elpa1_auxiliary.F90 132 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
!    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,
!    - Max-Plack-Institut für Mathematik in den Naturwissenschaftrn,
!      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
122
123
124
125
!> \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
126
!> \param result                logical, reports success or failure
127
128

  interface invert_trm_real
129
    module procedure elpa_invert_trm_real_double
130
131
132
  end interface


133
!> \brief  old, deprecated interface cholesky_complex: Cholesky factorization of a double-precision complex hermitian matrix
134
135
136
137
138
139
140
141
142
143
144
145
146
!> \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
147
!> \result succes               logical, reports success or failure
148
149
150


  interface cholesky_complex
151
    module procedure elpa_cholesky_real_double
152
153
  end interface

154
!> \brief  old, deprecated interface invert_trm_complex: Inverts a double-precision complex upper triangular matrix
155
156
157
158
159
160
161
162
163
164
165
166
!> \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
167
!> \result succes               logical, reports success or failure
168
169

  interface invert_trm_complex
170
    module procedure elpa_invert_trm_complex_double
171
172
  end interface

173
!> \brief  mult_at_b_real: Performs C : = A**T * B for double matrices
174
!> this is the old, deprecated interface for the newer elpa_mult_at_b_real
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
!>         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
205
    module procedure elpa_mult_at_b_real_double
206
207
  end interface

208
!> \brief  Old, deprecated interface mult_ah_b_complex: Performs C : = A**H * B for double-precision matrices
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
238
!>         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
239
    module procedure elpa_mult_ah_b_complex_double
240
241
242
  end interface


243
!> \brief  solve_tridi: Old, deprecated interface to solve a double-precision tridiagonal eigensystem for a double-precision matrix with divide and conquer method
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
!> \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
260
    module procedure elpa_solve_tridi_double
261
  end interface
262

263
264
  contains

265
!> \brief  cholesky_real_double: Cholesky factorization of a double-precision real symmetric matrix
266
!> \details
267
268
269
270
271
272
273
274
275
276
277
278
279
280
!>
!> \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
!> \param succes                logical, reports success or failure
281
282
283
284
285
286

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

287
288
   function elpa_cholesky_real_double(na, a, lda, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols, &
                                            wantDebug) result(success)
289
290
291
     use elpa1_compute
     use elpa_utilities
     use elpa_mpi
292
293
294
295
296
297
298
#ifdef HAVE_DETAILED_TIMINGS
      use timings
#endif
      use precision
      implicit none

      integer(kind=ik)              :: na, lda, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols
299
      real(kind=REAL_DATATYPE)                 :: a(lda,matrixCols)
300
301
302
303
304
305
306
307
308
      ! 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

309
      real(kind=REAL_DATATYPE), allocatable    :: tmp1(:), tmp2(:,:), tmatr(:,:), tmatc(:,:)
310
311

      logical, intent(in)           :: wantDebug
312
      logical                       :: success
313
314
315
316
      integer(kind=ik)              :: istat
      character(200)                :: errorMessage

#ifdef HAVE_DETAILED_TIMINGS
317
#ifdef DOUBLE_PRECISION_REAL
318
      call timer%start("elpa_cholesky_real_double")
319
#else
320
      call timer%start("elpa_cholesky_real_single")
321
#endif
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
#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
342
        print *,"elpa_cholesky_real: error when allocating tmp1 "//errorMessage
343
344
345
346
347
        stop
      endif

      allocate(tmp2(nblk,nblk), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
348
        print *,"elpa_cholesky_real: error when allocating tmp2 "//errorMessage
349
350
351
352
353
354
355
356
        stop
      endif

      tmp1 = 0
      tmp2 = 0

      allocate(tmatr(l_rows,nblk), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
357
        print *,"elpa_cholesky_real: error when allocating tmatr "//errorMessage
358
359
360
361
362
        stop
      endif

      allocate(tmatc(l_cols,nblk), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
363
        print *,"elpa_cholesky_real: error when allocating tmatc "//errorMessage
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
        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
387
388
389
390
391
#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
392
            if (info/=0) then
393
              if (wantDebug) write(error_unit,*) "elpa_cholesky_real: Error in dpotrf"
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
              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
410
411
412
413
414
#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
415
            if (info/=0) then
416
              if (wantDebug) write(error_unit,*) "elpa_cholesky_real: Error in dpotrf"
417
418
419
420
421
422
423
424
425
426
427
              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
428
429
430
431
432

#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)
433
#endif
434
435

#endif /* WITH_MPI */
436
437
438
439
440
441
442
          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) &
443
444
445
446
447
#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
448
449
450
451
452
453
454
        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) &
455
456
457
458
#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)
459
#endif
460
461

#endif /* WITH_MPI */
462
463
        enddo
        ! this has to be checked since it was changed substantially when doing type safe
464
465
466
467
468
469
#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, &
470
471
                                      tmatr, ubound(tmatr,dim=1), mpi_comm_rows, &
                                      n, na, nblk, nblk)
472
#endif
473
474
475
476
477
478
479

        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
480
481
482
483
484
485
486
487
488
#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
489
490
491
492
493
494
        enddo

      enddo

      deallocate(tmp1, tmp2, tmatr, tmatc, stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
495
        print *,"elpa_cholesky_real: error when deallocating tmp1 "//errorMessage
496
497
498
499
500
501
502
503
504
505
506
507
508
509
        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
510
#ifdef DOUBLE_PRECISION_REAL
511
      call timer%stop("elpa_cholesky_real_double")
512
#else
513
      call timer%stop("elpa_cholesky_real_single")
514
#endif
515
516
#endif

517
    end function elpa_cholesky_real_double
518

519
520
521
522
#ifdef WANT_SINGLE_PRECISION_REAL
#undef DOUBLE_PRECISION_REAL
#undef REAL_DATATYPE
#define REAL_DATATYPE rk4
523

524
!> \brief  cholesky_real_single: Cholesky factorization of a single-precision real symmetric matrix
525
!> \details
526
!>
527
!> \param  na                   Order of matrix
528
!> \param  a(lda,matrixCols)    Distributed matrix which should be factorized.
529
530
!>                              Distribution is like in Scalapack.
!>                              Only upper triangle is needs to be set.
531
532
!>                              On return, the upper triangle contains the Cholesky factor
!>                              and the lower triangle is set to 0.
533
534
535
536
537
538
539
540
!> \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
!> \param succes                logical, reports success or failure

541
542
   function elpa_cholesky_real_single(na, a, lda, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols, &
                                            wantDebug) result(success)
543
544
545
546
547
     use elpa1_compute
     use elpa_utilities
     use elpa_mpi
#ifdef HAVE_DETAILED_TIMINGS
      use timings
548
#endif
549
550
      use precision
      implicit none
551

552
553
554
555
      integer(kind=ik)              :: na, lda, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols
      real(kind=REAL_DATATYPE)                 :: a(lda,matrixCols)
      ! was
      ! real a(lda, *)
556

557
558
559
560
561
      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
562

563
      real(kind=REAL_DATATYPE), allocatable    :: tmp1(:), tmp2(:,:), tmatr(:,:), tmatc(:,:)
564

565
      logical, intent(in)           :: wantDebug
566
      logical                       :: success
567
568
      integer(kind=ik)              :: istat
      character(200)                :: errorMessage
569

570
571
#ifdef HAVE_DETAILED_TIMINGS
#ifdef DOUBLE_PRECISION_REAL
572
      call timer%start("elpa_cholesky_real_double")
573
#else
574
      call timer%start("elpa_cholesky_real_single")
575
576
577
578
579
580
581
#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.
582

583
      ! Matrix is split into tiles; work is done only for tiles on the diagonal or above
584

585
586
      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
587
588
589
590
591
592
593
594
595

      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
596
        print *,"elpa_cholesky_real: error when allocating tmp1 "//errorMessage
597
598
599
600
601
        stop
      endif

      allocate(tmp2(nblk,nblk), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
602
        print *,"elpa_cholesky_real: error when allocating tmp2 "//errorMessage
603
604
605
606
607
608
609
610
        stop
      endif

      tmp1 = 0
      tmp2 = 0

      allocate(tmatr(l_rows,nblk), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
611
        print *,"elpa_cholesky_real: error when allocating tmatr "//errorMessage
612
613
614
615
616
        stop
      endif

      allocate(tmatc(l_cols,nblk), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
617
        print *,"elpa_cholesky_real: error when allocating tmatc "//errorMessage
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
        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
641
642
643
644
645
#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
646
            if (info/=0) then
647
              if (wantDebug) write(error_unit,*) "elpa_cholesky_real: Error in dpotrf"
648
649
650
651
652
653
654
              success = .false.
              return
            endif

          endif

          exit ! Loop
655

656
657
658
659
660
661
662
663
        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
664
665
666
667
668
#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
669
            if (info/=0) then
670
              if (wantDebug) write(error_unit,*) "elpa_cholesky_real: Error in dpotrf"
671
672
673
674
675
676
677
678
679
680
681
              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
682
683
684
685
686

#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)
687
#endif
688
689

#endif /* WITH_MPI */
690
691
692
693
694
695
696
          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) &
697
698
699
700
701
#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
702
703
704
705
        endif

        do i=1,nblk

706
          if (my_prow==prow(n, nblk, np_rows)) tmatc(l_colx:l_cols,i) = a(l_row1+i-1,l_colx:l_cols)
707
708
#ifdef WITH_MPI
          if (l_cols-l_colx+1>0) &
709
710
711
712
#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)
713
#endif
714
715

#endif /* WITH_MPI */
716
717
        enddo
        ! this has to be checked since it was changed substantially when doing type safe
718
719
720
721
722
723
724
725
726
727
#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

728
729
730
731
732
733
        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
734
735
736
737
738
739
740
741
742
#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
743
744
745
746
747
748
        enddo

      enddo

      deallocate(tmp1, tmp2, tmatr, tmatc, stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
749
        print *,"elpa_cholesky_real: error when deallocating tmp1 "//errorMessage
750
751
752
753
754
755
756
757
758
759
760
761
762
763
        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
764
#ifdef DOUBLE_PRECISION_REAL
765
      call timer%stop("elpa_cholesky_real_double")
766
#else
767
      call timer%stop("elpa_cholesky_real_single")
768
#endif
769
770
#endif

771
    end function elpa_cholesky_real_single
772

773
#endif /* WANT_SINGLE_PRECSION_REAL */
774

775
776
777
778
779
#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
780
!> \details
781
782
783
784
785
786
787
788
789
790
791
!> \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
792
793
!> \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)
794
795
796
797
798
799
       use precision
       use elpa1_compute
       use elpa_utilities
       use elpa_mpi
       implicit none

800
       integer(kind=ik)             :: na, lda, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols
801
#ifdef DESPERATELY_WANT_ASSUMED_SIZE
802
       real(kind=REAL_DATATYPE)                :: a(lda,*)
803
#else
804
       real(kind=REAL_DATATYPE)                :: a(lda,matrixCols)
805
#endif
806
807
808
       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
809

810
       real(kind=REAL_DATATYPE), allocatable   :: tmp1(:), tmp2(:,:), tmat1(:,:), tmat2(:,:)
811

812
       logical, intent(in)          :: wantDebug
813
       logical                      :: success
814
815
       integer(kind=ik)             :: istat
       character(200)               :: errorMessage
816

817
818
819
820
821
822
823
824
825
826
827
       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
828
         print *,"elpa_invert_trm_real: error when allocating tmp1 "//errorMessage
829
830
831
832
833
         stop
       endif

       allocate(tmp2(nblk,nblk), stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
834
         print *,"elpa_invert_trm_real: error when allocating tmp2 "//errorMessage
835
836
837
838
839
840
841
842
         stop
       endif

       tmp1 = 0
       tmp2 = 0

       allocate(tmat1(l_rows,nblk), stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
843
         print *,"elpa_invert_trm_real: error when allocating tmat1 "//errorMessage
844
845
846
847
848
         stop
       endif

       allocate(tmat2(nblk,l_cols), stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
849
         print *,"elpa_invert_trm_real: error when allocating tmat2 "//errorMessage
850
851
852
         stop
       endif

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
       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
879
               if (wantDebug) write(error_unit,*) "elpa_invert_trm_real: Error in DTRTRI"
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
955
956
               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
957
         print *,"elpa_invert_trm_real: error when deallocating tmp1 "//errorMessage
958
959
         stop
       endif
960
     end function elpa_invert_trm_real_double
961
962
963
964
965

#if WANT_SINGLE_PRECISION_REAL
#undef DOUBLE_PRECISION_REAL
#undef REAL_DATATYPE
#define REAL_DATATYPE rk4
966
!> \brief  elpa_invert_trm_real_single: Inverts a single-precision real upper triangular matrix
967
968
969
970
971
972
973
974
975
976
977
978
!> \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
979
980
!> \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)
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
       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
1000
       logical                      :: success
1001
1002
       integer(kind=ik)             :: istat
       character(200)               :: errorMessage
1003

1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
       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
1015
         print *,"elpa_invert_trm_real: error when allocating tmp1 "//errorMessage
1016
1017
1018
1019
1020
         stop
       endif

       allocate(tmp2(nblk,nblk), stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
1021
         print *,"elpa_invert_trm_real: error when allocating tmp2 "//errorMessage
1022
1023
1024
1025
1026
1027
1028
1029
         stop
       endif

       tmp1 = 0
       tmp2 = 0

       allocate(tmat1(l_rows,nblk), stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
1030
         print *,"elpa_invert_trm_real: error when allocating tmat1 "//errorMessage
1031
1032
1033
1034
1035
         stop
       endif

       allocate(tmat2(nblk,l_cols), stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
1036
         print *,"elpa_invert_trm_real: error when allocating tmat2 "//errorMessage
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
         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
1066
               if (wantDebug) write(error_unit,*) "elpa_invert_trm_real: Error in DTRTRI"
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
               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
1144
         print *,"elpa_invert_trm_real: error when deallocating tmp1 "//errorMessage
1145
1146
         stop
       endif
1147
     end function elpa_invert_trm_real_single
1148
1149
1150

#endif /* WANT_SINGLE_PRECISION_REAL */

1151
1152
1153
1154
1155
1156

#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
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
!> \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
1170
1171
!> \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)
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

      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
1197
      logical                          :: success
1198
1199
1200
1201
1202
      integer(kind=ik)                 :: istat
      character(200)                   :: errorMessage

#ifdef HAVE_DETAILED_TIMINGS
#ifdef DOUBLE_PRECISION_COMPLEX
1203
      call timer%start("elpa_cholesky_complex_double")
1204
#else
1205
      call timer%start("elpa_cholesky_complex_single")
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
#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
1226
        print *,"elpa_cholesky_complex: error when allocating tmp1 "//errorMessage
1227
1228
1229
1230
1231
        stop
      endif

      allocate(tmp2(nblk,nblk), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
1232
        print *,"elpa_cholesky_complex: error when allocating tmp2 "//errorMessage
1233
1234
1235
1236
1237
1238
1239
1240
        stop
      endif

      tmp1 = 0
      tmp2 = 0

      allocate(tmatr(l_rows,nblk), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
1241
        print *,"elpa_cholesky_complex: error when allocating tmatr "//errorMessage
1242
1243
1244
1245
1246
        stop
      endif

      allocate(tmatc(l_cols,nblk), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
1247
        print *,"elpa_cholesky_complex: error when allocating tmatc "//errorMessage
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
1275
1276
        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
#ifdef DOUBLE_PRECISION_COMPLEX
            call zpotrf('U', na-n+1, a(l_row1,l_col1),lda, info)
#else
            call cpotrf('U', na-n+1, a(l_row1,l_col1),lda, info)
#endif
            if (info/=0) then
1277
              if (wantDebug) write(error_unit,*) "elpa_cholesky_complex: Error in zpotrf"
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
              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
#ifdef DOUBLE_PRECISION_COMPLEX
            call zpotrf('U', nblk, a(l_row1,l_col1),lda, info)
#else
            call cpotrf('U', nblk, a(l_row1,l_col1),lda, info)
#endif
            if (info/=0) then
1299
              if (wantDebug) write(error_unit,*) "elpa_cholesky_complex: Error in zpotrf"
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
              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

#ifdef DOUBLE_PRECISION_COMPLEX
          call MPI_Bcast(tmp1, nblk*(nblk+1)/2, MPI_DOUBLE_COMPLEX, pcol(n, nblk, np_cols), mpi_comm_cols, mpierr)
#else
          call MPI_Bcast(tmp1, nblk*(nblk+1)/2, MPI_COMPLEX, pcol(n, nblk, np_cols), mpi_comm_cols, mpierr)
#endif

#endif /* WITH_MPI */

          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) &
#ifdef DOUBLE_PRECISION_COMPLEX
                call ztrsm('L', 'U', 'C', 'N', nblk, l_cols-l_colx+1, (1.0_rk8,0.0_rk8), tmp2, ubound(tmp2,dim=1), &
                           a(l_row1,l_colx), lda)
#else
                call ctrsm('L', 'U', 'C', 'N', nblk, l_cols-l_colx+1, (1.0_rk4,0.0_rk4), tmp2, ubound(tmp2,dim=1), &
                           a(l_row1,l_colx), lda)
#endif
        endif

        do i=1,nblk

          if (my_prow==prow(n, nblk, np_rows)) tmatc(l_colx:l_cols,i) = conjg(a(l_row1+i-1,l_colx:l_cols))
#ifdef WITH_MPI
          if (l_cols-l_colx+1>0) &
#ifdef DOUBLE_PRECISION_COMPLEX
                call MPI_Bcast(tmatc(l_colx,i), l_cols-l_colx+1, MPI_DOUBLE_COMPLEX, prow(n, nblk, np_rows), &
                               mpi_comm_rows, mpierr)
#else
                call MPI_Bcast(tmatc(l_colx,i), l_cols-l_colx+1, MPI_COMPLEX, prow(n, nblk, np_rows), &
                               mpi_comm_rows, mpierr)
#endif

#endif /* WITH_MPI */
        enddo
        ! this has to be checked since it was changed substantially when doing type safe
#ifdef DOUBLE_PRECISION_COMPLEX
        call elpa_transpose_vectors_complex_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_complex_single  (tmatc, ubound(tmatc,dim=1), mpi_comm_cols, &
                                        tmatr, ubound(tmatr,dim=1), mpi_comm_rows, &
                                        n, na, nblk, nblk)
#endif
        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
#ifdef DOUBLE_PRECISION_COMPLEX
          call ZGEMM('N', 'C', lre-lrs+1, lce-lcs+1, nblk, (-1.0_rk8,0.0_rk8),               &
                     tmatr(lrs,1), ubound(tmatr,dim=1), tmatc(lcs,1), ubound(tmatc,dim=1), &
                     (1.0_rk8,0.0_rk8), a(lrs,lcs), lda)
#else
          call CGEMM('N', 'C', lre-lrs+1, lce-lcs+1, nblk, (-1.0_rk4,0.0_rk4),               &
                     tmatr(lrs,1), ubound(tmatr,dim=1), tmatc(lcs,1), ubound(tmatc,dim=1), &
                     (1.0_rk4,0.0_rk4), a(lrs,lcs), lda)
#endif
        enddo

      enddo

      deallocate(tmp1, tmp2, tmatr, tmatc, stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
1382
        print *,"elpa_cholesky_complex: error when deallocating tmatr "//errorMessage
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396