P**H in VT,
957: * storing the result in A, copying to VT
958: * (Cworkspace: need 0)
959: * (Rworkspace: need 3*N*N)
960: *
961: CALL ZLARCM( N, N, RWORK( IRVT ), N, VT, LDVT, A, LDA,
962: $ RWORK( NRWORK ) )
963: CALL ZLACPY( 'F', N, N, A, LDA, VT, LDVT )
964: *
965: * Multiply Q in U by real matrix RWORK(IRU), storing the
966: * result in A, copying to U
967: * (CWorkspace: need 0)
968: * (Rworkspace: need N*N+2*M*N)
969: *
970: NRWORK = IRVT
971: CALL ZLACRM( M, N, U, LDU, RWORK( IRU ), N, A, LDA,
972: $ RWORK( NRWORK ) )
973: CALL ZLACPY( 'F', M, N, A, LDA, U, LDU )
974: ELSE
975: *
976: * Copy A to VT, generate P**H
977: * (Cworkspace: need 2*N, prefer N+N*NB)
978: * (Rworkspace: 0)
979: *
980: CALL ZLACPY( 'U', N, N, A, LDA, VT, LDVT )
981: CALL ZUNGBR( 'P', N, N, N, VT, LDVT, WORK( ITAUP ),
982: $ WORK( NWORK ), LWORK-NWORK+1, IERR )
983: *
984: * Copy A to U, generate Q
985: * (Cworkspace: need 2*N, prefer N+N*NB)
986: * (Rworkspace: 0)
987: *
988: CALL ZLACPY( 'L', M, N, A, LDA, U, LDU )
989: CALL ZUNGBR( 'Q', M, M, N, U, LDU, WORK( ITAUQ ),
990: $ WORK( NWORK ), LWORK-NWORK+1, IERR )
991: *
992: * Perform bidiagonal SVD, computing left singular vectors
993: * of bidiagonal matrix in RWORK(IRU) and computing right
994: * singular vectors of bidiagonal matrix in RWORK(IRVT)
995: * (CWorkspace: need 0)
996: * (RWorkspace: need BDSPAC)
997: *
998: IRU = NRWORK
999: IRVT = IRU + N*N
1000: NRWORK = IRVT + N*N
1001: CALL DBDSDC( 'U', 'I', N, S, RWORK( IE ), RWORK( IRU ),
1002: $ N, RWORK( IRVT ), N, DUM, IDUM,
1003: $ RWORK( NRWORK ), IWORK, INFO )
1004: *
1005: * Multiply real matrix RWORK(IRVT) by P**H in VT,
1006: * storing the result in A, copying to VT
1007: * (Cworkspace: need 0)
1008: * (Rworkspace: need 3*N*N)
1009: *
1010: CALL ZLARCM( N, N, RWORK( IRVT ), N, VT, LDVT, A, LDA,
1011: $ RWORK( NRWORK ) )
1012: CALL ZLACPY( 'F', N, N, A, LDA, VT, LDVT )
1013: *
1014: * Multiply Q in U by real matrix RWORK(IRU), storing the
1015: * result in A, copying to U
1016: * (CWorkspace: 0)
1017: * (Rworkspace: need 3*N*N)
1018: *
1019: NRWORK = IRVT
1020: CALL ZLACRM( M, N, U, LDU, RWORK( IRU ), N, A, LDA,
1021: $ RWORK( NRWORK ) )
1022: CALL ZLACPY( 'F', M, N, A, LDA, U, LDU )
1023: END IF
1024: *
1025: ELSE
1026: *
1027: * M .LT. MNTHR2
1028: *
1029: * Path 6 (M at least N, but not much larger)
1030: * Reduce to bidiagonal form without QR decomposition
1031: * Use ZUNMBR to compute singular vectors
1032: *
1033: IE = 1
1034: NRWORK = IE + N
1035: ITAUQ = 1
1036: ITAUP = ITAUQ + N
1037: NWORK = ITAUP + N
1038: *
1039: * Bidiagonalize A
1040: * (CWorkspace: need 2*N+M, prefer 2*N+(M+N)*NB)
1041: * (RWorkspace: need N)
1042: *
1043: CALL ZGEBRD( M, N, A, LDA, S, RWORK( IE ), WORK( ITAUQ ),
1044: $ WORK( ITAUP ), WORK( NWORK ), LWORK-NWORK+1,
1045: $ IERR )
1046: IF( WNTQN ) THEN
1047: *
1048: * Compute singular values only
1049: * (Cworkspace: 0)
1050: * (Rworkspace: need BDSPAN)
1051: *
1052: CALL DBDSDC( 'U', 'N', N, S, RWORK( IE ), DUM, 1, DUM, 1,
1053: $ DUM, IDUM, RWORK( NRWORK ), IWORK, INFO )
1054: ELSE IF( WNTQO ) THEN
1055: IU = NWORK
1056: IRU = NRWORK
1057: IRVT = IRU + N*N
1058: NRWORK = IRVT + N*N
1059: IF( LWORK.GE.M*N+3*N ) THEN
1060: *
1061: * WORK( IU ) is M by N
1062: *
1063: LDWRKU = M
1064: ELSE
1065: *
1066: * WORK( IU ) is LDWRKU by N
1067: *
1068: LDWRKU = ( LWORK-3*N ) / N
1069: END IF
1070: NWORK = IU + LDWRKU*N
1071: *
1072: * Perform bidiagonal SVD, computing left singular vectors
1073: * of bidiagonal matrix in RWORK(IRU) and computing right
1074: * singular vectors of bidiagonal matrix in RWORK(IRVT)
1075: * (CWorkspace: need 0)
1076: * (RWorkspace: need BDSPAC)
1077: *
1078: CALL DBDSDC( 'U', 'I', N, S, RWORK( IE ), RWORK( IRU ),
1079: $ N, RWORK( IRVT ), N, DUM, IDUM,
1080: $ RWORK( NRWORK ), IWORK, INFO )
1081: *
1082: * Copy real matrix RWORK(IRVT) to complex matrix VT
1083: * Overwrite VT by right singular vectors of A
1084: * (Cworkspace: need 2*N, prefer N+N*NB)
1085: * (Rworkspace: need 0)
1086: *
1087: CALL ZLACP2( 'F', N, N, RWORK( IRVT ), N, VT, LDVT )
1088: CALL ZUNMBR( 'P', 'R', 'C', N, N, N, A, LDA,
1089: $ WORK( ITAUP ), VT, LDVT, WORK( NWORK ),
1090: $ LWORK-NWORK+1, IERR )
1091: *
1092: IF( LWORK.GE.M*N+3*N ) THEN
1093: *
1094: * Copy real matrix RWORK(IRU) to complex matrix WORK(IU)
1095: * Overwrite WORK(IU) by left singular vectors of A, copying
1096: * to A
1097: * (Cworkspace: need M*N+2*N, prefer M*N+N+N*NB)
1098: * (Rworkspace: need 0)
1099: *
1100: CALL ZLASET( 'F', M, N, CZERO, CZERO, WORK( IU ),
1101: $ LDWRKU )
1102: CALL ZLACP2( 'F', N, N, RWORK( IRU ), N, WORK( IU ),
1103: $ LDWRKU )
1104: CALL ZUNMBR( 'Q', 'L', 'N', M, N, N, A, LDA,
1105: $ WORK( ITAUQ ), WORK( IU ), LDWRKU,
1106: $ WORK( NWORK ), LWORK-NWORK+1, IERR )
1107: CALL ZLACPY( 'F', M, N, WORK( IU ), LDWRKU, A, LDA )
1108: ELSE
1109: *
1110: * Generate Q in A
1111: * (Cworkspace: need 2*N, prefer N+N*NB)
1112: * (Rworkspace: need 0)
1113: *
1114: CALL ZUNGBR( 'Q', M, N, N, A, LDA, WORK( ITAUQ ),
1115: $ WORK( NWORK ), LWORK-NWORK+1, IERR )
1116: *
1117: * Multiply Q in A by real matrix RWORK(IRU), storing the
1118: * result in WORK(IU), copying to A
1119: * (CWorkspace: need N*N, prefer M*N)
1120: * (Rworkspace: need 3*N*N, prefer N*N+2*M*N)
1121: *
1122: NRWORK = IRVT
1123: DO 30 I = 1, M, LDWRKU
1124: CHUNK = MIN( M-I+1, LDWRKU )
1125: CALL ZLACRM( CHUNK, N, A( I, 1 ), LDA,
1126: $ RWORK( IRU ), N, WORK( IU ), LDWRKU,
1127: $ RWORK( NRWORK ) )
1128: CALL ZLACPY( 'F', CHUNK, N, WORK( IU ), LDWRKU,
1129: $ A( I, 1 ), LDA )
1130: 30 CONTINUE
1131: END IF
1132: *
1133: ELSE IF( WNTQS ) THEN
1134: *
1135: * Perform bidiagonal SVD, computing left singular vectors
1136: * of bidiagonal matrix in RWORK(IRU) and computing right
1137: * singular vectors of bidiagonal matrix in RWORK(IRVT)
1138: * (CWorkspace: need 0)
1139: * (RWorkspace: need BDSPAC)
1140: *
1141: IRU = NRWORK
1142: IRVT = IRU + N*N
1143: NRWORK = IRVT + N*N
1144: CALL DBDSDC( 'U', 'I', N, S, RWORK( IE ), RWORK( IRU ),
1145: $ N, RWORK( IRVT ), N, DUM, IDUM,
1146: $ RWORK( NRWORK ), IWORK, INFO )
1147: *
1148: * Copy real matrix RWORK(IRU) to complex matrix U
1149: * Overwrite U by left singular vectors of A
1150: * (CWorkspace: need 3*N, prefer 2*N+N*NB)
1151: * (RWorkspace: 0)
1152: *
1153: CALL ZLASET( 'F', M, N, CZERO, CZERO, U, LDU )
1154: CALL ZLACP2( 'F', N, N, RWORK( IRU ), N, U, LDU )
1155: CALL ZUNMBR( 'Q', 'L', 'N', M, N, N, A, LDA,
1156: $ WORK( ITAUQ ), U, LDU, WORK( NWORK ),
1157: $ LWORK-NWORK+1, IERR )
1158: *
1159: * Copy real matrix RWORK(IRVT) to complex matrix VT
1160: * Overwrite VT by right singular vectors of A
1161: * (CWorkspace: need 3*N, prefer 2*N+N*NB)
1162: * (RWorkspace: 0)
1163: *
1164: CALL ZLACP2( 'F', N, N, RWORK( IRVT ), N, VT, LDVT )
1165: CALL ZUNMBR( 'P', 'R', 'C', N, N, N, A, LDA,
1166: $ WORK( ITAUP ), VT, LDVT, WORK( NWORK ),
1167: $ LWORK-NWORK+1, IERR )
1168: ELSE
1169: *
1170: * Perform bidiagonal SVD, computing left singular vectors
1171: * of bidiagonal matrix in RWORK(IRU) and computing right
1172: * singular vectors of bidiagonal matrix in RWORK(IRVT)
1173: * (CWorkspace: need 0)
1174: * (RWorkspace: need BDSPAC)
1175: *
1176: IRU = NRWORK
1177: IRVT = IRU + N*N
1178: NRWORK = IRVT + N*N
1179: CALL DBDSDC( 'U', 'I', N, S, RWORK( IE ), RWORK( IRU ),
1180: $ N, RWORK( IRVT ), N, DUM, IDUM,
1181: $ RWORK( NRWORK ), IWORK, INFO )
1182: *
1183: * Set the right corner of U to identity matrix
1184: *
1185: CALL ZLASET( 'F', M, M, CZERO, CZERO, U, LDU )
1186: IF( M.GT.N ) THEN
1187: CALL ZLASET( 'F', M-N, M-N, CZERO, CONE,
1188: $ U( N+1, N+1 ), LDU )
1189: END IF
1190: *
1191: * Copy real matrix RWORK(IRU) to complex matrix U
1192: * Overwrite U by left singular vectors of A
1193: * (CWorkspace: need 2*N+M, prefer 2*N+M*NB)
1194: * (RWorkspace: 0)
1195: *
1196: CALL ZLACP2( 'F', N, N, RWORK( IRU ), N, U, LDU )
1197: CALL ZUNMBR( 'Q', 'L', 'N', M, M, N, A, LDA,
1198: $ WORK( ITAUQ ), U, LDU, WORK( NWORK ),
1199: $ LWORK-NWORK+1, IERR )
1200: *
1201: * Copy real matrix RWORK(IRVT) to complex matrix VT
1202: * Overwrite VT by right singular vectors of A
1203: * (CWorkspace: need 3*N, prefer 2*N+N*NB)
1204: * (RWorkspace: 0)
1205: *
1206: CALL ZLACP2( 'F', N, N, RWORK( IRVT ), N, VT, LDVT )
1207: CALL ZUNMBR( 'P', 'R', 'C', N, N, N, A, LDA,
1208: $ WORK( ITAUP ), VT, LDVT, WORK( NWORK ),
1209: $ LWORK-NWORK+1, IERR )
1210: END IF
1211: *
1212: END IF
1213: *
1214: ELSE
1215: *
1216: * A has more columns than rows. If A has sufficiently more
1217: * columns than rows, first reduce using the LQ decomposition (if
1218: * sufficient workspace available)
1219: *
1220: IF( N.GE.MNTHR1 ) THEN
1221: *
1222: IF( WNTQN ) THEN
1223: *
1224: * Path 1t (N much larger than M, JOBZ='N')
1225: * No singular vectors to be computed
1226: *
1227: ITAU = 1
1228: NWORK = ITAU + M
1229: *
1230: * Compute A=L*Q
1231: * (CWorkspace: need 2*M, prefer M+M*NB)
1232: * (RWorkspace: 0)
1233: *
1234: CALL ZGELQF( M, N, A, LDA, WORK( ITAU ), WORK( NWORK ),
1235: $ LWORK-NWORK+1, IERR )
1236: *
1237: * Zero out above L
1238: *
1239: CALL ZLASET( 'U', M-1, M-1, CZERO, CZERO, A( 1, 2 ),
1240: $ LDA )
1241: IE = 1
1242: ITAUQ = 1
1243: ITAUP = ITAUQ + M
1244: NWORK = ITAUP + M
1245: *
1246: * Bidiagonalize L in A
1247: * (CWorkspace: need 3*M, prefer 2*M+2*M*NB)
1248: * (RWorkspace: need M)
1249: *
1250: CALL ZGEBRD( M, M, A, LDA, S, RWORK( IE ), WORK( ITAUQ ),
1251: $ WORK( ITAUP ), WORK( NWORK ), LWORK-NWORK+1,
1252: $ IERR )
1253: NRWORK = IE + M
1254: *
1255: * Perform bidiagonal SVD, compute singular values only
1256: * (CWorkspace: 0)
1257: * (RWorkspace: need BDSPAN)
1258: *
1259: CALL DBDSDC( 'U', 'N', M, S, RWORK( IE ), DUM, 1, DUM, 1,
1260: $ DUM, IDUM, RWORK( NRWORK ), IWORK, INFO )
1261: *
1262: ELSE IF( WNTQO ) THEN
1263: *
1264: * Path 2t (N much larger than M, JOBZ='O')
1265: * M right singular vectors to be overwritten on A and
1266: * M left singular vectors to be computed in U
1267: *
1268: IVT = 1
1269: LDWKVT = M
1270: *
1271: * WORK(IVT) is M by M
1272: *
1273: IL = IVT + LDWKVT*M
1274: IF( LWORK.GE.M*N+M*M+3*M ) THEN
1275: *
1276: * WORK(IL) M by N
1277: *
1278: LDWRKL = M
1279: CHUNK = N
1280: ELSE
1281: *
1282: * WORK(IL) is M by CHUNK
1283: *
1284: LDWRKL = M
1285: CHUNK = ( LWORK-M*M-3*M ) / M
1286: END IF
1287: ITAU = IL + LDWRKL*CHUNK
1288: NWORK = ITAU + M
1289: *
1290: * Compute A=L*Q
1291: * (CWorkspace: need 2*M, prefer M+M*NB)
1292: * (RWorkspace: 0)
1293: *
1294: CALL ZGELQF( M, N, A, LDA, WORK( ITAU ), WORK( NWORK ),
1295: $ LWORK-NWORK+1, IERR )
1296: *
1297: * Copy L to WORK(IL), zeroing about above it
1298: *
1299: CALL ZLACPY( 'L', M, M, A, LDA, WORK( IL ), LDWRKL )
1300: CALL ZLASET( 'U', M-1, M-1, CZERO, CZERO,
1301: $ WORK( IL+LDWRKL ), LDWRKL )
1302: *
1303: * Generate Q in A
1304: * (CWorkspace: need M*M+2*M, prefer M*M+M+M*NB)
1305: * (RWorkspace: 0)
1306: *
1307: CALL ZUNGLQ( M, N, M, A, LDA, WORK( ITAU ),
1308: $ WORK( NWORK ), LWORK-NWORK+1, IERR )
1309: IE = 1
1310: ITAUQ = ITAU
1311: ITAUP = ITAUQ + M
1312: NWORK = ITAUP + M
1313: *
1314: * Bidiagonalize L in WORK(IL)
1315: * (CWorkspace: need M*M+3*M, prefer M*M+2*M+2*M*NB)
1316: * (RWorkspace: need M)
1317: *
1318: CALL ZGEBRD( M, M, WORK( IL ), LDWRKL, S, RWORK( IE ),
1319: $ WORK( ITAUQ ), WORK( ITAUP ), WORK( NWORK ),
1320: $ LWORK-NWORK+1, IERR )
1321: *
1322: * Perform bidiagonal SVD, computing left singular vectors
1323: * of bidiagonal matrix in RWORK(IRU) and computing right
1324: * singular vectors of bidiagonal matrix in RWORK(IRVT)
1325: * (CWorkspace: need 0)
1326: * (RWorkspace: need BDSPAC)
1327: *
1328: IRU = IE + M
1329: IRVT = IRU + M*M
1330: NRWORK = IRVT + M*M
1331: CALL DBDSDC( 'U', 'I', M, S, RWORK( IE ), RWORK( IRU ),
1332: $ M, RWORK( IRVT ), M, DUM, IDUM,
1333: $ RWORK( NRWORK ), IWORK, INFO )
1334: *
1335: * Copy real matrix RWORK(IRU) to complex matrix WORK(IU)
1336: * Overwrite WORK(IU) by the left singular vectors of L
1337: * (CWorkspace: need N*N+3*N, prefer M*N+2*N+N*NB)
1338: * (RWorkspace: 0)
1339: *
1340: CALL ZLACP2( 'F', M, M, RWORK( IRU ), M, U, LDU )
1341: CALL ZUNMBR( 'Q', 'L', 'N', M, M, M, WORK( IL ), LDWRKL,
1342: $ WORK( ITAUQ ), U, LDU, WORK( NWORK ),
1343: $ LWORK-NWORK+1, IERR )
1344: *
1345: * Copy real matrix RWORK(IRVT) to complex matrix WORK(IVT)
1346: * Overwrite WORK(IVT) by the right singular vectors of L
1347: * (CWorkspace: need N*N+3*N, prefer M*N+2*N+N*NB)
1348: * (RWorkspace: 0)
1349: *
1350: CALL ZLACP2( 'F', M, M, RWORK( IRVT ), M, WORK( IVT ),
1351: $ LDWKVT )
1352: CALL ZUNMBR( 'P', 'R', 'C', M, M, M, WORK( IL ), LDWRKL,
1353: $ WORK( ITAUP ), WORK( IVT ), LDWKVT,
1354: $ WORK( NWORK ), LWORK-NWORK+1, IERR )
1355: *
1356: * Multiply right singular vectors of L in WORK(IL) by Q
1357: * in A, storing result in WORK(IL) and copying to A
1358: * (CWorkspace: need 2*M*M, prefer M*M+M*N))
1359: * (RWorkspace: 0)
1360: *
1361: DO 40 I = 1, N, CHUNK
1362: BLK = MIN( N-I+1, CHUNK )
1363: CALL ZGEMM( 'N', 'N', M, BLK, M, CONE, WORK( IVT ), M,
1364: $ A( 1, I ), LDA, CZERO, WORK( IL ),
1365: $ LDWRKL )
1366: CALL ZLACPY( 'F', M, BLK, WORK( IL ), LDWRKL,
1367: $ A( 1, I ), LDA )
1368: 40 CONTINUE
1369: *
1370: ELSE IF( WNTQS ) THEN
1371: *
1372: * Path 3t (N much larger than M, JOBZ='S')
1373: * M right singular vectors to be computed in VT and
1374: * M left singular vectors to be computed in U
1375: *
1376: IL = 1
1377: *
1378: * WORK(IL) is M by M
1379: *
1380: LDWRKL = M
1381: ITAU = IL + LDWRKL*M
1382: NWORK = ITAU + M
1383: *
1384: * Compute A=L*Q
1385: * (CWorkspace: need 2*M, prefer M+M*NB)
1386: * (RWorkspace: 0)
1387: *
1388: CALL ZGELQF( M, N, A, LDA, WORK( ITAU ), WORK( NWORK ),
1389: $ LWORK-NWORK+1, IERR )
1390: *
1391: * Copy L to WORK(IL), zeroing out above it
1392: *
1393: CALL ZLACPY( 'L', M, M, A, LDA, WORK( IL ), LDWRKL )
1394: CALL ZLASET( 'U', M-1, M-1, CZERO, CZERO,
1395: $ WORK( IL+LDWRKL ), LDWRKL )
1396: *
1397: * Generate Q in A
1398: * (CWorkspace: need M*M+2*M, prefer M*M+M+M*NB)
1399: * (RWorkspace: 0)
1400: *
1401: CALL ZUNGLQ( M, N, M, A, LDA, WORK( ITAU ),
1402: $ WORK( NWORK ), LWORK-NWORK+1, IERR )
1403: IE = 1
1404: ITAUQ = ITAU
1405: ITAUP = ITAUQ + M
1406: NWORK = ITAUP + M
1407: *
1408: * Bidiagonalize L in WORK(IL)
1409: * (CWorkspace: need M*M+3*M, prefer M*M+2*M+2*M*NB)
1410: * (RWorkspace: need M)
1411: *
1412: CALL ZGEBRD( M, M, WORK( IL ), LDWRKL, S, RWORK( IE ),
1413: $ WORK( ITAUQ ), WORK( ITAUP ), WORK( NWORK ),
1414: $ LWORK-NWORK+1, IERR )
1415: *
1416: * Perform bidiagonal SVD, computing left singular vectors
1417: * of bidiagonal matrix in RWORK(IRU) and computing right
1418: * singular vectors of bidiagonal matrix in RWORK(IRVT)
1419: * (CWorkspace: need 0)
1420: * (RWorkspace: need BDSPAC)
1421: *
1422: IRU = IE + M
1423: IRVT = IRU + M*M
1424: NRWORK = IRVT + M*M
1425: CALL DBDSDC( 'U', 'I', M, S, RWORK( IE ), RWORK( IRU ),
1426: $ M, RWORK( IRVT ), M, DUM, IDUM,
1427: $ RWORK( NRWORK ), IWORK, INFO )
1428: *
1429: * Copy real matrix RWORK(IRU) to complex matrix U
1430: * Overwrite U by left singular vectors of L
1431: * (CWorkspace: need M*M+3*M, prefer M*M+2*M+M*NB)
1432: * (RWorkspace: 0)
1433: *
1434: CALL ZLACP2( 'F', M, M, RWORK( IRU ), M, U, LDU )
1435: CALL ZUNMBR( 'Q', 'L', 'N', M, M, M, WORK( IL ), LDWRKL,
1436: $ WORK( ITAUQ ), U, LDU, WORK( NWORK ),
1437: $ LWORK-NWORK+1, IERR )
1438: *
1439: * Copy real matrix RWORK(IRVT) to complex matrix VT
1440: * Overwrite VT by left singular vectors of L
1441: * (CWorkspace: need M*M+3*M, prefer M*M+2*M+M*NB)
1442: * (RWorkspace: 0)
1443: *
1444: CALL ZLACP2( 'F', M, M, RWORK( IRVT ), M, VT, LDVT )
1445: CALL ZUNMBR( 'P', 'R', 'C', M, M, M, WORK( IL ), LDWRKL,
1446: $ WORK( ITAUP ), VT, LDVT, WORK( NWORK ),
1447: $ LWORK-NWORK+1, IERR )
1448: *
1449: * Copy VT to WORK(IL), multiply right singular vectors of L
1450: * in WORK(IL) by Q in A, storing result in VT
1451: * (CWorkspace: need M*M)
1452: * (RWorkspace: 0)
1453: *
1454: CALL ZLACPY( 'F', M, M, VT, LDVT, WORK( IL ), LDWRKL )
1455: CALL ZGEMM( 'N', 'N', M, N, M, CONE, WORK( IL ), LDWRKL,
1456: $ A, LDA, CZERO, VT, LDVT )
1457: *
1458: ELSE IF( WNTQA ) THEN
1459: *
1460: * Path 9t (N much larger than M, JOBZ='A')
1461: * N right singular vectors to be computed in VT and
1462: * M left singular vectors to be computed in U
1463: *
1464: IVT = 1
1465: *
1466: * WORK(IVT) is M by M
1467: *
1468: LDWKVT = M
1469: ITAU = IVT + LDWKVT*M
1470: NWORK = ITAU + M
1471: *
1472: * Compute A=L*Q, copying result to VT
1473: * (CWorkspace: need 2*M, prefer M+M*NB)
1474: * (RWorkspace: 0)
1475: *
1476: CALL ZGELQF( M, N, A, LDA, WORK( ITAU ), WORK( NWORK ),
1477: $ LWORK-NWORK+1, IERR )
1478: CALL ZLACPY( 'U', M, N, A, LDA, VT, LDVT )
1479: *
1480: * Generate Q in VT
1481: * (CWorkspace: need M+N, prefer M+N*NB)
1482: * (RWorkspace: 0)
1483: *
1484: CALL ZUNGLQ( N, N, M, VT, LDVT, WORK( ITAU ),
1485: $ WORK( NWORK ), LWORK-NWORK+1, IERR )
1486: *
1487: * Produce L in A, zeroing out above it
1488: *
1489: CALL ZLASET( 'U', M-1, M-1, CZERO, CZERO, A( 1, 2 ),
1490: $ LDA )
1491: IE = 1
1492: ITAUQ = ITAU
1493: ITAUP = ITAUQ + M
1494: NWORK = ITAUP + M
1495: *
1496: * Bidiagonalize L in A
1497: * (CWorkspace: need M*M+3*M, prefer M*M+2*M+2*M*NB)
1498: * (RWorkspace: need M)
1499: *
1500: CALL ZGEBRD( M, M, A, LDA, S, RWORK( IE ), WORK( ITAUQ ),
1501: $ WORK( ITAUP ), WORK( NWORK ), LWORK-NWORK+1,
1502: $ IERR )
1503: *
1504: * Perform bidiagonal SVD, computing left singular vectors
1505: * of bidiagonal matrix in RWORK(IRU) and computing right
1506: * singular vectors of bidiagonal matrix in RWORK(IRVT)
1507: * (CWorkspace: need 0)
1508: * (RWorkspace: need BDSPAC)
1509: *
1510: IRU = IE + M
1511: IRVT = IRU + M*M
1512: NRWORK = IRVT + M*M
1513: CALL DBDSDC( 'U', 'I', M, S, RWORK( IE ), RWORK( IRU ),
1514: $ M, RWORK( IRVT ), M, DUM, IDUM,
1515: $ RWORK( NRWORK ), IWORK, INFO )
1516: *
1517: * Copy real matrix RWORK(IRU) to complex matrix U
1518: * Overwrite U by left singular vectors of L
1519: * (CWorkspace: need 3*M, prefer 2*M+M*NB)
1520: * (RWorkspace: 0)
1521: *
1522: CALL ZLACP2( 'F', M, M, RWORK( IRU ), M, U, LDU )
1523: CALL ZUNMBR( 'Q', 'L', 'N', M, M, M, A, LDA,
1524: $ WORK( ITAUQ ), U, LDU, WORK( NWORK ),
1525: $ LWORK-NWORK+1, IERR )
1526: *
1527: * Copy real matrix RWORK(IRVT) to complex matrix WORK(IVT)
1528: * Overwrite WORK(IVT) by right singular vectors of L
1529: * (CWorkspace: need M*M+3*M, prefer M*M+2*M+M*NB)
1530: * (RWorkspace: 0)
1531: *
1532: CALL ZLACP2( 'F', M, M, RWORK( IRVT ), M, WORK( IVT ),
1533: $ LDWKVT )
1534: CALL ZUNMBR( 'P', 'R', 'C', M, M, M, A, LDA,
1535: $ WORK( ITAUP ), WORK( IVT ), LDWKVT,
1536: $ WORK( NWORK ), LWORK-NWORK+1, IERR )
1537: *
1538: * Multiply right singular vectors of L in WORK(IVT) by
1539: * Q in VT, storing result in A
1540: * (CWorkspace: need M*M)
1541: * (RWorkspace: 0)
1542: *
1543: CALL ZGEMM( 'N', 'N', M, N, M, CONE, WORK( IVT ), LDWKVT,
1544: $ VT, LDVT, CZERO, A, LDA )
1545: *
1546: * Copy right singular vectors of A from A to VT
1547: *
1548: CALL ZLACPY( 'F', M, N, A, LDA, VT, LDVT )
1549: *
1550: END IF
1551: *
1552: ELSE IF( N.GE.MNTHR2 ) THEN
1553: *
1554: * MNTHR2 <= N < MNTHR1
1555: *
1556: * Path 5t (N much larger than M, but not as much as MNTHR1)
1557: * Reduce to bidiagonal form without QR decomposition, use
1558: * ZUNGBR and matrix multiplication to compute singular vectors
1559: *
1560: *
1561: IE = 1
1562: NRWORK = IE + M
1563: ITAUQ = 1
1564: ITAUP = ITAUQ + M
1565: NWORK = ITAUP + M
1566: *
1567: * Bidiagonalize A
1568: * (CWorkspace: need 2*M+N, prefer 2*M+(M+N)*NB)
1569: * (RWorkspace: M)
1570: *
1571: CALL ZGEBRD( M, N, A, LDA, S, RWORK( IE ), WORK( ITAUQ ),
1572: $ WORK( ITAUP ), WORK( NWORK ), LWORK-NWORK+1,
1573: $ IERR )
1574: *
1575: IF( WNTQN ) THEN
1576: *
1577: * Compute singular values only
1578: * (Cworkspace: 0)
1579: * (Rworkspace: need BDSPAN)
1580: *
1581: CALL DBDSDC( 'L', 'N', M, S, RWORK( IE ), DUM, 1, DUM, 1,
1582: $ DUM, IDUM, RWORK( NRWORK ), IWORK, INFO )
1583: ELSE IF( WNTQO ) THEN
1584: IRVT = NRWORK
1585: IRU = IRVT + M*M
1586: NRWORK = IRU + M*M
1587: IVT = NWORK
1588: *
1589: * Copy A to U, generate Q
1590: * (Cworkspace: need 2*M, prefer M+M*NB)
1591: * (Rworkspace: 0)
1592: *
1593: CALL ZLACPY( 'L', M, M, A, LDA, U, LDU )
1594: CALL ZUNGBR( 'Q', M, M, N, U, LDU, WORK( ITAUQ ),
1595: $ WORK( NWORK ), LWORK-NWORK+1, IERR )
1596: *
1597: * Generate P**H in A
1598: * (Cworkspace: need 2*M, prefer M+M*NB)
1599: * (Rworkspace: 0)
1600: *
1601: CALL ZUNGBR( 'P', M, N, M, A, LDA, WORK( ITAUP ),
1602: $ WORK( NWORK ), LWORK-NWORK+1, IERR )
1603: *
1604: LDWKVT = M
1605: IF( LWORK.GE.M*N+3*M ) THEN
1606: *
1607: * WORK( IVT ) is M by N
1608: *
1609: NWORK = IVT + LDWKVT*N
1610: CHUNK = N
1611: ELSE
1612: *
1613: * WORK( IVT ) is M by CHUNK
1614: *
1615: CHUNK = ( LWORK-3*M ) / M
1616: NWORK = IVT + LDWKVT*CHUNK
1617: END IF
1618: *
1619: * Perform bidiagonal SVD, computing left singular vectors
1620: * of bidiagonal matrix in RWORK(IRU) and computing right
1621: * singular vectors of bidiagonal matrix in RWORK(IRVT)
1622: * (CWorkspace: need 0)
1623: * (RWorkspace: need BDSPAC)
1624: *
1625: CALL DBDSDC( 'L', 'I', M, S, RWORK( IE ), RWORK( IRU ),
1626: $ M, RWORK( IRVT ), M, DUM, IDUM,
1627: $ RWORK( NRWORK ), IWORK, INFO )
1628: *
1629: * Multiply Q in U by real matrix RWORK(IRVT)
1630: * storing the result in WORK(IVT), copying to U
1631: * (Cworkspace: need 0)
1632: * (Rworkspace: need 2*M*M)
1633: *
1634: CALL ZLACRM( M, M, U, LDU, RWORK( IRU ), M, WORK( IVT ),
1635: $ LDWKVT, RWORK( NRWORK ) )
1636: CALL ZLACPY( 'F', M, M, WORK( IVT ), LDWKVT, U, LDU )
1637: *
1638: * Multiply RWORK(IRVT) by P**H in A, storing the
1639: * result in WORK(IVT), copying to A
1640: * (CWorkspace: need M*M, prefer M*N)
1641: * (Rworkspace: need 2*M*M, prefer 2*M*N)
1642: *
1643: NRWORK = IRU
1644: DO 50 I = 1, N, CHUNK
1645: BLK = MIN( N-I+1, CHUNK )
1646: CALL ZLARCM( M, BLK, RWORK( IRVT ), M, A( 1, I ), LDA,
1647: $ WORK( IVT ), LDWKVT, RWORK( NRWORK ) )
1648: CALL ZLACPY( 'F', M, BLK, WORK( IVT ), LDWKVT,
1649: $ A( 1, I ), LDA )
1650: 50 CONTINUE
1651: ELSE IF( WNTQS ) THEN
1652: *
1653: * Copy A to U, generate Q
1654: * (Cworkspace: need 2*M, prefer M+M*NB)
1655: * (Rworkspace: 0)
1656: *
1657: CALL ZLACPY( 'L', M, M, A, LDA, U, LDU )
1658: CALL ZUNGBR( 'Q', M, M, N, U, LDU, WORK( ITAUQ ),
1659: $ WORK( NWORK ), LWORK-NWORK+1, IERR )
1660: *
1661: * Copy A to VT, generate P**H
1662: * (Cworkspace: need 2*M, prefer M+M*NB)
1663: * (Rworkspace: 0)
1664: *
1665: CALL ZLACPY( 'U', M, N, A, LDA, VT, LDVT )
1666: CALL ZUNGBR( 'P', M, N, M, VT, LDVT, WORK( ITAUP ),
1667: $ WORK( NWORK ), LWORK-NWORK+1, IERR )
1668: *
1669: * Perform bidiagonal SVD, computing left singular vectors
1670: * of bidiagonal matrix in RWORK(IRU) and computing right
1671: * singular vectors of bidiagonal matrix in RWORK(IRVT)
1672: * (CWorkspace: need 0)
1673: * (RWorkspace: need BDSPAC)
1674: *
1675: IRVT = NRWORK
1676: IRU = IRVT + M*M
1677: NRWORK = IRU + M*M
1678: CALL DBDSDC( 'L', 'I', M, S, RWORK( IE ), RWORK( IRU ),
1679: $ M, RWORK( IRVT ), M, DUM, IDUM,
1680: $ RWORK( NRWORK ), IWORK, INFO )
1681: *
1682: * Multiply Q in U by real matrix RWORK(IRU), storing the
1683: * result in A, copying to U
1684: * (CWorkspace: need 0)
1685: * (Rworkspace: need 3*M*M)
1686: *
1687: CALL ZLACRM( M, M, U, LDU, RWORK( IRU ), M, A, LDA,
1688: $ RWORK( NRWORK ) )
1689: CALL ZLACPY( 'F', M, M, A, LDA, U, LDU )
1690: *
1691: * Multiply real matrix RWORK(IRVT) by P**H in VT,
1692: * storing the result in A, copying to VT
1693: * (Cworkspace: need 0)
1694: * (Rworkspace: need M*M+2*M*N)
1695: *
1696: NRWORK = IRU
1697: CALL ZLARCM( M, N, RWORK( IRVT ), M, VT, LDVT, A, LDA,
1698: $ RWORK( NRWORK ) )
1699: CALL ZLACPY( 'F', M, N, A, LDA, VT, LDVT )
1700: ELSE
1701: *
1702: * Copy A to U, generate Q
1703: * (Cworkspace: need 2*M, prefer M+M*NB)
1704: * (Rworkspace: 0)
1705: *
1706: CALL ZLACPY( 'L', M, M, A, LDA, U, LDU )
1707: CALL ZUNGBR( 'Q', M, M, N, U, LDU, WORK( ITAUQ ),
1708: $ WORK( NWORK ), LWORK-NWORK+1, IERR )
1709: *
1710: * Copy A to VT, generate P**H
1711: * (Cworkspace: need 2*M, prefer M+M*NB)
1712: * (Rworkspace: 0)
1713: *
1714: CALL ZLACPY( 'U', M, N, A, LDA, VT, LDVT )
1715: CALL ZUNGBR( 'P', N, N, M, VT, LDVT, WORK( ITAUP ),
1716: $ WORK( NWORK ), LWORK-NWORK+1, IERR )
1717: *
1718: * Perform bidiagonal SVD, computing left singular vectors
1719: * of bidiagonal matrix in RWORK(IRU) and computing right
1720: * singular vectors of bidiagonal matrix in RWORK(IRVT)
1721: * (CWorkspace: need 0)
1722: * (RWorkspace: need BDSPAC)
1723: *
1724: IRVT = NRWORK
1725: IRU = IRVT + M*M
1726: NRWORK = IRU + M*M
1727: CALL DBDSDC( 'L', 'I', M, S, RWORK( IE ), RWORK( IRU ),
1728: $ M, RWORK( IRVT ), M, DUM, IDUM,
1729: $ RWORK( NRWORK ), IWORK, INFO )
1730: *
1731: * Multiply Q in U by real matrix RWORK(IRU), storing the
1732: * result in A, copying to U
1733: * (CWorkspace: need 0)
1734: * (Rworkspace: need 3*M*M)
1735: *
1736: CALL ZLACRM( M, M, U, LDU, RWORK( IRU ), M, A, LDA,
1737: $ RWORK( NRWORK ) )
1738: CALL ZLACPY( 'F', M, M, A, LDA, U, LDU )
1739: *
1740: * Multiply real matrix RWORK(IRVT) by P**H in VT,
1741: * storing the result in A, copying to VT
1742: * (Cworkspace: need 0)
1743: * (Rworkspace: need M*M+2*M*N)
1744: *
1745: CALL ZLARCM( M, N, RWORK( IRVT ), M, VT, LDVT, A, LDA,
1746: $ RWORK( NRWORK ) )
1747: CALL ZLACPY( 'F', M, N, A, LDA, VT, LDVT )
1748: END IF
1749: *
1750: ELSE
1751: *
1752: * N .LT. MNTHR2
1753: *
1754: * Path 6t (N greater than M, but not much larger)
1755: * Reduce to bidiagonal form without LQ decomposition
1756: * Use ZUNMBR to compute singular vectors
1757: *
1758: IE = 1
1759: NRWORK = IE + M
1760: ITAUQ = 1
1761: ITAUP = ITAUQ + M
1762: NWORK = ITAUP + M
1763: *
1764: * Bidiagonalize A
1765: * (CWorkspace: need 2*M+N, prefer 2*M+(M+N)*NB)
1766: * (RWorkspace: M)
1767: *
1768: CALL ZGEBRD( M, N, A, LDA, S, RWORK( IE ), WORK( ITAUQ ),
1769: $ WORK( ITAUP ), WORK( NWORK ), LWORK-NWORK+1,
1770: $ IERR )
1771: IF( WNTQN ) THEN
1772: *
1773: * Compute singular values only
1774: * (Cworkspace: 0)
1775: * (Rworkspace: need BDSPAN)
1776: *
1777: CALL DBDSDC( 'L', 'N', M, S, RWORK( IE ), DUM, 1, DUM, 1,
1778: $ DUM, IDUM, RWORK( NRWORK ), IWORK, INFO )
1779: ELSE IF( WNTQO ) THEN
1780: LDWKVT = M
1781: IVT = NWORK
1782: IF( LWORK.GE.M*N+3*M ) THEN
1783: *
1784: * WORK( IVT ) is M by N
1785: *
1786: CALL ZLASET( 'F', M, N, CZERO, CZERO, WORK( IVT ),
1787: $ LDWKVT )
1788: NWORK = IVT + LDWKVT*N
1789: ELSE
1790: *
1791: * WORK( IVT ) is M by CHUNK
1792: *
1793: CHUNK = ( LWORK-3*M ) / M
1794: NWORK = IVT + LDWKVT*CHUNK
1795: END IF
1796: *
1797: * Perform bidiagonal SVD, computing left singular vectors
1798: * of bidiagonal matrix in RWORK(IRU) and computing right
1799: * singular vectors of bidiagonal matrix in RWORK(IRVT)
1800: * (CWorkspace: need 0)
1801: * (RWorkspace: need BDSPAC)
1802: *
1803: IRVT = NRWORK
1804: IRU = IRVT + M*M
1805: NRWORK = IRU + M*M
1806: CALL DBDSDC( 'L', 'I', M, S, RWORK( IE ), RWORK( IRU ),
1807: $ M, RWORK( IRVT ), M, DUM, IDUM,
1808: $ RWORK( NRWORK ), IWORK, INFO )
1809: *
1810: * Copy real matrix RWORK(IRU) to complex matrix U
1811: * Overwrite U by left singular vectors of A
1812: * (Cworkspace: need 2*M, prefer M+M*NB)
1813: * (Rworkspace: need 0)
1814: *
1815: CALL ZLACP2( 'F', M, M, RWORK( IRU ), M, U, LDU )
1816: CALL ZUNMBR( 'Q', 'L', 'N', M, M, N, A, LDA,
1817: $ WORK( ITAUQ ), U, LDU, WORK( NWORK ),
1818: $ LWORK-NWORK+1, IERR )
1819: *
1820: IF( LWORK.GE.M*N+3*M ) THEN
1821: *
1822: * Copy real matrix RWORK(IRVT) to complex matrix WORK(IVT)
1823: * Overwrite WORK(IVT) by right singular vectors of A,
1824: * copying to A
1825: * (Cworkspace: need M*N+2*M, prefer M*N+M+M*NB)
1826: * (Rworkspace: need 0)
1827: *
1828: CALL ZLACP2( 'F', M, M, RWORK( IRVT ), M, WORK( IVT ),
1829: $ LDWKVT )
1830: CALL ZUNMBR( 'P', 'R', 'C', M, N, M, A, LDA,
1831: $ WORK( ITAUP ), WORK( IVT ), LDWKVT,
1832: $ WORK( NWORK ), LWORK-NWORK+1, IERR )
1833: CALL ZLACPY( 'F', M, N, WORK( IVT ), LDWKVT, A, LDA )
1834: ELSE
1835: *
1836: * Generate P**H in A
1837: * (Cworkspace: need 2*M, prefer M+M*NB)
1838: * (Rworkspace: need 0)
1839: *
1840: CALL ZUNGBR( 'P', M, N, M, A, LDA, WORK( ITAUP ),
1841: $ WORK( NWORK ), LWORK-NWORK+1, IERR )
1842: *
1843: * Multiply Q in A by real matrix RWORK(IRU), storing the
1844: * result in WORK(IU), copying to A
1845: * (CWorkspace: need M*M, prefer M*N)
1846: * (Rworkspace: need 3*M*M, prefer M*M+2*M*N)
1847: *
1848: NRWORK = IRU
1849: DO 60 I = 1, N, CHUNK
1850: BLK = MIN( N-I+1, CHUNK )
1851: CALL ZLARCM( M, BLK, RWORK( IRVT ), M, A( 1, I ),
1852: $ LDA, WORK( IVT ), LDWKVT,
1853: $ RWORK( NRWORK ) )
1854: CALL ZLACPY( 'F', M, BLK, WORK( IVT ), LDWKVT,
1855: $ A( 1, I ), LDA )
1856: 60 CONTINUE
1857: END IF
1858: ELSE IF( WNTQS ) THEN
1859: *
1860: * Perform bidiagonal SVD, computing left singular vectors
1861: * of bidiagonal matrix in RWORK(IRU) and computing right
1862: * singular vectors of bidiagonal matrix in RWORK(IRVT)
1863: * (CWorkspace: need 0)
1864: * (RWorkspace: need BDSPAC)
1865: *
1866: IRVT = NRWORK
1867: IRU = IRVT + M*M
1868: NRWORK = IRU + M*M
1869: CALL DBDSDC( 'L', 'I', M, S, RWORK( IE ), RWORK( IRU ),
1870: $ M, RWORK( IRVT ), M, DUM, IDUM,
1871: $ RWORK( NRWORK ), IWORK, INFO )
1872: *
1873: * Copy real matrix RWORK(IRU) to complex matrix U
1874: * Overwrite U by left singular vectors of A
1875: * (CWorkspace: need 3*M, prefer 2*M+M*NB)
1876: * (RWorkspace: M*M)
1877: *
1878: CALL ZLACP2( 'F', M, M, RWORK( IRU ), M, U, LDU )
1879: CALL ZUNMBR( 'Q', 'L', 'N', M, M, N, A, LDA,
1880: $ WORK( ITAUQ ), U, LDU, WORK( NWORK ),
1881: $ LWORK-NWORK+1, IERR )
1882: *
1883: * Copy real matrix RWORK(IRVT) to complex matrix VT
1884: * Overwrite VT by right singular vectors of A
1885: * (CWorkspace: need 3*M, prefer 2*M+M*NB)
1886: * (RWorkspace: M*M)
1887: *
1888: CALL ZLASET( 'F', M, N, CZERO, CZERO, VT, LDVT )
1889: CALL ZLACP2( 'F', M, M, RWORK( IRVT ), M, VT, LDVT )
1890: CALL ZUNMBR( 'P', 'R', 'C', M, N, M, A, LDA,
1891: $ WORK( ITAUP ), VT, LDVT, WORK( NWORK ),
1892: $ LWORK-NWORK+1, IERR )
1893: ELSE
1894: *
1895: * Perform bidiagonal SVD, computing left singular vectors
1896: * of bidiagonal matrix in RWORK(IRU) and computing right
1897: * singular vectors of bidiagonal matrix in RWORK(IRVT)
1898: * (CWorkspace: need 0)
1899: * (RWorkspace: need BDSPAC)
1900: *
1901: IRVT = NRWORK
1902: IRU = IRVT + M*M
1903: NRWORK = IRU + M*M
1904: *
1905: CALL DBDSDC( 'L', 'I', M, S, RWORK( IE ), RWORK( IRU ),
1906: $ M, RWORK( IRVT ), M, DUM, IDUM,
1907: $ RWORK( NRWORK ), IWORK, INFO )
1908: *
1909: * Copy real matrix RWORK(IRU) to complex matrix U
1910: * Overwrite U by left singular vectors of A
1911: * (CWorkspace: need 3*M, prefer 2*M+M*NB)
1912: * (RWorkspace: M*M)
1913: *
1914: CALL ZLACP2( 'F', M, M, RWORK( IRU ), M, U, LDU )
1915: CALL ZUNMBR( 'Q', 'L', 'N', M, M, N, A, LDA,
1916: $ WORK( ITAUQ ), U, LDU, WORK( NWORK ),
1917: $ LWORK-NWORK+1, IERR )
1918: *
1919: * Set all of VT to identity matrix
1920: *
1921: CALL ZLASET( 'F', N, N, CZERO, CONE, VT, LDVT )
1922: *
1923: * Copy real matrix RWORK(IRVT) to complex matrix VT
1924: * Overwrite VT by right singular vectors of A
1925: * (CWorkspace: need 2*M+N, prefer 2*M+N*NB)
1926: * (RWorkspace: M*M)
1927: *
1928: CALL ZLACP2( 'F', M, M, RWORK( IRVT ), M, VT, LDVT )
1929: CALL ZUNMBR( 'P', 'R', 'C', N, N, M, A, LDA,
1930: $ WORK( ITAUP ), VT, LDVT, WORK( NWORK ),
1931: $ LWORK-NWORK+1, IERR )
1932: END IF
1933: *
1934: END IF
1935: *
1936: END IF
1937: *
1938: * Undo scaling if necessary
1939: *
1940: IF( ISCL.EQ.1 ) THEN
1941: IF( ANRM.GT.BIGNUM )
1942: $ CALL DLASCL( 'G', 0, 0, BIGNUM, ANRM, MINMN, 1, S, MINMN,
1943: $ IERR )
1944: IF( INFO.NE.0 .AND. ANRM.GT.BIGNUM )
1945: $ CALL DLASCL( 'G', 0, 0, BIGNUM, ANRM, MINMN-1, 1,
1946: $ RWORK( IE ), MINMN, IERR )
1947: IF( ANRM.LT.SMLNUM )
1948: $ CALL DLASCL( 'G', 0, 0, SMLNUM, ANRM, MINMN, 1, S, MINMN,
1949: $ IERR )
1950: IF( INFO.NE.0 .AND. ANRM.LT.SMLNUM )
1951: $ CALL DLASCL( 'G', 0, 0, SMLNUM, ANRM, MINMN-1, 1,
1952: $ RWORK( IE ), MINMN, IERR )
1953: END IF
1954: *
1955: * Return optimal workspace in WORK(1)
1956: *
1957: WORK( 1 ) = MAXWRK
1958: *
1959: RETURN
1960: *
1961: * End of ZGESDD
1962: *
1963: END
CVSweb interface <joel.bertrand@systella.fr>