elpa2_trans_ev_tridi_to_band_template.F90 98.2 KB
Newer Older
Andreas Marek's avatar
Andreas Marek committed
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
!    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.
!
!
! 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".

53
#include "../general/sanity.F90"
Andreas Marek's avatar
Andreas Marek committed
54
55

  subroutine trans_ev_tridi_to_band_&
56
57
58
    &MATH_DATATYPE&
    &_&
    &PRECISION &
59
    (obj, na, nev, nblk, nbw, q, q_dev, ldq, matrixCols,         &
60
     hh_trans, mpi_comm_rows, mpi_comm_cols, wantDebug, useGPU, success, &
61
     kernel)
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

    !-------------------------------------------------------------------------------
    !  trans_ev_tridi_to_band_real/complex:
    !  Transforms the eigenvectors of a tridiagonal matrix back to the eigenvectors of the band matrix
    !
    !  Parameters
    !
    !  na          Order of matrix a, number of rows of matrix q
    !
    !  nev         Number eigenvectors to compute (= columns of matrix q)
    !
    !  nblk        blocksize of cyclic distribution, must be the same in both directions!
    !
    !  nb          semi bandwith
    !
    !  q           On input: Eigenvectors of tridiagonal matrix
    !              On output: Transformed eigenvectors
    !              Distribution is like in Scalapack.
    !
    !  q_dev       GPU device pointer to q
    !
    !  ldq         Leading dimension of q
    !  matrixCols  local columns of matrix q
    !
    !  mpi_comm_rows
    !  mpi_comm_cols
    !              MPI-Communicators for rows/columns/both
    !
    !-------------------------------------------------------------------------------
91
      use elpa_abstract_impl
92
      use elpa2_workload
Andreas Marek's avatar
Andreas Marek committed
93
      use pack_unpack_cpu
Andreas Marek's avatar
Andreas Marek committed
94
      use pack_unpack_gpu
95
      use compute_hh_trafo
96
97
98
99
      use cuda_functions
      use precision
      use iso_c_binding
      implicit none
100
      class(elpa_abstract_impl_t), intent(inout) :: obj
Andreas Marek's avatar
Andreas Marek committed
101
      logical, intent(in)                        :: useGPU
102

Andreas Marek's avatar
Andreas Marek committed
103
104
      integer(kind=ik), intent(in)               :: kernel
      integer(kind=ik), intent(in)               :: na, nev, nblk, nbw, ldq, matrixCols, mpi_comm_rows, mpi_comm_cols
105
106
107

#if REALCASE == 1
#ifdef USE_ASSUMED_SIZE
Andreas Marek's avatar
Andreas Marek committed
108
      real(kind=REAL_DATATYPE)                   :: q(ldq,*)
109
#else
Andreas Marek's avatar
Andreas Marek committed
110
      real(kind=REAL_DATATYPE)                   :: q(ldq,matrixCols)
111
112
#endif

Andreas Marek's avatar
Andreas Marek committed
113
      real(kind=REAL_DATATYPE), intent(in)       :: hh_trans(:,:)
114
#endif
Andreas Marek's avatar
Andreas Marek committed
115
      integer(kind=c_intptr_t)                   :: q_dev
Andreas Marek's avatar
Andreas Marek committed
116

117
118
#if COMPLEXCASE == 1
#ifdef USE_ASSUMED_SIZE
Andreas Marek's avatar
Andreas Marek committed
119
      complex(kind=COMPLEX_DATATYPE)             :: q(ldq,*)
120
#else
Andreas Marek's avatar
Andreas Marek committed
121
      complex(kind=COMPLEX_DATATYPE)             :: q(ldq,matrixCols)
122
#endif
Andreas Marek's avatar
Andreas Marek committed
123
      complex(kind=COMPLEX_DATATYPE)             :: hh_trans(:,:)
124
125
#endif

Andreas Marek's avatar
Andreas Marek committed
126
127
      integer(kind=ik)                           :: np_rows, my_prow, np_cols, my_pcol
      integer(kind=ik)                           :: tmp
128

Andreas Marek's avatar
Andreas Marek committed
129
130
131
132
133
      integer(kind=ik)                           :: i, j, ip, sweep, nbuf, l_nev, a_dim2
      integer(kind=ik)                           :: current_n, current_local_n, current_n_start, current_n_end
      integer(kind=ik)                           :: next_n, next_local_n, next_n_start, next_n_end
      integer(kind=ik)                           :: bottom_msg_length, top_msg_length, next_top_msg_length
      integer(kind=ik)                           :: stripe_width, last_stripe_width, stripe_count
134
#ifdef WITH_OPENMP
Andreas Marek's avatar
Andreas Marek committed
135
      integer(kind=ik)                           :: thread_width, csw, b_off, b_len
136
#endif
Andreas Marek's avatar
Andreas Marek committed
137
138
139
      integer(kind=ik)                           :: num_result_blocks, num_result_buffers, num_bufs_recvd
      integer(kind=ik)                           :: a_off, current_tv_off, max_blk_size
      integer(kind=ik)                           :: mpierr, src, src_offset, dst, offset, nfact, num_blk
140

Andreas Marek's avatar
Andreas Marek committed
141
      logical                                    :: flag
142
143
#if REALCASE == 1
#ifdef WITH_OPENMP
Andreas Marek's avatar
Andreas Marek committed
144
      real(kind=REAL_DATATYPE), pointer          :: aIntern(:,:,:,:)
145
#else
Andreas Marek's avatar
Andreas Marek committed
146
      real(kind=REAL_DATATYPE), pointer          :: aIntern(:,:,:)
147
#endif
Andreas Marek's avatar
Andreas Marek committed
148
      real(kind=REAL_DATATYPE)                   :: a_real
149
150
151
152
#endif

#if COMPLEXCASE == 1
#ifdef WITH_OPENMP
Andreas Marek's avatar
Andreas Marek committed
153
      complex(kind=COMPLEX_DATATYPE), pointer    :: aIntern(:,:,:,:)
154
#else
Andreas Marek's avatar
Andreas Marek committed
155
      complex(kind=COMPLEX_DATATYPE), pointer    :: aIntern(:,:,:)
156
#endif
Andreas Marek's avatar
Andreas Marek committed
157
      complex(kind=COMPLEX_DATATYPE)             :: a_complex
158
#endif
Andreas Marek's avatar
Andreas Marek committed
159
      type(c_ptr)                                :: aIntern_ptr
160
161

#if REALCASE == 1
Andreas Marek's avatar
Andreas Marek committed
162
163
      real(kind=REAL_DATATYPE)   , allocatable   :: row(:)
      real(kind=REAL_DATATYPE)   , allocatable   :: row_group(:,:)
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
#endif
#if COMPLEXCASE == 1
      complex(kind=COMPLEX_DATATYPE), allocatable :: row(:)
      complex(kind=COMPLEX_DATATYPE), allocatable :: row_group(:,:)
#endif

#if REALCASE == 1
#ifdef WITH_OPENMP
      real(kind=REAL_DATATYPE), allocatable    :: top_border_send_buffer(:,:), top_border_recv_buffer(:,:)
      real(kind=REAL_DATATYPE), allocatable    :: bottom_border_send_buffer(:,:), bottom_border_recv_buffer(:,:)
#else
      real(kind=REAL_DATATYPE), allocatable    :: top_border_send_buffer(:,:,:), top_border_recv_buffer(:,:,:)
      real(kind=REAL_DATATYPE), allocatable    :: bottom_border_send_buffer(:,:,:), bottom_border_recv_buffer(:,:,:)
#endif
#endif

#if COMPLEXCASE == 1
#ifdef WITH_OPENMP
      complex(kind=COMPLEX_DATATYPE), allocatable :: top_border_send_buffer(:,:), top_border_recv_buffer(:,:)
      complex(kind=COMPLEX_DATATYPE), allocatable :: bottom_border_send_buffer(:,:), bottom_border_recv_buffer(:,:)
#else
      complex(kind=COMPLEX_DATATYPE), allocatable :: top_border_send_buffer(:,:,:), top_border_recv_buffer(:,:,:)
      complex(kind=COMPLEX_DATATYPE), allocatable :: bottom_border_send_buffer(:,:,:), bottom_border_recv_buffer(:,:,:)
#endif
#endif

      integer(kind=c_intptr_t)                 :: aIntern_dev
      integer(kind=c_intptr_t)                 :: bcast_buffer_dev
192
193
      integer(kind=c_intptr_t)                   :: num
      integer(kind=c_intptr_t)                   :: dev_offset, dev_offset_1, dev_offset_2
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
226
227
228
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
      integer(kind=c_intptr_t)                 :: row_dev
      integer(kind=c_intptr_t)                 :: row_group_dev
      integer(kind=c_intptr_t)                 :: hh_tau_dev
      integer(kind=c_intptr_t)                 :: hh_dot_dev
      integer(kind=ik)                         :: row_group_size, unpack_idx

      integer(kind=ik)                         :: n_times
      integer(kind=ik)                         :: top, chunk, this_chunk

#if REALCASE == 1
      real(kind=REAL_DATATYPE), allocatable    :: result_buffer(:,:,:)
      real(kind=REAL_DATATYPE), allocatable    :: bcast_buffer(:,:)
#endif
#if COMPLEXCASE == 1
      complex(kind=COMPLEX_DATATYPE), allocatable :: result_buffer(:,:,:)
      complex(kind=COMPLEX_DATATYPE), allocatable :: bcast_buffer(:,:)
#endif


      integer(kind=ik)                         :: n_off

      integer(kind=ik), allocatable            :: result_send_request(:), result_recv_request(:), limits(:)
      integer(kind=ik), allocatable            :: top_send_request(:), bottom_send_request(:)
      integer(kind=ik), allocatable            :: top_recv_request(:), bottom_recv_request(:)
#ifdef WITH_OPENMP
!      integer(kind=ik), allocatable           :: mpi_statuses(:,:)
#endif

#ifdef WITH_OPENMP
#ifdef WITH_MPI
!      integer(kind=ik)                        :: my_MPI_STATUS_(MPI_STATUS_SIZE)
#endif
#endif

#if COMPLEXCASE == 1
#ifdef WITH_MPI
      integer(kind=ik), external               :: numroc
#endif
      integer(kind=ik)                         :: na_rows, na_cols
#endif

      ! MPI send/recv tags, arbitrary

      integer(kind=ik), parameter              :: bottom_recv_tag = 111
      integer(kind=ik), parameter              :: top_recv_tag    = 222
      integer(kind=ik), parameter              :: result_recv_tag = 333
#ifdef WITH_OPENMP
      integer(kind=ik)                         :: max_threads, my_thread
      integer(kind=ik)                         :: omp_get_max_threads
#endif


      ! Just for measuring the kernel performance
      real(kind=c_double)                      :: kernel_time, kernel_time_recv ! MPI_WTIME always needs double
      ! long integer
      integer(kind=lik)                        :: kernel_flops, kernel_flops_recv



      logical, intent(in)                      :: wantDebug
      logical                                  :: success
255
      integer(kind=ik)                         :: istat, print_flops
256
257
258
259
260
      character(200)                           :: errorMessage
      logical                                  :: successCUDA
#ifndef WITH_MPI
      integer(kind=ik)                         :: j1
#endif
261
      integer(kind=c_intptr_t), parameter      :: size_of_datatype = size_of_&
262
263
264
265
                                                                   &PRECISION&
                                                                   &_&
                                                                   &MATH_DATATYPE

266
      call obj%timer%start("trans_ev_tridi_to_band_&
267
      &MATH_DATATYPE&
Andreas Marek's avatar
Cleanup  
Andreas Marek committed
268
      &" // &
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
      &PRECISION_SUFFIX &
      )

      if (useGPU) then
#if COMPLEXCASE == 1
        n_times = 0
#endif
        unpack_idx = 0
        row_group_size = 0
      endif

      success = .true.
      kernel_time = 0.0
      kernel_flops = 0

#ifdef WITH_OPENMP
      max_threads = 1
      max_threads = omp_get_max_threads()
#endif
288
      if (wantDebug) call obj%timer%start("mpi_communication")
289
290
291
292
      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)
293
      if (wantDebug) call obj%timer%stop("mpi_communication")
294
295
296
297
298
299
300
301
302
303
304

#if COMPLEXCASE == 1
      if (useGPU) then
#ifdef WITH_MPI
        na_rows = numroc(na, nblk, my_prow, 0, np_rows)
        na_cols = numroc(na, nblk, my_pcol, 0, np_cols)
#else
        na_rows = na
        na_cols = na
#endif
      endif
Andreas Marek's avatar
Cleanup  
Andreas Marek committed
305
#endif /* COMPLEXCASE */
306
307
308
309
310

      if (mod(nbw,nblk)/=0) then
        if (my_prow==0 .and. my_pcol==0) then
          if (wantDebug) then
            write(error_unit,*) 'ELPA2_trans_ev_tridi_to_band_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
311
312
      &MATH_DATATYPE&
      &: ERROR: nbw=',nbw,', nblk=',nblk
313
            write(error_unit,*) 'ELPA2_trans_ev_tridi_to_band_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
314
315
      &MATH_DATATYPE&
      &: band backtransform works only for nbw==n*nblk'
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
          endif
          success = .false.
          return
        endif
      endif

      nfact = nbw / nblk


      ! local number of eigenvectors
      l_nev = local_index(nev, my_pcol, np_cols, nblk, -1)

      if (l_nev==0) then
#ifdef WITH_OPENMP
        thread_width = 0
#endif
        stripe_width = 0
        stripe_count = 0
        last_stripe_width = 0

      else ! l_nev

#if WITH_OPENMP
        ! Suggested stripe width is 48 since 48*64 real*8 numbers should fit into
        ! every primary cache
Andreas Marek's avatar
Retab  
Andreas Marek committed
341
  ! Suggested stripe width is 48 - should this be reduced for the complex case ???
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

        if (useGPU) then
          stripe_width = 256 ! Must be a multiple of 4
          stripe_count = (l_nev - 1) / stripe_width + 1
        else ! useGPU
          ! openmp only in non-GPU case
          thread_width = (l_nev-1)/max_threads + 1 ! number of eigenvectors per OMP thread
#if REALCASE == 1
#ifdef DOUBLE_PRECISION_REAL
          stripe_width = 48 ! Must be a multiple of 4
#else
          stripe_width = 96 ! Must be a multiple of 8
#endif
#endif /* REALCASE */

#if COMPLEXCASE == 1
#ifdef DOUBLE_PRECISION_COMPLEX
          stripe_width = 48 ! Must be a multiple of 2
#else
          stripe_width = 48 ! Must be a multiple of 4
#endif
#endif /* COMPLEXCASE */

          stripe_count = (thread_width-1)/stripe_width + 1

          ! Adapt stripe width so that last one doesn't get too small

          stripe_width = (thread_width-1)/stripe_count + 1

#if REALCASE == 1
#ifdef DOUBLE_PRECISION_REAL
373
          if (kernel .eq. ELPA_2STAGE_REAL_AVX512_BLOCK2 .or. &
Andreas Marek's avatar
Retab  
Andreas Marek committed
374
        kernel .eq. ELPA_2STAGE_REAL_AVX512_BLOCK4 .or. &
375
              kernel .eq. ELPA_2STAGE_REAL_AVX512_BLOCK6) then
376
377
378
379

            stripe_width = ((stripe_width+7)/8)*8 ! Must be a multiple of 8 because of AVX-512 memory alignment of 64 bytes
                                                  ! (8 * sizeof(double) == 64)

Andreas Marek's avatar
Retab  
Andreas Marek committed
380
    else
381
            stripe_width = ((stripe_width+3)/4)*4 ! Must be a multiple of 4 because of AVX/SSE memory alignment of 32 bytes
Andreas Marek's avatar
Retab  
Andreas Marek committed
382
                                            ! (4 * sizeof(double) == 32)
383
          endif
384
#else
385
          if (kernel .eq. ELPA_2STAGE_REAL_AVX512_BLOCK2 .or. &
Andreas Marek's avatar
Retab  
Andreas Marek committed
386
        kernel .eq. ELPA_2STAGE_REAL_AVX512_BLOCK4 .or. &
387
              kernel .eq. ELPA_2STAGE_REAL_AVX512_BLOCK6) then
388
389
390


            stripe_width = ((stripe_width+15)/16)*16 ! Must be a multiple of 16 because of AVX-512 memory alignment of 64 bytes
Andreas Marek's avatar
Retab  
Andreas Marek committed
391
                                               ! (16 * sizeof(float) == 64)
392

Andreas Marek's avatar
Retab  
Andreas Marek committed
393
    else
394
            stripe_width = ((stripe_width+7)/8)*8 ! Must be a multiple of 8 because of AVX/SSE memory alignment of 32 bytes
Andreas Marek's avatar
Retab  
Andreas Marek committed
395
                                            ! (8 * sizeof(float) == 32)
396
          endif
397
398
399
400
401
#endif
#endif /* REALCASE */

#if COMPLEXCASE == 1
#ifdef DOUBLE_PRECISION_COMPLEX
402
          if (kernel .eq. ELPA_2STAGE_COMPLEX_AVX512_BLOCK1 .or. &
Andreas Marek's avatar
Retab  
Andreas Marek committed
403
        kernel .eq. ELPA_2STAGE_COMPLEX_AVX512_BLOCK2) then
404

Lorenz Huedepohl's avatar
Lorenz Huedepohl committed
405
            stripe_width = ((stripe_width+7)/8)*8 ! Must be a multiple of 4 because of AVX-512 memory alignment of 64 bytes
Andreas Marek's avatar
Retab  
Andreas Marek committed
406
                                            ! (4 * sizeof(double complex) == 64)
407

Andreas Marek's avatar
Retab  
Andreas Marek committed
408
    else
409
410

            stripe_width = ((stripe_width+3)/4)*4 ! Must be a multiple of 2 because of AVX/SSE memory alignment of 32 bytes
Andreas Marek's avatar
Retab  
Andreas Marek committed
411
412
                                            ! (2 * sizeof(double complex) == 32)
    endif
413
#else
414

415
          if (kernel .eq. ELPA_2STAGE_COMPLEX_AVX512_BLOCK1 .or. &
Andreas Marek's avatar
Retab  
Andreas Marek committed
416
        kernel .eq. ELPA_2STAGE_COMPLEX_AVX512_BLOCK2) then
417
418

            stripe_width = ((stripe_width+7)/8)*8 ! Must be a multiple of 8 because of AVX-512 memory alignment of 64 bytes
Andreas Marek's avatar
Retab  
Andreas Marek committed
419
                                            ! (8 * sizeof(float complex) == 64)
420
421
422

          else
            stripe_width = ((stripe_width+3)/4)*4 ! Must be a multiple of 4 because of AVX/SSE memory alignment of 32 bytes
Andreas Marek's avatar
Retab  
Andreas Marek committed
423
424
                                            ! (4 * sizeof(float complex) == 32)
    endif
425
426
427
#endif
#endif /* COMPLEXCASE */

428
#if REALCASE == 1
429
        last_stripe_width = l_nev - (stripe_count-1)*stripe_width
430
431
432
433
434
#endif
#if COMPLEXCASE == 1
! only needed in no OMP case check thsis
! last_stripe_width = l_nev - (stripe_count-1)*stripe_width
#endif
435
436

        endif ! useGPU
Andreas Marek's avatar
Cleanup  
Andreas Marek committed
437

438
#else /* WITH_OPENMP */
Andreas Marek's avatar
Cleanup  
Andreas Marek committed
439

440
441
        ! Suggested stripe width is 48 since 48*64 real*8 numbers should fit into
        ! every primary cache
Andreas Marek's avatar
Retab  
Andreas Marek committed
442
  ! Suggested stripe width is 48 - should this be reduced for the complex case ???
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


        if (useGPU) then
          stripe_width = 256 ! Must be a multiple of 4
          stripe_count = (l_nev - 1) / stripe_width + 1

        else ! useGPU
#if REALCASE == 1
#ifdef DOUBLE_PRECISION_REAL
          stripe_width = 48 ! Must be a multiple of 4
#else
          stripe_width = 96 ! Must be a multiple of 8
#endif
#endif /* REALCASE */

#if COMPLEXCASE == 1
#ifdef DOUBLE_PRECISION_COMPLEX
          stripe_width = 48 ! Must be a multiple of 2
#else
          stripe_width = 48 ! Must be a multiple of 4
#endif
#endif /* COMPLEXCASE */

          stripe_count = (l_nev-1)/stripe_width + 1

          ! Adapt stripe width so that last one doesn't get too small

          stripe_width = (l_nev-1)/stripe_count + 1

#if REALCASE == 1
#ifdef DOUBLE_PRECISION_REAL
474
          if (kernel .eq. ELPA_2STAGE_REAL_AVX512_BLOCK2 .or. &
Andreas Marek's avatar
Retab  
Andreas Marek committed
475
        kernel .eq. ELPA_2STAGE_REAL_AVX512_BLOCK4 .or. &
476
              kernel .eq. ELPA_2STAGE_REAL_AVX512_BLOCK6) then
477
478
479
480

            stripe_width = ((stripe_width+7)/8)*8 ! Must be a multiple of 8 because of AVX-512 memory alignment of 64 bytes
                                                  ! (8 * sizeof(double) == 64)

Andreas Marek's avatar
Retab  
Andreas Marek committed
481
    else
482
            stripe_width = ((stripe_width+3)/4)*4 ! Must be a multiple of 4 because of AVX/SSE memory alignment of 32 bytes
Andreas Marek's avatar
Retab  
Andreas Marek committed
483
                                            ! (4 * sizeof(double) == 32)
484
          endif
485
#else
486
          if (kernel .eq. ELPA_2STAGE_REAL_AVX512_BLOCK2 .or. &
Andreas Marek's avatar
Retab  
Andreas Marek committed
487
        kernel .eq. ELPA_2STAGE_REAL_AVX512_BLOCK4 .or. &
488
              kernel .eq. ELPA_2STAGE_REAL_AVX512_BLOCK6) then
489
490
491


            stripe_width = ((stripe_width+15)/16)*16 ! Must be a multiple of 16 because of AVX-512 memory alignment of 64 bytes
Andreas Marek's avatar
Retab  
Andreas Marek committed
492
                                               ! (16 * sizeof(float) == 64)
493

Andreas Marek's avatar
Retab  
Andreas Marek committed
494
    else
495
            stripe_width = ((stripe_width+7)/8)*8 ! Must be a multiple of 8 because of AVX/SSE memory alignment of 32 bytes
Andreas Marek's avatar
Retab  
Andreas Marek committed
496
                                            ! (8 * sizeof(float) == 32)
497
          endif
498
499
500
501
502
#endif
#endif /* REALCASE */

#if COMPLEXCASE == 1
#ifdef DOUBLE_PRECISION_COMPLEX
503

504
          if (kernel .eq. ELPA_2STAGE_COMPLEX_AVX512_BLOCK1 .or. &
Andreas Marek's avatar
Retab  
Andreas Marek committed
505
        kernel .eq. ELPA_2STAGE_COMPLEX_AVX512_BLOCK2) then
506

Lorenz Huedepohl's avatar
Lorenz Huedepohl committed
507
            stripe_width = ((stripe_width+7)/8)*8 ! Must be a multiple of 4 because of AVX-512 memory alignment of 64 bytes
Andreas Marek's avatar
Retab  
Andreas Marek committed
508
                                            ! (4 * sizeof(double complex) == 64)
509

Andreas Marek's avatar
Retab  
Andreas Marek committed
510
    else
511
512

            stripe_width = ((stripe_width+3)/4)*4 ! Must be a multiple of 2 because of AVX/SSE memory alignment of 32 bytes
Andreas Marek's avatar
Retab  
Andreas Marek committed
513
514
                                            ! (2 * sizeof(double complex) == 32)
    endif
515
#else
516

517
          if (kernel .eq. ELPA_2STAGE_COMPLEX_AVX512_BLOCK1 .or. &
Andreas Marek's avatar
Retab  
Andreas Marek committed
518
        kernel .eq. ELPA_2STAGE_COMPLEX_AVX512_BLOCK2) then
519

Lorenz Huedepohl's avatar
Lorenz Huedepohl committed
520
            stripe_width = ((stripe_width+15)/16)*16 ! Must be a multiple of 8 because of AVX-512 memory alignment of 64 bytes
Andreas Marek's avatar
Retab  
Andreas Marek committed
521
                                            ! (8 * sizeof(float complex) == 64)
522
523
524

          else
            stripe_width = ((stripe_width+3)/4)*4 ! Must be a multiple of 4 because of AVX/SSE memory alignment of 32 bytes
Andreas Marek's avatar
Retab  
Andreas Marek committed
525
526
                                            ! (4 * sizeof(float complex) == 32)
    endif
527
528
529
530
531
532
533
534
535
536
537
538
539
540
#endif
#endif /* COMPLEXCASE */
        endif ! useGPU

        last_stripe_width = l_nev - (stripe_count-1)*stripe_width

#endif /* WITH_OPENMP */
      endif ! l_nev

      ! Determine the matrix distribution at the beginning

      allocate(limits(0:np_rows), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
        print *,"trans_ev_tridi_to_band_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
541
542
  &MATH_DATATYPE&
  &: error when allocating limits"//errorMessage
Andreas Marek's avatar
Andreas Marek committed
543
        stop 1
544
      endif
545
      call determine_workload(obj,na, nbw, np_rows, limits)
546
547
548
549
550
551

      max_blk_size = maxval(limits(1:np_rows) - limits(0:np_rows-1))

      a_dim2 = max_blk_size + nbw

      if (useGPU) then
552
553
        num =  (stripe_width*a_dim2*stripe_count)* size_of_datatype
        successCUDA = cuda_malloc(aIntern_dev, stripe_width*a_dim2*stripe_count* size_of_datatype)
554
555
556
557
        if (.not.(successCUDA)) then
          print *,"trans_ev_tridi_to_band_&
          &MATH_DATATYPE&
          &: error in cudaMalloc"//errorMessage
Andreas Marek's avatar
Andreas Marek committed
558
          stop 1
559
560
561
562
563
564
565
        endif

        successCUDA = cuda_memset(aIntern_dev , 0, num)
        if (.not.(successCUDA)) then
          print *,"trans_ev_tridi_to_band_&
          &MATH_DATATYPE&
          &: error in cudaMemset"//errorMessage
Andreas Marek's avatar
Andreas Marek committed
566
          stop 1
567
568
        endif

569
        num =  (l_nev)* size_of_datatype
570
571
572
573
574
        successCUDA = cuda_malloc( row_dev,num)
        if (.not.(successCUDA)) then
          print *,"trans_ev_tridi_to_band_&
          &MATH_DATATYPE&
          &: error in cudaMalloc "
Andreas Marek's avatar
Andreas Marek committed
575
          stop 1
576
577
578
579
580
581
582
        endif

        successCUDA = cuda_memset(row_dev , 0, num)
        if (.not.(successCUDA)) then
          print *,"trans_ev_tridi_to_band_&
          &MATH_DATATYPE&
          &: error in cudaMemset "
Andreas Marek's avatar
Andreas Marek committed
583
          stop 1
584
585
586
587
588
589
590
591
        endif

        ! "row_group" and "row_group_dev" are needed for GPU optimizations
        allocate(row_group(l_nev, nblk), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
          print *,"trans_ev_tridi_to_band_&
          &MATH_DATATYPE&
          &: error when allocating row_group"//errorMessage
Andreas Marek's avatar
Andreas Marek committed
592
          stop 1
593
594
595
596
597
598
599
600
601
        endif

#if REALCASE == 1
        row_group(:, :) = CONST_0_0
#endif
#if COMPLEXCASE == 1
        row_group(:, :) = CONST_COMPLEX_0_0
#endif

602
        num =  (l_nev*nblk)* size_of_datatype
603
604
605
606
607
        successCUDA = cuda_malloc(row_group_dev, num)
        if (.not.(successCUDA)) then
          print *,"trans_ev_tridi_to_band_&
          &MATH_DATATYPE&
          &: error in cudaMalloc"//errorMessage
Andreas Marek's avatar
Andreas Marek committed
608
          stop 1
609
610
611
612
613
614
615
        endif

        successCUDA = cuda_memset(row_group_dev , 0, num)
        if (.not.(successCUDA)) then
          print *,"trans_ev_tridi_to_band_&
          &MATH_DATATYPE&
          &: error in cudaMemset"//errorMessage
Andreas Marek's avatar
Andreas Marek committed
616
          stop 1
617
618
619
620
621
622
623
624
625
626
        endif

      else ! GPUs are not used

#if 0
! realcase or complexcase
!DEC$ ATTRIBUTES ALIGN: 64:: aIntern
#endif

#ifdef WITH_OPENMP
627
        if (posix_memalign(aIntern_ptr, 64_c_intptr_t, stripe_width*a_dim2*stripe_count*max_threads*     &
628
629
630
631
632
633
634
635
#if REALCASE == 1
               C_SIZEOF(a_real)) /= 0) then
#endif
#if COMPLEXCASE == 1
               C_SIZEOF(a_complex)) /= 0) then
#endif
          print *,"trans_ev_tridi_to_band_&
          &MATH_DATATYPE&
Andreas Marek's avatar
Cleanup  
Andreas Marek committed
636
          &: error when allocating aIntern"//errorMessage
Andreas Marek's avatar
Andreas Marek committed
637
          stop 1
638
639
640
641
642
643
        endif

        call c_f_pointer(aIntern_ptr, aIntern, [stripe_width,a_dim2,stripe_count,max_threads])
        ! allocate(aIntern(stripe_width,a_dim2,stripe_count,max_threads), stat=istat, errmsg=errorMessage)

        ! aIntern(:,:,:,:) should be set to 0 in a parallel region, not here!
644
645
646

#else /* WITH_OPENMP */

647
        if (posix_memalign(aIntern_ptr, 64_c_intptr_t, stripe_width*a_dim2*stripe_count*  &
648
649
650
651
652
653
654
#if REALCASE == 1
            C_SIZEOF(a_real)) /= 0) then
#endif
#if COMPLEXCASE == 1
            C_SIZEOF(a_complex)) /= 0) then
#endif
          print *,"trans_ev_tridi_to_band_real: error when allocating aIntern"//errorMessage
Andreas Marek's avatar
Andreas Marek committed
655
          stop 1
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
        endif

        call c_f_pointer(aIntern_ptr, aIntern,[stripe_width,a_dim2,stripe_count] )
        !allocate(aIntern(stripe_width,a_dim2,stripe_count), stat=istat, errmsg=errorMessage)

#if REALCASE == 1
        aIntern(:,:,:) = CONST_0_0
#endif
#if COMPLEXCASE == 1
        aIntern(:,:,:) = 0
#endif
#endif /* WITH_OPENMP */
      endif !useGPU

      allocate(row(l_nev), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
        print *,"trans_ev_tridi_to_band_&
        &MATH_DATATYPE&
        &: error when allocating row"//errorMessage
Andreas Marek's avatar
Andreas Marek committed
675
        stop 1
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
      endif

#if REALCASE == 1
      row(:) = CONST_0_0
#endif
#if COMPLEXCASE == 1
      row(:) = 0
#endif


      ! Copy q from a block cyclic distribution into a distribution with contiguous rows,
      ! and transpose the matrix using stripes of given stripe_width for cache blocking.

      ! The peculiar way it is done below is due to the fact that the last row should be
      ! ready first since it is the first one to start below

#ifdef WITH_OPENMP
      ! Please note about the OMP usage below:
      ! This is not for speed, but because we want the matrix a in the memory and
      ! in the cache of the correct thread (if possible)

697
      call obj%timer%start("OpenMP parallel" // PRECISION_SUFFIX)
698
699
700
701
702
703
704
705
706
707
708
      !$omp parallel do private(my_thread), schedule(static, 1)
      do my_thread = 1, max_threads
#if REALCASE == 1
        aIntern(:,:,:,my_thread) = CONST_0_0 ! if possible, do first touch allocation!
#endif
#if COMPLEXCASE == 1
        aIntern(:,:,:,my_thread) = CONST_COMPLEX_0_0 ! if possible, do first touch allocation!
#endif
      enddo
      !$omp end parallel do

709
      call obj%timer%stop("OpenMP parallel" // PRECISION_SUFFIX)
710
711
712
713
714
715
716
717
718
719
720
721
722
#endif /* WITH_OPENMP */

      do ip = np_rows-1, 0, -1
        if (my_prow == ip) then
          ! Receive my rows which have not yet been received
          src_offset = local_index(limits(ip), my_prow, np_rows, nblk, -1)
          do i=limits(ip)+1,limits(ip+1)
            src = mod((i-1)/nblk, np_rows)

            if (src < my_prow) then
#ifdef WITH_OPENMP

#ifdef WITH_MPI
723
              if (wantDebug) call obj%timer%start("mpi_communication")
724
725
726
727
728
729
730
731
              call MPI_Recv(row, l_nev,      &
#if REALCASE == 1
                            MPI_REAL_PRECISION, &
#endif
#if COMPLEXCASE == 1
                            MPI_COMPLEX_EXPLICIT_PRECISION, &
#endif
                            src, 0, mpi_comm_rows, MPI_STATUS_IGNORE, mpierr)
732
              if (wantDebug) call obj%timer%stop("mpi_communication")
733
734
735
736
737
738
739

#else /* WITH_MPI */

!              row(1:l_nev) = row(1:l_nev)

#endif /* WITH_MPI */

740
              call obj%timer%start("OpenMP parallel" // PRECISION_SUFFIX)
741
742
743

!$omp parallel do private(my_thread), schedule(static, 1)
              do my_thread = 1, max_threads
Andreas Marek's avatar
Retab  
Andreas Marek committed
744
          call unpack_row_&
745
746
747
&MATH_DATATYPE&
&_cpu_openmp_&
&PRECISION &
748
                                  (obj,aIntern, row, i-limits(ip), my_thread, stripe_count, &
749
750
                                   thread_width, stripe_width, l_nev)

751
752
753
              enddo
!$omp end parallel do

754
              call obj%timer%stop("OpenMP parallel" // PRECISION_SUFFIX)
755
756
757
758

#else /* WITH_OPENMP */
              if (useGPU) then
                ! An unpacking of the current row group may occur before queuing the next row
Andreas Marek's avatar
Retab  
Andreas Marek committed
759
760
761
762
763
764
    call unpack_and_prepare_row_group_&
    &MATH_DATATYPE&
    &_gpu_&
    &PRECISION &
                              ( &
                              row_group, row_group_dev, aIntern_dev, stripe_count, &
Andreas Marek's avatar
Andreas Marek committed
765
766
767
                                          stripe_width, last_stripe_width, a_dim2, l_nev,&
                                          row_group_size, nblk, unpack_idx, &
                                           i - limits(ip), .false.)
768
#ifdef WITH_MPI
769
                if (wantDebug) call obj%timer%start("mpi_communication")
770
771
772
773
774
775
776
777
                call MPI_Recv(row_group(:, row_group_size), l_nev,     &
#if REALCASE == 1
                              MPI_REAL_PRECISION,    &
#endif
#if COMPLEXCASE == 1
                              MPI_COMPLEX_EXPLICIT_PRECISION, &
#endif
                              src, 0, mpi_comm_rows, MPI_STATUS_IGNORE, mpierr)
778
                if (wantDebug) call obj%timer%stop("mpi_communication")
779
780
781
782
783
784
785

#else /* WITH_MPI */
                row_group(1:l_nev, row_group_size) = row(1:l_nev) ! is this correct?
#endif /* WITH_MPI */

              else ! useGPU
#ifdef WITH_MPI
786
                if (wantDebug) call obj%timer%start("mpi_communication")
787
788
789
790
791
792
793
794
                call MPI_Recv(row, l_nev,         &
#if REALCASE == 1
                              MPI_REAL_PRECISION, &
#endif
#if COMPLEXCASE == 1
                              MPI_COMPLEX_EXPLICIT_PRECISION, &
#endif
                              src, 0, mpi_comm_rows, MPI_STATUS_IGNORE, mpierr)
795
                if (wantDebug) call obj%timer%stop("mpi_communication")
796
797
798
799
800
801
802

#else /* WITH_MPI */

!                row(1:l_nev) = row(1:l_nev)

#endif /* WITH_MPI */

803
                call unpack_row_&
Andreas Marek's avatar
Cleanup  
Andreas Marek committed
804
805
806
                &MATH_DATATYPE&
                &_cpu_&
                &PRECISION &
807
                                (obj,aIntern, row,i-limits(ip), stripe_count, stripe_width, last_stripe_width)
808
809
810
811
812
813
814
815
816
817
              endif ! useGPU
#endif /* WITH_OPENMP */

            elseif (src == my_prow) then

              src_offset = src_offset+1

              if (useGPU) then
#ifndef WITH_OPENMP

Andreas Marek's avatar
Andreas Marek committed
818
819
                 ! An unpacking of the current row group may occur before queuing the next row
                 call unpack_and_prepare_row_group_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
820
821
822
823
     &MATH_DATATYPE&
     &_gpu_&
     &PRECISION &
                  ( &
Andreas Marek's avatar
Andreas Marek committed
824
825
826
827
828
                               row_group, row_group_dev, aIntern_dev, stripe_count, &
                               stripe_width, last_stripe_width, a_dim2, l_nev,&
                               row_group_size, nblk, unpack_idx, &
                               i - limits(ip), .false.)

829
830
831
832
833
834
835
836
#if REALCASE == 1
                row_group(:, row_group_size) = q(src_offset, 1:l_nev)
#endif

#if COMPLEXCASE == 1
                row_group(:, row_group_size) = q(src_offset, 1:l_nev)
#endif

837
#else /* WITH_OPENMP */
Andreas Marek's avatar
Andreas Marek committed
838

839
#if COMPLEXCASE == 1
Andreas Marek's avatar
Andreas Marek committed
840
! why is an cuda call in the openmp region?
Andreas Marek's avatar
Andreas Marek committed
841
                call unpack_and_prepare_row_group_complex_gpu_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
842
843
844
845
    &PRECISION&
    &(row_group, row_group_dev, aIntern_dev, stripe_count, stripe_width, &
      last_stripe_width, a_dim2, l_nev, row_group_size, nblk,      &
      unpack_idx, i - limits(ip),.false.)
846
847
848
849
850
851
852
853
854
                row_group(:, row_group_size) = q(src_offset, 1:l_nev)
#endif

#endif /* not OpenMP */
              else
                row(:) = q(src_offset, 1:l_nev)
              endif

#ifdef WITH_OPENMP
855
              call obj%timer%start("OpenMP parallel" // PRECISION_SUFFIX)
856
857
858

!$omp parallel do private(my_thread), schedule(static, 1)
              do my_thread = 1, max_threads
Andreas Marek's avatar
Retab  
Andreas Marek committed
859
          call unpack_row_&
Andreas Marek's avatar
Cleanup  
Andreas Marek committed
860
861
862
                &MATH_DATATYPE&
                &_cpu_openmp_&
                &PRECISION &
863
                                   (obj,aIntern, row, i-limits(ip), my_thread, stripe_count, thread_width, stripe_width, l_nev)
864

865
866
867
              enddo
!$omp end parallel do

868
              call obj%timer%stop("OpenMP parallel" // PRECISION_SUFFIX)
869
870
871
872
873
874

#else /* WITH_OPENMP */

              if (useGPU) then

              else
Andreas Marek's avatar
Retab  
Andreas Marek committed
875
          call unpack_row_&
Andreas Marek's avatar
Cleanup  
Andreas Marek committed
876
877
878
                &MATH_DATATYPE&
                &_cpu_&
                &PRECISION &
879
                                (obj,aIntern, row,i-limits(ip),  stripe_count, stripe_width, last_stripe_width)
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
              endif

#endif /* WITH_OPENMP */

            endif
          enddo

          ! Send all rows which have not yet been send
          src_offset = 0
          do dst = 0, ip-1
            do i=limits(dst)+1,limits(dst+1)
              if (mod((i-1)/nblk, np_rows) == my_prow) then
                src_offset = src_offset+1
                row(:) = q(src_offset, 1:l_nev)

#ifdef WITH_MPI
896
                if (wantDebug) call obj%timer%start("mpi_communication")
897
898
899
900
901
902
903
904
                call MPI_Send(row, l_nev,        &
#if REALCASE == 1
                              MPI_REAL_PRECISION, &
#endif
#if COMPLEXCASE == 1
                              MPI_COMPLEX_EXPLICIT_PRECISION, &
#endif
                              dst, 0, mpi_comm_rows, mpierr)
905
                if (wantDebug) call obj%timer%stop("mpi_communication")
906
907
908
909
910
911
#endif /* WITH_MPI */
              endif
            enddo
          enddo

        else if (my_prow < ip) then
Andreas Marek's avatar
Cleanup  
Andreas Marek committed
912

913
914
915
916
917
918
919
920
          ! Send all rows going to PE ip
          src_offset = local_index(limits(ip), my_prow, np_rows, nblk, -1)
          do i=limits(ip)+1,limits(ip+1)
            src = mod((i-1)/nblk, np_rows)
            if (src == my_prow) then
              src_offset = src_offset+1
              row(:) = q(src_offset, 1:l_nev)
#ifdef WITH_MPI
921
              if (wantDebug) call obj%timer%start("mpi_communication")
922
923
924
925
926
927
928
929
              call MPI_Send(row, l_nev,        &
#if REALCASE == 1
                            MPI_REAL_PRECISION, &
#endif
#if COMPLEXCASE == 1
                            MPI_COMPLEX_EXPLICIT_PRECISION, &
#endif
                            ip, 0, mpi_comm_rows, mpierr)
930
              if (wantDebug) call obj%timer%stop("mpi_communication")
931
932
933
934
935
936
937
938
939
940
941
#endif /* WITH_MPI */
            endif
          enddo

          ! Receive all rows from PE ip
          do i=limits(my_prow)+1,limits(my_prow+1)
            src = mod((i-1)/nblk, np_rows)
            if (src == ip) then
#ifdef WITH_OPENMP

#ifdef WITH_MPI
942
              if (wantDebug) call obj%timer%start("mpi_communication")
943
944
945
946
947
948
949
950
              call MPI_Recv(row, l_nev,     &
#if REALCASE == 1
                            MPI_REAL_PRECISION, &
#endif
#if COMPLEXCASE == 1
                            MPI_COMPLEX_EXPLICIT_PRECISION, &
#endif
                            src, 0, mpi_comm_rows, MPI_STATUS_IGNORE, mpierr)
951
              if (wantDebug) call obj%timer%stop("mpi_communication")
952
953
954
955
956
957
#else /* WITH_MPI */

!              row(1:l_nev) = row(1:l_nev)

#endif /* WITH_MPI */

958
              call obj%timer%start("OpenMP parallel" // PRECISION_SUFFIX)
959
960
!$omp parallel do private(my_thread), schedule(static, 1)
              do my_thread = 1, max_threads
Andreas Marek's avatar
Retab  
Andreas Marek committed
961
          call unpack_row_&
Andreas Marek's avatar
Cleanup  
Andreas Marek committed
962
963
964
                &MATH_DATATYPE&
                &_cpu_openmp_&
                &PRECISION &
965
                                 (obj,aIntern, row, i-limits(my_prow), my_thread, stripe_count, thread_width, stripe_width, l_nev)
966
967
              enddo
!$omp end parallel do
968
              call obj%timer%stop("OpenMP parallel" // PRECISION_SUFFIX)
969
970
971
972

#else /* WITH_OPENMP */
              if (useGPU) then
                ! An unpacking of the current row group may occur before queuing the next row
Andreas Marek's avatar
Andreas Marek committed
973
                call unpack_and_prepare_row_group_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
974
975
976
977
    &MATH_DATATYPE&
    &_gpu_&
    &PRECISION&
    &( &
Andreas Marek's avatar
Andreas Marek committed
978
979
980
981
                  row_group, row_group_dev, aIntern_dev, stripe_count,  &
                  stripe_width, last_stripe_width, a_dim2, l_nev,       &
                  row_group_size, nblk, unpack_idx,                     &
                  i - limits(my_prow), .false.)
982
983

#ifdef WITH_MPI
984
               if (wantDebug) call obj%timer%start("mpi_communication")
985
986
987
988
989
990
991
992
               call MPI_Recv(row_group(:, row_group_size), l_nev,   &
#if REALCASE == 1
                             MPI_REAL_PRECISION,    &
#endif
#if COMPLEXCASE == 1
                             MPI_COMPLEX_EXPLICIT_PRECISION, &
#endif
                             src, 0, mpi_comm_rows, MPI_STATUS_IGNORE, mpierr)
993
               if (wantDebug) call obj%timer%stop("mpi_communication")
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
#else /* WITH_MPI */

#if REALCASE == 1
               row_group(1:l_nev,row_group_size) = row(1:l_nev) ! is this correct ?
#endif
#if COMPLEXCASE == 1
! todo: what of this is correct? Was different for double/single precision
#ifdef DOUBLE_PRECISION_COMPLEX
                row_group(1:l_nev,row_group_size) = row(1:l_nev) ! is this correct ?
#else
                row_group(1:l_nev,row_group_size) = row_group(1:l_nev,row_group_size) ! is this correct
#endif
#endif
#endif /* WITH_MPI */

              else ! useGPU
#ifdef WITH_MPI
1011
                if (wantDebug) call obj%timer%start("mpi_communication")
1012
1013
1014
1015
1016
1017
1018
1019
                call MPI_Recv(row, l_nev,    &
#if REALCASE == 1
                              MPI_REAL_PRECISION,    &
#endif
#if COMPLEXCASE == 1
                              MPI_COMPLEX_EXPLICIT_PRECISION, &
#endif
                              src, 0, mpi_comm_rows, MPI_STATUS_IGNORE, mpierr)
1020
                if (wantDebug) call obj%timer%stop("mpi_communication")
1021
1022
1023
1024
1025
#else /* WITH_MPI */

!                row(1:l_nev) = row(1:l_nev)

#endif
1026
                call unpack_row_&
Andreas Marek's avatar
Cleanup  
Andreas Marek committed
1027
1028
1029
                &MATH_DATATYPE&
                &_cpu_&
                &PRECISION &
1030
                                (obj,aIntern, row,i-limits(my_prow), stripe_count, stripe_width, last_stripe_width)
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
              endif ! useGPU

#endif /* WITH_OPENMP */

            endif
          enddo
        endif
      enddo

      if (useGPU) then
        ! Force an unpacking of all remaining rows that haven't been unpacked yet
Andreas Marek's avatar
Andreas Marek committed
1042
        call unpack_and_prepare_row_group_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
1043
1044
1045
1046
  &MATH_DATATYPE&
  &_gpu_&
  &PRECISION&
  &( &
Andreas Marek's avatar
Andreas Marek committed
1047
1048
1049
1050
1051
          row_group, row_group_dev, aIntern_dev, stripe_count, &
          stripe_width, last_stripe_width, &
          a_dim2, l_nev, row_group_size, nblk, unpack_idx,     &
          -1, .true.)

1052
1053
1054
1055
1056
1057
        successCUDA = cuda_devicesynchronize()

         if (.not.(successCUDA)) then
           print *,"trans_ev_tridi_to_band_&
           &MATH_DATATYPE&
           &: error in cudaDeviceSynchronize"//errorMessage
Andreas Marek's avatar
Andreas Marek committed
1058
           stop 1
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
         endif
      endif

      ! Set up result buffer queue

      num_result_blocks = ((na-1)/nblk + np_rows - my_prow) / np_rows

      num_result_buffers = 4*nfact
      allocate(result_buffer(l_nev,nblk,num_result_buffers), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
        print *,"trans_ev_tridi_to_band_&
        &MATH_DATATYPE&
        &: error when allocating result_buffer"//errorMessage
Andreas Marek's avatar
Andreas Marek committed
1072
        stop 1
1073
1074
1075
1076
1077
1078
1079
      endif

      allocate(result_send_request(num_result_buffers), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
        print *,"trans_ev_tridi_to_band_&
        &MATH_DATATYPE&
        &: error when allocating result_send_request"//errorMessage
Andreas Marek's avatar
Andreas Marek committed
1080
        stop 1
1081
1082
1083
1084
1085
1086
1087
      endif

      allocate(result_recv_request(num_result_buffers), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
        print *,"trans_ev_tridi_to_band_&
        &MATH_DATATYPE&
        &: error when allocating result_recv_request"//errorMessage
Andreas Marek's avatar
Andreas Marek committed
1088
        stop 1
1089
1090
1091
1092
1093
1094
1095
1096
1097
      endif

#ifdef WITH_MPI
      result_send_request(:) = MPI_REQUEST_NULL
      result_recv_request(:) = MPI_REQUEST_NULL
#endif

      ! Queue up buffers
#ifdef WITH_MPI
1098
      if (wantDebug) call obj%timer%start("mpi_communication")
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111

      if (my_prow > 0 .and. l_nev>0) then ! note: row 0 always sends
        do j = 1, min(num_result_buffers, num_result_blocks)
          call MPI_Irecv(result_buffer(1,1,j), l_nev*nblk,    &
#if REALCASE == 1
                         MPI_REAL_PRECISION,     &
#endif
#if COMPLEXCASE == 1
                         MPI_COMPLEX_EXPLICIT_PRECISION, &
#endif
                         0, result_recv_tag, mpi_comm_rows, result_recv_request(j), mpierr)
        enddo
      endif
1112
      if (wantDebug) call obj%timer%stop("mpi_communication")
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
#else /* WITH_MPI */

      ! carefull the "recv" has to be done at the corresponding wait or send
      ! result_buffer(1: l_nev*nblk,1,j) =result_buffer(1:l_nev*nblk,1,nbuf)

#endif /* WITH_MPI */

      num_bufs_recvd = 0 ! No buffers received yet

      ! Initialize top/bottom requests

      allocate(top_send_request(stripe_count), stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
         print *,"trans_ev_tridi_to_band_&
         &MPI_DATATYPE&
         &: error when allocating top_send_request"//errorMessage
Andreas Marek's avatar
Andreas Marek committed
1129
         stop 1
1130
1131
1132
1133
1134
1135
1136
       endif

      allocate(top_recv_request(stripe_count), stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
         print *,"trans_ev_tridi_to_band_&
         &MATH_DATATYPE&
         &: error when allocating top_recv_request"//errorMessage
Andreas Marek's avatar
Andreas Marek committed
1137
         stop 1
1138
1139
1140
1141
1142
1143
1144
       endif

      allocate(bottom_send_request(stripe_count), stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
         print *,"trans_ev_tridi_to_band_&
         &MATH_DATATYPE&
         &: error when allocating bottom_send_request"//errorMessage
Andreas Marek's avatar
Andreas Marek committed
1145
         stop 1
1146
1147
1148
1149
1150
1151
1152
       endif

      allocate(bottom_recv_request(stripe_count), stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
         print *,"trans_ev_tridi_to_band_&
         &MATH_DATATYPE&
         &: error when allocating bottom_recv_request"//errorMessage
Andreas Marek's avatar
Andreas Marek committed
1153
         stop 1
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
       endif

#ifdef WITH_MPI
      top_send_request(:) = MPI_REQUEST_NULL
      top_recv_request(:) = MPI_REQUEST_NULL
      bottom_send_request(:) = MPI_REQUEST_NULL
      bottom_recv_request(:) = MPI_REQUEST_NULL
#endif

#ifdef WITH_OPENMP
      allocate(top_border_send_buffer(stripe_width*nbw*max_threads, stripe_count), stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
         print *,"trans_ev_tridi_to_band_&
         &MATH_DATATYPE&
         &: error when allocating top_border_send_buffer"//errorMessage
Andreas Marek's avatar
Andreas Marek committed
1169
         stop 1
1170
1171
1172
1173
1174
1175
1176
       endif

      allocate(top_border_recv_buffer(stripe_width*nbw*max_threads, stripe_count), stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
         print *,"trans_ev_tridi_to_band_&
         &MATH_DATATYPE&
         &: error when allocating top_border_recv_buffer"//errorMessage
Andreas Marek's avatar
Andreas Marek committed
1177
         stop 1
1178
1179
1180
1181
1182
1183
1184
       endif

      allocate(bottom_border_send_buffer(stripe_width*nbw*max_threads, stripe_count), stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
         print *,"trans_ev_tridi_to_band_&
         &MATH_DATATYPE&
         &: error when allocating bottom_border_send_buffer"//errorMessage
Andreas Marek's avatar
Andreas Marek committed
1185
         stop 1
1186
1187
1188
1189
1190
1191
1192
       endif

      allocate(bottom_border_recv_buffer(stripe_width*nbw*max_threads, stripe_count), stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
         print *,"trans_ev_tridi_to_band_&
         &MATH_DATATYPE&
         &: error when allocating bottom_border_recv_buffer"//errorMessage
Andreas Marek's avatar
Andreas Marek committed
1193
         stop 1
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
       endif

#if REALCASE == 1
      top_border_send_buffer(:,:) = CONST_0_0
      top_border_recv_buffer(:,:) = CONST_0_0
      bottom_border_send_buffer(:,:) = CONST_0_0
      bottom_border_recv_buffer(:,:) = CONST_0_0
#endif
#if COMPLEXCASE == 1
      top_border_send_buffer(:,:) = CONST_COMPLEX_0_0
      top_border_recv_buffer(:,:) = CONST_COMPLEX_0_0
      bottom_border_send_buffer(:,:) = CONST_COMPLEX_0_0
      bottom_border_recv_buffer(:,:) = CONST_COMPLEX_0_0
#endif
      ! Initialize broadcast buffer

#else /* WITH_OPENMP */

       allocate(top_border_send_buffer(stripe_width, nbw, stripe_count), stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
         print *,"trans_ev_tridi_to_band_&
         &MATH_DATATYPE&
         &: error when allocating top_border_send_bufer"//errorMessage
Andreas Marek's avatar
Andreas Marek committed
1217
         stop 1
1218
1219
1220
1221
1222
1223
1224