elpa1.F90 18.8 KB
Newer Older
1
2
!    This file is part of ELPA.
!
3
!    The ELPA library was originally created by the ELPA consortium,
4
5
!    consisting of the following organizations:
!
6
!    - Rechenzentrum Garching der Max-Planck-Gesellschaft (RZG),
7
8
9
!    - Bergische Universität Wuppertal, Lehrstuhl für angewandte
!      Informatik,
!    - Technische Universität München, Lehrstuhl für Informatik mit
10
11
12
13
14
!      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
15
16
!    - IBM Deutschland GmbH
!
17
!    This particular source code file contains additions, changes and
18
!    enhancements authored by Intel Corporation which is not part of
19
!    the ELPA consortium.
20
21
!
!    More information can be found here:
22
!    http://elpa.mpcdf.mpg.de/
23
24
!
!    ELPA is free software: you can redistribute it and/or modify
25
26
!    it under the terms of the version 3 of the license of the
!    GNU Lesser General Public License as published by the Free
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
!    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
46
!
47
48
49
50
51
! 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".

52
53
54
!> \mainpage
!> Eigenvalue SoLvers for Petaflop-Applications (ELPA)
!> \par
55
!> http://elpa.mpcdf.mpg.de
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
!>
!> \par
!>    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 Naturwissenschaftrn,
!>      Leipzig, Abt. Komplexe Strukutren in Biologie und Kognition,
!>      and
!>    - IBM Deutschland GmbH
!>
!>   Some parts and enhancements of ELPA have been contributed and authored
!>   by the Intel Corporation which is not part of the ELPA consortium.
!>
!>   Contributions to the ELPA source have been authored by (in alphabetical order):
!>
!> \author T. Auckenthaler, Volker Blum, A. Heinecke, L. Huedepohl, R. Johanni, Werner Jürgens, and A. Marek

80

81
82
#include "config-f90.h"
!> \brief Fortran module which provides the routines to use the one-stage ELPA solver
83
module ELPA1
Andreas Marek's avatar
Andreas Marek committed
84
  use precision
85
86
  use elpa_utilities

87
88
89
90
91
92
  implicit none

  PRIVATE ! By default, all routines contained are private

  ! The following routines are public:

93
94
  public :: get_elpa_row_col_comms     !< old, deprecated interface: Sets MPI row/col communicators
  public :: get_elpa_communicators     !< Sets MPI row/col communicators
95

96
97
98
99
  public :: solve_evp_real             !< old, deprecated interface: Driver routine for real eigenvalue problem
  public :: solve_evp_real_1stage      !< Driver routine for real eigenvalue problem
  public :: solve_evp_complex          !< old, deprecated interface:  Driver routine for complex eigenvalue problem
  public :: solve_evp_complex_1stage   !< Driver routine for complex eigenvalue problem
100
101
102

  ! Timing results, set by every call to solve_evp_xxx

Andreas Marek's avatar
Andreas Marek committed
103
104
105
  real(kind=rk), public :: time_evp_fwd    !< time for forward transformations (to tridiagonal form)
  real(kind=rk), public :: time_evp_solve  !< time for solving the tridiagonal system
  real(kind=rk), public :: time_evp_back   !< time for back transformations of eigenvectors
106

107
  logical, public :: elpa_print_times = .false. !< Set elpa_print_times to .true. for explicit timing outputs
108
109


110
111
112
113
114
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
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
!> \brief get_elpa_row_col_comms:  old, deprecated Fortran function to create the MPI communicators for ELPA. Better use "elpa_get_communicators"
!> \detail
!> The interface and variable definition is the same as in "elpa_get_communicators"
!> \param  mpi_comm_global   Global communicator for the calculations (in)
!>
!> \param  my_prow           Row coordinate of the calling process in the process grid (in)
!>
!> \param  my_pcol           Column coordinate of the calling process in the process grid (in)
!>
!> \param  mpi_comm_rows     Communicator for communicating within rows of processes (out)
!>
!> \param  mpi_comm_cols     Communicator for communicating within columns of processes (out)
!> \result mpierr            integer error value of mpi_comm_split function
  interface get_elpa_row_col_comms
    module procedure get_elpa_communicators
  end interface

!> \brief solve_evp_real: old, deprecated Fortran function to solve the real eigenvalue problem with 1-stage solver. Better use "solve_evp_real_1stage"
!>
!> \detail
!>  The interface and variable definition is the same as in "elpa_solve_evp_real_1stage"
!  Parameters
!
!> \param  na                   Order of matrix a
!>
!> \param  nev                  Number of eigenvalues needed.
!>                              The smallest nev eigenvalues/eigenvectors are calculated.
!>
!> \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           distributed number of matrix columns
!>
!>  \param mpi_comm_rows        MPI-Communicator for rows
!>  \param mpi_comm_cols        MPI-Communicator for columns
!>
!>  \result                     success


  interface solve_evp_real
    module procedure solve_evp_real_1stage
  end interface

!> \brief solve_evp_complex: old, deprecated Fortran function to solve the complex eigenvalue problem with 1-stage solver. Better use "solve_evp_complex_1stage"
!>
!> \detail
!> The interface and variable definition is the same as in "elpa_solve_evp_complex_1stage"
!  Parameters
!
!> \param  na                   Order of matrix a
!>
!> \param  nev                  Number of eigenvalues needed.
!>                              The smallest nev eigenvalues/eigenvectors are calculated.
!>
!> \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           distributed number of matrix columns
!>
!>  \param mpi_comm_rows        MPI-Communicator for rows
!>  \param mpi_comm_cols        MPI-Communicator for columns
!>
!>  \result                     success


  interface solve_evp_complex
    module procedure solve_evp_complex_1stage
  end interface

209
210
211
212
contains

!-------------------------------------------------------------------------------

213
!> \brief Fortran function to create the MPI communicators for ELPA.
214
215
216
217
218
219
! All ELPA routines need MPI communicators for communicating within
! rows or columns of processes, these are set here.
! mpi_comm_rows/mpi_comm_cols can be free'd with MPI_Comm_free if not used any more.
!
!  Parameters
!
220
221
222
223
224
225
226
227
228
229
230
231
!> \param  mpi_comm_global   Global communicator for the calculations (in)
!>
!> \param  my_prow           Row coordinate of the calling process in the process grid (in)
!>
!> \param  my_pcol           Column coordinate of the calling process in the process grid (in)
!>
!> \param  mpi_comm_rows     Communicator for communicating within rows of processes (out)
!>
!> \param  mpi_comm_cols     Communicator for communicating within columns of processes (out)
!> \result mpierr            integer error value of mpi_comm_split function


232
function get_elpa_communicators(mpi_comm_global, my_prow, my_pcol, mpi_comm_rows, mpi_comm_cols) result(mpierr)
Andreas Marek's avatar
Andreas Marek committed
233
   use precision
234
235
   implicit none

Andreas Marek's avatar
Andreas Marek committed
236
237
   integer(kind=ik), intent(in)  :: mpi_comm_global, my_prow, my_pcol
   integer(kind=ik), intent(out) :: mpi_comm_rows, mpi_comm_cols
238

Andreas Marek's avatar
Andreas Marek committed
239
   integer(kind=ik)              :: mpierr
240
241
242
243
244
245
246
247
248

   ! mpi_comm_rows is used for communicating WITHIN rows, i.e. all processes
   ! having the same column coordinate share one mpi_comm_rows.
   ! So the "color" for splitting is my_pcol and the "key" is my row coordinate.
   ! Analogous for mpi_comm_cols

   call mpi_comm_split(mpi_comm_global,my_pcol,my_prow,mpi_comm_rows,mpierr)
   call mpi_comm_split(mpi_comm_global,my_prow,my_pcol,mpi_comm_cols,mpierr)

249
end function get_elpa_communicators
250
251


252
!> \brief solve_evp_real_1stage: Fortran function to solve the real eigenvalue problem with 1-stage solver
253
!>
254
255
!  Parameters
!
256
257
258
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
!> \param  na                   Order of matrix a
!>
!> \param  nev                  Number of eigenvalues needed.
!>                              The smallest nev eigenvalues/eigenvectors are calculated.
!>
!> \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           distributed number of matrix columns
!>
!>  \param mpi_comm_rows        MPI-Communicator for rows
!>  \param mpi_comm_cols        MPI-Communicator for columns
!>
!>  \result                     success


287
function solve_evp_real_1stage(na, nev, a, lda, ev, q, ldq, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols) result(success)
Andreas Marek's avatar
Andreas Marek committed
288
   use precision
289
#ifdef HAVE_DETAILED_TIMINGS
Andreas Marek's avatar
Andreas Marek committed
290
   use timings
291
#endif
292
293
   use elpa_mpi
   use elpa1_compute
294
295
   implicit none

Andreas Marek's avatar
Andreas Marek committed
296
297
   integer(kind=ik), intent(in)  :: na, nev, lda, ldq, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols
   real(kind=rk)                 :: a(lda,matrixCols), ev(na), q(ldq,matrixCols)
298
299
   ! was
   ! real a(lda,*), q(ldq,*)
300

Andreas Marek's avatar
Andreas Marek committed
301
302
303
304
305
306
   integer(kind=ik)              :: my_prow, my_pcol, mpierr
   real(kind=rk), allocatable    :: e(:), tau(:)
   real(kind=rk)                 :: ttt0, ttt1
   logical                       :: success
   logical, save                 :: firstCall = .true.
   logical                       :: wantDebug
307

308
#ifdef HAVE_DETAILED_TIMINGS
309
   call timer%start("solve_evp_real_1stage")
310
311
#endif

312
313
314
   call mpi_comm_rank(mpi_comm_rows,my_prow,mpierr)
   call mpi_comm_rank(mpi_comm_cols,my_pcol,mpierr)

315
316
   success = .true.

317
318
319
320
321
322
323
   wantDebug = .false.
   if (firstCall) then
     ! are debug messages desired?
     wantDebug = debug_messages_via_environment_variable()
     firstCall = .false.
   endif

324
325
326
   allocate(e(na), tau(na))

   ttt0 = MPI_Wtime()
327
   call tridiag_real(na, a, lda, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols, ev, e, tau)
328

329
   ttt1 = MPI_Wtime()
330
   if(my_prow==0 .and. my_pcol==0 .and. elpa_print_times) write(error_unit,*) 'Time tridiag_real :',ttt1-ttt0
331
332
333
   time_evp_fwd = ttt1-ttt0

   ttt0 = MPI_Wtime()
334
   call solve_tridi(na, nev, ev, e, q, ldq, nblk, matrixCols, mpi_comm_rows, &
335
                    mpi_comm_cols, wantDebug, success)
336
337
   if (.not.(success)) return

338
   ttt1 = MPI_Wtime()
339
   if(my_prow==0 .and. my_pcol==0 .and. elpa_print_times) write(error_unit,*) 'Time solve_tridi  :',ttt1-ttt0
340
341
342
   time_evp_solve = ttt1-ttt0

   ttt0 = MPI_Wtime()
343
   call trans_ev_real(na, nev, a, lda, tau, q, ldq, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols)
344
   ttt1 = MPI_Wtime()
345
   if(my_prow==0 .and. my_pcol==0 .and. elpa_print_times) write(error_unit,*) 'Time trans_ev_real:',ttt1-ttt0
346
347
348
349
   time_evp_back = ttt1-ttt0

   deallocate(e, tau)

350
#ifdef HAVE_DETAILED_TIMINGS
351
   call timer%stop("solve_evp_real_1stage")
352
353
#endif

354
end function solve_evp_real_1stage
355
356


357
!> \brief solve_evp_complex_1stage: Fortran function to solve the complex eigenvalue problem with 1-stage solver
358
!>
359
360
!  Parameters
!
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
!> \param  na                   Order of matrix a
!>
!> \param  nev                  Number of eigenvalues needed.
!>                              The smallest nev eigenvalues/eigenvectors are calculated.
!>
!> \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           distributed number of matrix columns
!>
!>  \param mpi_comm_rows        MPI-Communicator for rows
!>  \param mpi_comm_cols        MPI-Communicator for columns
!>
!>  \result                     success

391
function solve_evp_complex_1stage(na, nev, a, lda, ev, q, ldq, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols) result(success)
392
#ifdef HAVE_DETAILED_TIMINGS
Andreas Marek's avatar
Andreas Marek committed
393
   use timings
394
#endif
Andreas Marek's avatar
Andreas Marek committed
395
   use precision
396
397
   use elpa_mpi
   use elpa1_compute
398
399
   implicit none

Andreas Marek's avatar
Andreas Marek committed
400
401
   integer(kind=ik), intent(in)     :: na, nev, lda, ldq, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols
   complex(kind=ck)                 :: a(lda,matrixCols), q(ldq,matrixCols)
402
403
   ! was
   ! complex a(lda,*), q(ldq,*)
Andreas Marek's avatar
Andreas Marek committed
404
   real(kind=rk)                    :: ev(na)
405

Andreas Marek's avatar
Andreas Marek committed
406
407
408
409
410
   integer(kind=ik)                 :: my_prow, my_pcol, np_rows, np_cols, mpierr
   integer(kind=ik)                 :: l_rows, l_cols, l_cols_nev
   real(kind=rk), allocatable       :: q_real(:,:), e(:)
   complex(kind=ck), allocatable    :: tau(:)
   real(kind=rk)                    :: ttt0, ttt1
411

Andreas Marek's avatar
Andreas Marek committed
412
413
414
   logical                          :: success
   logical, save                    :: firstCall = .true.
   logical                          :: wantDebug
415

416
#ifdef HAVE_DETAILED_TIMINGS
417
   call timer%start("solve_evp_complex_1stage")
418
#endif
419

420
421
422
423
424
   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)

425
426
   success = .true.

427
428
429
430
431
432
433
434
   wantDebug = .false.
   if (firstCall) then
     ! are debug messages desired?
     wantDebug = debug_messages_via_environment_variable()
     firstCall = .false.
   endif


435
436
437
438
439
440
441
442
443
   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(e(na), tau(na))
   allocate(q_real(l_rows,l_cols))

   ttt0 = MPI_Wtime()
444
   call tridiag_complex(na, a, lda, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols, ev, e, tau)
445
   ttt1 = MPI_Wtime()
446
   if(my_prow==0 .and. my_pcol==0 .and. elpa_print_times) write(error_unit,*) 'Time tridiag_complex :',ttt1-ttt0
447
448
449
   time_evp_fwd = ttt1-ttt0

   ttt0 = MPI_Wtime()
450
   call solve_tridi(na, nev, ev, e, q_real, l_rows, nblk, matrixCols, mpi_comm_rows, &
451
                    mpi_comm_cols, wantDebug, success)
452
453
   if (.not.(success)) return

454
   ttt1 = MPI_Wtime()
455
   if(my_prow==0 .and. my_pcol==0 .and. elpa_print_times) write(error_unit,*) 'Time solve_tridi     :',ttt1-ttt0
456
457
458
459
460
   time_evp_solve = ttt1-ttt0

   ttt0 = MPI_Wtime()
   q(1:l_rows,1:l_cols_nev) = q_real(1:l_rows,1:l_cols_nev)

461
   call trans_ev_complex(na, nev, a, lda, tau, q, ldq, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols)
462
   ttt1 = MPI_Wtime()
463
   if(my_prow==0 .and. my_pcol==0 .and. elpa_print_times) write(error_unit,*) 'Time trans_ev_complex:',ttt1-ttt0
464
465
466
467
   time_evp_back = ttt1-ttt0

   deallocate(q_real)
   deallocate(e, tau)
468
#ifdef HAVE_DETAILED_TIMINGS
469
   call timer%stop("solve_evp_complex_1stage")
470
#endif
471

472
end function solve_evp_complex_1stage
473

474

475
476

end module ELPA1