elpa2_trans_ev_band_to_full_template.F90 34.2 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
#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

52
#include "../general/sanity.F90"
53
54
55
56
57

    subroutine trans_ev_band_to_full_&
    &MATH_DATATYPE&
    &_&
    &PRECISION &
58
    (obj, na, nqc, nblk, nbw, a, a_dev, lda, tmat, tmat_dev, q, &
Andreas Marek's avatar
Andreas Marek committed
59
     q_dev, ldq, matrixCols, numBlocks, mpi_comm_rows, mpi_comm_cols, useGPU &
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
#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
    !
    !-------------------------------------------------------------------------------
      use precision
      use cuda_functions
      use iso_c_binding
103
      use elpa_abstract_impl
104
      implicit none
105
#include "../general/precision_kinds.F90"
106
      class(elpa_abstract_impl_t), intent(inout) :: obj
107
108
109
110
111
112
      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
#ifdef USE_ASSUMED_SIZE
113
      MATH_DATATYPE(kind=rck)               :: a(lda,*), q(ldq,*), tmat(nbw,nbw,*)
114
#else
115
      MATH_DATATYPE(kind=rck)               :: a(lda,matrixCols), q(ldq,matrixCols), tmat(nbw, nbw, numBlocks)
116
117
118
119
120
121
122
123
124
#endif
      integer(kind=C_intptr_T)               :: a_dev ! passed from bandred_real at the moment not used since copied in bandred_real

      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

125
      MATH_DATATYPE(kind=rck), allocatable  :: tmp1(:), tmp2(:), hvb(:), hvm(:,:)
126
127
128
129
130
131
132
133
      ! 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
134
#ifdef BAND_TO_FULL_BLOCKING
135
      MATH_DATATYPE(kind=rck), allocatable  :: tmat_complete(:,:), t_tmp(:,:), t_tmp2(:,:)
136
137
      integer(kind=ik)                       :: cwy_blocking, t_blocking, t_cols, t_rows
#endif
Andreas Marek's avatar
Andreas Marek committed
138

139
140
141
      integer(kind=ik)                       :: istat
      character(200)                         :: errorMessage
      logical                                :: successCUDA
142
      integer(kind=c_intptr_t), parameter      :: size_of_datatype = size_of_&
143
144
145
                                                                   &PRECISION&
                                                                   &_&
                                                                   &MATH_DATATYPE
146
      call obj%timer%start("trans_ev_band_to_full_&
147
      &MATH_DATATYPE&
Andreas Marek's avatar
Andreas Marek committed
148
      &" // &
149
150
      &PRECISION_SUFFIX &
      )
151
      call obj%timer%start("mpi_communication")
152
153
154
155
156
157

      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)

158
      call obj%timer%stop("mpi_communication")
159
160
161
162
163
164
165
166
167
168
169
170
171
172

      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
173
174
175
        ! the GPU version does not (yet) support blocking
        ! but the handling is the same for real/complex case

176
177
178
        allocate(tmp1(max_local_cols*nbw), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab    
Andreas Marek committed
179
180
    &MATH_DATATYPE&
    &: error when allocating tmp1 "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
181
          stop 1
182
183
184
185
186
        endif

        allocate(tmp2(max_local_cols*nbw), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab    
Andreas Marek committed
187
188
    &MATH_DATATYPE&
    &: error when allocating tmp2 "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
189
          stop 1
190
191
192
193
194
        endif

        allocate(hvb(max_local_rows*nbw), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab    
Andreas Marek committed
195
196
    &MATH_DATATYPE&
    &: error when allocating hvb "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
197
          stop 1
198
199
200
201
202
        endif

        allocate(hvm(max_local_rows,nbw), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab    
Andreas Marek committed
203
204
    &MATH_DATATYPE&
    &: error when allocating hvm "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
205
          stop 1
206
207
        endif

208
        successCUDA = cuda_malloc(hvm_dev, (max_local_rows)*nbw* size_of_datatype)
209
210
        if (.not.(successCUDA)) then
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab    
Andreas Marek committed
211
212
    &MATH_DATATYPE&
    &: error in cudaMalloc"
Andreas Marek's avatar
Andreas Marek committed
213
          stop 1
214
215
        endif

216
        successCUDA = cuda_malloc(tmp_dev, (max_local_cols)*nbw* size_of_datatype)
217
218
        if (.not.(successCUDA)) then
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab    
Andreas Marek committed
219
220
    &MATH_DATATYPE&
    &: error in cudaMalloc"
Andreas Marek's avatar
Andreas Marek committed
221
          stop 1
222
223
224
        endif

!#ifdef WITH_MPI
Andreas Marek's avatar
Andreas Marek committed
225
226
227
228
!! 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
Andreas Marek's avatar
Retab    
Andreas Marek committed
229
!  size_of_PRECISION_real)
Andreas Marek's avatar
Andreas Marek committed
230
231
232
233
234
235
236
!#endif
!#if COMPLEXCASE == 1
!        size_of_PRECISION_complex)
!#endif
!
!        if (.not.(successCUDA)) then
!          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab    
Andreas Marek committed
237
238
!    &MATH_DATATYPE&
!    &: error in cudaMalloc"
Andreas Marek's avatar
Andreas Marek committed
239
!          stop 1
Andreas Marek's avatar
Andreas Marek committed
240
!        endif
241
242
243
244
!#endif

#if REALCASE == 1
! q_dev already living on device
245
!        successCUDA = cuda_malloc(q_dev, ldq*matrixCols*size_of_datatype)
246
247
!        if (.not.(successCUDA)) then
!          print *,"trans_ev_band_to_full_real: error in cudaMalloc"
Andreas Marek's avatar
Andreas Marek committed
248
!          stop 1
249
250
251
252
253
254
255
256
!        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"
Andreas Marek's avatar
Andreas Marek committed
257
!          stop 1
258
259
260
!        endif
#endif
#if COMPLEXCASE == 1
Andreas Marek's avatar
Andreas Marek committed
261
262
263
!         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"
Andreas Marek's avatar
Andreas Marek committed
264
!           stop 1
Andreas Marek's avatar
Andreas Marek committed
265
266
267
268
269
!         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"
Andreas Marek's avatar
Andreas Marek committed
270
!            stop 1
Andreas Marek's avatar
Andreas Marek committed
271
!          endif
272
273
274
#endif

        ! if MPI is NOT used the following steps could be done on the GPU and memory transfers could be avoided
275
        successCUDA = cuda_memset(hvm_dev, 0, (max_local_rows)*(nbw)* size_of_datatype)
276
277
        if (.not.(successCUDA)) then
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab    
Andreas Marek committed
278
279
    &MATH_DATATYPE&
    &: error in cudaMalloc"
Andreas Marek's avatar
Andreas Marek committed
280
          stop 1
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
        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
303
            ncol = istep*nbw + lc ! absolute column number of householder Vector
304
305
306
307
308
309
310
311
312
313
314
            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
315
              call obj%timer%start("mpi_communication")
316
317
              call MPI_Bcast(hvb(ns+1), nb-ns,  &
#if REALCASE == 1
Andreas Marek's avatar
Retab    
Andreas Marek committed
318
                       MPI_REAL_PRECISION,&
319
320
321
322
#endif
#if COMPLEXCASE == 1
                             MPI_COMPLEX_PRECISION, &
#endif
Andreas Marek's avatar
Retab    
Andreas Marek committed
323
           pcol(ncol, nblk, np_cols), mpi_comm_cols, mpierr)
324

325
              call obj%timer%stop("mpi_communication")
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348

#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

349
          successCUDA = cuda_memcpy(hvm_dev, loc(hvm), max_local_rows*nbw* size_of_datatype, cudaMemcpyHostToDevice)
350
351
352

          if (.not.(successCUDA)) then
            print *,"trans_ev_band_to_full_real: error in cudaMemcpy"
Andreas Marek's avatar
Andreas Marek committed
353
            stop 1
354
355
356
357
358
359
360
361

          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
Andreas Marek's avatar
Retab    
Andreas Marek committed
362
      call obj%timer%start("cublas")
363
364
365
366
367
368
#if REALCASE == 1
            call cublas_PRECISION_GEMM('T', 'N',        &
#endif
#if COMPLEXCASE == 1
            call cublas_PRECISION_GEMM('C', 'N',        &
#endif
Andreas Marek's avatar
Retab    
Andreas Marek committed
369
                                 n_cols, l_cols, l_rows, ONE, hvm_dev, max_local_rows, &
370
                                       q_dev, ldq , ZERO, tmp_dev, n_cols)
Andreas Marek's avatar
Retab    
Andreas Marek committed
371
      call obj%timer%stop("cublas")
372

Andreas Marek's avatar
Andreas Marek committed
373
#if REALCASE == 1
374
375
376
377
#ifdef WITH_MPI

            ! copy data from device to host for a later MPI_ALLREDUCE
            ! copy to host maybe this can be avoided this is needed if MPI is used (allreduce)
378
            successCUDA = cuda_memcpy(loc(tmp1), tmp_dev, l_cols*n_cols*size_of_datatype, cudaMemcpyDeviceToHost)
379
380
            if (.not.(successCUDA)) then
              print *,"trans_ev_band_to_full_real: error in cudaMemcpy"
Andreas Marek's avatar
Andreas Marek committed
381
              stop 1
382
383
384
            endif

#else /* WITH_MPI */
Andreas Marek's avatar
Andreas Marek committed
385
386
387
388
            ! in real case no copy needed. maybe also in complexcase ?
#endif /* WITH_MPI */

#endif /* REALCASE */
389
390

#if COMPLEXCASE == 1
391
            successCUDA = cuda_memcpy(loc(tmp1), tmp_dev, n_cols*l_cols*size_of_datatype, &
392
393
394
395
                                       cudaMemcpyDeviceToHost)

            if (.not.(successCUDA)) then
              print *,"trans_ev_band_to_full_complex: error in cudaMemcpy"
Andreas Marek's avatar
Andreas Marek committed
396
              stop 1
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
            endif
#endif

          else ! l_rows>0

#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

          endif ! l_rows>0

          !#ifdef WITH_GPU_VERSION
412
          !       istat = cuda_memcpy(loc(tmp1), tmp_dev, max_local_cols*nbw*size_of_datatype,cudaMemcpyDeviceToHost)
413
414
          !       if (istat .ne. 0) then
          !         print *,"error in cudaMemcpy"
Andreas Marek's avatar
Andreas Marek committed
415
          !         stop 1
416
417
418
419
          !       endif
          !#endif

#ifdef WITH_MPI
420
          call obj%timer%start("mpi_communication")
421
422
423
424
425
426
427
428
          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)
429
          call obj%timer%stop("mpi_communication")
430
431
432
433
434
435

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

           !#ifdef WITH_GPU_VERSION
436
          !       istat = cuda_memcpy(tmp_dev, loc(tmp2), max_local_cols*nbw*size_of_datatype,cudaMemcpyHostToDevice)
437
438
          !       if (istat .ne. 0) then
          !         print *,"error in cudaMemcpy"
Andreas Marek's avatar
Andreas Marek committed
439
          !         stop 1
440
441
442
443
444
445
446
          !       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
447
            successCUDA = cuda_memcpy(tmp_dev, loc(tmp2), n_cols*l_cols* size_of_datatype, &
Andreas Marek's avatar
Retab    
Andreas Marek committed
448
              cudaMemcpyHostToDevice)
449
450
            if (.not.(successCUDA)) then
              print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab    
Andreas Marek committed
451
452
        &MATH_DATATYPE&
        &: error in cudaMemcpy"
Andreas Marek's avatar
Andreas Marek committed
453
              stop 1
454
455
456
457
458
            endif
#else /* WITH_MPI */

#if COMPLEXCASE == 1
            ! check whether this could be avoided like in the real case
459
            successCUDA = cuda_memcpy(tmp_dev,loc(tmp1),l_cols*n_cols*size_of_datatype,cudaMemcpyHostToDevice)
460
461
            if (.not.(successCUDA)) then
              print *,"trans_ev_band_to_full_complex: error in cudaMemcpy"
Andreas Marek's avatar
Andreas Marek committed
462
              stop 1
463
464
465
466
            endif
#endif
#endif /* WITH_MPI */

467
!#ifdef WITH_MPI
468
469
470
471
472
473
474
           ! IMPORTANT: even though tmat_dev is transfered from the previous rutine, we have to copy from tmat again 
           ! tmat is 3-dimensional array, while tmat_dev contains only one 2-dimensional slice of it - and here we 
           ! need to upload another slice
           successCUDA = cuda_memcpy(tmat_dev, loc(tmat(1,1,istep)), nbw*nbw*size_of_datatype, cudaMemcpyHostToDevice)

           if (.not.(successCUDA)) then
             print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab    
Andreas Marek committed
475
476
        &MATH_DATATYPE&
        &: error in cudaMemcpy"
477
478
             stop 1
           endif
479
!#endif /* WITH_MPI */
480

Andreas Marek's avatar
Retab    
Andreas Marek committed
481
      call obj%timer%start("cublas")
482
483
484
485
486
487
488
489
490
491
#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)
Andreas Marek's avatar
Retab    
Andreas Marek committed
492
      call obj%timer%stop("cublas")
493
494
495
496

#if REALCASE == 1
            ! copy to host maybe this can be avoided
            ! this is not necessary hvm is not used anymore
497
            successCUDA = cuda_memcpy(loc(hvm), hvm_dev, ((max_local_rows)*nbw*size_of_datatype),cudaMemcpyDeviceToHost)
498
499
            if (.not.(successCUDA)) then
              print *,"trans_ev_band_to_full_real: error in cudaMemcpy"
Andreas Marek's avatar
Andreas Marek committed
500
              stop 1
501
502
503
504
505
            endif
#endif
          endif ! l_rows > 0

          !#ifdef WITH_GPU_VERSION
506
          !       istat = cuda_memcpy(loc(hvm), hvm_dev, ((max_local_rows)*nbw*size_of_datatype),cudaMemcpyDeviceToHost)
507
508
          !       if (istat .ne. 0) then
          !         print *,"error in cudaMemcpy"
Andreas Marek's avatar
Andreas Marek committed
509
          !         stop 1
510
511
512
513
514
515
516
517
          !       endif
          !
          !#endif
        enddo ! istep



      else ! do not useGPU
Andreas Marek's avatar
Andreas Marek committed
518
519

#ifdef BAND_TO_FULL_BLOCKING
520
521
522
523
524
525
526
527
528
        ! 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
529
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab    
Andreas Marek committed
530
531
    &MATH_DATATYPE&
    &: error when allocating tmp1 "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
532
          stop 1
533
534
535
536
        endif

        allocate(tmp2(max_local_cols*cwy_blocking), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
537
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab    
Andreas Marek committed
538
539
    &MATH_DATATYPE&
    &: error when allocating tmp2 "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
540
          stop 1
541
542
543
544
        endif

        allocate(hvb(max_local_rows*cwy_blocking), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
545
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab    
Andreas Marek committed
546
547
    &MATH_DATATYPE&
    &: error when allocating hvb "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
548
          stop 1
549
550
551
552
        endif

        allocate(hvm(max_local_rows,cwy_blocking), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
553
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab    
Andreas Marek committed
554
555
    &MATH_DATATYPE&
    &: error when allocating hvm "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
556
          stop 1
557
558
        endif

Andreas Marek's avatar
Andreas Marek committed
559
560
#else /* BAND_TO_FULL_BLOCKING */

561
562
        allocate(tmp1(max_local_cols*nbw), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
563
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab    
Andreas Marek committed
564
565
    &MATH_DATATYPE&
    &: error when allocating tmp1 "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
566
          stop 1
567
568
569
570
        endif

        allocate(tmp2(max_local_cols*nbw), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
571
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab    
Andreas Marek committed
572
    &MATH_DATATYPE&: error when allocating tmp2 "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
573
          stop 1
574
575
576
577
        endif

        allocate(hvb(max_local_rows*nbw), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
578
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab    
Andreas Marek committed
579
580
    &MATH_DATATYPE&
    &: error when allocating hvb "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
581
          stop 1
582
583
584
585
        endif

        allocate(hvm(max_local_rows,nbw), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
586
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab    
Andreas Marek committed
587
588
    &MATH_DATATYPE&
    &: error when allocating hvm "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
589
          stop 1
590
        endif
Andreas Marek's avatar
Andreas Marek committed
591
#endif /* BAND_TO_FULL_BLOCKING */
592

Andreas Marek's avatar
Andreas Marek committed
593
#ifdef BAND_TO_FULL_BLOCKING
594
595
        allocate(tmat_complete(cwy_blocking,cwy_blocking), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
596
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab    
Andreas Marek committed
597
598
    &MATH_DATATYPE&
    &: error when allocating tmat_complete "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
599
          stop 1
600
601
602
        endif
        allocate(t_tmp(cwy_blocking,nbw), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
603
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab    
Andreas Marek committed
604
605
    &MATH_DATATYPE&
    &: error when allocating t_tmp "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
606
          stop 1
607
608
609
        endif
        allocate(t_tmp2(cwy_blocking,nbw), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
610
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab    
Andreas Marek committed
611
612
    &MATH_DATATYPE&
    &: error when allocating t_tmp2 "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
613
          stop 1
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
        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
635
#ifdef BAND_TO_FULL_BLOCKING
636
        do istep=1,((na-1)/nbw-1)/t_blocking + 1
Andreas Marek's avatar
Andreas Marek committed
637
#else
638
639
640
        do istep=1,(na-1)/nbw
#endif

Andreas Marek's avatar
Andreas Marek committed
641
#ifdef BAND_TO_FULL_BLOCKING
642
643
644
645
646
647
648
649
650
651
652
          ! 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
653
#else /* BAND_TO_FULL_BLOCKING */
654
          n_cols = MIN(na,(istep+1)*nbw) - istep*nbw ! Number of columns in current step
Andreas Marek's avatar
Andreas Marek committed
655
#endif /* BAND_TO_FULL_BLOCKING */
656
657
658
659
660
661
          ! 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
662
#ifdef BAND_TO_FULL_BLOCKING
663
            ncol = (istep-1)*cwy_blocking + nbw + lc ! absolute column number of householder Vector
Andreas Marek's avatar
Andreas Marek committed
664
#else
665
            ncol = istep*nbw + lc ! absolute column number of householder Vector
666
667
668
669
670
671
672
673
674
675
676
677
#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
678
              call obj%timer%start("mpi_communication")
679
680
              call MPI_Bcast(hvb(ns+1), nb-ns,      &
#if REALCASE == 1
Andreas Marek's avatar
Retab    
Andreas Marek committed
681
                       MPI_REAL_PRECISION,    &
682
683
684
685
#endif
#if COMPLEXCASE == 1
                             MPI_COMPLEX_PRECISION, &
#endif
Andreas Marek's avatar
Retab    
Andreas Marek committed
686
           pcol(ncol, nblk, np_cols), mpi_comm_cols, mpierr)
687

688
              call obj%timer%stop("mpi_communication")
689
690
691
692
693
694
695
696
697
698

#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
699
#ifdef BAND_TO_FULL_BLOCKING
700
            nrow = (istep-1)*cwy_blocking + lc ! absolute number of pivot row
Andreas Marek's avatar
Andreas Marek committed
701
#else
702
703
704
705
706
707
708
709
710
711
712
713
714
715
            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
716
#ifdef BAND_TO_FULL_BLOCKING
717
718
719
720
721
722
723
724
725
          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)
Andreas Marek's avatar
Andreas Marek committed
726

727
            if (i > 1) then
Andreas Marek's avatar
Retab    
Andreas Marek committed
728
        call obj%timer%start("blas")
Andreas Marek's avatar
Andreas Marek committed
729
730
731
732
733
734
#if REALCASE == 1
              call PRECISION_GEMM('T', 'N',      &
#endif
#if COMPLEXCASE == 1
              call PRECISION_GEMM('C', 'N',      &
#endif
Andreas Marek's avatar
Retab    
Andreas Marek committed
735
                           t_rows, t_cols, l_rows, ONE, hvm(1,1), max_local_rows, hvm(1,(i-1)*nbw+1), &
Andreas Marek's avatar
Andreas Marek committed
736
737
                                 max_local_rows, ZERO, t_tmp, cwy_blocking)

738
            call obj%timer%stop("blas")
739
#ifdef WITH_MPI
740
              call obj%timer%start("mpi_communication")
741

Andreas Marek's avatar
Andreas Marek committed
742
743
              call mpi_allreduce(t_tmp, t_tmp2, cwy_blocking*nbw,     &
#if  REALCASE == 1
Andreas Marek's avatar
Retab    
Andreas Marek committed
744
                           MPI_REAL_PRECISION,    &
Andreas Marek's avatar
Andreas Marek committed
745
746
747
748
#endif
#if COMPLEXCASE == 1
                                 MPI_COMPLEX_PRECISION,       &
#endif
Andreas Marek's avatar
Retab    
Andreas Marek committed
749
         MPI_SUM, mpi_comm_rows, mpierr)
750
              call obj%timer%stop("mpi_communication")
Andreas Marek's avatar
Retab    
Andreas Marek committed
751
        call obj%timer%start("blas")
Andreas Marek's avatar
Andreas Marek committed
752
753
              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, &
754
                         t_tmp2, cwy_blocking)
Andreas Marek's avatar
Retab    
Andreas Marek committed
755
             call obj%timer%stop("blas")
756
757
758

              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
759
#else /* WITH_MPI */
760
!              t_tmp2(1:cwy_blocking,1:nbw) = t_tmp(1:cwy_blocking,1:nbw)
Andreas Marek's avatar
Retab    
Andreas Marek committed
761
        call obj%timer%start("blas")
Andreas Marek's avatar
Andreas Marek committed
762
763
              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, &
764
                         t_tmp, cwy_blocking)
Andreas Marek's avatar
Retab    
Andreas Marek committed
765
        call obj%timer%stop("blas")
766
767
768

              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
769
#endif /* WITH_MPI */
770
771
772
773
774
775
776
777

!              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
778
#else /* BAND_TO_FULL_BLOCKING */
779
780
781
782
783
784
          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
785
            call obj%timer%start("blas")
786
787
788
789
790
791
792

#if REALCASE == 1
            call PRECISION_GEMM('T', 'N',         &
#endif
#if COMPLEXCASE == 1
            call PRECISION_GEMM('C', 'N',         &
#endif
Andreas Marek's avatar
Retab    
Andreas Marek committed
793
                          n_cols, l_cols, l_rows, ONE, hvm, ubound(hvm,dim=1), &
794
                                q, ldq, ZERO, tmp1, n_cols)
Andreas Marek's avatar
Retab    
Andreas Marek committed
795
      call obj%timer%stop("blas")
796
797
798
799
800
801
802
803
804
805
806
807

          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
808
          call obj%timer%start("mpi_communication")
809
810
          call mpi_allreduce(tmp1, tmp2, n_cols*l_cols,  &
#if REALCASE == 1
Andreas Marek's avatar
Retab    
Andreas Marek committed
811
                       MPI_REAL_PRECISION,         &
812
813
#endif
#if COMPLEXCASE == 1
Andreas Marek's avatar
Retab    
Andreas Marek committed
814
                       MPI_COMPLEX_PRECISION,       &
815
816

#endif
Andreas Marek's avatar
Retab    
Andreas Marek committed
817
           MPI_SUM, mpi_comm_rows ,mpierr)
818

819
          call obj%timer%stop("mpi_communication")
820

Andreas Marek's avatar
Retab    
Andreas Marek committed
821
    call obj%timer%start("blas")
822
823

          if (l_rows>0) then
Andreas Marek's avatar
Andreas Marek committed
824
825
#ifdef BAND_TO_FULL_BLOCKING

826
#if REALCASE == 1
Andreas Marek's avatar
Andreas Marek committed
827
828
829
830
831
            call PRECISION_TRMM('L', 'U', 'T', 'N',        &
#endif
#if COMPLEXCASE == 1
            call PRECISION_TRMM('L', 'U', 'C', 'N',        &
#endif
Andreas Marek's avatar
Retab    
Andreas Marek committed
832
                          n_cols, l_cols, ONE, tmat_complete, cwy_blocking, tmp2, n_cols)
833
            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
834
835
836
837
838

#else /* BAND_TO_FULL_BLOCKING */

#if REALCASE == 1
            call PRECISION_TRMM('L', 'U', 'T', 'N',        &
839
840
#endif
#if COMPLEXCASE == 1
Andreas Marek's avatar
Andreas Marek committed
841
842
843
            call PRECISION_TRMM('L', 'U', 'C', 'N',        &
#endif
                                n_cols, l_cols, ONE, tmat(1,1,istep), ubound(tmat,dim=1), tmp2, n_cols)
844
845
            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
846
847

#endif /* BAND_TO_FULL_BLOCKING */
848
849

          endif
Andreas Marek's avatar
Retab    
Andreas Marek committed
850
    call obj%timer%stop("blas")
851
852
#else /* WITH_MPI */
!          tmp2 = tmp1
Andreas Marek's avatar
Retab    
Andreas Marek committed
853
    call obj%timer%start("blas")
854
          if (l_rows>0) then
Andreas Marek's avatar
Andreas Marek committed
855
856
#ifdef BAND_TO_FULL_BLOCKING

857
#if REALCASE == 1
Andreas Marek's avatar
Andreas Marek committed
858
859
860
861
862
863
            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)
864
            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
865
866
867
868
869
870


#else /* BAND_TO_FULL_BLOCKING */

#if REALCASE == 1
            call PRECISION_TRMM('L', 'U', 'T', 'N',        &
871
872
#endif
#if COMPLEXCASE == 1
Andreas Marek's avatar
Andreas Marek committed
873
874
875
            call PRECISION_TRMM('L', 'U', 'C', 'N',        &
#endif
                                n_cols, l_cols, ONE, tmat(1,1,istep), ubound(tmat,dim=1), tmp1, n_cols)
876
877
            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
878
879

#endif  /* BAND_TO_FULL_BLOCKING */
880
          endif
Andreas Marek's avatar
Retab    
Andreas Marek committed
881
    call obj%timer%stop("blas")
882
883
884
885
886
887
888
889
890
891
892
893
894
895
#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_&
Andreas Marek's avatar
Retab    
Andreas Marek committed
896
897
  &MATH_DATATYPE&
  &: error when deallocating tmp1 tmp2 hvb "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
898
        stop 1
899
900
901
902
903
904
      endif

      if (useGPU) then
        successCUDA = cuda_free(hvm_dev)
        if (.not.(successCUDA)) then
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab    
Andreas Marek committed
905
    &MATH_DATATYPE&
906
          &: error in cudaFree"
Andreas Marek's avatar
Andreas Marek committed
907
          stop 1
908
909
910
911
912
        endif

        successCUDA = cuda_free(tmp_dev)
        if (.not.(successCUDA)) then
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab    
Andreas Marek committed
913
    &MATH_DATATYPE&
914
          &: error in cudaFree"
Andreas Marek's avatar
Andreas Marek committed
915
          stop 1
916
917
918
919
920
        endif

        successCUDA = cuda_free(tmat_dev)
        if (.not.(successCUDA)) then
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab    
Andreas Marek committed
921
    &MATH_DATATYPE&
922
          &: error in cudaFree"
Andreas Marek's avatar
Andreas Marek committed
923
          stop 1
924
925
926
        endif

         ! final transfer of q_dev
927
         successCUDA = cuda_memcpy(loc(q), q_dev, ldq*matrixCols* size_of_datatype, cudaMemcpyDeviceToHost)
928
929
930

         if (.not.(successCUDA)) then
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab    
Andreas Marek committed
931
    &MATH_DATATYPE&
932
          &: error in cudamemcpu q_dev"
Andreas Marek's avatar
Andreas Marek committed
933
          stop 1
934
935
936
937
938
939
940
         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_&
Andreas Marek's avatar
Retab    
Andreas Marek committed
941
     &MATH_DATATYPE&
942
           &: error in cudaFree"
Andreas Marek's avatar
Andreas Marek committed
943
           stop 1
944
945
946
947
948
         endif

         !   deallocate(q_temp, stat=istat, errmsg=errorMessage)
         !   if (istat .ne. 0) then
         !     print *,"error when deallocating q_temp "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
949
         !     stop 1
950
951
952
953
         !   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
Andreas Marek's avatar
Andreas Marek committed
954
         !     stop 1
955
956
957
958
959
960
961
         !   endif

      endif ! useGPU

      deallocate(hvm, stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
        print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab    
Andreas Marek committed
962
963
  &MATH_DATATYPE&
  &: error when deallocating hvm "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
964
        stop 1
965
966
      endif

Andreas Marek's avatar
Andreas Marek committed
967
#if BAND_TO_FULL_BLOCKING
968
969
970
971
      if (.not.(useGPU)) then
        deallocate(tmat_complete, t_tmp, t_tmp2, stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab    
Andreas Marek committed
972
973
    &MATH_DATATYPE&
    &: error when deallocating tmat_complete, t_tmp, t_tmp2 "//errorMessage
974
975
          stop 1
        endif
976
977
978
      endif
#endif

979
      call obj%timer%stop("trans_ev_band_to_full_&
980
981
982
983
984
985
986
987
988
989
990
      &MATH_DATATYPE&
      &" // &
      &PRECISION_SUFFIX&
      )

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