elpa2.F90 56.2 KB
Newer Older
1
!   This file is part of ELPA.
2 3 4 5
!
!    The ELPA library was originally created by the ELPA consortium,
!    consisting of the following organizations:
!
6 7
!    - Max Planck Computing and Data Facility (MPCDF), fomerly known as
!      Rechenzentrum Garching der Max-Planck-Gesellschaft (RZG),
8 9 10 11 12 13 14 15 16 17
!    - 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 Naturwissenschaftrn,
!      Leipzig, Abt. Komplexe Strukutren in Biologie und Kognition,
!      and
!    - IBM Deutschland GmbH
!
18
!    This particular source code file contains additions, changes and
Andreas Marek's avatar
Andreas Marek committed
19
!    enhancements authored by Intel Corporation which is not part of
20
!    the ELPA consortium.
21 22
!
!    More information can be found here:
23
!    http://elpa.mpcdf.mpg.de/
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 57 58 59 60 61 62 63
!
!    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".



! ELPA2 -- 2-stage solver for ELPA
!
! 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".


#include "config-f90.h"
64
!> \brief Fortran module which provides the routines to use the two-stage ELPA solver
65 66 67 68
module ELPA2

! Version 1.1.2, 2011-02-21

69
  use elpa_utilities
Andreas Marek's avatar
Andreas Marek committed
70
  use elpa1_compute
71
  use elpa1, only : elpa_print_times, time_evp_back, time_evp_fwd, time_evp_solve
72
  use elpa2_utilities
73
  use elpa2_compute
74 75
  use elpa_pdgeqrf

76
  use iso_c_binding
77 78 79 80 81

#ifdef WITH_GPU_VERSION
!  use cuda_routines
!  use cuda_c_kernel
!  use iso_c_binding
82
#endif
83
  use elpa_mpi
84

85 86 87 88 89 90
  implicit none

  PRIVATE ! By default, all routines contained are private

  ! The following routines are public:

91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109
  public :: solve_evp_real_2stage_double
  public :: solve_evp_complex_2stage_double

  interface solve_evp_real_2stage
    module procedure solve_evp_real_2stage_double
  end interface

  interface solve_evp_complex_2stage
    module procedure solve_evp_complex_2stage_double
  end interface

#ifdef WANT_SINGLE_PRECISION_REAL
  public :: solve_evp_real_2stage_single
#endif

#ifdef WANT_SINGLE_PRECISION_COMPLEX
  public :: solve_evp_complex_2stage_single
#endif

110 111 112

!******
contains
113
!-------------------------------------------------------------------------------
114
!>  \brief solve_evp_real_2stage_double: Fortran function to solve the double-precision real eigenvalue problem with a 2 stage approach
115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151
!>
!>  Parameters
!>
!>  \param na                                   Order of matrix a
!>
!>  \param nev                                  Number of eigenvalues needed
!>
!>  \param a(lda,matrixCols)                    Distributed matrix for which eigenvalues are to be computed.
!>                                              Distribution is like in Scalapack.
!>                                              The full matrix must be set (not only one half like in scalapack).
!>                                              Destroyed on exit (upper and lower half).
!>
!>  \param lda                                  Leading dimension of a
!>
!>  \param ev(na)                               On output: eigenvalues of a, every processor gets the complete set
!>
!>  \param q(ldq,matrixCols)                    On output: Eigenvectors of a
!>                                              Distribution is like in Scalapack.
!>                                              Must be always dimensioned to the full size (corresponding to (na,na))
!>                                              even if only a part of the eigenvalues is needed.
!>
!>  \param ldq                                  Leading dimension of q
!>
!>  \param nblk                                 blocksize of cyclic distribution, must be the same in both directions!
!>
!>  \param matrixCols                           local columns of matrix a and q
!>
!>  \param mpi_comm_rows                        MPI communicator for rows
!>  \param mpi_comm_cols                        MPI communicator for columns
!>  \param mpi_comm_all                         MPI communicator for the total processor set
!>
!>  \param THIS_REAL_ELPA_KERNEL_API (optional) specify used ELPA2 kernel via API
!>
!>  \param use_qr (optional)                    use QR decomposition
!>
!>  \result success                             logical, false if error occured
!-------------------------------------------------------------------------------
152

153 154 155 156 157 158 159 160 161 162
#define DOUBLE_PRECISION_REAL

#ifdef DOUBLE_PRECISION_REAL
  function solve_evp_real_2stage_double(na, nev, a, lda, ev, q, ldq, nblk,        &
                               matrixCols,                               &
                                 mpi_comm_rows, mpi_comm_cols,           &
                                 mpi_comm_all, THIS_REAL_ELPA_KERNEL_API,&
                                 useQR) result(success)
#else
  function solve_evp_real_2stage_single(na, nev, a, lda, ev, q, ldq, nblk,        &
163
                               matrixCols,                               &
164 165 166
                                 mpi_comm_rows, mpi_comm_cols,           &
                                 mpi_comm_all, THIS_REAL_ELPA_KERNEL_API,&
                                 useQR) result(success)
167
#endif
168

169

170
#ifdef HAVE_DETAILED_TIMINGS
171
    use timings
172
#endif
173

Andreas Marek's avatar
Andreas Marek committed
174
   use precision
175 176
   use cuda_functions
   use mod_check_for_gpu
177
   use iso_c_binding
178
   implicit none
Andreas Marek's avatar
Andreas Marek committed
179 180 181 182 183 184 185 186
   logical, intent(in), optional          :: useQR
   logical                                :: useQRActual, useQREnvironment
   integer(kind=ik), intent(in), optional :: THIS_REAL_ELPA_KERNEL_API
   integer(kind=ik)                       :: THIS_REAL_ELPA_KERNEL

   integer(kind=ik), intent(in)           :: na, nev, lda, ldq, matrixCols, mpi_comm_rows, &
                                             mpi_comm_cols, mpi_comm_all
   integer(kind=ik), intent(in)           :: nblk
187
   real(kind=rk8), intent(inout)           :: a(lda,matrixCols), ev(na), q(ldq,matrixCols)
188 189
   ! was
   ! real a(lda,*), q(ldq,*)
190
   real(kind=rk8), allocatable             :: hh_trans_real(:,:)
Andreas Marek's avatar
Andreas Marek committed
191 192 193

   integer(kind=ik)                       :: my_pe, n_pes, my_prow, my_pcol, np_rows, np_cols, mpierr
   integer(kind=ik)                       :: nbw, num_blocks
194
   real(kind=rk8), allocatable             :: tmat(:,:,:), e(:)
195
   real(kind=c_double)                    :: ttt0, ttt1, ttts  ! MPI_WTIME always needs double
Andreas Marek's avatar
Andreas Marek committed
196 197 198 199
   integer(kind=ik)                       :: i
   logical                                :: success
   logical, save                          :: firstCall = .true.
   logical                                :: wantDebug
200 201 202 203
   integer(kind=ik)                       :: istat
   character(200)                         :: errorMessage
   logical                                :: useGPU
   integer(kind=ik)                       :: numberOfGPUDevices
Andreas Marek's avatar
Andreas Marek committed
204

205
#ifdef HAVE_DETAILED_TIMINGS
206
    call timer%start("solve_evp_real_2stage_double")
207
#endif
208

209 210
    call mpi_comm_rank(mpi_comm_all,my_pe,mpierr)
    call mpi_comm_size(mpi_comm_all,n_pes,mpierr)
211

212 213 214 215
    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)
216

217

218 219 220 221 222 223
    wantDebug = .false.
    if (firstCall) then
      ! are debug messages desired?
      wantDebug = debug_messages_via_environment_variable()
      firstCall = .false.
    endif
224

225
    success = .true.
226

227 228
    useQRActual = .false.
    useGPU      = .false.
229

230 231 232 233 234
    ! set usage of qr decomposition via API call
    if (present(useQR)) then
      if (useQR) useQRActual = .true.
        if (.not.(useQR)) useQRACtual = .false.
    endif
235

236 237 238 239
    ! overwrite this with environment variable settings
    if (qr_decomposition_via_environment_variable(useQREnvironment)) then
      useQRActual = useQREnvironment
    endif
240

241 242 243 244 245 246 247 248 249 250
    if (useQRActual) then
      if (mod(na,nblk) .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
251

252

253 254 255 256
    if (present(THIS_REAL_ELPA_KERNEL_API)) then
      ! user defined kernel via the optional argument in the API call
      THIS_REAL_ELPA_KERNEL = THIS_REAL_ELPA_KERNEL_API
    else
257

258 259 260 261
      ! if kernel is not choosen via api
      ! check whether set by environment variable
      THIS_REAL_ELPA_KERNEL = get_actual_real_kernel()
    endif
Andreas Marek's avatar
Andreas Marek committed
262

263 264 265 266 267 268 269 270 271 272 273 274 275 276
    ! check whether choosen kernel is allowed
    if (check_allowed_real_kernels(THIS_REAL_ELPA_KERNEL)) then

      if (my_pe == 0) then
        write(error_unit,*) " "
        write(error_unit,*) "The choosen kernel ",REAL_ELPA_KERNEL_NAMES(THIS_REAL_ELPA_KERNEL)
        write(error_unit,*) "is not in the list of the allowed kernels!"
        write(error_unit,*) " "
        write(error_unit,*) "Allowed kernels are:"
        do i=1,size(REAL_ELPA_KERNEL_NAMES(:))
          if (AVAILABLE_REAL_ELPA_KERNELS(i) .ne. 0) then
            write(error_unit,*) REAL_ELPA_KERNEL_NAMES(i)
          endif
        enddo
Andreas Marek's avatar
Andreas Marek committed
277

278 279 280 281
        write(error_unit,*) " "
        write(error_unit,*) "The defaul kernel REAL_ELPA_KERNEL_GENERIC will be used !"
      endif
      THIS_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_GENERIC
Andreas Marek's avatar
Andreas Marek committed
282

283 284 285
    endif

    if (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_GPU) then
286
      if (check_for_gpu(my_pe,numberOfGPUDevices, wantDebug=wantDebug)) then
287 288 289 290 291 292
        useGPU = .true.
      endif
      if (nblk .ne. 128) then
        print *,"At the moment GPU version needs blocksize 128"
        stop
      endif
293

294 295 296 297 298 299 300
      ! set the neccessary parameters
      cudaMemcpyHostToDevice   = cuda_memcpyHostToDevice()
      cudaMemcpyDeviceToHost   = cuda_memcpyDeviceToHost()
      cudaMemcpyDeviceToDevice = cuda_memcpyDeviceToDevice()
      cudaHostRegisterPortable = cuda_hostRegisterPortable()
      cudaHostRegisterMapped   = cuda_hostRegisterMapped()
    endif
301

302 303
    ! some temporarilly checks until single precision works with all kernels
#ifndef DOUBLE_PRECISION_REAL
304 305 306 307 308 309 310 311
    if ( (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_GENERIC) .or. &
         (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_GENERIC_SIMPLE) .or. &
         (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_SSE) .or. &
         (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_GPU) ) then
    else
      print *,"At the moment single precision only works with the generic kernels"
      stop
    endif
312 313
#endif

314
    ! Choose bandwidth, must be a multiple of nblk, set to a value >= 32
315 316 317 318
    ! 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
319 320 321
    if (useGPU) then
      nbw = nblk
    else
322
      nbw = (63/nblk+1)*nblk
323
    endif
324

325
    num_blocks = (na-1)/nbw + 1
326

327 328 329 330 331
    allocate(tmat(nbw,nbw,num_blocks), stat=istat, errmsg=errorMessage)
    if (istat .ne. 0) then
      print *,"solve_evp_real_2stage: error when allocating tmat "//errorMessage
      stop
    endif
332

333
    ! Reduction full -> band
334

335 336
    ttt0 = MPI_Wtime()
    ttts = ttt0
337 338
#ifdef DOUBLE_PRECISION_REAL
    call bandred_real_double(na, a, lda, nblk, nbw, matrixCols, num_blocks, mpi_comm_rows, mpi_comm_cols, &
339
                      tmat, wantDebug, useGPU, success, useQRActual)
340 341 342 343
#else
    call bandred_real_single(na, a, lda, nblk, nbw, matrixCols, num_blocks, mpi_comm_rows, mpi_comm_cols, &
                      tmat, wantDebug, useGPU, success, useQRActual)
#endif
344 345 346 347
    if (.not.(success)) return
    ttt1 = MPI_Wtime()
    if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
       write(error_unit,*) 'Time bandred_real               :',ttt1-ttt0
348

349
     ! Reduction band -> tridiagonal
350

351 352 353 354 355
     allocate(e(na), stat=istat, errmsg=errorMessage)
     if (istat .ne. 0) then
       print *,"solve_evp_real_2stage: error when allocating e "//errorMessage
       stop
     endif
356

357
     ttt0 = MPI_Wtime()
358 359
#ifdef DOUBLE_PRECISION_REAL
     call tridiag_band_real_double(na, nbw, nblk, a, lda, ev, e, matrixCols, hh_trans_real, &
360
                          mpi_comm_rows, mpi_comm_cols, mpi_comm_all)
361 362 363 364
#else
     call tridiag_band_real_single(na, nbw, nblk, a, lda, ev, e, matrixCols, hh_trans_real, &
                          mpi_comm_rows, mpi_comm_cols, mpi_comm_all)
#endif
365

366 367 368
     ttt1 = MPI_Wtime()
     if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
       write(error_unit,*) 'Time tridiag_band_real          :',ttt1-ttt0
369

370
#ifdef WITH_MPI
371

372
#ifdef DOUBLE_PRECISION_REAL
373 374
     call mpi_bcast(ev,na,MPI_REAL8,0,mpi_comm_all,mpierr)
     call mpi_bcast(e,na,MPI_REAL8,0,mpi_comm_all,mpierr)
375 376 377 378
#else
     call mpi_bcast(ev,na,MPI_REAL4,0,mpi_comm_all,mpierr)
     call mpi_bcast(e,na,MPI_REAL4,0,mpi_comm_all,mpierr)
#endif
379

380
#endif /* WITH_MPI */
381 382
     ttt1 = MPI_Wtime()
     time_evp_fwd = ttt1-ttts
383

384
     ! Solve tridiagonal system
385

386
     ttt0 = MPI_Wtime()
387 388
#ifdef DOUBLE_PRECISION_REAL
     call solve_tridi_double(na, nev, ev, e, q, ldq, nblk, matrixCols, mpi_comm_rows,  &
389
                      mpi_comm_cols, wantDebug, success)
390 391 392 393
#else
     call solve_tridi_single(na, nev, ev, e, q, ldq, nblk, matrixCols, mpi_comm_rows,  &
                      mpi_comm_cols, wantDebug, success)
#endif
394 395 396 397
     if (.not.(success)) return

     ttt1 = MPI_Wtime()
     if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
398
     write(error_unit,*) 'Time solve_tridi                :',ttt1-ttt0
399 400
     time_evp_solve = ttt1-ttt0
     ttts = ttt1
401

402 403 404 405 406 407 408 409
     deallocate(e, stat=istat, errmsg=errorMessage)
     if (istat .ne. 0) then
       print *,"solve_evp_real_2stage: error when deallocating e "//errorMessage
       stop
     endif
     ! Backtransform stage 1

     ttt0 = MPI_Wtime()
410 411 412 413 414 415
#ifdef DOUBLE_PRECISION_REAL
     call trans_ev_tridi_to_band_real_double(na, nev, nblk, nbw, q, ldq, matrixCols, hh_trans_real, &
                                    mpi_comm_rows, mpi_comm_cols, wantDebug, useGPU, success,      &
                                    THIS_REAL_ELPA_KERNEL)
#else
     call trans_ev_tridi_to_band_real_single(na, nev, nblk, nbw, q, ldq, matrixCols, hh_trans_real, &
416
                                    mpi_comm_rows, mpi_comm_cols, wantDebug, useGPU, success,      &
417
                                    THIS_REAL_ELPA_KERNEL)
418
#endif
419

420 421 422 423
     if (.not.(success)) return
     ttt1 = MPI_Wtime()
     if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
       write(error_unit,*) 'Time trans_ev_tridi_to_band_real:',ttt1-ttt0
424

425 426 427 428 429 430
     ! We can now deallocate the stored householder vectors
     deallocate(hh_trans_real, stat=istat, errmsg=errorMessage)
     if (istat .ne. 0) then
       print *,"solve_evp_real_2stage: error when deallocating hh_trans_real "//errorMessage
       stop
     endif
431 432


433 434 435
     ! Backtransform stage 2
     print *,"useGPU== ",useGPU
     ttt0 = MPI_Wtime()
436 437 438 439 440
#ifdef DOUBLE_PRECISION_REAL
     call trans_ev_band_to_full_real_double(na, nev, nblk, nbw, a, lda, tmat, q, ldq, matrixCols, num_blocks, mpi_comm_rows, &
                                     mpi_comm_cols, useGPU, useQRActual)
#else
     call trans_ev_band_to_full_real_single(na, nev, nblk, nbw, a, lda, tmat, q, ldq, matrixCols, num_blocks, mpi_comm_rows, &
441
                                     mpi_comm_cols, useGPU, useQRActual)
442
#endif
443

444 445 446 447 448 449 450 451 452 453
     ttt1 = MPI_Wtime()
     if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
       write(error_unit,*) 'Time trans_ev_band_to_full_real :',ttt1-ttt0
     time_evp_back = ttt1-ttts

     deallocate(tmat, stat=istat, errmsg=errorMessage)
     if (istat .ne. 0) then
       print *,"solve_evp_real_2stage: error when deallocating tmat"//errorMessage
       stop
     endif
454

455
#ifdef HAVE_DETAILED_TIMINGS
456
     call timer%stop("solve_evp_real_2stage_double")
457
#endif
458
1    format(a,f10.3)
459

460 461 462 463 464
#ifdef DOUBLE_PRECISION_REAL
   end function solve_evp_real_2stage_double
#else
   end function solve_evp_real_2stage_single
#endif
465

466 467 468 469
#ifdef WANT_SINGLE_PRECISION_REAL
#undef DOUBLE_PRECISION_REAL
!-------------------------------------------------------------------------------
!>  \brief solve_evp_real_2stage_single: Fortran function to solve the single-precision real eigenvalue problem with a 2 stage approach
470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502
!>
!>  Parameters
!>
!>  \param na                                   Order of matrix a
!>
!>  \param nev                                  Number of eigenvalues needed
!>
!>  \param a(lda,matrixCols)                    Distributed matrix for which eigenvalues are to be computed.
!>                                              Distribution is like in Scalapack.
!>                                              The full matrix must be set (not only one half like in scalapack).
!>                                              Destroyed on exit (upper and lower half).
!>
!>  \param lda                                  Leading dimension of a
!>
!>  \param ev(na)                               On output: eigenvalues of a, every processor gets the complete set
!>
!>  \param q(ldq,matrixCols)                    On output: Eigenvectors of a
!>                                              Distribution is like in Scalapack.
!>                                              Must be always dimensioned to the full size (corresponding to (na,na))
!>                                              even if only a part of the eigenvalues is needed.
!>
!>  \param ldq                                  Leading dimension of q
!>
!>  \param nblk                                 blocksize of cyclic distribution, must be the same in both directions!
!>
!>  \param matrixCols                           local columns of matrix a and q
!>
!>  \param mpi_comm_rows                        MPI communicator for rows
!>  \param mpi_comm_cols                        MPI communicator for columns
!>  \param mpi_comm_all                         MPI communicator for the total processor set
!>
!>  \param THIS_REAL_ELPA_KERNEL_API (optional) specify used ELPA2 kernel via API
!>
503 504
!>  \param use_qr (optional)                    use QR decomposition
!>
505
!>  \result success                             logical, false if error occured
506
!-------------------------------------------------------------------------------
507 508 509 510 511 512 513 514 515 516 517 518 519 520

#ifdef DOUBLE_PRECISION_REAL
  function solve_evp_real_2stage_double(na, nev, a, lda, ev, q, ldq, nblk,        &
                               matrixCols,                               &
                                 mpi_comm_rows, mpi_comm_cols,           &
                                 mpi_comm_all, THIS_REAL_ELPA_KERNEL_API,&
                                 useQR) result(success)
#else
  function solve_evp_real_2stage_single(na, nev, a, lda, ev, q, ldq, nblk,        &
                               matrixCols,                               &
                                 mpi_comm_rows, mpi_comm_cols,           &
                                 mpi_comm_all, THIS_REAL_ELPA_KERNEL_API,&
                                 useQR) result(success)
#endif
521

522
#ifdef HAVE_DETAILED_TIMINGS
523
    use timings
524
#endif
525

Andreas Marek's avatar
Andreas Marek committed
526
   use precision
527 528
   use cuda_functions
   use mod_check_for_gpu
529
   use iso_c_binding
530
   implicit none
531 532 533 534 535 536 537 538 539
   logical, intent(in), optional          :: useQR
   logical                                :: useQRActual, useQREnvironment
   integer(kind=ik), intent(in), optional :: THIS_REAL_ELPA_KERNEL_API
   integer(kind=ik)                       :: THIS_REAL_ELPA_KERNEL

   integer(kind=ik), intent(in)           :: na, nev, lda, ldq, matrixCols, mpi_comm_rows, &
                                             mpi_comm_cols, mpi_comm_all
   integer(kind=ik), intent(in)           :: nblk
   real(kind=rk4), intent(inout)           :: a(lda,matrixCols), ev(na), q(ldq,matrixCols)
540
   ! was
541 542
   ! real a(lda,*), q(ldq,*)
   real(kind=rk4), allocatable             :: hh_trans_real(:,:)
Andreas Marek's avatar
Andreas Marek committed
543

544 545 546
   integer(kind=ik)                       :: my_pe, n_pes, my_prow, my_pcol, np_rows, np_cols, mpierr
   integer(kind=ik)                       :: nbw, num_blocks
   real(kind=rk4), allocatable             :: tmat(:,:,:), e(:)
547
   real(kind=c_double)                    :: ttt0, ttt1, ttts  ! MPI_WTIME always needs double
Andreas Marek's avatar
Andreas Marek committed
548
   integer(kind=ik)                       :: i
549
   logical                                :: success
Andreas Marek's avatar
Andreas Marek committed
550
   logical, save                          :: firstCall = .true.
551
   logical                                :: wantDebug
552 553 554 555
   integer(kind=ik)                       :: istat
   character(200)                         :: errorMessage
   logical                                :: useGPU
   integer(kind=ik)                       :: numberOfGPUDevices
Andreas Marek's avatar
Andreas Marek committed
556

557
#ifdef HAVE_DETAILED_TIMINGS
558
    call timer%start("solve_evp_real_2stage_single")
559
#endif
560

561 562 563 564 565 566 567 568 569 570 571 572 573 574
    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)

    wantDebug = .false.
    if (firstCall) then
      ! are debug messages desired?
      wantDebug = debug_messages_via_environment_variable()
      firstCall = .false.
    endif
575

576
    success = .true.
577

578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604
    useQRActual = .false.
    useGPU      = .false.

    ! set usage of qr decomposition via API call
    if (present(useQR)) then
      if (useQR) useQRActual = .true.
        if (.not.(useQR)) useQRACtual = .false.
    endif

    ! overwrite this with environment variable settings
    if (qr_decomposition_via_environment_variable(useQREnvironment)) then
      useQRActual = useQREnvironment
    endif

    if (useQRActual) then
      if (mod(na,nblk) .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


    if (present(THIS_REAL_ELPA_KERNEL_API)) then
605
      ! user defined kernel via the optional argument in the API call
606
      THIS_REAL_ELPA_KERNEL = THIS_REAL_ELPA_KERNEL_API
607
    else
608

609 610
      ! if kernel is not choosen via api
      ! check whether set by environment variable
611
      THIS_REAL_ELPA_KERNEL = get_actual_real_kernel()
612
    endif
613

614
    ! check whether choosen kernel is allowed
615
    if (check_allowed_real_kernels(THIS_REAL_ELPA_KERNEL)) then
616 617 618

      if (my_pe == 0) then
        write(error_unit,*) " "
619
        write(error_unit,*) "The choosen kernel ",REAL_ELPA_KERNEL_NAMES(THIS_REAL_ELPA_KERNEL)
620 621 622
        write(error_unit,*) "is not in the list of the allowed kernels!"
        write(error_unit,*) " "
        write(error_unit,*) "Allowed kernels are:"
623 624 625
        do i=1,size(REAL_ELPA_KERNEL_NAMES(:))
          if (AVAILABLE_REAL_ELPA_KERNELS(i) .ne. 0) then
            write(error_unit,*) REAL_ELPA_KERNEL_NAMES(i)
626 627
          endif
        enddo
628

629
        write(error_unit,*) " "
630
        write(error_unit,*) "The defaul kernel REAL_ELPA_KERNEL_GENERIC will be used !"
631
      endif
632 633
      THIS_REAL_ELPA_KERNEL = REAL_ELPA_KERNEL_GENERIC

634
    endif
635

636 637 638
    if (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_GPU) then
      if (check_for_gpu(my_pe,numberOfGPUDevices, wantDebug=wantDebug)) then
        useGPU = .true.
639 640 641 642 643
      endif
      if (nblk .ne. 128) then
        print *,"At the moment GPU version needs blocksize 128"
        stop
      endif
644
    ! some temporarilly checks until single precision works with all kernels
645

646 647 648 649 650 651 652
      ! set the neccessary parameters
      cudaMemcpyHostToDevice   = cuda_memcpyHostToDevice()
      cudaMemcpyDeviceToHost   = cuda_memcpyDeviceToHost()
      cudaMemcpyDeviceToDevice = cuda_memcpyDeviceToDevice()
      cudaHostRegisterPortable = cuda_hostRegisterPortable()
      cudaHostRegisterMapped   = cuda_hostRegisterMapped()
    endif
653

654 655 656 657 658 659 660 661 662 663
#ifndef DOUBLE_PRECISION_REAL
    if ( (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_GENERIC) .or. &
         (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_GENERIC_SIMPLE) .or. &
         (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_SSE) .or. &
         (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_GPU) ) then
    else
      print *,"At the moment single precision only works with the generic kernels"
      stop
    endif
#endif
664
    ! Choose bandwidth, must be a multiple of nblk, set to a value >= 32
665 666 667 668 669 670 671 672 673
    ! 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 (useGPU) then
      nbw = nblk
    else
      nbw = (63/nblk+1)*nblk
    endif
674

675 676 677 678
    num_blocks = (na-1)/nbw + 1

    allocate(tmat(nbw,nbw,num_blocks), stat=istat, errmsg=errorMessage)
    if (istat .ne. 0) then
679
      print *,"solve_evp_real_2stage: error when allocating tmat "//errorMessage
680 681
      stop
    endif
682

683
    ! Reduction full -> band
684

685 686
    ttt0 = MPI_Wtime()
    ttts = ttt0
687 688 689 690 691 692
#ifdef DOUBLE_PRECISION_REAL
    call bandred_real_double(na, a, lda, nblk, nbw, matrixCols, num_blocks, mpi_comm_rows, mpi_comm_cols, &
                      tmat, wantDebug, useGPU, success, useQRActual)
#else
    call bandred_real_single(na, a, lda, nblk, nbw, matrixCols, num_blocks, mpi_comm_rows, mpi_comm_cols, &
                      tmat, wantDebug, useGPU, success, useQRActual)
693
#endif
694
    if (.not.(success)) return
695 696
    ttt1 = MPI_Wtime()
    if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
697
       write(error_unit,*) 'Time bandred_real               :',ttt1-ttt0
698

699
     ! Reduction band -> tridiagonal
700

701 702 703 704 705
     allocate(e(na), stat=istat, errmsg=errorMessage)
     if (istat .ne. 0) then
       print *,"solve_evp_real_2stage: error when allocating e "//errorMessage
       stop
     endif
706

707 708 709 710 711 712 713 714
     ttt0 = MPI_Wtime()
#ifdef DOUBLE_PRECISION_REAL
     call tridiag_band_real_double(na, nbw, nblk, a, lda, ev, e, matrixCols, hh_trans_real, &
                          mpi_comm_rows, mpi_comm_cols, mpi_comm_all)
#else
     call tridiag_band_real_single(na, nbw, nblk, a, lda, ev, e, matrixCols, hh_trans_real, &
                          mpi_comm_rows, mpi_comm_cols, mpi_comm_all)
#endif
715

716 717 718
     ttt1 = MPI_Wtime()
     if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
       write(error_unit,*) 'Time tridiag_band_real          :',ttt1-ttt0
719

720
#ifdef WITH_MPI
721

722 723 724
#ifdef DOUBLE_PRECISION_REAL
     call mpi_bcast(ev,na,MPI_REAL8,0,mpi_comm_all,mpierr)
     call mpi_bcast(e,na,MPI_REAL8,0,mpi_comm_all,mpierr)
725
#else
726 727
     call mpi_bcast(ev,na,MPI_REAL4,0,mpi_comm_all,mpierr)
     call mpi_bcast(e,na,MPI_REAL4,0,mpi_comm_all,mpierr)
728
#endif
729

730
#endif /* WITH_MPI */
731 732
     ttt1 = MPI_Wtime()
     time_evp_fwd = ttt1-ttts
733

734
     ! Solve tridiagonal system
735

736 737 738 739 740 741 742 743 744
     ttt0 = MPI_Wtime()
#ifdef DOUBLE_PRECISION_REAL
     call solve_tridi_double(na, nev, ev, e, q, ldq, nblk, matrixCols, mpi_comm_rows,  &
                      mpi_comm_cols, wantDebug, success)
#else
     call solve_tridi_single(na, nev, ev, e, q, ldq, nblk, matrixCols, mpi_comm_rows,  &
                      mpi_comm_cols, wantDebug, success)
#endif
     if (.not.(success)) return
745

746 747 748 749 750
     ttt1 = MPI_Wtime()
     if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
     write(error_unit,*) 'Time solve_tridi                :',ttt1-ttt0
     time_evp_solve = ttt1-ttt0
     ttts = ttt1
751

752 753 754 755 756 757
     deallocate(e, stat=istat, errmsg=errorMessage)
     if (istat .ne. 0) then
       print *,"solve_evp_real_2stage: error when deallocating e "//errorMessage
       stop
     endif
     ! Backtransform stage 1
758

759 760 761 762 763 764 765 766 767 768
     ttt0 = MPI_Wtime()
#ifdef DOUBLE_PRECISION_REAL
     call trans_ev_tridi_to_band_real_double(na, nev, nblk, nbw, q, ldq, matrixCols, hh_trans_real, &
                                    mpi_comm_rows, mpi_comm_cols, wantDebug, useGPU, success,      &
                                    THIS_REAL_ELPA_KERNEL)
#else
     call trans_ev_tridi_to_band_real_single(na, nev, nblk, nbw, q, ldq, matrixCols, hh_trans_real, &
                                    mpi_comm_rows, mpi_comm_cols, wantDebug, useGPU, success,      &
                                    THIS_REAL_ELPA_KERNEL)
#endif
769

770 771 772 773
     if (.not.(success)) return
     ttt1 = MPI_Wtime()
     if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
       write(error_unit,*) 'Time trans_ev_tridi_to_band_real:',ttt1-ttt0
774

775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949
     ! We can now deallocate the stored householder vectors
     deallocate(hh_trans_real, stat=istat, errmsg=errorMessage)
     if (istat .ne. 0) then
       print *,"solve_evp_real_2stage: error when deallocating hh_trans_real "//errorMessage
       stop
     endif


     ! Backtransform stage 2
     print *,"useGPU== ",useGPU
     ttt0 = MPI_Wtime()
#ifdef DOUBLE_PRECISION_REAL
     call trans_ev_band_to_full_real_double(na, nev, nblk, nbw, a, lda, tmat, q, ldq, matrixCols, num_blocks, mpi_comm_rows, &
                                     mpi_comm_cols, useGPU, useQRActual)
#else
     call trans_ev_band_to_full_real_single(na, nev, nblk, nbw, a, lda, tmat, q, ldq, matrixCols, num_blocks, mpi_comm_rows, &
                                     mpi_comm_cols, useGPU, useQRActual)
#endif

     ttt1 = MPI_Wtime()
     if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
       write(error_unit,*) 'Time trans_ev_band_to_full_real :',ttt1-ttt0
     time_evp_back = ttt1-ttts

     deallocate(tmat, stat=istat, errmsg=errorMessage)
     if (istat .ne. 0) then
       print *,"solve_evp_real_2stage: error when deallocating tmat"//errorMessage
       stop
     endif

#ifdef HAVE_DETAILED_TIMINGS
     call timer%stop("solve_evp_real_2stage_single")
#endif
1    format(a,f10.3)

#ifdef DOUBLE_PRECISION_REAL
   end function solve_evp_real_2stage_double
#else
   end function solve_evp_real_2stage_single
#endif

#endif /* WANT_SINGLE_PRECISION_REAL */

   !>  \brief solve_evp_complex_2stage_double: Fortran function to solve the double-precision complex eigenvalue problem with a 2 stage approach
!>
!>  Parameters
!>
!>  \param na                                   Order of matrix a
!>
!>  \param nev                                  Number of eigenvalues needed
!>
!>  \param a(lda,matrixCols)                    Distributed matrix for which eigenvalues are to be computed.
!>                                              Distribution is like in Scalapack.
!>                                              The full matrix must be set (not only one half like in scalapack).
!>                                              Destroyed on exit (upper and lower half).
!>
!>  \param lda                                  Leading dimension of a
!>
!>  \param ev(na)                               On output: eigenvalues of a, every processor gets the complete set
!>
!>  \param q(ldq,matrixCols)                    On output: Eigenvectors of a
!>                                              Distribution is like in Scalapack.
!>                                              Must be always dimensioned to the full size (corresponding to (na,na))
!>                                              even if only a part of the eigenvalues is needed.
!>
!>  \param ldq                                  Leading dimension of q
!>
!>  \param nblk                                 blocksize of cyclic distribution, must be the same in both directions!
!>
!>  \param matrixCols                           local columns of matrix a and q
!>
!>  \param mpi_comm_rows                        MPI communicator for rows
!>  \param mpi_comm_cols                        MPI communicator for columns
!>  \param mpi_comm_all                         MPI communicator for the total processor set
!>
!>  \param THIS_REAL_ELPA_KERNEL_API (optional) specify used ELPA2 kernel via API
!>
!>  \result success                             logical, false if error occured
!-------------------------------------------------------------------------------
#define DOUBLE_PRECISION_COMPLEX 1

#ifdef DOUBLE_PRECISION_COMPLEX
function solve_evp_complex_2stage_double(na, nev, a, lda, ev, q, ldq, nblk, &
                                  matrixCols, mpi_comm_rows, mpi_comm_cols,      &
                                    mpi_comm_all, THIS_COMPLEX_ELPA_KERNEL_API) result(success)
#else
function solve_evp_complex_2stage_single(na, nev, a, lda, ev, q, ldq, nblk, &
                                  matrixCols, mpi_comm_rows, mpi_comm_cols,      &
                                    mpi_comm_all, THIS_COMPLEX_ELPA_KERNEL_API) result(success)
#endif


#ifdef HAVE_DETAILED_TIMINGS
   use timings
#endif
   use precision
   use cuda_functions
   use mod_check_for_gpu
   use iso_c_binding
   implicit none
   integer(kind=ik), intent(in), optional :: THIS_COMPLEX_ELPA_KERNEL_API
   integer(kind=ik)                       :: THIS_COMPLEX_ELPA_KERNEL
   integer(kind=ik), intent(in)           :: na, nev, lda, ldq, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols, mpi_comm_all
   complex(kind=ck8), intent(inout)        :: a(lda,matrixCols), q(ldq,matrixCols)
   ! was
   ! complex a(lda,*), q(ldq,*)
   real(kind=rk8), intent(inout)           :: ev(na)
   complex(kind=ck8), allocatable          :: hh_trans_complex(:,:)

   integer(kind=ik)                       :: my_prow, my_pcol, np_rows, np_cols, mpierr, my_pe, n_pes
   integer(kind=ik)                       :: l_cols, l_rows, l_cols_nev, nbw, num_blocks
   complex(kind=ck8), allocatable          :: tmat(:,:,:)
   real(kind=rk8), allocatable             :: q_real(:,:), e(:)
   real(kind=c_double)                    :: ttt0, ttt1, ttts  ! MPI_WTIME always needs double
   integer(kind=ik)                       :: i

   logical                                :: success, wantDebug
   logical, save                          :: firstCall = .true.
   integer(kind=ik)                       :: istat
   character(200)                         :: errorMessage
   logical                                :: useGPU
   integer(kind=ik)                       :: numberOfGPUDevices

#ifdef HAVE_DETAILED_TIMINGS
    call timer%start("solve_evp_complex_2stage_double")
#endif

    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)

    useGPU = .false.
    wantDebug = .false.
    if (firstCall) then
      ! are debug messages desired?
      wantDebug = debug_messages_via_environment_variable()
      firstCall = .false.
    endif


    success = .true.

    if (present(THIS_COMPLEX_ELPA_KERNEL_API)) then
      ! user defined kernel via the optional argument in the API call
      THIS_COMPLEX_ELPA_KERNEL = THIS_COMPLEX_ELPA_KERNEL_API
    else
      ! if kernel is not choosen via api
      ! check whether set by environment variable
      THIS_COMPLEX_ELPA_KERNEL = get_actual_complex_kernel()
    endif

    ! check whether choosen kernel is allowed
    if (check_allowed_complex_kernels(THIS_COMPLEX_ELPA_KERNEL)) then

      if (my_pe == 0) then
        write(error_unit,*) " "
        write(error_unit,*) "The choosen kernel ",COMPLEX_ELPA_KERNEL_NAMES(THIS_COMPLEX_ELPA_KERNEL)
        write(error_unit,*) "is not in the list of the allowed kernels!"
        write(error_unit,*) " "
        write(error_unit,*) "Allowed kernels are:"
        do i=1,size(COMPLEX_ELPA_KERNEL_NAMES(:))
          if (AVAILABLE_COMPLEX_ELPA_KERNELS(i) .ne. 0) then
            write(error_unit,*) COMPLEX_ELPA_KERNEL_NAMES(i)
          endif
        enddo

        write(error_unit,*) " "
        write(error_unit,*) "The defaul kernel COMPLEX_ELPA_KERNEL_GENERIC will be used !"
      endif
      THIS_COMPLEX_ELPA_KERNEL = COMPLEX_ELPA_KERNEL_GENERIC
    endif
950

951
#ifndef DOUBLE_PRECISION_COMPLEX
952 953 954 955 956 957 958
    if ( (THIS_COMPLEX_ELPA_KERNEL .eq. COMPLEX_ELPA_KERNEL_GENERIC) .or. &
         (THIS_COMPLEX_ELPA_KERNEL .eq. COMPLEX_ELPA_KERNEL_GENERIC_SIMPLE) .or. &
         (THIS_COMPLEX_ELPA_KERNEL .eq. COMPLEX_ELPA_KERNEL_SSE) ) then
    else
      print *,"At the moment single precision only works with the generic kernels"
      stop
    endif
959
#endif
960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274
    if (THIS_COMPLEX_ELPA_KERNEL .eq. COMPLEX_ELPA_KERNEL_GPU) then
      if (check_for_gpu(my_pe, numberOfGPUDevices, wantDebug=wantDebug)) then
        useGPU=.true.
      endif
      if (nblk .ne. 128) then
        print *,"At the moment GPU version needs blocksize 128"
        stop
      endif

      ! set the neccessary parameters
      cudaMemcpyHostToDevice   = cuda_memcpyHostToDevice()
      cudaMemcpyDeviceToHost   = cuda_memcpyDeviceToHost()
      cudaMemcpyDeviceToDevice = cuda_memcpyDeviceToDevice()
      cudaHostRegisterPortable = cuda_hostRegisterPortable()
      cudaHostRegisterMapped   = cuda_hostRegisterMapped()
    endif

    ! Choose bandwidth, must be a multiple of nblk, set to a value >= 32

    nbw = (31/nblk+1)*nblk

    num_blocks = (na-1)/nbw + 1

    allocate(tmat(nbw,nbw,num_blocks), stat=istat, errmsg=errorMessage)
    if (istat .ne. 0) then
      print *,"solve_evp_complex_2stage: error when allocating tmat"//errorMessage
      stop
    endif
    ! Reduction full -> band

    ttt0 = MPI_Wtime()
    ttts = ttt0
#ifdef DOUBLE_PRECISION_COMPLEX
    call bandred_complex_double(na, a, lda, nblk, nbw, matrixCols, num_blocks, mpi_comm_rows, mpi_comm_cols, &
                         tmat, wantDebug, useGPU, success)
#else
    call bandred_complex_single(na, a, lda, nblk, nbw, matrixCols, num_blocks, mpi_comm_rows, mpi_comm_cols, &
                         tmat, wantDebug, useGPU, success)
#endif
    if (.not.(success)) then

#ifdef HAVE_DETAILED_TIMINGS
      call timer%stop("solve_evp_complex_2stage_double")
#endif
      return
    endif
    ttt1 = MPI_Wtime()
    if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
       write(error_unit,*) 'Time bandred_complex               :',ttt1-ttt0

    ! Reduction band -> tridiagonal

    allocate(e(na), stat=istat, errmsg=errorMessage)
    if (istat .ne. 0) then
      print *,"solve_evp_complex_2stage: error when allocating e"//errorMessage
      stop
    endif


    ttt0 = MPI_Wtime()
#ifdef DOUBLE_PRECISION_COMPLEX
   call tridiag_band_complex_double(na, nbw, nblk, a, lda, ev, e, matrixCols, hh_trans_complex, &
                             mpi_comm_rows, mpi_comm_cols, mpi_comm_all)
#else
   call tridiag_band_complex_single(na, nbw, nblk, a, lda, ev, e, matrixCols, hh_trans_complex, &
                             mpi_comm_rows, mpi_comm_cols, mpi_comm_all)
#endif

    ttt1 = MPI_Wtime()
    if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
       write(error_unit,*) 'Time tridiag_band_complex          :',ttt1-ttt0

#ifdef WITH_MPI

#ifdef DOUBLE_PRECISION_COMPLEX
    call mpi_bcast(ev, na, mpi_real8, 0, mpi_comm_all, mpierr)
    call mpi_bcast(e, na, mpi_real8, 0, mpi_comm_all, mpierr)
#else
    call mpi_bcast(ev, na, mpi_real4, 0, mpi_comm_all, mpierr)
    call mpi_bcast(e, na, mpi_real4, 0, mpi_comm_all, mpierr)
#endif

#endif /* WITH_MPI */
    ttt1 = MPI_Wtime()
    time_evp_fwd = ttt1-ttts

    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_complex_2stage: error when allocating q_real"//errorMessage
      stop
    endif

    ! Solve tridiagonal system

    ttt0 = MPI_Wtime()
#ifdef DOUBLE_PRECISION_COMPLEX
    call solve_tridi_double(na, nev, ev, e, q_real, ubound(q_real,dim=1), nblk, matrixCols, &
                     mpi_comm_rows, mpi_comm_cols, wantDebug, success)
#else
    call solve_tridi_single(na, nev, ev, e, q_real, ubound(q_real,dim=1), nblk, matrixCols, &
                     mpi_comm_rows, mpi_comm_cols, wantDebug, success)
#endif
    if (.not.(success)) return

    ttt1 = MPI_Wtime()
    if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times)  &
       write(error_unit,*) 'Time solve_tridi                   :',ttt1-ttt0
    time_evp_solve = ttt1-ttt0
    ttts = ttt1

    q(1:l_rows,1:l_cols_nev) = q_real(1:l_rows,1:l_cols_nev)

    deallocate(e, q_real, stat=istat, errmsg=errorMessage)
    if (istat .ne. 0) then
      print *,"solve_evp_complex_2stage: error when deallocating e, q_real"//errorMessage
      stop
    endif


    ! Backtransform stage 1

    ttt0 = MPI_Wtime()
#ifdef DOUBLE_PRECISION_COMPLEX
    call trans_ev_tridi_to_band_complex_double(na, nev, nblk, nbw, q, ldq,  &
                                       matrixCols, hh_trans_complex, &
                                       mpi_comm_rows, mpi_comm_cols, &
                                       wantDebug, useGPU, success,THIS_COMPLEX_ELPA_KERNEL)
#else
    call trans_ev_tridi_to_band_complex_single(na, nev, nblk, nbw, q, ldq,  &
                                       matrixCols, hh_trans_complex, &
                                       mpi_comm_rows, mpi_comm_cols, &
                                       wantDebug, useGPU, success,THIS_COMPLEX_ELPA_KERNEL)
#endif
    if (.not.(success)) return
    ttt1 = MPI_Wtime()
    if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
       write(error_unit,*) 'Time trans_ev_tridi_to_band_complex:',ttt1-ttt0

    ! We can now deallocate the stored householder vectors
    deallocate(hh_trans_complex, stat=istat, errmsg=errorMessage)
    if (istat .ne. 0) then
      print *,"solve_evp_complex_2stage: error when deallocating hh_trans_complex"//errorMessage
      stop
    endif

    ! Backtransform stage 2

    ttt0 = MPI_Wtime()
#ifdef DOUBLE_PRECISION_COMPLEX
   call trans_ev_band_to_full_complex_double(na, nev, nblk, nbw, a, lda, tmat, q, ldq, matrixCols, num_blocks, &
                                      mpi_comm_rows, mpi_comm_cols, useGPU)
#else
   call trans_ev_band_to_full_complex_single(na, nev, nblk, nbw, a, lda, tmat, q, ldq, matrixCols, num_blocks, &
                                      mpi_comm_rows, mpi_comm_cols, useGPU)
#endif
    ttt1 = MPI_Wtime()
    if (my_prow==0 .and. my_pcol==0 .and. elpa_print_times) &
       write(error_unit,*) 'Time trans_ev_band_to_full_complex :',ttt1-ttt0
    time_evp_back = ttt1-ttts

    deallocate(tmat, stat=istat, errmsg=errorMessage)
    if (istat .ne. 0) then
      print *,"solve_evp_complex_2stage: error when deallocating tmat "//errorMessage
      stop
    endif

#ifdef HAVE_DETAILED_TIMINGS
    call timer%stop("solve_evp_complex_2stage_double")
#endif

1   format(a,f10.3)
#ifdef DOUBLE_PRECISION_COMPLEX
end function solve_evp_complex_2stage_double
#else
end function solve_evp_complex_2stage_single
#endif

#ifdef WANT_SINGLE_PRECISION_COMPLEX
#undef DOUBLE_PRECISION_COMPLEX

!>  \brief solve_evp_complex_2stage_single: Fortran function to solve the single-precision complex eigenvalue problem with a 2 stage approach
!>
!>  Parameters
!>
!>  \param na                                   Order of matrix a
!>
!>  \param nev                                  Number of eigenvalues needed
!>
!>  \param a(lda,matrixCols)                    Distributed matrix for which eigenvalues are to be computed.
!>                                              Distribution is like in Scalapack.
!>                                              The full matrix must be set (not only one half like in scalapack).
!>                                              Destroyed on exit (upper and lower half).
!>
!>  \param lda                                  Leading dimension of a
!>
!>  \param ev(na)                               On output: eigenvalues of a, every processor gets the complete set
!>
!>  \param q(ldq,matrixCols)                    On output: Eigenvectors of a
!>                                              Distribution is like in Scalapack.
!>                                              Must be always dimensioned to the full size (corresponding to (na,na))
!>                                              even if only a part of the eigenvalues is needed.
!>
!>  \param ldq                                  Leading dimension of q
!>
!>  \param nblk                                 blocksize of cyclic distribution, must be the same in both directions!
!>
!>  \param matrixCols                           local columns of matrix a and q
!>
!>  \param mpi_comm_rows                        MPI communicator for rows
!>  \param mpi_comm_cols                        MPI communicator for columns
!>  \param mpi_comm_all                         MPI communicator for the total processor set
!>
!>  \param THIS_REAL_ELPA_KERNEL_API (optional) specify used ELPA2 kernel via API
!>
!>  \result success                             logical, false if error occured
!-------------------------------------------------------------------------------

#ifdef DOUBLE_PRECISION_COMPLEX
function solve_evp_complex_2stage_double(na, nev, a, lda, ev, q, ldq, nblk, &
                                  matrixCols, mpi_comm_rows, mpi_comm_cols,      &
                                    mpi_comm_all, THIS_COMPLEX_ELPA_KERNEL_API) result(success)
#else
function solve_evp_complex_2stage_single(na, nev, a, lda, ev, q, ldq, nblk, &
                                  matrixCols, mpi_comm_rows, mpi_comm_cols,      &
                                    mpi_comm_all, THIS_COMPLEX_ELPA_KERNEL_API) result(success)
#endif


#ifdef HAVE_DETAILED_TIMINGS
   use timings
#endif
   use precision
   use cuda_functions
   use mod_check_for_gpu
   use iso_c_binding
   implicit none
   integer(kind=ik), intent(in), optional :: THIS_COMPLEX_ELPA_KERNEL_API
   integer(kind=ik)                       :: THIS_COMPLEX_ELPA_KERNEL
   integer(kind=ik), intent(in)           :: na, nev, lda, ldq, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols, mpi_comm_all
   complex(kind=ck4), intent(inout)        :: a(lda,matrixCols), q(ldq,matrixCols)
   ! was
   ! complex a(lda,*), q(ldq,*)
   real(kind=rk4), intent(inout)           :: ev(na)
   complex(kind=ck4), allocatable          :: hh_trans_complex(:,:)

   integer(kind=ik)                       :: my_prow, my_pcol, np_rows, np_cols, mpierr, my_pe, n_pes
   integer(kind=ik)                       :: l_cols, l_rows, l_cols_nev, nbw, num_blocks
   complex(kind=ck4), allocatable          :: tmat(:,:,:)
   real(kind=rk4), allocatable             :: q_real(:,:), e(:)
   real(kind=c_double)                    :: ttt0, ttt1, ttts  ! MPI_WTIME always needs double
   integer(kind=ik)                       :: i

   logical                                :: success, wantDebug
   logical, save                          :: firstCall = .true.
   integer(kind=ik)                       :: istat
   character(200)                         :: errorMessage
   logical                                :: useGPU
   integer(kind=ik)                       :: numberOfGPUDevices

#ifdef HAVE_DETAILED_TIMINGS
    call timer%start("solve_evp_complex_2stage_single")
#endif

    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)

    useGPU = .false.
    wantDebug = .false.
    if (firstCall) then
      ! are debug messages desired?
      wantDebug = debug_messages_via_environment_variable()
      firstCall = .false.
    endif


    success = .true.

    if (present(THIS_COMPLEX_ELPA_KERNEL_API)) then
      ! user defined kernel via the optional argument in the API call
      THIS_COMPLEX_ELPA_KERNEL = THIS_COMPLEX_ELPA_KERNEL_API
    else
      ! if kernel is not choosen via api
      ! check whether set by environment variable
      THIS_COMPLEX_ELPA_KERNEL = get_actual_complex_kernel()
    endif

    ! check whether choosen kernel is allowed
    if (check_allowed_complex_kernels(THIS_COMPLEX_ELPA_KERNEL)) then

      if (my_pe == 0) then
        write(