elpa1_auxiliary.F90 132 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
!    This file is part of ELPA.
!
!    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,
13
!    - Max-Plack-Institut für Mathematik in den Naturwissenschaften,
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
57
58
!      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".

#include "config-f90.h"


module elpa1_auxiliary
  implicit none

59
60
  public :: elpa_mult_at_b_real_double      !< Multiply double-precision real matrices A**T * B
  public :: mult_at_b_real                  !< Old, deprecated interface to multiply double-precision real matrices A**T * B. DO NOT USE
61

62
63
  public :: elpa_mult_ah_b_complex_double   !< Multiply double-precision complex matrices A**H * B
  public :: mult_ah_b_complex               !< Old, deprecated interface to multiply double-precision complex matrices A**H * B. DO NOT USE
64

65
66
  public :: elpa_invert_trm_real_double     !< Invert double-precision real triangular matrix
  public :: invert_trm_real                 !< Old, deprecated interface for inversion of double-precision real triangular matrix. DO NOT USE
67

68
69
  public :: elpa_invert_trm_complex_double  !< Invert double-precision complex triangular matrix
  public :: invert_trm_complex              !< Old, deprecated interface to invert double-precision complex triangular matrix. DO NOT USE
70

71
72
  public :: elpa_cholesky_real_double       !< Cholesky factorization of a double-precision real matrix
  public :: cholesky_real                   !< Old, deprecated name for Cholesky factorization of a double-precision real matrix. DO NOT USE
73

74
75
  public :: elpa_cholesky_complex_double    !< Cholesky factorization of a double-precision complex matrix
  public :: cholesky_complex                !< Old, deprecated interface for a Cholesky factorization of a double-precision complex matrix. DO NOT USE
76

77
78
  public :: elpa_solve_tridi_double         !< Solve tridiagonal eigensystem for a double-precision matrix with divide and conquer method
  public :: solve_tridi                     !< Old, deprecated interface to solve tridiagonal eigensystem for a double-precision matrix with divide and conquer method
79
80

#ifdef WANT_SINGLE_PRECISION_REAL
81
82
83
84
  public :: elpa_cholesky_real_single       !< Cholesky factorization of a single-precision real matrix
  public :: elpa_invert_trm_real_single     !< Invert single-precision real triangular matrix
  public :: elpa_mult_at_b_real_single      !< Multiply single-precision real matrices A**T * B
  public :: elpa_solve_tridi_single         !< Solve tridiagonal eigensystem for a single-precision matrix with divide and conquer method
85
86
87
#endif

#ifdef WANT_SINGLE_PRECISION_COMPLEX
88
89
90
  public :: elpa_cholesky_complex_single    !< Cholesky factorization of a single-precision complex matrix
  public :: elpa_invert_trm_complex_single  !< Invert single-precision complex triangular matrix
  public :: elpa_mult_ah_b_complex_single   !< Multiply single-precision complex matrices A**H * B
91
92
#endif

93
!> \brief  cholesky_real: old, deprecated interface for Cholesky factorization of a double-precision real symmetric matrix
94
95
96
97
98
99
100
101
102
103
104
105
106
107
!> \details
!>
!> \param  na                   Order of matrix
!> \param  a(lda,matrixCols)    Distributed matrix which should be factorized.
!>                              Distribution is like in Scalapack.
!>                              Only upper triangle is needs to be set.
!>                              On return, the upper triangle contains the Cholesky factor
!>                              and the lower triangle is set to 0.
!> \param  lda                  Leading dimension of a
!> \param                       matrixCols  local columns of matrix a
!> \param  nblk                 blocksize of cyclic distribution, must be the same in both directions!
!> \param  mpi_comm_rows        MPI communicator for rows
!> \param  mpi_comm_cols        MPI communicator for columns
!> \param wantDebug             logical, more debug information on failure
108
!> \result succes                logical, reports success or failure
109
  interface cholesky_real
110
    module procedure elpa_cholesky_real_double
111
112
  end interface

113
!> \brief  Old, deprecated interface invert_trm_real: Inverts a upper double-precision triangular matrix
114
115
116
117
118
119
120
121
!> \details
!> \param  na                   Order of matrix
!> \param  a(lda,matrixCols)    Distributed matrix which should be inverted
!>                              Distribution is like in Scalapack.
!>                              Only upper triangle is needs to be set.
!>                              The lower triangle is not referenced.
!> \param  lda                  Leading dimension of a
!> \param  nblk                 blocksize of cyclic distribution, must be the same in both directions!
Andreas Marek's avatar
Andreas Marek committed
122
!> \param  matrixCols           local columns of matrix a
123
124
125
!> \param  mpi_comm_rows        MPI communicator for rows
!> \param  mpi_comm_cols        MPI communicator for columns
!> \param wantDebug             logical, more debug information on failure
126
!> \param result                logical, reports success or failure
127
128

  interface invert_trm_real
129
    module procedure elpa_invert_trm_real_double
130
131
  end interface

132
!> \brief  old, deprecated interface cholesky_complex: Cholesky factorization of a double-precision complex hermitian matrix
133
134
135
136
137
138
139
140
141
!> \details
!> \param  na                   Order of matrix
!> \param  a(lda,matrixCols)    Distributed matrix which should be factorized.
!>                              Distribution is like in Scalapack.
!>                              Only upper triangle is needs to be set.
!>                              On return, the upper triangle contains the Cholesky factor
!>                              and the lower triangle is set to 0.
!> \param  lda                  Leading dimension of a
!> \param  nblk                 blocksize of cyclic distribution, must be the same in both directions!
Andreas Marek's avatar
Andreas Marek committed
142
!> \param  matrixCols           local columns of matrix a
143
144
145
!> \param  mpi_comm_rows        MPI communicator for rows
!> \param  mpi_comm_cols        MPI communicator for columns
!> \param wantDebug             logical, more debug information on failure
146
!> \result succes               logical, reports success or failure
147
148
149


  interface cholesky_complex
150
    module procedure elpa_cholesky_real_double
151
152
  end interface

153
!> \brief  old, deprecated interface invert_trm_complex: Inverts a double-precision complex upper triangular matrix
154
155
156
157
158
159
160
161
!> \details
!> \param  na                   Order of matrix
!> \param  a(lda,matrixCols)    Distributed matrix which should be inverted
!>                              Distribution is like in Scalapack.
!>                              Only upper triangle is needs to be set.
!>                              The lower triangle is not referenced.
!> \param  lda                  Leading dimension of a
!> \param  nblk                 blocksize of cyclic distribution, must be the same in both directions!
Andreas Marek's avatar
Andreas Marek committed
162
!> \param  matrixCols           local columns of matrix a
163
164
165
!> \param  mpi_comm_rows        MPI communicator for rows
!> \param  mpi_comm_cols        MPI communicator for columns
!> \param wantDebug             logical, more debug information on failure
166
!> \result succes               logical, reports success or failure
167
168

  interface invert_trm_complex
169
    module procedure elpa_invert_trm_complex_double
170
171
  end interface

172
!> \brief  mult_at_b_real: Performs C : = A**T * B for double matrices
173
!> this is the old, deprecated interface for the newer elpa_mult_at_b_real
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
!>         where   A is a square matrix (na,na) which is optionally upper or lower triangular
!>                 B is a (na,ncb) matrix
!>                 C is a (na,ncb) matrix where optionally only the upper or lower
!>                   triangle may be computed
!> \details

!> \param  uplo_a               'U' if A is upper triangular
!>                              'L' if A is lower triangular
!>                              anything else if A is a full matrix
!>                              Please note: This pertains to the original A (as set in the calling program)
!>                                           whereas the transpose of A is used for calculations
!>                              If uplo_a is 'U' or 'L', the other triangle is not used at all,
!>                              i.e. it may contain arbitrary numbers
!> \param uplo_c                'U' if only the upper diagonal part of C is needed
!>                              'L' if only the upper diagonal part of C is needed
!>                              anything else if the full matrix C is needed
!>                              Please note: Even when uplo_c is 'U' or 'L', the other triangle may be
!>                                            written to a certain extent, i.e. one shouldn't rely on the content there!
!> \param na                    Number of rows/columns of A, number of rows of B and C
!> \param ncb                   Number of columns  of B and C
!> \param a                     matrix a
!> \param lda                   leading dimension of matrix a
!> \param b                     matrix b
!> \param ldb                   leading dimension of matrix b
!> \param nblk                  blocksize of cyclic distribution, must be the same in both directions!
!> \param  mpi_comm_rows        MPI communicator for rows
!> \param  mpi_comm_cols        MPI communicator for columns
!> \param c                     matrix c
!> \param ldc                   leading dimension of matrix c
  interface mult_at_b_real
204
    module procedure elpa_mult_at_b_real_double
205
206
  end interface

207
!> \brief  Old, deprecated interface mult_ah_b_complex: Performs C : = A**H * B for double-precision matrices
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
!>         where   A is a square matrix (na,na) which is optionally upper or lower triangular
!>                 B is a (na,ncb) matrix
!>                 C is a (na,ncb) matrix where optionally only the upper or lower
!>                   triangle may be computed
!> \details
!>
!> \param  uplo_a               'U' if A is upper triangular
!>                              'L' if A is lower triangular
!>                              anything else if A is a full matrix
!>                              Please note: This pertains to the original A (as set in the calling program)
!>                                           whereas the transpose of A is used for calculations
!>                              If uplo_a is 'U' or 'L', the other triangle is not used at all,
!>                              i.e. it may contain arbitrary numbers
!> \param uplo_c                'U' if only the upper diagonal part of C is needed
!>                              'L' if only the upper diagonal part of C is needed
!>                              anything else if the full matrix C is needed
!>                              Please note: Even when uplo_c is 'U' or 'L', the other triangle may be
!>                                            written to a certain extent, i.e. one shouldn't rely on the content there!
!> \param na                    Number of rows/columns of A, number of rows of B and C
!> \param ncb                   Number of columns  of B and C
!> \param a                     matrix a
!> \param lda                   leading dimension of matrix a
!> \param b                     matrix b
!> \param ldb                   leading dimension of matrix b
!> \param nblk                  blocksize of cyclic distribution, must be the same in both directions!
!> \param  mpi_comm_rows        MPI communicator for rows
!> \param  mpi_comm_cols        MPI communicator for columns
!> \param c                     matrix c
!> \param ldc                   leading dimension of matrix c
  interface mult_ah_b_complex
238
    module procedure elpa_mult_ah_b_complex_double
239
240
241
  end interface


242
!> \brief  solve_tridi: Old, deprecated interface to solve a double-precision tridiagonal eigensystem for a double-precision matrix with divide and conquer method
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
!> \details
!>
!> \param na                    Matrix dimension
!> \param nev                   number of eigenvalues/vectors to be computed
!> \param d                     array d(na) on input diagonal elements of tridiagonal matrix, on
!>                              output the eigenvalues in ascending order
!> \param e                     array e(na) on input subdiagonal elements of matrix, on exit destroyed
!> \param q                     on exit : matrix q(ldq,matrixCols) contains the eigenvectors
!> \param ldq                   leading dimension of matrix q
!> \param nblk                  blocksize of cyclic distribution, must be the same in both directions!
!> \param matrixCols            columns of matrix q
!> \param mpi_comm_rows         MPI communicator for rows
!> \param mpi_comm_cols         MPI communicator for columns
!> \param wantDebug             logical, give more debug information if .true.
!> \result success              logical, .true. on success, else .false.
  interface solve_tridi
259
    module procedure elpa_solve_tridi_double
260
  end interface
261

262
263
  contains

264
!> \brief  cholesky_real_double: Cholesky factorization of a double-precision real symmetric matrix
265
!> \details
266
267
268
269
270
271
272
273
274
!>
!> \param  na                   Order of matrix
!> \param  a(lda,matrixCols)    Distributed matrix which should be factorized.
!>                              Distribution is like in Scalapack.
!>                              Only upper triangle is needs to be set.
!>                              On return, the upper triangle contains the Cholesky factor
!>                              and the lower triangle is set to 0.
!> \param  lda                  Leading dimension of a
!> \param  nblk                 blocksize of cyclic distribution, must be the same in both directions!
275
!> \param  matrixCols           local columns of matrix a
276
277
278
279
!> \param  mpi_comm_rows        MPI communicator for rows
!> \param  mpi_comm_cols        MPI communicator for columns
!> \param wantDebug             logical, more debug information on failure
!> \param succes                logical, reports success or failure
280
281
282
283
284
285

#undef DOUBLE_PRECISION_REAL
#undef REAL_DATATYPE
#define DOUBLE_PRECISION_REAL 1
#define REAL_DATATYPE rk8

286
287
   function elpa_cholesky_real_double(na, a, lda, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols, &
                                            wantDebug) result(success)
288
289
290
     use elpa1_compute
     use elpa_utilities
     use elpa_mpi
291
292
293
294
295
296
297
#ifdef HAVE_DETAILED_TIMINGS
      use timings
#endif
      use precision
      implicit none

      integer(kind=ik)              :: na, lda, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols
298
      real(kind=REAL_DATATYPE)                 :: a(lda,matrixCols)
299
300
301
302
303
304
305
306
307
      ! was
      ! real a(lda, *)

      integer(kind=ik)              :: my_prow, my_pcol, np_rows, np_cols, mpierr
      integer(kind=ik)              :: l_cols, l_rows, l_col1, l_row1, l_colx, l_rowx
      integer(kind=ik)              :: n, nc, i, info
      integer(kind=ik)              :: lcs, lce, lrs, lre
      integer(kind=ik)              :: tile_size, l_rows_tile, l_cols_tile

308
      real(kind=REAL_DATATYPE), allocatable    :: tmp1(:), tmp2(:,:), tmatr(:,:), tmatc(:,:)
309
310

      logical, intent(in)           :: wantDebug
311
      logical                       :: success
312
313
314
315
      integer(kind=ik)              :: istat
      character(200)                :: errorMessage

#ifdef HAVE_DETAILED_TIMINGS
316
#ifdef DOUBLE_PRECISION_REAL
317
      call timer%start("elpa_cholesky_real_double")
318
#else
319
      call timer%start("elpa_cholesky_real_single")
320
#endif
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
#endif
      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)
      success = .true.

      ! Matrix is split into tiles; work is done only for tiles on the diagonal or above

      tile_size = nblk*least_common_multiple(np_rows,np_cols) ! minimum global tile size
      tile_size = ((128*max(np_rows,np_cols)-1)/tile_size+1)*tile_size ! make local tiles at least 128 wide

      l_rows_tile = tile_size/np_rows ! local rows of a tile
      l_cols_tile = tile_size/np_cols ! local cols of a tile

      l_rows = local_index(na, my_prow, np_rows, nblk, -1) ! Local rows of a
      l_cols = local_index(na, my_pcol, np_cols, nblk, -1) ! Local cols of a

      allocate(tmp1(nblk*nblk), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
341
        print *,"elpa_cholesky_real: error when allocating tmp1 "//errorMessage
342
343
344
345
346
        stop
      endif

      allocate(tmp2(nblk,nblk), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
347
        print *,"elpa_cholesky_real: error when allocating tmp2 "//errorMessage
348
349
350
351
352
353
354
355
        stop
      endif

      tmp1 = 0
      tmp2 = 0

      allocate(tmatr(l_rows,nblk), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
356
        print *,"elpa_cholesky_real: error when allocating tmatr "//errorMessage
357
358
359
360
361
        stop
      endif

      allocate(tmatc(l_cols,nblk), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
362
        print *,"elpa_cholesky_real: error when allocating tmatc "//errorMessage
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
        stop
      endif

      tmatr = 0
      tmatc = 0

      do n = 1, na, nblk
        ! Calculate first local row and column of the still remaining matrix
        ! on the local processor

        l_row1 = local_index(n, my_prow, np_rows, nblk, +1)
        l_col1 = local_index(n, my_pcol, np_cols, nblk, +1)

        l_rowx = local_index(n+nblk, my_prow, np_rows, nblk, +1)
        l_colx = local_index(n+nblk, my_pcol, np_cols, nblk, +1)

        if (n+nblk > na) then

          ! This is the last step, just do a Cholesky-Factorization
          ! of the remaining block

          if (my_prow==prow(n, nblk, np_rows) .and. my_pcol==pcol(n, nblk, np_cols)) then
385
386
387
388
389
#ifdef DOUBLE_PRECISION_REAL
            call dpotrf('U', na-n+1, a(l_row1,l_col1), lda, info)
#else
            call spotrf('U', na-n+1, a(l_row1,l_col1), lda, info)
#endif
390
            if (info/=0) then
391
              if (wantDebug) write(error_unit,*) "elpa_cholesky_real: Error in dpotrf 1: ",info
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
              success = .false.
              return
            endif

          endif

          exit ! Loop

        endif

        if (my_prow==prow(n, nblk, np_rows)) then

          if (my_pcol==pcol(n, nblk, np_cols)) then

            ! The process owning the upper left remaining block does the
            ! Cholesky-Factorization of this block
408
409
410
411
412
#ifdef DOUBLE_PRECISION_REAL
            call dpotrf('U', nblk, a(l_row1,l_col1), lda, info)
#else
            call spotrf('U', nblk, a(l_row1,l_col1), lda, info)
#endif
413
            if (info/=0) then
414
              if (wantDebug) write(error_unit,*) "elpa_cholesky_real: Error in dpotrf 2: ",info
415
416
417
418
419
420
421
422
423
424
425
              success = .false.
              return
            endif

            nc = 0
            do i=1,nblk
              tmp1(nc+1:nc+i) = a(l_row1:l_row1+i-1,l_col1+i-1)
              nc = nc+i
            enddo
          endif
#ifdef WITH_MPI
426
427
428
429
430

#ifdef DOUBLE_PRECISION_REAL
          call MPI_Bcast(tmp1, nblk*(nblk+1)/2, MPI_REAL8, pcol(n, nblk, np_cols), mpi_comm_cols, mpierr)
#else
          call MPI_Bcast(tmp1, nblk*(nblk+1)/2, MPI_REAL4, pcol(n, nblk, np_cols), mpi_comm_cols, mpierr)
431
#endif
432
433

#endif /* WITH_MPI */
434
435
436
437
438
439
440
          nc = 0
          do i=1,nblk
            tmp2(1:i,i) = tmp1(nc+1:nc+i)
            nc = nc+i
          enddo

          if (l_cols-l_colx+1>0) &
441
442
443
444
445
#ifdef DOUBLE_PRECISION_REAL
              call dtrsm('L', 'U', 'T', 'N', nblk, l_cols-l_colx+1, 1.0_rk8, tmp2, ubound(tmp2,dim=1), a(l_row1,l_colx), lda)
#else
              call strsm('L', 'U', 'T', 'N', nblk, l_cols-l_colx+1, 1.0_rk4, tmp2, ubound(tmp2,dim=1), a(l_row1,l_colx), lda)
#endif
446
447
448
449
450
451
452
        endif

        do i=1,nblk

          if (my_prow==prow(n, nblk, np_rows)) tmatc(l_colx:l_cols,i) = a(l_row1+i-1,l_colx:l_cols)
#ifdef WITH_MPI
          if (l_cols-l_colx+1>0) &
453
454
455
456
#ifdef DOUBLE_PRECISION_REAL
              call MPI_Bcast(tmatc(l_colx,i), l_cols-l_colx+1, MPI_REAL8, prow(n, nblk, np_rows), mpi_comm_rows, mpierr)
#else
              call MPI_Bcast(tmatc(l_colx,i), l_cols-l_colx+1, MPI_REAL4, prow(n, nblk, np_rows), mpi_comm_rows, mpierr)
457
#endif
458
459

#endif /* WITH_MPI */
460
461
        enddo
        ! this has to be checked since it was changed substantially when doing type safe
462
463
464
465
466
467
#ifdef DOUBLE_PRECISION_REAL
        call elpa_transpose_vectors_real_double  (tmatc, ubound(tmatc,dim=1), mpi_comm_cols, &
                                      tmatr, ubound(tmatr,dim=1), mpi_comm_rows, &
                                      n, na, nblk, nblk)
#else
        call elpa_transpose_vectors_real_single  (tmatc, ubound(tmatc,dim=1), mpi_comm_cols, &
468
469
                                      tmatr, ubound(tmatr,dim=1), mpi_comm_rows, &
                                      n, na, nblk, nblk)
470
#endif
471
472
473
474
475
476
477

        do i=0,(na-1)/tile_size
          lcs = max(l_colx,i*l_cols_tile+1)
          lce = min(l_cols,(i+1)*l_cols_tile)
          lrs = l_rowx
          lre = min(l_rows,(i+1)*l_rows_tile)
          if (lce<lcs .or. lre<lrs) cycle
478
479
480
481
482
483
484
485
486
#ifdef DOUBLE_PRECISION_REAL
          call DGEMM('N', 'T', lre-lrs+1, lce-lcs+1, nblk, -1.0_rk8,                        &
                     tmatr(lrs,1), ubound(tmatr,dim=1), tmatc(lcs,1), ubound(tmatc,dim=1), &
                     1.0_rk8, a(lrs,lcs), lda)
#else
          call SGEMM('N', 'T', lre-lrs+1, lce-lcs+1, nblk, -1.0_rk4,                        &
                     tmatr(lrs,1), ubound(tmatr,dim=1), tmatc(lcs,1), ubound(tmatc,dim=1), &
                     1.0_rk4, a(lrs,lcs), lda)
#endif
487
488
489
490
491
492
        enddo

      enddo

      deallocate(tmp1, tmp2, tmatr, tmatc, stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
493
        print *,"elpa_cholesky_real: error when deallocating tmp1 "//errorMessage
494
495
496
497
498
499
500
501
502
503
504
505
506
507
        stop
      endif

      ! Set the lower triangle to 0, it contains garbage (form the above matrix multiplications)

      do i=1,na
        if (my_pcol==pcol(i, nblk, np_cols)) then
          ! column i is on local processor
          l_col1 = local_index(i  , my_pcol, np_cols, nblk, +1) ! local column number
          l_row1 = local_index(i+1, my_prow, np_rows, nblk, +1) ! first row below diagonal
          a(l_row1:l_rows,l_col1) = 0
        endif
      enddo
#ifdef HAVE_DETAILED_TIMINGS
508
#ifdef DOUBLE_PRECISION_REAL
509
      call timer%stop("elpa_cholesky_real_double")
510
#else
511
      call timer%stop("elpa_cholesky_real_single")
512
#endif
513
514
#endif

515
    end function elpa_cholesky_real_double
516

517
518
519
520
#ifdef WANT_SINGLE_PRECISION_REAL
#undef DOUBLE_PRECISION_REAL
#undef REAL_DATATYPE
#define REAL_DATATYPE rk4
521

522
!> \brief  cholesky_real_single: Cholesky factorization of a single-precision real symmetric matrix
523
!> \details
524
!>
525
!> \param  na                   Order of matrix
526
!> \param  a(lda,matrixCols)    Distributed matrix which should be factorized.
527
528
!>                              Distribution is like in Scalapack.
!>                              Only upper triangle is needs to be set.
529
530
!>                              On return, the upper triangle contains the Cholesky factor
!>                              and the lower triangle is set to 0.
531
532
!> \param  lda                  Leading dimension of a
!> \param  nblk                 blocksize of cyclic distribution, must be the same in both directions!
Andreas Marek's avatar
Andreas Marek committed
533
!> \param  matrixCols           local columns of matrix a
534
535
536
537
538
!> \param  mpi_comm_rows        MPI communicator for rows
!> \param  mpi_comm_cols        MPI communicator for columns
!> \param wantDebug             logical, more debug information on failure
!> \param succes                logical, reports success or failure

539
540
   function elpa_cholesky_real_single(na, a, lda, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols, &
                                            wantDebug) result(success)
541
542
543
544
545
     use elpa1_compute
     use elpa_utilities
     use elpa_mpi
#ifdef HAVE_DETAILED_TIMINGS
      use timings
546
#endif
547
548
      use precision
      implicit none
549

550
551
552
553
      integer(kind=ik)              :: na, lda, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols
      real(kind=REAL_DATATYPE)                 :: a(lda,matrixCols)
      ! was
      ! real a(lda, *)
554

555
556
557
558
559
      integer(kind=ik)              :: my_prow, my_pcol, np_rows, np_cols, mpierr
      integer(kind=ik)              :: l_cols, l_rows, l_col1, l_row1, l_colx, l_rowx
      integer(kind=ik)              :: n, nc, i, info
      integer(kind=ik)              :: lcs, lce, lrs, lre
      integer(kind=ik)              :: tile_size, l_rows_tile, l_cols_tile
560

561
      real(kind=REAL_DATATYPE), allocatable    :: tmp1(:), tmp2(:,:), tmatr(:,:), tmatc(:,:)
562

563
      logical, intent(in)           :: wantDebug
564
      logical                       :: success
565
566
      integer(kind=ik)              :: istat
      character(200)                :: errorMessage
567

568
569
#ifdef HAVE_DETAILED_TIMINGS
#ifdef DOUBLE_PRECISION_REAL
570
      call timer%start("elpa_cholesky_real_double")
571
#else
572
      call timer%start("elpa_cholesky_real_single")
573
574
575
576
577
578
579
#endif
#endif
      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)
      success = .true.
580

581
      ! Matrix is split into tiles; work is done only for tiles on the diagonal or above
582

583
584
      tile_size = nblk*least_common_multiple(np_rows,np_cols) ! minimum global tile size
      tile_size = ((128*max(np_rows,np_cols)-1)/tile_size+1)*tile_size ! make local tiles at least 128 wide
585
586
587
588
589
590
591
592
593

      l_rows_tile = tile_size/np_rows ! local rows of a tile
      l_cols_tile = tile_size/np_cols ! local cols of a tile

      l_rows = local_index(na, my_prow, np_rows, nblk, -1) ! Local rows of a
      l_cols = local_index(na, my_pcol, np_cols, nblk, -1) ! Local cols of a

      allocate(tmp1(nblk*nblk), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
594
        print *,"elpa_cholesky_real: error when allocating tmp1 "//errorMessage
595
596
597
598
599
        stop
      endif

      allocate(tmp2(nblk,nblk), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
600
        print *,"elpa_cholesky_real: error when allocating tmp2 "//errorMessage
601
602
603
604
605
606
607
608
        stop
      endif

      tmp1 = 0
      tmp2 = 0

      allocate(tmatr(l_rows,nblk), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
609
        print *,"elpa_cholesky_real: error when allocating tmatr "//errorMessage
610
611
612
613
614
        stop
      endif

      allocate(tmatc(l_cols,nblk), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
615
        print *,"elpa_cholesky_real: error when allocating tmatc "//errorMessage
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
        stop
      endif

      tmatr = 0
      tmatc = 0

      do n = 1, na, nblk

        ! Calculate first local row and column of the still remaining matrix
        ! on the local processor

        l_row1 = local_index(n, my_prow, np_rows, nblk, +1)
        l_col1 = local_index(n, my_pcol, np_cols, nblk, +1)

        l_rowx = local_index(n+nblk, my_prow, np_rows, nblk, +1)
        l_colx = local_index(n+nblk, my_pcol, np_cols, nblk, +1)

        if (n+nblk > na) then

          ! This is the last step, just do a Cholesky-Factorization
          ! of the remaining block

          if (my_prow==prow(n, nblk, np_rows) .and. my_pcol==pcol(n, nblk, np_cols)) then
639
640
641
642
643
#ifdef DOUBLE_PRECISION_REAL
            call dpotrf('U', na-n+1, a(l_row1,l_col1), lda, info)
#else
            call spotrf('U', na-n+1, a(l_row1,l_col1), lda, info)
#endif
644
            if (info/=0) then
645
              if (wantDebug) write(error_unit,*) "elpa_cholesky_real: Error in dpotrf"
646
647
648
649
650
651
652
              success = .false.
              return
            endif

          endif

          exit ! Loop
653

654
655
656
657
658
659
660
661
        endif

        if (my_prow==prow(n, nblk, np_rows)) then

          if (my_pcol==pcol(n, nblk, np_cols)) then

            ! The process owning the upper left remaining block does the
            ! Cholesky-Factorization of this block
662
663
664
665
666
#ifdef DOUBLE_PRECISION_REAL
            call dpotrf('U', nblk, a(l_row1,l_col1), lda, info)
#else
            call spotrf('U', nblk, a(l_row1,l_col1), lda, info)
#endif
667
            if (info/=0) then
668
              if (wantDebug) write(error_unit,*) "elpa_cholesky_real: Error in dpotrf"
669
670
671
672
673
674
675
676
677
678
679
              success = .false.
              return
            endif

            nc = 0
            do i=1,nblk
              tmp1(nc+1:nc+i) = a(l_row1:l_row1+i-1,l_col1+i-1)
              nc = nc+i
            enddo
          endif
#ifdef WITH_MPI
680
681
682
683
684

#ifdef DOUBLE_PRECISION_REAL
          call MPI_Bcast(tmp1, nblk*(nblk+1)/2, MPI_REAL8, pcol(n, nblk, np_cols), mpi_comm_cols, mpierr)
#else
          call MPI_Bcast(tmp1, nblk*(nblk+1)/2, MPI_REAL4, pcol(n, nblk, np_cols), mpi_comm_cols, mpierr)
685
#endif
686
687

#endif /* WITH_MPI */
688
689
690
691
692
693
694
          nc = 0
          do i=1,nblk
            tmp2(1:i,i) = tmp1(nc+1:nc+i)
            nc = nc+i
          enddo

          if (l_cols-l_colx+1>0) &
695
696
697
698
699
#ifdef DOUBLE_PRECISION_REAL
              call dtrsm('L', 'U', 'T', 'N', nblk, l_cols-l_colx+1, 1.0_rk8, tmp2, ubound(tmp2,dim=1), a(l_row1,l_colx), lda)
#else
              call strsm('L', 'U', 'T', 'N', nblk, l_cols-l_colx+1, 1.0_rk4, tmp2, ubound(tmp2,dim=1), a(l_row1,l_colx), lda)
#endif
700
701
702
703
        endif

        do i=1,nblk

704
          if (my_prow==prow(n, nblk, np_rows)) tmatc(l_colx:l_cols,i) = a(l_row1+i-1,l_colx:l_cols)
705
706
#ifdef WITH_MPI
          if (l_cols-l_colx+1>0) &
707
708
709
710
#ifdef DOUBLE_PRECISION_REAL
              call MPI_Bcast(tmatc(l_colx,i), l_cols-l_colx+1, MPI_REAL8, prow(n, nblk, np_rows), mpi_comm_rows, mpierr)
#else
              call MPI_Bcast(tmatc(l_colx,i), l_cols-l_colx+1, MPI_REAL4, prow(n, nblk, np_rows), mpi_comm_rows, mpierr)
711
#endif
712
713

#endif /* WITH_MPI */
714
715
        enddo
        ! this has to be checked since it was changed substantially when doing type safe
716
717
718
719
720
721
722
723
724
725
#ifdef DOUBLE_PRECISION_REAL
        call elpa_transpose_vectors_real_double  (tmatc, ubound(tmatc,dim=1), mpi_comm_cols, &
                                      tmatr, ubound(tmatr,dim=1), mpi_comm_rows, &
                                      n, na, nblk, nblk)
#else
        call elpa_transpose_vectors_real_single  (tmatc, ubound(tmatc,dim=1), mpi_comm_cols, &
                                      tmatr, ubound(tmatr,dim=1), mpi_comm_rows, &
                                      n, na, nblk, nblk)
#endif

726
727
728
729
730
731
        do i=0,(na-1)/tile_size
          lcs = max(l_colx,i*l_cols_tile+1)
          lce = min(l_cols,(i+1)*l_cols_tile)
          lrs = l_rowx
          lre = min(l_rows,(i+1)*l_rows_tile)
          if (lce<lcs .or. lre<lrs) cycle
732
733
734
735
736
737
738
739
740
#ifdef DOUBLE_PRECISION_REAL
          call DGEMM('N', 'T', lre-lrs+1, lce-lcs+1, nblk, -1.0_rk8,                        &
                     tmatr(lrs,1), ubound(tmatr,dim=1), tmatc(lcs,1), ubound(tmatc,dim=1), &
                     1.0_rk8, a(lrs,lcs), lda)
#else
          call SGEMM('N', 'T', lre-lrs+1, lce-lcs+1, nblk, -1.0_rk4,                        &
                     tmatr(lrs,1), ubound(tmatr,dim=1), tmatc(lcs,1), ubound(tmatc,dim=1), &
                     1.0_rk4, a(lrs,lcs), lda)
#endif
741
742
743
744
745
746
        enddo

      enddo

      deallocate(tmp1, tmp2, tmatr, tmatc, stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
747
        print *,"elpa_cholesky_real: error when deallocating tmp1 "//errorMessage
748
749
750
751
752
753
754
755
756
757
758
759
760
761
        stop
      endif

      ! Set the lower triangle to 0, it contains garbage (form the above matrix multiplications)

      do i=1,na
        if (my_pcol==pcol(i, nblk, np_cols)) then
          ! column i is on local processor
          l_col1 = local_index(i  , my_pcol, np_cols, nblk, +1) ! local column number
          l_row1 = local_index(i+1, my_prow, np_rows, nblk, +1) ! first row below diagonal
          a(l_row1:l_rows,l_col1) = 0
        endif
      enddo
#ifdef HAVE_DETAILED_TIMINGS
762
#ifdef DOUBLE_PRECISION_REAL
763
      call timer%stop("elpa_cholesky_real_double")
764
#else
765
      call timer%stop("elpa_cholesky_real_single")
766
#endif
767
768
#endif

769
    end function elpa_cholesky_real_single
770

771
#endif /* WANT_SINGLE_PRECSION_REAL */
772

773
774
775
776
777
#undef DOUBLE_PRECISION_REAL
#undef REAL_DATATYPE
#define DOUBLE_PRECISION_REAL 1
#define REAL_DATATYPE rk8
!> \brief  elpa_invert_trm_real_double: Inverts a double-precision real upper triangular matrix
778
!> \details
779
780
781
782
783
784
785
!> \param  na                   Order of matrix
!> \param  a(lda,matrixCols)    Distributed matrix which should be inverted
!>                              Distribution is like in Scalapack.
!>                              Only upper triangle is needs to be set.
!>                              The lower triangle is not referenced.
!> \param  lda                  Leading dimension of a
!> \param  nblk                 blocksize of cyclic distribution, must be the same in both directions!
Andreas Marek's avatar
Andreas Marek committed
786
!> \param  matrixCols           local columns of matrix a
787
788
789
!> \param  mpi_comm_rows        MPI communicator for rows
!> \param  mpi_comm_cols        MPI communicator for columns
!> \param wantDebug             logical, more debug information on failure
790
791
!> \result succes               logical, reports success or failure
    function elpa_invert_trm_real_double(na, a, lda, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols, wantDebug) result(success)
792
793
794
795
796
797
       use precision
       use elpa1_compute
       use elpa_utilities
       use elpa_mpi
       implicit none

798
       integer(kind=ik)             :: na, lda, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols
799
#ifdef DESPERATELY_WANT_ASSUMED_SIZE
800
       real(kind=REAL_DATATYPE)                :: a(lda,*)
801
#else
802
       real(kind=REAL_DATATYPE)                :: a(lda,matrixCols)
803
#endif
804
805
806
       integer(kind=ik)             :: my_prow, my_pcol, np_rows, np_cols, mpierr
       integer(kind=ik)             :: l_cols, l_rows, l_col1, l_row1, l_colx, l_rowx
       integer(kind=ik)             :: n, nc, i, info, ns, nb
807

808
       real(kind=REAL_DATATYPE), allocatable   :: tmp1(:), tmp2(:,:), tmat1(:,:), tmat2(:,:)
809

810
       logical, intent(in)          :: wantDebug
811
       logical                      :: success
812
813
       integer(kind=ik)             :: istat
       character(200)               :: errorMessage
814

815
816
817
818
819
820
821
822
823
824
825
       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)
       success = .true.

       l_rows = local_index(na, my_prow, np_rows, nblk, -1) ! Local rows of a
       l_cols = local_index(na, my_pcol, np_cols, nblk, -1) ! Local cols of a

       allocate(tmp1(nblk*nblk), stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
826
         print *,"elpa_invert_trm_real: error when allocating tmp1 "//errorMessage
827
828
829
830
831
         stop
       endif

       allocate(tmp2(nblk,nblk), stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
832
         print *,"elpa_invert_trm_real: error when allocating tmp2 "//errorMessage
833
834
835
836
837
838
839
840
         stop
       endif

       tmp1 = 0
       tmp2 = 0

       allocate(tmat1(l_rows,nblk), stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
841
         print *,"elpa_invert_trm_real: error when allocating tmat1 "//errorMessage
842
843
844
845
846
         stop
       endif

       allocate(tmat2(nblk,l_cols), stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
847
         print *,"elpa_invert_trm_real: error when allocating tmat2 "//errorMessage
848
849
850
         stop
       endif

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
       tmat1 = 0
       tmat2 = 0


       ns = ((na-1)/nblk)*nblk + 1

       do n = ns,1,-nblk

         l_row1 = local_index(n, my_prow, np_rows, nblk, +1)
         l_col1 = local_index(n, my_pcol, np_cols, nblk, +1)

         nb = nblk
         if (na-n+1 < nblk) nb = na-n+1

         l_rowx = local_index(n+nb, my_prow, np_rows, nblk, +1)
         l_colx = local_index(n+nb, my_pcol, np_cols, nblk, +1)

         if (my_prow==prow(n, nblk, np_rows)) then

           if (my_pcol==pcol(n, nblk, np_cols)) then
#ifdef DOUBLE_PRECISION_REAL
             call DTRTRI('U', 'N', nb, a(l_row1,l_col1), lda, info)
#else
             call STRTRI('U', 'N', nb, a(l_row1,l_col1), lda, info)
#endif
             if (info/=0) then
877
               if (wantDebug) write(error_unit,*) "elpa_invert_trm_real: Error in DTRTRI"
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
950
951
952
953
954
               success = .false.
               return
             endif

             nc = 0
             do i=1,nb
               tmp1(nc+1:nc+i) = a(l_row1:l_row1+i-1,l_col1+i-1)
               nc = nc+i
             enddo
           endif
#ifdef WITH_MPI

#ifdef DOUBLE_PRECISION_REAL
           call MPI_Bcast(tmp1, nb*(nb+1)/2, MPI_REAL8, pcol(n, nblk, np_cols), mpi_comm_cols, mpierr)
#else
           call MPI_Bcast(tmp1, nb*(nb+1)/2, MPI_REAL4, pcol(n, nblk, np_cols), mpi_comm_cols, mpierr)
#endif

#endif /* WITH_MPI */
           nc = 0
           do i=1,nb
             tmp2(1:i,i) = tmp1(nc+1:nc+i)
             nc = nc+i
           enddo

           if (l_cols-l_colx+1>0) &
#ifdef DOUBLE_PRECISION_REAL
               call DTRMM('L', 'U', 'N', 'N', nb, l_cols-l_colx+1, 1.0_rk8, tmp2, ubound(tmp2,dim=1), a(l_row1,l_colx), lda)
#else
               call STRMM('L', 'U', 'N', 'N', nb, l_cols-l_colx+1, 1.0_rk4, tmp2, ubound(tmp2,dim=1), a(l_row1,l_colx), lda)
#endif
           if (l_colx<=l_cols)   tmat2(1:nb,l_colx:l_cols) = a(l_row1:l_row1+nb-1,l_colx:l_cols)
           if (my_pcol==pcol(n, nblk, np_cols)) tmat2(1:nb,l_col1:l_col1+nb-1) = tmp2(1:nb,1:nb) ! tmp2 has the lower left triangle 0

         endif

         if (l_row1>1) then
           if (my_pcol==pcol(n, nblk, np_cols)) then
             tmat1(1:l_row1-1,1:nb) = a(1:l_row1-1,l_col1:l_col1+nb-1)
             a(1:l_row1-1,l_col1:l_col1+nb-1) = 0
           endif

           do i=1,nb
#ifdef WITH_MPI

#ifdef DOUBLE_PRECISION_REAL
             call MPI_Bcast(tmat1(1,i), l_row1-1, MPI_REAL8, pcol(n, nblk, np_cols), mpi_comm_cols, mpierr)
#else
             call MPI_Bcast(tmat1(1,i), l_row1-1, MPI_REAL4, pcol(n, nblk, np_cols), mpi_comm_cols, mpierr)
#endif

#endif /* WITH_MPI */
           enddo
         endif
#ifdef WITH_MPI
         if (l_cols-l_col1+1>0) &
#ifdef DOUBLE_PRECISION_REAL
            call MPI_Bcast(tmat2(1,l_col1), (l_cols-l_col1+1)*nblk, MPI_REAL8, prow(n, nblk, np_rows), mpi_comm_rows, mpierr)
#else
            call MPI_Bcast(tmat2(1,l_col1), (l_cols-l_col1+1)*nblk, MPI_REAL4, prow(n, nblk, np_rows), mpi_comm_rows, mpierr)
#endif

#endif /* WITH_MPI */
         if (l_row1>1 .and. l_cols-l_col1+1>0) &
#ifdef DOUBLE_PRECISION_REAL
            call dgemm('N', 'N', l_row1-1, l_cols-l_col1+1, nb, -1.0_rk8,                 &
                       tmat1, ubound(tmat1,dim=1), tmat2(1,l_col1), ubound(tmat2,dim=1), &
                       1.0_rk8, a(1,l_col1), lda)
#else
            call sgemm('N', 'N', l_row1-1, l_cols-l_col1+1, nb, -1.0_rk4,                 &
                       tmat1, ubound(tmat1,dim=1), tmat2(1,l_col1), ubound(tmat2,dim=1), &
                       1.0_rk4, a(1,l_col1), lda)
#endif
       enddo

       deallocate(tmp1, tmp2, tmat1, tmat2, stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
955
         print *,"elpa_invert_trm_real: error when deallocating tmp1 "//errorMessage
956
957
         stop
       endif
958
     end function elpa_invert_trm_real_double
959
960
961
962
963

#if WANT_SINGLE_PRECISION_REAL
#undef DOUBLE_PRECISION_REAL
#undef REAL_DATATYPE
#define REAL_DATATYPE rk4
964
!> \brief  elpa_invert_trm_real_single: Inverts a single-precision real upper triangular matrix
965
966
967
968
969
970
971
972
973
974
975
976
!> \details
!> \param  na                   Order of matrix
!> \param  a(lda,matrixCols)    Distributed matrix which should be inverted
!>                              Distribution is like in Scalapack.
!>                              Only upper triangle is needs to be set.
!>                              The lower triangle is not referenced.
!> \param  lda                  Leading dimension of a
!> \param                       matrixCols  local columns of matrix a
!> \param  nblk                 blocksize of cyclic distribution, must be the same in both directions!
!> \param  mpi_comm_rows        MPI communicator for rows
!> \param  mpi_comm_cols        MPI communicator for columns
!> \param wantDebug             logical, more debug information on failure
977
978
!> \result succes               logical, reports success or failure
    function elpa_invert_trm_real_single(na, a, lda, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols, wantDebug) result(success)
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
       use precision
       use elpa1_compute
       use elpa_utilities
       use elpa_mpi
       implicit none

       integer(kind=ik)             :: na, lda, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols
#ifdef DESPERATELY_WANT_ASSUMED_SIZE
       real(kind=REAL_DATATYPE)                :: a(lda,*)
#else
       real(kind=REAL_DATATYPE)                :: a(lda,matrixCols)
#endif
       integer(kind=ik)             :: my_prow, my_pcol, np_rows, np_cols, mpierr
       integer(kind=ik)             :: l_cols, l_rows, l_col1, l_row1, l_colx, l_rowx
       integer(kind=ik)             :: n, nc, i, info, ns, nb

       real(kind=REAL_DATATYPE), allocatable   :: tmp1(:), tmp2(:,:), tmat1(:,:), tmat2(:,:)

       logical, intent(in)          :: wantDebug
998
       logical                      :: success
999
1000
       integer(kind=ik)             :: istat
       character(200)               :: errorMessage
1001

1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
       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)
       success = .true.

       l_rows = local_index(na, my_prow, np_rows, nblk, -1) ! Local rows of a
       l_cols = local_index(na, my_pcol, np_cols, nblk, -1) ! Local cols of a

       allocate(tmp1(nblk*nblk), stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
1013
         print *,"elpa_invert_trm_real: error when allocating tmp1 "//errorMessage
1014
1015
1016
1017
1018
         stop
       endif

       allocate(tmp2(nblk,nblk), stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
1019
         print *,"elpa_invert_trm_real: error when allocating tmp2 "//errorMessage
1020
1021
1022
1023
1024
1025
1026
1027
         stop
       endif

       tmp1 = 0
       tmp2 = 0

       allocate(tmat1(l_rows,nblk), stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
1028
         print *,"elpa_invert_trm_real: error when allocating tmat1 "//errorMessage
1029
1030
1031
1032
1033
         stop
       endif

       allocate(tmat2(nblk,l_cols), stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
1034
         print *,"elpa_invert_trm_real: error when allocating tmat2 "//errorMessage
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
         stop
       endif

       tmat1 = 0
       tmat2 = 0


       ns = ((na-1)/nblk)*nblk + 1

       do n = ns,1,-nblk

         l_row1 = local_index(n, my_prow, np_rows, nblk, +1)
         l_col1 = local_index(n, my_pcol, np_cols, nblk, +1)

         nb = nblk
         if (na-n+1 < nblk) nb = na-n+1

         l_rowx = local_index(n+nb, my_prow, np_rows, nblk, +1)
         l_colx = local_index(n+nb, my_pcol, np_cols, nblk, +1)

         if (my_prow==prow(n, nblk, np_rows)) then

           if (my_pcol==pcol(n, nblk, np_cols)) then
#ifdef DOUBLE_PRECISION_REAL
             call DTRTRI('U', 'N', nb, a(l_row1,l_col1), lda, info)
#else
             call STRTRI('U', 'N', nb, a(l_row1,l_col1), lda, info)
#endif
             if (info/=0) then
1064
               if (wantDebug) write(error_unit,*) "elpa_invert_trm_real: Error in DTRTRI"
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
               success = .false.
               return
             endif

             nc = 0
             do i=1,nb
               tmp1(nc+1:nc+i) = a(l_row1:l_row1+i-1,l_col1+i-1)
               nc = nc+i
             enddo
           endif
#ifdef WITH_MPI

#ifdef DOUBLE_PRECISION_REAL
           call MPI_Bcast(tmp1, nb*(nb+1)/2, MPI_REAL8, pcol(n, nblk, np_cols), mpi_comm_cols, mpierr)
#else
           call MPI_Bcast(tmp1, nb*(nb+1)/2, MPI_REAL4, pcol(n, nblk, np_cols), mpi_comm_cols, mpierr)
#endif

#endif /* WITH_MPI */
           nc = 0
           do i=1,nb
             tmp2(1:i,i) = tmp1(nc+1:nc+i)
             nc = nc+i
           enddo

           if (l_cols-l_colx+1>0) &
#ifdef DOUBLE_PRECISION_REAL
               call DTRMM('L', 'U', 'N', 'N', nb, l_cols-l_colx+1, 1.0_rk8, tmp2, ubound(tmp2,dim=1), a(l_row1,l_colx), lda)
#else
               call STRMM('L', 'U', 'N', 'N', nb, l_cols-l_colx+1, 1.0_rk4, tmp2, ubound(tmp2,dim=1), a(l_row1,l_colx), lda)
#endif
           if (l_colx<=l_cols)   tmat2(1:nb,l_colx:l_cols) = a(l_row1:l_row1+nb-1,l_colx:l_cols)
           if (my_pcol==pcol(n, nblk, np_cols)) tmat2(1:nb,l_col1:l_col1+nb-1) = tmp2(1:nb,1:nb) ! tmp2 has the lower left triangle 0

         endif

         if (l_row1>1) then
           if (my_pcol==pcol(n, nblk, np_cols)) then
             tmat1(1:l_row1-1,1:nb) = a(1:l_row1-1,l_col1:l_col1+nb-1)
             a(1:l_row1-1,l_col1:l_col1+nb-1) = 0
           endif

           do i=1,nb
#ifdef WITH_MPI

#ifdef DOUBLE_PRECISION_REAL
             call MPI_Bcast(tmat1(1,i), l_row1-1, MPI_REAL8, pcol(n, nblk, np_cols), mpi_comm_cols, mpierr)
#else
             call MPI_Bcast(tmat1(1,i), l_row1-1, MPI_REAL4, pcol(n, nblk, np_cols), mpi_comm_cols, mpierr)
#endif

#endif /* WITH_MPI */
           enddo
         endif
#ifdef WITH_MPI
         if (l_cols-l_col1+1>0) &
#ifdef DOUBLE_PRECISION_REAL
            call MPI_Bcast(tmat2(1,l_col1), (l_cols-l_col1+1)*nblk, MPI_REAL8, prow(n, nblk, np_rows), mpi_comm_rows, mpierr)
#else
            call MPI_Bcast(tmat2(1,l_col1), (l_cols-l_col1+1)*nblk, MPI_REAL4, prow(n, nblk, np_rows), mpi_comm_rows, mpierr)
#endif

#endif /* WITH_MPI */
         if (l_row1>1 .and. l_cols-l_col1+1>0) &
#ifdef DOUBLE_PRECISION_REAL
            call dgemm('N', 'N', l_row1-1, l_cols-l_col1+1, nb, -1.0_rk8,                 &
                       tmat1, ubound(tmat1,dim=1), tmat2(1,l_col1), ubound(tmat2,dim=1), &
                       1.0_rk8, a(1,l_col1), lda)
#else
            call sgemm('N', 'N', l_row1-1, l_cols-l_col1+1, nb, -1.0_rk4,                 &
                       tmat1, ubound(tmat1,dim=1), tmat2(1,l_col1), ubound(tmat2,dim=1), &
                       1.0_rk4, a(1,l_col1), lda)
#endif
       enddo

       deallocate(tmp1, tmp2, tmat1, tmat2, stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
1142
         print *,"elpa_invert_trm_real: error when deallocating tmp1 "//errorMessage
1143
1144
         stop
       endif
1145
     end function elpa_invert_trm_real_single
1146
1147
1148

#endif /* WANT_SINGLE_PRECISION_REAL */

1149
1150
1151
1152
1153
1154

#undef DOUBLE_PRECISION_COMPLEX
#undef COMPLEX_DATATYPE
#define DOUBLE_PRECISION_COMPLEX 1
#define COMPLEX_DATATYPE CK8
!> \brief  elpa_cholesky_complex_double: Cholesky factorization of a double-precision complex hermitian matrix
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
!> \details
!> \param  na                   Order of matrix
!> \param  a(lda,matrixCols)    Distributed matrix which should be factorized.
!>                              Distribution is like in Scalapack.
!>                              Only upper triangle is needs to be set.
!>                              On return, the upper triangle contains the Cholesky factor
!>                              and the lower triangle is set to 0.
!> \param  lda                  Leading dimension of a
!> \param                       matrixCols  local columns of matrix a
!> \param  nblk                 blocksize of cyclic distribution, must be the same in both directions!
!> \param  mpi_comm_rows        MPI communicator for rows
!> \param  mpi_comm_cols        MPI communicator for columns
!> \param wantDebug             logical, more debug information on failure
1168
1169
!> \result succes               logical, reports success or failure
    function elpa_cholesky_complex_double(na, a, lda, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols, wantDebug) result(success)
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

      use elpa1_compute
      use elpa_utilities
      use elpa_mpi
#ifdef HAVE_DETAILED_TIMINGS
      use timings
#endif
      use precision
      implicit none

      integer(kind=ik)                 :: na, lda, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols
#ifdef DESPERATELY_WANT_ASSUMED_SIZE
      complex(kind=COMPLEX_DATATYPE)                 :: a(lda,*)
#else
      complex(kind=COMPLEX_DATATYPE)                 :: a(lda,matrixCols)
#endif
      integer(kind=ik)                 :: my_prow, my_pcol, np_rows, np_cols, mpierr
      integer(kind=ik)                 :: l_cols, l_rows, l_col1, l_row1, l_colx, l_rowx
      integer(kind=ik)                 :: n, nc, i, info
      integer(kind=ik)                 :: lcs, lce, lrs, lre
      integer(kind=ik)                 :: tile_size, l_rows_tile, l_cols_tile

      complex(kind=COMPLEX_DATATYPE), allocatable    :: tmp1(:), tmp2(:,:), tmatr(:,:), tmatc(:,:)

      logical, intent(in)              :: wantDebug
1195
      logical                          :: success
1196
1197
1198
1199
1200
      integer(kind=ik)                 :: istat
      character(200)                   :: errorMessage

#ifdef HAVE_DETAILED_TIMINGS
#ifdef DOUBLE_PRECISION_COMPLEX
1201
      call timer%start("elpa_cholesky_complex_double")
1202
#else
1203
      call timer%start("elpa_cholesky_complex_single")
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
#endif
#endif
      success = .true.
      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)
      ! Matrix is split into tiles; work is done only for tiles on the diagonal or above

      tile_size = nblk*least_common_multiple(np_rows,np_cols) ! minimum global tile size
      tile_size = ((128*max(np_rows,np_cols)-1)/tile_size+1)*tile_size ! make local tiles at least 128 wide

      l_rows_tile = tile_size/np_rows ! local rows of a tile
      l_cols_tile = tile_size/np_cols ! local cols of a tile

      l_rows = local_index(na, my_prow, np_rows, nblk, -1) ! Local rows of a
      l_cols = local_index(na, my_pcol, np_cols, nblk, -1) ! Local cols of a

      allocate(tmp1(nblk*nblk), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
1224
        print *,"elpa_cholesky_complex: error when allocating tmp1 "//errorMessage
1225
1226
1227
1228
1229
        stop
      endif

      allocate(tmp2(nblk,nblk), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
1230
        print *,"elpa_cholesky_complex: error when allocating tmp2 "//errorMessage
1231
1232
1233
1234
1235
1236
1237
1238
        stop
      endif

      tmp1 = 0
      tmp2 = 0

      allocate(tmatr(l_rows,nblk), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
1239
        print *,"elpa_cholesky_complex: error when allocating tmatr "//errorMessage
1240
1241
1242
1243
1244
        stop
      endif

      allocate(tmatc(l_cols,nblk), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
1245
        print *,"elpa_cholesky_complex: error when allocating tmatc "//errorMessage
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
        stop
      endif

      tmatr = 0
      tmatc = 0

      do n = 1, na, nblk

        ! Calculate first local row and column of the still remaining matrix
        ! on the local processor