elpa2.F90 208 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
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
!    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
!
!
!    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

65
  use elpa_utilities
66
  USE ELPA1
67
  use elpa2_utilities
68
69
  use elpa_pdgeqrf

70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
  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
88

89
90
91
92
93
94
  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)
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
!-------------------------------------------------------------------------------

  ! 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
113

114
function solve_evp_real_2stage(na, nev, a, lda, ev, q, ldq, nblk,        &
115
                               matrixCols,                               &
116
117
118
                                 mpi_comm_rows, mpi_comm_cols,           &
                                 mpi_comm_all, THIS_REAL_ELPA_KERNEL_API,&
                                 useQR) result(success)
119
120
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

!-------------------------------------------------------------------------------
!  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
!
!-------------------------------------------------------------------------------
154
155
156
#ifdef HAVE_DETAILED_TIMINGS
 use timings
#endif
157
   implicit none
158
159
   logical, intent(in), optional :: useQR
   logical                       :: useQRActual, useQREnvironment
Andreas Marek's avatar
Andreas Marek committed
160
   integer, intent(in), optional :: THIS_REAL_ELPA_KERNEL_API
161
   integer                       :: THIS_REAL_ELPA_KERNEL
162

163
   integer, intent(in)           :: na, nev, lda, ldq, matrixCols, mpi_comm_rows, &
164
                                    mpi_comm_cols, mpi_comm_all
165
   integer, intent(in)           :: nblk
166
   real*8, intent(inout)         :: a(lda,matrixCols), ev(na), q(ldq,matrixCols)
167

168
169
170
171
172
173
   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
174
175
   logical, save                 :: firstCall = .true.
   logical                       :: wantDebug
176

177
178
179
#ifdef HAVE_DETAILED_TIMINGS
   call timer%start("solve_evp_real_2stage")
#endif
180
181
182
183
184
185
186
   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)
187

188
189
190
191
192
193
194
195

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

196
197
   success = .true.

198
199
200
201
202
203
204
205
206
207
208
209
210
   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

211
   if (useQRActual) then
212
213
214
215
     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
216
     print *, "Do not use QR-decomposition for this matrix and blocksize."
Andreas Marek's avatar
Andreas Marek committed
217
218
     success = .false.
     return
219
     endif
220
221
   endif

222

223
224
225
   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
226
   else
227

228
229
230
     ! 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
231
232
233
234
   endif

   ! check whether choosen kernel is allowed
   if (check_allowed_real_kernels(THIS_REAL_ELPA_KERNEL)) then
235

236
237
238
239
240
241
242
243
244
245
246
     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
247

248
249
250
251
       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
252
253

   endif
254
255
256
257
258
259
260
261
262
263
264
265
266

   ! 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
267
   call bandred_real(na, a, lda, nblk, nbw, mpi_comm_rows, mpi_comm_cols, &
268
                     tmat, wantDebug, success, useQRActual)
269
   if (.not.(success)) return
270
   ttt1 = MPI_Wtime()
271
   if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
272
      write(error_unit,*) 'Time bandred_real               :',ttt1-ttt0
273
274
275
276
277
278

   ! Reduction band -> tridiagonal

   allocate(e(na))

   ttt0 = MPI_Wtime()
Andreas Marek's avatar
Andreas Marek committed
279
280
   call tridiag_band_real(na, nbw, nblk, a, lda, ev, e, mpi_comm_rows, &
                          mpi_comm_cols, mpi_comm_all)
281
   ttt1 = MPI_Wtime()
282
   if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
283
      write(error_unit,*) 'Time tridiag_band_real          :',ttt1-ttt0
284
285
286
287
288
289
290
291
292
293

   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()
294
   call solve_tridi(na, nev, ev, e, q, ldq, nblk, matrixCols, mpi_comm_rows,  &
295
                    mpi_comm_cols, wantDebug, success)
296
297
   if (.not.(success)) return

298
   ttt1 = MPI_Wtime()
299
300
   if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
     write(error_unit,*) 'Time solve_tridi                :',ttt1-ttt0
301
302
303
304
305
306
307
308
   time_evp_solve = ttt1-ttt0
   ttts = ttt1

   deallocate(e)

   ! Backtransform stage 1

   ttt0 = MPI_Wtime()
309
   call trans_ev_tridi_to_band_real(na, nev, nblk, nbw, q, ldq, mpi_comm_rows, &
310
                                    mpi_comm_cols, wantDebug, success, THIS_REAL_ELPA_KERNEL)
311
   if (.not.(success)) return
312
   ttt1 = MPI_Wtime()
313
   if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
314
      write(error_unit,*) 'Time trans_ev_tridi_to_band_real:',ttt1-ttt0
315
316
317
318
319
320
321

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

   ! Backtransform stage 2

   ttt0 = MPI_Wtime()
322
323
   call trans_ev_band_to_full_real(na, nev, nblk, nbw, a, lda, tmat, q, ldq, mpi_comm_rows, &
                                   mpi_comm_cols, useQRActual)
324
   ttt1 = MPI_Wtime()
325
   if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
326
      write(error_unit,*) 'Time trans_ev_band_to_full_real :',ttt1-ttt0
327
328
329
   time_evp_back = ttt1-ttts

   deallocate(tmat)
330
331
332
#ifdef HAVE_DETAILED_TIMINGS
   call timer%stop("solve_evp_real_2stage")
#endif
333
334
1  format(a,f10.3)

335
end function solve_evp_real_2stage
336
337
338
339
340

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

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

341
function solve_evp_complex_2stage(na, nev, a, lda, ev, q, ldq, nblk, &
342
                                  matrixCols, mpi_comm_rows, mpi_comm_cols,      &
343
                                    mpi_comm_all, THIS_COMPLEX_ELPA_KERNEL_API) result(success)
344
345
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

!-------------------------------------------------------------------------------
!  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
!
!-------------------------------------------------------------------------------
379
380
381
#ifdef HAVE_DETAILED_TIMINGS
 use timings
#endif
382
   implicit none
Andreas Marek's avatar
Andreas Marek committed
383
384
   integer, intent(in), optional :: THIS_COMPLEX_ELPA_KERNEL_API
   integer                       :: THIS_COMPLEX_ELPA_KERNEL
385
386
   integer, intent(in)           :: na, nev, lda, ldq, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols, mpi_comm_all
   complex*16, intent(inout)     :: a(lda,matrixCols), q(ldq,matrixCols)
387
388
389
390
391
392
393
394
   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
395

396
397
398
   logical                       :: success, wantDebug
   logical, save                 :: firstCall = .true.

399
400
401
#ifdef HAVE_DETAILED_TIMINGS
   call timer%start("solve_evp_complex_2stage")
#endif
Andreas Marek's avatar
Andreas Marek committed
402
403
   call mpi_comm_rank(mpi_comm_all,my_pe,mpierr)
   call mpi_comm_size(mpi_comm_all,n_pes,mpierr)
404
405
406
407
408

   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)
409

410
411
412
413
414
415
416
417
   wantDebug = .false.
   if (firstCall) then
     ! are debug messages desired?
     wantDebug = debug_messages_via_environment_variable()
     firstCall = .false.
   endif


418
419
   success = .true.

420
421
422
   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
423
   else
424
425
426
     ! 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
427
   endif
428

Andreas Marek's avatar
Andreas Marek committed
429
430
   ! check whether choosen kernel is allowed
   if (check_allowed_complex_kernels(THIS_COMPLEX_ELPA_KERNEL)) then
431

432
433
434
435
436
437
438
439
440
441
442
     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
443

444
445
446
447
       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
448
   endif
449
450
451
452
453
454
455
456
457
458
459
460
   ! 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
461
   call bandred_complex(na, a, lda, nblk, nbw, mpi_comm_rows, mpi_comm_cols, &
462
                        tmat, wantDebug, success)
463
464
465
466
467
468
   if (.not.(success)) then
#ifdef HAVE_DETAILED_TIMINGS
     call timer%stop()
#endif
     return
   endif
469
   ttt1 = MPI_Wtime()
470
   if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
471
      write(error_unit,*) 'Time bandred_complex               :',ttt1-ttt0
472
473
474
475
476
477
478
479

   ! 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()
480
   if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
481
      write(error_unit,*) 'Time tridiag_band_complex          :',ttt1-ttt0
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497

   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()
498
   call solve_tridi(na, nev, ev, e, q_real, ubound(q_real,dim=1), nblk, matrixCols, &
499
                    mpi_comm_rows, mpi_comm_cols, wantDebug, success)
500
501
   if (.not.(success)) return

502
   ttt1 = MPI_Wtime()
503
   if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times)  &
504
      write(error_unit,*) 'Time solve_tridi                   :',ttt1-ttt0
505
506
507
508
509
510
511
512
513
514
   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
515
   call trans_ev_tridi_to_band_complex(na, nev, nblk, nbw, q, ldq,  &
516
                                       mpi_comm_rows, mpi_comm_cols,&
517
                                       wantDebug, success,THIS_COMPLEX_ELPA_KERNEL)
518
   if (.not.(success)) return
519
   ttt1 = MPI_Wtime()
520
   if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
521
      write(error_unit,*) 'Time trans_ev_tridi_to_band_complex:',ttt1-ttt0
522
523
524
525
526
527
528
529
530

   ! 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()
531
   if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
532
      write(error_unit,*) 'Time trans_ev_band_to_full_complex :',ttt1-ttt0
533
534
535
   time_evp_back = ttt1-ttts

   deallocate(tmat)
536
537
538
#ifdef HAVE_DETAILED_TIMINGS
   call timer%stop("solve_evp_complex_2stage")
#endif
539
540
541

1  format(a,f10.3)

542
end function solve_evp_complex_2stage
543
544
545

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

546
subroutine bandred_real(na, a, lda, nblk, nbw, mpi_comm_rows, mpi_comm_cols, &
547
                        tmat, wantDebug, success, useQR)
548
549
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

!-------------------------------------------------------------------------------
!  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
!
!-------------------------------------------------------------------------------
576
577
578
#ifdef HAVE_DETAILED_TIMINGS
 use timings
#endif
579
   implicit none
580
   
581
582
   integer             :: na, lda, nblk, nbw, mpi_comm_rows, mpi_comm_cols
   real*8              :: a(lda,*), tmat(nbw,nbw,*)
583

584
585
586
587
588
   integer             :: my_prow, my_pcol, np_rows, np_cols, mpierr
   integer             :: l_cols, l_rows
   integer             :: i, j, lcs, lce, lre, lc, lr, cur_pcol, n_cols, nrow
   integer             :: istep, ncol, lch, lcx, nlc
   integer             :: tile_size, l_rows_tile, l_cols_tile
589

590
   real*8              :: vnorm2, xf, aux1(nbw), aux2(nbw), vrl, tau, vav(nbw,nbw)
591

592
   real*8, allocatable :: tmp(:,:), vr(:), vmr(:,:), umc(:,:)
593

594
595
596
597
598
   ! needed for blocked QR decomposition
   integer             :: PQRPARAM(11), work_size
   real*8              :: dwork_size(1)
   real*8, allocatable :: work_blocked(:), tauvector(:), blockheuristic(:)

599
   logical, intent(in) :: wantDebug
600
601
   logical, intent(out):: success

602
603
   logical, intent(in) :: useQR

604
605
606
#ifdef HAVE_DETAILED_TIMINGS
   call timer%start("bandred_real")
#endif
607
608
609
610
   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)
611
   success = .true.
612
613


614
   ! Semibandwith nbw must be a multiple of blocksize nblk
615
616
   if (mod(nbw,nblk)/=0) then
     if (my_prow==0 .and. my_pcol==0) then
617
618
619
620
       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
621
       success = .false.
Lorenz Huedepohl's avatar
Lorenz Huedepohl committed
622
       return
623
     endif
624
625
626
627
628
629
630
631
632
633
   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

634
635
636
637
638
639
640
   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))
641

642
       call qr_pdgeqrf_2dcomm(a, lda, vmr, max(l_rows,1), tauvector(1), tmat(1,1,1), nbw, dwork_size(1), -1, na, &
643
                             nbw, nblk, nblk, na, na, 1, 0, PQRPARAM, mpi_comm_rows, mpi_comm_cols, blockheuristic)
644
645
       work_size = dwork_size(1)
       allocate(work_blocked(work_size))
646

647
648
649
       work_blocked = 0.0d0
       deallocate(vmr)
     endif
650
651
   endif

652
653
   do istep = (na-1)/nbw, 1, -1

654
     n_cols = MIN(na,(istep+1)*nbw) - istep*nbw ! Number of columns in current step
655

656
657
658
     ! 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)
659

660
     ! Allocate vmr and umc to their exact sizes so that they can be used in bcasts and reduces
661

662
663
     allocate(vmr(max(l_rows,1),2*n_cols))
     allocate(umc(max(l_cols,1),2*n_cols))
664

665
     allocate(vr(l_rows+1))
666

667
668
669
     vmr(1:l_rows,1:n_cols) = 0.
     vr(:) = 0
     tmat(:,:,istep) = 0
670

671
     ! Reduce current block to lower triangular form
672
673
674
675
676
677
678
679
680
681

     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
682
     else
683

684
       do lc = n_cols, 1, -1
685

686
687
         ncol = istep*nbw + lc ! absolute column number of householder vector
         nrow = ncol - nbw ! Absolute number of pivot row
688

689
690
         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
691

692
         tau = 0
693

694
         if (nrow == 1) exit ! Nothing to do
695

696
         cur_pcol = pcol(ncol, nblk, np_cols) ! Processor column owning current block
697

698
         if (my_pcol==cur_pcol) then
699

700
701
           ! Get vector to be transformed; distribute last element and norm of
           ! remaining elements to all procs in current column
702

703
           vr(1:lr) = a(1:lr,lch) ! vector to be transformed
704

705
           if (my_prow==prow(nrow, nblk, np_rows)) then
706
707
708
709
710
711
             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
712

713
           call mpi_allreduce(aux1,aux2,2,MPI_REAL8,MPI_SUM,mpi_comm_rows,mpierr)
714

715
716
           vnorm2 = aux2(1)
           vrl    = aux2(2)
717

718
           ! Householder transformation
719

720
           call hh_transform_real(vrl, vnorm2, xf, tau)
721

722
           ! Scale vr and store Householder vector for back transformation
723

724
           vr(1:lr) = vr(1:lr) * xf
725
           if (my_prow==prow(nrow, nblk, np_rows)) then
726
727
728
729
730
             a(1:lr-1,lch) = vr(1:lr-1)
             a(lr,lch) = vrl
             vr(lr) = 1.
           else
             a(1:lr,lch) = vr(1:lr)
731
           endif
732

733
         endif
734

735
         ! Broadcast Householder vector and tau along columns
736

737
738
739
740
741
         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
742

743
744
         ! Transform remaining columns in current block with Householder vector
         ! Local dot product
745

746
         aux1 = 0
747

748
749
750
751
752
753
754
755
         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
756

757
758
         ! Get global dot products
         if (nlc>0) call mpi_allreduce(aux1,aux2,nlc,MPI_REAL8,MPI_SUM,mpi_comm_rows,mpierr)
759

760
         ! Transform
761

762
763
764
765
766
767
768
769
770
771
         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

       enddo
772

773
774
       ! Calculate scalar products of stored Householder vectors.
       ! This can be done in different ways, we use dsyrk
775

776
777
       vav = 0
       if (l_rows>0) &
778
779
           call dsyrk('U','T',n_cols,l_rows,1.d0,vmr,ubound(vmr,dim=1),0.d0,vav,ubound(vav,dim=1))
       call symm_matrix_allreduce(n_cols,vav,ubound(vav,dim=1),mpi_comm_rows)
780

781
       ! Calculate triangular matrix T for block Householder Transformation
782

783
784
785
       do lc=n_cols,1,-1
         tau = tmat(lc,lc,istep)
         if (lc<n_cols) then
786
           call dtrmv('U','T','N',n_cols-lc,tmat(lc+1,lc+1,istep),ubound(tmat,dim=1),vav(lc+1,lc),1)
787
788
789
           tmat(lc,lc+1:n_cols,istep) = -tau * vav(lc+1:n_cols,lc)
         endif
       enddo
790
     endif
791

792
    ! Transpose vmr -> vmc (stored in umc, second half)
793

794
795
    call elpa_transpose_vectors_real  (vmr, ubound(vmr,dim=1), mpi_comm_rows, &
                                    umc(1,n_cols+1), ubound(umc,dim=1), mpi_comm_cols, &
796
797
                                    1, istep*nbw, n_cols, nblk)

798
799
800
801
    ! 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
802

803
804
805
806
    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
807

808
809
810
        lcs = i*l_cols_tile+1
        lce = min(l_cols,(i+1)*l_cols_tile)
        if (lce<lcs) cycle
811

812
        lre = min(l_rows,(i+1)*l_rows_tile)
813
814
        call DGEMM('T','N',lce-lcs+1,n_cols,lre,1.d0,a(1,lcs),ubound(a,dim=1), &
                     vmr,ubound(vmr,dim=1),1.d0,umc(lcs,1),ubound(umc,dim=1))
815

816
817
818
        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, &
819
                     umc(lcs,n_cols+1),ubound(umc,dim=1),1.d0,vmr(1,n_cols+1),ubound(vmr,dim=1))
820
821
      enddo
    endif
822

823
824
825
826
    ! 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
827

828
    if (tile_size < istep*nbw) then
829
830
       call elpa_reduce_add_vectors_real  (vmr(1,n_cols+1),ubound(vmr,dim=1),mpi_comm_rows, &
                                      umc, ubound(umc,dim=1), mpi_comm_cols, &
831
832
                                      istep*nbw, n_cols, nblk)
    endif
833

834
835
836
837
838
839
    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
840

841
    ! U = U * Tmat**T
842

843
    call dtrmm('Right','Upper','Trans','Nonunit',l_cols,n_cols,1.d0,tmat(1,1,istep),ubound(tmat,dim=1),umc,ubound(umc,dim=1))
844

845
    ! VAV = Tmat * V**T * A * V * Tmat**T = (U*Tmat**T)**T * V * Tmat**T
846

847
848
    call dgemm('T','N',n_cols,n_cols,l_cols,1.d0,umc,ubound(umc,dim=1),umc(1,n_cols+1),ubound(umc,dim=1),0.d0,vav,ubound(vav,dim=1))
    call dtrmm('Right','Upper','Trans','Nonunit',n_cols,n_cols,1.d0,tmat(1,1,istep),ubound(tmat,dim=1),vav,ubound(vav,dim=1))
849

850
    call symm_matrix_allreduce(n_cols,vav,ubound(vav,dim=1),mpi_comm_cols)
851

852
    ! U = U - 0.5 * V * VAV
853
    call dgemm('N','N',l_cols,n_cols,n_cols,-0.5d0,umc(1,n_cols+1),ubound(umc,dim=1),vav,ubound(vav,dim=1),1.d0,umc,ubound(umc,dim=1))
854

855
    ! Transpose umc -> umr (stored in vmr, second half)
856

857
858
    call elpa_transpose_vectors_real  (umc, ubound(umc,dim=1), mpi_comm_cols, &
                                   vmr(1,n_cols+1), ubound(vmr,dim=1), mpi_comm_rows, &
859
                                   1, istep*nbw, n_cols, nblk)
860

861
    ! A = A - V*U**T - U*V**T
862

863
864
865
866
867
868
    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, &
869
                  vmr,ubound(vmr,dim=1),umc(lcs,1),ubound(umc,dim=1), &
870
871
                  1.d0,a(1,lcs),lda)
    enddo
872

873
    deallocate(vmr, umc, vr)
874

875
  enddo
876

877
878
879
880
881
  if (useQR) then
    if (which_qr_decomposition == 1) then
      deallocate(work_blocked)
      deallocate(tauvector)
    endif
882
  endif
883

Andreas Marek's avatar
Andreas Marek committed
884
885
886
#ifdef HAVE_DETAILED_TIMINGS
  call timer%stop("bandred_real")
#endif
887
888
889
890
891
892
893
894
895
896
897
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
898
899
900
#ifdef HAVE_DETAILED_TIMINGS
 use timings
#endif
901
   implicit none
Andreas Marek's avatar
Andreas Marek committed
902
903
904
905
906
   integer  :: n, lda, comm
   real*8   :: a(lda,*)

   integer  :: i, nc, mpierr
   real*8   :: h1(n*n), h2(n*n)
907

Andreas Marek's avatar
Andreas Marek committed
908
909
910
#ifdef HAVE_DETAILED_TIMINGS
  call timer%start("symm_matrix_allreduce")
#endif
911
912
913

   nc = 0
   do i=1,n
914
915
     h1(nc+1:nc+i) = a(1:i,i)
     nc = nc+i
916
917
918
919
920
921
   enddo

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

   nc = 0
   do i=1,n
922
923
924
     a(1:i,i) = h2(nc+1:nc+i)
     a(i,1:i-1) = a(1:i-1,i)
     nc = nc+i
925
926
   enddo

Andreas Marek's avatar
Andreas Marek committed
927
928
929
930
#ifdef HAVE_DETAILED_TIMINGS
  call timer%stop("symm_matrix_allreduce")
#endif

931
932
933
934
end subroutine symm_matrix_allreduce

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

935
936
subroutine trans_ev_band_to_full_real(na, nqc, nblk, nbw, a, lda, tmat, q, ldq, mpi_comm_rows, &
                                      mpi_comm_cols, useQR)
937

Andreas Marek's avatar
Andreas Marek committed
938

939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
!-------------------------------------------------------------------------------
!  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
!
!-------------------------------------------------------------------------------
971
972
973
#ifdef HAVE_DETAILED_TIMINGS
 use timings
#endif
974
975
   implicit none

976
977
   integer              :: na, nqc, lda, ldq, nblk, nbw, mpi_comm_rows, mpi_comm_cols
   real*8               :: a(lda,*), q(ldq,*), tmat(nbw, nbw, *)
978

979
980
981
982
983
   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
984

985
   real*8, allocatable  :: tmp1(:), tmp2(:), hvb(:), hvm(:,:)
986

987
   integer              :: i
988
989

   real*8, allocatable  :: tmat_complete(:,:), t_tmp(:,:), t_tmp2(:,:)
990
991
992
   integer              :: cwy_blocking, t_blocking, t_cols, t_rows
   logical, intent(in)  :: useQR

993
994
995
#ifdef HAVE_DETAILED_TIMINGS
   call timer%start("trans_ev_band_to_full_real")
#endif
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007

   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)

   max_blocks_row = ((na -1)/nblk)/np_rows + 1  ! Rows of A
   max_blocks_col = ((nqc-1)/nblk)/np_cols + 1  ! Columns of q!

   max_local_rows = max_blocks_row*nblk
   max_local_cols = max_blocks_col*nblk

1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
   if (useQR) then
     t_blocking = 2 ! number of matrices T (tmat) which are aggregated into a new (larger) T matrix (tmat_complete) and applied at once
     cwy_blocking = t_blocking * nbw

     allocate(tmp1(max_local_cols*cwy_blocking))
     allocate(tmp2(max_local_cols*cwy_blocking))
     allocate(hvb(max_local_rows*cwy_blocking))
     allocate(hvm(max_local_rows,cwy_blocking))
     allocate(tmat_complete(cwy_blocking,cwy_blocking))
     allocate(t_tmp(cwy_blocking,nbw))
     allocate(t_tmp2(cwy_blocking,nbw))
   else
     allocate(tmp1(max_local_cols*nbw))
     allocate(tmp2(max_local_cols*nbw))
     allocate(hvb(max_local_rows*nbw))
     allocate(hvm(max_local_rows,nbw))
   endif
1025
1026
1027
1028
1029
1030

   hvm = 0   ! Must be set to 0 !!!
   hvb = 0   ! Safety only

   l_cols = local_index(nqc, my_pcol, np_cols, nblk, -1) ! Local columns of q

1031
   if (useQR) then
1032

1033
1034
     do istep=1,((na-1)/nbw-1)/t_blocking + 1
       n_cols = MIN(na,istep*cwy_blocking+nbw) - (istep-1)*cwy_blocking - nbw ! Number of columns in current step
1035

1036
       ! Broadcast all Householder vectors for current step compressed in hvb
1037

1038
1039
       nb = 0
       ns = 0
1040

1041
1042
1043
       do lc = 1, n_cols
         ncol = (istep-1)*cwy_blocking + nbw + lc ! absolute column number of householder vector
         nrow = ncol - nbw ! absolute number of pivot row
1044

1045
1046
         l_rows = local_index(nrow-1, my_prow, np_rows, nblk, -1) ! row length for bcast
         l_colh = local_index(ncol  , my_pcol, np_cols, nblk, -1) ! HV local column number
1047

1048
         if (my_pcol==pcol(ncol, nblk, np_cols)) hvb(nb+1:nb+l_rows) = a(1:l_rows,l_colh)
1049

1050
         nb = nb+l_rows
1051

1052
         if (lc==n_cols .or. mod(ncol,nblk)==0) then
1053
           call MPI_Bcast(hvb(ns+1),nb-ns,MPI_REAL8,pcol(ncol, nblk, np_cols),mpi_comm_cols,mpierr)
1054
1055
1056
           ns = nb
         endif
       enddo
1057

1058
       ! Expand compressed Householder vectors into matrix hvm
1059

1060
1061
1062
1063
       nb = 0
       do lc = 1, n_cols
         nrow = (istep-1)*cwy_blocking + lc ! absolute number of pivot row
         l_rows = local_index(nrow-1, my_prow, np_rows, nblk, -1) ! row length for bcast
1064

1065
         hvm(1:l_rows,lc) = hvb(nb+1:nb+l_rows)
1066
         if (my_prow==prow(nrow, nblk, np_rows)) hvm(l_rows+1,lc) = 1.
1067

1068
1069
         nb = nb+l_rows
       enddo
1070

1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
       l_rows = local_index(MIN(na,(istep+1)*cwy_blocking), my_prow, np_rows, nblk, -1)

       ! compute tmat2 out of tmat(:,:,)
       tmat_complete = 0
       do i = 1, t_blocking
         t_cols = MIN(nbw, n_cols - (i-1)*nbw)
         if (t_cols <= 0) exit
         t_rows = (i - 1) * nbw
         tmat_complete(t_rows+1:t_rows+t_cols,t_rows+1:t_rows+t_cols) = tmat(1:t_cols,1:t_cols,(istep-1)*t_blocking + i)
         if (i > 1) then
           call dgemm('T', 'N', t_rows, t_cols, l_rows, 1.d0, hvm(1,1), max_local_rows, hvm(1,(i-1)*nbw+1), &
1082
                     max_local_rows, 0.d0, t_tmp, cwy_blocking)
1083
1084
1085
1086
1087
1088
           call mpi_allreduce(t_tmp,t_tmp2,cwy_blocking*nbw,MPI_REAL8,MPI_SUM,mpi_comm_rows,mpierr)
           call dtrmm('L','U','N','N',t_rows,t_cols,1.0d0,tmat_complete,cwy_blocking,t_tmp2,cwy_blocking)
           call dtrmm('R','U','N','N',t_rows,t_cols,-1.0d0,tmat_complete(t_rows+1,t_rows+1),cwy_blocking,t_tmp2,cwy_blocking)
           tmat_complete(1:t_rows,t_rows+1:t_rows+t_cols) = t_tmp2(1:t_rows,1:t_cols)
         endif
       enddo
1089

1090
       ! Q = Q - V * T**T * V**T * Q
1091

1092
       if (l_rows>0) then
Andreas Marek's avatar