elpa2_template.F90 25.8 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
Andreas Marek's avatar
Andreas Marek committed
66
67
68
#ifdef WITH_OPENMP
   use omp_lib
#endif
69
70
   use iso_c_binding
   implicit none
71
#include "../general/precision_kinds.F90"
Andreas Marek's avatar
Andreas Marek committed
72
73
   class(elpa_abstract_impl_t), intent(inout)                         :: obj
   logical                                                            :: useGPU
74
#if REALCASE == 1
Andreas Marek's avatar
Andreas Marek committed
75
76
   logical                                                            :: useQR
   logical                                                            :: useQRActual
Andreas Marek's avatar
Andreas Marek committed
77
#endif
Andreas Marek's avatar
Andreas Marek committed
78
   integer(kind=c_int)                                                :: kernel
79
80

#ifdef USE_ASSUMED_SIZE
Andreas Marek's avatar
Andreas Marek committed
81
   MATH_DATATYPE(kind=C_DATATYPE_KIND), intent(inout)                 :: a(obj%local_nrows,*)
Andreas Marek's avatar
Andreas Marek committed
82
   MATH_DATATYPE(kind=C_DATATYPE_KIND), optional, target, intent(out) :: q(obj%local_nrows,*)
83
#else
Andreas Marek's avatar
Andreas Marek committed
84
   MATH_DATATYPE(kind=C_DATATYPE_KIND), intent(inout)                 :: a(obj%local_nrows,obj%local_ncols)
Andreas Marek's avatar
Andreas Marek committed
85
   MATH_DATATYPE(kind=C_DATATYPE_KIND), optional, target, intent(out) :: q(obj%local_nrows,obj%local_ncols)
86
#endif
Andreas Marek's avatar
Andreas Marek committed
87
88
   real(kind=C_DATATYPE_KIND), intent(inout)                          :: ev(obj%na)
   MATH_DATATYPE(kind=C_DATATYPE_KIND), allocatable                   :: hh_trans(:,:)
89

Andreas Marek's avatar
Andreas Marek committed
90
   integer(kind=c_int)                                                :: my_pe, n_pes, my_prow, my_pcol, np_rows, np_cols, mpierr
Andreas Marek's avatar
Andreas Marek committed
91
92
93
94
   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
95
96
   MATH_DATATYPE(kind=C_DATATYPE_KIND), allocatable                   :: tmat(:,:,:)
   real(kind=C_DATATYPE_KIND), allocatable                            :: e(:)
97
#if COMPLEXCASE == 1
Andreas Marek's avatar
Andreas Marek committed
98
   real(kind=C_DATATYPE_KIND), allocatable                            :: q_real(:,:)
99
#endif
Andreas Marek's avatar
Andreas Marek committed
100
101
102
103
104
105
106
107
108
109
110
   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
111
   logical                                                            :: do_useGPU, do_useGPU_bandred, &
112
                                                                         do_useGPU_tridiag_band, do_useGPU_solve_tridi, &
113
114
                                                                         do_useGPU_trans_ev_tridi_to_band, &
                                                                         do_useGPU_trans_ev_band_to_full
Andreas Marek's avatar
Andreas Marek committed
115
116
117
118
119
120
121
   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
122
                                                                         mpi_comm_all, check_pd, error
Andreas Marek's avatar
Andreas Marek committed
123
124
125

    logical                                                           :: do_bandred, do_tridiag, do_solve_tridi,  &
                                                                         do_trans_to_band, do_trans_to_full
126

Andreas Marek's avatar
Andreas Marek committed
127
    integer(kind=ik)                                                  :: nrThreads
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
#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

145
    call obj%timer%start("elpa_solve_evp_&
146
    &MATH_DATATYPE&
147
148
149
    &_2stage_&
    &PRECISION&
    &")
150

Andreas Marek's avatar
Andreas Marek committed
151
152

#ifdef WITH_OPENMP
Andreas Marek's avatar
Andreas Marek committed
153
154
155
    !nrThreads = omp_get_max_threads()
    call obj%get("omp_threads",nrThreads,error)
    call omp_set_num_threads(nrThreads)
Andreas Marek's avatar
Andreas Marek committed
156
157
158
159
#else
    nrThreads = 1
#endif

Andreas Marek's avatar
Andreas Marek committed
160
161
    success = .true.

Andreas Marek's avatar
Andreas Marek committed
162
    if (present(q)) then
163
      obj%eigenvalues_only = .false.
Andreas Marek's avatar
Andreas Marek committed
164
    else
165
      obj%eigenvalues_only = .true.
Andreas Marek's avatar
Andreas Marek committed
166
167
    endif

168
169
170
171
172
173
174
    na         = obj%na
    nev        = obj%nev
    lda        = obj%local_nrows
    ldq        = obj%local_nrows
    nblk       = obj%nblk
    matrixCols = obj%local_ncols

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
    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
201
202
203
204
205
206
207
   ! 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))
208
#endif
Andreas Marek's avatar
Andreas Marek committed
209
     if (.not.(obj%eigenvalues_only)) then
210
       q(1,1) = ONE
Andreas Marek's avatar
Andreas Marek committed
211
212
213
214
215
216
217
218
219
     endif
     call obj%timer%stop("elpa_solve_evp_&
     &MATH_DATATYPE&
     &_2stage_&
     &PRECISION&
     &")
     return
   endif

220
221
222
223
224
   if (nev == 0) then
     nev = 1
     obj%eigenvalues_only = .true.
   endif

225
    call obj%get(KERNEL_STRING,kernel,error)
Andreas Marek's avatar
Andreas Marek committed
226
227
228
229
    if (error .ne. ELPA_OK) then
      print *,"Problem getting option. Aborting..."
      stop
    endif
230
231

    ! GPU settings
Andreas Marek's avatar
Andreas Marek committed
232
233
234
235
236
    call obj%get("gpu", gpu,error)
    if (error .ne. ELPA_OK) then
      print *,"Problem getting option. Aborting..."
      stop
    endif
237

238
    useGPU = (gpu == 1)
239

240
    do_useGPU = .false.
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
    if (useGPU) then
      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
    endif

259
260
261
262
263
264
265
266
267
268
269
270
271
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
    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

306
    ! check consistency between request for GPUs and defined kernel
307
308
309

    if (do_useGPU_trans_ev_tridi_to_band) then
      if (kernel .ne. GPU_KERNEL) then
310
        write(error_unit,*) "ELPA: Warning, GPU usage has been requested but compute kernel is defined as non-GPU!"
311
        write(error_unit,*) "The compute kernel will be executed on CPUs!"
312
        do_useGPU_trans_ev_tridi_to_band = .false.
313
      else if (nblk .ne. 128) then
314
315
316
317
        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
318
319
      endif
    endif
320

321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
    ! 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
341
342
      endif
    endif
343

344

345
#if REALCASE == 1
346
347
348
349
350
351
352
353
354
#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
355
356
357
358
359
360
361
362
    ! 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
363
364
#endif

365
366
367
368
#endif


#if REALCASE == 1
Andreas Marek's avatar
Andreas Marek committed
369
370
371
372
373
    call obj%get("qr",qr,error)
    if (error .ne. ELPA_OK) then
      print *,"Problem getting option. Aborting..."
      stop
    endif
374
    if (qr .eq. 1) then
375
376
377
378
379
380
      useQR = .true.
    else
      useQR = .false.
    endif

#endif
381

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



#if REALCASE == 1
    useQRActual = .false.
    ! set usage of qr decomposition via API call
394
395
    if (useQR) useQRActual = .true.
    if (.not.(useQR)) useQRACtual = .false.
396
397
398
399
400
401
402
403
404
405
406
407
408
409

    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
410

411
    if (.not. obj%eigenvalues_only) then
Andreas Marek's avatar
Andreas Marek committed
412
413
414
415
416
417
      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

418
419
420
421
422
423
424
425
426
427
428
429
430

    ! 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
431
    if (obj%is_set("bandwidth") == 1) then
Andreas Marek's avatar
Andreas Marek committed
432
      call obj%get("bandwidth",nbw,error)
433
      if (nbw == 0) then
434
        if (wantDebug) then
435
          write(error_unit,*) "Specified bandwidth = 0; ELPA refuses to solve the eigenvalue problem ", &
Andreas Marek's avatar
Andreas Marek committed
436
437
                              "for a diagonal matrix! This is too simple"
          endif
438
        print *, "Specified bandwidth = 0; ELPA refuses to solve the eigenvalue problem ", &
Andreas Marek's avatar
Andreas Marek committed
439
                 "for a diagonal matrix! This is too simple"
440
441
442
        success = .false.
        return
      endif
443
444
      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
445
        ! such that effective bandwidth is a multiply of nblk. which is a prerequiste for ELPA
446
447
448
        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
449
        if (nbw .gt. na) then
450
451
          if (wantDebug) then
            write(error_unit,*) "Specified bandwidth ",nbw," leads internaly to a computed bandwidth ", &
Andreas Marek's avatar
Andreas Marek committed
452
453
454
                                "which is larger than the matrix size ",na," ! ELPA will abort! Try to", &
                                "solve your problem by not specifing a bandwidth"
          endif
455
          print *, "Specified bandwidth ",nbw," leads internaly to a computed bandwidth ", &
Andreas Marek's avatar
Andreas Marek committed
456
457
                                "which is larger than the matrix size ",na," ! ELPA will abort! Try to", &
                                "solve your problem by not specifing a bandwidth"
458
459
          success = .false.
          return
Andreas Marek's avatar
Andreas Marek committed
460
        endif
461
      endif
462
463
464
465
466
      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
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487

      ! 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_&
488
489
490
491
        &MATH_DATATYPE&
        &_2stage_&
        &PRECISION&
        &" // ": error when allocating tmat "//errorMessage
492
493
494
        stop 1
      endif

495
496
497
498
499
500
501
502
503
504
      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")
505
506
507
508
509
      ! Reduction full -> band
      call bandred_&
      &MATH_DATATYPE&
      &_&
      &PRECISION &
510
      (obj, na, a, &
511
      a_dev, lda, nblk, nbw, matrixCols, num_blocks, mpi_comm_rows, mpi_comm_cols, tmat, &
512
      tmat_dev,  wantDebug, do_useGPU_bandred, success, &
513
#if REALCASE == 1
Andreas Marek's avatar
Andreas Marek committed
514
      useQRActual, &
515
#endif
Andreas Marek's avatar
Andreas Marek committed
516
       nrThreads)
517
      call obj%timer%stop("bandred")
518
      if (.not.(success)) return
519
520
    endif

521
522

     ! Reduction band -> tridiagonal
523
524
525
526
527
528
529
530
531
     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
532

533
534
       call obj%timer%start("tridiag")
       call tridiag_band_&
535
       &MATH_DATATYPE&
536
537
       &_&
       &PRECISION&
538
       (obj, na, nbw, nblk, a, a_dev, lda, ev, e, matrixCols, hh_trans, mpi_comm_rows, mpi_comm_cols, mpi_comm_all, &
539
       do_useGPU_tridiag_band, wantDebug, nrThreads)
540
541

#ifdef WITH_MPI
542
543
544
545
       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")
546
#endif /* WITH_MPI */
547
548
       call obj%timer%stop("tridiag")
     endif ! do_tridiag
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564

#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
565
566
567
568
569
     if (do_solve_tridi) then
       call obj%timer%start("solve")
       call solve_tridi_&
       &PRECISION &
       (obj, na, nev, ev, e, &
570
#if REALCASE == 1
571
       q_actual, ldq,   &
572
573
#endif
#if COMPLEXCASE == 1
574
       q_real, ubound(q_real,dim=1), &
575
#endif
576
       nblk, matrixCols, mpi_comm_rows, mpi_comm_cols, do_useGPU_solve_tridi, wantDebug, success, nrThreads)
577
578
579
       call obj%timer%stop("solve")
       if (.not.(success)) return
     endif ! do_solve_tridi
580
581
582
583
584
585
586
587
588

     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

589
     if (obj%eigenvalues_only) then
Andreas Marek's avatar
Andreas Marek committed
590
591
592
593
       do_trans_to_band = .false.
       do_trans_to_full = .false.
     else

Andreas Marek's avatar
Andreas Marek committed
594
595
596
597
598
       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
599
600
601
602
603
604
605
606
607
608
609
       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
610
         else
Andreas Marek's avatar
Andreas Marek committed
611
612
613
614
615
           do_trans_to_band = .false.
           do_trans_to_full = .false.
         endif
       endif
     endif ! eigenvalues only
616

617
     if (do_trans_to_band) then
618
#if COMPLEXCASE == 1
Andreas Marek's avatar
Andreas Marek committed
619
       ! q must be given thats why from here on we can use q and not q_actual
620

Andreas Marek's avatar
Andreas Marek committed
621
       q(1:l_rows,1:l_cols_nev) = q_real(1:l_rows,1:l_cols_nev)
622

Andreas Marek's avatar
Andreas Marek committed
623
624
625
626
627
628
629
630
       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
631

Andreas Marek's avatar
Andreas Marek committed
632
633
       ! Backtransform stage 1
       call obj%timer%start("trans_ev_to_band")
634

Andreas Marek's avatar
Andreas Marek committed
635
       call trans_ev_tridi_to_band_&
636
       &MATH_DATATYPE&
Andreas Marek's avatar
Andreas Marek committed
637
638
639
640
       &_&
       &PRECISION &
       (obj, na, nev, nblk, nbw, q, &
       q_dev, &
641
       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
642
       nrThreads, success=success, kernel=kernel)
Andreas Marek's avatar
Andreas Marek committed
643
       call obj%timer%stop("trans_ev_to_band")
644

Andreas Marek's avatar
Andreas Marek committed
645
       if (.not.(success)) return
646

Andreas Marek's avatar
Andreas Marek committed
647
648
649
650
651
652
653
654
       ! 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
655
       endif
656
     endif ! do_trans_to_band
657

658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
     ! 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)
       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

680
     if (do_trans_to_full) then
Andreas Marek's avatar
Andreas Marek committed
681
       call obj%timer%start("trans_ev_to_full")
682
       if ( (do_useGPU_trans_ev_band_to_full) .and. .not.(do_useGPU_trans_ev_tridi_to_band) ) then
683
684
         ! 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
685

686
687
         successCUDA = cuda_memcpy(q_dev, loc(q), ldq*matrixCols* size_of_datatype, cudaMemcpyHostToDevice)
       endif
Andreas Marek's avatar
Andreas Marek committed
688

689
       ! Backtransform stage 2
Andreas Marek's avatar
Andreas Marek committed
690

691
692
693
694
695
696
697
       call trans_ev_band_to_full_&
       &MATH_DATATYPE&
       &_&
       &PRECISION &
       (obj, na, nev, nblk, nbw, a, &
       a_dev, lda, tmat, tmat_dev,  q,  &
       q_dev, &
698
       ldq, matrixCols, num_blocks, mpi_comm_rows, mpi_comm_cols, do_useGPU_trans_ev_band_to_full &
699
#if REALCASE == 1
700
       , useQRActual  &
701
#endif
702
       )
703

704

705
706
707
708
709
710
711
712
       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
713
       call obj%timer%stop("trans_ev_to_full")
714
     endif ! do_trans_to_full
Andreas Marek's avatar
Andreas Marek committed
715

716
     if (obj%eigenvalues_only) then
Andreas Marek's avatar
Andreas Marek committed
717
       deallocate(q_dummy, stat=istat, errmsg=errorMessage)
718
719
       if (istat .ne. 0) then
         print *,"solve_evp_&
Andreas Marek's avatar
Andreas Marek committed
720
721
722
723
         &MATH_DATATYPE&
         &_1stage_&
         &PRECISION&
         &" // ": error when deallocating q_dummy "//errorMessage
724
725
726
727
         stop 1
       endif
     endif

728
     call obj%timer%stop("elpa_solve_evp_&
729
     &MATH_DATATYPE&
730
731
732
     &_2stage_&
    &PRECISION&
    &")
733
734
735
736
737
738
1    format(a,f10.3)

   end function elpa_solve_evp_&
   &MATH_DATATYPE&
   &_2stage_&
   &PRECISION&
739
   &_impl
740
741

! vim: syntax=fortran