elpa2_bandred_template.X90 73.8 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
59
60
61
62
#if 0
!    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 Naturwissenschaften,
!      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".
#endif
63
64
65
66
67
68
69
70
71
72
73
74
75
    subroutine bandred_&
    &MATH_DATATYPE&
    &_&
    &PRECISION &
    (na, a, &
#if REALCASE == 1
     a_dev, &
#endif
     lda, nblk, nbw, matrixCols, numBlocks, mpi_comm_rows, mpi_comm_cols, tmat, &
#if REALCASE == 1
     tmat_dev, &
#endif
     wantDebug, useGPU, success &
76
#if REALCASE == 1
77
     , useQR)
78
79
#endif
#if COMPLEXCASE == 1
80
     )
81
#endif
82

83
  !-------------------------------------------------------------------------------
84
  !  bandred_real/complex: Reduces a distributed symmetric matrix to band form
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
  !
  !  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
      use elpa1_compute
#ifdef HAVE_DETAILED_TIMINGS
      use timings
Andreas Marek's avatar
Andreas Marek committed
117
118
#else
      use timings_dummy
119
120
121
122
123
124
125
#endif
#ifdef WITH_OPENMP
      use omp_lib
#endif
      use precision
      implicit none

126
127
128
      integer(kind=ik)                            :: na, lda, nblk, nbw, matrixCols, numBlocks, mpi_comm_rows, mpi_comm_cols

#if REALCASE == 1
129
#ifdef USE_ASSUMED_SIZE
130
      real(kind=REAL_DATATYPE)                    :: a(lda,*), tmat(nbw,nbw,*)
131
#else
132
      real(kind=REAL_DATATYPE)                    :: a(lda,matrixCols), tmat(nbw,nbw,numBlocks)
133
#endif
134
135
136
137
138
139
140
#endif
#if COMPLEXCASE == 1
#ifdef USE_ASSUMED_SIZE
      complex(kind=COMPLEX_DATATYPE)              :: a(lda,*), tmat(nbw,nbw,*)
#else
      complex(kind=COMPLEX_DATATYPE)              :: a(lda,matrixCols), tmat(nbw,nbw,numBlocks)
#endif
Andreas Marek's avatar
Andreas Marek committed
141
142
143
144
145
146
147
148
149
150
151
152
#endif /* COMPLEXCASE */

#if REALCASE == 1
#ifdef DOUBLE_PRECISION_REAL
      real(kind=REAL_DATATYPE), parameter         :: ZERO = 0.0_rk8, ONE = 1.0_rk8

#else
      real(kind=REAL_DATATYPE), parameter         :: ZERO = 0.0_rk4, ONE = 1.0_rk4
#endif
#endif

#if COMPLEXCASE == 1
153
#ifdef DOUBLE_PRECISION_COMPLEX
Andreas Marek's avatar
Andreas Marek committed
154
      complex(kind=COMPLEX_DATATYPE), parameter   :: ZERO = (0.0_rk8, 0.0_rk8), ONE = (1.0_rk8, 0.0_rk8)
155
#else
Andreas Marek's avatar
Andreas Marek committed
156
      complex(kind=COMPLEX_DATATYPE), parameter   :: ZERO = (0.0_rk4, 0.0_rk4), ONE = (1.0_rk4, 0.0_rk4)
157
158
#endif
#endif /* COMPLEXCASE == 1 */
159

160
161
162
163
#if REALCASE == 1
      real(kind=REAL_DATATYPE)                    :: eps
#endif
      logical, intent(in)                         :: useGPU
164

165
166
167
168
169
170
171
172
      integer(kind=ik)                            :: my_prow, my_pcol, np_rows, np_cols, mpierr
      integer(kind=ik)                            :: l_cols, l_rows
#if REALCASE == 1
      integer(kind=ik)                            :: vmrCols, mynlc
#endif
      integer(kind=ik)                            :: i, j, lcs, lce, lrs, lre, lc, lr, cur_pcol, n_cols, nrow
      integer(kind=ik)                            :: istep, ncol, lch, lcx, nlc
      integer(kind=ik)                            :: tile_size, l_rows_tile, l_cols_tile
173

174
175
176
177
178
179
180
      real(kind=REAL_DATATYPE)                    :: vnorm2
#if REALCASE == 1
      real(kind=REAL_DATATYPE)                    :: xf, aux1(nbw), aux2(nbw), vrl, tau, vav(nbw,nbw)
#endif
#if COMPLEXCASE == 1
      complex(kind=COMPLEX_DATATYPE)              :: xf, aux1(nbw), aux2(nbw), vrl, tau, vav(nbw,nbw)

181
      complex(kind=COMPLEX_DATATYPE), allocatable :: tmp_CPU(:,:), vmrCPU(:,:), umcCPU(:,:)
Andreas Marek's avatar
Andreas Marek committed
182
      complex(kind=COMPLEX_DATATYPE), allocatable :: vr(:)
183
#endif
184

185
186
187
188
189
190
191
192
193
194
195
196
#if REALCASE == 1
      real(kind=REAL_DATATYPE), allocatable       :: tmpCUDA(:),  vmrCUDA(:),  umcCUDA(:)
      real(kind=REAL_DATATYPE), allocatable       :: tmpCPU(:,:), vmrCPU(:,:), umcCPU(:,:)
      real(kind=REAL_DATATYPE), allocatable       :: vr(:)
#endif

#if REALCASE == 1
      ! needed for blocked QR decomposition
      integer(kind=ik)                            :: PQRPARAM(11), work_size
      real(kind=REAL_DATATYPE)                    :: dwork_size(1)
      real(kind=REAL_DATATYPE), allocatable       :: work_blocked(:), tauvector(:), blockheuristic(:)
#endif
197
      ! a_dev is passed from bandred_real to trans_ev_band
198
      integer(kind=C_intptr_T)                    :: a_dev, vmr_dev, umc_dev, tmat_dev, vav_dev
199
#ifdef WITH_MPI
200
201
202
203
204
205
206
207
208
209
210
211
      integer(kind=ik), external                  :: numroc
#endif
      integer(kind=ik)                            :: ierr
      integer(kind=ik)                            :: cur_l_rows, cur_l_cols, vmr_size, umc_size
      integer(kind=c_size_t)                      :: lc_start, lc_end
#if COMPLEXCASE == 1
      integer(kind=c_size_t)                      :: lce_1, lcs_1, lre_1
#endif
      integer(kind=ik)                            :: lr_end
      integer(kind=ik)                            :: na_cols
#if COMPLEXCASE == 1
      integer(kind=ik)                            :: na_rows
212
213
#endif

214
215
216
217
218
      logical, intent(in)                         :: wantDebug
      logical, intent(out)                        :: success
      logical                                     :: successCUDA
      integer(kind=ik)                            :: istat
      character(200)                              :: errorMessage
219

220
221
222
223
224
225
226
#if REALCASE == 1
      logical, intent(in)                         :: useQR
#endif
#if REALCASE == 1
      integer(kind=ik)                            :: mystart, myend, m_way, n_way, work_per_thread, m_id, n_id, n_threads, &
                                                    ii, pp, transformChunkSize
#endif
Andreas Marek's avatar
Andreas Marek committed
227

Andreas Marek's avatar
Andreas Marek committed
228
229
230
231
232
      call timer%start("bandred_&
      &MATH_DATATYPE&
      &" // &
      &PRECISION_SUFFIX &
      )
233
234
235
236
237
238
      call timer%start("mpi_communication")

      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)
Andreas Marek's avatar
Andreas Marek committed
239

240
241
242
243
244
245
246
247
      call timer%stop("mpi_communication")
      success = .true.


      ! 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
Andreas Marek's avatar
Andreas Marek committed
248
249
250
251
252
253
            write(error_unit,*) 'ELPA2_bandred_&
	    &MATH_DATATYPE&
	    &: ERROR: nbw=',nbw,', nblk=',nblk
            write(error_unit,*) 'ELPA2_bandred_&
	    &MATH_DATATYPE&
	    &: ELPA2 works only for nbw==n*nblk'
254
255
256
257
258
259
260
261
262
263
          endif
          success = .false.
          return
        endif
      endif

! na_rows in used nowhere; only na_cols
      if (useGPU) then
#ifdef WITH_MPI
!        na_rows = numroc(na, nblk, my_prow, 0, np_rows)
264
265
266
#if COMPLEXCASE == 1
         na_rows = numroc(na, nblk, my_prow, 0, np_rows)
#endif
267
268
269
        na_cols = numroc(na, nblk, my_pcol, 0, np_cols)
#else
!        na_rows = na
270
271
272
#if COMPLEXCASE == 1
         na_rows = na
#endif
273
        na_cols = na
Andreas Marek's avatar
Cleanup    
Andreas Marek committed
274
#endif /* WITH_MPI */
275
276

        ! Here we convert the regular host array into a pinned host array
Andreas Marek's avatar
Andreas Marek committed
277
278
279
        successCUDA = cuda_malloc(a_dev, lda*na_cols*  &
#if REALCASE == 1
	                          size_of_PRECISION_real)
280
281
#endif
#if COMPLEXCASE == 1
Andreas Marek's avatar
Andreas Marek committed
282
283
                                  size_of_PRECISION_complex)
#endif
284
        if (.not.(successCUDA)) then
Andreas Marek's avatar
Andreas Marek committed
285
286
287
          print *,"bandred_&
	  &MATH_DATATYPE&
	  &: error in cudaMalloc"
288
289
290
          stop
        endif

Andreas Marek's avatar
Andreas Marek committed
291
        successCUDA = cuda_malloc(tmat_dev, nbw*nbw*   &
292
#if REALCASE == 1
Andreas Marek's avatar
Andreas Marek committed
293
	                          size_of_PRECISION_real)
294
295
#endif
#if COMPLEXCASE == 1
Andreas Marek's avatar
Andreas Marek committed
296
297
                                  size_of_PRECISION_complex)
#endif
298
        if (.not.(successCUDA)) then
Andreas Marek's avatar
Andreas Marek committed
299
300
301
          print *,"bandred_&
	  &MATH_DATATYPE&
	  &: error in cudaMalloc"
302
303
304
          stop
        endif

Andreas Marek's avatar
Andreas Marek committed
305
        successCUDA = cuda_malloc(vav_dev, nbw*nbw*   &
306
#if REALCASE == 1
Andreas Marek's avatar
Andreas Marek committed
307
	                          size_of_PRECISION_real)
308
309
#endif
#if COMPLEXCASE == 1
Andreas Marek's avatar
Andreas Marek committed
310
311
                                  size_of_PRECISION_complex)
#endif
312
        if (.not.(successCUDA)) then
Andreas Marek's avatar
Andreas Marek committed
313
314
315
          print *,"bandred_&
	  &MATH_DATATYPE&
	  &: error in cudaMalloc"
316
317
          stop
        endif
318
319
320
321
322
323
324
325
326
327
      endif ! useGPU

      ! 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

328
#if REALCASE == 1
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
      if (useQR) then

        if (useGPU) then
          print *,"qr decomposition at the moment not supported with GPU"
          stop
        endif

        if (which_qr_decomposition == 1) then
          call qr_pqrparam_init(pqrparam(1:11),    nblk,'M',0,   nblk,'M',0,   nblk,'M',1,'s')
          allocate(tauvector(na), stat=istat, errmsg=errorMessage)
          if (istat .ne. 0) then
            print *,"bandred_real: error when allocating tauvector "//errorMessage
            stop
          endif

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

          l_rows = local_index(na, my_prow, np_rows, nblk, -1)
          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

          vmrCols = na

#ifdef USE_ASSUMED_SIZE_QR
Andreas Marek's avatar
Andreas Marek committed
360
361
362
          call qr_pdgeqrf_2dcomm_&
	  &PRECISION&
	  &(a, lda, matrixCols, vmrCPU, max(l_rows,1), vmrCols, tauvector(1), na, tmat(1,1,1), &
363
364
365
366
                                 nbw, nbw, dwork_size, 1, -1, na, nbw, nblk, nblk, na, na, 1, 0, PQRPARAM(1:11), &
                                 mpi_comm_rows, mpi_comm_cols, blockheuristic)

#else
Andreas Marek's avatar
Andreas Marek committed
367
368
369
          call qr_pdgeqrf_2dcomm_&
	  &PRECISION&
	  &(a(1:lda,1:matrixCols), matrixCols, lda, vmrCPU(1:max(l_rows,1),1:vmrCols), max(l_rows,1), &
370
371
372
373
374
375
376
377
378
379
380
                                 vmrCols, tauvector(1:na), na, tmat(1:nbw,1:nbw,1), nbw, &
                                 nbw, dwork_size(1:1), 1, -1, na, nbw, nblk, nblk, na, na, 1, 0, PQRPARAM(1:11), &
                                 mpi_comm_rows, mpi_comm_cols, blockheuristic)
#endif

          work_size = dwork_size(1)
          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
381
          work_blocked = CONST_0_0
382
383
384
385
386
387
388
389
390
          deallocate(vmrCPU, stat=istat, errmsg=errorMessage)
          if (istat .ne. 0) then
            print *,"bandred_real: error when deallocating vmrCPU "//errorMessage
            stop
          endif

        endif ! which_qr_decomposition

      endif ! useQr
Andreas Marek's avatar
Andreas Marek committed
391
#endif /* REALCASE */
392
393

      if (useGPU) then
394
395
396
397
398
!#if !defined(USE_ASSUMED_SIZE)
!        if (size(a,dim=1) .ne. lda .or. size(a,dim=2) .ne. na_cols) then
!          print *,"bandred_complex: sizes of a wrong ? ",lda,size(a,dim=1),na_cols,size(a,dim=2)
!        endif
!#endif
399
400
401

        cur_l_rows = 0
        cur_l_cols = 0
Andreas Marek's avatar
Andreas Marek committed
402
403

        successCUDA = cuda_memcpy(a_dev, loc(a(1,1)), (lda)*(na_cols)*   &
404
#if REALCASE == 1
Andreas Marek's avatar
Andreas Marek committed
405
	                          size_of_PRECISION_real,    &
406
407
#endif
#if COMPLEXCASE == 1
Andreas Marek's avatar
Andreas Marek committed
408
409
410
                                  size_of_PRECISION_complex, &
#endif
				  cudaMemcpyHostToDevice)
411
        if (.not.(successCUDA)) then
Andreas Marek's avatar
Andreas Marek committed
412
413
414
          print *,"bandred_&
	  &MATH_DATATYPE&
	  &: error in cudaMemcpy"
415
416
          stop
        endif
417
418
419
420
421
422
423
424
425
426
427
      endif ! useGPU


      do istep = (na-1)/nbw, 1, -1

        n_cols = MIN(na,(istep+1)*nbw) - istep*nbw ! Number of columns in current step

        ! 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)

428
429
        ! Allocate vmr and umc to their exact sizes so that they can be used in bcasts and reduces

430
431
432
433
434
435
436
437
438
439
440
441
442
        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
Andreas Marek's avatar
Andreas Marek committed
443
444
445
                print *,"bandred_&
		&MATH_DATATYPE&
		&: error when deallocating vr "//errorMessage
446
447
448
449
450
                stop
              endif
            endif
            allocate(vr(l_rows + 1), stat=istat, errmsg=errorMessage)
            if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
451
452
453
              print *,"bandred_&
	      &MATH_DATATYPE&
	      &: error when allocating vr "//errorMessage
454
455
456
457
458
              stop
            endif

          endif

459
#if REALCASE == 1
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
          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

              successCUDA = cuda_free(vmr_dev)
              if (.not.(successCUDA)) then
                print *,"bandred_real: error in cuda_free"
                stop
              endif
            endif

            allocate(vmrCUDA(vmr_size), stat=istat, errmsg=errorMessage)
            if (istat .ne. 0) then
              print *,"bandred_real: error when allocating vmrCUDA "//errorMessage
              stop
            endif
480
            successCUDA = cuda_malloc(vmr_dev, vmr_size*size_of_PRECISION_real)
481
482
483
484
485
486
            if (.not.(successCUDA)) then
              print *,"bandred_real: error in cudaMalloc"
              stop
            endif

          endif
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
#endif

#if COMPLEXCASE == 1
          if ((.not. allocated(vmrCPU)) .or. (vmr_size .gt. ubound(vmrCPU, dim=1))) then
            if (allocated(vmrCPU)) then
              deallocate(vmrCPU, stat=istat, errmsg=errorMessage)
              if (istat .ne. 0) then
                print *,"bandred_complex: error when deallocating vmrCPU "//errorMessage
                stop
              endif

              successCUDA = cuda_free(vmr_dev)
              if (.not.(successCUDA))then
                print *,"bandred_complex: error in cudaFree"
                stop
              endif
            endif

            allocate(vmrCPU(max(l_rows,1),2*n_cols), stat=istat, errmsg=errorMessage)
            if (istat .ne. 0) then
              print *,"bandred_complex: error when allocating vmrCPU "//errorMessage
              stop
            endif

            if (max(l_rows,1) * 2*n_cols .gt. vmr_size) then
              print *,"bandred_complex: vmc_size ",max(l_rows,1) * 2*n_cols,vmr_size
            endif

            successCUDA = cuda_malloc(vmr_dev, vmr_size*size_of_PRECISION_complex)
            if (.not.(successCUDA)) then
              print *, "bandred_complex:  cuda malloc failed vmr_dev ", istat
              stop
            endif

          endif
#endif
523

524
#if REALCASE == 1
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
          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

              successCUDA = cuda_free(umc_dev)
              if (.not.(successCUDA)) then
                 print *,"bandred_real: error in cudaFree"
                 stop
              endif

            endif
            allocate(umcCUDA(umc_size), stat=istat, errmsg=errorMessage)
            if (istat .ne. 0) then
              print *,"bandred_real: error when deallocating umcCUDA "//errorMessage
              stop
            endif
545

546
            successCUDA = cuda_malloc(umc_dev, umc_size*size_of_PRECISION_real)
547
548
549
550
551
552
            if (.not.(successCUDA)) then
              print *,"bandred_real: error in cudaMalloc"
              stop
            endif

          endif
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
587
#endif /* REALCASE == 1 */

#if COMPLEXCASE == 1
          if ((.not. allocated(umcCPU)) .or. (umc_size .gt. ubound(umcCPU, dim=1))) then
            if (allocated(umcCPU)) then
              deallocate(umcCPU, stat=istat, errmsg=errorMessage)


              if (istat .ne. 0) then
                print *,"bandred_complex: error when allocating umcCPU "//errorMessage
                stop
              endif
              successCUDA = cuda_free(umc_dev)
              if (.not.(successCUDA))then
                print *,"bandred_complex: error in cudaFree"
                stop
              endif
            endif

            allocate(umcCPU(max(l_cols,1),2*n_cols), stat=istat, errmsg=errorMessage)
            if (istat .ne. 0) then
              print *,"bandred_complex: error when allocating umcCPU "//errorMessage
              stop
            endif

            if (max(l_cols,1) * 2*n_cols .gt. umc_size) then
              print *,"bandred_complex: umc_size ",max(l_cols,1) * 2*n_cols,umc_size
            endif
            successCUDA = cuda_malloc(umc_dev, umc_size*size_of_PRECISION_complex)
            if (.not.(successCUDA)) then
              print *, "bandred_complex:  cuda malloc failed umc_dev ", istat
              stop
            endif
          endif
#endif
588
589
590

        else ! GPU not used

591
592
593
          ! unify the the name vmr and vmrCPU, as well as vmrGPU
          ! the same for umcCPU and umcGPU
          ! Allocate vmr and umcCPU to their exact sizes so that they can be used in bcasts and reduces
594
595
596

          allocate(vmrCPU(max(l_rows,1),2*n_cols), stat=istat, errmsg=errorMessage)
          if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
597
598
599
            print *,"bandred_&
	    &MATH_DATATYPE&
	    &: error when allocating vmrCPU "//errorMessage
600
601
602
            stop
          endif

603
604
          allocate(umcCPU(max(l_cols,1),2*n_cols), stat=istat, errmsg=errorMessage)
          if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
605
606
607
            print *,"bandred_&
	    &MATH_DATATYPE&
	    &: error when allocating umcCPU "//errorMessage
608
609
            stop
          endif
Andreas Marek's avatar
Andreas Marek committed
610

611
612
          allocate(vr(l_rows+1), stat=istat, errmsg=errorMessage)
          if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
613
614
615
            print *,"bandred_&
	    &MATH_DATATYPE&
	    &: error when allocating vr "//errorMessage
616
617
618
619
620
621
            stop
          endif

        endif ! use GPU

        if (useGPU) then
622
#if REALCASE == 1
623
          vmrCUDA(1 : cur_l_rows * n_cols) = CONST_0_0
624
#endif
625
        else
626
#if REALCASE == 1
627
          vmrCPU(1:l_rows,1:n_cols) = CONST_0_0
628
629
630
631
632
633
634
#endif
#if COMPLEXCASE == 1
          vmrCPU(1:l_rows,1:n_cols) = CONST_COMPLEX_0_0
#endif
        endif ! useGPU

#if REALCASE == 1
635
636
        vr(:) = CONST_0_0
        tmat(:,:,istep) = CONST_0_0
637
638
639
640
641
#endif
#if COMPLEXCASE == 1
        vr(:) = CONST_COMPLEX_0_0
        tmat(:,:,istep) = CONST_COMPLEX_0_0
#endif
642
        if (useGPU) then
643
#if REALCASE == 1
644
          umcCUDA(1 : umc_size) = CONST_0_0
645
#endif
646
647
648
649
          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)

650
          if (lc_start .le. 0) lc_start = 1
651
652
653
654
655

          ! Here we assume that the processor grid and the block grid are aligned
          cur_pcol = pcol(istep*nbw+1, nblk, np_cols)

          if(my_pcol == cur_pcol) then
656
#if REALCASE == 1
657
658
659
            successCUDA = cuda_memcpy2d(loc(a(1, lc_start)), lda*size_of_PRECISION_real,         &
                                       (a_dev + ((lc_start-1) * lda*size_of_PRECISION_real)),    &
                                       lda*size_of_PRECISION_real, lr_end*size_of_PRECISION_real, &
660
                                       (lc_end - lc_start+1), cudaMemcpyDeviceToHost)
661
662
663
664
665
666
667
668
#endif
#if COMPLEXCASE == 1
            successCUDA = cuda_memcpy2d(loc(a(1, lc_start)), int(lda*size_of_PRECISION_complex,kind=c_size_t),            &
                                        (a_dev + int( ( (lc_start-1) * lda*size_of_PRECISION_complex),kind=c_size_t )),      &
                                        int(lda*size_of_PRECISION_complex,kind=c_size_t),              &
                                    int(lr_end*size_of_PRECISION_complex,kind=c_size_t),               &
                                      int((lc_end - lc_start+1),kind=c_size_t),int(cudaMemcpyDeviceToHost,kind=c_int))
#endif
669
            if (.not.(successCUDA)) then
Andreas Marek's avatar
Andreas Marek committed
670
671
672
              print *,"bandred_&
	      &MATH_DATATYPE&
	      &: error in cudaMemcpy2d"
673
674
675
676
677
678
679
              stop
            endif

          endif
        endif ! useGPU

        ! Reduce current block to lower triangular form
680
#if REALCASE == 1
681
682
683
684
        if (useQR) then
          if (which_qr_decomposition == 1) then
            vmrCols = 2*n_cols
#ifdef USE_ASSUMED_SIZE_QR
Andreas Marek's avatar
Andreas Marek committed
685
686
687
            call qr_pdgeqrf_2dcomm_&
	    &PRECISION&
	    &(a, lda, matrixCols, vmrCPU, max(l_rows,1), vmrCols, tauvector(1), &
688
689
690
691
692
693
694
                                   na, tmat(1,1,istep), nbw, nbw, work_blocked, work_size,        &
                                     work_size, na, n_cols, nblk, nblk,        &
                                     istep*nbw+n_cols-nbw, istep*nbw+n_cols, 1,&
                                     0, PQRPARAM(1:11), mpi_comm_rows, mpi_comm_cols,&
                                     blockheuristic)

#else
Andreas Marek's avatar
Andreas Marek committed
695
696
697
            call qr_pdgeqrf_2dcomm_&
	    &PRECISION&
	    &(a(1:lda,1:matrixCols), lda, matrixCols, vmrCPU(1:max(l_rows,1),1:vmrCols) ,   &
698
699
700
701
702
703
704
705
706
707
                                    max(l_rows,1), vmrCols, tauvector(1:na), na, &
                                     tmat(1:nbw,1:nbw,istep), nbw, nbw, work_blocked(1:work_size), work_size, &
                                     work_size, na, n_cols, nblk, nblk,        &
                                     istep*nbw+n_cols-nbw, istep*nbw+n_cols, 1,&
                                     0, PQRPARAM(1:11), mpi_comm_rows, mpi_comm_cols,&
                                     blockheuristic)
#endif
          endif

       else !useQR
708
#endif /* REALCASE == 1 */
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
         do lc = n_cols, 1, -1

           ncol = istep*nbw + lc ! absolute column number of householder vector
           nrow = ncol - nbw ! Absolute number of pivot row

           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

           tau = 0

           if (nrow == 1) exit ! Nothing to do

           cur_pcol = pcol(ncol, nblk, np_cols) ! Processor column owning current block

           if (my_pcol==cur_pcol) then

             ! Get vector to be transformed; distribute last element and norm of
             ! remaining elements to all procs in current column

             vr(1:lr) = a(1:lr,lch) ! vector to be transformed

             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))
735
#if REALCASE == 1
736
               aux1(2) = CONST_0_0
737
738
739
740
#endif
#if COMPLEXCASE == 1
               aux1(2) = CONST_COMPLEX_0_0
#endif
741
742
743
744
             endif

#ifdef WITH_MPI
             call timer%start("mpi_communication")
Andreas Marek's avatar
Andreas Marek committed
745
             call mpi_allreduce(aux1, aux2, 2, &
746
#if REALCASE == 1
Andreas Marek's avatar
Andreas Marek committed
747
                                MPI_REAL_PRECISION, &
748
749
#endif
#if COMPLEXCASE == 1
Andreas Marek's avatar
Andreas Marek committed
750
                                MPI_COMPLEX_PRECISION, &
751
#endif
Andreas Marek's avatar
Andreas Marek committed
752
                                MPI_SUM, mpi_comm_rows, mpierr)
753
754
755
756
             call timer%stop("mpi_communication")

#else /* WITH_MPI */
              aux2 = aux1 ! this should be optimized
Andreas Marek's avatar
Andreas Marek committed
757
#endif
758
759
760
761
762

             vnorm2 = aux2(1)
             vrl    = aux2(2)

             ! Householder transformation
763
#if REALCASE == 1
764
	     call hh_transform_real_&
765
766
#endif
#if COMPLEXCASE == 1
767
	     call hh_transform_complex_&
768
#endif
Andreas Marek's avatar
Andreas Marek committed
769
             &PRECISION &
770
                              (vrl, vnorm2, xf, tau)
771
772
773
774
775
776
             ! Scale vr and store Householder vector for back transformation

             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
777
#if REALCASE == 1
778
               vr(lr) = CONST_1_0
779
780
781
782
#endif
#if COMPLEXCASE == 1
               vr(lr) = CONST_COMPLEX_1_0
#endif
783
784
785
786
787
788
789
790
791
792
793
             else
               a(1:lr,lch) = vr(1:lr)
             endif

           endif

           ! Broadcast Householder vector and tau along columns

           vr(lr+1) = tau
#ifdef WITH_MPI
           call timer%start("mpi_communication")
Andreas Marek's avatar
Andreas Marek committed
794
	   call MPI_Bcast(vr, lr+1, &
795
#if REALCASE == 1
Andreas Marek's avatar
Andreas Marek committed
796
                          MPI_REAL_PRECISION, &
797
798
#endif
#if COMPLEXCASE == 1
Andreas Marek's avatar
Andreas Marek committed
799
                          MPI_COMPLEX_PRECISION, &
800
#endif
Andreas Marek's avatar
Andreas Marek committed
801
                          cur_pcol, mpi_comm_cols, mpierr)
802
803
804
           call timer%stop("mpi_communication")

#endif /* WITH_MPI */
805
806

#if REALCASE == 1
807
808
809
810
811
           if (useGPU) then
             vmrCUDA(cur_l_rows * (lc - 1) + 1 : cur_l_rows * (lc - 1) + lr) = vr(1:lr)
           else
             vmrCPU(1:lr,lc) = vr(1:lr)
           endif
812
813
814
815
#endif
#if COMPLEXCASE == 1
           vmrCPU(1:lr,lc) = vr(1:lr)
#endif
816
817
           tau = vr(lr+1)

818
819
820
821
822
823
#if REALCASE == 1
           tmat(lc,lc,istep) = tau ! Store tau in diagonal of tmat
#endif
#if COMPLEXCASE == 1
           tmat(lc,lc,istep) = conjg(tau) ! Store tau in diagonal of tmat
#endif
824
825
826
           ! Transform remaining columns in current block with Householder vector
           ! Local dot product

827
#if REALCASE == 1
828
           aux1 = 0
829
830
831
832
#endif
#if COMPLEXCASE == 1
          aux1 = CONST_COMPLEX_0_0
#endif
833

834
#if REALCASE == 1
835
#ifdef WITH_OPENMP
Andreas Marek's avatar
Andreas Marek committed
836

837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
           !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

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

           !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

           ! Get global dot products

           !$omp barrier
           !$omp single
#ifdef WITH_MPI
           call timer%start("mpi_communication")
865
           if (mynlc>0) call mpi_allreduce(aux1, aux2, mynlc, MPI_REAL_PRECISION, MPI_SUM, mpi_comm_rows, mpierr)
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
           call timer%stop("mpi_communication")
#else /* WITH_MPI */
           if (mynlc>0) aux2 = aux1
#endif /* WITH_MPI */
           !$omp end single
           !$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
Andreas Marek's avatar
Andreas Marek committed
891

892
893
894
895
896
897
898
899
900
901
902
903
904
905
#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

           ! Get global dot products
#ifdef WITH_MPI
           call timer%start("mpi_communication")
906
           if (nlc>0) call mpi_allreduce(aux1, aux2, nlc, MPI_REAL_PRECISION, MPI_SUM, mpi_comm_rows, mpierr)
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
           call timer%stop("mpi_communication")
#else /* WITH_MPI */
           if (nlc>0) aux2=aux1
#endif /* WITH_MPI */
           ! Transform

           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 */
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
#endif /* REALCASE == 1 */
#if COMPLEXCASE == 1
           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
               aux1(nlc) = dot_product(vr(1:lr),a(1:lr,lcx))
             endif
           enddo

           ! Get global dot products
#ifdef WITH_MPI
           call timer%start("mpi_communication")
           if (nlc>0) call mpi_allreduce(aux1, aux2, nlc, MPI_COMPLEX_PRECISION, MPI_SUM, mpi_comm_rows, mpierr)

           ! Transform

           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) - conjg(tau)*aux2(nlc)*vr(1:lr)
Andreas Marek's avatar
Andreas Marek committed
946

947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
             endif
           enddo

           call timer%stop("mpi_communication")

#else /* WITH_MPI */
!          if (nlc>0) aux2=aux1

           ! Transform

           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) - conjg(tau)*aux1(nlc)*vr(1:lr)
             endif
           enddo

#endif /* WITH_MPI */
!
!           ! Transform
!
!           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) - conjg(tau)*aux2(nlc)*vr(1:lr)
!             endif
!           enddo
#endif
979
980
981
982
983
984
         enddo ! lc

         if (useGPU) then
           ! store column tiles back to GPU
           cur_pcol = pcol(istep*nbw+1, nblk, np_cols)
           if (my_pcol == cur_pcol) then
985
#if REALCASE == 1
986
987
988
             successCUDA = cuda_memcpy2d((a_dev+((lc_start-1)*lda*size_of_PRECISION_real)),          &
                                          lda*size_of_PRECISION_real, loc(a(1, lc_start)),           &
                                          lda*size_of_PRECISION_real,  lr_end*size_of_PRECISION_real, &
989
990
991
992
993
                                          (lc_end - lc_start+1),cudaMemcpyHostToDevice)
             if (.not.(successCUDA)) then
               print *,"bandred_real: error in cudaMemcpy2d"
               stop
             endif
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
#endif
#if COMPLEXCASE == 1
             successCUDA = cuda_memcpy2d((a_dev+int(((lc_start-1)*lda*size_of_PRECISION_complex),kind=c_size_t)),    &
                                        int(lda*size_of_PRECISION_complex,kind=c_size_t), loc(a(1,lc_start)),       &
                                        int(lda*size_of_PRECISION_complex,kind=c_size_t),                           &
                                        int(lr_end*size_of_PRECISION_complex,kind=c_size_t),                        &
                                        int((lc_end - lc_start+1),kind=c_size_t) &
                                        ,int(cudaMemcpyHostToDevice,kind=c_int))
             if (.not.(successCUDA)) then
               print *, "bandred_complex: cuda memcpy a_dev  failed ", istat
               stop
             endif
#endif
1007
1008
1009
1010
1011
1012
1013
           endif
         endif

         ! Calculate scalar products of stored Householder vectors.
         ! This can be done in different ways, we use dsyrk

         vav = 0
1014
	 call timer%start("blas")
1015
1016
         if (useGPU) then
           if (l_rows>0) &
Andreas Marek's avatar
Andreas Marek committed
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
#if REALCASE == 1
             call PRECISION_SYRK('U', 'T',            &
#endif
#if COMPLEXCASE == 1
             call PRECISION_HERK('U', 'C',            &
#endif
	                         n_cols, l_rows, ONE, &
#if REALCASE == 1
				 vmrCUDA, cur_l_rows, &
#endif
1027
#if COMPLEXCASE == 1
Andreas Marek's avatar
Andreas Marek committed
1028
1029
1030
                                 vmrCPU, ubound(vmrCPU,dim=1), &
#endif
				 ZERO, vav, ubound(vav,dim=1))
1031
1032
         else
           if (l_rows>0) &
Andreas Marek's avatar
Andreas Marek committed
1033
1034
#if REALCASE == 1
             call PRECISION_SYRK('U', 'T',           &
1035
1036
#endif
#if COMPLEXCASE == 1
Andreas Marek's avatar
Andreas Marek committed
1037
             call PRECISION_HERK('U', 'C',           &
1038
#endif
Andreas Marek's avatar
Andreas Marek committed
1039
1040
	                         n_cols, l_rows, ONE, vmrCPU, ubound(vmrCPU,dim=1), ZERO, vav, ubound(vav,dim=1))
         endif
1041
	 call timer%stop("blas")
1042
#if REALCASE == 1
1043
	 call symm_matrix_allreduce_&
1044
1045
#endif
#if COMPLEXCASE == 1
1046
         call herm_matrix_allreduce_&
1047
#endif
Andreas Marek's avatar
Andreas Marek committed
1048
         &PRECISION &
1049
                         (n_cols,vav, nbw, nbw,mpi_comm_rows)
1050
         ! Calculate triangular matrix T for block Householder Transformation
1051
	 call timer%start("blas")
1052
1053
1054
         do lc=n_cols,1,-1
           tau = tmat(lc,lc,istep)
           if (lc<n_cols) then
1055
#if REALCASE == 1
Andreas Marek's avatar
Andreas Marek committed
1056
             call PRECISION_TRMV('U', 'T', 'N',          &
1057
1058
#endif
#if COMPLEXCASE == 1
Andreas Marek's avatar
Andreas Marek committed
1059
             call PRECISION_TRMV('U', 'C', 'N',          &
1060
#endif
Andreas Marek's avatar
Andreas Marek committed
1061
                                 n_cols-lc, tmat(lc+1,lc+1,istep), ubound(tmat,dim=1), vav(lc+1,lc), 1)
1062
1063

#if REALCASE == 1
1064
             tmat(lc,lc+1:n_cols,istep) = -tau * vav(lc+1:n_cols,lc)
1065
1066
1067
1068
#endif
#if COMPLEXCASE == 1
             tmat(lc,lc+1:n_cols,istep) = -tau * conjg(vav(lc+1:n_cols,lc))
#endif
1069
1070
           endif
         enddo
1071
 	 call timer%stop("blas")
1072
1073
1074
#if REALCASE == 1
       endif !useQR
#endif
1075
       ! Transpose vmr -> vmc (stored in umc, second half)
1076
#if REALCASE == 1
1077
       if (useGPU) then
1078
         call elpa_transpose_vectors_&
Andreas Marek's avatar
Andreas Marek committed
1079
1080
1081
              &MATH_DATATYPE&
              &_&
              &PRECISION &
1082
	                                   (vmrCUDA, cur_l_rows, mpi_comm_rows, &
1083
1084
1085
                                            umcCUDA(cur_l_cols * n_cols + 1), cur_l_cols, mpi_comm_cols, &
                                            1, istep*nbw, n_cols, nblk)
       else
1086
         call elpa_transpose_vectors_&
Andreas Marek's avatar
Andreas Marek committed
1087
1088
1089
              &MATH_DATATYPE&
              &_&
              &PRECISION &
1090
                                           (vmrCPU, ubound(vmrCPU,dim=1), mpi_comm_rows, &
1091
1092
1093
                                            umcCPU(1,n_cols+1), ubound(umcCPU,dim=1), mpi_comm_cols, &
                                            1, istep*nbw, n_cols, nblk)
       endif
1094
1095
#endif
#if COMPLEXCASE == 1
1096
         call elpa_transpose_vectors_&
Andreas Marek's avatar
Andreas Marek committed
1097
1098
1099
              &MATH_DATATYPE&
              &_&
              &PRECISION &
1100
                                      (vmrCPU, ubound(vmrCPU,dim=1), mpi_comm_rows, &
1101
1102
1103
                                      umcCPU(1,n_cols+1), ubound(umcCPU,dim=1), mpi_comm_cols, &
                                      1, istep*nbw, n_cols, nblk)
#endif
1104
1105
1106
1107
1108

       ! 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
1109
#if REALCASE == 1
1110
1111
1112
       ! 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
       if (useGPU) then
1113
1114
         umcCUDA(1 : l_cols * n_cols) = CONST_0_0
         vmrCUDA(cur_l_rows * n_cols + 1 : cur_l_rows * n_cols * 2) = CONST_0_0
1115
1116

         if (l_cols>0 .and. l_rows>0) then
1117
           successCUDA = cuda_memcpy(vmr_dev, loc(vmrCUDA(1)), vmr_size*size_of_PRECISION_real,cudaMemcpyHostToDevice)
1118
1119
1120
1121
           if (.not.(successCUDA)) then
             print *,"bandred_real: error in cudaMemcpy"
             stop
           endif
1122
           successCUDA = cuda_memcpy(umc_dev, loc(umcCUDA(1)), umc_size*size_of_PRECISION_real,cudaMemcpyHostToDevice)
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
           if (.not.(successCUDA)) then
             print *,"bandred_real: error in cudaMemcpy"
             stop
           endif

           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
1133
             call timer%start("cublas")
1134
             lre = min(l_rows,(i+1)*l_rows_tile)
1135
             call cublas_PRECISION_GEMM('T', 'N', lce-lcs+1, n_cols, lre, &
1136
1137
                               CONST_1_0, (a_dev + ((lcs-1)*lda*size_of_PRECISION_real)), lda, vmr_dev,cur_l_rows, &
                               CONST_1_0, (umc_dev+ (lcs-1)*size_of_PRECISION_real), cur_l_cols)
1138
1139
             if(i==0) cycle
             lre = min(l_rows,i*l_rows_tile)
1140
             call cublas_PRECISION_GEMM('N', 'N', lre,n_cols, lce-lcs+1,&
1141
1142
1143
                               CONST_1_0, (a_dev+ ((lcs-1)*lda*size_of_PRECISION_real)), lda,                  &
                               (umc_dev+(cur_l_cols * n_cols+lcs-1)*size_of_PRECISION_real), cur_l_cols, &
                               CONST_1_0, (vmr_dev+(cur_l_rows * n_cols)*size_of_PRECISION_real), cur_l_rows)
1144
             call timer%stop("cublas")
1145
           enddo
1146
           successCUDA = cuda_memcpy(loc(vmrCUDA(1)), vmr_dev,vmr_size*size_of_PRECISION_real,cudaMemcpyDeviceToHost)
1147
1148
1149
1150
           if (.not.(successCUDA)) then
             print *,"bandred_real: error in cudaMemcpy"
             stop
           endif
1151
           successCUDA = cuda_memcpy(loc(umcCUDA(1)), umc_dev, umc_size*size_of_PRECISION_real,cudaMemcpyDeviceToHost)
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
           if (.not.(successCUDA)) then
             print *,"bandred_real: error in cudaMemcpy"
             stop
           endif

         endif ! l_cols>0 .and. l_rows>0

       else ! do not useGPU version

         !Code for Algorithm 4

         n_way = 1
#ifdef WITH_OPENMP
         n_way = omp_get_max_threads()
#endif
1167
1168
         !umcCPU(1:l_cols,1:n_cols) = 0.d0
         !vmrCPU(1:l_rows,n_cols+1:2*n_cols) = 0
1169
1170
1171
1172
1173
1174
#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)
1175
             umcCPU(i,1:n_cols) = CONST_0_0
1176
1177
1178
1179
           enddo

           !$omp do
           do i=1,l_rows
1180
             vmrCPU(i,n_cols+1:2*n_cols) = CONST_0_0
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
           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
1210
	         call timer%start("blas")
1211
1212
                 call PRECISION_GEMM('N', 'N', lre-lrs+1, n_cols, l_cols-lcs+1,          &
                            CONST_1_0, a(lrs,lcs), ubound(a,dim=1),                 &
1213
                                  umcCPU(lcs,n_cols+1), ubound(umcCPU,dim=1),  &
1214
                            CONST_0_0, vmrCPU(lrs,n_cols+1), ubound(vmrCPU,dim=1))
1215
	         call timer%stop("blas")
1216
1217
1218
1219
               endif

               ! C1 += A10' B0
               if ( lce > lcs .and. i > 0 ) then
1220
	       	 call timer%start("blas")
1221
1222
                 call PRECISION_GEMM('T', 'N', lce-lcs+1, n_cols, lrs-1,           &
                            CONST_1_0, a(1,lcs),   ubound(a,dim=1),           &
1223
                                  vmrCPU(1,1),   ubound(vmrCPU,dim=1),   &
1224
                            CONST_0_0, umcCPU(lcs,1), ubound(umcCPU,dim=1))
1225
	       	 call timer%stop("blas")
1226
1227
1228
1229
               endif
             enddo
           endif ! l_cols>0 .and. l_rows>0
         else ! n_way > 1
1230
1231
           umcCPU(1:l_cols,1:n_cols) = CONST_0_0
           vmrCPU(1:l_rows,n_cols+1:2*n_cols) = CONST_0_0
1232
1233
1234
1235
1236
1237
           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
1238
	       call timer%start("blas")
1239
               lre = min(l_rows,(i+1)*l_rows_tile)
1240
1241
               call PRECISION_GEMM('T', 'N', lce-lcs+1, n_cols, lre, CONST_1_0, a(1,lcs), ubound(a,dim=1), &
                            vmrCPU, ubound(vmrCPU,dim=1), CONST_1_0, umcCPU(lcs,1), ubound(umcCPU,dim=1))
1242
	       call timer%stop("blas")
1243
1244
               if (i==0) cycle
                 lre = min(l_rows,i*l_rows_tile)
1245
	       	 call timer%start("blas")
1246
1247
                 call PRECISION_GEMM('N', 'N', lre, n_cols, lce-lcs+1, CONST_1_0, a(1,lcs), lda, &
                            umcCPU(lcs,n_cols+1), ubound(umcCPU,dim=1), CONST_1_0, vmrCPU(1,n_cols+1), ubound(vmrCPU,dim=1))
1248
	       	 call timer%stop("blas")
1249
1250
1251
1252
1253
1254
1255
             enddo
           endif
         endif ! n_way > 1
#ifdef WITH_OPENMP
        !$omp end parallel
#endif
       endif ! do not useGPU version
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
#endif /* REALCASE == 1 */

#if COMPLEXCASE == 1
        umcCPU(1:l_cols,1:n_cols) = CONST_COMPLEX_0_0
        vmrCPU(1:l_rows,n_cols+1:2*n_cols) = CONST_COMPLEX_0_0
        if (l_cols>0 .and. l_rows>0) then
          if (useGPU) then
            if (size(vmrCPU,dim=1)*size(vmrCPU,dim=2) .gt. vmr_size) then
              print *,"bandred_complex: vmr size 2 :",size(vmrCPU,dim=1)*size(vmrCPU,dim=2),vmr_size
              stop
            endif
            successCUDA = cuda_memcpy(vmr_dev, loc(vmrCPU(1,1)),vmr_size*size_of_PRECISION_complex,cudaMemcpyHostToDevice)

            if (.not.(successCUDA)) then
              print *, "bandred_complex:  cuda memcpy vmr_dev failed ", istat
              stop
            endif
            if (size(umcCPU,dim=1)*size(umcCPU,dim=2) .gt. umc_size) then
              print *,"bandred_complex: umc size 2 :",size(umcCPU,dim=1)*size(umcCPU,dim=2),umc_size
              stop
            endif
            successCUDA = cuda_memcpy(umc_dev, loc(umcCPU(1,1)),umc_size*size_of_PRECISION_complex,cudaMemcpyHostToDevice)
            if (.not.(successCUDA)) then
              print *, "bandred_complex:  cuda memcpy umc_dev failed  ", istat
              stop
            endif
          endif
          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)

            if (useGPU) then
              call timer%start("cublas")
Andreas Marek's avatar
Andreas Marek committed
1293
              call cublas_PRECISION_GEMM('C', 'N', lce-lcs+1, n_cols, lre, ONE, (a_dev + ((lcs-1)*lda* &
1294
                        size_of_PRECISION_complex)), lda, &
Andreas Marek's avatar
Andreas Marek committed
1295
                        vmr_dev, cur_l_rows, ONE, (umc_dev +(lcs-1)*size_of_PRECISION_complex), cur_l_cols)
1296
1297
1298
              call timer%stop("cublas")
            else
              call timer%start("blas")
Andreas Marek's avatar
Andreas Marek committed
1299
1300
              call PRECISION_GEMM('C', 'N', lce-lcs+1, n_cols, lre, ONE, a(1,lcs), ubound(a,dim=1), &
                         vmrCPU, ubound(vmrCPU,dim=1), ONE, umcCPU(lcs,1), ubound(umcCPU,dim=1))
1301
1302
1303
1304
1305
1306
1307
              call timer%stop("blas")
            endif

            if (i==0) cycle
            lre = min(l_rows,i*l_rows_tile)
            if (useGPU) then
              call timer%start("cublas")
Andreas Marek's avatar
Andreas Marek committed
1308
              call cublas_PRECISION_GEMM('N', 'N', lre, n_cols, lce-lcs+1, ONE, (a_dev+((lcs-1)*lda* &
1309
                        size_of_PRECISION_complex)),lda,  &
Andreas Marek's avatar
Andreas Marek committed
1310
                        (umc_dev+(cur_l_cols * n_cols+lcs-1)*size_of_PRECISION_complex), cur_l_cols,ONE,  &
1311
1312
1313
1314
                        (vmr_dev+(cur_l_rows * n_cols)*size_of_PRECISION_complex), cur_l_rows)
              call timer%stop("cublas")
            else
              call timer%start("blas")
Andreas Marek's avatar
Andreas Marek committed
1315
1316
              call PRECISION_GEMM('N', 'N', lre, n_cols, lce-lcs+1, ONE, a(1,lcs), lda, &
                         umcCPU(lcs,n_cols+1), ubound(umcCPU,dim=1), ONE, vmrCPU(1,n_cols+1), ubound(vmrCPU,dim=1))
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
              call timer%stop("blas")
            endif
          enddo

          if (useGPU) then
            if (size(vmrCPU,dim=1)*size(vmrCPU,dim=2) .gt. vmr_size) then
              print *,"bandred_complex: vmr size 3 :",size(vmrCPU,dim=1)*size(vmrCPU,dim=2),vmr_size
              stop
            endif
            successCUDA = cuda_memcpy(loc(vmrCPU(1,1)),vmr_dev,vmr_size*size_of_PRECISION_complex,cudaMemcpyDeviceToHost)
            if (.not.(successCUDA)) then
              print *, "bandred_complex:  cuad memcpy failed vmrCPU ", istat
              stop
            endif
            if (size(umcCPU,dim=1)*size(umcCPU,dim=2) .gt. umc_size) then
              print *,"bandred_complex: umc size 3 :",size(umcCPU,dim=1)*size(umcCPU,dim=2),umc_size
              stop
            endif
            successCUDA = cuda_memcpy(loc(umcCPU(1,1)), umc_dev,umc_size*size_of_PRECISION_complex,cudaMemcpyDeviceToHost)
            if (.not.(successCUDA)) then
              print *, "bandred_complex:  cuad memcpy failed umcCPU ", istat
              stop
            endif
          endif ! useGPU
1341
        endif ! (l_cols>0 .and. l_rows>0)
1342
#endif /* COMPLEXCASE == 1 */
1343
1344
1345
1346
1347
1348

       ! 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

1349
#if REALCASE == 1
1350
1351
1352
1353
       if (useGPU) then
         ! here the GPU version and CPU version divereged due to the same reasons as above

         if (tile_size < istep*nbw) then
1354
           call elpa_reduce_add_vectors_&
Andreas Marek's avatar
Cleanup    
Andreas Marek committed
1355
1356
1357
                &MATH_DATATYPE&
                &_&
                &PRECISION &
1358
                                               (vmrCUDA(cur_l_rows * n_cols + 1),cur_l_rows,mpi_comm_rows, &
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
                                               umcCUDA, cur_l_cols, mpi_comm_cols, &
                                               istep*nbw, n_cols, nblk)
         endif

         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

#ifdef WITH_MPI
           call timer%start("mpi_communication")

1373
           call mpi_allreduce(umcCUDA, tmpCUDA, l_cols*n_cols, MPI_REAL_PRECISION, MPI_SUM, mpi_comm_rows, ierr)
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
           umcCUDA(1 : l_cols * n_cols) = tmpCUDA(1 : l_cols * n_cols)
           call timer%stop("mpi_communication")
#else /* WITH_MPI */

!           tmpCUDA(1 : l_cols * n_cols) = umcCUDA(1 : l_cols * n_cols)

#endif /* WITH_MPI */

           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

         ! U = U * Tmat**T
1392
         successCUDA = cuda_memcpy(umc_dev, loc(umcCUDA(1)), umc_size*size_of_PRECISION_real, cudaMemcpyHostToDevice)
1393
1394
1395
1396
         if (.not.(successCUDA)) then
           print *,"bandred_real: error in cudaMemcpy"
           stop
         endif
1397
         successCUDA = cuda_memcpy(tmat_dev,loc(tmat(1,1,istep)),nbw*nbw*size_of_PRECISION_real,cudaMemcpyHostToDevice)
1398
1399
1400
1401
         if (.not.(successCUDA)) then
           print *,"bandred_real: error in cudaMemcpy"
           stop
         endif
1402
	 call timer%start("cublas")
1403
         call cublas_PRECISION_TRMM('Right', 'Upper', 'Trans', 'Nonunit', l_cols, n_cols, &
1404
                           CONST_1_0, tmat_dev, nbw, umc_dev, cur_l_cols)
1405
1406
	 call timer%start("cublas")

1407
         ! VAV = Tmat * V**T * A * V * Tmat**T = (U*Tmat**T)**T * V * Tmat**T
1408
         successCUDA = cuda_memcpy(vav_dev,loc(vav(1,1)), nbw*nbw*size_of_PRECISION_real,cudaMemcpyHostToDevice)
1409
1410
1411
1412
         if (.not.(successCUDA)) then
           print *,"bandred_real: error in cudaMemcpy"
           stop
         endif
1413
1414
	 call timer%start("cublas")

1415
         call cublas_PRECISION_GEMM('T', 'N', n_cols, n_cols, l_cols, &
1416
1417
                           CONST_1_0, umc_dev, cur_l_cols, (umc_dev+(cur_l_cols * n_cols )*size_of_PRECISION_real),cur_l_cols, &
                           CONST_0_0, vav_dev, nbw)
1418

1419
         call cublas_PRECISION_TRMM('Right', 'Upper', 'Trans', 'Nonunit', n_cols, n_cols, &
1420
                           CONST_1_0, tmat_dev, nbw, vav_dev, nbw)
1421
	 call timer%stop("cublas")
1422

1423
         successCUDA = cuda_memcpy(loc(vav(1,1)), vav_dev, nbw*nbw*size_of_PRECISION_real, cudaMemcpyDeviceToHost)
1424
1425
1426
1427
1428
         if (.not.(successCUDA)) then
           print *,"bandred_real: error in cudaMemcpy"
           stop
         endif

1429
         call symm_matrix_allreduce_&
Andreas Marek's avatar
Cleanup    
Andreas Marek committed
1430
         &PRECISION &
1431
	                          (n_cols,vav, nbw,nbw,mpi_comm_cols)
1432

1433
         successCUDA = cuda_memcpy(vav_dev, loc(vav(1,1)), nbw*nbw*size_of_PRECISION_real,cudaMemcpyHostToDevice)
1434
1435
1436
1437
1438
1439
         if (.not.(successCUDA)) then
           print *,"bandred_real: error in cudaMemcpy"
           stop
         endif

         ! U = U - 0.5 * V * VAV
1440
1441
 	 call timer%start("cublas")

1442
         call cublas_PRECISION_GEMM('N', 'N', l_cols, n_cols, n_cols,&
1443
1444
                           -CONST_0_5, (umc_dev+(cur_l_cols * n_cols )*size_of_PRECISION_real),cur_l_cols, vav_dev,nbw,&
                           CONST_1_0, umc_dev, cur_l_cols)
1445
	 call timer%stop("cublas")
1446

1447
         successCUDA = cuda_memcpy(loc(umcCUDA(1)), umc_dev, umc_size*size_of_PRECISION_real, cudaMemcpyDeviceToHost)
1448
1449
1450
1451
1452
1453
1454

         if (.not.(successCUDA)) then
           print *,"bandred_real: error in cudaMemcpy"
           stop
         endif

         ! Transpose umc -> umr (stored in vmr, second half)
1455
         call elpa_transpose_vectors_&
Andreas Marek's avatar
Cleanup    
Andreas Marek committed
1456
1457
1458
         &MATH_DATATYPE&
         &_&
         &PRECISION &
1459
                                           (umcCUDA, cur_l_cols, mpi_comm_cols, &
1460
1461
1462
                                            vmrCUDA(cur_l_rows * n_cols + 1), cur_l_rows, mpi_comm_rows, &
                                            1, istep*nbw, n_cols, nblk)

1463
         successCUDA = cuda_memcpy(vmr_dev, loc(vmrCUDA(1)), vmr_size*size_of_PRECISION_real, cudaMemcpyHostToDevice)
1464
1465
1466
1467
1468
         if (.not.(successCUDA)) then
           print *,"bandred_real: error in cudaMemcpy"
           stop
         endif

1469
         successCUDA = cuda_memcpy(umc_dev, loc(umcCUDA(1)), umc_size*size_of_PRECISION_real, cudaMemcpyHostToDevice)
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
         if (.not.(successCUDA)) then
           print *,"bandred_real: error in cudaMemcpy"
           stop
         endif

         ! 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
1481
1482
	   call timer%start("cublas")

1483
           call cublas_PRECISION_GEMM('N', 'T', lre, lce-lcs+1, 2*n_cols, -CONST_1_0, &
1484
1485
                             vmr_dev, cur_l_rows, (umc_dev +(lcs-1)*size_of_PRECISION_real), cur_l_cols, &
                             CONST_1_0, (a_dev+(lcs-1)*lda*size_of_PRECISION_real), lda)
1486
1487
	   call timer%stop("cublas")

1488
1489
1490
1491
1492
1493
         enddo

       else ! do not useGPU

         ! Or if we used the Algorithm 4
         if (tile_size < istep*nbw .or. n_way > 1) then
1494
           call elpa_reduce_add_vectors_&
Andreas Marek's avatar
Cleanup    
Andreas Marek committed
1495
1496
1497
           &MATH_DATATYPE&
           &_&
           &PRECISION &
1498
                                            (vmrCPU(1,n_cols+1),ubound(vmrCPU,dim=1),mpi_comm_rows, &
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
                                             umcCPU, ubound(umcCPU,dim=1), mpi_comm_cols, &
                                             istep*nbw, n_cols, nblk)
         endif

         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

#ifdef WITH_MPI
           call timer%start("mpi_communication")
1512
           call mpi_allreduce(umcCPU, tmpCPU, l_cols*n_cols, MPI_REAL_PRECISION, MPI_SUM, mpi_comm_rows, mpierr)
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
           umcCPU(1:l_cols,1:n_cols) = tmpCPU(1:l_cols,1:n_cols)
           call timer%stop("mpi_communication")
#else /* WITH_MPI */
!           tmpCPU(1:l_cols,1:n_cols) = umcCPU(1:l_cols,1:n_cols)
#endif /* WITH_MPI */

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

         ! U = U * Tmat**T
1527
1528
	 call timer%start("blas")

1529
         call PRECISION_TRMM('Right', 'Upper', 'Trans', 'Nonunit', l_cols,n_cols, CONST_1_0, tmat(1,1,istep), ubound(tmat,dim=1), &
1530
1531
1532
1533
                    umcCPU, ubound(umcCPU,dim=1))

         ! VAV = Tmat * V**T * A * V * Tmat**T = (U*Tmat**T)**T * V * Tmat**T

1534
1535
         call PRECISION_GEMM('T', 'N', n_cols, n_cols, l_cols, CONST_1_0, umcCPU, ubound(umcCPU,dim=1), umcCPU(1,n_cols+1), &
                    ubound(umcCPU,dim=1), CONST_0_0, vav, ubound(vav,dim=1))
1536

1537
         call PRECISION_TRMM('Right', 'Upper', 'Trans', 'Nonunit', n_cols, n_cols, CONST_1_0, tmat(1,1,istep),    &
1538
                    ubound(tmat,dim=1), vav, ubound(vav,dim=1))
1539
	 call timer%stop("blas")
1540
         call symm_matrix_allreduce_&
Andreas Marek's avatar
Cleanup    
Andreas Marek committed
1541
         &PRECISION &
1542
	                            (n_cols,vav, nbw, nbw ,mpi_comm_cols)
1543
1544

         ! U = U - 0.5 * V * VAV
1545
	 call timer%start("blas")
1546
1547
         call PRECISION_GEMM('N', 'N', l_cols, n_cols, n_cols, -CONST_0_5, umcCPU(1,n_cols+1), ubound(umcCPU,dim=1), vav, &
                     ubound(vav,dim=1), CONST_1_0, umcCPU, ubound(umcCPU,dim=1))
1548
	 call timer%stop("blas")
1549
         ! Transpose umc -> umr (stored in vmr, second half)
1550
         call elpa_transpose_vectors_&
Andreas Marek's avatar
Cleanup    
Andreas Marek committed
1551
1552
1553
         &MATH_DATATYPE&
         &_&
         &PRECISION &