elpa2_compute.F90 347 KB
Newer Older
Andreas Marek's avatar
Andreas Marek committed
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
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
!    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), fomerly 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".



! ELPA2 -- 2-stage solver for ELPA
!
! Copyright of the original code rests with the authors inside the ELPA
! consortium. The copyright of any additional modifications shall rest
! with their original authors, but shall adhere to the licensing terms
! distributed along with the original code in the file "COPYING".


#include "config-f90.h"

module ELPA2_compute

! Version 1.1.2, 2011-02-21

  use elpa_utilities
  USE ELPA1_compute
  use elpa1, only : elpa_print_times, time_evp_back, time_evp_fwd, time_evp_solve
  use elpa2_utilities
  use elpa_pdgeqrf

  implicit none

  PRIVATE ! By default, all routines contained are private

  public :: bandred_real
  public :: tridiag_band_real
  public :: trans_ev_tridi_to_band_real
  public :: trans_ev_band_to_full_real

  public :: bandred_complex
  public :: tridiag_band_complex
  public :: trans_ev_tridi_to_band_complex
  public :: trans_ev_band_to_full_complex

  public :: band_band_real
  public :: divide_band

  integer, public :: which_qr_decomposition = 1     ! defines, which QR-decomposition algorithm will be used
                                                    ! 0 for unblocked
                                                    ! 1 for blocked (maxrank: nblk)
  include 'mpif.h'

  contains

99
    subroutine bandred_real(na, a, lda, nblk, nbw, matrixCols, numBlocks, mpi_comm_rows, mpi_comm_cols, &
100
                            tmat, wantDebug, useGPU, success, useQR)
Andreas Marek's avatar
Andreas Marek committed
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134

  !-------------------------------------------------------------------------------
  !  bandred_real: Reduces a distributed symmetric matrix to band form
  !
  !  Parameters
  !
  !  na          Order of matrix
  !
  !  a(lda,matrixCols)    Distributed matrix which should be reduced.
  !              Distribution is like in Scalapack.
  !              Opposed to Scalapack, a(:,:) must be set completely (upper and lower half)
  !              a(:,:) is overwritten on exit with the band and the Householder vectors
  !              in the upper half.
  !
  !  lda         Leading dimension of a
  !  matrixCols  local columns of matrix a
  !
  !  nblk        blocksize of cyclic distribution, must be the same in both directions!
  !
  !  nbw         semi bandwith of output matrix
  !
  !  mpi_comm_rows
  !  mpi_comm_cols
  !              MPI-Communicators for rows/columns
  !
  !  tmat(nbw,nbw,numBlocks)    where numBlocks = (na-1)/nbw + 1
  !              Factors for the Householder vectors (returned), needed for back transformation
  !
  !-------------------------------------------------------------------------------

    use cuda_functions
    use iso_c_binding

#ifdef HAVE_DETAILED_TIMINGS
135
      use timings
Andreas Marek's avatar
Andreas Marek committed
136
#endif
137
138
139
140
#ifdef WITH_OPENMP
      use omp_lib
#endif
      implicit none
Andreas Marek's avatar
Andreas Marek committed
141

142
143
      integer             :: na, lda, nblk, nbw, matrixCols, numBlocks, mpi_comm_rows, mpi_comm_cols
      real*8              :: a(lda,matrixCols), tmat(nbw,nbw,numBlocks)
Andreas Marek's avatar
Andreas Marek committed
144

145
      real*8              :: eps
Andreas Marek's avatar
Andreas Marek committed
146

147
      logical, intent(in) :: useGPU
Andreas Marek's avatar
Andreas Marek committed
148

149
150
151
152
153
      integer             :: my_prow, my_pcol, np_rows, np_cols, mpierr
      integer             :: l_cols, l_rows
      integer             :: i, j, lcs, lce, lrs, lre, lc, lr, cur_pcol, n_cols, nrow
      integer             :: istep, ncol, lch, lcx, nlc, mynlc
      integer             :: tile_size, l_rows_tile, l_cols_tile
Andreas Marek's avatar
Andreas Marek committed
154

155
      real*8              :: vnorm2, xf, aux1(nbw), aux2(nbw), vrl, tau, vav(nbw,nbw)
Andreas Marek's avatar
Andreas Marek committed
156

157
158
159
      real*8, allocatable :: tmpCUDA(:),  vmrCUDA(:),  umcCUDA(:)
      real*8, allocatable :: tmpCPU(:,:), vmrCPU(:,:), umcCPU(:,:)
      real*8, allocatable :: vr(:)
Andreas Marek's avatar
Andreas Marek committed
160

161
      ! needed for blocked QR decomposition
162
163
164
      integer                  :: PQRPARAM(11), work_size
      real*8                   :: dwork_size(1)
      real*8, allocatable      :: work_blocked(:), tauvector(:), blockheuristic(:)
Andreas Marek's avatar
Andreas Marek committed
165

166
167
168
169
170
171
172
      integer(kind=C_intptr_T) :: a_dev, vmr_dev, umc_dev, tmat_dev, vav_dev
      integer, external        :: numroc
      integer                  :: ierr
      integer                  :: cur_l_rows, cur_l_cols, vmr_size, umc_size
      integer(kind=c_size_t)   :: lc_start, lc_end
      integer                  :: lr_end
      integer                  :: na_rows, na_cols
Andreas Marek's avatar
Andreas Marek committed
173
174


175
176
177
      logical             :: successCUDA
      integer             :: istat
      character(200)      :: errorMessage
178
179
      logical, intent(in) :: wantDebug
      logical, intent(out):: success
Andreas Marek's avatar
Andreas Marek committed
180

181
      logical, intent(in) :: useQR
Andreas Marek's avatar
Andreas Marek committed
182

183
      integer :: mystart, myend, m_way, n_way, work_per_thread, m_id, n_id, n_threads, ii, pp, transformChunkSize
Andreas Marek's avatar
Andreas Marek committed
184
185

#ifdef HAVE_DETAILED_TIMINGS
186
      call timer%start("bandred_real")
Andreas Marek's avatar
Andreas Marek committed
187
#endif
188
189
190
191
192
      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.
Andreas Marek's avatar
Andreas Marek committed
193
194


195
196
197
198
199
200
201
202
203
204
205
      ! Semibandwith nbw must be a multiple of blocksize nblk
      if (mod(nbw,nblk)/=0) then
        if (my_prow==0 .and. my_pcol==0) then
          if (wantDebug) then
            write(error_unit,*) 'ELPA2_bandred_real: ERROR: nbw=',nbw,', nblk=',nblk
            write(error_unit,*) 'ELPA2_bandred_real: ELPA2 works only for nbw==n*nblk'
          endif
          success = .false.
          return
        endif
      endif
Andreas Marek's avatar
Andreas Marek committed
206

207
208
209
210
      if (useGPU) then
        na_rows = numroc(na, nblk, my_prow, 0, np_rows)
        na_cols = numroc(na, nblk, my_pcol, 0, np_cols)
      endif
Andreas Marek's avatar
Andreas Marek committed
211

212
      ! Matrix is split into tiles; work is done only for tiles on the diagonal or above
Andreas Marek's avatar
Andreas Marek committed
213

214
215
      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
Andreas Marek's avatar
Andreas Marek committed
216

217
218
      l_rows_tile = tile_size/np_rows ! local rows of a tile
      l_cols_tile = tile_size/np_cols ! local cols of a tile
Andreas Marek's avatar
Andreas Marek committed
219

220
      if (useQR) then
Andreas Marek's avatar
Andreas Marek committed
221

222
223
224
225
        if (useGPU) then
          print *,"qr decomposition at the moment not supported with GPU"
          stop
        endif
Andreas Marek's avatar
Andreas Marek committed
226

227
228
        if (which_qr_decomposition == 1) then
          call qr_pqrparam_init(pqrparam,    nblk,'M',0,   nblk,'M',0,   nblk,'M',1,'s')
229
230
231
232
233
          allocate(tauvector(na), stat=istat, errmsg=errorMessage)
          if (istat .ne. 0) then
            print *,"bandred_real: error when allocating tauvector "//errorMessage
            stop
          endif
Andreas Marek's avatar
Andreas Marek committed
234

235
236
237
238
239
          allocate(blockheuristic(nblk), stat=istat, errmsg=errorMessage)
          if (istat .ne. 0) then
            print *,"bandred_real: error when allocating blockheuristic "//errorMessage
            stop
          endif
Andreas Marek's avatar
Andreas Marek committed
240

241
          l_rows = local_index(na, my_prow, np_rows, nblk, -1)
242
243
244
245
246
          allocate(vmrCPU(max(l_rows,1),na), stat=istat, errmsg=errorMessage)
          if (istat .ne. 0) then
            print *,"bandred_real: error when allocating vmrCPU "//errorMessage
            stop
          endif
Andreas Marek's avatar
Andreas Marek committed
247

248
          call qr_pdgeqrf_2dcomm(a, lda, vmrCPU, max(l_rows,1), tauvector, tmat(1,1,1), nbw, dwork_size(1), -1, na, &
249
250
                                nbw, nblk, nblk, na, na, 1, 0, PQRPARAM, mpi_comm_rows, mpi_comm_cols, blockheuristic)
          work_size = dwork_size(1)
251
252
253
254
255
          allocate(work_blocked(work_size), stat=istat, errmsg=errorMessage)
          if (istat .ne. 0) then
            print *,"bandred_real: error when allocating work_blocked "//errorMessage
            stop
          endif
Andreas Marek's avatar
Andreas Marek committed
256

257
          work_blocked = 0.0d0
258
259
260
261
262
          deallocate(vmrCPU, stat=istat, errmsg=errorMessage)
          if (istat .ne. 0) then
            print *,"bandred_real: error when deallocating vmrCPU "//errorMessage
            stop
          endif
Andreas Marek's avatar
Andreas Marek committed
263

264
        endif ! which_qr_decomposition
Andreas Marek's avatar
Andreas Marek committed
265

266
      endif ! useQr
Andreas Marek's avatar
Andreas Marek committed
267

268
269
270
271
272
273
      if (useGPU) then
        ! Here we convert the regular host array into a pinned host array
        successCUDA = cuda_malloc(a_dev, lda*na_cols*size_of_real_datatype)
        if (.not.(successCUDA)) then
          print *,"bandred_real: error in cudaMalloc"
          stop
274
        endif
Andreas Marek's avatar
Andreas Marek committed
275

276
277
278
279
280
        successCUDA = cuda_malloc(tmat_dev, nbw*nbw*size_of_real_datatype)
        if (.not.(successCUDA)) then
          print *,"bandred_real: error in cudaMalloc"
          stop
        endif
Andreas Marek's avatar
Andreas Marek committed
281

282
283
284
285
286
        successCUDA = cuda_malloc(vav_dev, nbw*nbw*size_of_real_datatype)
        if (.not.(successCUDA)) then
          print *,"bandred_real: error in cudaMalloc"
          stop
        endif
Andreas Marek's avatar
Andreas Marek committed
287

288
289
        cur_l_rows = 0
        cur_l_cols = 0
Andreas Marek's avatar
Andreas Marek committed
290

291
292
293
294
295
296
        successCUDA = cuda_memcpy(a_dev, loc(a(1,1)), (lda)*(na_cols)*size_of_real_datatype,cudaMemcpyHostToDevice)
        if (.not.(successCUDA)) then
          print *,"bandred_real: error in cudaMemcpy"
          stop
        endif
      endif ! useGPU
Andreas Marek's avatar
Andreas Marek committed
297
298


299
      do istep = (na-1)/nbw, 1, -1
Andreas Marek's avatar
Andreas Marek committed
300

301
        n_cols = MIN(na,(istep+1)*nbw) - istep*nbw ! Number of columns in current step
Andreas Marek's avatar
Andreas Marek committed
302

303
304
305
        ! Number of local columns/rows of remaining matrix
        l_cols = local_index(istep*nbw, my_pcol, np_cols, nblk, -1)
        l_rows = local_index(istep*nbw, my_prow, np_rows, nblk, -1)
Andreas Marek's avatar
Andreas Marek committed
306

307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
        if (useGPU) then
          cur_l_rows = max(l_rows, 1)
          cur_l_cols = max(l_cols, 1)

          vmr_size = cur_l_rows * 2 * n_cols
          umc_size = cur_l_cols * 2 * n_cols

          ! Allocate vmr and umc only if the inew size exceeds their current capacity
          ! Added for FORTRAN CALLS
          if ((.not. allocated(vr)) .or. (l_rows + 1 .gt. ubound(vr, dim=1))) then
            if (allocated(vr)) then
              deallocate(vr, stat=istat, errmsg=errorMessage)
              if (istat .ne. 0) then
                print *,"bandred_real: error when deallocating vr "//errorMessage
                stop
              endif
            endif
            allocate(vr(l_rows + 1), stat=istat, errmsg=errorMessage)
            if (istat .ne. 0) then
              print *,"bandred_real: error when allocating vr "//errorMessage
              stop
            endif
Andreas Marek's avatar
Andreas Marek committed
329

330
          endif
Andreas Marek's avatar
Andreas Marek committed
331

332
333
334
335
336
337
338
          if ((.not. allocated(vmrCUDA)) .or. (vmr_size .gt. ubound(vmrCUDA, dim=1))) then
            if (allocated(vmrCUDA)) then
              deallocate(vmrCUDA, stat=istat, errmsg=errorMessage)
              if (istat .ne. 0) then
                print *,"bandred_real: error when allocating vmrCUDA "//errorMessage
                stop
              endif
Andreas Marek's avatar
Andreas Marek committed
339

340
341
342
343
344
345
              successCUDA = cuda_free(vmr_dev)
              if (.not.(successCUDA)) then
                print *,"bandred_real: error in cuda_free"
                stop
              endif
            endif
Andreas Marek's avatar
Andreas Marek committed
346

347
348
349
350
351
            allocate(vmrCUDA(vmr_size), stat=istat, errmsg=errorMessage)
            if (istat .ne. 0) then
              print *,"bandred_real: error when allocating vmrCUDA "//errorMessage
              stop
            endif
Andreas Marek's avatar
Andreas Marek committed
352

353
354
355
356
357
            successCUDA = cuda_malloc(vmr_dev, vmr_size*size_of_real_datatype)
            if (.not.(successCUDA)) then
              print *,"bandred_real: error in cudaMalloc"
              stop
            endif
Andreas Marek's avatar
Andreas Marek committed
358

359
          endif
Andreas Marek's avatar
Andreas Marek committed
360

361
362
363
364
365
366
367
          if ((.not. allocated(umcCUDA)) .or. (umc_size .gt. ubound(umcCUDA, dim=1))) then
            if (allocated(umcCUDA)) then
              deallocate(umcCUDA, stat=istat, errmsg=errorMessage)
              if (istat .ne. 0) then
                print *,"bandred_real: error when deallocating umcCUDA "//errorMessage
                stop
              endif
Andreas Marek's avatar
Andreas Marek committed
368

369
370
371
372
373
              successCUDA = cuda_free(umc_dev)
              if (.not.(successCUDA)) then
                 print *,"bandred_real: error in cudaFree"
                 stop
              endif
Andreas Marek's avatar
Andreas Marek committed
374

375
            endif
Andreas Marek's avatar
Andreas Marek committed
376

377
378
379
380
381
            allocate(umcCUDA(umc_size), stat=istat, errmsg=errorMessage)
            if (istat .ne. 0) then
              print *,"bandred_real: error when deallocating umcCUDA "//errorMessage
              stop
            endif
Andreas Marek's avatar
Andreas Marek committed
382

383
384
385
386
387
            successCUDA = cuda_malloc(umc_dev, umc_size*size_of_real_datatype)
            if (.not.(successCUDA)) then
              print *,"bandred_real: error in cudaMalloc"
              stop
            endif
Andreas Marek's avatar
Andreas Marek committed
388

389
390
391
          endif
        else ! GPU not used
          ! Allocate vmr and umc to their exact sizes so that they can be used in bcasts and reduces
Andreas Marek's avatar
Andreas Marek committed
392

393
394
395
396
397
          allocate(vmrCPU(max(l_rows,1),2*n_cols), stat=istat, errmsg=errorMessage)
          if (istat .ne. 0) then
            print *,"bandred_real: error when allocating vmrCPU "//errorMessage
            stop
          endif
Andreas Marek's avatar
Andreas Marek committed
398

399
400
401
402
403
          allocate(umcCPU(max(l_cols,1),2*n_cols), stat=istat, errmsg=errorMessage)
          if (istat .ne. 0) then
            print *,"bandred_real: error when allocating umcCPU "//errorMessage
            stop
          endif
Andreas Marek's avatar
Andreas Marek committed
404

405
406
407
408
409
410
          allocate(vr(l_rows+1), stat=istat, errmsg=errorMessage)
          if (istat .ne. 0) then
            print *,"bandred_real: error when allocating vr "//errorMessage
            stop
          endif
        endif ! use GPU
Andreas Marek's avatar
Andreas Marek committed
411

412
413
414
415
416
        if (useGPU) then
          vmrCUDA(1 : cur_l_rows * n_cols) = 0.
        else
          vmrCPU(1:l_rows,1:n_cols) = 0.
        endif
Andreas Marek's avatar
Andreas Marek committed
417

418
419
        vr(:) = 0
        tmat(:,:,istep) = 0
Andreas Marek's avatar
Andreas Marek committed
420

421
422
        if (useGPU) then
          umcCUDA(1 : umc_size) = 0.
Andreas Marek's avatar
Andreas Marek committed
423

424
425
426
          lc_start = local_index(istep*nbw+1, my_pcol, np_cols, nblk, -1)
          lc_end   = local_index(istep*nbw+n_cols, my_pcol, np_cols, nblk, -1)
          lr_end   = local_index((istep-1)*nbw + n_cols, my_prow, np_rows, nblk, -1)
Andreas Marek's avatar
Andreas Marek committed
427

428
          if(lc_start .le. 0) lc_start = 1
Andreas Marek's avatar
Andreas Marek committed
429

430
431
          ! Here we assume that the processor grid and the block grid are aligned
          cur_pcol = pcol(istep*nbw+1, nblk, np_cols)
Andreas Marek's avatar
Andreas Marek committed
432

433
          if(my_pcol == cur_pcol) then
Andreas Marek's avatar
Andreas Marek committed
434

435
436
437
438
439
440
441
442
            successCUDA = cuda_memcpy2d(loc(a(1, lc_start)), lda*size_of_real_datatype,         &
                                       (a_dev + ((lc_start-1) * lda*size_of_real_datatype)),    &
                                       lda*size_of_real_datatype, lr_end*size_of_real_datatype, &
                                       (lc_end - lc_start+1), cudaMemcpyDeviceToHost)
            if (.not.(successCUDA)) then
              print *,"bandred_real: error in cudaMemcpy2d"
              stop
            endif
Andreas Marek's avatar
Andreas Marek committed
443

444
445
          endif
        endif ! useGPU
Andreas Marek's avatar
Andreas Marek committed
446

447
        ! Reduce current block to lower triangular form
Andreas Marek's avatar
Andreas Marek committed
448

449
450
        if (useQR) then
          if (which_qr_decomposition == 1) then
451
            call qr_pdgeqrf_2dcomm(a, lda, vmrCPU, max(l_rows,1), tauvector(1), &
Andreas Marek's avatar
Andreas Marek committed
452
453
454
455
456
                                  tmat(1,1,istep), nbw, work_blocked,       &
                                  work_size, na, n_cols, nblk, nblk,        &
                                  istep*nbw+n_cols-nbw, istep*nbw+n_cols, 1,&
                                  0, PQRPARAM, mpi_comm_rows, mpi_comm_cols,&
                                  blockheuristic)
457
          endif
458
       else !useQR
Andreas Marek's avatar
Andreas Marek committed
459

460
         do lc = n_cols, 1, -1
Andreas Marek's avatar
Andreas Marek committed
461

462
463
           ncol = istep*nbw + lc ! absolute column number of householder vector
           nrow = ncol - nbw ! Absolute number of pivot row
Andreas Marek's avatar
Andreas Marek committed
464

465
466
           lr  = local_index(nrow, my_prow, np_rows, nblk, -1) ! current row length
           lch = local_index(ncol, my_pcol, np_cols, nblk, -1) ! HV local column number
Andreas Marek's avatar
Andreas Marek committed
467

468
           tau = 0
Andreas Marek's avatar
Andreas Marek committed
469

470
           if (nrow == 1) exit ! Nothing to do
Andreas Marek's avatar
Andreas Marek committed
471

472
           cur_pcol = pcol(ncol, nblk, np_cols) ! Processor column owning current block
Andreas Marek's avatar
Andreas Marek committed
473

474
           if (my_pcol==cur_pcol) then
Andreas Marek's avatar
Andreas Marek committed
475

476
477
             ! Get vector to be transformed; distribute last element and norm of
             ! remaining elements to all procs in current column
Andreas Marek's avatar
Andreas Marek committed
478

479
             vr(1:lr) = a(1:lr,lch) ! vector to be transformed
Andreas Marek's avatar
Andreas Marek committed
480

481
482
483
484
485
486
487
             if (my_prow==prow(nrow, nblk, np_rows)) then
               aux1(1) = dot_product(vr(1:lr-1),vr(1:lr-1))
               aux1(2) = vr(lr)
             else
               aux1(1) = dot_product(vr(1:lr),vr(1:lr))
               aux1(2) = 0.
             endif
Andreas Marek's avatar
Andreas Marek committed
488

489
             call mpi_allreduce(aux1,aux2,2,MPI_REAL8,MPI_SUM,mpi_comm_rows,mpierr)
Andreas Marek's avatar
Andreas Marek committed
490

491
492
             vnorm2 = aux2(1)
             vrl    = aux2(2)
Andreas Marek's avatar
Andreas Marek committed
493

494
             ! Householder transformation
Andreas Marek's avatar
Andreas Marek committed
495

496
             call hh_transform_real(vrl, vnorm2, xf, tau)
Andreas Marek's avatar
Andreas Marek committed
497

498
             ! Scale vr and store Householder vector for back transformation
Andreas Marek's avatar
Andreas Marek committed
499

500
501
502
503
504
505
506
507
             vr(1:lr) = vr(1:lr) * xf
             if (my_prow==prow(nrow, nblk, np_rows)) then
               a(1:lr-1,lch) = vr(1:lr-1)
               a(lr,lch) = vrl
               vr(lr) = 1.
             else
               a(1:lr,lch) = vr(1:lr)
             endif
508

509
           endif
510

511
           ! Broadcast Householder vector and tau along columns
512

513
514
           vr(lr+1) = tau
           call MPI_Bcast(vr,lr+1,MPI_REAL8,cur_pcol,mpi_comm_cols,mpierr)
515

516
517
           if (useGPU) then
             vmrCUDA(cur_l_rows * (lc - 1) + 1 : cur_l_rows * (lc - 1) + lr) = vr(1:lr)
Andreas Marek's avatar
Andreas Marek committed
518
           else
519
             vmrCPU(1:lr,lc) = vr(1:lr)
Andreas Marek's avatar
Andreas Marek committed
520
521
           endif

522
523
           tau = vr(lr+1)
           tmat(lc,lc,istep) = tau ! Store tau in diagonal of tmat
524

525
526
           ! Transform remaining columns in current block with Householder vector
           ! Local dot product
527

528
           aux1 = 0
529

530
531
532
533
#ifdef WITH_OPENMP
           !Open up one omp region to avoid paying openmp overhead.
           !This does not help performance due to the addition of two openmp barriers around the MPI call,
           !But in the future this may be beneficial if these barriers are replaced with a faster implementation
534

535
536
           !$omp parallel private(mynlc, j, lcx, ii, pp ) shared(aux1)
           mynlc = 0 ! number of local columns
537

538
539
540
541
542
543
544
545
546
547
548
549
550
551
           !This loop does not have independent iterations,
           !'mynlc' is incremented each iteration, and it is difficult to remove this dependency
           !Thus each thread executes every iteration of the loop, except it only does the work if it 'owns' that iteration
           !That is, a thread only executes the work associated with an iteration if its thread id is congruent to
           !the iteration number modulo the number of threads
           do j=1,lc-1
             lcx = local_index(istep*nbw+j, my_pcol, np_cols, nblk, 0)
             if (lcx>0 ) then
               mynlc = mynlc+1
               if ( mod((j-1), omp_get_num_threads()) .eq. omp_get_thread_num() ) then
                   if (lr>0) aux1(mynlc) = dot_product(vr(1:lr),a(1:lr,lcx))
               endif
             endif
           enddo
552

553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
           ! Get global dot products
           !$omp barrier
           !$omp single
           if (mynlc>0) call mpi_allreduce(aux1,aux2,mynlc,MPI_REAL8,MPI_SUM,mpi_comm_rows,mpierr)
           !$omp end single
           !$omp barrier

           ! Transform
           transformChunkSize=32
           mynlc = 0
           do j=1,lc-1
             lcx = local_index(istep*nbw+j, my_pcol, np_cols, nblk, 0)
             if (lcx>0) then
               mynlc = mynlc+1
               !This loop could be parallelized with an openmp pragma with static scheduling and chunk size 32
               !However, for some reason this is slower than doing it manually, so it is parallelized as below.
               do ii=omp_get_thread_num()*transformChunkSize,lr,omp_get_num_threads()*transformChunkSize
                  do pp = 1,transformChunkSize
                      if (pp + ii > lr) exit
                          a(ii+pp,lcx) = a(ii+pp,lcx) - tau*aux2(mynlc)*vr(ii+pp)
                  enddo
               enddo
             endif
           enddo
           !$omp end parallel
#else /* WITH_OPENMP */
           nlc = 0 ! number of local columns
           do j=1,lc-1
             lcx = local_index(istep*nbw+j, my_pcol, np_cols, nblk, 0)
             if (lcx>0) then
               nlc = nlc+1
               if (lr>0) aux1(nlc) = dot_product(vr(1:lr),a(1:lr,lcx))
             endif
           enddo
587

588
589
           ! Get global dot products
           if (nlc>0) call mpi_allreduce(aux1,aux2,nlc,MPI_REAL8,MPI_SUM,mpi_comm_rows,mpierr)
Andreas Marek's avatar
Andreas Marek committed
590

591
           ! Transform
Andreas Marek's avatar
Andreas Marek committed
592

593
594
595
596
597
598
599
600
601
602
           nlc = 0
           do j=1,lc-1
             lcx = local_index(istep*nbw+j, my_pcol, np_cols, nblk, 0)
             if (lcx>0) then
               nlc = nlc+1
               a(1:lr,lcx) = a(1:lr,lcx) - tau*aux2(nlc)*vr(1:lr)
             endif
           enddo
#endif /* WITH_OPENMP */
         enddo ! lc
Andreas Marek's avatar
Andreas Marek committed
603
604

         if (useGPU) then
605
606
607
608
609
610
611
612
613
614
615
           ! store column tiles back to GPU
           cur_pcol = pcol(istep*nbw+1, nblk, np_cols)
           if (my_pcol == cur_pcol) then
             successCUDA = cuda_memcpy2d((a_dev+((lc_start-1)*lda*size_of_real_datatype)),          &
                                          lda*size_of_real_datatype, loc(a(1, lc_start)),           &
                                          lda*size_of_real_datatype,  lr_end*size_of_real_datatype, &
                                          (lc_end - lc_start+1),cudaMemcpyHostToDevice)
             if (.not.(successCUDA)) then
               print *,"bandred_real: error in cudaMemcpy2d"
               stop
             endif
616

617
           endif
Andreas Marek's avatar
Andreas Marek committed
618
619
         endif

620
621
         ! Calculate scalar products of stored Householder vectors.
         ! This can be done in different ways, we use dsyrk
Andreas Marek's avatar
Andreas Marek committed
622

623
         vav = 0
Andreas Marek's avatar
Andreas Marek committed
624

625
626
627
628
629
630
         if (useGPU) then
           if (l_rows>0) &
             call dsyrk('U','T',n_cols,l_rows,1.d0,vmrCUDA,cur_l_rows,0.d0,vav,ubound(vav,dim=1))
         else
           if (l_rows>0) &
             call dsyrk('U','T',n_cols,l_rows,1.d0,vmrCPU,ubound(vmrCPU,dim=1),0.d0,vav,ubound(vav,dim=1))
Andreas Marek's avatar
Andreas Marek committed
631

632
         endif
Andreas Marek's avatar
Andreas Marek committed
633

634
         call symm_matrix_allreduce(n_cols,vav, nbw, nbw,mpi_comm_rows)
Andreas Marek's avatar
Andreas Marek committed
635

636
         ! Calculate triangular matrix T for block Householder Transformation
Andreas Marek's avatar
Andreas Marek committed
637

638
639
640
641
642
         do lc=n_cols,1,-1
           tau = tmat(lc,lc,istep)
           if (lc<n_cols) then
             call dtrmv('U','T','N',n_cols-lc,tmat(lc+1,lc+1,istep),ubound(tmat,dim=1),vav(lc+1,lc),1)
             tmat(lc,lc+1:n_cols,istep) = -tau * vav(lc+1:n_cols,lc)
Andreas Marek's avatar
Andreas Marek committed
643
644
           endif
         enddo
645
       endif
Andreas Marek's avatar
Andreas Marek committed
646

647
       ! Transpose vmr -> vmc (stored in umc, second half)
Andreas Marek's avatar
Andreas Marek committed
648
649

       if (useGPU) then
650
651
652
653
654
655
656
         call elpa_transpose_vectors_real  (vmrCUDA, cur_l_rows, mpi_comm_rows, &
                                            umcCUDA(cur_l_cols * n_cols + 1), cur_l_cols, mpi_comm_cols, &
                                            1, istep*nbw, n_cols, nblk)
       else
         call elpa_transpose_vectors_real  (vmrCPU, ubound(vmrCPU,dim=1), mpi_comm_rows, &
                                            umcCPU(1,n_cols+1), ubound(umcCPU,dim=1), mpi_comm_cols, &
                                            1, istep*nbw, n_cols, nblk)
Andreas Marek's avatar
Andreas Marek committed
657
658
       endif

659
660
661
662
663
664
665
       ! Calculate umc = A**T * vmr
       ! Note that the distributed A has to be transposed
       ! Opposed to direct tridiagonalization there is no need to use the cache locality
       ! of the tiles, so we can use strips of the matrix

       ! here the GPU version and CPU version diverged substantially, due to the newest
       ! optimizations due to Intel. The GPU version has to be re-written
Andreas Marek's avatar
Andreas Marek committed
666
       if (useGPU) then
667
668
         umcCUDA(1 : l_cols * n_cols) = 0.d0
         vmrCUDA(cur_l_rows * n_cols + 1 : cur_l_rows * n_cols * 2) = 0
Andreas Marek's avatar
Andreas Marek committed
669

670
671
672
673
674
675
         if (l_cols>0 .and. l_rows>0) then
           successCUDA = cuda_memcpy(vmr_dev, loc(vmrCUDA(1)), vmr_size*size_of_real_datatype,cudaMemcpyHostToDevice)
           if (.not.(successCUDA)) then
             print *,"bandred_real: error in cudaMemcpy"
             stop
           endif
Andreas Marek's avatar
Andreas Marek committed
676

677
678
679
680
681
           successCUDA = cuda_memcpy(umc_dev, loc(umcCUDA(1)), umc_size*size_of_real_datatype,cudaMemcpyHostToDevice)
           if (.not.(successCUDA)) then
             print *,"bandred_real: error in cudaMemcpy"
             stop
           endif
Andreas Marek's avatar
Andreas Marek committed
682

683
           do i=0,(istep*nbw-1)/tile_size
Andreas Marek's avatar
Andreas Marek committed
684

685
686
687
             lcs = i*l_cols_tile+1
             lce = min(l_cols,(i+1)*l_cols_tile)
             if (lce<lcs) cycle
Andreas Marek's avatar
Andreas Marek committed
688

689
             lre = min(l_rows,(i+1)*l_rows_tile)
Andreas Marek's avatar
Andreas Marek committed
690

691
692
693
             call cublas_dgemm('T','N',lce-lcs+1,n_cols,lre, &
                               1.d0, (a_dev + ((lcs-1)*lda*size_of_real_datatype)), lda, vmr_dev,cur_l_rows, &
                               1.d0, (umc_dev+ (lcs-1)*size_of_real_datatype), cur_l_cols)
Andreas Marek's avatar
Andreas Marek committed
694

695
696
             if(i==0) cycle
             lre = min(l_rows,i*l_rows_tile)
Andreas Marek's avatar
Andreas Marek committed
697

698
699
700
701
702
             call cublas_dgemm('N','N',lre,n_cols,lce-lcs+1,&
                               1.d0, (a_dev+ ((lcs-1)*lda*size_of_real_datatype)),lda,                  &
                               (umc_dev+(cur_l_cols * n_cols+lcs-1)*size_of_real_datatype), cur_l_cols, &
                               1.d0, (vmr_dev+(cur_l_rows * n_cols)*size_of_real_datatype), cur_l_rows)
           enddo
Andreas Marek's avatar
Andreas Marek committed
703

704
705
706
707
708
           successCUDA = cuda_memcpy(loc(vmrCUDA(1)), vmr_dev,vmr_size*size_of_real_datatype,cudaMemcpyDeviceToHost)
           if (.not.(successCUDA)) then
             print *,"bandred_real: error in cudaMemcpy"
             stop
           endif
Andreas Marek's avatar
Andreas Marek committed
709

710
711
712
713
714
           successCUDA = cuda_memcpy(loc(umcCUDA(1)), umc_dev, umc_size*size_of_real_datatype,cudaMemcpyDeviceToHost)
           if (.not.(successCUDA)) then
             print *,"bandred_real: error in cudaMemcpy"
             stop
           endif
Andreas Marek's avatar
Andreas Marek committed
715

716
         endif ! l_cols>0 .and. l_rows>0
Andreas Marek's avatar
Andreas Marek committed
717

718
719
       else ! do not useGPU version
         !Code for Algorithm 4
Andreas Marek's avatar
Andreas Marek committed
720

721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
         n_way = 1
#ifdef WITH_OPENMP
         n_way = omp_get_max_threads()
#endif
         !umc(1:l_cols,1:n_cols) = 0.d0
         !vmr(1:l_rows,n_cols+1:2*n_cols) = 0
#ifdef WITH_OPENMP
         !$omp parallel private( i,lcs,lce,lrs,lre)
#endif
         if (n_way > 1) then
           !$omp do
           do i=1,min(l_cols_tile, l_cols)
             umcCPU(i,1:n_cols) = 0.d0
           enddo
           !$omp do
           do i=1,l_rows
             vmrCPU(i,n_cols+1:2*n_cols) = 0.d0
           enddo
           if (l_cols>0 .and. l_rows>0) then

             !SYMM variant 4
             !Partitioned Matrix Expression:
             ! Ct = Atl Bt + Atr Bb
             ! Cb = Atr' Bt + Abl Bb
             !
             !Loop invariant:
             ! Ct = Atl Bt + Atr Bb
             !
             !Update:
             ! C1 = A10'B0 + A11B1 + A21 B2
             !
             !This algorithm chosen because in this algoirhtm, the loop around the dgemm calls
             !is easily parallelized, and regardless of choise of algorithm,
             !the startup cost for parallelizing the dgemms inside the loop is too great

             !$omp do schedule(static,1)
             do i=0,(istep*nbw-1)/tile_size
               lcs = i*l_cols_tile+1                   ! local column start
               lce = min(l_cols, (i+1)*l_cols_tile)    ! local column end

               lrs = i*l_rows_tile+1                   ! local row start
               lre = min(l_rows, (i+1)*l_rows_tile)    ! local row end

               !C1 += [A11 A12] [B1
               !                 B2]
               if ( lre > lrs .and. l_cols > lcs ) then
                 call DGEMM('N','N', lre-lrs+1, n_cols, l_cols-lcs+1,          &
                            1.d0, a(lrs,lcs), ubound(a,dim=1),                 &
                                  umcCPU(lcs,n_cols+1), ubound(umcCPU,dim=1),  &
                            0.d0, vmrCPU(lrs,n_cols+1), ubound(vmrCPU,dim=1))
               endif
Andreas Marek's avatar
Andreas Marek committed
772

773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
               ! C1 += A10' B0
               if ( lce > lcs .and. i > 0 ) then
                 call DGEMM('T','N', lce-lcs+1, n_cols, lrs-1,           &
                            1.d0, a(1,lcs),   ubound(a,dim=1),           &
                                  vmrCPU(1,1),   ubound(vmrCPU,dim=1),   &
                            0.d0, umcCPU(lcs,1), ubound(umcCPU,dim=1))
               endif
             enddo
           endif ! l_cols>0 .and. l_rows>0
         else ! n_way > 1
           umcCPU(1:l_cols,1:n_cols) = 0.d0
           vmrCPU(1:l_rows,n_cols+1:2*n_cols) = 0
           if (l_cols>0 .and. l_rows>0) then
             do i=0,(istep*nbw-1)/tile_size

               lcs = i*l_cols_tile+1
               lce = min(l_cols,(i+1)*l_cols_tile)
               if (lce<lcs) cycle

               lre = min(l_rows,(i+1)*l_rows_tile)
               call DGEMM('T','N',lce-lcs+1,n_cols,lre,1.d0,a(1,lcs),ubound(a,dim=1), &
                            vmrCPU,ubound(vmrCPU,dim=1),1.d0,umcCPU(lcs,1),ubound(umcCPU,dim=1))

               if (i==0) cycle
                 lre = min(l_rows,i*l_rows_tile)
                 call DGEMM('N','N',lre,n_cols,lce-lcs+1,1.d0,a(1,lcs),lda, &
                            umcCPU(lcs,n_cols+1),ubound(umcCPU,dim=1),1.d0,vmrCPU(1,n_cols+1),ubound(vmrCPU,dim=1))
             enddo
           endif
         endif ! n_way > 1
#ifdef WITH_OPENMP
        !$omp end parallel
805
#endif
806
       endif ! do not useGPU version
Andreas Marek's avatar
Andreas Marek committed
807

808
809
810
811
       ! Sum up all ur(:) parts along rows and add them to the uc(:) parts
       ! on the processors containing the diagonal
       ! This is only necessary if ur has been calculated, i.e. if the
       ! global tile size is smaller than the global remaining matrix
Andreas Marek's avatar
Andreas Marek committed
812

813
814
       if (useGPU) then
         ! here the GPU version and CPU version divereged due to the same reasons as above
Andreas Marek's avatar
Andreas Marek committed
815

816
817
818
819
820
         if (tile_size < istep*nbw) then
           call elpa_reduce_add_vectors_real  (vmrCUDA(cur_l_rows * n_cols + 1),cur_l_rows,mpi_comm_rows, &
                                               umcCUDA, cur_l_cols, mpi_comm_cols, &
                                               istep*nbw, n_cols, nblk)
         endif
Andreas Marek's avatar
Andreas Marek committed
821

822
823
824
825
826
827
         if (l_cols>0) then
           allocate(tmpCUDA(l_cols * n_cols), stat=istat, errmsg=errorMessage)
           if (istat .ne. 0) then
             print *,"bandred_real: error when allocating tmpCUDA "//errorMessage
             stop
           endif
Andreas Marek's avatar
Andreas Marek committed
828

829
830
           call mpi_allreduce(umcCUDA,tmpCUDA,l_cols*n_cols,MPI_REAL8,MPI_SUM,mpi_comm_rows,ierr)
           umcCUDA(1 : l_cols * n_cols) = tmpCUDA(1 : l_cols * n_cols)
Andreas Marek's avatar
Andreas Marek committed
831

832
833
834
835
836
837
838
839
           if (allocated(tmpCUDA)) then
             deallocate(tmpCUDA, stat=istat, errmsg=errorMessage)
             if (istat .ne. 0) then
               print *,"bandred_real: error when deallocating tmpCUDA "//errorMessage
               stop
             endif
           endif
         endif ! l_cols
Andreas Marek's avatar
Andreas Marek committed
840

841
842
843
844
845
846
         ! U = U * Tmat**T
         successCUDA = cuda_memcpy(umc_dev, loc(umcCUDA(1)), umc_size*size_of_real_datatype, cudaMemcpyHostToDevice)
         if (.not.(successCUDA)) then
           print *,"bandred_real: error in cudaMemcpy"
           stop
         endif
Andreas Marek's avatar
Andreas Marek committed
847

848
849
850
851
852
         successCUDA = cuda_memcpy(tmat_dev,loc(tmat(1,1,istep)),nbw*nbw*size_of_real_datatype,cudaMemcpyHostToDevice)
         if (.not.(successCUDA)) then
           print *,"bandred_real: error in cudaMemcpy"
           stop
         endif
Andreas Marek's avatar
Andreas Marek committed
853

854
855
         call cublas_dtrmm('Right','Upper','Trans','Nonunit',l_cols,n_cols, &
                           1.d0, tmat_dev,nbw,umc_dev,cur_l_cols)
Andreas Marek's avatar
Andreas Marek committed
856

857
         ! VAV = Tmat * V**T * A * V * Tmat**T = (U*Tmat**T)**T * V * Tmat**T
Andreas Marek's avatar
Andreas Marek committed
858

859
860
861
862
863
         successCUDA = cuda_memcpy(vav_dev,loc(vav(1,1)), nbw*nbw*size_of_real_datatype,cudaMemcpyHostToDevice)
         if (.not.(successCUDA)) then
           print *,"bandred_real: error in cudaMemcpy"
           stop
         endif
Andreas Marek's avatar
Andreas Marek committed
864

865
866
867
         call cublas_dgemm('T','N',n_cols,n_cols,l_cols, &
                           1.d0, umc_dev,cur_l_cols,(umc_dev+(cur_l_cols * n_cols )*size_of_real_datatype),cur_l_cols, &
                           0.d0, vav_dev,nbw)
Andreas Marek's avatar
Andreas Marek committed
868

869
870
         call cublas_dtrmm('Right','Upper','Trans','Nonunit',n_cols,n_cols, &
                           1.d0, tmat_dev,nbw, vav_dev, nbw)
Andreas Marek's avatar
Andreas Marek committed
871
872


873
874
875
876
877
         successCUDA = cuda_memcpy(loc(vav(1,1)), vav_dev, nbw*nbw*size_of_real_datatype, cudaMemcpyDeviceToHost)
         if (.not.(successCUDA)) then
           print *,"bandred_real: error in cudaMemcpy"
           stop
         endif
Andreas Marek's avatar
Andreas Marek committed
878

879
         call symm_matrix_allreduce(n_cols,vav, nbw,nbw,mpi_comm_cols)
Andreas Marek's avatar
Andreas Marek committed
880

881
882
883
884
885
         successCUDA = cuda_memcpy(vav_dev, loc(vav(1,1)), nbw*nbw*size_of_real_datatype,cudaMemcpyHostToDevice)
         if (.not.(successCUDA)) then
           print *,"bandred_real: error in cudaMemcpy"
           stop
         endif
Andreas Marek's avatar
Andreas Marek committed
886

887
888
889
890
         ! U = U - 0.5 * V * VAV
         call cublas_dgemm('N','N',l_cols,n_cols,n_cols,&
                           -0.5d0, (umc_dev+(cur_l_cols * n_cols )*size_of_real_datatype),cur_l_cols, vav_dev,nbw,&
                           1.0d0, umc_dev,cur_l_cols)
Andreas Marek's avatar
Andreas Marek committed
891

892
893
894
895
896
         successCUDA = cuda_memcpy(loc(umcCUDA(1)), umc_dev, umc_size*size_of_real_datatype, cudaMemcpyDeviceToHost)
         if (.not.(successCUDA)) then
           print *,"bandred_real: error in cudaMemcpy"
           stop
         endif
Andreas Marek's avatar
Andreas Marek committed
897

898
         ! Transpose umc -> umr (stored in vmr, second half)
Andreas Marek's avatar
Andreas Marek committed
899

900
901
902
903
904
905
906
907
         call elpa_transpose_vectors_real  (umcCUDA, cur_l_cols, mpi_comm_cols, &
                                            vmrCUDA(cur_l_rows * n_cols + 1), cur_l_rows, mpi_comm_rows, &
                                            1, istep*nbw, n_cols, nblk)
         successCUDA = cuda_memcpy(vmr_dev, loc(vmrCUDA(1)), vmr_size*size_of_real_datatype, cudaMemcpyHostToDevice)
         if (.not.(successCUDA)) then
           print *,"bandred_real: error in cudaMemcpy"
           stop
         endif
Andreas Marek's avatar
Andreas Marek committed
908

909
910
911
912
913
         successCUDA = cuda_memcpy(umc_dev, loc(umcCUDA(1)), umc_size*size_of_real_datatype, cudaMemcpyHostToDevice)
         if (.not.(successCUDA)) then
           print *,"bandred_real: error in cudaMemcpy"
           stop
         endif
Andreas Marek's avatar
Andreas Marek committed
914

915
916
917
918
919
920
         ! A = A - V*U**T - U*V**T
         do i=0,(istep*nbw-1)/tile_size
           lcs = i*l_cols_tile+1
           lce = min(l_cols,(i+1)*l_cols_tile)
           lre = min(l_rows,(i+1)*l_rows_tile)
           if (lce<lcs .or. lre<1) cycle
Andreas Marek's avatar
Andreas Marek committed
921

922
923
924
925
926
927
928
929
930
931
932
           call cublas_dgemm('N', 'T', lre, lce-lcs+1, 2*n_cols, -1.d0, &
                             vmr_dev,cur_l_rows,(umc_dev +(lcs-1)*size_of_real_datatype),cur_l_cols, &
                             1.d0,(a_dev+(lcs-1)*lda*size_of_real_datatype),lda)
         enddo
       else ! do not useGPU
         ! Or if we used the Algorithm 4
         if (tile_size < istep*nbw .or. n_way > 1) then
         call elpa_reduce_add_vectors_real  (vmrCPU(1,n_cols+1),ubound(vmrCPU,dim=1),mpi_comm_rows, &
                                             umcCPU, ubound(umcCPU,dim=1), mpi_comm_cols, &
                                             istep*nbw, n_cols, nblk)
         endif
Andreas Marek's avatar
Andreas Marek committed
933

934
935
936
937
938
939
         if (l_cols>0) then
           allocate(tmpCPU(l_cols,n_cols), stat=istat, errmsg=errorMessage)
           if (istat .ne. 0) then
             print *,"bandred_real: error when allocating tmpCPU "//errorMessage
             stop
           endif
Andreas Marek's avatar
Andreas Marek committed
940

941
942
           call mpi_allreduce(umcCPU,tmpCPU,l_cols*n_cols,MPI_REAL8,MPI_SUM,mpi_comm_rows,mpierr)
           umcCPU(1:l_cols,1:n_cols) = tmpCPU(1:l_cols,1:n_cols)
Andreas Marek's avatar
Andreas Marek committed
943

944
945
946
947
948
949
           deallocate(tmpCPU, stat=istat, errmsg=errorMessage)
           if (istat .ne. 0) then
             print *,"bandred_real: error when deallocating tmpCPU "//errorMessage
             stop
           endif
         endif
Andreas Marek's avatar
Andreas Marek committed
950

951
         ! U = U * Tmat**T
Andreas Marek's avatar
Andreas Marek committed
952

953
954
         call dtrmm('Right','Upper','Trans','Nonunit',l_cols,n_cols,1.d0,tmat(1,1,istep),ubound(tmat,dim=1), &
                    umcCPU,ubound(umcCPU,dim=1))
Andreas Marek's avatar
Andreas Marek committed
955

956
         ! VAV = Tmat * V**T * A * V * Tmat**T = (U*Tmat**T)**T * V * Tmat**T
Andreas Marek's avatar
Andreas Marek committed
957

958
959
         call dgemm('T','N',n_cols,n_cols,l_cols,1.d0,umcCPU,ubound(umcCPU,dim=1),umcCPU(1,n_cols+1), &
                    ubound(umcCPU,dim=1),0.d0,vav,ubound(vav,dim=1))
Andreas Marek's avatar
Andreas Marek committed
960

961
962
963
964
         call dtrmm('Right','Upper','Trans','Nonunit',n_cols,n_cols,1.d0,tmat(1,1,istep),    &
                    ubound(tmat,dim=1),vav,ubound(vav,dim=1))

         call symm_matrix_allreduce(n_cols,vav, nbw, nbw ,mpi_comm_cols)
Andreas Marek's avatar
Andreas Marek committed
965

966
967
968
         ! U = U - 0.5 * V * VAV
         call dgemm('N','N',l_cols,n_cols,n_cols,-0.5d0,umcCPU(1,n_cols+1),ubound(umcCPU,dim=1),vav, &
                     ubound(vav,dim=1),1.d0,umcCPU,ubound(umcCPU,dim=1))
Andreas Marek's avatar
Andreas Marek committed
969

970
         ! Transpose umc -> umr (stored in vmr, second half)
Andreas Marek's avatar
Andreas Marek committed
971

972
         call elpa_transpose_vectors_real(umcCPU, ubound(umcCPU,dim=1), mpi_comm_cols, &
Andreas Marek's avatar
Andreas Marek committed
973
974
975
                                         vmrCPU(1,n_cols+1), ubound(vmrCPU,dim=1), mpi_comm_rows, &
                                         1, istep*nbw, n_cols, nblk)

976
977
978
979
980
981
982
983
984
         ! A = A - V*U**T - U*V**T
#ifdef WITH_OPENMP
         !$omp parallel private( ii, i, lcs, lce, lre, n_way, m_way, m_id, n_id, work_per_thread, mystart, myend  )
         n_threads = omp_get_num_threads()
         if (mod(n_threads, 2) == 0) then
           n_way = 2
         else
           n_way = 1
         endif
Andreas Marek's avatar
Andreas Marek committed
985

986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
         m_way = n_threads / n_way

         m_id = mod(omp_get_thread_num(),  m_way)
         n_id = omp_get_thread_num() / m_way

         do ii=n_id*tile_size,(istep*nbw-1),tile_size*n_way
           i = ii / tile_size
           lcs = i*l_cols_tile+1
           lce = min(l_cols,(i+1)*l_cols_tile)
           lre = min(l_rows,(i+1)*l_rows_tile)
           if (lce<lcs .or. lre<1) cycle

           !Figure out this thread's range
           work_per_thread = lre / m_way
           if (work_per_thread * m_way < lre) work_per_thread = work_per_thread + 1
           mystart = m_id * work_per_thread + 1
           myend   = mystart + work_per_thread - 1
           if ( myend > lre ) myend = lre
           if ( myend-mystart+1 < 1) cycle

           call dgemm('N','T',myend-mystart+1, lce-lcs+1, 2*n_cols, -1.d0, &
                      vmrCPU(mystart, 1), ubound(vmrCPU,1), umcCPU(lcs,1), ubound(umcCPU,1), &
                       1.d0,a(mystart,lcs),ubound(a,1))
         enddo
         !$omp end parallel
Andreas Marek's avatar
Andreas Marek committed
1011

1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
#else /* WITH_OPENMP */
         do i=0,(istep*nbw-1)/tile_size
           lcs = i*l_cols_tile+1
           lce = min(l_cols,(i+1)*l_cols_tile)
           lre = min(l_rows,(i+1)*l_rows_tile)
           if (lce<lcs .or. lre<1) cycle
           call dgemm('N','T',lre,lce-lcs+1,2*n_cols,-1.d0, &
                       vmrCPU,ubound(vmrCPU,dim=1),umcCPU(lcs,1),ubound(umcCPU,dim=1), &
                       1.d0,a(1,lcs),lda)
         enddo
#endif /* WITH_OPENMP */
Andreas Marek's avatar
Andreas Marek committed
1023

1024
       endif ! useGPU
Andreas Marek's avatar
Andreas Marek committed
1025

1026
1027
1028
1029
1030
1031
1032
1033
       if (.not.(useGPU)) then
         if (allocated(vr)) then
           deallocate(vr, stat=istat, errmsg=errorMessage)
           if (istat .ne. 0) then
             print *,"bandred_real: error when deallocating vr "//errorMessage
             stop
           endif
         endif
Andreas Marek's avatar
Andreas Marek committed
1034

1035
1036
1037
1038
1039
1040
1041
         if (allocated(umcCPU)) then
           deallocate(umcCPU, stat=istat, errmsg=errorMessage)
           if (istat .ne. 0) then
             print *,"bandred_real: error when deallocating vmrCPU "//errorMessage
             stop
           endif
         endif
Andreas Marek's avatar
Andreas Marek committed
1042

1043
1044
1045
1046
1047
1048
1049
         if (allocated(vmrCPU)) then
           deallocate(vmrCPU, stat=istat, errmsg=errorMessage)
           if (istat .ne. 0) then
             print *,"bandred_real: error when deallocating vmrCPU "//errorMessage
             stop
           endif
         endif
Andreas Marek's avatar
Andreas Marek committed
1050

1051
       endif !useGPU
Andreas Marek's avatar
Andreas Marek committed
1052

1053
     enddo ! istep
Andreas Marek's avatar
Andreas Marek committed
1054

1055
1056
1057
1058
1059
1060
     if (useGPU) then
       successCUDA = cuda_memcpy ( loc (a), a_dev, lda*na_cols*size_of_real_datatype,cudaMemcpyDeviceToHost)
       if (.not.(successCUDA)) then
         print *,"bandred_real: error in cudaMemcpy"
         stop
       endif
Andreas Marek's avatar
Andreas Marek committed
1061

1062
1063
1064
1065
1066
       successCUDA = cuda_free(a_dev)
       if (.not.(successCUDA)) then
         print *,"bandred_real: error in cudaFree"
         stop
       endif
Andreas Marek's avatar
Andreas Marek committed
1067

1068
1069
1070
1071
1072
       successCUDA = cuda_free(tmat_dev)
       if (.not.(successCUDA)) then
         print *,"bandred_real: error in cudaFree"
         stop
       endif
Andreas Marek's avatar
Andreas Marek committed
1073

1074
1075
1076
1077
1078
1079
       successCUDA = cuda_free(vav_dev)
       if (.not.(successCUDA)) then
         print *,"bandred_real: error in cudaFree"
         stop
       endif
     endif ! useGPU
Andreas Marek's avatar
Andreas Marek committed
1080

1081
1082
1083
1084
1085
1086
1087
     if (allocated(vr)) then
       deallocate(vr, stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
         print *,"bandred_real: error when deallocating vr "//errorMessage
         stop
       endif
     endif
Andreas Marek's avatar
Andreas Marek committed
1088

1089
1090
1091
1092
1093
1094
1095
     if (allocated(umcCPU)) then
       deallocate(umcCPU, stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
         print *,"bandred_real: error when deallocating umcCPU "//errorMessage
         stop
       endif
     endif
Andreas Marek's avatar
Andreas Marek committed
1096

1097
1098
1099
1100
1101
1102
1103
     if (allocated(vmrCPU)) then
       deallocate(vmrCPU, stat=istat, errmsg=errorMessage)
       if (istat .ne. 0) then
         print *,"bandred_real: error when deallocating vmrCPU "//errorMessage
         stop
       endif
     endif
Andreas Marek's avatar
Andreas Marek committed
1104

1105
1106
1107
1108
1109
1110
     if (useGPU) then
       successCUDA = cuda_free(vmr_dev)
       if (.not.(successCUDA)) then
         print *,"bandred_real: error in cudaFree"
         stop
       endif
Andreas Marek's avatar
Andreas Marek committed
1111

1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
       successCUDA = cuda_free(umc_dev)
       if (.not.(successCUDA)) then
         print *,"bandred_real: error in cudaFree"
         stop
       endif
       if (allocated(umcCUDA)) then
         deallocate(umcCUDA, stat=istat, errmsg=errorMessage)
         if (istat .ne. 0) then