elpa2_trans_ev_band_to_full_template.X90 36.5 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
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
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
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
#if 0
!    This file is part of ELPA.
!
!    The ELPA library was originally created by the ELPA consortium,
!    consisting of the following organizations:
!
!    - Max Planck Computing and Data Facility (MPCDF), formerly known as
!      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 Naturwissenschaften,
!      Leipzig, Abt. Komplexe Strukutren in Biologie und Kognition,
!      and
!    - IBM Deutschland GmbH
!
!    This particular source code file contains additions, changes and
!    enhancements authored by Intel Corporation which is not part of
!    the ELPA consortium.
!
!    More information can be found here:
!    http://elpa.mpcdf.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.
!
! 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".
#endif


    subroutine trans_ev_band_to_full_&
    &MATH_DATATYPE&
    &_&
    &PRECISION &
    (na, nqc, nblk, nbw, a, &
#if REALCASE == 1
    a_dev,                  &
#endif
    lda, tmat,              &
#if REALCASE == 1
    tmat_dev,               &
#endif
    q,                      &
#if REALCASE == 1
    q_dev,                  &
#endif
     ldq, matrixCols, numBlocks, mpi_comm_rows, mpi_comm_cols, useGPU &
#if REALCASE == 1
     ,useQr)
#endif
#if COMPLEXCASE == 1
     )
#endif

    !-------------------------------------------------------------------------------
    !  trans_ev_band_to_full_real/complex:
    !  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,matrixCols)    Matrix containing the Householder vectors (i.e. matrix a after bandred_real/complex)
    !              Distribution is like in Scalapack.
    !
    !  lda         Leading dimension of a
    !  matrixCols  local columns of matrix a and q
    !
    !  tmat(nbw,nbw,numBlocks) Factors returned by bandred_real/complex
    !
    !  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
    !
    !-------------------------------------------------------------------------------
#ifdef HAVE_DETAILED_TIMINGS
      use timings
#else
      use timings_dummy
#endif
      use precision
      use cuda_functions
      use iso_c_binding

      implicit none

      logical, intent(in)                    :: useGPU
#if REALCASE == 1
     logical, intent(in)                     :: useQR
#endif
      integer(kind=ik)                       :: na, nqc, lda, ldq, nblk, nbw, matrixCols, numBlocks, mpi_comm_rows, mpi_comm_cols
#if REALCASE == 1
#ifdef USE_ASSUMED_SIZE
      real(kind=REAL_DATATYPE)               :: a(lda,*), q(ldq,*), tmat(nbw,nbw,*)
      !real(kind=rk8)               :: a(lda,*), q(ldq,*), tmat(nbw,nbw,*)
#else
      real(kind=REAL_DATATYPE)               :: a(lda,matrixCols), q(ldq,matrixCols), tmat(nbw, nbw, numBlocks)
      !real(kind=rk8)               :: a(lda,matrixCols), q(ldq,matrixCols), tmat(nbw, nbw, numBlocks)
#endif
#endif
#if COMPLEXCASE == 1
#ifdef USE_ASSUMED_SIZE
      complex(kind=COMPLEX_DATATYPE)         :: a(lda,*), q(ldq,*), tmat(nbw,nbw,*)
      !complex(kind=ck8)         :: a(lda,*), q(ldq,*), tmat(nbw,nbw,*)
#else
      complex(kind=COMPLEX_DATATYPE)         :: a(lda,matrixCols), q(ldq,matrixCols), tmat(nbw, nbw, numBlocks)
      !complex(kind=ck8)         :: a(lda,matrixCols), q(ldq,matrixCols), tmat(nbw, nbw, numBlocks)
#endif
#endif

#if REALCASE == 1
#ifdef DOUBLE_PRECISION_COMPLEX
       real(kind=REAL_DATATYPE), parameter   :: ZERO = 0.0_rk8, ONE = 1.0_rk8
#else
       real(kind=REAL_DATATYPE), parameter   :: ZERO = 0.0_rk4, ONE = 1.0_rk4
#endif

#endif
#if COMPLEXCASE == 1
#ifdef DOUBLE_PRECISION_COMPLEX
       complex(kind=COMPLEX_DATATYPE), parameter   :: ZERO = (0.0_rk8,0.0_rk8), ONE = (1.0_rk8,0.0_rk8)
#else
       complex(kind=COMPLEX_DATATYPE), parameter   :: ZERO = (0.0_rk4,0.0_rk4), ONE = (1.0_rk4,0.0_rk4)
#endif
#endif

#if REALCASE == 1
      integer(kind=C_intptr_T)               :: a_dev ! passed from bandred_real at the moment not used since copied in bandred_real
#endif

      integer(kind=ik)                       :: my_prow, my_pcol, np_rows, np_cols, mpierr
      integer(kind=ik)                       :: max_blocks_row, max_blocks_col, max_local_rows, &
                                                max_local_cols
      integer(kind=ik)                       :: l_cols, l_rows, l_colh, n_cols
      integer(kind=ik)                       :: istep, lc, ncol, nrow, nb, ns

#if REALCASE ==1
      real(kind=REAL_DATATYPE), allocatable  :: tmp1(:), tmp2(:), hvb(:), hvm(:,:)
#endif
#if COMPLEXCASE == 1
      complex(kind=COMPLEX_DATATYPE), allocatable :: tmp1(:), tmp2(:), hvb(:), hvm(:,:)
#endif
      ! hvm_dev is fist used and set in this routine
      ! q is changed in trans_ev_tridi on the host, copied to device and passed here. this can be adapted
      ! tmp_dev is first used in this routine
      ! tmat_dev is passed along from bandred_real
      integer(kind=C_intptr_T)               :: hvm_dev, q_dev, tmp_dev, tmat_dev

      integer(kind=ik)                       :: i

Andreas Marek's avatar
Andreas Marek committed
185
#ifdef BAND_TO_FULL_BLOCKING
186
187
#if REALCASE == 1
      real(kind=REAL_DATATYPE), allocatable  :: tmat_complete(:,:), t_tmp(:,:), t_tmp2(:,:)
Andreas Marek's avatar
Andreas Marek committed
188
189
190
191
#endif
#if COMPLEXCASE == 1
      complex(kind=COMPLEX_DATATYPE), allocatable  :: tmat_complete(:,:), t_tmp(:,:), t_tmp2(:,:)
#endif
192
193
      integer(kind=ik)                       :: cwy_blocking, t_blocking, t_cols, t_rows
#endif
Andreas Marek's avatar
Andreas Marek committed
194

195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
      integer(kind=ik)                       :: istat
      character(200)                         :: errorMessage
      logical                                :: successCUDA

      call timer%start("trans_ev_band_to_full_&
      &MATH_DATATYPE&
      &_" // &
      &PRECISION_SUFFIX &
      )
      call timer%start("mpi_communication")

      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)

      call timer%stop("mpi_communication")

      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

      if (useGPU) then

#if REALCASE == 1
        ! here the GPU and CPU version diverged: the CPU version now always uses the useQR path which
        ! is not implemented in the GPU version
#endif

Andreas Marek's avatar
Andreas Marek committed
226
227
228
        ! the GPU version does not (yet) support blocking
        ! but the handling is the same for real/complex case

229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
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
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
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
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
        allocate(tmp1(max_local_cols*nbw), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
          print *,"trans_ev_band_to_full_&
	  &MATH_DATATYPE&
	  &: error when allocating tmp1 "//errorMessage
          stop
        endif

        allocate(tmp2(max_local_cols*nbw), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
          print *,"trans_ev_band_to_full_&
	  &MATH_DATATYPE&
	  &: error when allocating tmp2 "//errorMessage
          stop
        endif

        allocate(hvb(max_local_rows*nbw), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
          print *,"trans_ev_band_to_full_&
	  &MATH_DATATYPE&
	  &: error when allocating hvb "//errorMessage
          stop
        endif

        allocate(hvm(max_local_rows,nbw), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
          print *,"trans_ev_band_to_full_&
	  &MATH_DATATYPE&
	  &: error when allocating hvm "//errorMessage
          stop
        endif

        successCUDA = cuda_malloc(hvm_dev, (max_local_rows)*nbw* &
#if REALCASE == 1
	size_of_PRECISION_real)
#endif
#if COMPLEXCASE == 1
        size_of_PRECISION_complex)
#endif
        if (.not.(successCUDA)) then
          print *,"trans_ev_band_to_full_&
	  &MATH_DATATYPE&
	  &: error in cudaMalloc"
          stop
        endif

        successCUDA = cuda_malloc(tmp_dev, (max_local_cols)*nbw* &
#if REALCASE == 1
	size_of_PRECISION_real)
#endif
#if COMPLEXCASE == 1
        size_of_PRECISION_complex)
#endif
        if (.not.(successCUDA)) then
          print *,"trans_ev_band_to_full_&
	  &MATH_DATATYPE&
	  &: error in cudaMalloc"
          stop
        endif

!#ifdef WITH_MPI
! it should be possible to keep tmat dev on the device and not copy it around
! already existent on GPU
        successCUDA = cuda_malloc(tmat_dev, nbw*nbw* &
#if REALCASE == 1
	size_of_PRECISION_real)
#endif
#if COMPLEXCASE == 1
        size_of_PRECISION_complex)
#endif

        if (.not.(successCUDA)) then
          print *,"trans_ev_band_to_full_&
	  &MATH_DATATYPE&
	  &: error in cudaMalloc"
          stop
        endif
!#endif

#if REALCASE == 1
! q_dev already living on device
!        successCUDA = cuda_malloc(q_dev, ldq*matrixCols*size_of_PRECISION_real)
!        if (.not.(successCUDA)) then
!          print *,"trans_ev_band_to_full_real: error in cudaMalloc"
!          stop
!        endif
  !      q_temp(:,:) = 0.0
  !      q_temp(1:ldq,1:na_cols) = q(1:ldq,1:na_cols)

!        ! copy q_dev to device, maybe this can be avoided if q_dev can be kept on device in trans_ev_tridi_to_band
!        successCUDA = cuda_memcpy(q_dev, loc(q), (ldq)*(matrixCols)*size_of_PRECISION_real, cudaMemcpyHostToDevice)
!        if (.not.(successCUDA)) then
!          print *,"trans_ev_band_to_full_real: error in cudaMalloc"
!          stop
!        endif
#endif
#if COMPLEXCASE == 1
         successCUDA = cuda_malloc(q_dev, ldq*matrixCols*size_of_PRECISION_complex)
         if (.not.(successCUDA)) then
           print *,"trans_ev_band_to_full_complex: error in cudaMalloc"
           stop
         endif

         successCUDA = cuda_memcpy(q_dev, loc(q),ldq*matrixCols*size_of_PRECISION_complex, cudaMemcpyHostToDevice)
          if (.not.(successCUDA)) then
            print *,"trans_ev_band_to_full_complex: error in cudaMemcpy"
            stop
          endif
#endif

        ! if MPI is NOT used the following steps could be done on the GPU and memory transfers could be avoided
        successCUDA = cuda_memset(hvm_dev, 0, (max_local_rows)*(nbw)* &
#if REALCASE == 1
	                          size_of_PRECISION_real)
#endif
#if COMPLEXCASE == 1
                                  size_of_PRECISION_complex)
#endif
        if (.not.(successCUDA)) then
          print *,"trans_ev_band_to_full_&
	  &MATH_DATATYPE&
	  &: error in cudaMalloc"
          stop
        endif

#if REALCASE == 1
        hvm = CONST_0_0   ! Must be set to 0 !!!
        hvb = CONST_0_0   ! Safety only
#endif
#if COMPLEXCASE == 1
        hvm = CONST_COMPLEX_0_0   ! Must be set to 0 !!!
        hvb = CONST_COMPLEX_0_0   ! Safety only
#endif
        l_cols = local_index(nqc, my_pcol, np_cols, nblk, -1) ! Local columns of q

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

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

          ! Broadcast all Householder vectors for current step compressed in hvb

          nb = 0
          ns = 0

          do lc = 1, n_cols
            ncol = istep*nbw + lc ! absolute column number of householder vector
            nrow = ncol - nbw ! absolute number of pivot row

            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

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

            nb = nb+l_rows

            if (lc==n_cols .or. mod(ncol,nblk)==0) then
#ifdef WITH_MPI
              call timer%start("mpi_communication")
              call MPI_Bcast(hvb(ns+1), nb-ns,  &
#if REALCASE == 1
	                     MPI_REAL_PRECISION,&
#endif
#if COMPLEXCASE == 1
                             MPI_COMPLEX_PRECISION, &
#endif
			     pcol(ncol, nblk, np_cols), mpi_comm_cols, mpierr)

              call timer%stop("mpi_communication")

#endif /* WITH_MPI */
              ns = nb
            endif
          enddo

          ! Expand compressed Householder vectors into matrix hvm

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

            hvm(1:l_rows,lc) = hvb(nb+1:nb+l_rows)
#if REALCASE == 1
            if (my_prow==prow(nrow, nblk, np_rows)) hvm(l_rows+1,lc) = CONST_1_0
#endif
#if COMPLEXCASE == 1
            if (my_prow==prow(nrow, nblk, np_rows)) hvm(l_rows+1,lc) = 1.
#endif
            nb = nb+l_rows
          enddo

          successCUDA = cuda_memcpy(hvm_dev, loc(hvm), ((max_local_rows)*nbw* &
#if REALCASE == 1
	                            size_of_PRECISION_real), &
#endif
#if COMPLEXCASE == 1
                                    size_of_PRECISION_complex), &
#endif
				    cudaMemcpyHostToDevice)

          if (.not.(successCUDA)) then
            print *,"trans_ev_band_to_full_real: error in cudaMemcpy"
            stop

          endif

          l_rows = local_index(MIN(na,(istep+1)*nbw), my_prow, np_rows, nblk, -1)

          ! Q = Q - V * T**T * V**T * Q

          if (l_rows>0) then
	    call timer%start("cublas")
#if REALCASE == 1
            call cublas_PRECISION_GEMM('T', 'N',        &
#endif
#if COMPLEXCASE == 1
            call cublas_PRECISION_GEMM('C', 'N',        &
#endif
	                               n_cols, l_cols, l_rows, ONE, hvm_dev, max_local_rows, &
                                       q_dev, ldq , ZERO, tmp_dev, n_cols)
	    call timer%stop("cublas")

#ifdef WITH_MPI

            ! copy data from device to host for a later MPI_ALLREDUCE
#if REALCASE == 1
            ! copy to host maybe this can be avoided this is needed if MPI is used (allreduce)
            successCUDA = cuda_memcpy(loc(tmp1), tmp_dev, l_cols*n_cols*size_of_PRECISION_real, cudaMemcpyDeviceToHost)
            if (.not.(successCUDA)) then
              print *,"trans_ev_band_to_full_real: error in cudaMemcpy"
              stop
            endif
#endif

#else /* WITH_MPI */
! check whether in the complex case this ist also only necessary for MPI

#if COMPLEXCASE == 1
            successCUDA = cuda_memcpy(loc(tmp1), tmp_dev, n_cols*l_cols*size_of_PRECISION_complex, &
                                       cudaMemcpyDeviceToHost)

            if (.not.(successCUDA)) then
              print *,"trans_ev_band_to_full_complex: error in cudaMemcpy"
              stop
            endif
#endif

#endif /* WITH_MPI */

          else ! l_rows>0

#ifdef WITH_MPI
#if REALCASE == 1
            tmp1(1:l_cols*n_cols) = 0
#endif
#if COMPLEXCASE == 1
            tmp1(1:l_cols*n_cols) = CONST_COMPLEX_0_0
#endif

#else /* WITH_MPI */

            ! if MPI is not used (we do not need to transfer because of MPI_ALLREDUCE) we can set to zero on the device

            successCUDA = cuda_memset(tmp_dev, 0, l_cols*n_cols* &
#if REALCASE == 1
                                      size_of_PRECISION_real)
#endif
#if COMPLEXCASE == 1
                                      size_of_PRECISION_complex)
#endif
            if (.not.(successCUDA)) then
              print *,"trans_ev_band_to_full_&
	      &MATH_DATATYPE&
	      &: error in cudaMemset"
              stop
             endif

#endif /* WITH_MPI */

          endif ! l_rows>0

          !#ifdef WITH_GPU_VERSION
          !       istat = cuda_memcpy(loc(tmp1), tmp_dev, max_local_cols*nbw*size_of_real_datatype,cudaMemcpyDeviceToHost)
          !       if (istat .ne. 0) then
          !         print *,"error in cudaMemcpy"
          !         stop
          !       endif
          !#endif

#ifdef WITH_MPI
          call timer%start("mpi_communication")
          call mpi_allreduce(tmp1, tmp2, n_cols*l_cols, &
#if REALCASE == 1
                             MPI_REAL_PRECISION,        &
#endif
#if COMPLEXCASE == 1
                             MPI_COMPLEX_PRECISION,        &
#endif
                             MPI_SUM, mpi_comm_rows, mpierr)
          call timer%stop("mpi_communication")

#else /* WITH_MPI */
!          tmp2(1:n_cols*l_cols) = tmp1(1:n_cols*l_cols)
#endif /* WITH_MPI */

           !#ifdef WITH_GPU_VERSION
          !       istat = cuda_memcpy(tmp_dev, loc(tmp2), max_local_cols*nbw*size_of_real_datatype,cudaMemcpyHostToDevice)
          !       if (istat .ne. 0) then
          !         print *,"error in cudaMemcpy"
          !         stop
          !       endif
          !#endif

          if (l_rows>0) then
#ifdef WITH_MPI
            ! after the mpi_allreduce we have to copy back to the device
            ! copy back to device
            successCUDA = cuda_memcpy(tmp_dev, loc(tmp2), n_cols*l_cols* &
#if REALCASE == 1
	                              size_of_PRECISION_real, &
#endif
#if COMPLEXCASE == 1
	                              size_of_PRECISION_complex, &
#endif
				      cudaMemcpyHostToDevice)
            if (.not.(successCUDA)) then
              print *,"trans_ev_band_to_full_&
	      &MATH_DATATYPE&
	      &: error in cudaMemcpy"
              stop
            endif
#else /* WITH_MPI */

#if COMPLEXCASE == 1
            ! check whether this could be avoided like in the real case
            successCUDA = cuda_memcpy(tmp_dev,loc(tmp1),l_cols*n_cols*size_of_PRECISION_complex,cudaMemcpyHostToDevice)
            if (.not.(successCUDA)) then
              print *,"trans_ev_band_to_full_complex: error in cudaMemcpy"
              stop
            endif
#endif
#endif /* WITH_MPI */

!#ifdef WITH_MPI
! it should be possible to keep tmat on the device and not copy it aroud
!            ! copy to device, maybe this can be avoided tmat is input from bandred_real

            successCUDA = cuda_memcpy(tmat_dev, loc(tmat(1,1,istep)), nbw*nbw* &
#if REALCASE == 1
	                              size_of_PRECISION_real,                  &
#endif
#if COMPLEXCASE == 1
	                              size_of_PRECISION_complex,               &
#endif
				      cudaMemcpyHostToDevice)

            if (.not.(successCUDA)) then
              print *,"trans_ev_band_to_full_&
	      &MATH_DATATYPE&
	      &: error in cudaMemcpy"
              stop
            endif
!#endif /* WITH_MPI */

	    call timer%start("cublas")
#if REALCASE == 1
            call cublas_PRECISION_TRMM('L', 'U', 'T', 'N',       &
#endif
#if COMPLEXCASE == 1
            call cublas_PRECISION_TRMM('L', 'U', 'C', 'N',      &
#endif
                                        n_cols, l_cols, ONE, tmat_dev, nbw, tmp_dev, n_cols)

            call cublas_PRECISION_GEMM('N', 'N', l_rows, l_cols, n_cols, -ONE, hvm_dev, max_local_rows, &
                                       tmp_dev, n_cols, one, q_dev, ldq)
	    call timer%stop("cublas")

#if REALCASE == 1
            ! copy to host maybe this can be avoided
            ! this is not necessary hvm is not used anymore
            successCUDA = cuda_memcpy(loc(hvm), hvm_dev, ((max_local_rows)*nbw*size_of_PRECISION_real),cudaMemcpyDeviceToHost)
            if (.not.(successCUDA)) then
              print *,"trans_ev_band_to_full_real: error in cudaMemcpy"
              stop
            endif
#endif
          endif ! l_rows > 0

          !#ifdef WITH_GPU_VERSION
          !       istat = cuda_memcpy(loc(hvm), hvm_dev, ((max_local_rows)*nbw*size_of_real_datatype),cudaMemcpyDeviceToHost)
          !       if (istat .ne. 0) then
          !         print *,"error in cudaMemcpy"
          !         stop
          !       endif
          !
          !#endif
        enddo ! istep



      else ! do not useGPU
Andreas Marek's avatar
Andreas Marek committed
630
631

#ifdef BAND_TO_FULL_BLOCKING
632
633
634
635
636
637
638
639
640
        ! t_blocking was formerly 2; 3 is a better choice
        t_blocking = 3 ! number of matrices T (tmat) which are aggregated into a new (larger) T matrix (tmat_complete) and applied at once

        ! we only use the t_blocking if we could call it fully, this is might be better but needs to benchmarked.
!       if ( na >= ((t_blocking+1)*nbw) ) then
        cwy_blocking = t_blocking * nbw

        allocate(tmp1(max_local_cols*cwy_blocking), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
641
642
643
          print *,"trans_ev_band_to_full_&
	  &MATH_DATATYPE&
	  &: error when allocating tmp1 "//errorMessage
644
645
646
647
648
          stop
        endif

        allocate(tmp2(max_local_cols*cwy_blocking), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
649
650
651
          print *,"trans_ev_band_to_full_&
	  &MATH_DATATYPE&
	  &: error when allocating tmp2 "//errorMessage
652
653
654
655
656
          stop
        endif

        allocate(hvb(max_local_rows*cwy_blocking), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
657
658
659
          print *,"trans_ev_band_to_full_&
	  &MATH_DATATYPE&
	  &: error when allocating hvb "//errorMessage
660
661
662
663
664
          stop
        endif

        allocate(hvm(max_local_rows,cwy_blocking), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
665
666
667
          print *,"trans_ev_band_to_full_&
	  &MATH_DATATYPE&
	  &: error when allocating hvm "//errorMessage
668
669
670
          stop
        endif

Andreas Marek's avatar
Andreas Marek committed
671
672
#else /* BAND_TO_FULL_BLOCKING */

673
674
        allocate(tmp1(max_local_cols*nbw), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
675
676
677
          print *,"trans_ev_band_to_full_&
	  &MATH_DATATYPE&
	  &: error when allocating tmp1 "//errorMessage
678
679
680
681
682
          stop
        endif

        allocate(tmp2(max_local_cols*nbw), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
683
684
          print *,"trans_ev_band_to_full_&
	  &MATH_DATATYPE&: error when allocating tmp2 "//errorMessage
685
686
687
688
689
          stop
        endif

        allocate(hvb(max_local_rows*nbw), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
690
691
692
          print *,"trans_ev_band_to_full_&
	  &MATH_DATATYPE&
	  &: error when allocating hvb "//errorMessage
693
694
695
696
697
          stop
        endif

        allocate(hvm(max_local_rows,nbw), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
698
699
700
          print *,"trans_ev_band_to_full_&
	  &MATH_DATATYPE&
	  &: error when allocating hvm "//errorMessage
701
702
703
704
          stop
        endif
#endif

Andreas Marek's avatar
Andreas Marek committed
705
#ifdef BAND_TO_FULL_BLOCKING
706
707
        allocate(tmat_complete(cwy_blocking,cwy_blocking), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
708
709
710
          print *,"trans_ev_band_to_full_&
	  &MATH_DATATYPE&
	  &: error when allocating tmat_complete "//errorMessage
711
712
713
714
          stop
        endif
        allocate(t_tmp(cwy_blocking,nbw), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
715
716
717
          print *,"trans_ev_band_to_full_&
	  &MATH_DATATYPE&
	  &: error when allocating t_tmp "//errorMessage
718
719
720
721
          stop
        endif
        allocate(t_tmp2(cwy_blocking,nbw), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
722
723
724
          print *,"trans_ev_band_to_full_&
	  &MATH_DATATYPE&
	  &: error when allocating t_tmp2 "//errorMessage
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
          stop
        endif
#endif
!        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

#if REALCASE == 1
        hvm = CONST_0_0   ! Must be set to 0 !!!
        hvb = CONST_0_0   ! Safety only
#endif
#if COMPLEXCASE == 1
        hvm = CONST_COMPLEX_0_0   ! Must be set to 0 !!!
        hvb = CONST_COMPLEX_0_0   ! Safety only
#endif
        l_cols = local_index(nqc, my_pcol, np_cols, nblk, -1) ! Local columns of q

!       if ( na >= ((t_blocking+1)*nbw) ) then

Andreas Marek's avatar
Andreas Marek committed
747
#ifdef BAND_TO_FULL_BLOCKING
748
        do istep=1,((na-1)/nbw-1)/t_blocking + 1
Andreas Marek's avatar
Andreas Marek committed
749
#else
750
751
752
        do istep=1,(na-1)/nbw
#endif

Andreas Marek's avatar
Andreas Marek committed
753
#ifdef BAND_TO_FULL_BLOCKING
754
755
756
757
758
759
760
761
762
763
764
          ! This the call when using  na >= ((t_blocking+1)*nbw)
          !      n_cols = MIN(na,istep*cwy_blocking+nbw) - (istep-1)*cwy_blocking - nbw ! Number of columns in current step
          ! As an alternative we add some special case handling if na < cwy_blocking
          IF (na < cwy_blocking) THEN
            n_cols = MAX(0, na-nbw)
            IF ( n_cols .eq. 0 ) THEN
              EXIT
            END IF
          ELSE
            n_cols = MIN(na,istep*cwy_blocking+nbw) - (istep-1)*cwy_blocking - nbw ! Number of columns in current step
          END IF
Andreas Marek's avatar
Andreas Marek committed
765
#else /* BAND_TO_FULL_BLOCKING */
766
          n_cols = MIN(na,(istep+1)*nbw) - istep*nbw ! Number of columns in current step
Andreas Marek's avatar
Andreas Marek committed
767
#endif /* BAND_TO_FULL_BLOCKING */
768
769
770
771
772
773
          ! Broadcast all Householder vectors for current step compressed in hvb

          nb = 0
          ns = 0

          do lc = 1, n_cols
Andreas Marek's avatar
Andreas Marek committed
774
#ifdef BAND_TO_FULL_BLOCKING
775
            ncol = (istep-1)*cwy_blocking + nbw + lc ! absolute column number of householder vector
Andreas Marek's avatar
Andreas Marek committed
776
#else
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
            ncol = istep*nbw + lc ! absolute column number of householder vector
#endif
            nrow = ncol - nbw ! absolute number of pivot row

            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

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

            nb = nb+l_rows

            if (lc==n_cols .or. mod(ncol,nblk)==0) then
#ifdef WITH_MPI
              call timer%start("mpi_communication")
              call MPI_Bcast(hvb(ns+1), nb-ns,      &
#if REALCASE == 1
	                     MPI_REAL_PRECISION,    &
#endif
#if COMPLEXCASE == 1
                             MPI_COMPLEX_PRECISION, &
#endif
			     pcol(ncol, nblk, np_cols), mpi_comm_cols, mpierr)

              call timer%stop("mpi_communication")

#endif /* WITH_MPI */
              ns = nb
            endif
          enddo ! lc

          ! Expand compressed Householder vectors into matrix hvm

          nb = 0
          do lc = 1, n_cols
Andreas Marek's avatar
Andreas Marek committed
811
#ifdef BAND_TO_FULL_BLOCKING
812
            nrow = (istep-1)*cwy_blocking + lc ! absolute number of pivot row
Andreas Marek's avatar
Andreas Marek committed
813
#else
814
815
816
817
818
819
820
821
822
823
824
825
826
827
            nrow = (istep-1)*nbw+lc ! absolute number of pivot row
#endif
            l_rows = local_index(nrow-1, my_prow, np_rows, nblk, -1) ! row length for bcast

            hvm(1:l_rows,lc) = hvb(nb+1:nb+l_rows)
#if REALCASE == 1
            if (my_prow==prow(nrow, nblk, np_rows)) hvm(l_rows+1,lc) = CONST_1_0
#endif
#if COMPLEXCASE == 1
            if (my_prow==prow(nrow, nblk, np_rows)) hvm(l_rows+1,lc) = 1.
#endif
            nb = nb+l_rows
          enddo

Andreas Marek's avatar
Andreas Marek committed
828
#ifdef BAND_TO_FULL_BLOCKING
829
830
831
832
833
834
835
836
837
838
839
          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)
	    call timer%start("blas")
            if (i > 1) then
Andreas Marek's avatar
Andreas Marek committed
840
841
842
843
844
845
846
847
848
#if REALCASE == 1
              call PRECISION_GEMM('T', 'N',      &
#endif
#if COMPLEXCASE == 1
              call PRECISION_GEMM('C', 'N',      &
#endif
	                         t_rows, t_cols, l_rows, ONE, hvm(1,1), max_local_rows, hvm(1,(i-1)*nbw+1), &
                                 max_local_rows, ZERO, t_tmp, cwy_blocking)

849
850
851
852
            call timer%stop("blas")
#ifdef WITH_MPI
              call timer%start("mpi_communication")

Andreas Marek's avatar
Andreas Marek committed
853
854
855
856
857
858
859
860
              call mpi_allreduce(t_tmp, t_tmp2, cwy_blocking*nbw,     &
#if  REALCASE == 1
	                         MPI_REAL_PRECISION,    &
#endif
#if COMPLEXCASE == 1
                                 MPI_COMPLEX_PRECISION,       &
#endif
				 MPI_SUM, mpi_comm_rows, mpierr)
861
862
              call timer%stop("mpi_communication")
	      call timer%start("blas")
Andreas Marek's avatar
Andreas Marek committed
863
864
              call PRECISION_TRMM('L', 'U', 'N', 'N', t_rows, t_cols, ONE, tmat_complete, cwy_blocking, t_tmp2, cwy_blocking)
              call PRECISION_TRMM('R', 'U', 'N', 'N', t_rows, t_cols, -ONE, tmat_complete(t_rows+1,t_rows+1), cwy_blocking, &
865
866
867
868
869
                         t_tmp2, cwy_blocking)
     	      call timer%stop("blas")

              tmat_complete(1:t_rows,t_rows+1:t_rows+t_cols) = t_tmp2(1:t_rows,1:t_cols)

Andreas Marek's avatar
Andreas Marek committed
870
#else /* WITH_MPI */
871
872
!              t_tmp2(1:cwy_blocking,1:nbw) = t_tmp(1:cwy_blocking,1:nbw)
	      call timer%start("blas")
Andreas Marek's avatar
Andreas Marek committed
873
874
              call PRECISION_TRMM('L', 'U', 'N', 'N', t_rows, t_cols, ONE, tmat_complete, cwy_blocking, t_tmp, cwy_blocking)
              call PRECISION_TRMM('R', 'U', 'N', 'N', t_rows, t_cols, -ONE, tmat_complete(t_rows+1,t_rows+1), cwy_blocking, &
875
876
877
878
879
                         t_tmp, cwy_blocking)
	      call timer%stop("blas")

              tmat_complete(1:t_rows,t_rows+1:t_rows+t_cols) = t_tmp(1:t_rows,1:t_cols)

Andreas Marek's avatar
Andreas Marek committed
880
#endif /* WITH_MPI */
881
882
883
884
885
886
887
888

!              call PRECISION_TRMM('L', 'U', 'N', 'N', t_rows, t_cols, CONST_1_0, tmat_complete, cwy_blocking, t_tmp2, cwy_blocking)
!              call PRECISION_TRMM('R', 'U', 'N', 'N', t_rows, t_cols, -CONST_1_0, 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
Andreas Marek's avatar
Andreas Marek committed
889
#else /* BAND_TO_FULL_BLOCKING */
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
          l_rows = local_index(MIN(na,(istep+1)*nbw), my_prow, np_rows, nblk, -1)
#endif

          ! Q = Q - V * T**T * V**T * Q

          if (l_rows>0) then
            call timer%start("blas")

#if REALCASE == 1
            call PRECISION_GEMM('T', 'N',         &
#endif
#if COMPLEXCASE == 1
            call PRECISION_GEMM('C', 'N',         &
#endif
	                        n_cols, l_cols, l_rows, ONE, hvm, ubound(hvm,dim=1), &
                                q, ldq, ZERO, tmp1, n_cols)
	    call timer%stop("blas")

          else ! l_rows>0

#if REALCASE == 1
            tmp1(1:l_cols*n_cols) = CONST_0_0
#endif
#if COMPLEXCASE == 1
            tmp1(1:l_cols*n_cols) = CONST_COMPLEX_0_0
#endif
          endif ! l_rows>0

#ifdef WITH_MPI
          call timer%start("mpi_communication")
          call mpi_allreduce(tmp1, tmp2, n_cols*l_cols,  &
#if REALCASE == 1
	                     MPI_REAL_PRECISION,         &
#endif
#if COMPLEXCASE == 1
	                     MPI_COMPLEX_PRECISION,       &

#endif
			     MPI_SUM, mpi_comm_rows ,mpierr)

          call timer%stop("mpi_communication")

	  call timer%start("blas")

          if (l_rows>0) then
Andreas Marek's avatar
Andreas Marek committed
935
936
#ifdef BAND_TO_FULL_BLOCKING

937
#if REALCASE == 1
Andreas Marek's avatar
Andreas Marek committed
938
939
940
941
942
943
            call PRECISION_TRMM('L', 'U', 'T', 'N',        &
#endif
#if COMPLEXCASE == 1
            call PRECISION_TRMM('L', 'U', 'C', 'N',        &
#endif
	                        n_cols, l_cols, ONE, tmat_complete, cwy_blocking, tmp2, n_cols)
944
            call PRECISION_GEMM('N', 'N', l_rows, l_cols, n_cols, -ONE, hvm, ubound(hvm,dim=1), tmp2, n_cols, ONE, q, ldq)
Andreas Marek's avatar
Andreas Marek committed
945
946
947
948
949

#else /* BAND_TO_FULL_BLOCKING */

#if REALCASE == 1
            call PRECISION_TRMM('L', 'U', 'T', 'N',        &
950
951
#endif
#if COMPLEXCASE == 1
Andreas Marek's avatar
Andreas Marek committed
952
953
954
            call PRECISION_TRMM('L', 'U', 'C', 'N',        &
#endif
                                n_cols, l_cols, ONE, tmat(1,1,istep), ubound(tmat,dim=1), tmp2, n_cols)
955
956
            call PRECISION_GEMM('N', 'N', l_rows, l_cols, n_cols, -ONE, hvm, ubound(hvm,dim=1), &
                                tmp2, n_cols, ONE, q, ldq)
Andreas Marek's avatar
Andreas Marek committed
957
958

#endif /* BAND_TO_FULL_BLOCKING */
959
960
961
962
963
964
965

          endif
	  call timer%stop("blas")
#else /* WITH_MPI */
!          tmp2 = tmp1
	  call timer%start("blas")
          if (l_rows>0) then
Andreas Marek's avatar
Andreas Marek committed
966
967
#ifdef BAND_TO_FULL_BLOCKING

968
#if REALCASE == 1
Andreas Marek's avatar
Andreas Marek committed
969
970
971
972
973
974
            call PRECISION_TRMM('L', 'U', 'T', 'N',        &
#endif
#if COMPLEXCASE == 1
            call PRECISION_TRMM('L', 'U', 'C', 'N',        &
#endif
                                n_cols, l_cols, ONE, tmat_complete, cwy_blocking, tmp1, n_cols)
975
            call PRECISION_GEMM('N', 'N', l_rows, l_cols, n_cols, -ONE, hvm, ubound(hvm,dim=1), tmp1, n_cols, ONE, q, ldq)
Andreas Marek's avatar
Andreas Marek committed
976
977
978
979
980
981


#else /* BAND_TO_FULL_BLOCKING */

#if REALCASE == 1
            call PRECISION_TRMM('L', 'U', 'T', 'N',        &
982
983
#endif
#if COMPLEXCASE == 1
Andreas Marek's avatar
Andreas Marek committed
984
985
986
            call PRECISION_TRMM('L', 'U', 'C', 'N',        &
#endif
                                n_cols, l_cols, ONE, tmat(1,1,istep), ubound(tmat,dim=1), tmp1, n_cols)
987
988
            call PRECISION_GEMM('N', 'N', l_rows, l_cols, n_cols, -ONE, hvm, ubound(hvm,dim=1), &
                                tmp1, n_cols, ONE, q, ldq)
Andreas Marek's avatar
Andreas Marek committed
989
990

#endif  /* BAND_TO_FULL_BLOCKING */
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
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
          endif
	  call timer%stop("blas")
#endif /* WITH_MPI */

!          if (l_rows>0) then
!            call PRECISION_TRMM('L', 'U', 'T', 'N', n_cols, l_cols, CONST_1_0, tmat_complete, cwy_blocking, tmp2, n_cols)
!            call PRECISION_GEMM('N', 'N', l_rows, l_cols, n_cols, -CONST_1_0, hvm, ubound(hvm,dim=1), tmp2, n_cols, CONST_1_0, q, ldq)
!          endif

        enddo ! istep

      endif ! useGPU

      deallocate(tmp1, tmp2, hvb, stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
        print *,"trans_ev_band_to_full_&
	&MATH_DATATYPE&
	&: error when deallocating tmp1 tmp2 hvb "//errorMessage
        stop
      endif

      if (useGPU) then
        successCUDA = cuda_free(hvm_dev)
        if (.not.(successCUDA)) then
          print *,"trans_ev_band_to_full_&
	  &MATH_DATATYPE&
          &: error in cudaFree"
          stop
        endif

        successCUDA = cuda_free(tmp_dev)
        if (.not.(successCUDA)) then
          print *,"trans_ev_band_to_full_&
	  &MATH_DATATYPE&
          &: error in cudaFree"
          stop
        endif

        successCUDA = cuda_free(tmat_dev)
        if (.not.(successCUDA)) then
          print *,"trans_ev_band_to_full_&
	  &MATH_DATATYPE&
          &: error in cudaFree"
          stop
        endif

         ! final transfer of q_dev
         successCUDA = cuda_memcpy(loc(q), q_dev, ldq*matrixCols* &
#if REALCASE == 1
         size_of_PRECISION_real,    &
#endif
#if COMPLEXCASE == 1
         size_of_PRECISION_complex, &
#endif
         cudaMemcpyDeviceToHost)

         if (.not.(successCUDA)) then
          print *,"trans_ev_band_to_full_&
	  &MATH_DATATYPE&
          &: error in cudamemcpu q_dev"
          stop
         endif

         !   q(1:ldq,1:na_cols) = q_temp(1:ldq,1:na_cols)

         successCUDA = cuda_free(q_dev)
         if (.not.(successCUDA)) then
           print *,"trans_ev_band_to_full_&
	   &MATH_DATATYPE&
           &: error in cudaFree"
           stop
         endif

         !   deallocate(q_temp, stat=istat, errmsg=errorMessage)
         !   if (istat .ne. 0) then
         !     print *,"error when deallocating q_temp "//errorMessage
         !     stop
         !   endif
         !   deallocate(tmat_temp, stat=istat, errmsg=errorMessage)
         !   if (istat .ne. 0) then
         !     print *,"trans_ev_band_to_full_real: error when deallocating tmat_temp "//errorMessage
         !     stop
         !   endif

      endif ! useGPU

      deallocate(hvm, stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
        print *,"trans_ev_band_to_full_&
	&MATH_DATATYPE&
	&: error when deallocating hvm "//errorMessage
        stop
      endif

Andreas Marek's avatar
Andreas Marek committed
1085
1086
1087
1088
1089
1090
1091
#if BAND_TO_FULL_BLOCKING
      deallocate(tmat_complete, t_tmp, t_tmp2, stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
        print *,"trans_ev_band_to_full_&
	MATH_DATATYPE&
	&: error when deallocating tmat_complete, t_tmp, t_tmp2 "//errorMessage
        stop
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
      endif
#endif

      call timer%stop("trans_ev_band_to_full_&
      &MATH_DATATYPE&
      &" // &
      &PRECISION_SUFFIX&
      )

    end subroutine trans_ev_band_to_full_&
    &MATH_DATATYPE&
    &_&
    &PRECISION