elpa2_trans_ev_tridi_to_band_template.F90 98 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
#include "../general/precision_kinds.F90"
101
      class(elpa_abstract_impl_t), intent(inout) :: obj
Andreas Marek's avatar
Andreas Marek committed
102
      logical, intent(in)                        :: useGPU
103

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

#ifdef USE_ASSUMED_SIZE
108
      MATH_DATATYPE(kind=rck)                   :: q(ldq,*)
109
#else
110
      MATH_DATATYPE(kind=rck)                   :: q(ldq,matrixCols)
111
112
#endif

113
      MATH_DATATYPE(kind=rck), intent(in)       :: hh_trans(:,:)
Andreas Marek's avatar
Andreas Marek committed
114
      integer(kind=c_intptr_t)                   :: q_dev
Andreas Marek's avatar
Andreas Marek committed
115

Andreas Marek's avatar
Andreas Marek committed
116
117
      integer(kind=ik)                           :: np_rows, my_prow, np_cols, my_pcol
      integer(kind=ik)                           :: tmp
118

Andreas Marek's avatar
Andreas Marek committed
119
120
121
122
123
      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
124
#ifdef WITH_OPENMP
Andreas Marek's avatar
Andreas Marek committed
125
      integer(kind=ik)                           :: thread_width, csw, b_off, b_len
126
#endif
Andreas Marek's avatar
Andreas Marek committed
127
128
129
      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
130

Andreas Marek's avatar
Andreas Marek committed
131
      logical                                    :: flag
132
#ifdef WITH_OPENMP
133
      MATH_DATATYPE(kind=rck), pointer          :: aIntern(:,:,:,:)
134
#else
135
      MATH_DATATYPE(kind=rck), pointer          :: aIntern(:,:,:)
136
#endif
137
      MATH_DATATYPE(kind=rck)                   :: a_var
138

Andreas Marek's avatar
Andreas Marek committed
139
      type(c_ptr)                                :: aIntern_ptr
140

141
142
      MATH_DATATYPE(kind=rck)   , allocatable   :: row(:)
      MATH_DATATYPE(kind=rck)   , allocatable   :: row_group(:,:)
143
144

#ifdef WITH_OPENMP
145
146
      MATH_DATATYPE(kind=rck), allocatable    :: top_border_send_buffer(:,:), top_border_recv_buffer(:,:)
      MATH_DATATYPE(kind=rck), allocatable    :: bottom_border_send_buffer(:,:), bottom_border_recv_buffer(:,:)
147
#else
148
149
      MATH_DATATYPE(kind=rck), allocatable    :: top_border_send_buffer(:,:,:), top_border_recv_buffer(:,:,:)
      MATH_DATATYPE(kind=rck), allocatable    :: bottom_border_send_buffer(:,:,:), bottom_border_recv_buffer(:,:,:)
150
151
152
153
#endif

      integer(kind=c_intptr_t)                 :: aIntern_dev
      integer(kind=c_intptr_t)                 :: bcast_buffer_dev
154
155
      integer(kind=c_intptr_t)                   :: num
      integer(kind=c_intptr_t)                   :: dev_offset, dev_offset_1, dev_offset_2
156
157
158
159
160
161
162
163
164
      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

165
166
      MATH_DATATYPE(kind=rck), allocatable    :: result_buffer(:,:,:)
      MATH_DATATYPE(kind=rck), allocatable    :: bcast_buffer(:,:)
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
192
193
194
195
196
197
198
199
200
201
202

      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

      ! 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
203
      integer(kind=ik)                         :: istat, print_flops
204
205
206
207
208
      character(200)                           :: errorMessage
      logical                                  :: successCUDA
#ifndef WITH_MPI
      integer(kind=ik)                         :: j1
#endif
209
      integer(kind=c_intptr_t), parameter      :: size_of_datatype = size_of_&
210
211
212
213
                                                                   &PRECISION&
                                                                   &_&
                                                                   &MATH_DATATYPE

214
      call obj%timer%start("trans_ev_tridi_to_band_&
215
      &MATH_DATATYPE&
Andreas Marek's avatar
Cleanup  
Andreas Marek committed
216
      &" // &
217
218
219
      &PRECISION_SUFFIX &
      )

220
      n_times = 0
221
222
223
224
225
226
227
228
229
230
231
232
233
      if (useGPU) then
        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
234
      if (wantDebug) call obj%timer%start("mpi_communication")
235
236
237
238
      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)
239
      if (wantDebug) call obj%timer%stop("mpi_communication")
240
241
242
243
244

      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_&
245
246
                                &MATH_DATATYPE&
                                &: ERROR: nbw=',nbw,', nblk=',nblk
247
            write(error_unit,*) 'ELPA2_trans_ev_tridi_to_band_&
248
249
                                &MATH_DATATYPE&
                                &: band backtransform works only for nbw==n*nblk'
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
          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
275
        ! Suggested stripe width is 48 - should this be reduced for the complex case ???
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306

        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
307
          if (kernel .eq. ELPA_2STAGE_REAL_AVX512_BLOCK2 .or. &
308
              kernel .eq. ELPA_2STAGE_REAL_AVX512_BLOCK4 .or. &
309
              kernel .eq. ELPA_2STAGE_REAL_AVX512_BLOCK6) then
310
311
312
313

            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
314
    else
315
            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
316
                                            ! (4 * sizeof(double) == 32)
317
          endif
318
#else
319
          if (kernel .eq. ELPA_2STAGE_REAL_AVX512_BLOCK2 .or. &
320
              kernel .eq. ELPA_2STAGE_REAL_AVX512_BLOCK4 .or. &
321
              kernel .eq. ELPA_2STAGE_REAL_AVX512_BLOCK6) then
322
323
324


            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
325
                                               ! (16 * sizeof(float) == 64)
326

Andreas Marek's avatar
Retab  
Andreas Marek committed
327
    else
328
            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
329
                                            ! (8 * sizeof(float) == 32)
330
          endif
331
332
333
334
335
#endif
#endif /* REALCASE */

#if COMPLEXCASE == 1
#ifdef DOUBLE_PRECISION_COMPLEX
336
          if (kernel .eq. ELPA_2STAGE_COMPLEX_AVX512_BLOCK1 .or. &
337
              kernel .eq. ELPA_2STAGE_COMPLEX_AVX512_BLOCK2) then
338

Lorenz Huedepohl's avatar
Lorenz Huedepohl committed
339
            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
340
                                            ! (4 * sizeof(double complex) == 64)
341

Andreas Marek's avatar
Retab  
Andreas Marek committed
342
    else
343
344

            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
345
346
                                            ! (2 * sizeof(double complex) == 32)
    endif
347
#else
348

349
          if (kernel .eq. ELPA_2STAGE_COMPLEX_AVX512_BLOCK1 .or. &
350
              kernel .eq. ELPA_2STAGE_COMPLEX_AVX512_BLOCK2) then
351
352

            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
353
                                            ! (8 * sizeof(float complex) == 64)
354
355
356

          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
357
358
                                            ! (4 * sizeof(float complex) == 32)
    endif
359
360
361
#endif
#endif /* COMPLEXCASE */

362
#if REALCASE == 1
363
        last_stripe_width = l_nev - (stripe_count-1)*stripe_width
364
365
366
367
368
#endif
#if COMPLEXCASE == 1
! only needed in no OMP case check thsis
! last_stripe_width = l_nev - (stripe_count-1)*stripe_width
#endif
369
370

        endif ! useGPU
Andreas Marek's avatar
Cleanup  
Andreas Marek committed
371

372
#else /* WITH_OPENMP */
Andreas Marek's avatar
Cleanup  
Andreas Marek committed
373

374
375
        ! Suggested stripe width is 48 since 48*64 real*8 numbers should fit into
        ! every primary cache
376
        ! Suggested stripe width is 48 - should this be reduced for the complex case ???
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406

        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
407
          if (kernel .eq. ELPA_2STAGE_REAL_AVX512_BLOCK2 .or. &
408
              kernel .eq. ELPA_2STAGE_REAL_AVX512_BLOCK4 .or. &
409
              kernel .eq. ELPA_2STAGE_REAL_AVX512_BLOCK6) then
410
411
412
413

            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
414
    else
415
            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
416
                                            ! (4 * sizeof(double) == 32)
417
          endif
418
#else
419
          if (kernel .eq. ELPA_2STAGE_REAL_AVX512_BLOCK2 .or. &
420
              kernel .eq. ELPA_2STAGE_REAL_AVX512_BLOCK4 .or. &
421
              kernel .eq. ELPA_2STAGE_REAL_AVX512_BLOCK6) then
422
423
424


            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
425
                                               ! (16 * sizeof(float) == 64)
426

Andreas Marek's avatar
Retab  
Andreas Marek committed
427
    else
428
            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
429
                                            ! (8 * sizeof(float) == 32)
430
          endif
431
432
433
434
435
#endif
#endif /* REALCASE */

#if COMPLEXCASE == 1
#ifdef DOUBLE_PRECISION_COMPLEX
436

437
          if (kernel .eq. ELPA_2STAGE_COMPLEX_AVX512_BLOCK1 .or. &
438
              kernel .eq. ELPA_2STAGE_COMPLEX_AVX512_BLOCK2) then
439

Lorenz Huedepohl's avatar
Lorenz Huedepohl committed
440
            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
441
                                            ! (4 * sizeof(double complex) == 64)
442

Andreas Marek's avatar
Retab  
Andreas Marek committed
443
    else
444
445

            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
446
447
                                            ! (2 * sizeof(double complex) == 32)
    endif
448
#else
449

450
          if (kernel .eq. ELPA_2STAGE_COMPLEX_AVX512_BLOCK1 .or. &
451
              kernel .eq. ELPA_2STAGE_COMPLEX_AVX512_BLOCK2) then
452

Lorenz Huedepohl's avatar
Lorenz Huedepohl committed
453
            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
454
                                            ! (8 * sizeof(float complex) == 64)
455
456
457

          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
458
459
                                            ! (4 * sizeof(float complex) == 32)
    endif
460
461
462
463
464
465
466
467
468
469
470
471
472
473
#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_&
474
475
                &MATH_DATATYPE&
                &: error when allocating limits"//errorMessage
Andreas Marek's avatar
Andreas Marek committed
476
        stop 1
477
      endif
478
      call determine_workload(obj,na, nbw, np_rows, limits)
479
480
481
482
483
484

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

      a_dim2 = max_blk_size + nbw

      if (useGPU) then
485
486
        num =  (stripe_width*a_dim2*stripe_count)* size_of_datatype
        successCUDA = cuda_malloc(aIntern_dev, stripe_width*a_dim2*stripe_count* size_of_datatype)
487
488
489
490
        if (.not.(successCUDA)) then
          print *,"trans_ev_tridi_to_band_&
          &MATH_DATATYPE&
          &: error in cudaMalloc"//errorMessage
Andreas Marek's avatar
Andreas Marek committed
491
          stop 1
492
493
494
495
496
497
498
        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
499
          stop 1
500
501
        endif

502
        num =  (l_nev)* size_of_datatype
503
504
505
506
507
        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
508
          stop 1
509
510
511
512
513
514
515
        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
516
          stop 1
517
518
519
520
521
522
523
524
        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
525
          stop 1
526
527
528
529
530
531
532
533
534
        endif

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

535
        num =  (l_nev*nblk)* size_of_datatype
536
537
538
539
540
        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
541
          stop 1
542
543
544
545
546
547
548
        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
549
          stop 1
550
551
552
553
554
555
556
557
558
559
        endif

      else ! GPUs are not used

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

#ifdef WITH_OPENMP
560
        if (posix_memalign(aIntern_ptr, 64_c_intptr_t, stripe_width*a_dim2*stripe_count*max_threads*     &
561
#if REALCASE == 1
562
               C_SIZEOF(a_var)) /= 0) then
563
564
#endif
#if COMPLEXCASE == 1
565
               C_SIZEOF(a_var)) /= 0) then
566
567
568
#endif
          print *,"trans_ev_tridi_to_band_&
          &MATH_DATATYPE&
Andreas Marek's avatar
Cleanup  
Andreas Marek committed
569
          &: error when allocating aIntern"//errorMessage
Andreas Marek's avatar
Andreas Marek committed
570
          stop 1
571
572
573
574
575
576
        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!
577
578
579

#else /* WITH_OPENMP */

580
        if (posix_memalign(aIntern_ptr, 64_c_intptr_t, stripe_width*a_dim2*stripe_count*  &
581
#if REALCASE == 1
582
            C_SIZEOF(a_var)) /= 0) then
583
584
#endif
#if COMPLEXCASE == 1
585
            C_SIZEOF(a_var)) /= 0) then
586
587
#endif
          print *,"trans_ev_tridi_to_band_real: error when allocating aIntern"//errorMessage
Andreas Marek's avatar
Andreas Marek committed
588
          stop 1
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
        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
608
        stop 1
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
      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)

630
      call obj%timer%start("OpenMP parallel" // PRECISION_SUFFIX)
631
632
633
634
635
636
637
638
639
640
641
      !$omp parallel do private(my_thread), schedule(static, 1)
      do my_thread = 1, max_threads
#if REALCASE == 1
        aIntern(:,:,:,my_thread) = CONST_0_0 ! if possible, do first touch allocation!
#endif
#if COMPLEXCASE == 1
        aIntern(:,:,:,my_thread) = CONST_COMPLEX_0_0 ! if possible, do first touch allocation!
#endif
      enddo
      !$omp end parallel do

642
      call obj%timer%stop("OpenMP parallel" // PRECISION_SUFFIX)
643
644
645
646
647
648
649
650
651
652
653
654
655
#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
656
              if (wantDebug) call obj%timer%start("mpi_communication")
657
658
659
660
661
662
663
664
              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)
665
              if (wantDebug) call obj%timer%stop("mpi_communication")
666
667
668
669
670
671
672

#else /* WITH_MPI */

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

#endif /* WITH_MPI */

673
              call obj%timer%start("OpenMP parallel" // PRECISION_SUFFIX)
674
675
676

!$omp parallel do private(my_thread), schedule(static, 1)
              do my_thread = 1, max_threads
677
678
679
680
                call unpack_row_&
                     &MATH_DATATYPE&
                     &_cpu_openmp_&
                     &PRECISION &
681
                                  (obj,aIntern, row, i-limits(ip), my_thread, stripe_count, &
682
683
                                   thread_width, stripe_width, l_nev)

684
685
686
              enddo
!$omp end parallel do

687
              call obj%timer%stop("OpenMP parallel" // PRECISION_SUFFIX)
688
689
690
691

#else /* WITH_OPENMP */
              if (useGPU) then
                ! An unpacking of the current row group may occur before queuing the next row
692
693
694
695
                call unpack_and_prepare_row_group_&
                &MATH_DATATYPE&
                &_gpu_&
                &PRECISION &
Andreas Marek's avatar
Retab  
Andreas Marek committed
696
697
                              ( &
                              row_group, row_group_dev, aIntern_dev, stripe_count, &
Andreas Marek's avatar
Andreas Marek committed
698
699
700
                                          stripe_width, last_stripe_width, a_dim2, l_nev,&
                                          row_group_size, nblk, unpack_idx, &
                                           i - limits(ip), .false.)
701
#ifdef WITH_MPI
702
                if (wantDebug) call obj%timer%start("mpi_communication")
703
704
705
706
707
708
709
710
                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)
711
                if (wantDebug) call obj%timer%stop("mpi_communication")
712
713
714
715
716
717
718

#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
719
                if (wantDebug) call obj%timer%start("mpi_communication")
720
721
722
723
724
725
726
727
                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)
728
                if (wantDebug) call obj%timer%stop("mpi_communication")
729
730
731
732
733
734
735

#else /* WITH_MPI */

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

#endif /* WITH_MPI */

736
                call unpack_row_&
737
738
739
                     &MATH_DATATYPE&
                     &_cpu_&
                     &PRECISION &
740
                                (obj,aIntern, row,i-limits(ip), stripe_count, stripe_width, last_stripe_width)
741
742
743
744
745
746
747
748
749
750
              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
751
752
                 ! An unpacking of the current row group may occur before queuing the next row
                 call unpack_and_prepare_row_group_&
753
754
755
                      &MATH_DATATYPE&
                      &_gpu_&
                      &PRECISION &
Andreas Marek's avatar
Retab  
Andreas Marek committed
756
                  ( &
Andreas Marek's avatar
Andreas Marek committed
757
758
759
760
761
                               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.)

762
763
764
765
766
767
768
769
#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

770
#else /* WITH_OPENMP */
Andreas Marek's avatar
Andreas Marek committed
771

772
#if COMPLEXCASE == 1
Andreas Marek's avatar
Andreas Marek committed
773
! why is an cuda call in the openmp region?
Andreas Marek's avatar
Andreas Marek committed
774
                call unpack_and_prepare_row_group_complex_gpu_&
775
776
777
778
779
                     &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.)
                      row_group(:, row_group_size) = q(src_offset, 1:l_nev)
780
781
782
783
784
785
786
787
#endif

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

#ifdef WITH_OPENMP
788
              call obj%timer%start("OpenMP parallel" // PRECISION_SUFFIX)
789
790
791

!$omp parallel do private(my_thread), schedule(static, 1)
              do my_thread = 1, max_threads
792
793
794
795
                call unpack_row_&
                     &MATH_DATATYPE&
                     &_cpu_openmp_&
                     &PRECISION &
796
                                   (obj,aIntern, row, i-limits(ip), my_thread, stripe_count, thread_width, stripe_width, l_nev)
797

798
799
800
              enddo
!$omp end parallel do

801
              call obj%timer%stop("OpenMP parallel" // PRECISION_SUFFIX)
802
803
804
805
806
807

#else /* WITH_OPENMP */

              if (useGPU) then

              else
808
809
810
811
                call unpack_row_&
                     &MATH_DATATYPE&
                     &_cpu_&
                     &PRECISION &
812
                                (obj,aIntern, row,i-limits(ip),  stripe_count, stripe_width, last_stripe_width)
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
              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
829
                if (wantDebug) call obj%timer%start("mpi_communication")
830
831
832
833
834
835
836
837
                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)
838
                if (wantDebug) call obj%timer%stop("mpi_communication")
839
840
841
842
843
844
#endif /* WITH_MPI */
              endif
            enddo
          enddo

        else if (my_prow < ip) then
Andreas Marek's avatar
Cleanup  
Andreas Marek committed
845

846
847
848
849
850
851
852
853
          ! 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
854
              if (wantDebug) call obj%timer%start("mpi_communication")
855
856
857
858
859
860
861
862
              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)
863
              if (wantDebug) call obj%timer%stop("mpi_communication")
864
865
866
867
868
869
870
871
872
873
874
#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
875
              if (wantDebug) call obj%timer%start("mpi_communication")
876
877
878
879
880
881
882
883
              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)
884
              if (wantDebug) call obj%timer%stop("mpi_communication")
885
886
887
888
889
890
#else /* WITH_MPI */

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

#endif /* WITH_MPI */

891
              call obj%timer%start("OpenMP parallel" // PRECISION_SUFFIX)
892
893
!$omp parallel do private(my_thread), schedule(static, 1)
              do my_thread = 1, max_threads
894
895
896
897
                call unpack_row_&
                     &MATH_DATATYPE&
                     &_cpu_openmp_&
                     &PRECISION &
898
                                 (obj,aIntern, row, i-limits(my_prow), my_thread, stripe_count, thread_width, stripe_width, l_nev)
899
900
              enddo
!$omp end parallel do
901
              call obj%timer%stop("OpenMP parallel" // PRECISION_SUFFIX)
902
903
904
905

#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
906
                call unpack_and_prepare_row_group_&
907
908
909
910
                     &MATH_DATATYPE&
                     &_gpu_&
                     &PRECISION&
                     &( &
Andreas Marek's avatar
Andreas Marek committed
911
912
913
914
                  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.)
915
916

#ifdef WITH_MPI
917
               if (wantDebug) call obj%timer%start("mpi_communication")
918
919
920
921
922
923
924
925
               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)
926
               if (wantDebug) call obj%timer%stop("mpi_communication")
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
#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
944
                if (wantDebug) call obj%timer%start("mpi_communication")
945
946
947
948
949
950
951
952
                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)
953
                if (wantDebug) call obj%timer%stop("mpi_communication")
954
955
956
957
958
#else /* WITH_MPI */

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

#endif
959
                call unpack_row_&
960
961
962
                     &MATH_DATATYPE&
                     &_cpu_&
                     &PRECISION &
963
                                (obj,aIntern, row,i-limits(my_prow), stripe_count, stripe_width, last_stripe_width)
964
965
966
967
968
969
970
971
972
973
974
              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
975
        call unpack_and_prepare_row_group_&
976
977
978
979
             &MATH_DATATYPE&
             &_gpu_&
             &PRECISION&
             &( &
Andreas Marek's avatar
Andreas Marek committed
980
981
982
983
984
          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.)

985
986
987
988
989
990
        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
991
           stop 1
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
         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
1005
        stop 1
1006
1007
1008
1009
1010
1011
1012
      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
1013
        stop 1
1014
1015
1016
1017
1018
1019
1020
      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
1021
        stop 1
1022
1023
1024
1025
1026
1027
1028
1029
1030
      endif

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

      ! Queue up buffers
#ifdef WITH_MPI
1031
      if (wantDebug) call obj%timer%start("mpi_communication")
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044

      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
1045
      if (wantDebug) call obj%timer%stop("mpi_communication")
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
#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
1062
         stop 1
1063
1064
1065
1066
1067
1068
1069
       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
1070
         stop 1
1071
1072
1073
1074
1075
1076
1077
       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
1078
         stop 1
1079
1080
1081
1082
1083
1084
1085
       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
1086
         stop 1
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
       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
1102
         stop 1
1103
1104
1105
1106
1107
1108
1109
       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
1110
         stop 1
1111
1112
1113
1114
1115
1116
1117
       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
1118
         stop 1
1119
1120
1121
1122
1123
1124
1125
       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
1126
         stop 1
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
       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
1150
         stop 1
1151
1152
1153
1154
1155
1156
1157
       endif

      allocate(top_border_recv_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_recv_buffer"//errorMessage
Andreas Marek's avatar
Andreas Marek committed
1158
         stop 1
1159
1160
1161
1162
1163
1164
1165
       endif

      allocate(bottom_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 bottom_border_send_buffer"//errorMessage
Andreas Marek's avatar
Andreas Marek committed
1166
         stop 1
1167
1168
1169
1170
1171
1172
1173
       endif

      allocate(bottom_border_recv_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 bottom_border_recv_buffer"//errorMessage
Andreas Marek's avatar
Andreas Marek committed
1174
         stop 1
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
       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