elpa2_template.F90 28.5 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
   use elpa1_compute
   use elpa2_compute
   use elpa_mpi
   use cuda_functions
   use mod_check_for_gpu
66
67
   use elpa_omp

68
69
   use iso_c_binding
   implicit none
70
#include "../general/precision_kinds.F90"
Andreas Marek's avatar
Andreas Marek committed
71
72
   class(elpa_abstract_impl_t), intent(inout)                         :: obj
   logical                                                            :: useGPU
73
#if REALCASE == 1
Andreas Marek's avatar
Andreas Marek committed
74
75
   logical                                                            :: useQR
   logical                                                            :: useQRActual
Andreas Marek's avatar
Andreas Marek committed
76
#endif
Andreas Marek's avatar
Andreas Marek committed
77
   integer(kind=c_int)                                                :: kernel
78
79

#ifdef USE_ASSUMED_SIZE
Andreas Marek's avatar
Andreas Marek committed
80
   MATH_DATATYPE(kind=C_DATATYPE_KIND), intent(inout)                 :: a(obj%local_nrows,*)
Andreas Marek's avatar
Andreas Marek committed
81
   MATH_DATATYPE(kind=C_DATATYPE_KIND), optional, target, intent(out) :: q(obj%local_nrows,*)
82
#else
Andreas Marek's avatar
Andreas Marek committed
83
   MATH_DATATYPE(kind=C_DATATYPE_KIND), intent(inout)                 :: a(obj%local_nrows,obj%local_ncols)
Andreas Marek's avatar
Andreas Marek committed
84
   MATH_DATATYPE(kind=C_DATATYPE_KIND), optional, target, intent(out) :: q(obj%local_nrows,obj%local_ncols)
85
#endif
Andreas Marek's avatar
Andreas Marek committed
86
87
   real(kind=C_DATATYPE_KIND), intent(inout)                          :: ev(obj%na)
   MATH_DATATYPE(kind=C_DATATYPE_KIND), allocatable                   :: hh_trans(:,:)
88

Andreas Marek's avatar
Andreas Marek committed
89
   integer(kind=c_int)                                                :: my_pe, n_pes, my_prow, my_pcol, np_rows, np_cols, mpierr
Andreas Marek's avatar
Andreas Marek committed
90
91
92
93
   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
94
95
   MATH_DATATYPE(kind=C_DATATYPE_KIND), allocatable                   :: tmat(:,:,:)
   real(kind=C_DATATYPE_KIND), allocatable                            :: e(:)
96
#if COMPLEXCASE == 1
Andreas Marek's avatar
Andreas Marek committed
97
   real(kind=C_DATATYPE_KIND), allocatable                            :: q_real(:,:)
98
#endif
Andreas Marek's avatar
Andreas Marek committed
99
100
101
102
103
104
105
106
107
108
109
   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
110
   logical                                                            :: do_useGPU, do_useGPU_bandred, &
111
                                                                         do_useGPU_tridiag_band, do_useGPU_solve_tridi, &
112
113
                                                                         do_useGPU_trans_ev_tridi_to_band, &
                                                                         do_useGPU_trans_ev_band_to_full
Andreas Marek's avatar
Andreas Marek committed
114
115
116
117
118
119
120
   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,        &
Andreas Marek's avatar
Andreas Marek committed
121
                                                                         mpi_comm_all, check_pd, error
Andreas Marek's avatar
Andreas Marek committed
122
123
124

    logical                                                           :: do_bandred, do_tridiag, do_solve_tridi,  &
                                                                         do_trans_to_band, do_trans_to_full
125

Andreas Marek's avatar
Andreas Marek committed
126
    integer(kind=ik)                                                  :: nrThreads
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
#if REALCASE == 1
#undef GPU_KERNEL
#undef GENERIC_KERNEL
#undef KERNEL_STRING
#define GPU_KERNEL ELPA_2STAGE_REAL_GPU
#define GENERIC_KERNEL ELPA_2STAGE_REAL_GENERIC
#define KERNEL_STRING "real_kernel"
#endif
#if COMPLEXCASE == 1
#undef GPU_KERNEL
#undef GENERIC_KERNEL
#undef KERNEL_STRING
#define GPU_KERNEL ELPA_2STAGE_COMPLEX_GPU
#define GENERIC_KERNEL ELPA_2STAGE_COMPLEX_GENERIC
#define KERNEL_STRING "complex_kernel"
#endif

144
    call obj%timer%start("elpa_solve_evp_&
145
    &MATH_DATATYPE&
146
147
148
    &_2stage_&
    &PRECISION&
    &")
149

Andreas Marek's avatar
Andreas Marek committed
150
151

#ifdef WITH_OPENMP
152
153
154
155
156
    ! store the number of OpenMP threads used in the calling function
    ! restore this at the end of ELPA 2
    omp_threads_caller = omp_get_max_threads()

    ! check the number of threads that ELPA should use internally
Andreas Marek's avatar
Andreas Marek committed
157
158
    call obj%get("omp_threads",nrThreads,error)
    call omp_set_num_threads(nrThreads)
Andreas Marek's avatar
Andreas Marek committed
159
160
161
162
#else
    nrThreads = 1
#endif

Andreas Marek's avatar
Andreas Marek committed
163
164
    success = .true.

Andreas Marek's avatar
Andreas Marek committed
165
    if (present(q)) then
166
      obj%eigenvalues_only = .false.
Andreas Marek's avatar
Andreas Marek committed
167
    else
168
      obj%eigenvalues_only = .true.
Andreas Marek's avatar
Andreas Marek committed
169
170
    endif

171
172
173
174
175
176
177
    na         = obj%na
    nev        = obj%nev
    lda        = obj%local_nrows
    ldq        = obj%local_nrows
    nblk       = obj%nblk
    matrixCols = obj%local_ncols

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
203
    call obj%get("mpi_comm_rows",mpi_comm_rows,error)
    if (error .ne. ELPA_OK) then
      print *,"Problem getting option. Aborting..."
      stop
    endif
    call obj%get("mpi_comm_cols",mpi_comm_cols,error)
    if (error .ne. ELPA_OK) then
      print *,"Problem getting option. Aborting..."
      stop
    endif
    call obj%get("mpi_comm_parent",mpi_comm_all,error)
    if (error .ne. ELPA_OK) then
      print *,"Problem getting option. Aborting..."
      stop
    endif

    call obj%timer%start("mpi_communication")
    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)
    call obj%timer%stop("mpi_communication")

Andreas Marek's avatar
Andreas Marek committed
204
205
206
207
208
209
210
   ! special case na = 1
   if (na .eq. 1) then
#if REALCASE == 1
     ev(1) = a(1,1)
#endif
#if COMPLEXCASE == 1
     ev(1) = real(a(1,1))
211
#endif
Andreas Marek's avatar
Andreas Marek committed
212
     if (.not.(obj%eigenvalues_only)) then
213
       q(1,1) = ONE
Andreas Marek's avatar
Andreas Marek committed
214
     endif
215
216
217
218
219
220
221
222

     ! restore original OpenMP settings
#ifdef WITH_OPENMP
     ! store the number of OpenMP threads used in the calling function
     ! restore this at the end of ELPA 2
     call omp_set_num_threads(omp_threads_caller)
#endif

Andreas Marek's avatar
Andreas Marek committed
223
224
225
226
227
228
229
230
     call obj%timer%stop("elpa_solve_evp_&
     &MATH_DATATYPE&
     &_2stage_&
     &PRECISION&
     &")
     return
   endif

231
232
233
234
235
   if (nev == 0) then
     nev = 1
     obj%eigenvalues_only = .true.
   endif

236
    call obj%get(KERNEL_STRING,kernel,error)
Andreas Marek's avatar
Andreas Marek committed
237
238
239
240
    if (error .ne. ELPA_OK) then
      print *,"Problem getting option. Aborting..."
      stop
    endif
241
242

    ! GPU settings
Andreas Marek's avatar
Andreas Marek committed
243
244
245
246
247
    call obj%get("gpu", gpu,error)
    if (error .ne. ELPA_OK) then
      print *,"Problem getting option. Aborting..."
      stop
    endif
248

249
    useGPU = (gpu == 1)
250

251
    do_useGPU = .false.
252
    if (useGPU) then
253
      call obj%timer%start("check_for_gpu")
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
      if (check_for_gpu(my_pe,numberOfGPUDevices, wantDebug=wantDebug)) then

         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
269
      call obj%timer%stop("check_for_gpu")
270
271
    endif

272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
    do_useGPU_bandred = do_useGPU
    do_useGPU_tridiag_band = do_useGPU
    do_useGPU_solve_tridi = do_useGPU
    do_useGPU_trans_ev_tridi_to_band = do_useGPU
    do_useGPU_trans_ev_band_to_full = do_useGPU

    ! only if we want (and can) use GPU in general, look what are the
    ! requirements for individual routines. Implicitly they are all set to 1, so
    ! unles specified otherwise by the user, GPU versions of all individual
    ! routines should be used
    if(do_useGPU) then
      call obj%get("gpu_bandred", gpu, error)
      if (error .ne. ELPA_OK) then
        print *,"Problem getting option. Aborting..."
        stop
      endif
      do_useGPU_bandred = (gpu == 1)

      call obj%get("gpu_tridiag_band", gpu, error)
      if (error .ne. ELPA_OK) then
        print *,"Problem getting option. Aborting..."
        stop
      endif
      do_useGPU_tridiag_band = (gpu == 1)

      call obj%get("gpu_solve_tridi", gpu, error)
      if (error .ne. ELPA_OK) then
        print *,"Problem getting option. Aborting..."
        stop
      endif
      do_useGPU_solve_tridi = (gpu == 1)

      call obj%get("gpu_trans_ev_tridi_to_band", gpu, error)
      if (error .ne. ELPA_OK) then
        print *,"Problem getting option. Aborting..."
        stop
      endif
      do_useGPU_trans_ev_tridi_to_band = (gpu == 1)

      call obj%get("gpu_trans_ev_band_to_full", gpu, error)
      if (error .ne. ELPA_OK) then
        print *,"Problem getting option. Aborting..."
        stop
      endif
      do_useGPU_trans_ev_band_to_full = (gpu == 1)
    endif

319
    ! check consistency between request for GPUs and defined kernel
320
321
322

    if (do_useGPU_trans_ev_tridi_to_band) then
      if (kernel .ne. GPU_KERNEL) then
323
        write(error_unit,*) "ELPA: Warning, GPU usage has been requested but compute kernel is defined as non-GPU!"
324
        write(error_unit,*) "The compute kernel will be executed on CPUs!"
325
        do_useGPU_trans_ev_tridi_to_band = .false.
326
      else if (nblk .ne. 128) then
327
328
329
330
        write(error_unit,*) "ELPA: Warning, GPU kernel can run only with scalapack block size 128!"
        write(error_unit,*) "The compute kernel will be executed on CPUs!"
        do_useGPU_trans_ev_tridi_to_band = .false.
        kernel = GENERIC_KERNEL
331
332
      endif
    endif
333

334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
    ! check again, now kernel and do_useGPU_trans_ev_tridi_to_band sould be
    ! finally consistent
    if (do_useGPU_trans_ev_tridi_to_band) then
      if (kernel .ne. GPU_KERNEL) then
        ! this should never happen, checking as an assert
        write(error_unit,*) "ELPA: INTERNAL ERROR setting GPU kernel!  Aborting..."
        stop
      endif
      if (nblk .ne. 128) then
        ! this should never happen, checking as an assert
        write(error_unit,*) "ELPA: INTERNAL ERROR setting GPU kernel and blocksize!  Aborting..."
        stop
      endif
    else
      if (kernel .eq. GPU_KERNEL) then
        ! combination not allowed
        write(error_unit,*) "ELPA: Warning, GPU usage has NOT been requested but compute kernel &
                            &is defined as the GPU kernel!  Aborting..."
        stop
        !TODO do error handling properly
354
355
      endif
    endif
356

357

358
#if REALCASE == 1
359
360
361
362
363
364
365
366
367
#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
368
369
370
371
372
373
374
375
    ! special case at the moment NO single precision kernels on SPARC64 -> set GENERIC for now
    if (kernel .eq. ELPA_2STAGE_REAL_SPARC64_BLOCK2 .or. &
        kernel .eq. ELPA_2STAGE_REAL_SPARC64_BLOCK4 .or. &
        kernel .eq. ELPA_2STAGE_REAL_SPARC64_BLOCK6        ) then
        write(error_unit,*) "ELPA: At the moment there exist no specific SINGLE precision kernels for SPARC64"
        write(error_unit,*) "The GENERIC kernel will be used at the moment"
        kernel = ELPA_2STAGE_REAL_GENERIC
    endif
376
377
#endif

378
379
380
381
#endif


#if REALCASE == 1
Andreas Marek's avatar
Andreas Marek committed
382
383
384
385
386
    call obj%get("qr",qr,error)
    if (error .ne. ELPA_OK) then
      print *,"Problem getting option. Aborting..."
      stop
    endif
387
    if (qr .eq. 1) then
388
389
390
391
392
393
      useQR = .true.
    else
      useQR = .false.
    endif

#endif
394

Andreas Marek's avatar
Andreas Marek committed
395
396
397
398
399
    call obj%get("debug",debug,error)
    if (error .ne. ELPA_OK) then
      print *,"Problem getting option. Aborting..."
      stop
    endif
400
    wantDebug = debug == 1
401
402
403
404
405
406



#if REALCASE == 1
    useQRActual = .false.
    ! set usage of qr decomposition via API call
407
408
    if (useQR) useQRActual = .true.
    if (.not.(useQR)) useQRACtual = .false.
409
410
411
412
413
414
415
416
417
418
419
420
421
422

    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 */


Andreas Marek's avatar
Andreas Marek committed
423

424
    if (.not. obj%eigenvalues_only) then
Andreas Marek's avatar
Andreas Marek committed
425
426
427
428
429
430
      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

431
432
433
434
435
436
437
438
439
440
441
442
443

    ! 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
444
    if (obj%is_set("bandwidth") == 1) then
445
446
      ! bandwidth is set. That means, that the inputed matrix is actually banded and thus the 
      ! first step of ELPA2 should be skipped
Andreas Marek's avatar
Andreas Marek committed
447
      call obj%get("bandwidth",nbw,error)
448
      if (nbw == 0) then
449
        if (wantDebug) then
450
          write(error_unit,*) "Specified bandwidth = 0; ELPA refuses to solve the eigenvalue problem ", &
Andreas Marek's avatar
Andreas Marek committed
451
452
                              "for a diagonal matrix! This is too simple"
          endif
453
        print *, "Specified bandwidth = 0; ELPA refuses to solve the eigenvalue problem ", &
Andreas Marek's avatar
Andreas Marek committed
454
                 "for a diagonal matrix! This is too simple"
455
456
457
        success = .false.
        return
      endif
458
459
      if (mod(nbw, nblk) .ne. 0) then
        ! treat matrix with an effective bandwidth slightly bigger than specified bandwidth
Andreas Marek's avatar
Andreas Marek committed
460
        ! such that effective bandwidth is a multiply of nblk. which is a prerequiste for ELPA
461
462
463
        nbw = nblk * ceiling(real(nbw,kind=c_double)/real(nblk,kind=c_double))

        ! just check that effective bandwidth is NOT larger than matrix size
Andreas Marek's avatar
Andreas Marek committed
464
        if (nbw .gt. na) then
465
466
          if (wantDebug) then
            write(error_unit,*) "Specified bandwidth ",nbw," leads internaly to a computed bandwidth ", &
Andreas Marek's avatar
Andreas Marek committed
467
468
469
                                "which is larger than the matrix size ",na," ! ELPA will abort! Try to", &
                                "solve your problem by not specifing a bandwidth"
          endif
470
          print *, "Specified bandwidth ",nbw," leads internaly to a computed bandwidth ", &
Andreas Marek's avatar
Andreas Marek committed
471
472
                                "which is larger than the matrix size ",na," ! ELPA will abort! Try to", &
                                "solve your problem by not specifing a bandwidth"
473
474
          success = .false.
          return
Andreas Marek's avatar
Andreas Marek committed
475
        endif
476
      endif
477
478
479
480
      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
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
    else ! matrix is not banded, determine the intermediate bandwidth for full->banded->tridi
      !first check if the intermediate bandwidth was set by the user
      call obj%get("intermediate_bandwidth", nbw, error)
      if (error .ne. ELPA_OK) then
        print *,"Problem getting option. Aborting..."
        stop
      endif

      if(nbw == 0) then
        ! intermediate bandwidth was not specified, select one of the defaults

        ! 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
500
#if REALCASE == 1
501
          nbw = (63/nblk+1)*nblk
502
#elif COMPLEXCASE == 1
503
          nbw = (31/nblk+1)*nblk
504
#endif
505
506
507
508
509
510
511
512
513
514
        endif

      else
        ! intermediate bandwidth has been specified by the user, check, whether correctly
        if (mod(nbw, nblk) .ne. 0) then
          print *, "Specified bandwidth ",nbw," has to be mutiple of the blocksize ", nblk, ". Aborting..."
          success = .false.
          return
        endif
      endif !nbw == 0
515
516
517

      num_blocks = (na-1)/nbw + 1

Pavel Kus's avatar
Pavel Kus committed
518
519
      ! tmat is needed only in full->band and band->full steps, so alocate here
      ! (not allocated for banded matrix on input)
520
521
522
      allocate(tmat(nbw,nbw,num_blocks), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
        print *,"solve_evp_&
523
524
525
526
        &MATH_DATATYPE&
        &_2stage_&
        &PRECISION&
        &" // ": error when allocating tmat "//errorMessage
527
528
529
        stop 1
      endif

Pavel Kus's avatar
Pavel Kus committed
530
531
532
533
534
535
536
537
538
539
540
541
      ! if either of full->band or band->full steps are to be done on GPU,
      ! allocate also corresponding array on GPU.
      if (do_useGPU_bandred .or.  do_useGPU_trans_ev_band_to_full) then
        successCUDA = cuda_malloc(tmat_dev, nbw*nbw* size_of_datatype)
        if (.not.(successCUDA)) then
          print *,"bandred_&
                  &MATH_DATATYPE&
                  &: error in cudaMalloc tmat_dev 1"
          stop 1
        endif
      endif

542
543
544
545
      do_bandred       = .true.
      do_solve_tridi   = .true.
      do_trans_to_band = .true.
      do_trans_to_full = .true.
546
    endif  ! matrix not already banded on input
547
548
549
550
551

    ! start the computations in 5 steps

    if (do_bandred) then
      call obj%timer%start("bandred")
552
553
554
555
556
      ! Reduction full -> band
      call bandred_&
      &MATH_DATATYPE&
      &_&
      &PRECISION &
557
      (obj, na, a, &
558
      a_dev, lda, nblk, nbw, matrixCols, num_blocks, mpi_comm_rows, mpi_comm_cols, tmat, &
559
      tmat_dev,  wantDebug, do_useGPU_bandred, success, &
560
#if REALCASE == 1
Andreas Marek's avatar
Andreas Marek committed
561
      useQRActual, &
562
#endif
Andreas Marek's avatar
Andreas Marek committed
563
       nrThreads)
564
      call obj%timer%stop("bandred")
565
      if (.not.(success)) return
566
567
    endif

568
569

     ! Reduction band -> tridiagonal
570
571
572
573
574
575
576
577
578
     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
579

580
581
       call obj%timer%start("tridiag")
       call tridiag_band_&
582
       &MATH_DATATYPE&
583
584
       &_&
       &PRECISION&
585
       (obj, na, nbw, nblk, a, a_dev, lda, ev, e, matrixCols, hh_trans, mpi_comm_rows, mpi_comm_cols, mpi_comm_all, &
586
       do_useGPU_tridiag_band, wantDebug, nrThreads)
587
588

#ifdef WITH_MPI
589
590
591
592
       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")
593
#endif /* WITH_MPI */
594
595
       call obj%timer%stop("tridiag")
     endif ! do_tridiag
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611

#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
612
613
614
615
616
     if (do_solve_tridi) then
       call obj%timer%start("solve")
       call solve_tridi_&
       &PRECISION &
       (obj, na, nev, ev, e, &
617
#if REALCASE == 1
618
       q_actual, ldq,   &
619
620
#endif
#if COMPLEXCASE == 1
621
       q_real, ubound(q_real,dim=1), &
622
#endif
623
       nblk, matrixCols, mpi_comm_rows, mpi_comm_cols, do_useGPU_solve_tridi, wantDebug, success, nrThreads)
624
625
626
       call obj%timer%stop("solve")
       if (.not.(success)) return
     endif ! do_solve_tridi
627
628
629
630
631
632
633
634
635

     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

636
     if (obj%eigenvalues_only) then
Andreas Marek's avatar
Andreas Marek committed
637
638
639
640
       do_trans_to_band = .false.
       do_trans_to_full = .false.
     else

Andreas Marek's avatar
Andreas Marek committed
641
642
643
644
645
       call obj%get("check_pd",check_pd,error)
       if (error .ne. ELPA_OK) then
         print *,"Problem getting option. Aborting..."
         stop
       endif
Andreas Marek's avatar
Andreas Marek committed
646
647
648
649
650
651
652
653
654
655
656
       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.
Andreas Marek's avatar
Andreas Marek committed
657
         else
Andreas Marek's avatar
Andreas Marek committed
658
659
660
661
662
           do_trans_to_band = .false.
           do_trans_to_full = .false.
         endif
       endif
     endif ! eigenvalues only
663

664
     if (do_trans_to_band) then
665
#if COMPLEXCASE == 1
Andreas Marek's avatar
Andreas Marek committed
666
       ! q must be given thats why from here on we can use q and not q_actual
667

Andreas Marek's avatar
Andreas Marek committed
668
       q(1:l_rows,1:l_cols_nev) = q_real(1:l_rows,1:l_cols_nev)
669

Andreas Marek's avatar
Andreas Marek committed
670
671
672
673
674
675
676
677
       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
678

Andreas Marek's avatar
Andreas Marek committed
679
680
       ! Backtransform stage 1
       call obj%timer%start("trans_ev_to_band")
681

Andreas Marek's avatar
Andreas Marek committed
682
       call trans_ev_tridi_to_band_&
683
       &MATH_DATATYPE&
Andreas Marek's avatar
Andreas Marek committed
684
685
686
687
       &_&
       &PRECISION &
       (obj, na, nev, nblk, nbw, q, &
       q_dev, &
688
       ldq, matrixCols, hh_trans, mpi_comm_rows, mpi_comm_cols, wantDebug, do_useGPU_trans_ev_tridi_to_band, &
Andreas Marek's avatar
Andreas Marek committed
689
       nrThreads, success=success, kernel=kernel)
Andreas Marek's avatar
Andreas Marek committed
690
       call obj%timer%stop("trans_ev_to_band")
691

Andreas Marek's avatar
Andreas Marek committed
692
       if (.not.(success)) return
693

Andreas Marek's avatar
Andreas Marek committed
694
695
696
697
698
699
700
701
       ! 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
702
       endif
703
     endif ! do_trans_to_band
704

705
706
707
708
709
710
711
712
713
714
     ! the array q might reside on device or host, depending on whether GPU is
     ! used or not. We thus have to transfer he data manually, if one of the
     ! routines is run on GPU and the other not.

     ! first deal with the situation that first backward step was on GPU
     if(do_useGPU_trans_ev_tridi_to_band) then
       ! if the second backward step is to be performed, but not on GPU, we have
       ! to transfer q to the host
       if(do_trans_to_full .and. (.not. do_useGPU_trans_ev_band_to_full)) then
         successCUDA = cuda_memcpy(loc(q), q_dev, ldq*matrixCols* size_of_datatype, cudaMemcpyDeviceToHost)
Pavel Kus's avatar
Pavel Kus committed
715
716
717
718
         if (.not.(successCUDA)) then
           print *,"elpa2_template, error in copy to host"
           stop 1
         endif
719
720
721
722
723
724
725
726
727
728
729
730
       endif

       ! if the last step is not required at all, or will be performed on CPU,
       ! release the memmory allocated on the device
       if((.not. do_trans_to_full) .or. (.not. do_useGPU_trans_ev_band_to_full)) then
         successCUDA = cuda_free(q_dev)
       endif
     endif

     !TODO check that the memory is properly dealocated on the host in case that
     !the last step is not required

731
     if (do_trans_to_full) then
Andreas Marek's avatar
Andreas Marek committed
732
       call obj%timer%start("trans_ev_to_full")
733
       if ( (do_useGPU_trans_ev_band_to_full) .and. .not.(do_useGPU_trans_ev_tridi_to_band) ) then
734
735
         ! 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
736

737
         successCUDA = cuda_memcpy(q_dev, loc(q), ldq*matrixCols* size_of_datatype, cudaMemcpyHostToDevice)
Pavel Kus's avatar
Pavel Kus committed
738
739
740
741
         if (.not.(successCUDA)) then
           print *,"elpa2_template, error in copy to device"
           stop 1
         endif
742
       endif
Andreas Marek's avatar
Andreas Marek committed
743

744
       ! Backtransform stage 2
Andreas Marek's avatar
Andreas Marek committed
745

746
747
748
749
750
751
752
       call trans_ev_band_to_full_&
       &MATH_DATATYPE&
       &_&
       &PRECISION &
       (obj, na, nev, nblk, nbw, a, &
       a_dev, lda, tmat, tmat_dev,  q,  &
       q_dev, &
753
       ldq, matrixCols, num_blocks, mpi_comm_rows, mpi_comm_cols, do_useGPU_trans_ev_band_to_full &
754
#if REALCASE == 1
755
       , useQRActual  &
756
#endif
757
       )
758

759

760
761
762
763
764
765
766
767
       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
768
       call obj%timer%stop("trans_ev_to_full")
769
     endif ! do_trans_to_full
Andreas Marek's avatar
Andreas Marek committed
770

Pavel Kus's avatar
Pavel Kus committed
771
772
773
774
775
776
777
778
779
780
     if(do_bandred .or. do_trans_to_full) then
       if (do_useGPU_bandred .or. do_useGPU_trans_ev_band_to_full) then
         successCUDA = cuda_free(tmat_dev)
         if (.not.(successCUDA)) then
           print *,"elpa2_template: error in cudaFree, tmat_dev"
           stop 1
         endif
       endif
     endif

781
     if (obj%eigenvalues_only) then
Andreas Marek's avatar
Andreas Marek committed
782
       deallocate(q_dummy, stat=istat, errmsg=errorMessage)
783
784
       if (istat .ne. 0) then
         print *,"solve_evp_&
Andreas Marek's avatar
Andreas Marek committed
785
786
787
788
         &MATH_DATATYPE&
         &_1stage_&
         &PRECISION&
         &" // ": error when deallocating q_dummy "//errorMessage
789
790
791
792
         stop 1
       endif
     endif

793
794
795
796
797
798
799
     ! restore original OpenMP settings
#ifdef WITH_OPENMP
    ! store the number of OpenMP threads used in the calling function
    ! restore this at the end of ELPA 2
    call omp_set_num_threads(omp_threads_caller)
#endif

800
     call obj%timer%stop("elpa_solve_evp_&
801
     &MATH_DATATYPE&
802
803
804
     &_2stage_&
    &PRECISION&
    &")
805
806
807
808
809
810
1    format(a,f10.3)

   end function elpa_solve_evp_&
   &MATH_DATATYPE&
   &_2stage_&
   &PRECISION&
811
   &_impl
812
813

! vim: syntax=fortran