elpa2_template.F90 21.7 KB
Newer Older
1
!    This file is part of ELPA.
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
!
!    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".
 function elpa_solve_evp_&
  &MATH_DATATYPE&
  &_&
  &2stage_&
  &PRECISION&
57
  &_impl (obj, a, ev, q) result(success)
58

59
   use elpa_abstract_impl
60
   use elpa_utilities
61
62
63
64
65
66
67
   use elpa1_compute
   use elpa2_compute
   use elpa_mpi
   use cuda_functions
   use mod_check_for_gpu
   use iso_c_binding
   implicit none
Andreas Marek's avatar
Andreas Marek committed
68
69
   class(elpa_abstract_impl_t), intent(inout)                         :: obj
   logical                                                            :: useGPU
70
#if REALCASE == 1
Andreas Marek's avatar
Andreas Marek committed
71
72
   logical                                                            :: useQR
   logical                                                            :: useQRActual
Andreas Marek's avatar
Andreas Marek committed
73
#endif
Andreas Marek's avatar
Andreas Marek committed
74
   integer(kind=c_int)                                                :: kernel
75
76

#ifdef USE_ASSUMED_SIZE
Andreas Marek's avatar
Andreas Marek committed
77
   MATH_DATATYPE(kind=C_DATATYPE_KIND), intent(inout)                 :: a(obj%local_nrows,*)
Andreas Marek's avatar
Andreas Marek committed
78
   MATH_DATATYPE(kind=C_DATATYPE_KIND), optional, target, intent(out) :: q(obj%local_nrows,*)
79
#else
Andreas Marek's avatar
Andreas Marek committed
80
   MATH_DATATYPE(kind=C_DATATYPE_KIND), intent(inout)                 :: a(obj%local_nrows,obj%local_ncols)
Andreas Marek's avatar
Andreas Marek committed
81
   MATH_DATATYPE(kind=C_DATATYPE_KIND), optional, target, intent(out) :: q(obj%local_nrows,obj%local_ncols)
82
#endif
Andreas Marek's avatar
Andreas Marek committed
83
84
   real(kind=C_DATATYPE_KIND), intent(inout)                          :: ev(obj%na)
   MATH_DATATYPE(kind=C_DATATYPE_KIND), allocatable                   :: hh_trans(:,:)
85

Andreas Marek's avatar
Andreas Marek committed
86
   integer(kind=c_int)                                                :: my_pe, n_pes, my_prow, my_pcol, np_rows, np_cols, mpierr
Andreas Marek's avatar
Andreas Marek committed
87
88
89
90
   integer(kind=c_int)                                                :: nbw, num_blocks
#if COMPLEXCASE == 1
   integer(kind=c_int)                                                :: l_cols_nev, l_rows, l_cols
#endif
Andreas Marek's avatar
Andreas Marek committed
91
92
   MATH_DATATYPE(kind=C_DATATYPE_KIND), allocatable                   :: tmat(:,:,:)
   real(kind=C_DATATYPE_KIND), allocatable                            :: e(:)
93
#if COMPLEXCASE == 1
Andreas Marek's avatar
Andreas Marek committed
94
   real(kind=C_DATATYPE_KIND), allocatable                            :: q_real(:,:)
95
#endif
Andreas Marek's avatar
Andreas Marek committed
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
   MATH_DATATYPE(kind=C_DATATYPE_KIND), allocatable, target           :: q_dummy(:,:)
   MATH_DATATYPE(kind=C_DATATYPE_KIND), pointer                       :: q_actual(:,:)


   integer(kind=c_intptr_t)                                           :: tmat_dev, q_dev, a_dev

   integer(kind=c_int)                                                :: i
   logical                                                            :: success, successCUDA
   logical                                                            :: wantDebug
   integer(kind=c_int)                                                :: istat, gpu, debug, qr
   character(200)                                                     :: errorMessage
   logical                                                            :: do_useGPU, do_useGPU_trans_ev_tridi
   integer(kind=c_int)                                                :: numberOfGPUDevices
   integer(kind=c_intptr_t), parameter                                :: size_of_datatype = size_of_&
                                                                                            &PRECISION&
                                                                                            &_&
                                                                                            &MATH_DATATYPE
    integer(kind=ik)                                                  :: na, nev, lda, ldq, nblk, matrixCols, &
                                                                         mpi_comm_rows, mpi_comm_cols,        &
					                                 mpi_comm_all, check_pd

    logical                                                           :: do_bandred, do_tridiag, do_solve_tridi,  &
                                                                         do_trans_to_band, do_trans_to_full
119

120
    call obj%timer%start("elpa_solve_evp_&
121
    &MATH_DATATYPE&
122
123
124
    &_2stage_&
    &PRECISION&
    &")
125

Andreas Marek's avatar
Andreas Marek committed
126
127
    success = .true.

Andreas Marek's avatar
Andreas Marek committed
128
    if (present(q)) then
129
      obj%eigenvalues_only = .false.
Andreas Marek's avatar
Andreas Marek committed
130
    else
131
      obj%eigenvalues_only = .true.
Andreas Marek's avatar
Andreas Marek committed
132
133
    endif

134
135
136
137
138
139
140
    na         = obj%na
    nev        = obj%nev
    lda        = obj%local_nrows
    ldq        = obj%local_nrows
    nblk       = obj%nblk
    matrixCols = obj%local_ncols

Andreas Marek's avatar
Andreas Marek committed
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
   ! special case na = 1
   if (na .eq. 1) then
#if REALCASE == 1
     ev(1) = a(1,1)
     if (.not.(obj%eigenvalues_only)) then
       q(1,1) = CONST_REAL_1_0
     endif
#endif
#if COMPLEXCASE == 1
     ev(1) = real(a(1,1))
     if (.not.(obj%eigenvalues_only)) then
       q(1,1) = CONST_COMPLEX_PAIR_1_0
     endif
#endif
     call obj%timer%stop("elpa_solve_evp_&
     &MATH_DATATYPE&
     &_2stage_&
     &PRECISION&
     &")
     return
   endif

163
164
165
166
167
   if (nev == 0) then
     nev = 1
     obj%eigenvalues_only = .true.
   endif

168
#if REALCASE == 1
169
    call obj%get("real_kernel",kernel)
170
    ! check consistency between request for GPUs and defined kernel
171
172
    call obj%get("gpu", gpu)
    if (gpu == 1) then
173
174
      if (kernel .ne. ELPA_2STAGE_REAL_GPU) then
        write(error_unit,*) "ELPA: Warning, GPU usage has been requested but compute kernel is defined as non-GPU!"
175
        write(error_unit,*) "The compute kernel will be executed on CPUs!"
176
      else if (nblk .ne. 128) then
177
178
179
        kernel = ELPA_2STAGE_REAL_GENERIC
      endif
    endif
180
181
182
183
184
    if (kernel .eq. ELPA_2STAGE_REAL_GPU) then
      if (gpu .ne. 1) then
        write(error_unit,*) "ELPA: Warning, GPU usage has been requested but compute kernel is defined as non-GPU!"
      endif
    endif
185
186
187
188
189
190
191
192
193
194
195
196

#ifdef SINGLE_PRECISION_REAL
    ! special case at the moment NO single precision kernels on POWER 8 -> set GENERIC for now
    if (kernel .eq. ELPA_2STAGE_REAL_VSX_BLOCK2 .or. &
        kernel .eq. ELPA_2STAGE_REAL_VSX_BLOCK4 .or. &
        kernel .eq. ELPA_2STAGE_REAL_VSX_BLOCK6        ) then
        write(error_unit,*) "ELPA: At the moment there exist no specific SINGLE precision kernels for POWER8"
        write(error_unit,*) "The GENERIC kernel will be used at the moment"
        kernel = ELPA_2STAGE_REAL_GENERIC
    endif
#endif

197
198
199
#endif

#if COMPLEXCASE == 1
200
    call obj%get("complex_kernel",kernel)
201
    ! check consistency between request for GPUs and defined kernel
202
203
    call obj%get("gpu", gpu)
    if (gpu == 1) then
204
205
      if (kernel .ne. ELPA_2STAGE_COMPLEX_GPU) then
        write(error_unit,*) "ELPA: Warning, GPU usage has been requested but compute kernel is defined as non-GPU!"
206
        write(error_unit,*) "The compute kernel will be executed on CPUs!"
207
      else if (nblk .ne. 128) then
208
209
210
        kernel = ELPA_2STAGE_COMPLEX_GENERIC
      endif
    endif
211
212
213
214
215
216
    if (kernel .eq. ELPA_2STAGE_COMPLEX_GPU) then
      if (gpu .ne. 1) then
        write(error_unit,*) "ELPA: Warning, GPU usage has been requested but compute kernel is defined as non-GPU!"
      endif
    endif

217
#endif
218
219
220
    call obj%get("mpi_comm_rows",mpi_comm_rows)
    call obj%get("mpi_comm_cols",mpi_comm_cols)
    call obj%get("mpi_comm_parent",mpi_comm_all)
221

222
    if (gpu .eq. 1) then
223
224
225
226
227
228
      useGPU = .true.
    else
      useGPU = .false.
    endif

#if REALCASE == 1
229
230
    call obj%get("qr",qr)
    if (qr .eq. 1) then
231
232
233
234
235
236
      useQR = .true.
    else
      useQR = .false.
    endif

#endif
237
    call obj%timer%start("mpi_communication")
238
239
240
241
242
243
244
    call mpi_comm_rank(mpi_comm_all,my_pe,mpierr)
    call mpi_comm_size(mpi_comm_all,n_pes,mpierr)

    call mpi_comm_rank(mpi_comm_rows,my_prow,mpierr)
    call mpi_comm_size(mpi_comm_rows,np_rows,mpierr)
    call mpi_comm_rank(mpi_comm_cols,my_pcol,mpierr)
    call mpi_comm_size(mpi_comm_cols,np_cols,mpierr)
245
    call obj%timer%stop("mpi_communication")
246

247
248
    call obj%get("debug",debug)
    wantDebug = debug == 1
249
250
251
252
253
254
255
256

    do_useGPU      = .false.
    do_useGPU_trans_ev_tridi =.false.


#if REALCASE == 1
    useQRActual = .false.
    ! set usage of qr decomposition via API call
257
258
    if (useQR) useQRActual = .true.
    if (.not.(useQR)) useQRACtual = .false.
259
260
261
262
263
264
265
266
267
268
269
270
271

    if (useQRActual) then
      if (mod(na,2) .ne. 0) then
        if (wantDebug) then
          write(error_unit,*) "solve_evp_real_2stage: QR-decomposition: blocksize does not fit with matrixsize"
        endif
        print *, "Do not use QR-decomposition for this matrix and blocksize."
        success = .false.
        return
      endif
    endif
#endif /* REALCASE */

272
273
    if (useGPU) then
      if (check_for_gpu(my_pe,numberOfGPUDevices, wantDebug=wantDebug)) then
274

275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
         do_useGPU = .true.

         ! set the neccessary parameters
         cudaMemcpyHostToDevice   = cuda_memcpyHostToDevice()
         cudaMemcpyDeviceToHost   = cuda_memcpyDeviceToHost()
         cudaMemcpyDeviceToDevice = cuda_memcpyDeviceToDevice()
         cudaHostRegisterPortable = cuda_hostRegisterPortable()
         cudaHostRegisterMapped   = cuda_hostRegisterMapped()
      else
        print *,"GPUs are requested but not detected! Aborting..."
        success = .false.
        return
      endif
    else
      ! check whether set by environment variable
290
291
      call obj%get("gpu",gpu)
      do_useGPU = gpu == 1
292
293
      if (do_useGPU) then
        if (check_for_gpu(my_pe,numberOfGPUDevices, wantDebug=wantDebug)) then
294
295
296
297
298
299
300
301
302
303
304
305
306

           ! set the neccessary parameters
           cudaMemcpyHostToDevice   = cuda_memcpyHostToDevice()
           cudaMemcpyDeviceToHost   = cuda_memcpyDeviceToHost()
           cudaMemcpyDeviceToDevice = cuda_memcpyDeviceToDevice()
           cudaHostRegisterPortable = cuda_hostRegisterPortable()
           cudaHostRegisterMapped   = cuda_hostRegisterMapped()
        else
          print *,"GPUs are requested but not detected! Aborting..."
          success = .false.
          return
        endif
      endif
307
    endif
308
309
310

    ! check consistency between request for GPUs and defined kernel
    if (do_useGPU) then
311
312
313
314
315
      if (nblk .ne. 128) then
        ! cannot run on GPU with this blocksize
        ! disable GPU usage for trans_ev_tridi
        do_useGPU_trans_ev_tridi = .false.
      else
316
317
318
319
320
321
322
323
324
325
#if REALCASE == 1
        if (kernel .eq. ELPA_2STAGE_REAL_GPU) then
#endif
#if COMPLEXCASE == 1
        if (kernel .eq. ELPA_2STAGE_COMPLEX_GPU) then
#endif
          do_useGPU_trans_ev_tridi = .true.
	else
          do_useGPU_trans_ev_tridi = .false.
	endif
326
327
      endif
    endif
328

329

Andreas Marek's avatar
Andreas Marek committed
330

331
    if (.not. obj%eigenvalues_only) then
Andreas Marek's avatar
Andreas Marek committed
332
333
334
335
336
337
      q_actual => q(1:obj%local_nrows,1:obj%local_ncols)
    else
     allocate(q_dummy(1:obj%local_nrows,1:obj%local_ncols))
     q_actual => q_dummy(1:obj%local_nrows,1:obj%local_ncols)
    endif

338
339
340
341
342
343
344
345
346
347
348
349
350

    ! set the default values for each of the 5 compute steps
    do_bandred        = .true.
    do_tridiag        = .true.
    do_solve_tridi    = .true.
    do_trans_to_band  = .true.
    do_trans_to_full  = .true.

    if (obj%eigenvalues_only) then
      do_trans_to_band  = .false.
      do_trans_to_full  = .false.
    endif

Andreas Marek's avatar
Andreas Marek committed
351
    if (obj%is_set("bandwidth") == 1) then
352
      call obj%get("bandwidth",nbw)
353
      if (nbw == 0) then
354
        if (wantDebug) then
355
356
357
358
359
          write(error_unit,*) "Specified bandwidth = 0; ELPA refuses to solve the eigenvalue problem ", &
	                      "for a diagonal matrix! This is too simple"
	  endif
        print *, "Specified bandwidth = 0; ELPA refuses to solve the eigenvalue problem ", &
	         "for a diagonal matrix! This is too simple"
360
361
362
        success = .false.
        return
      endif
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
      if (mod(nbw, nblk) .ne. 0) then
        ! treat matrix with an effective bandwidth slightly bigger than specified bandwidth
	! such that effective bandwidth is a multiply of nblk. which is a prerequiste for ELPA
        nbw = nblk * ceiling(real(nbw,kind=c_double)/real(nblk,kind=c_double))

        ! just check that effective bandwidth is NOT larger than matrix size
	if (nbw .gt. na) then
          if (wantDebug) then
            write(error_unit,*) "Specified bandwidth ",nbw," leads internaly to a computed bandwidth ", &
	                        "which is larger than the matrix size ",na," ! ELPA will abort! Try to", &
				"solve your problem by not specifing a bandwidth"
	  endif
          print *, "Specified bandwidth ",nbw," leads internaly to a computed bandwidth ", &
	                        "which is larger than the matrix size ",na," ! ELPA will abort! Try to", &
				"solve your problem by not specifing a bandwidth"
          success = .false.
          return
	endif
      endif
382
383
384
385
386
      do_bandred       = .false. ! we already have a banded matrix
      do_solve_tridi   = .true.  ! we also have to solve something :-)
      do_trans_to_band = .true.  ! and still we have to backsub to banded
      do_trans_to_full = .false. ! but not to full since we have a banded matrix
    else ! bandwidth is not set
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407

      ! Choose bandwidth, must be a multiple of nblk, set to a value >= 32
      ! On older systems (IBM Bluegene/P, Intel Nehalem) a value of 32 was optimal.
      ! For Intel(R) Xeon(R) E5 v2 and v3, better use 64 instead of 32!
      ! For IBM Bluegene/Q this is not clear at the moment. We have to keep an eye
      ! on this and maybe allow a run-time optimization here
      if (do_useGPU) then
        nbw = nblk
      else
#if REALCASE == 1
        nbw = (63/nblk+1)*nblk
#elif COMPLEXCASE == 1
        nbw = (31/nblk+1)*nblk
#endif
      endif

      num_blocks = (na-1)/nbw + 1

      allocate(tmat(nbw,nbw,num_blocks), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
        print *,"solve_evp_&
408
409
410
411
        &MATH_DATATYPE&
        &_2stage_&
        &PRECISION&
        &" // ": error when allocating tmat "//errorMessage
412
413
414
        stop 1
      endif

415
416
417
418
419
420
421
422
423
424
      do_bandred       = .true.
      do_solve_tridi   = .true.
      do_trans_to_band = .true.
      do_trans_to_full = .true.
    end if  ! matrix not already banded on input

    ! start the computations in 5 steps

    if (do_bandred) then
      call obj%timer%start("bandred")
425
426
427
428
429
      ! Reduction full -> band
      call bandred_&
      &MATH_DATATYPE&
      &_&
      &PRECISION &
430
      (obj, na, a, &
431
432
      a_dev, lda, nblk, nbw, matrixCols, num_blocks, mpi_comm_rows, mpi_comm_cols, tmat, &
      tmat_dev,  wantDebug, do_useGPU, success &
433
#if REALCASE == 1
434
      , useQRActual &
435
#endif
436
437
      )
      call obj%timer%stop("bandred")
438
      if (.not.(success)) return
439
440
    endif

441
442

     ! Reduction band -> tridiagonal
443
444
445
446
447
448
449
450
451
     if (do_tridiag) then
       allocate(e(na), stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
         print *,"solve_evp_&
         &MATH_DATATYPE&
         &_2stage_&
         &PRECISION " // ": error when allocating e "//errorMessage
         stop 1
       endif
452

453
454
       call obj%timer%start("tridiag")
       call tridiag_band_&
455
       &MATH_DATATYPE&
456
457
       &_&
       &PRECISION&
458
459
       (obj, na, nbw, nblk, a, a_dev, lda, ev, e, matrixCols, hh_trans, mpi_comm_rows, mpi_comm_cols, mpi_comm_all, &
        do_useGPU, wantDebug)
460
461

#ifdef WITH_MPI
462
463
464
465
       call obj%timer%start("mpi_communication")
       call mpi_bcast(ev, na, MPI_REAL_PRECISION, 0, mpi_comm_all, mpierr)
       call mpi_bcast(e, na, MPI_REAL_PRECISION, 0, mpi_comm_all, mpierr)
       call obj%timer%stop("mpi_communication")
466
#endif /* WITH_MPI */
467
468
       call obj%timer%stop("tridiag")
     endif ! do_tridiag
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484

#if COMPLEXCASE == 1
     l_rows = local_index(na, my_prow, np_rows, nblk, -1) ! Local rows of a and q
     l_cols = local_index(na, my_pcol, np_cols, nblk, -1) ! Local columns of q
     l_cols_nev = local_index(nev, my_pcol, np_cols, nblk, -1) ! Local columns corresponding to nev

     allocate(q_real(l_rows,l_cols), stat=istat, errmsg=errorMessage)
     if (istat .ne. 0) then
       print *,"solve_evp_&
       &MATH_DATATYPE&
       &_2stage: error when allocating q_real"//errorMessage
       stop 1
     endif
#endif

     ! Solve tridiagonal system
485
486
487
488
489
     if (do_solve_tridi) then
       call obj%timer%start("solve")
       call solve_tridi_&
       &PRECISION &
       (obj, na, nev, ev, e, &
490
#if REALCASE == 1
491
       q_actual, ldq,   &
492
493
#endif
#if COMPLEXCASE == 1
494
       q_real, ubound(q_real,dim=1), &
495
#endif
496
497
498
499
       nblk, matrixCols, mpi_comm_rows, mpi_comm_cols, wantDebug, success)
       call obj%timer%stop("solve")
       if (.not.(success)) return
     endif ! do_solve_tridi
500
501
502
503
504
505
506
507
508

     deallocate(e, stat=istat, errmsg=errorMessage)
     if (istat .ne. 0) then
       print *,"solve_evp_&
       &MATH_DATATYPE&
       &_2stage: error when deallocating e "//errorMessage
       stop 1
     endif

509
     if (obj%eigenvalues_only) then
Andreas Marek's avatar
Andreas Marek committed
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
       do_trans_to_band = .false.
       do_trans_to_full = .false.
     else

       call obj%get("check_pd",check_pd)
       if (check_pd .eq. 1) then
         check_pd = 0
         do i = 1, na
           if (ev(i) .gt. THRESHOLD) then
             check_pd = check_pd + 1
           endif
         enddo
         if (check_pd .lt. na) then
           ! not positiv definite => eigenvectors needed
           do_trans_to_band = .true.
           do_trans_to_full = .true.
	 else
           do_trans_to_band = .false.
           do_trans_to_full = .false.
         endif
       endif
     endif ! eigenvalues only
532

533
     if (do_trans_to_band) then
534
#if COMPLEXCASE == 1
Andreas Marek's avatar
Andreas Marek committed
535
       ! q must be given thats why from here on we can use q and not q_actual
536

Andreas Marek's avatar
Andreas Marek committed
537
       q(1:l_rows,1:l_cols_nev) = q_real(1:l_rows,1:l_cols_nev)
538

Andreas Marek's avatar
Andreas Marek committed
539
540
541
542
543
544
545
546
       deallocate(q_real, stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
         print *,"solve_evp_&
         &MATH_DATATYPE&
         &_2stage: error when deallocating q_real"//errorMessage
         stop 1
       endif
#endif
Andreas Marek's avatar
Andreas Marek committed
547

Andreas Marek's avatar
Andreas Marek committed
548
549
       ! Backtransform stage 1
       call obj%timer%start("trans_ev_to_band")
550

Andreas Marek's avatar
Andreas Marek committed
551
       call trans_ev_tridi_to_band_&
552
       &MATH_DATATYPE&
Andreas Marek's avatar
Andreas Marek committed
553
554
555
556
557
558
559
       &_&
       &PRECISION &
       (obj, na, nev, nblk, nbw, q, &
       q_dev, &
       ldq, matrixCols, hh_trans, mpi_comm_rows, mpi_comm_cols, wantDebug, do_useGPU_trans_ev_tridi, &
       success=success, kernel=kernel)
       call obj%timer%stop("trans_ev_to_band")
560

Andreas Marek's avatar
Andreas Marek committed
561
       if (.not.(success)) return
562

Andreas Marek's avatar
Andreas Marek committed
563
564
565
566
567
568
569
570
       ! We can now deallocate the stored householder vectors
       deallocate(hh_trans, stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
         print *, "solve_evp_&
         &MATH_DATATYPE&
         &_2stage_&
         &PRECISION " // ": error when deallocating hh_trans "//errorMessage
         stop 1
571
       endif
572
     endif ! do_trans_to_band
573

574
     if (do_trans_to_full) then
Andreas Marek's avatar
Andreas Marek committed
575
       call obj%timer%start("trans_ev_to_full")
576
577
578
       if ( (do_useGPU) .and. .not.(do_useGPU_trans_ev_tridi) ) then
         ! copy to device if we want to continue on GPU
         successCUDA = cuda_malloc(q_dev, ldq*matrixCols*size_of_datatype)
Andreas Marek's avatar
Andreas Marek committed
579

580
581
         successCUDA = cuda_memcpy(q_dev, loc(q), ldq*matrixCols* size_of_datatype, cudaMemcpyHostToDevice)
       endif
Andreas Marek's avatar
Andreas Marek committed
582

583
       ! Backtransform stage 2
Andreas Marek's avatar
Andreas Marek committed
584

585
586
587
588
589
590
591
592
       call trans_ev_band_to_full_&
       &MATH_DATATYPE&
       &_&
       &PRECISION &
       (obj, na, nev, nblk, nbw, a, &
       a_dev, lda, tmat, tmat_dev,  q,  &
       q_dev, &
       ldq, matrixCols, num_blocks, mpi_comm_rows, mpi_comm_cols, do_useGPU &
593
#if REALCASE == 1
594
       , useQRActual  &
595
#endif
596
       )
597

598

599
600
601
602
603
604
605
606
       deallocate(tmat, stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
         print *,"solve_evp_&
         &MATH_DATATYPE&
         &_2stage_&
         &PRECISION " // ": error when deallocating tmat"//errorMessage
         stop 1
       endif
Andreas Marek's avatar
Andreas Marek committed
607
       call obj%timer%stop("trans_ev_to_full")
608
     endif ! do_trans_to_full
Andreas Marek's avatar
Andreas Marek committed
609

610
     if (obj%eigenvalues_only) then
Andreas Marek's avatar
Andreas Marek committed
611
       deallocate(q_dummy, stat=istat, errmsg=errorMessage)
612
613
       if (istat .ne. 0) then
         print *,"solve_evp_&
Andreas Marek's avatar
Andreas Marek committed
614
615
616
617
         &MATH_DATATYPE&
         &_1stage_&
         &PRECISION&
         &" // ": error when deallocating q_dummy "//errorMessage
618
619
620
621
         stop 1
       endif
     endif

622
     call obj%timer%stop("elpa_solve_evp_&
623
     &MATH_DATATYPE&
624
625
626
     &_2stage_&
    &PRECISION&
    &")
627
628
629
630
631
632
1    format(a,f10.3)

   end function elpa_solve_evp_&
   &MATH_DATATYPE&
   &_2stage_&
   &PRECISION&
633
   &_impl
634
635

! vim: syntax=fortran