elpa2.F90 214 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
!    This file is part of ELPA.
!
!    The ELPA library was originally created by the ELPA consortium,
!    consisting of the following organizations:
!
!    - Rechenzentrum Garching der Max-Planck-Gesellschaft (RZG),
!    - Bergische Universität Wuppertal, Lehrstuhl für angewandte
!      Informatik,
!    - Technische Universität München, Lehrstuhl für Informatik mit
!      Schwerpunkt Wissenschaftliches Rechnen ,
!    - Fritz-Haber-Institut, Berlin, Abt. Theorie,
!    - Max-Plack-Institut für Mathematik in den Naturwissenschaftrn,
!      Leipzig, Abt. Komplexe Strukutren in Biologie und Kognition,
!      and
!    - IBM Deutschland GmbH
!
17
18
19
!    This particular source code file contains additions, changes and
!    enhancements authored by Intel Corporation which is not part of 
!    the ELPA consortium.
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
!
!    More information can be found here:
!    http://elpa.rzg.mpg.de/
!
!    ELPA is free software: you can redistribute it and/or modify
!    it under the terms of the version 3 of the license of the
!    GNU Lesser General Public License as published by the Free
!    Software Foundation.
!
!    ELPA is distributed in the hope that it will be useful,
!    but WITHOUT ANY WARRANTY; without even the implied warranty of
!    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
!    GNU Lesser General Public License for more details.
!
!    You should have received a copy of the GNU Lesser General Public License
!    along with ELPA.  If not, see <http://www.gnu.org/licenses/>
!
!    ELPA reflects a substantial effort on the part of the original
!    ELPA consortium, and we ask you to respect the spirit of the
!    license that we chose: i.e., please contribute any changes you
!    may have back to the original ELPA library distribution, and keep
!    any derivatives of ELPA under the same license that we chose for
!    the original distribution, the GNU Lesser General Public License.
!
!
! ELPA1 -- Faster replacements for ScaLAPACK symmetric eigenvalue routines
!
! Copyright of the original code rests with the authors inside the ELPA
! consortium. The copyright of any additional modifications shall rest
! with their original authors, but shall adhere to the licensing terms
! distributed along with the original code in the file "COPYING".



! ELPA2 -- 2-stage solver for ELPA
!
! Copyright of the original code rests with the authors inside the ELPA
! consortium. The copyright of any additional modifications shall rest
! with their original authors, but shall adhere to the licensing terms
! distributed along with the original code in the file "COPYING".


#include "config-f90.h"

module ELPA2

! Version 1.1.2, 2011-02-21

68
  use elpa_utilities
69
  USE ELPA1
70
  use elpa2_utilities
71
72
  use elpa_pdgeqrf

73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
  implicit none

  PRIVATE ! By default, all routines contained are private

  ! The following routines are public:

  public :: solve_evp_real_2stage
  public :: solve_evp_complex_2stage

  public :: bandred_real
  public :: tridiag_band_real
  public :: trans_ev_tridi_to_band_real
  public :: trans_ev_band_to_full_real

  public :: bandred_complex
  public :: tridiag_band_complex
  public :: trans_ev_tridi_to_band_complex
  public :: trans_ev_band_to_full_complex
91

92
93
94
95
96
97
  public :: band_band_real
  public :: divide_band

  integer, public :: which_qr_decomposition = 1     ! defines, which QR-decomposition algorithm will be used
                                                    ! 0 for unblocked
                                                    ! 1 for blocked (maxrank: nblk)
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
!-------------------------------------------------------------------------------

  ! The following array contains the Householder vectors of the
  ! transformation band -> tridiagonal.
  ! It is allocated and set in tridiag_band_real and used in
  ! trans_ev_tridi_to_band_real.
  ! It must be deallocated by the user after trans_ev_tridi_to_band_real!

  real*8, allocatable :: hh_trans_real(:,:)
  complex*16, allocatable :: hh_trans_complex(:,:)

!-------------------------------------------------------------------------------

  include 'mpif.h'


!******
contains
116

117
118
119
120
function solve_evp_real_2stage(na, nev, a, lda, ev, q, ldq, nblk,        &
                                 mpi_comm_rows, mpi_comm_cols,           &
                                 mpi_comm_all, THIS_REAL_ELPA_KERNEL_API,&
                                 useQR) result(success)
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155

!-------------------------------------------------------------------------------
!  solve_evp_real_2stage: Solves the real eigenvalue problem with a 2 stage approach
!
!  Parameters
!
!  na          Order of matrix a
!
!  nev         Number of eigenvalues needed
!
!  a(lda,*)    Distributed matrix for which eigenvalues are to be computed.
!              Distribution is like in Scalapack.
!              The full matrix must be set (not only one half like in scalapack).
!              Destroyed on exit (upper and lower half).
!
!  lda         Leading dimension of a
!
!  ev(na)      On output: eigenvalues of a, every processor gets the complete set
!
!  q(ldq,*)    On output: Eigenvectors of a
!              Distribution is like in Scalapack.
!              Must be always dimensioned to the full size (corresponding to (na,na))
!              even if only a part of the eigenvalues is needed.
!
!  ldq         Leading dimension of q
!
!  nblk        blocksize of cyclic distribution, must be the same in both directions!
!
!  mpi_comm_rows
!  mpi_comm_cols
!              MPI-Communicators for rows/columns
!  mpi_comm_all
!              MPI-Communicator for the total processor set
!
!-------------------------------------------------------------------------------
156
157
158
#ifdef HAVE_DETAILED_TIMINGS
 use timings
#endif
159
   implicit none
160
161
   logical, intent(in), optional :: useQR
   logical                       :: useQRActual, useQREnvironment
Andreas Marek's avatar
Andreas Marek committed
162
   integer, intent(in), optional :: THIS_REAL_ELPA_KERNEL_API
163
   integer                       :: THIS_REAL_ELPA_KERNEL
164

165
   integer, intent(in)           :: na, nev, lda, ldq, mpi_comm_rows, &
166
                                    mpi_comm_cols, mpi_comm_all
167
   integer, intent(in)           :: nblk
168
   real*8, intent(inout)         :: a(lda,*), ev(na), q(ldq,*)
169

170
171
172
173
174
175
   integer                       :: my_pe, n_pes, my_prow, my_pcol, np_rows, np_cols, mpierr
   integer                       :: nbw, num_blocks
   real*8, allocatable           :: tmat(:,:,:), e(:)
   real*8                        :: ttt0, ttt1, ttts
   integer                       :: i
   logical                       :: success
176
177
   logical, save                 :: firstCall = .true.
   logical                       :: wantDebug
178

179
180
181
#ifdef HAVE_DETAILED_TIMINGS
   call timer%start("solve_evp_real_2stage")
#endif
182
183
184
185
186
187
188
   call mpi_comm_rank(mpi_comm_all,my_pe,mpierr)
   call mpi_comm_size(mpi_comm_all,n_pes,mpierr)

   call mpi_comm_rank(mpi_comm_rows,my_prow,mpierr)
   call mpi_comm_size(mpi_comm_rows,np_rows,mpierr)
   call mpi_comm_rank(mpi_comm_cols,my_pcol,mpierr)
   call mpi_comm_size(mpi_comm_cols,np_cols,mpierr)
189

190
191
192
193
194
195
196
197

   wantDebug = .false.
   if (firstCall) then
     ! are debug messages desired?
     wantDebug = debug_messages_via_environment_variable()
     firstCall = .false.
   endif

198
199
   success = .true.

200
201
202
203
204
205
206
207
208
209
210
211
212
   useQRActual = .false.

   ! set usage of qr decomposition via API call
   if (present(useQR)) then
     if (useQR) useQRActual = .true.
     if (.not.(useQR)) useQRACtual = .false.
   endif

   ! overwrite this with environment variable settings
   if (qr_decomposition_via_environment_variable(useQREnvironment)) then
     useQRActual = useQREnvironment
   endif

213
   if (useQRActual) then
214
215
216
217
     if (mod(na,nblk) .ne. 0) then
       if (wantDebug) then
         write(error_unit,*) "solve_evp_real_2stage: QR-decomposition: blocksize does not fit with matrixsize"
       endif
Andreas Marek's avatar
Andreas Marek committed
218
     print *, "Do not use QR-decomposition for this matrix and blocksize."
Andreas Marek's avatar
Andreas Marek committed
219
220
     success = .false.
     return
221
     endif
222
223
   endif

224

225
226
227
   if (present(THIS_REAL_ELPA_KERNEL_API)) then
     ! user defined kernel via the optional argument in the API call
     THIS_REAL_ELPA_KERNEL = THIS_REAL_ELPA_KERNEL_API
Andreas Marek's avatar
Andreas Marek committed
228
   else
229

230
231
232
     ! if kernel is not choosen via api
     ! check whether set by environment variable
     THIS_REAL_ELPA_KERNEL = get_actual_real_kernel()
Andreas Marek's avatar
Andreas Marek committed
233
234
235
236
   endif

   ! check whether choosen kernel is allowed
   if (check_allowed_real_kernels(THIS_REAL_ELPA_KERNEL)) then
237

238
239
240
241
242
243
244
245
246
247
248
     if (my_pe == 0) then
       write(error_unit,*) " "
       write(error_unit,*) "The choosen kernel ",REAL_ELPA_KERNEL_NAMES(THIS_REAL_ELPA_KERNEL)
       write(error_unit,*) "is not in the list of the allowed kernels!"
       write(error_unit,*) " "
       write(error_unit,*) "Allowed kernels are:"
       do i=1,size(REAL_ELPA_KERNEL_NAMES(:))
         if (AVAILABLE_REAL_ELPA_KERNELS(i) .ne. 0) then
           write(error_unit,*) REAL_ELPA_KERNEL_NAMES(i)
         endif
       enddo
Andreas Marek's avatar
Andreas Marek committed
249

250
251
252
253
       write(error_unit,*) " "
       write(error_unit,*) "The defaul kernel REAL_ELPA_KERNEL_GENERIC will be used !"
     endif
     THIS_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_GENERIC
Andreas Marek's avatar
Andreas Marek committed
254
255

   endif
256
257

   ! Choose bandwidth, must be a multiple of nblk, set to a value >= 32
258
259
   ! For Intel(R) Xeon(R) E5 v2 and v3, better use 64 instead of 32!
   nbw = (63/nblk+1)*nblk
260
261
262
263
264
265
266
267
268

   num_blocks = (na-1)/nbw + 1

   allocate(tmat(nbw,nbw,num_blocks))

   ! Reduction full -> band

   ttt0 = MPI_Wtime()
   ttts = ttt0
269
   call bandred_real(na, a, lda, nblk, nbw, mpi_comm_rows, mpi_comm_cols, &
270
                     tmat, wantDebug, success, useQRActual)
271
   if (.not.(success)) return
272
   ttt1 = MPI_Wtime()
273
   if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
274
      write(error_unit,*) 'Time bandred_real               :',ttt1-ttt0
275
276
277
278
279
280

   ! Reduction band -> tridiagonal

   allocate(e(na))

   ttt0 = MPI_Wtime()
Andreas Marek's avatar
Andreas Marek committed
281
282
   call tridiag_band_real(na, nbw, nblk, a, lda, ev, e, mpi_comm_rows, &
                          mpi_comm_cols, mpi_comm_all)
283
   ttt1 = MPI_Wtime()
284
   if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
285
      write(error_unit,*) 'Time tridiag_band_real          :',ttt1-ttt0
286
287
288
289
290
291
292
293
294
295

   call mpi_bcast(ev,na,MPI_REAL8,0,mpi_comm_all,mpierr)
   call mpi_bcast(e,na,MPI_REAL8,0,mpi_comm_all,mpierr)

   ttt1 = MPI_Wtime()
   time_evp_fwd = ttt1-ttts

   ! Solve tridiagonal system

   ttt0 = MPI_Wtime()
296
   call solve_tridi(na, nev, ev, e, q, ldq, nblk, mpi_comm_rows,  &
297
                    mpi_comm_cols, wantDebug, success)
298
299
   if (.not.(success)) return

300
   ttt1 = MPI_Wtime()
301
302
   if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
     write(error_unit,*) 'Time solve_tridi                :',ttt1-ttt0
303
304
305
306
307
308
309
310
   time_evp_solve = ttt1-ttt0
   ttts = ttt1

   deallocate(e)

   ! Backtransform stage 1

   ttt0 = MPI_Wtime()
311
   call trans_ev_tridi_to_band_real(na, nev, nblk, nbw, q, ldq, mpi_comm_rows, &
312
                                    mpi_comm_cols, wantDebug, success, THIS_REAL_ELPA_KERNEL)
313
   if (.not.(success)) return
314
   ttt1 = MPI_Wtime()
315
   if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
316
      write(error_unit,*) 'Time trans_ev_tridi_to_band_real:',ttt1-ttt0
317
318
319
320
321
322
323

   ! We can now deallocate the stored householder vectors
   deallocate(hh_trans_real)

   ! Backtransform stage 2

   ttt0 = MPI_Wtime()
324
325
   call trans_ev_band_to_full_real(na, nev, nblk, nbw, a, lda, tmat, q, ldq, mpi_comm_rows, &
                                   mpi_comm_cols, useQRActual)
326
   ttt1 = MPI_Wtime()
327
   if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
328
      write(error_unit,*) 'Time trans_ev_band_to_full_real :',ttt1-ttt0
329
330
331
   time_evp_back = ttt1-ttts

   deallocate(tmat)
332
333
334
#ifdef HAVE_DETAILED_TIMINGS
   call timer%stop("solve_evp_real_2stage")
#endif
335
336
1  format(a,f10.3)

337
end function solve_evp_real_2stage
338
339
340
341
342

!-------------------------------------------------------------------------------

!-------------------------------------------------------------------------------

343
function solve_evp_complex_2stage(na, nev, a, lda, ev, q, ldq, nblk, &
Andreas Marek's avatar
Andreas Marek committed
344
                                    mpi_comm_rows, mpi_comm_cols,      &
345
                                    mpi_comm_all, THIS_COMPLEX_ELPA_KERNEL_API) result(success)
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380

!-------------------------------------------------------------------------------
!  solve_evp_complex_2stage: Solves the complex eigenvalue problem with a 2 stage approach
!
!  Parameters
!
!  na          Order of matrix a
!
!  nev         Number of eigenvalues needed
!
!  a(lda,*)    Distributed matrix for which eigenvalues are to be computed.
!              Distribution is like in Scalapack.
!              The full matrix must be set (not only one half like in scalapack).
!              Destroyed on exit (upper and lower half).
!
!  lda         Leading dimension of a
!
!  ev(na)      On output: eigenvalues of a, every processor gets the complete set
!
!  q(ldq,*)    On output: Eigenvectors of a
!              Distribution is like in Scalapack.
!              Must be always dimensioned to the full size (corresponding to (na,na))
!              even if only a part of the eigenvalues is needed.
!
!  ldq         Leading dimension of q
!
!  nblk        blocksize of cyclic distribution, must be the same in both directions!
!
!  mpi_comm_rows
!  mpi_comm_cols
!              MPI-Communicators for rows/columns
!  mpi_comm_all
!              MPI-Communicator for the total processor set
!
!-------------------------------------------------------------------------------
381
382
383
#ifdef HAVE_DETAILED_TIMINGS
 use timings
#endif
384
   implicit none
Andreas Marek's avatar
Andreas Marek committed
385
386
   integer, intent(in), optional :: THIS_COMPLEX_ELPA_KERNEL_API
   integer                       :: THIS_COMPLEX_ELPA_KERNEL
387
388
389
390
391
392
393
394
395
396
   integer, intent(in)           :: na, nev, lda, ldq, nblk, mpi_comm_rows, mpi_comm_cols, mpi_comm_all
   complex*16, intent(inout)     :: a(lda,*), q(ldq,*)
   real*8, intent(inout)         :: ev(na)

   integer                       :: my_prow, my_pcol, np_rows, np_cols, mpierr, my_pe, n_pes
   integer                       :: l_cols, l_rows, l_cols_nev, nbw, num_blocks
   complex*16, allocatable       :: tmat(:,:,:)
   real*8, allocatable           :: q_real(:,:), e(:)
   real*8                        :: ttt0, ttt1, ttts
   integer                       :: i
397

398
399
400
   logical                       :: success, wantDebug
   logical, save                 :: firstCall = .true.

401
402
403
#ifdef HAVE_DETAILED_TIMINGS
   call timer%start("solve_evp_complex_2stage")
#endif
Andreas Marek's avatar
Andreas Marek committed
404
405
   call mpi_comm_rank(mpi_comm_all,my_pe,mpierr)
   call mpi_comm_size(mpi_comm_all,n_pes,mpierr)
406
407
408
409
410

   call mpi_comm_rank(mpi_comm_rows,my_prow,mpierr)
   call mpi_comm_size(mpi_comm_rows,np_rows,mpierr)
   call mpi_comm_rank(mpi_comm_cols,my_pcol,mpierr)
   call mpi_comm_size(mpi_comm_cols,np_cols,mpierr)
411

412
413
414
415
416
417
418
419
   wantDebug = .false.
   if (firstCall) then
     ! are debug messages desired?
     wantDebug = debug_messages_via_environment_variable()
     firstCall = .false.
   endif


420
421
   success = .true.

422
423
424
   if (present(THIS_COMPLEX_ELPA_KERNEL_API)) then
     ! user defined kernel via the optional argument in the API call
     THIS_COMPLEX_ELPA_KERNEL = THIS_COMPLEX_ELPA_KERNEL_API
Andreas Marek's avatar
Andreas Marek committed
425
   else
426
427
428
     ! if kernel is not choosen via api
     ! check whether set by environment variable
     THIS_COMPLEX_ELPA_KERNEL = get_actual_complex_kernel()
Andreas Marek's avatar
Andreas Marek committed
429
   endif
430

Andreas Marek's avatar
Andreas Marek committed
431
432
   ! check whether choosen kernel is allowed
   if (check_allowed_complex_kernels(THIS_COMPLEX_ELPA_KERNEL)) then
433

434
435
436
437
438
439
440
441
442
443
444
     if (my_pe == 0) then
       write(error_unit,*) " "
       write(error_unit,*) "The choosen kernel ",COMPLEX_ELPA_KERNEL_NAMES(THIS_COMPLEX_ELPA_KERNEL)
       write(error_unit,*) "is not in the list of the allowed kernels!"
       write(error_unit,*) " "
       write(error_unit,*) "Allowed kernels are:"
       do i=1,size(COMPLEX_ELPA_KERNEL_NAMES(:))
         if (AVAILABLE_COMPLEX_ELPA_KERNELS(i) .ne. 0) then
           write(error_unit,*) COMPLEX_ELPA_KERNEL_NAMES(i)
         endif
       enddo
Andreas Marek's avatar
Andreas Marek committed
445

446
447
448
449
       write(error_unit,*) " "
       write(error_unit,*) "The defaul kernel COMPLEX_ELPA_KERNEL_GENERIC will be used !"
     endif
     THIS_COMPLEX_ELPA_KERNEL = COMPLEX_ELPA_KERNEL_GENERIC
Andreas Marek's avatar
Andreas Marek committed
450
   endif
451
452
453
454
455
456
457
458
459
460
461
462
   ! Choose bandwidth, must be a multiple of nblk, set to a value >= 32

   nbw = (31/nblk+1)*nblk

   num_blocks = (na-1)/nbw + 1

   allocate(tmat(nbw,nbw,num_blocks))

   ! Reduction full -> band

   ttt0 = MPI_Wtime()
   ttts = ttt0
463
   call bandred_complex(na, a, lda, nblk, nbw, mpi_comm_rows, mpi_comm_cols, &
464
                        tmat, wantDebug, success)
465
466
467
468
469
470
   if (.not.(success)) then
#ifdef HAVE_DETAILED_TIMINGS
     call timer%stop()
#endif
     return
   endif
471
   ttt1 = MPI_Wtime()
472
   if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
473
      write(error_unit,*) 'Time bandred_complex               :',ttt1-ttt0
474
475
476
477
478
479
480
481

   ! Reduction band -> tridiagonal

   allocate(e(na))

   ttt0 = MPI_Wtime()
   call tridiag_band_complex(na, nbw, nblk, a, lda, ev, e, mpi_comm_rows, mpi_comm_cols, mpi_comm_all)
   ttt1 = MPI_Wtime()
482
   if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
483
      write(error_unit,*) 'Time tridiag_band_complex          :',ttt1-ttt0
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499

   call mpi_bcast(ev,na,MPI_REAL8,0,mpi_comm_all,mpierr)
   call mpi_bcast(e,na,MPI_REAL8,0,mpi_comm_all,mpierr)

   ttt1 = MPI_Wtime()
   time_evp_fwd = ttt1-ttts

   l_rows = local_index(na, my_prow, np_rows, nblk, -1) ! Local rows of a and q
   l_cols = local_index(na, my_pcol, np_cols, nblk, -1) ! Local columns of q
   l_cols_nev = local_index(nev, my_pcol, np_cols, nblk, -1) ! Local columns corresponding to nev

   allocate(q_real(l_rows,l_cols))

   ! Solve tridiagonal system

   ttt0 = MPI_Wtime()
500
   call solve_tridi(na, nev, ev, e, q_real, ubound(q_real,1), nblk, &
501
                    mpi_comm_rows, mpi_comm_cols, wantDebug, success)
502
503
   if (.not.(success)) return

504
   ttt1 = MPI_Wtime()
505
   if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times)  &
506
      write(error_unit,*) 'Time solve_tridi                   :',ttt1-ttt0
507
508
509
510
511
512
513
514
515
516
   time_evp_solve = ttt1-ttt0
   ttts = ttt1

   q(1:l_rows,1:l_cols_nev) = q_real(1:l_rows,1:l_cols_nev)

   deallocate(e, q_real)

   ! Backtransform stage 1

   ttt0 = MPI_Wtime()
Andreas Marek's avatar
Andreas Marek committed
517
   call trans_ev_tridi_to_band_complex(na, nev, nblk, nbw, q, ldq,  &
518
                                       mpi_comm_rows, mpi_comm_cols,&
519
                                       wantDebug, success,THIS_COMPLEX_ELPA_KERNEL)
520
   if (.not.(success)) return
521
   ttt1 = MPI_Wtime()
522
   if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
523
      write(error_unit,*) 'Time trans_ev_tridi_to_band_complex:',ttt1-ttt0
524
525
526
527
528
529
530
531
532

   ! We can now deallocate the stored householder vectors
   deallocate(hh_trans_complex)

   ! Backtransform stage 2

   ttt0 = MPI_Wtime()
   call trans_ev_band_to_full_complex(na, nev, nblk, nbw, a, lda, tmat, q, ldq, mpi_comm_rows, mpi_comm_cols)
   ttt1 = MPI_Wtime()
533
   if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
534
      write(error_unit,*) 'Time trans_ev_band_to_full_complex :',ttt1-ttt0
535
536
537
   time_evp_back = ttt1-ttts

   deallocate(tmat)
538
539
540
#ifdef HAVE_DETAILED_TIMINGS
   call timer%stop("solve_evp_complex_2stage")
#endif
541
542
543

1  format(a,f10.3)

544
end function solve_evp_complex_2stage
545
546
547

!-------------------------------------------------------------------------------

548
subroutine bandred_real(na, a, lda, nblk, nbw, mpi_comm_rows, mpi_comm_cols, &
549
                        tmat, wantDebug, success, useQR)
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577

!-------------------------------------------------------------------------------
!  bandred_real: Reduces a distributed symmetric matrix to band form
!
!  Parameters
!
!  na          Order of matrix
!
!  a(lda,*)    Distributed matrix which should be reduced.
!              Distribution is like in Scalapack.
!              Opposed to Scalapack, a(:,:) must be set completely (upper and lower half)
!              a(:,:) is overwritten on exit with the band and the Householder vectors
!              in the upper half.
!
!  lda         Leading dimension of a
!
!  nblk        blocksize of cyclic distribution, must be the same in both directions!
!
!  nbw         semi bandwith of output matrix
!
!  mpi_comm_rows
!  mpi_comm_cols
!              MPI-Communicators for rows/columns
!
!  tmat(nbw,nbw,num_blocks)    where num_blocks = (na-1)/nbw + 1
!              Factors for the Householder vectors (returned), needed for back transformation
!
!-------------------------------------------------------------------------------
578
579
#ifdef HAVE_DETAILED_TIMINGS
 use timings
580
581
582
#endif
#ifdef WITH_OPENMP
   use omp_lib
583
#endif
584
   implicit none
585
   
586
587
   integer             :: na, lda, nblk, nbw, mpi_comm_rows, mpi_comm_cols
   real*8              :: a(lda,*), tmat(nbw,nbw,*)
588

589
590
   integer             :: my_prow, my_pcol, np_rows, np_cols, mpierr
   integer             :: l_cols, l_rows
591
592
   integer             :: i, j, lcs, lce, lrs, lre, lc, lr, cur_pcol, n_cols, nrow
   integer             :: istep, ncol, lch, lcx, nlc, mynlc
593
   integer             :: tile_size, l_rows_tile, l_cols_tile
594

595
   real*8              :: vnorm2, xf, aux1(nbw), aux2(nbw), vrl, tau, vav(nbw,nbw)
596

597
   real*8, allocatable :: tmp(:,:), vr(:), vmr(:,:), umc(:,:)
598

599
600
601
602
603
   ! needed for blocked QR decomposition
   integer             :: PQRPARAM(11), work_size
   real*8              :: dwork_size(1)
   real*8, allocatable :: work_blocked(:), tauvector(:), blockheuristic(:)

604
   logical, intent(in) :: wantDebug
605
606
   logical, intent(out):: success

607
608
   logical, intent(in) :: useQR

609
610
   integer :: mystart, myend, m_way, n_way, work_per_thread, m_id, n_id, n_threads, ii, pp, transformChunkSize

611
612
613
#ifdef HAVE_DETAILED_TIMINGS
   call timer%start("bandred_real")
#endif
614
615
616
617
   call mpi_comm_rank(mpi_comm_rows,my_prow,mpierr)
   call mpi_comm_size(mpi_comm_rows,np_rows,mpierr)
   call mpi_comm_rank(mpi_comm_cols,my_pcol,mpierr)
   call mpi_comm_size(mpi_comm_cols,np_cols,mpierr)
618
   success = .true.
619
620


621
   ! Semibandwith nbw must be a multiple of blocksize nblk
622
623
   if (mod(nbw,nblk)/=0) then
     if (my_prow==0 .and. my_pcol==0) then
624
625
626
627
       if (wantDebug) then
         write(error_unit,*) 'ELPA2_bandred_real: ERROR: nbw=',nbw,', nblk=',nblk
         write(error_unit,*) 'ELPA2_bandred_real: ELPA2 works only for nbw==n*nblk'
       endif
628
       success = .false.
Lorenz Huedepohl's avatar
Lorenz Huedepohl committed
629
       return
630
     endif
631
632
633
634
635
636
637
638
639
640
   endif

   ! Matrix is split into tiles; work is done only for tiles on the diagonal or above

   tile_size = nblk*least_common_multiple(np_rows,np_cols) ! minimum global tile size
   tile_size = ((128*max(np_rows,np_cols)-1)/tile_size+1)*tile_size ! make local tiles at least 128 wide

   l_rows_tile = tile_size/np_rows ! local rows of a tile
   l_cols_tile = tile_size/np_cols ! local cols of a tile

641
642
643
644
645
646
647
   if (useQR) then
     if (which_qr_decomposition == 1) then
       call qr_pqrparam_init(pqrparam,    nblk,'M',0,   nblk,'M',0,   nblk,'M',1,'s')
       allocate(tauvector(na))
       allocate(blockheuristic(nblk))
       l_rows = local_index(na, my_prow, np_rows, nblk, -1)
       allocate(vmr(max(l_rows,1),na))
648

649
       call qr_pdgeqrf_2dcomm(a, lda, vmr, max(l_rows,1), tauvector(1), tmat(1,1,1), nbw, dwork_size(1), -1, na, &
650
                             nbw, nblk, nblk, na, na, 1, 0, PQRPARAM, mpi_comm_rows, mpi_comm_cols, blockheuristic)
651
652
       work_size = dwork_size(1)
       allocate(work_blocked(work_size))
653

654
655
656
       work_blocked = 0.0d0
       deallocate(vmr)
     endif
657
658
   endif

659
660
   do istep = (na-1)/nbw, 1, -1

661
     n_cols = MIN(na,(istep+1)*nbw) - istep*nbw ! Number of columns in current step
662

663
664
665
     ! Number of local columns/rows of remaining matrix
     l_cols = local_index(istep*nbw, my_pcol, np_cols, nblk, -1)
     l_rows = local_index(istep*nbw, my_prow, np_rows, nblk, -1)
666

667
     ! Allocate vmr and umc to their exact sizes so that they can be used in bcasts and reduces
668

669
670
     allocate(vmr(max(l_rows,1),2*n_cols))
     allocate(umc(max(l_cols,1),2*n_cols))
671

672
     allocate(vr(l_rows+1))
673

674
675
676
     vmr(1:l_rows,1:n_cols) = 0.
     vr(:) = 0
     tmat(:,:,istep) = 0
677

678
     ! Reduce current block to lower triangular form
679
680
681
682
683
684
685
686
687
688

     if (useQR) then
       if (which_qr_decomposition == 1) then
         call qr_pdgeqrf_2dcomm(a, lda, vmr, max(l_rows,1), tauvector(1), &
                                  tmat(1,1,istep), nbw, work_blocked,       &
                                  work_size, na, n_cols, nblk, nblk,        &
                                  istep*nbw+n_cols-nbw, istep*nbw+n_cols, 1,&
                                  0, PQRPARAM, mpi_comm_rows, mpi_comm_cols,&
                                  blockheuristic)
       endif
689
     else
690

691
       do lc = n_cols, 1, -1
692

693
694
         ncol = istep*nbw + lc ! absolute column number of householder vector
         nrow = ncol - nbw ! Absolute number of pivot row
695

696
697
         lr  = local_index(nrow, my_prow, np_rows, nblk, -1) ! current row length
         lch = local_index(ncol, my_pcol, np_cols, nblk, -1) ! HV local column number
698

699
         tau = 0
700

701
         if (nrow == 1) exit ! Nothing to do
702

703
         cur_pcol = pcol(ncol, nblk, np_cols) ! Processor column owning current block
704

705
         if (my_pcol==cur_pcol) then
706

707
708
           ! Get vector to be transformed; distribute last element and norm of
           ! remaining elements to all procs in current column
709

710
           vr(1:lr) = a(1:lr,lch) ! vector to be transformed
711

712
           if (my_prow==prow(nrow, nblk, np_rows)) then
713
714
715
716
717
718
             aux1(1) = dot_product(vr(1:lr-1),vr(1:lr-1))
             aux1(2) = vr(lr)
           else
             aux1(1) = dot_product(vr(1:lr),vr(1:lr))
             aux1(2) = 0.
           endif
719

720
           call mpi_allreduce(aux1,aux2,2,MPI_REAL8,MPI_SUM,mpi_comm_rows,mpierr)
721

722
723
           vnorm2 = aux2(1)
           vrl    = aux2(2)
724

725
           ! Householder transformation
726

727
           call hh_transform_real(vrl, vnorm2, xf, tau)
728

729
           ! Scale vr and store Householder vector for back transformation
730

731
           vr(1:lr) = vr(1:lr) * xf
732
           if (my_prow==prow(nrow, nblk, np_rows)) then
733
734
735
736
737
             a(1:lr-1,lch) = vr(1:lr-1)
             a(lr,lch) = vrl
             vr(lr) = 1.
           else
             a(1:lr,lch) = vr(1:lr)
738
           endif
739

740
         endif
741

742
         ! Broadcast Householder vector and tau along columns
743

744
745
746
747
748
         vr(lr+1) = tau
         call MPI_Bcast(vr,lr+1,MPI_REAL8,cur_pcol,mpi_comm_cols,mpierr)
         vmr(1:lr,lc) = vr(1:lr)
         tau = vr(lr+1)
         tmat(lc,lc,istep) = tau ! Store tau in diagonal of tmat
749

750
751
         ! Transform remaining columns in current block with Householder vector
         ! Local dot product
752

753
         aux1 = 0
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
#ifdef WITH_OPENMP
         !Open up one omp region to avoid paying openmp overhead.
         !This does not help performance due to the addition of two openmp barriers around the MPI call,
         !But in the future this may be beneficial if these barriers are replaced with a faster implementation

         !$omp parallel private(mynlc, j, lcx, ii, pp ) shared(aux1)
         mynlc = 0 ! number of local columns

         !This loop does not have independent iterations,
         !'mynlc' is incremented each iteration, and it is difficult to remove this dependency 
         !Thus each thread executes every iteration of the loop, except it only does the work if it 'owns' that iteration
         !That is, a thread only executes the work associated with an iteration if its thread id is congruent to 
         !the iteration number modulo the number of threads
         do j=1,lc-1
           lcx = local_index(istep*nbw+j, my_pcol, np_cols, nblk, 0)
           if (lcx>0 ) then
             mynlc = mynlc+1
             if ( mod((j-1), omp_get_num_threads()) .eq. omp_get_thread_num() ) then
                 if (lr>0) aux1(mynlc) = dot_product(vr(1:lr),a(1:lr,lcx))
             endif
           endif
         enddo
         
         ! Get global dot products
         !$omp barrier
         !$omp single 
         if (mynlc>0) call mpi_allreduce(aux1,aux2,mynlc,MPI_REAL8,MPI_SUM,mpi_comm_rows,mpierr)
         !$omp end single 
         !$omp barrier
783

784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
         ! Transform
         transformChunkSize=32
         mynlc = 0
         do j=1,lc-1
           lcx = local_index(istep*nbw+j, my_pcol, np_cols, nblk, 0)
           if (lcx>0) then
             mynlc = mynlc+1
             !This loop could be parallelized with an openmp pragma with static scheduling and chunk size 32
             !However, for some reason this is slower than doing it manually, so it is parallelized as below.
             do ii=omp_get_thread_num()*transformChunkSize,lr,omp_get_num_threads()*transformChunkSize
                do pp = 1,transformChunkSize
                    if (pp + ii > lr) exit
                        a(ii+pp,lcx) = a(ii+pp,lcx) - tau*aux2(mynlc)*vr(ii+pp)
                enddo
             enddo
           endif
         enddo
         !$omp end parallel
#else
803
804
805
806
807
808
809
810
         nlc = 0 ! number of local columns
         do j=1,lc-1
           lcx = local_index(istep*nbw+j, my_pcol, np_cols, nblk, 0)
           if (lcx>0) then
             nlc = nlc+1
             if (lr>0) aux1(nlc) = dot_product(vr(1:lr),a(1:lr,lcx))
           endif
         enddo
811

812
813
         ! Get global dot products
         if (nlc>0) call mpi_allreduce(aux1,aux2,nlc,MPI_REAL8,MPI_SUM,mpi_comm_rows,mpierr)
814

815
         ! Transform
816

817
818
819
820
821
822
823
824
         nlc = 0
         do j=1,lc-1
           lcx = local_index(istep*nbw+j, my_pcol, np_cols, nblk, 0)
           if (lcx>0) then
             nlc = nlc+1
             a(1:lr,lcx) = a(1:lr,lcx) - tau*aux2(nlc)*vr(1:lr)
           endif
         enddo
825
#endif
826
       enddo
827

828
829
       ! Calculate scalar products of stored Householder vectors.
       ! This can be done in different ways, we use dsyrk
830

831
832
       vav = 0
       if (l_rows>0) &
833
           call dsyrk('U','T',n_cols,l_rows,1.d0,vmr,ubound(vmr,1),0.d0,vav,ubound(vav,1))
834
       call symm_matrix_allreduce(n_cols,vav,ubound(vav,1),mpi_comm_rows)
835

836
       ! Calculate triangular matrix T for block Householder Transformation
837

838
839
840
841
842
843
844
       do lc=n_cols,1,-1
         tau = tmat(lc,lc,istep)
         if (lc<n_cols) then
           call dtrmv('U','T','N',n_cols-lc,tmat(lc+1,lc+1,istep),ubound(tmat,1),vav(lc+1,lc),1)
           tmat(lc,lc+1:n_cols,istep) = -tau * vav(lc+1:n_cols,lc)
         endif
       enddo
845
     endif
846

847
    ! Transpose vmr -> vmc (stored in umc, second half)
848

849
    call elpa_transpose_vectors  (vmr, ubound(vmr,1), mpi_comm_rows, &
850
851
852
                                    umc(1,n_cols+1), ubound(umc,1), mpi_comm_cols, &
                                    1, istep*nbw, n_cols, nblk)

853
854
855
856
    ! Calculate umc = A**T * vmr
    ! Note that the distributed A has to be transposed
    ! Opposed to direct tridiagonalization there is no need to use the cache locality
    ! of the tiles, so we can use strips of the matrix
857
    !Code for Algorithm 4
858

859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
    n_way = 1
#ifdef WITH_OPENMP
    n_way = omp_get_max_threads()   
#endif
    !umc(1:l_cols,1:n_cols) = 0.d0
    !vmr(1:l_rows,n_cols+1:2*n_cols) = 0

    !$omp parallel private( i,lcs,lce,lrs,lre)

    if(n_way > 1) then
        !$omp do
        do i=1,min(l_cols_tile, l_cols)
            umc(i,1:n_cols) = 0.d0
        enddo
        !$omp do
        do i=1,l_rows
            vmr(i,n_cols+1:2*n_cols) = 0.d0
        enddo
        if (l_cols>0 .and. l_rows>0) then

          !SYMM variant 4
          !Partitioned Matrix Expression:
          ! Ct = Atl Bt + Atr Bb
          ! Cb = Atr' Bt + Abl Bb
          !
          !Loop invariant:
          ! Ct = Atl Bt + Atr Bb
          !
          !Update:
          ! C1 = A10'B0 + A11B1 + A21 B2
          !
          !This algorithm chosen because in this algoirhtm, the loop around the dgemm calls
          !is easily parallelized, and regardless of choise of algorithm,
          !the startup cost for parallelizing the dgemms inside the loop is too great

          !$omp do schedule(static,1)
          do i=0,(istep*nbw-1)/tile_size
            lcs = i*l_cols_tile+1                   ! local column start
            lce = min(l_cols, (i+1)*l_cols_tile)    ! local column end

            lrs = i*l_rows_tile+1                   ! local row start
            lre = min(l_rows, (i+1)*l_rows_tile)    ! local row end

            !C1 += [A11 A12] [B1
            !                 B2]
            if( lre > lrs .and. l_cols > lcs ) then
            call DGEMM('N','N', lre-lrs+1, n_cols, l_cols-lcs+1,    &
                       1.d0, a(lrs,lcs), ubound(a,1),               &
                             umc(lcs,n_cols+1), ubound(umc,1),      &
                       0.d0, vmr(lrs,n_cols+1), ubound(vmr,1))
            endif

            ! C1 += A10' B0
            if( lce > lcs .and. i > 0 ) then
            call DGEMM('T','N', lce-lcs+1, n_cols, lrs-1,           &
                       1.d0, a(1,lcs),   ubound(a,1),               &
                             vmr(1,1),   ubound(vmr,1),             &
                       0.d0, umc(lcs,1), ubound(umc,1))
            endif
          enddo
        endif
    else
        umc(1:l_cols,1:n_cols) = 0.d0
        vmr(1:l_rows,n_cols+1:2*n_cols) = 0
        if (l_cols>0 .and. l_rows>0) then
          do i=0,(istep*nbw-1)/tile_size

            lcs = i*l_cols_tile+1
            lce = min(l_cols,(i+1)*l_cols_tile)
            if (lce<lcs) cycle

            lre = min(l_rows,(i+1)*l_rows_tile)
            call DGEMM('T','N',lce-lcs+1,n_cols,lre,1.d0,a(1,lcs),ubound(a,1), &
                         vmr,ubound(vmr,1),1.d0,umc(lcs,1),ubound(umc,1))

            if (i==0) cycle
            lre = min(l_rows,i*l_rows_tile)
            call DGEMM('N','N',lre,n_cols,lce-lcs+1,1.d0,a(1,lcs),lda, &
                         umc(lcs,n_cols+1),ubound(umc,1),1.d0,vmr(1,n_cols+1),ubound(vmr,1))
          enddo
        endif
940
    endif
941
    !$omp end parallel
942

943
944
945
946
    ! Sum up all ur(:) parts along rows and add them to the uc(:) parts
    ! on the processors containing the diagonal
    ! This is only necessary if ur has been calculated, i.e. if the
    ! global tile size is smaller than the global remaining matrix
947
948
949
950
951
    ! Or if we used the Algorithm 4
    if (tile_size < istep*nbw .or. n_way > 1) then
    call elpa_reduce_add_vectors  (vmr(1,n_cols+1),ubound(vmr,1),mpi_comm_rows, &
                                   umc, ubound(umc,1), mpi_comm_cols, &
                                   istep*nbw, n_cols, nblk)
952
    endif
953

954
955
956
957
958
959
    if (l_cols>0) then
      allocate(tmp(l_cols,n_cols))
      call mpi_allreduce(umc,tmp,l_cols*n_cols,MPI_REAL8,MPI_SUM,mpi_comm_rows,mpierr)
      umc(1:l_cols,1:n_cols) = tmp(1:l_cols,1:n_cols)
      deallocate(tmp)
    endif
960

961
    ! U = U * Tmat**T
962

963
    call dtrmm('Right','Upper','Trans','Nonunit',l_cols,n_cols,1.d0,tmat(1,1,istep),ubound(tmat,1),umc,ubound(umc,1))
964

965
    ! VAV = Tmat * V**T * A * V * Tmat**T = (U*Tmat**T)**T * V * Tmat**T
966

967
968
    call dgemm('T','N',n_cols,n_cols,l_cols,1.d0,umc,ubound(umc,1),umc(1,n_cols+1),ubound(umc,1),0.d0,vav,ubound(vav,1))
    call dtrmm('Right','Upper','Trans','Nonunit',n_cols,n_cols,1.d0,tmat(1,1,istep),ubound(tmat,1),vav,ubound(vav,1))
969

970
    call symm_matrix_allreduce(n_cols,vav,ubound(vav,1),mpi_comm_cols)
971

972
973
    ! U = U - 0.5 * V * VAV
    call dgemm('N','N',l_cols,n_cols,n_cols,-0.5d0,umc(1,n_cols+1),ubound(umc,1),vav,ubound(vav,1),1.d0,umc,ubound(umc,1))
974

975
    ! Transpose umc -> umr (stored in vmr, second half)
976

977
978
979
    call elpa_transpose_vectors  (umc, ubound(umc,1), mpi_comm_cols, &
                                   vmr(1,n_cols+1), ubound(vmr,1), mpi_comm_rows, &
                                   1, istep*nbw, n_cols, nblk)
980

981
    ! A = A - V*U**T - U*V**T
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
#ifdef WITH_OPENMP
    !$omp parallel private( ii, i, lcs, lce, lre, n_way, m_way, m_id, n_id, work_per_thread, mystart, myend  )
    n_threads = omp_get_num_threads()
    if(mod(n_threads, 2) == 0) then
        n_way = 2
    else
        n_way = 1
    endif
    
    m_way = n_threads / n_way
    
    m_id = mod(omp_get_thread_num(),  m_way)
    n_id = omp_get_thread_num() / m_way
    
    do ii=n_id*tile_size,(istep*nbw-1),tile_size*n_way
      i = ii / tile_size
      lcs = i*l_cols_tile+1
      lce = min(l_cols,(i+1)*l_cols_tile)
      lre = min(l_rows,(i+1)*l_rows_tile)
      if (lce<lcs .or. lre<1) cycle

      !Figure out this thread's range
      work_per_thread = lre / m_way
      if (work_per_thread * m_way < lre) work_per_thread = work_per_thread + 1
      mystart = m_id * work_per_thread + 1
      myend   = mystart + work_per_thread - 1
      if( myend > lre ) myend = lre
      if( myend-mystart+1 < 1) cycle

      call dgemm('N','T',myend-mystart+1, lce-lcs+1, 2*n_cols, -1.d0, &
                  vmr(mystart, 1), ubound(vmr,1), umc(lcs,1), ubound(umc,1), &
                  1.d0,a(mystart,lcs),ubound(a,1))
    enddo
    !$omp end parallel
1016

1017
#else
1018
1019
1020
1021
1022
1023
1024
1025
1026
    do i=0,(istep*nbw-1)/tile_size
      lcs = i*l_cols_tile+1
      lce = min(l_cols,(i+1)*l_cols_tile)
      lre = min(l_rows,(i+1)*l_rows_tile)
      if (lce<lcs .or. lre<1) cycle
      call dgemm('N','T',lre,lce-lcs+1,2*n_cols,-1.d0, &
                  vmr,ubound(vmr,1),umc(lcs,1),ubound(umc,1), &
                  1.d0,a(1,lcs),lda)
    enddo
1027
#endif
1028
    deallocate(vmr, umc, vr)
1029

1030
  enddo
1031

1032
1033
1034
1035
1036
  if (useQR) then
    if (which_qr_decomposition == 1) then
      deallocate(work_blocked)
      deallocate(tauvector)
    endif
1037
  endif
1038

Andreas Marek's avatar
Andreas Marek committed
1039
1040
1041
#ifdef HAVE_DETAILED_TIMINGS
  call timer%stop("bandred_real")
#endif
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
end subroutine bandred_real

!-------------------------------------------------------------------------------

subroutine symm_matrix_allreduce(n,a,lda,comm)

!-------------------------------------------------------------------------------
!  symm_matrix_allreduce: Does an mpi_allreduce for a symmetric matrix A.
!  On entry, only the upper half of A needs to be set
!  On exit, the complete matrix is set
!-------------------------------------------------------------------------------
Andreas Marek's avatar
Andreas Marek committed
1053
1054
1055
#ifdef HAVE_DETAILED_TIMINGS
 use timings
#endif
1056
   implicit none
Andreas Marek's avatar
Andreas Marek committed
1057
1058
1059
1060
1061
   integer  :: n, lda, comm
   real*8   :: a(lda,*)

   integer  :: i, nc, mpierr
   real*8   :: h1(n*n), h2(n*n)
1062

Andreas Marek's avatar
Andreas Marek committed
1063
1064
1065
#ifdef HAVE_DETAILED_TIMINGS
  call timer%start("symm_matrix_allreduce")
#endif
1066
1067
1068

   nc = 0
   do i=1,n
1069
1070
     h1(nc+1:nc+i) = a(1:i,i)
     nc = nc+i
1071
1072
1073
1074
1075
1076
   enddo

   call mpi_allreduce(h1,h2,nc,MPI_REAL8,MPI_SUM,comm,mpierr)

   nc = 0
   do i=1,n
1077
1078
1079
     a(1:i,i) = h2(nc+1:nc+i)
     a(i,1:i-1) = a(1:i-1,i)
     nc = nc+i
1080
1081
   enddo

Andreas Marek's avatar
Andreas Marek committed
1082
1083
1084
1085
#ifdef HAVE_DETAILED_TIMINGS
  call timer%stop("symm_matrix_allreduce")
#endif

1086
1087
1088
1089
end subroutine symm_matrix_allreduce

!-------------------------------------------------------------------------------

1090
1091
subroutine trans_ev_band_to_full_real(na, nqc, nblk, nbw, a, lda, tmat, q, ldq, mpi_comm_rows, &
                                      mpi_comm_cols, useQR)
1092

Andreas Marek's avatar
Andreas Marek committed
1093

1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
!-------------------------------------------------------------------------------
!  trans_ev_band_to_full_real:
!  Transforms the eigenvectors of a band matrix back to the eigenvectors of the original matrix
!
!  Parameters
!
!  na          Order of matrix a, number of rows of matrix q
!
!  nqc         Number of columns of matrix q
!
!  nblk        blocksize of cyclic distribution, must be the same in both directions!
!
!  nbw         semi bandwith
!
!  a(lda,*)    Matrix containing the Householder vectors (i.e. matrix a after bandred_real)
!              Distribution is like in Scalapack.
!
!  lda         Leading dimension of a
!
!  tmat(nbw,nbw,.) Factors returned by bandred_real
!
!  q           On input: Eigenvectors of band matrix
!              On output: Transformed eigenvectors
!              Distribution is like in Scalapack.
!
!  ldq         Leading dimension of q
!
!  mpi_comm_rows
!  mpi_comm_cols
!              MPI-Communicators for rows/columns
!
!-------------------------------------------------------------------------------
1126
1127
1128
#ifdef HAVE_DETAILED_TIMINGS
 use timings
#endif
1129
1130
   implicit none

1131
1132
   integer              :: na, nqc, lda, ldq, nblk, nbw, mpi_comm_rows, mpi_comm_cols
   real*8               :: a(lda,*), q(ldq,*), tmat(nbw, nbw, *)
1133

1134
1135
1136
1137
1138
   integer              :: my_prow, my_pcol, np_rows, np_cols, mpierr
   integer              :: max_blocks_row, max_blocks_col, max_local_rows, &
                           max_local_cols
   integer              :: l_cols, l_rows, l_colh, n_cols
   integer              :: istep, lc, ncol, nrow, nb, ns
1139

1140
   real*8, allocatable  :: tmp1(:), tmp2(:), hvb(:), hvm(:,:)
1141

1142
   integer              :: i
1143
1144

   real*8, allocatable  :: tmat_complete(:,:), t_tmp(:,:), t_tmp2(:,:)
1145
1146
1147
   integer              :: cwy_blocking, t_blocking, t_cols, t_rows
   logical, intent(in)  :: useQR

1148
1149
1150
#ifdef HAVE_DETAILED_TIMINGS
   call timer%start("trans_ev_band_to_full_real")
#endif
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162

   call mpi_comm_rank(mpi_comm_rows,my_prow,mpierr)
   call mpi_comm_size(mpi_comm_rows,np_rows