elpa2_compute.F90 215 KB
Newer Older
1 2 3 4 5
!    This file is part of ELPA.
!
!    The ELPA library was originally created by the ELPA consortium,
!    consisting of the following organizations:
!
6 7
!    - Max Planck Computing and Data Facility (MPCDF), fomerly known as
!      Rechenzentrum Garching der Max-Planck-Gesellschaft (RZG),
8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
!    - 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:
23
!    http://elpa.mpcdf.mpg.de/
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
!
!    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
74
  use elpa_mpi
75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134

  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)
  contains

    subroutine bandred_real(na, a, lda, nblk, nbw, matrixCols, numBlocks, mpi_comm_rows, mpi_comm_cols, &
                            tmat, wantDebug, success, useQR)

    !-------------------------------------------------------------------------------
    !  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
    !
    !-------------------------------------------------------------------------------
#ifdef HAVE_DETAILED_TIMINGS
      use timings
#endif
#ifdef WITH_OPENMP
      use omp_lib
#endif
Andreas Marek's avatar
Andreas Marek committed
135
      use precision
136 137
      implicit none

Andreas Marek's avatar
Andreas Marek committed
138
      integer(kind=ik)           :: na, lda, nblk, nbw, matrixCols, numBlocks, mpi_comm_rows, mpi_comm_cols
Andreas Marek's avatar
Andreas Marek committed
139 140 141
#ifdef DESPERATELY_WANT_ASSUMED_SIZE
      real(kind=rk)              :: a(lda,*), tmat(nbw,nbw,*)
#else
Andreas Marek's avatar
Andreas Marek committed
142
      real(kind=rk)              :: a(lda,matrixCols), tmat(nbw,nbw,numBlocks)
Andreas Marek's avatar
Andreas Marek committed
143
#endif
Andreas Marek's avatar
Andreas Marek committed
144
      integer(kind=ik)           :: my_prow, my_pcol, np_rows, np_cols, mpierr
145
      integer(kind=ik)           :: l_cols, l_rows, vmrCols
Andreas Marek's avatar
Andreas Marek committed
146 147 148
      integer(kind=ik)           :: i, j, lcs, lce, lrs, lre, lc, lr, cur_pcol, n_cols, nrow
      integer(kind=ik)           :: istep, ncol, lch, lcx, nlc, mynlc
      integer(kind=ik)           :: tile_size, l_rows_tile, l_cols_tile
149

Andreas Marek's avatar
Andreas Marek committed
150
      real(kind=rk)              :: vnorm2, xf, aux1(nbw), aux2(nbw), vrl, tau, vav(nbw,nbw)
151

Andreas Marek's avatar
Andreas Marek committed
152
      real(kind=rk), allocatable :: tmp(:,:), vr(:), vmr(:,:), umc(:,:)
153 154

      ! needed for blocked QR decomposition
Andreas Marek's avatar
Andreas Marek committed
155 156 157
      integer(kind=ik)           :: PQRPARAM(11), work_size
      real(kind=rk)              :: dwork_size(1)
      real(kind=rk), allocatable :: work_blocked(:), tauvector(:), blockheuristic(:)
158

Andreas Marek's avatar
Andreas Marek committed
159 160
      logical, intent(in)        :: wantDebug
      logical, intent(out)       :: success
161

Andreas Marek's avatar
Andreas Marek committed
162
      logical, intent(in)        :: useQR
163

Andreas Marek's avatar
Andreas Marek committed
164
      integer(kind=ik)           :: mystart, myend, m_way, n_way, work_per_thread, m_id, n_id, n_threads, ii, pp, transformChunkSize
165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197

#ifdef HAVE_DETAILED_TIMINGS
      call timer%start("bandred_real")
#endif
      call mpi_comm_rank(mpi_comm_rows,my_prow,mpierr)
      call mpi_comm_size(mpi_comm_rows,np_rows,mpierr)
      call mpi_comm_rank(mpi_comm_cols,my_pcol,mpierr)
      call mpi_comm_size(mpi_comm_cols,np_cols,mpierr)
      success = .true.


      ! 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

      ! 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

      if (useQR) then
        if (which_qr_decomposition == 1) then
198
          call qr_pqrparam_init(pqrparam(1:11),    nblk,'M',0,   nblk,'M',0,   nblk,'M',1,'s')
199 200 201 202 203
          allocate(tauvector(na))
          allocate(blockheuristic(nblk))
          l_rows = local_index(na, my_prow, np_rows, nblk, -1)
          allocate(vmr(max(l_rows,1),na))

204 205 206 207 208 209 210 211 212 213 214 215
          vmrCols = na
#ifdef DESPERATELY_WANT_ASSUMED_SIZE_QR
          call qr_pdgeqrf_2dcomm(a, lda, matrixCols, vmr, max(l_rows,1), vmrCols, tauvector(1), na, tmat(1,1,1), &
                                 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
          call qr_pdgeqrf_2dcomm(a(1:lda,1:matrixCols), matrixCols, lda, vmr(1:max(l_rows,1),1:vmrCols), max(l_rows,1), &
                                 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
216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246
          work_size = dwork_size(1)
          allocate(work_blocked(work_size))

          work_blocked = 0.0d0
          deallocate(vmr)
        endif
      endif

      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)

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

        allocate(vmr(max(l_rows,1),2*n_cols))
        allocate(umc(max(l_cols,1),2*n_cols))

        allocate(vr(l_rows+1))

        vmr(1:l_rows,1:n_cols) = 0.
        vr(:) = 0
        tmat(:,:,istep) = 0

        ! Reduce current block to lower triangular form

        if (useQR) then
          if (which_qr_decomposition == 1) then
247 248 249 250 251

            vmrCols = 2*n_cols
#ifdef DESPERATELY_WANT_ASSUMED_SIZE_QR
            call qr_pdgeqrf_2dcomm(a, lda, matrixCols, vmr, max(l_rows,1), vmrCols, tauvector(1), &
                                   na, tmat(1,1,istep), nbw, nbw, work_blocked, work_size,        &
252 253
                                     work_size, na, n_cols, nblk, nblk,        &
                                     istep*nbw+n_cols-nbw, istep*nbw+n_cols, 1,&
254
                                     0, PQRPARAM(1:11), mpi_comm_rows, mpi_comm_cols,&
255
                                     blockheuristic)
256 257 258 259 260 261 262 263 264 265

#else
            call qr_pdgeqrf_2dcomm(a(1:lda,1:matrixCols), lda, matrixCols, vmr(1:max(l_rows,1),1:vmrCols) ,   &
                                    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
266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296
          endif
        else

          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))
                aux1(2) = 0.
              endif
297
#ifdef WITH_MPI
298
              call mpi_allreduce(aux1,aux2,2,MPI_REAL8,MPI_SUM,mpi_comm_rows,mpierr)
299 300 301
#else
              aux2 = aux1
#endif
302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324
              vnorm2 = aux2(1)
              vrl    = aux2(2)

              ! Householder transformation

              call hh_transform_real(vrl, vnorm2, xf, tau)

              ! 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
                vr(lr) = 1.
              else
                a(1:lr,lch) = vr(1:lr)
              endif

            endif

            ! Broadcast Householder vector and tau along columns

            vr(lr+1) = tau
325
#ifdef WITH_MPI
326
            call MPI_Bcast(vr,lr+1,MPI_REAL8,cur_pcol,mpi_comm_cols,mpierr)
327
#endif
328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361
            vmr(1:lr,lc) = vr(1:lr)
            tau = vr(lr+1)
            tmat(lc,lc,istep) = tau ! Store tau in diagonal of tmat

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

            aux1 = 0
#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

            !$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
362
#ifdef WITH_MPI
363
            if (mynlc>0) call mpi_allreduce(aux1,aux2,mynlc,MPI_REAL8,MPI_SUM,mpi_comm_rows,mpierr)
364 365 366
#else
            if (mynlc>0) aux2 = aux1
#endif
367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387
            !$omp end single
            !$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
388
#else /* WITH_OPENMP */
389 390 391 392 393 394 395 396 397 398
            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
399
#ifdef WITH_MPI
400
            if (nlc>0) call mpi_allreduce(aux1,aux2,nlc,MPI_REAL8,MPI_SUM,mpi_comm_rows,mpierr)
401 402 403
#else
            if (nlc>0) aux2=aux1
#endif
404 405 406 407 408 409 410 411 412 413
            ! 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
414 415
#endif /* WITH_OPENMP */

416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543
          enddo

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

          vav = 0
          if (l_rows>0) &
              call dsyrk('U','T',n_cols,l_rows,1.d0,vmr,ubound(vmr,dim=1),0.d0,vav,ubound(vav,dim=1))
          call symm_matrix_allreduce(n_cols,vav, nbw, nbw,mpi_comm_rows)

          ! Calculate triangular matrix T for block Householder Transformation

          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)
            endif
          enddo
        endif

        ! Transpose vmr -> vmc (stored in umc, second half)

        call elpa_transpose_vectors_real  (vmr, ubound(vmr,dim=1), mpi_comm_rows, &
                                        umc(1,n_cols+1), ubound(umc,dim=1), mpi_comm_cols, &
                                        1, istep*nbw, n_cols, nblk)

        ! 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
        !Code for Algorithm 4

        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)
            umc(i,1:n_cols) = 0.d0
          enddo
          !$omp do
          do i=1,l_rows
            vmr(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),           &
                               umc(lcs,n_cols+1), ubound(umc,dim=1),  &
                         0.d0, vmr(lrs,n_cols+1), ubound(vmr,dim=1))
              endif

              ! 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),           &
                               vmr(1,1),   ubound(vmr,dim=1),         &
                         0.d0, umc(lcs,1), ubound(umc,dim=1))
              endif
            enddo
          endif
        else
          umc(1:l_cols,1:n_cols) = 0.d0
          vmr(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), &
                           vmr,ubound(vmr,dim=1),1.d0,umc(lcs,1),ubound(umc,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, &
                           umc(lcs,n_cols+1),ubound(umc,dim=1),1.d0,vmr(1,n_cols+1),ubound(vmr,dim=1))
            enddo
          endif
        endif
#ifdef WITH_OPENMP
        !$omp end parallel
#endif
        ! 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
        ! Or if we used the Algorithm 4
        if (tile_size < istep*nbw .or. n_way > 1) then
        call elpa_reduce_add_vectors_real  (vmr(1,n_cols+1),ubound(vmr,dim=1),mpi_comm_rows, &
                                            umc, ubound(umc,dim=1), mpi_comm_cols, &
                                            istep*nbw, n_cols, nblk)
        endif
544
#ifdef WITH_MPI
545 546 547 548 549 550
        if (l_cols>0) then
          allocate(tmp(l_cols,n_cols))
          call mpi_allreduce(umc,tmp,l_cols*n_cols,MPI_REAL8,MPI_SUM,mpi_comm_rows,mpierr)
          umc(1:l_cols,1:n_cols) = tmp(1:l_cols,1:n_cols)
          deallocate(tmp)
        endif
551
#endif
552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647
        ! U = U * Tmat**T

        call dtrmm('Right','Upper','Trans','Nonunit',l_cols,n_cols,1.d0,tmat(1,1,istep),ubound(tmat,dim=1),umc,ubound(umc,dim=1))

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

        call dgemm('T','N',n_cols,n_cols,l_cols,1.d0,umc,ubound(umc,dim=1),umc(1,n_cols+1), &
                   ubound(umc,dim=1),0.d0,vav,ubound(vav,dim=1))
        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)

        ! U = U - 0.5 * V * VAV
        call dgemm('N','N',l_cols,n_cols,n_cols,-0.5d0,umc(1,n_cols+1),ubound(umc,dim=1),vav, &
                    ubound(vav,dim=1),1.d0,umc,ubound(umc,dim=1))

        ! Transpose umc -> umr (stored in vmr, second half)

        call elpa_transpose_vectors_real  (umc, ubound(umc,dim=1), mpi_comm_cols, &
                                       vmr(1,n_cols+1), ubound(vmr,dim=1), mpi_comm_rows, &
                                       1, istep*nbw, n_cols, nblk)

        ! 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

        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, &
                      vmr(mystart, 1), ubound(vmr,1), umc(lcs,1), ubound(umc,1), &
                      1.d0,a(mystart,lcs),ubound(a,1))
        enddo
        !$omp end parallel

#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, &
                      vmr,ubound(vmr,dim=1),umc(lcs,1),ubound(umc,dim=1), &
                      1.d0,a(1,lcs),lda)
        enddo
#endif /* WITH_OPENMP */
        deallocate(vmr, umc, vr)

      enddo

      if (useQR) then
        if (which_qr_decomposition == 1) then
          deallocate(work_blocked)
          deallocate(tauvector)
        endif
      endif

#ifdef HAVE_DETAILED_TIMINGS
      call timer%stop("bandred_real")
#endif
    end subroutine bandred_real

    subroutine symm_matrix_allreduce(n,a,lda,ldb,comm)

    !-------------------------------------------------------------------------------
    !  symm_matrix_allreduce: Does an mpi_allreduce for a symmetric matrix A.
    !  On entry, only the upper half of A needs to be set
    !  On exit, the complete matrix is set
    !-------------------------------------------------------------------------------
#ifdef HAVE_DETAILED_TIMINGS
      use timings
#endif
Andreas Marek's avatar
Andreas Marek committed
648
      use precision
649
      implicit none
Andreas Marek's avatar
Andreas Marek committed
650
      integer(kind=ik)  :: n, lda, ldb, comm
Andreas Marek's avatar
Andreas Marek committed
651 652 653
#ifdef DESPERATELY_WANT_ASSUMED_SIZE
      real(kind=rk)     :: a(lda,*)
#else
Andreas Marek's avatar
Andreas Marek committed
654
      real(kind=rk)     :: a(lda,ldb)
Andreas Marek's avatar
Andreas Marek committed
655
#endif
Andreas Marek's avatar
Andreas Marek committed
656 657
      integer(kind=ik)  :: i, nc, mpierr
      real(kind=rk)     :: h1(n*n), h2(n*n)
658 659 660 661 662 663 664 665 666 667

#ifdef HAVE_DETAILED_TIMINGS
      call timer%start("symm_matrix_allreduce")
#endif

      nc = 0
      do i=1,n
        h1(nc+1:nc+i) = a(1:i,i)
        nc = nc+i
      enddo
668
#ifdef WITH_MPI
669
      call mpi_allreduce(h1,h2,nc,MPI_REAL8,MPI_SUM,comm,mpierr)
670 671 672
#else
      h2=h1
#endif
673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723
      nc = 0
      do i=1,n
        a(1:i,i) = h2(nc+1:nc+i)
        a(i,1:i-1) = a(1:i-1,i)
        nc = nc+i
      enddo

#ifdef HAVE_DETAILED_TIMINGS
      call timer%stop("symm_matrix_allreduce")
#endif

    end subroutine symm_matrix_allreduce

    subroutine trans_ev_band_to_full_real(na, nqc, nblk, nbw, a, lda, tmat, q, ldq, matrixCols, numBlocks, mpi_comm_rows, &
                                      mpi_comm_cols, useQR)
    !-------------------------------------------------------------------------------
    !  trans_ev_band_to_full_real:
    !  Transforms the eigenvectors of a band matrix back to the eigenvectors of the original matrix
    !
    !  Parameters
    !
    !  na          Order of matrix a, number of rows of matrix q
    !
    !  nqc         Number of columns of matrix q
    !
    !  nblk        blocksize of cyclic distribution, must be the same in both directions!
    !
    !  nbw         semi bandwith
    !
    !  a(lda,matrixCols)    Matrix containing the Householder vectors (i.e. matrix a after bandred_real)
    !              Distribution is like in Scalapack.
    !
    !  lda         Leading dimension of a
    !  matrixCols  local columns of matrix a and q
    !
    !  tmat(nbw,nbw,numBlocks) Factors returned by bandred_real
    !
    !  q           On input: Eigenvectors of band matrix
    !              On output: Transformed eigenvectors
    !              Distribution is like in Scalapack.
    !
    !  ldq         Leading dimension of q
    !
    !  mpi_comm_rows
    !  mpi_comm_cols
    !              MPI-Communicators for rows/columns
    !
    !-------------------------------------------------------------------------------
#ifdef HAVE_DETAILED_TIMINGS
      use timings
#endif
Andreas Marek's avatar
Andreas Marek committed
724
      use precision
725 726
      implicit none

Andreas Marek's avatar
Andreas Marek committed
727
      integer(kind=ik)            :: na, nqc, lda, ldq, nblk, nbw, matrixCols, numBlocks, mpi_comm_rows, mpi_comm_cols
Andreas Marek's avatar
Andreas Marek committed
728 729 730
#ifdef DESPERATELY_WANT_ASSUMED_SIZE
      real(kind=rk)               :: a(lda,*), q(ldq,*), tmat(nbw,nbw,*)
#else
Andreas Marek's avatar
Andreas Marek committed
731
      real(kind=rk)               :: a(lda,matrixCols), q(ldq,matrixCols), tmat(nbw, nbw, numBlocks)
Andreas Marek's avatar
Andreas Marek committed
732
#endif
Andreas Marek's avatar
Andreas Marek committed
733 734 735 736 737
      integer(kind=ik)            :: my_prow, my_pcol, np_rows, np_cols, mpierr
      integer(kind=ik)            :: max_blocks_row, max_blocks_col, max_local_rows, &
                                     max_local_cols
      integer(kind=ik)            :: l_cols, l_rows, l_colh, n_cols
      integer(kind=ik)            :: istep, lc, ncol, nrow, nb, ns
738

Andreas Marek's avatar
Andreas Marek committed
739
      real(kind=rk), allocatable  :: tmp1(:), tmp2(:), hvb(:), hvm(:,:)
740

Andreas Marek's avatar
Andreas Marek committed
741
      integer(kind=ik)            :: i
742

Andreas Marek's avatar
Andreas Marek committed
743 744 745
      real(kind=rk), allocatable  :: tmat_complete(:,:), t_tmp(:,:), t_tmp2(:,:)
      integer(kind=ik)            :: cwy_blocking, t_blocking, t_cols, t_rows
      logical, intent(in)         :: useQR
746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817

#ifdef HAVE_DETAILED_TIMINGS
      call timer%start("trans_ev_band_to_full_real")
#endif
      call mpi_comm_rank(mpi_comm_rows,my_prow,mpierr)
      call mpi_comm_size(mpi_comm_rows,np_rows,mpierr)
      call mpi_comm_rank(mpi_comm_cols,my_pcol,mpierr)
      call mpi_comm_size(mpi_comm_cols,np_cols,mpierr)
      max_blocks_row = ((na -1)/nblk)/np_rows + 1  ! Rows of A
      max_blocks_col = ((nqc-1)/nblk)/np_cols + 1  ! Columns of q!

      max_local_rows = max_blocks_row*nblk
      max_local_cols = max_blocks_col*nblk

      ! t_blocking was formerly 2; 3 is a better choice
      t_blocking = 3 ! number of matrices T (tmat) which are aggregated into a new (larger) T matrix (tmat_complete) and applied at once

      ! we only use the t_blocking if we could call it fully, this is might be better but needs to benchmarked.
!     if ( na >= ((t_blocking+1)*nbw) ) then
      cwy_blocking = t_blocking * nbw

      allocate(tmp1(max_local_cols*cwy_blocking))
      allocate(tmp2(max_local_cols*cwy_blocking))
      allocate(hvb(max_local_rows*cwy_blocking))
      allocate(hvm(max_local_rows,cwy_blocking))
      allocate(tmat_complete(cwy_blocking,cwy_blocking))
      allocate(t_tmp(cwy_blocking,nbw))
      allocate(t_tmp2(cwy_blocking,nbw))
!      else
!        allocate(tmp1(max_local_cols*nbw))
!        allocate(tmp2(max_local_cols*nbw))
!        allocate(hvb(max_local_rows*nbw))
!        allocate(hvm(max_local_rows,nbw))
!      endif

      hvm = 0   ! Must be set to 0 !!!
      hvb = 0   ! Safety only

      l_cols = local_index(nqc, my_pcol, np_cols, nblk, -1) ! Local columns of q

!     if ( na >= ((t_blocking+1)*nbw) ) then

      do istep=1,((na-1)/nbw-1)/t_blocking + 1
        ! This the call when using  na >= ((t_blocking+1)*nbw)
        !      n_cols = MIN(na,istep*cwy_blocking+nbw) - (istep-1)*cwy_blocking - nbw ! Number of columns in current step
        ! As an alternative we add some special case handling if na < cwy_blocking
        IF (na < cwy_blocking) THEN
          n_cols = MAX(0, na-nbw)
          IF ( n_cols .eq. 0 ) THEN
            EXIT
          END IF
        ELSE
          n_cols = MIN(na,istep*cwy_blocking+nbw) - (istep-1)*cwy_blocking - nbw ! Number of columns in current step
        END IF

        ! Broadcast all Householder vectors for current step compressed in hvb

        nb = 0
        ns = 0

        do lc = 1, n_cols
          ncol = (istep-1)*cwy_blocking + nbw + lc ! absolute column number of householder vector
          nrow = ncol - nbw ! absolute number of pivot row

          l_rows = local_index(nrow-1, my_prow, np_rows, nblk, -1) ! row length for bcast
          l_colh = local_index(ncol  , my_pcol, np_cols, nblk, -1) ! HV local column number

          if (my_pcol==pcol(ncol, nblk, np_cols)) hvb(nb+1:nb+l_rows) = a(1:l_rows,l_colh)

          nb = nb+l_rows

          if (lc==n_cols .or. mod(ncol,nblk)==0) then
818
#ifdef WITH_MPI
819
            call MPI_Bcast(hvb(ns+1),nb-ns,MPI_REAL8,pcol(ncol, nblk, np_cols),mpi_comm_cols,mpierr)
820
#endif
821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849
            ns = nb
          endif
        enddo

        ! Expand compressed Householder vectors into matrix hvm

        nb = 0
        do lc = 1, n_cols
          nrow = (istep-1)*cwy_blocking + lc ! absolute number of pivot row
          l_rows = local_index(nrow-1, my_prow, np_rows, nblk, -1) ! row length for bcast

          hvm(1:l_rows,lc) = hvb(nb+1:nb+l_rows)
          if (my_prow==prow(nrow, nblk, np_rows)) hvm(l_rows+1,lc) = 1.

          nb = nb+l_rows
        enddo

        l_rows = local_index(MIN(na,(istep+1)*cwy_blocking), my_prow, np_rows, nblk, -1)

        ! compute tmat2 out of tmat(:,:,)
        tmat_complete = 0
        do i = 1, t_blocking
          t_cols = MIN(nbw, n_cols - (i-1)*nbw)
          if (t_cols <= 0) exit
          t_rows = (i - 1) * nbw
          tmat_complete(t_rows+1:t_rows+t_cols,t_rows+1:t_rows+t_cols) = tmat(1:t_cols,1:t_cols,(istep-1)*t_blocking + i)
          if (i > 1) then
            call dgemm('T', 'N', t_rows, t_cols, l_rows, 1.d0, hvm(1,1), max_local_rows, hvm(1,(i-1)*nbw+1), &
                      max_local_rows, 0.d0, t_tmp, cwy_blocking)
850
#ifdef WITH_MPI
851
            call mpi_allreduce(t_tmp,t_tmp2,cwy_blocking*nbw,MPI_REAL8,MPI_SUM,mpi_comm_rows,mpierr)
852 853 854
#else
            t_tmp2 = t_tmp
#endif
855 856 857 858 859 860 861 862 863 864 865 866 867 868
            call dtrmm('L','U','N','N',t_rows,t_cols,1.0d0,tmat_complete,cwy_blocking,t_tmp2,cwy_blocking)
            call dtrmm('R','U','N','N',t_rows,t_cols,-1.0d0,tmat_complete(t_rows+1,t_rows+1),cwy_blocking,t_tmp2,cwy_blocking)
            tmat_complete(1:t_rows,t_rows+1:t_rows+t_cols) = t_tmp2(1:t_rows,1:t_cols)
          endif
        enddo

        ! Q = Q - V * T**T * V**T * Q

        if (l_rows>0) then
          call dgemm('T','N',n_cols,l_cols,l_rows,1.d0,hvm,ubound(hvm,dim=1), &
                     q,ldq,0.d0,tmp1,n_cols)
        else
          tmp1(1:l_cols*n_cols) = 0
        endif
869
#ifdef WITH_MPI
870
        call mpi_allreduce(tmp1,tmp2,n_cols*l_cols,MPI_REAL8,MPI_SUM,mpi_comm_rows,mpierr)
871 872 873
#else
        tmp2=tmp1
#endif
874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983

        if (l_rows>0) then
          call dtrmm('L','U','T','N',n_cols,l_cols,1.0d0,tmat_complete,cwy_blocking,tmp2,n_cols)
          call dgemm('N','N',l_rows,l_cols,n_cols,-1.d0,hvm,ubound(hvm,dim=1), tmp2,n_cols,1.d0,q,ldq)
        endif
      enddo

!   else 
!
!     do istep=1,(na-1)/nbw
!
!       n_cols = MIN(na,(istep+1)*nbw) - istep*nbw ! Number of columns in current step
!
!       ! Broadcast all Householder vectors for current step compressed in hvb
!
!       nb = 0
!       ns = 0
!
!       do lc = 1, n_cols
!         ncol = istep*nbw + lc ! absolute column number of householder vector
!         nrow = ncol - nbw ! absolute number of pivot row
!
!         l_rows = local_index(nrow-1, my_prow, np_rows, nblk, -1) ! row length for bcast
!         l_colh = local_index(ncol  , my_pcol, np_cols, nblk, -1) ! HV local column number
!
!         if (my_pcol==pcol(ncol, nblk, np_cols)) hvb(nb+1:nb+l_rows) = a(1:l_rows,l_colh)
!
!         nb = nb+l_rows
!
!         if (lc==n_cols .or. mod(ncol,nblk)==0) then
!           call MPI_Bcast(hvb(ns+1),nb-ns,MPI_REAL8,pcol(ncol, nblk, np_cols),mpi_comm_cols,mpierr)
!           ns = nb
!         endif
!       enddo
!
!       ! Expand compressed Householder vectors into matrix hvm
!
!       nb = 0
!       do lc = 1, n_cols
!         nrow = (istep-1)*nbw+lc ! absolute number of pivot row
!         l_rows = local_index(nrow-1, my_prow, np_rows, nblk, -1) ! row length for bcast
!
!         hvm(1:l_rows,lc) = hvb(nb+1:nb+l_rows)
!         if (my_prow==prow(nrow, nblk, np_rows)) hvm(l_rows+1,lc) = 1.
!
!         nb = nb+l_rows
!       enddo
!
!       l_rows = local_index(MIN(na,(istep+1)*nbw), my_prow, np_rows, nblk, -1)
!
!       ! Q = Q - V * T**T * V**T * Q
!
!       if (l_rows>0) then
!         call dgemm('T','N',n_cols,l_cols,l_rows,1.d0,hvm,ubound(hvm,dim=1), &
!                    q,ldq,0.d0,tmp1,n_cols)
!       else
!         tmp1(1:l_cols*n_cols) = 0
!       endif
!
!       call mpi_allreduce(tmp1,tmp2,n_cols*l_cols,MPI_REAL8,MPI_SUM,mpi_comm_rows,mpierr)
!
!       if (l_rows>0) then
!         call dtrmm('L','U','T','N',n_cols,l_cols,1.0d0,tmat(1,1,istep),ubound(tmat,dim=1),tmp2,n_cols)
!         call dgemm('N','N',l_rows,l_cols,n_cols,-1.d0,hvm,ubound(hvm,dim=1), &
!                    tmp2,n_cols,1.d0,q,ldq)
!       endif
!     enddo
!   endif

      deallocate(tmp1, tmp2, hvb, hvm)
!   if ( na >= ((t_blocking+1)*nbw) ) then
      deallocate(tmat_complete, t_tmp, t_tmp2)
!   endif

#ifdef HAVE_DETAILED_TIMINGS
      call timer%stop("trans_ev_band_to_full_real")
#endif
    end subroutine trans_ev_band_to_full_real

    subroutine tridiag_band_real(na, nb, nblk, a, lda, d, e, matrixCols, hh_trans_real, &
                                 mpi_comm_rows, mpi_comm_cols, mpi_comm)

    !-------------------------------------------------------------------------------
    ! tridiag_band_real:
    ! Reduces a real symmetric band matrix to tridiagonal form
    !
    !  na          Order of matrix a
    !
    !  nb          Semi bandwith
    !
    !  nblk        blocksize of cyclic distribution, must be the same in both directions!
    !
    !  a(lda,matrixCols)    Distributed system matrix reduced to banded form in the upper diagonal
    !
    !  lda         Leading dimension of a
    !  matrixCols  local columns of matrix a
    !
    !  d(na)       Diagonal of tridiagonal matrix, set only on PE 0 (output)
    !
    !  e(na)       Subdiagonal of tridiagonal matrix, set only on PE 0 (output)
    !
    !  mpi_comm_rows
    !  mpi_comm_cols
    !              MPI-Communicators for rows/columns
    !  mpi_comm
    !              MPI-Communicator for the total processor set
    !-------------------------------------------------------------------------------
#ifdef HAVE_DETAILED_TIMINGS
      use timings
#endif
Andreas Marek's avatar
Andreas Marek committed
984
      use precision
985 986
      implicit none

Andreas Marek's avatar
Andreas Marek committed
987
      integer(kind=ik), intent(in)  ::  na, nb, nblk, lda, matrixCols, mpi_comm_rows, mpi_comm_cols, mpi_comm
Andreas Marek's avatar
Andreas Marek committed
988 989 990
#ifdef DESPERATELY_WANT_ASSUMED_SIZE
      real(kind=rk), intent(in)     :: a(lda,*)
#else
Andreas Marek's avatar
Andreas Marek committed
991
      real(kind=rk), intent(in)     :: a(lda,matrixCols)
Andreas Marek's avatar
Andreas Marek committed
992
#endif
Andreas Marek's avatar
Andreas Marek committed
993 994 995
      real(kind=rk), intent(out)    :: d(na), e(na) ! set only on PE 0
      real(kind=rk), intent(out), &
          allocatable               :: hh_trans_real(:,:)
996

Andreas Marek's avatar
Andreas Marek committed
997 998
      real(kind=rk)                 :: vnorm2, hv(nb), tau, x, h(nb), ab_s(1+nb), hv_s(nb), hv_new(nb), tau_new, hf
      real(kind=rk)                 :: hd(nb), hs(nb)
999

Andreas Marek's avatar
Andreas Marek committed
1000 1001 1002 1003 1004
      integer(kind=ik)              :: i, j, n, nc, nr, ns, ne, istep, iblk, nblocks_total, nblocks, nt
      integer(kind=ik)              :: my_pe, n_pes, mpierr
      integer(kind=ik)              :: my_prow, np_rows, my_pcol, np_cols
      integer(kind=ik)              :: ireq_ab, ireq_hv
      integer(kind=ik)              :: na_s, nx, num_hh_vecs, num_chunks, local_size, max_blk_size, n_off
1005
#ifdef WITH_OPENMP
Andreas Marek's avatar
Andreas Marek committed
1006
      integer(kind=ik)              :: max_threads, my_thread, my_block_s, my_block_e, iter
1007
#ifdef WITH_MPI
Andreas Marek's avatar
Andreas Marek committed
1008
      integer(kind=ik)              :: mpi_status(MPI_STATUS_SIZE)
1009
#endif
Andreas Marek's avatar
Andreas Marek committed
1010 1011 1012
      integer(kind=ik), allocatable :: mpi_statuses(:,:), global_id_tmp(:,:)
      integer(kind=ik), allocatable :: omp_block_limits(:)
      real(kind=rk), allocatable    :: hv_t(:,:), tau_t(:)
1013
#endif
Andreas Marek's avatar
Andreas Marek committed
1014 1015 1016 1017
      integer(kind=ik), allocatable :: ireq_hhr(:), ireq_hhs(:), global_id(:,:), hh_cnt(:), hh_dst(:)
      integer(kind=ik), allocatable :: limits(:), snd_limits(:,:)
      integer(kind=ik), allocatable :: block_limits(:)
      real(kind=rk), allocatable    :: ab(:,:), hh_gath(:,:,:), hh_send(:,:,:)
1018 1019

#ifdef WITH_OPENMP
Andreas Marek's avatar
Andreas Marek committed
1020
      integer(kind=ik)              :: omp_get_max_threads
1021 1022
#endif

1023 1024 1025 1026
#ifndef WITH_MPI
      integer(kind=ik)             :: startAddr
#endif

1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045
#ifdef HAVE_DETAILED_TIMINGS
      call timer%start("tridiag_band_real")
#endif
      call mpi_comm_rank(mpi_comm,my_pe,mpierr)
      call mpi_comm_size(mpi_comm,n_pes,mpierr)

      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)
      ! Get global_id mapping 2D procssor coordinates to global id

      allocate(global_id(0:np_rows-1,0:np_cols-1))
      global_id(:,:) = 0
      global_id(my_prow, my_pcol) = my_pe
#ifdef WITH_OPENMP
      allocate(global_id_tmp(0:np_rows-1,0:np_cols-1))
#endif

1046 1047
#ifdef WITH_MPI

1048 1049 1050 1051 1052 1053 1054 1055
#ifndef WITH_OPENMP
      call mpi_allreduce(mpi_in_place, global_id, np_rows*np_cols, mpi_integer, mpi_sum, mpi_comm, mpierr)
#else
      global_id_tmp(:,:) = global_id(:,:)
      call mpi_allreduce(global_id_tmp, global_id, np_rows*np_cols, mpi_integer, mpi_sum, mpi_comm, mpierr)
      deallocate(global_id_tmp)
#endif

1056
#endif /* WITH_MPI */
1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119
      ! Total number of blocks in the band:

      nblocks_total = (na-1)/nb + 1

      ! Set work distribution

      allocate(block_limits(0:n_pes))
      call divide_band(nblocks_total, n_pes, block_limits)

      ! nblocks: the number of blocks for my task
      nblocks = block_limits(my_pe+1) - block_limits(my_pe)

      ! allocate the part of the band matrix which is needed by this PE
      ! The size is 1 block larger than needed to avoid extensive shifts
      allocate(ab(2*nb,(nblocks+1)*nb))
      ab = 0 ! needed for lower half, the extra block should also be set to 0 for safety

      ! n_off: Offset of ab within band
      n_off = block_limits(my_pe)*nb

      ! Redistribute band in a to ab
      call redist_band_real(a, lda, na, nblk, nb, matrixCols, mpi_comm_rows, mpi_comm_cols, mpi_comm, ab)

      ! Calculate the workload for each sweep in the back transformation
      ! and the space requirements to hold the HH vectors

      allocate(limits(0:np_rows))
      call determine_workload(na, nb, np_rows, limits)
      max_blk_size = maxval(limits(1:np_rows) - limits(0:np_rows-1))

      num_hh_vecs = 0
      num_chunks  = 0
      nx = na
      do n = 1, nblocks_total
        call determine_workload(nx, nb, np_rows, limits)
        local_size = limits(my_prow+1) - limits(my_prow)
        ! add to number of householder vectors
        ! please note: for nx==1 the one and only HH vector is 0 and is neither calculated nor send below!
        if (mod(n-1,np_cols) == my_pcol .and. local_size>0 .and. nx>1) then
          num_hh_vecs = num_hh_vecs + local_size
          num_chunks  = num_chunks+1
        endif
        nx = nx - nb
      enddo

      ! Allocate space for HH vectors

      allocate(hh_trans_real(nb,num_hh_vecs))

      ! Allocate and init MPI requests

      allocate(ireq_hhr(num_chunks)) ! Recv requests
      allocate(ireq_hhs(nblocks))    ! Send requests

      num_hh_vecs = 0
      num_chunks  = 0
      nx = na
      nt = 0
      do n = 1, nblocks_total
        call determine_workload(nx, nb, np_rows, limits)
        local_size = limits(my_prow+1) - limits(my_prow)
        if (mod(n-1,np_cols) == my_pcol .and. local_size>0 .and. nx>1) then
          num_chunks  = num_chunks+1
1120 1121
#ifdef WITH_MPI

1122 1123
          call mpi_irecv(hh_trans_real(1,num_hh_vecs+1), nb*local_size, mpi_real8, nt, &
                           10+n-block_limits(nt), mpi_comm, ireq_hhr(num_chunks), mpierr)
1124 1125 1126 1127
#else
          ! carefull non-block recv data copy must be done at wait or send
          ! hh_trans_real(1:nb*local_size,num_hh_vecs+1) = hh_send(1:nb*hh_cnt(iblk),1,iblk)
#endif
1128 1129 1130 1131 1132 1133 1134
          num_hh_vecs = num_hh_vecs + local_size
        endif
        nx = nx - nb
        if (n == block_limits(nt+1)) then
          nt = nt + 1
        endif
      enddo
1135
#ifdef WITH_MPI
1136
      ireq_hhs(:) = MPI_REQUEST_NULL
1137
#endif
1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151
      ! Buffers for gathering/sending the HH vectors

      allocate(hh_gath(nb,max_blk_size,nblocks)) ! gathers HH vectors
      allocate(hh_send(nb,max_blk_size,nblocks)) ! send buffer for HH vectors
      hh_gath(:,:,:) = 0
      hh_send(:,:,:) = 0

      ! Some counters

      allocate(hh_cnt(nblocks))
      allocate(hh_dst(nblocks))

      hh_cnt(:) = 1 ! The first transfomation vector is always 0 and not calculated at all
      hh_dst(:) = 0 ! PE number for receive
1152
#ifdef WITH_MPI
1153 1154
      ireq_ab = MPI_REQUEST_NULL
      ireq_hv = MPI_REQUEST_NULL
1155
#endif
1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192
      ! Limits for sending

      allocate(snd_limits(0:np_rows,nblocks))

      do iblk=1,nblocks
        call determine_workload(na-(iblk+block_limits(my_pe)-1)*nb, nb, np_rows, snd_limits(:,iblk))
      enddo

#ifdef WITH_OPENMP
      ! OpenMP work distribution:

      max_threads = 1
      max_threads = omp_get_max_threads()

      ! For OpenMP we need at least 2 blocks for every thread
      max_threads = MIN(max_threads, nblocks/2)
      if (max_threads==0) max_threads = 1

      allocate(omp_block_limits(0:max_threads))

      ! Get the OpenMP block limits
      call divide_band(nblocks, max_threads, omp_block_limits)

      allocate(hv_t(nb,max_threads), tau_t(max_threads))
      hv_t = 0
      tau_t = 0
#endif

      ! ---------------------------------------------------------------------------
      ! Start of calculations

      na_s = block_limits(my_pe)*nb + 1

      if (my_pe>0 .and. na_s<=na) then
        ! send first column to previous PE
        ! Only the PE owning the diagonal does that (sending 1 element of the subdiagonal block also)
        ab_s(1:nb+1) = ab(1:nb+1,na_s-n_off)
1193
#ifdef WITH_MPI
1194
        call mpi_isend(ab_s,nb+1,mpi_real8,my_pe-1,1,mpi_comm,ireq_ab,mpierr)
1195
#endif
1196 1197
      endif

1198 1199 1200 1201 1202

#ifndef WITH_MPI
          startAddr   = ubound(hh_trans_real,dim=2)
#endif

1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230
#ifdef WITH_OPENMP
      do istep=1,na-1-block_limits(my_pe)*nb
#else
      do istep=1,na-1
#endif

        if (my_pe==0) then
          n = MIN(na-na_s,nb) ! number of rows to be reduced
          hv(:) = 0
          tau = 0
          ! The last step (istep=na-1) is only needed for sending the last HH vectors.
          ! We don't want the sign of the last element flipped (analogous to the other sweeps)
          if (istep < na-1) then
            ! Transform first column of remaining matrix
            vnorm2 = sum(ab(3:n+1,na_s-n_off)**2)
            call hh_transform_real(ab(2,na_s-n_off),vnorm2,hf,tau)
            hv(1) = 1
            hv(2:n) = ab(3:n+1,na_s-n_off)*hf
          endif
          d(istep) = ab(1,na_s-n_off)
          e(istep) = ab(2,na_s-n_off)
          if (istep == na-1) then
            d(na) = ab(1,na_s+1-n_off)
            e(na) = 0
          endif
        else
          if (na>na_s) then
            ! Receive Householder vector from previous task, from PE owning subdiagonal
1231 1232
#ifdef WITH_MPI

1233 1234 1235 1236 1237
#ifdef WITH_OPENMP
            call mpi_recv(hv,nb,mpi_real8,my_pe-1,2,mpi_comm,MPI_STATUS,mpierr)
#else
            call mpi_recv(hv,nb,mpi_real8,my_pe-1,2,mpi_comm,MPI_STATUS_IGNORE,mpierr)
#endif
1238 1239 1240 1241 1242 1243 1244 1245 1246 1247

#else /* WITH_MPI */

#ifdef WITH_OPENMP
            hv(1:nb) = hv_s(1:nb)
#else
            hv(1:nb) = hv_s(1:nb)
#endif

#endif /* WITH_MPI */
1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390
            tau = hv(1)
            hv(1) = 1.
          endif
        endif

        na_s = na_s+1
        if (na_s-n_off > nb) then
          ab(:,1:nblocks*nb) = ab(:,nb+1:(nblocks+1)*nb)
          ab(:,nblocks*nb+1:(nblocks+1)*nb) = 0
          n_off = n_off + nb
        endif


#ifdef WITH_OPENMP
        if (max_threads > 1) then

          ! Codepath for OpenMP

          ! Please note that in this case it is absolutely necessary to have at least 2 blocks per thread!
          ! Every thread is one reduction cycle behind its predecessor and thus starts one step later.
          ! This simulates the behaviour of the MPI tasks which also work after each other.
          ! The code would be considerably easier, if the MPI communication would be made within
          ! the parallel region - this is avoided here since this would require
          ! MPI_Init_thread(MPI_THREAD_MULTIPLE) at the start of the program.

          hv_t(:,1) = hv
          tau_t(1) = tau

          do iter = 1, 2

            ! iter=1 : work on first block
            ! iter=2 : work on remaining blocks
            ! This is done in 2 iterations so that we have a barrier in between:
            ! After the first iteration, it is guaranteed that the last row of the last block
            ! is completed by the next thread.
            ! After the first iteration it is also the place to exchange the last row
            ! with MPI calls
#ifdef HAVE_DETAILED_TIMINGS
            call timer%start("OpenMP parallel")
#endif

!$omp parallel do private(my_thread, my_block_s, my_block_e, iblk, ns, ne, hv, tau, &
!$omp&                    nc, nr, hs, hd, vnorm2, hf, x, h, i), schedule(static,1), num_threads(max_threads)
            do my_thread = 1, max_threads

              if (iter == 1) then
                my_block_s = omp_block_limits(my_thread-1) + 1
                my_block_e = my_block_s
              else
                my_block_s = omp_block_limits(my_thread-1) + 2
                my_block_e = omp_block_limits(my_thread)
              endif

              do iblk = my_block_s, my_block_e

                ns = na_s + (iblk-1)*nb - n_off - my_thread + 1 ! first column in block
                ne = ns+nb-1                    ! last column in block

                if (istep<my_thread .or. ns+n_off>na) exit

                hv = hv_t(:,my_thread)
                tau = tau_t(my_thread)

                ! Store Householder vector for back transformation

                hh_cnt(iblk) = hh_cnt(iblk) + 1

                hh_gath(1   ,hh_cnt(iblk),iblk) = tau
                hh_gath(2:nb,hh_cnt(iblk),iblk) = hv(2:nb)

                nc = MIN(na-ns-n_off+1,nb) ! number of columns in diagonal block
                nr = MIN(na-nb-ns-n_off+1,nb) ! rows in subdiagonal block (may be < 0!!!)
                                          ! Note that nr>=0 implies that diagonal block is full (nc==nb)!

                ! Transform diagonal block

                call DSYMV('L',nc,tau,ab(1,ns),2*nb-1,hv,1,0.d0,hd,1)

                x = dot_product(hv(1:nc),hd(1:nc))*tau
                hd(1:nc) = hd(1:nc) - 0.5*x*hv(1:nc)

                call DSYR2('L',nc,-1.d0,hd,1,hv,1,ab(1,ns),2*nb-1)

                hv_t(:,my_thread) = 0
                tau_t(my_thread)  = 0

                if (nr<=0) cycle ! No subdiagonal block present any more

                ! Transform subdiagonal block

                call DGEMV('N',nr,nb,tau,ab(nb+1,ns),2*nb-1,hv,1,0.d0,hs,1)

                if (nr>1) then

                  ! complete (old) Householder transformation for first column

                  ab(nb+1:nb+nr,ns) = ab(nb+1:nb+nr,ns) - hs(1:nr) ! Note: hv(1) == 1

                  ! calculate new Householder transformation for first column
                  ! (stored in hv_t(:,my_thread) and tau_t(my_thread))

                  vnorm2 = sum(ab(nb+2:nb+nr,ns)**2)
                  call hh_transform_real(ab(nb+1,ns),vnorm2,hf,tau_t(my_thread))
                  hv_t(1   ,my_thread) = 1.
                  hv_t(2:nr,my_thread) = ab(nb+2:nb+nr,ns)*hf
                  ab(nb+2:,ns) = 0

                  ! update subdiagonal block for old and new Householder transformation
                  ! This way we can use a nonsymmetric rank 2 update which is (hopefully) faster

                  call DGEMV('T',nr,nb-1,tau_t(my_thread),ab(nb,ns+1),2*nb-1,hv_t(1,my_thread),1,0.d0,h(2),1)
                  x = dot_product(hs(1:nr),hv_t(1:nr,my_thread))*tau_t(my_thread)
                  h(2:nb) = h(2:nb) - x*hv(2:nb)
                  ! Unfortunately there is no BLAS routine like DSYR2 for a nonsymmetric rank 2 update ("DGER2")
                  do i=2,nb
                    ab(2+nb-i:1+nb+nr-i,i+ns-1) = ab(2+nb-i:1+nb+nr-i,i+ns-1) - hv_t(1:nr,my_thread)*h(i) - hs(1:nr)*hv(i)
                  enddo

                else

                  ! No new Householder transformation for nr=1, just complete the old one
                  ab(nb+1,ns) = ab(nb+1,ns) - hs(1) ! Note: hv(1) == 1
                  do i=2,nb
                    ab(2+nb-i,i+ns-1) = ab(2+nb-i,i+ns-1) - hs(1)*hv(i)
                  enddo
                  ! For safety: there is one remaining dummy transformation (but tau is 0 anyways)
                  hv_t(1,my_thread) = 1.

                endif

              enddo

            enddo ! my_thread
!$omp end parallel do
#ifdef HAVE_DETAILED_TIMINGS
            call timer%stop("OpenMP parallel")
#endif

            if (iter==1) then
              ! We are at the end of the first block

              ! Send our first column to previous PE
              if (my_pe>0 .and. na_s <= na) then
1391
#ifdef WITH_MPI
1392
                call mpi_wait(ireq_ab,mpi_status,mpierr)
1393
#endif
1394
                ab_s(1:nb+1) = ab(1:nb+1,na_s-n_off)
1395
#ifdef WITH_MPI