LAPACK  3.9.0
LAPACK: Linear Algebra PACKage
dsytrd_sb2st.F
Go to the documentation of this file.
1 *> \brief \b DSYTRD_SB2ST reduces a real symmetric band matrix A to real symmetric tridiagonal form T
2 *
3 * =========== DOCUMENTATION ===========
4 *
5 * Online html documentation available at
6 * http://www.netlib.org/lapack/explore-html/
7 *
8 *> \htmlonly
9 *> Download DSYTRD_SB2ST + dependencies
10 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/dsytrd_sb2st.f">
11 *> [TGZ]</a>
12 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/dsytrd_sb2st.f">
13 *> [ZIP]</a>
14 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/dsytrd_sb2st.f">
15 *> [TXT]</a>
16 *> \endhtmlonly
17 *
18 * Definition:
19 * ===========
20 *
21 * SUBROUTINE DSYTRD_SB2ST( STAGE1, VECT, UPLO, N, KD, AB, LDAB,
22 * D, E, HOUS, LHOUS, WORK, LWORK, INFO )
23 *
24 * #if defined(_OPENMP)
25 * use omp_lib
26 * #endif
27 *
28 * IMPLICIT NONE
29 *
30 * .. Scalar Arguments ..
31 * CHARACTER STAGE1, UPLO, VECT
32 * INTEGER N, KD, IB, LDAB, LHOUS, LWORK, INFO
33 * ..
34 * .. Array Arguments ..
35 * DOUBLE PRECISION D( * ), E( * )
36 * DOUBLE PRECISION AB( LDAB, * ), HOUS( * ), WORK( * )
37 * ..
38 *
39 *
40 *> \par Purpose:
41 * =============
42 *>
43 *> \verbatim
44 *>
45 *> DSYTRD_SB2ST reduces a real symmetric band matrix A to real symmetric
46 *> tridiagonal form T by a orthogonal similarity transformation:
47 *> Q**T * A * Q = T.
48 *> \endverbatim
49 *
50 * Arguments:
51 * ==========
52 *
53 *> \param[in] STAGE1
54 *> \verbatim
55 *> STAGE1 is CHARACTER*1
56 *> = 'N': "No": to mention that the stage 1 of the reduction
57 *> from dense to band using the dsytrd_sy2sb routine
58 *> was not called before this routine to reproduce AB.
59 *> In other term this routine is called as standalone.
60 *> = 'Y': "Yes": to mention that the stage 1 of the
61 *> reduction from dense to band using the dsytrd_sy2sb
62 *> routine has been called to produce AB (e.g., AB is
63 *> the output of dsytrd_sy2sb.
64 *> \endverbatim
65 *>
66 *> \param[in] VECT
67 *> \verbatim
68 *> VECT is CHARACTER*1
69 *> = 'N': No need for the Housholder representation,
70 *> and thus LHOUS is of size max(1, 4*N);
71 *> = 'V': the Householder representation is needed to
72 *> either generate or to apply Q later on,
73 *> then LHOUS is to be queried and computed.
74 *> (NOT AVAILABLE IN THIS RELEASE).
75 *> \endverbatim
76 *>
77 *> \param[in] UPLO
78 *> \verbatim
79 *> UPLO is CHARACTER*1
80 *> = 'U': Upper triangle of A is stored;
81 *> = 'L': Lower triangle of A is stored.
82 *> \endverbatim
83 *>
84 *> \param[in] N
85 *> \verbatim
86 *> N is INTEGER
87 *> The order of the matrix A. N >= 0.
88 *> \endverbatim
89 *>
90 *> \param[in] KD
91 *> \verbatim
92 *> KD is INTEGER
93 *> The number of superdiagonals of the matrix A if UPLO = 'U',
94 *> or the number of subdiagonals if UPLO = 'L'. KD >= 0.
95 *> \endverbatim
96 *>
97 *> \param[in,out] AB
98 *> \verbatim
99 *> AB is DOUBLE PRECISION array, dimension (LDAB,N)
100 *> On entry, the upper or lower triangle of the symmetric band
101 *> matrix A, stored in the first KD+1 rows of the array. The
102 *> j-th column of A is stored in the j-th column of the array AB
103 *> as follows:
104 *> if UPLO = 'U', AB(kd+1+i-j,j) = A(i,j) for max(1,j-kd)<=i<=j;
105 *> if UPLO = 'L', AB(1+i-j,j) = A(i,j) for j<=i<=min(n,j+kd).
106 *> On exit, the diagonal elements of AB are overwritten by the
107 *> diagonal elements of the tridiagonal matrix T; if KD > 0, the
108 *> elements on the first superdiagonal (if UPLO = 'U') or the
109 *> first subdiagonal (if UPLO = 'L') are overwritten by the
110 *> off-diagonal elements of T; the rest of AB is overwritten by
111 *> values generated during the reduction.
112 *> \endverbatim
113 *>
114 *> \param[in] LDAB
115 *> \verbatim
116 *> LDAB is INTEGER
117 *> The leading dimension of the array AB. LDAB >= KD+1.
118 *> \endverbatim
119 *>
120 *> \param[out] D
121 *> \verbatim
122 *> D is DOUBLE PRECISION array, dimension (N)
123 *> The diagonal elements of the tridiagonal matrix T.
124 *> \endverbatim
125 *>
126 *> \param[out] E
127 *> \verbatim
128 *> E is DOUBLE PRECISION array, dimension (N-1)
129 *> The off-diagonal elements of the tridiagonal matrix T:
130 *> E(i) = T(i,i+1) if UPLO = 'U'; E(i) = T(i+1,i) if UPLO = 'L'.
131 *> \endverbatim
132 *>
133 *> \param[out] HOUS
134 *> \verbatim
135 *> HOUS is DOUBLE PRECISION array, dimension LHOUS, that
136 *> store the Householder representation.
137 *> \endverbatim
138 *>
139 *> \param[in] LHOUS
140 *> \verbatim
141 *> LHOUS is INTEGER
142 *> The dimension of the array HOUS. LHOUS = MAX(1, dimension)
143 *> If LWORK = -1, or LHOUS=-1,
144 *> then a query is assumed; the routine
145 *> only calculates the optimal size of the HOUS array, returns
146 *> this value as the first entry of the HOUS array, and no error
147 *> message related to LHOUS is issued by XERBLA.
148 *> LHOUS = MAX(1, dimension) where
149 *> dimension = 4*N if VECT='N'
150 *> not available now if VECT='H'
151 *> \endverbatim
152 *>
153 *> \param[out] WORK
154 *> \verbatim
155 *> WORK is DOUBLE PRECISION array, dimension LWORK.
156 *> \endverbatim
157 *>
158 *> \param[in] LWORK
159 *> \verbatim
160 *> LWORK is INTEGER
161 *> The dimension of the array WORK. LWORK = MAX(1, dimension)
162 *> If LWORK = -1, or LHOUS=-1,
163 *> then a workspace query is assumed; the routine
164 *> only calculates the optimal size of the WORK array, returns
165 *> this value as the first entry of the WORK array, and no error
166 *> message related to LWORK is issued by XERBLA.
167 *> LWORK = MAX(1, dimension) where
168 *> dimension = (2KD+1)*N + KD*NTHREADS
169 *> where KD is the blocking size of the reduction,
170 *> FACTOPTNB is the blocking used by the QR or LQ
171 *> algorithm, usually FACTOPTNB=128 is a good choice
172 *> NTHREADS is the number of threads used when
173 *> openMP compilation is enabled, otherwise =1.
174 *> \endverbatim
175 *>
176 *> \param[out] INFO
177 *> \verbatim
178 *> INFO is INTEGER
179 *> = 0: successful exit
180 *> < 0: if INFO = -i, the i-th argument had an illegal value
181 *> \endverbatim
182 *
183 * Authors:
184 * ========
185 *
186 *> \author Univ. of Tennessee
187 *> \author Univ. of California Berkeley
188 *> \author Univ. of Colorado Denver
189 *> \author NAG Ltd.
190 *
191 *> \date November 2017
192 *
193 *> \ingroup real16OTHERcomputational
194 *
195 *> \par Further Details:
196 * =====================
197 *>
198 *> \verbatim
199 *>
200 *> Implemented by Azzam Haidar.
201 *>
202 *> All details are available on technical report, SC11, SC13 papers.
203 *>
204 *> Azzam Haidar, Hatem Ltaief, and Jack Dongarra.
205 *> Parallel reduction to condensed forms for symmetric eigenvalue problems
206 *> using aggregated fine-grained and memory-aware kernels. In Proceedings
207 *> of 2011 International Conference for High Performance Computing,
208 *> Networking, Storage and Analysis (SC '11), New York, NY, USA,
209 *> Article 8 , 11 pages.
210 *> http://doi.acm.org/10.1145/2063384.2063394
211 *>
212 *> A. Haidar, J. Kurzak, P. Luszczek, 2013.
213 *> An improved parallel singular value algorithm and its implementation
214 *> for multicore hardware, In Proceedings of 2013 International Conference
215 *> for High Performance Computing, Networking, Storage and Analysis (SC '13).
216 *> Denver, Colorado, USA, 2013.
217 *> Article 90, 12 pages.
218 *> http://doi.acm.org/10.1145/2503210.2503292
219 *>
220 *> A. Haidar, R. Solca, S. Tomov, T. Schulthess and J. Dongarra.
221 *> A novel hybrid CPU-GPU generalized eigensolver for electronic structure
222 *> calculations based on fine-grained memory aware tasks.
223 *> International Journal of High Performance Computing Applications.
224 *> Volume 28 Issue 2, Pages 196-209, May 2014.
225 *> http://hpc.sagepub.com/content/28/2/196
226 *>
227 *> \endverbatim
228 *>
229 * =====================================================================
230  SUBROUTINE dsytrd_sb2st( STAGE1, VECT, UPLO, N, KD, AB, LDAB,
231  $ D, E, HOUS, LHOUS, WORK, LWORK, INFO )
232 *
233 #if defined(_OPENMP)
234  use omp_lib
235 #endif
236 *
237  IMPLICIT NONE
238 *
239 * -- LAPACK computational routine (version 3.8.0) --
240 * -- LAPACK is a software package provided by Univ. of Tennessee, --
241 * -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
242 * November 2017
243 *
244 * .. Scalar Arguments ..
245  CHARACTER STAGE1, UPLO, VECT
246  INTEGER N, KD, LDAB, LHOUS, LWORK, INFO
247 * ..
248 * .. Array Arguments ..
249  DOUBLE PRECISION D( * ), E( * )
250  DOUBLE PRECISION AB( LDAB, * ), HOUS( * ), WORK( * )
251 * ..
252 *
253 * =====================================================================
254 *
255 * .. Parameters ..
256  DOUBLE PRECISION RZERO
257  DOUBLE PRECISION ZERO, ONE
258  parameter( rzero = 0.0d+0,
259  $ zero = 0.0d+0,
260  $ one = 1.0d+0 )
261 * ..
262 * .. Local Scalars ..
263  LOGICAL LQUERY, WANTQ, UPPER, AFTERS1
264  INTEGER I, M, K, IB, SWEEPID, MYID, SHIFT, STT, ST,
265  $ ed, stind, edind, blklastind, colpt, thed,
266  $ stepercol, grsiz, thgrsiz, thgrnb, thgrid,
267  $ nbtiles, ttype, tid, nthreads, debug,
268  $ abdpos, abofdpos, dpos, ofdpos, awpos,
269  $ inda, indw, apos, sizea, lda, indv, indtau,
270  $ sidev, sizetau, ldv, lhmin, lwmin
271 * ..
272 * .. External Subroutines ..
273  EXTERNAL dsb2st_kernels, dlacpy, dlaset, xerbla
274 * ..
275 * .. Intrinsic Functions ..
276  INTRINSIC min, max, ceiling, real
277 * ..
278 * .. External Functions ..
279  LOGICAL LSAME
280  INTEGER ILAENV2STAGE
281  EXTERNAL lsame, ilaenv2stage
282 * ..
283 * .. Executable Statements ..
284 *
285 * Determine the minimal workspace size required.
286 * Test the input parameters
287 *
288  debug = 0
289  info = 0
290  afters1 = lsame( stage1, 'Y' )
291  wantq = lsame( vect, 'V' )
292  upper = lsame( uplo, 'U' )
293  lquery = ( lwork.EQ.-1 ) .OR. ( lhous.EQ.-1 )
294 *
295 * Determine the block size, the workspace size and the hous size.
296 *
297  ib = ilaenv2stage( 2, 'DSYTRD_SB2ST', vect, n, kd, -1, -1 )
298  lhmin = ilaenv2stage( 3, 'DSYTRD_SB2ST', vect, n, kd, ib, -1 )
299  lwmin = ilaenv2stage( 4, 'DSYTRD_SB2ST', vect, n, kd, ib, -1 )
300 *
301  IF( .NOT.afters1 .AND. .NOT.lsame( stage1, 'N' ) ) THEN
302  info = -1
303  ELSE IF( .NOT.lsame( vect, 'N' ) ) THEN
304  info = -2
305  ELSE IF( .NOT.upper .AND. .NOT.lsame( uplo, 'L' ) ) THEN
306  info = -3
307  ELSE IF( n.LT.0 ) THEN
308  info = -4
309  ELSE IF( kd.LT.0 ) THEN
310  info = -5
311  ELSE IF( ldab.LT.(kd+1) ) THEN
312  info = -7
313  ELSE IF( lhous.LT.lhmin .AND. .NOT.lquery ) THEN
314  info = -11
315  ELSE IF( lwork.LT.lwmin .AND. .NOT.lquery ) THEN
316  info = -13
317  END IF
318 *
319  IF( info.EQ.0 ) THEN
320  hous( 1 ) = lhmin
321  work( 1 ) = lwmin
322  END IF
323 *
324  IF( info.NE.0 ) THEN
325  CALL xerbla( 'DSYTRD_SB2ST', -info )
326  RETURN
327  ELSE IF( lquery ) THEN
328  RETURN
329  END IF
330 *
331 * Quick return if possible
332 *
333  IF( n.EQ.0 ) THEN
334  hous( 1 ) = 1
335  work( 1 ) = 1
336  RETURN
337  END IF
338 *
339 * Determine pointer position
340 *
341  ldv = kd + ib
342  sizetau = 2 * n
343  sidev = 2 * n
344  indtau = 1
345  indv = indtau + sizetau
346  lda = 2 * kd + 1
347  sizea = lda * n
348  inda = 1
349  indw = inda + sizea
350  nthreads = 1
351  tid = 0
352 *
353  IF( upper ) THEN
354  apos = inda + kd
355  awpos = inda
356  dpos = apos + kd
357  ofdpos = dpos - 1
358  abdpos = kd + 1
359  abofdpos = kd
360  ELSE
361  apos = inda
362  awpos = inda + kd + 1
363  dpos = apos
364  ofdpos = dpos + 1
365  abdpos = 1
366  abofdpos = 2
367 
368  ENDIF
369 *
370 * Case KD=0:
371 * The matrix is diagonal. We just copy it (convert to "real" for
372 * real because D is double and the imaginary part should be 0)
373 * and store it in D. A sequential code here is better or
374 * in a parallel environment it might need two cores for D and E
375 *
376  IF( kd.EQ.0 ) THEN
377  DO 30 i = 1, n
378  d( i ) = ( ab( abdpos, i ) )
379  30 CONTINUE
380  DO 40 i = 1, n-1
381  e( i ) = rzero
382  40 CONTINUE
383 *
384  hous( 1 ) = 1
385  work( 1 ) = 1
386  RETURN
387  END IF
388 *
389 * Case KD=1:
390 * The matrix is already Tridiagonal. We have to make diagonal
391 * and offdiagonal elements real, and store them in D and E.
392 * For that, for real precision just copy the diag and offdiag
393 * to D and E while for the COMPLEX case the bulge chasing is
394 * performed to convert the hermetian tridiagonal to symmetric
395 * tridiagonal. A simpler coversion formula might be used, but then
396 * updating the Q matrix will be required and based if Q is generated
397 * or not this might complicate the story.
398 *
399  IF( kd.EQ.1 ) THEN
400  DO 50 i = 1, n
401  d( i ) = ( ab( abdpos, i ) )
402  50 CONTINUE
403 *
404  IF( upper ) THEN
405  DO 60 i = 1, n-1
406  e( i ) = ( ab( abofdpos, i+1 ) )
407  60 CONTINUE
408  ELSE
409  DO 70 i = 1, n-1
410  e( i ) = ( ab( abofdpos, i ) )
411  70 CONTINUE
412  ENDIF
413 *
414  hous( 1 ) = 1
415  work( 1 ) = 1
416  RETURN
417  END IF
418 *
419 * Main code start here.
420 * Reduce the symmetric band of A to a tridiagonal matrix.
421 *
422  thgrsiz = n
423  grsiz = 1
424  shift = 3
425  nbtiles = ceiling( real(n)/real(kd) )
426  stepercol = ceiling( real(shift)/real(grsiz) )
427  thgrnb = ceiling( real(n-1)/real(thgrsiz) )
428 *
429  CALL dlacpy( "A", kd+1, n, ab, ldab, work( apos ), lda )
430  CALL dlaset( "A", kd, n, zero, zero, work( awpos ), lda )
431 *
432 *
433 * openMP parallelisation start here
434 *
435 #if defined(_OPENMP)
436 !$OMP PARALLEL PRIVATE( TID, THGRID, BLKLASTIND )
437 !$OMP$ PRIVATE( THED, I, M, K, ST, ED, STT, SWEEPID )
438 !$OMP$ PRIVATE( MYID, TTYPE, COLPT, STIND, EDIND )
439 !$OMP$ SHARED ( UPLO, WANTQ, INDV, INDTAU, HOUS, WORK)
440 !$OMP$ SHARED ( N, KD, IB, NBTILES, LDA, LDV, INDA )
441 !$OMP$ SHARED ( STEPERCOL, THGRNB, THGRSIZ, GRSIZ, SHIFT )
442 !$OMP MASTER
443 #endif
444 *
445 * main bulge chasing loop
446 *
447  DO 100 thgrid = 1, thgrnb
448  stt = (thgrid-1)*thgrsiz+1
449  thed = min( (stt + thgrsiz -1), (n-1))
450  DO 110 i = stt, n-1
451  ed = min( i, thed )
452  IF( stt.GT.ed ) EXIT
453  DO 120 m = 1, stepercol
454  st = stt
455  DO 130 sweepid = st, ed
456  DO 140 k = 1, grsiz
457  myid = (i-sweepid)*(stepercol*grsiz)
458  $ + (m-1)*grsiz + k
459  IF ( myid.EQ.1 ) THEN
460  ttype = 1
461  ELSE
462  ttype = mod( myid, 2 ) + 2
463  ENDIF
464 
465  IF( ttype.EQ.2 ) THEN
466  colpt = (myid/2)*kd + sweepid
467  stind = colpt-kd+1
468  edind = min(colpt,n)
469  blklastind = colpt
470  ELSE
471  colpt = ((myid+1)/2)*kd + sweepid
472  stind = colpt-kd+1
473  edind = min(colpt,n)
474  IF( ( stind.GE.edind-1 ).AND.
475  $ ( edind.EQ.n ) ) THEN
476  blklastind = n
477  ELSE
478  blklastind = 0
479  ENDIF
480  ENDIF
481 *
482 * Call the kernel
483 *
484 #if defined(_OPENMP)
485  IF( ttype.NE.1 ) THEN
486 !$OMP TASK DEPEND(in:WORK(MYID+SHIFT-1))
487 !$OMP$ DEPEND(in:WORK(MYID-1))
488 !$OMP$ DEPEND(out:WORK(MYID))
489  tid = omp_get_thread_num()
490  CALL dsb2st_kernels( uplo, wantq, ttype,
491  $ stind, edind, sweepid, n, kd, ib,
492  $ work( inda ), lda,
493  $ hous( indv ), hous( indtau ), ldv,
494  $ work( indw + tid*kd ) )
495 !$OMP END TASK
496  ELSE
497 !$OMP TASK DEPEND(in:WORK(MYID+SHIFT-1))
498 !$OMP$ DEPEND(out:WORK(MYID))
499  tid = omp_get_thread_num()
500  CALL dsb2st_kernels( uplo, wantq, ttype,
501  $ stind, edind, sweepid, n, kd, ib,
502  $ work( inda ), lda,
503  $ hous( indv ), hous( indtau ), ldv,
504  $ work( indw + tid*kd ) )
505 !$OMP END TASK
506  ENDIF
507 #else
508  CALL dsb2st_kernels( uplo, wantq, ttype,
509  $ stind, edind, sweepid, n, kd, ib,
510  $ work( inda ), lda,
511  $ hous( indv ), hous( indtau ), ldv,
512  $ work( indw + tid*kd ) )
513 #endif
514  IF ( blklastind.GE.(n-1) ) THEN
515  stt = stt + 1
516  EXIT
517  ENDIF
518  140 CONTINUE
519  130 CONTINUE
520  120 CONTINUE
521  110 CONTINUE
522  100 CONTINUE
523 *
524 #if defined(_OPENMP)
525 !$OMP END MASTER
526 !$OMP END PARALLEL
527 #endif
528 *
529 * Copy the diagonal from A to D. Note that D is REAL thus only
530 * the Real part is needed, the imaginary part should be zero.
531 *
532  DO 150 i = 1, n
533  d( i ) = ( work( dpos+(i-1)*lda ) )
534  150 CONTINUE
535 *
536 * Copy the off diagonal from A to E. Note that E is REAL thus only
537 * the Real part is needed, the imaginary part should be zero.
538 *
539  IF( upper ) THEN
540  DO 160 i = 1, n-1
541  e( i ) = ( work( ofdpos+i*lda ) )
542  160 CONTINUE
543  ELSE
544  DO 170 i = 1, n-1
545  e( i ) = ( work( ofdpos+(i-1)*lda ) )
546  170 CONTINUE
547  ENDIF
548 *
549  hous( 1 ) = lhmin
550  work( 1 ) = lwmin
551  RETURN
552 *
553 * End of DSYTRD_SB2ST
554 *
555  END
556 
dsytrd_sb2st
subroutine dsytrd_sb2st(STAGE1, VECT, UPLO, N, KD, AB, LDAB, D, E, HOUS, LHOUS, WORK, LWORK, INFO)
DSYTRD_SB2ST reduces a real symmetric band matrix A to real symmetric tridiagonal form T
Definition: dsytrd_sb2st.F:232
xerbla
subroutine xerbla(SRNAME, INFO)
XERBLA
Definition: xerbla.f:62
dlaset
subroutine dlaset(UPLO, M, N, ALPHA, BETA, A, LDA)
DLASET initializes the off-diagonal elements and the diagonal elements of a matrix to given values.
Definition: dlaset.f:112
dsb2st_kernels
subroutine dsb2st_kernels(UPLO, WANTZ, TTYPE, ST, ED, SWEEP, N, NB, IB, A, LDA, V, TAU, LDVT, WORK)
DSB2ST_KERNELS
Definition: dsb2st_kernels.f:170
dlacpy
subroutine dlacpy(UPLO, M, N, A, LDA, B, LDB)
DLACPY copies all or part of one two-dimensional array to another.
Definition: dlacpy.f:105