elpa1_compute_private.F90 107 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), formerly known as
!      Rechenzentrum Garching der Max-Planck-Gesellschaft (RZG),
8 9 10 11 12
!    - Bergische Universität Wuppertal, Lehrstuhl für angewandte
!      Informatik,
!    - Technische Universität München, Lehrstuhl für Informatik mit
!      Schwerpunkt Wissenschaftliches Rechnen ,
!    - Fritz-Haber-Institut, Berlin, Abt. Theorie,
Andreas Marek's avatar
Andreas Marek committed
13
!    - Max-Plack-Institut für Mathematik in den Naturwissenschaften,
14 15 16 17 18 19 20 21 22
!      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
!
!    ELPA is free software: you can redistribute it and/or modify
!    it under the terms of the version 3 of the license of the
!    GNU Lesser General Public License as published by the Free
!    Software Foundation.
!
!    ELPA is distributed in the hope that it will be useful,
!    but WITHOUT ANY WARRANTY; without even the implied warranty of
!    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
!    GNU Lesser General Public License for more details.
!
!    You should have received a copy of the GNU Lesser General Public License
!    along with ELPA.  If not, see <http://www.gnu.org/licenses/>
!
!    ELPA reflects a substantial effort on the part of the original
!    ELPA consortium, and we ask you to respect the spirit of the
!    license that we chose: i.e., please contribute any changes you
!    may have back to the original ELPA library distribution, and keep
!    any derivatives of ELPA under the same license that we chose for
!    the original distribution, the GNU Lesser General Public License.
!
!
! ELPA1 -- Faster replacements for ScaLAPACK symmetric eigenvalue routines
!
! Copyright of the original code rests with the authors inside the ELPA
! consortium. The copyright of any additional modifications shall rest
! with their original authors, but shall adhere to the licensing terms
! distributed along with the original code in the file "COPYING".

#include "config-f90.h"

module ELPA1_compute
  use elpa_utilities
#ifdef HAVE_DETAILED_TIMINGS
  use timings
#endif
60
  use elpa_mpi
61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83
  implicit none

  PRIVATE ! set default to private

  public :: tridiag_real               ! Transform real symmetric matrix to tridiagonal form
  public :: trans_ev_real              ! Transform eigenvectors of a tridiagonal matrix back

  public :: tridiag_complex            ! Transform complex hermitian matrix to tridiagonal form
  public :: trans_ev_complex           ! Transform eigenvectors of a tridiagonal matrix back

  public :: solve_tridi                ! Solve tridiagonal eigensystem with divide and conquer method

  public :: local_index                ! Get local index of a block cyclic distributed matrix
  public :: least_common_multiple      ! Get least common multiple

  public :: hh_transform_real
  public :: hh_transform_complex

  public :: elpa_reduce_add_vectors_complex, elpa_reduce_add_vectors_real
  public :: elpa_transpose_vectors_complex, elpa_transpose_vectors_real

  contains

Andreas Marek's avatar
Andreas Marek committed
84
#define DATATYPE REAL(kind=rk)
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
#define BYTESIZE 8
#define REALCASE 1
#include "elpa_transpose_vectors.X90"
#include "elpa_reduce_add_vectors.X90"
#undef DATATYPE
#undef BYTESIZE
#undef REALCASE

    subroutine tridiag_real(na, a, lda, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols, d, e, tau)

    !-------------------------------------------------------------------------------
    !  tridiag_real: Reduces a distributed symmetric matrix to tridiagonal form
    !                (like Scalapack Routine PDSYTRD)
    !
    !  Parameters
    !
    !  na          Order of matrix
    !
    !  a(lda,matrixCols)    Distributed matrix which should be reduced.
    !              Distribution is like in Scalapack.
    !              Opposed to PDSYTRD, a(:,:) must be set completely (upper and lower half)
    !              a(:,:) is overwritten on exit with the Householder vectors
    !
    !  lda         Leading dimension of a
    !  matrixCols  local columns of matrix
    !
    !  nblk        blocksize of cyclic distribution, must be the same in both directions!
    !
    !  mpi_comm_rows
    !  mpi_comm_cols
    !              MPI-Communicators for rows/columns
    !
    !  d(na)       Diagonal elements (returned), identical on all processors
    !
    !  e(na)       Off-Diagonal elements (returned), identical on all processors
    !
    !  tau(na)     Factors for the Householder vectors (returned), needed for back transformation
    !
    !-------------------------------------------------------------------------------
#ifdef HAVE_DETAILED_TIMINGS
      use timings
#endif
Andreas Marek's avatar
Andreas Marek committed
127
      use precision
128 129
      implicit none

Andreas Marek's avatar
Andreas Marek committed
130
      integer(kind=ik)            :: na, lda, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols
Andreas Marek's avatar
Andreas Marek committed
131
      real(kind=rk)               :: d(na), e(na), tau(na)
132
#ifdef USE_ASSUMED_SIZE
Andreas Marek's avatar
Andreas Marek committed
133 134 135 136
      real(kind=rk)               :: a(lda,*)
#else
      real(kind=rk)               :: a(lda,matrixCols)
#endif
137

Andreas Marek's avatar
Andreas Marek committed
138
      integer(kind=ik), parameter :: max_stored_rows = 32
139

Andreas Marek's avatar
Andreas Marek committed
140 141 142 143 144
      integer(kind=ik)            :: my_prow, my_pcol, np_rows, np_cols, mpierr
      integer(kind=ik)            :: totalblocks, max_blocks_row, max_blocks_col, max_local_rows, max_local_cols
      integer(kind=ik)            :: l_cols, l_rows, nstor
      integer(kind=ik)            :: istep, i, j, lcs, lce, lrs, lre
      integer(kind=ik)            :: tile_size, l_rows_tile, l_cols_tile
145 146

#ifdef WITH_OPENMP
Andreas Marek's avatar
Andreas Marek committed
147 148
      integer(kind=ik)            :: my_thread, n_threads, max_threads, n_iter
      integer(kind=ik)            :: omp_get_thread_num, omp_get_num_threads, omp_get_max_threads
149 150
#endif

Andreas Marek's avatar
Andreas Marek committed
151
      real(kind=rk)               :: vav, vnorm2, x, aux(2*max_stored_rows), aux1(2), aux2(2), vrl, xf
152

Andreas Marek's avatar
Andreas Marek committed
153
      real(kind=rk), allocatable  :: tmp(:), vr(:), vc(:), ur(:), uc(:), vur(:,:), uvc(:,:)
154
#ifdef WITH_OPENMP
Andreas Marek's avatar
Andreas Marek committed
155
      real(kind=rk), allocatable  :: ur_p(:,:), uc_p(:,:)
156
#endif
157 158
      integer(kind=ik)            :: istat
      character(200)              :: errorMessage
159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183

#ifdef HAVE_DETAILED_TIMINGS
      call timer%start("tridiag_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)
      ! 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


      totalblocks = (na-1)/nblk + 1
      max_blocks_row = (totalblocks-1)/np_rows + 1
      max_blocks_col = (totalblocks-1)/np_cols + 1

      max_local_rows = max_blocks_row*nblk
      max_local_cols = max_blocks_col*nblk

184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212
      allocate(tmp(MAX(max_local_rows,max_local_cols)), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
        print *,"tridiag_real: error when allocating tmp "//errorMessage
        stop
      endif

      allocate(vr(max_local_rows+1), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
        print *,"tridiag_real: error when allocating vr "//errorMessage
        stop
      endif

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

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

      allocate(uc(max_local_cols), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
        print *,"tridiag_real: error when allocating uc "//errorMessage
        stop
      endif
213 214 215 216

#ifdef WITH_OPENMP
      max_threads = omp_get_max_threads()

217 218 219 220 221 222 223 224 225 226 227 228
      allocate(ur_p(max_local_rows,0:max_threads-1), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
        print *,"tridiag_real: error when allocating ur_p "//errorMessage
        stop
      endif

      allocate(uc_p(max_local_cols,0:max_threads-1), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
        print *,"tridiag_real: error when allocating uc_p "//errorMessage
        stop
      endif

229 230 231 232 233 234 235 236
#endif

      tmp = 0
      vr = 0
      ur = 0
      vc = 0
      uc = 0

237 238 239 240 241 242 243 244 245 246 247
      allocate(vur(max_local_rows,2*max_stored_rows), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
        print *,"tridiag_real: error when allocating vur "//errorMessage
        stop
      endif

      allocate(uvc(max_local_cols,2*max_stored_rows), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
        print *,"tridiag_real: error when allocating uvc "//errorMessage
        stop
      endif
248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288

      d(:) = 0
      e(:) = 0
      tau(:) = 0

      nstor = 0

      l_rows = local_index(na, my_prow, np_rows, nblk, -1) ! Local rows of a
      l_cols = local_index(na, my_pcol, np_cols, nblk, -1) ! Local cols of a
      if(my_prow==prow(na, nblk, np_rows) .and. my_pcol==pcol(na, nblk, np_cols)) d(na) = a(l_rows,l_cols)

      do istep=na,3,-1

         ! Calculate number of local rows and columns of the still remaining matrix
         ! on the local processor

         l_rows = local_index(istep-1, my_prow, np_rows, nblk, -1)
         l_cols = local_index(istep-1, my_pcol, np_cols, nblk, -1)

         ! Calculate vector for Householder transformation on all procs
         ! owning column istep

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

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

            vr(1:l_rows) = a(1:l_rows,l_cols+1)
            if(nstor>0 .and. l_rows>0) then
               call DGEMV('N',l_rows,2*nstor,1.d0,vur,ubound(vur,dim=1), &
                          uvc(l_cols+1,1),ubound(uvc,dim=1),1.d0,vr,1)
            endif

            if(my_prow==prow(istep-1, nblk, np_rows)) then
               aux1(1) = dot_product(vr(1:l_rows-1),vr(1:l_rows-1))
               aux1(2) = vr(l_rows)
            else
               aux1(1) = dot_product(vr(1:l_rows),vr(1:l_rows))
               aux1(2) = 0.
            endif

289
#ifdef WITH_MPI
290
            call mpi_allreduce(aux1,aux2,2,MPI_REAL8,MPI_SUM,mpi_comm_rows,mpierr)
291 292 293
#else
            aux2 = aux1
#endif
294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315

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

            ! Householder transformation

            call hh_transform_real(vrl, vnorm2, xf, tau(istep))

            ! Scale vr and store Householder vector for back transformation

            vr(1:l_rows) = vr(1:l_rows) * xf
            if(my_prow==prow(istep-1, nblk, np_rows)) then
               vr(l_rows) = 1.
               e(istep-1) = vrl
            endif
            a(1:l_rows,l_cols+1) = vr(1:l_rows) ! store Householder vector for back transformation

         endif

         ! Broadcast the Householder vector (and tau) along columns

         if(my_pcol==pcol(istep, nblk, np_cols)) vr(l_rows+1) = tau(istep)
316
#ifdef WITH_MPI
317
         call MPI_Bcast(vr,l_rows+1,MPI_REAL8,pcol(istep, nblk, np_cols),mpi_comm_cols,mpierr)
318
#endif
319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406
         tau(istep) =  vr(l_rows+1)

         ! Transpose Householder vector vr -> vc

         call elpa_transpose_vectors_real  (vr, ubound(vr,dim=1), mpi_comm_rows, &
                                            vc, ubound(vc,dim=1), mpi_comm_cols, &
                                            1, istep-1, 1, nblk)


         ! Calculate u = (A + VU**T + UV**T)*v

         ! For cache efficiency, we use only the upper half of the matrix tiles for this,
         ! thus the result is partly in uc(:) and partly in ur(:)

         uc(1:l_cols) = 0
         ur(1:l_rows) = 0
         if (l_rows>0 .and. l_cols>0) then

#ifdef WITH_OPENMP

#ifdef HAVE_DETAILED_TIMINGS
           call timer%start("OpenMP parallel")
#endif

!$OMP PARALLEL PRIVATE(my_thread,n_threads,n_iter,i,lcs,lce,j,lrs,lre)

           my_thread = omp_get_thread_num()
           n_threads = omp_get_num_threads()

           n_iter = 0

           uc_p(1:l_cols,my_thread) = 0.
           ur_p(1:l_rows,my_thread) = 0.
#endif
           do i=0,(istep-2)/tile_size
             lcs = i*l_cols_tile+1
             lce = min(l_cols,(i+1)*l_cols_tile)
             if (lce<lcs) cycle
             do j=0,i
               lrs = j*l_rows_tile+1
               lre = min(l_rows,(j+1)*l_rows_tile)
               if (lre<lrs) cycle
#ifdef WITH_OPENMP
               if (mod(n_iter,n_threads) == my_thread) then
                 call DGEMV('T',lre-lrs+1,lce-lcs+1,1.d0,a(lrs,lcs),lda,vr(lrs),1,1.d0,uc_p(lcs,my_thread),1)
                 if (i/=j) call DGEMV('N',lre-lrs+1,lce-lcs+1,1.d0,a(lrs,lcs),lda,vc(lcs),1,1.d0,ur_p(lrs,my_thread),1)
               endif
               n_iter = n_iter+1
#else
               call DGEMV('T',lre-lrs+1,lce-lcs+1,1.d0,a(lrs,lcs),lda,vr(lrs),1,1.d0,uc(lcs),1)
               if (i/=j) call DGEMV('N',lre-lrs+1,lce-lcs+1,1.d0,a(lrs,lcs),lda,vc(lcs),1,1.d0,ur(lrs),1)

#endif
             enddo
           enddo
#ifdef WITH_OPENMP
!$OMP END PARALLEL
#ifdef HAVE_DETAILED_TIMINGS
           call timer%stop("OpenMP parallel")
#endif

           do i=0,max_threads-1
             uc(1:l_cols) = uc(1:l_cols) + uc_p(1:l_cols,i)
             ur(1:l_rows) = ur(1:l_rows) + ur_p(1:l_rows,i)
           enddo
#endif
           if (nstor>0) then
             call DGEMV('T',l_rows,2*nstor,1.d0,vur,ubound(vur,dim=1),vr,1,0.d0,aux,1)
             call DGEMV('N',l_cols,2*nstor,1.d0,uvc,ubound(uvc,dim=1),aux,1,1.d0,uc,1)
           endif

         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

        if (tile_size < istep-1) then
          call elpa_reduce_add_vectors_REAL  (ur, ubound(ur,dim=1), mpi_comm_rows, &
                                        uc, ubound(uc,dim=1), mpi_comm_cols, &
                                        istep-1, 1, nblk)
        endif

        ! Sum up all the uc(:) parts, transpose uc -> ur

        if (l_cols>0) then
          tmp(1:l_cols) = uc(1:l_cols)
407
#ifdef WITH_MPI
408
          call mpi_allreduce(tmp,uc,l_cols,MPI_REAL8,MPI_SUM,mpi_comm_rows,mpierr)
409 410 411
#else
          uc = tmp
#endif
412 413 414 415 416 417 418 419 420 421
        endif

        call elpa_transpose_vectors_real  (uc, ubound(uc,dim=1), mpi_comm_cols, &
                                         ur, ubound(ur,dim=1), mpi_comm_rows, &
                                         1, istep-1, 1, nblk)

        ! calculate u**T * v (same as v**T * (A + VU**T + UV**T) * v )

        x = 0
        if (l_cols>0) x = dot_product(vc(1:l_cols),uc(1:l_cols))
422
#ifdef WITH_MPI
423
        call mpi_allreduce(x,vav,1,MPI_REAL8,MPI_SUM,mpi_comm_cols,mpierr)
424 425 426
#else
        vav = x
#endif
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
        ! store u and v in the matrices U and V
        ! these matrices are stored combined in one here

        do j=1,l_rows
          vur(j,2*nstor+1) = tau(istep)*vr(j)
          vur(j,2*nstor+2) = 0.5*tau(istep)*vav*vr(j) - ur(j)
        enddo
        do j=1,l_cols
          uvc(j,2*nstor+1) = 0.5*tau(istep)*vav*vc(j) - uc(j)
          uvc(j,2*nstor+2) = tau(istep)*vc(j)
        enddo

        nstor = nstor+1

        ! If the limit of max_stored_rows is reached, calculate A + VU**T + UV**T

        if (nstor==max_stored_rows .or. istep==3) then

          do i=0,(istep-2)/tile_size
            lcs = i*l_cols_tile+1
            lce = min(l_cols,(i+1)*l_cols_tile)
           lrs = 1
            lre = min(l_rows,(i+1)*l_rows_tile)
            if (lce<lcs .or. lre<lrs) cycle
            call dgemm('N','T',lre-lrs+1,lce-lcs+1,2*nstor,1.d0, &
                       vur(lrs,1),ubound(vur,dim=1),uvc(lcs,1),ubound(uvc,dim=1), &
                       1.d0,a(lrs,lcs),lda)
          enddo

          nstor = 0

        endif

        if (my_prow==prow(istep-1, nblk, np_rows) .and. my_pcol==pcol(istep-1, nblk, np_cols)) then
          if (nstor>0) a(l_rows,l_cols) = a(l_rows,l_cols) &
                        + dot_product(vur(l_rows,1:2*nstor),uvc(l_cols,1:2*nstor))
          d(istep-1) = a(l_rows,l_cols)
        endif

      enddo

      ! Store e(1) and d(1)

      if (my_prow==prow(1, nblk, np_rows) .and. my_pcol==pcol(2, nblk, np_cols)) e(1) = a(1,l_cols) ! use last l_cols value of loop above
      if (my_prow==prow(1, nblk, np_rows) .and. my_pcol==pcol(1, nblk, np_cols)) d(1) = a(1,1)

473 474 475 476 477 478
      deallocate(tmp, vr, ur, vc, uc, vur, uvc, stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
        print *,"tridiag_real: error when deallocating uvc "//errorMessage
        stop
      endif

479 480 481

      ! distribute the arrays d and e to all processors

482 483 484 485 486
      allocate(tmp(na),  stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
        print *,"tridiag_real: error when allocating tmp "//errorMessage
        stop
      endif
487
#ifdef WITH_MPI
488 489 490 491 492 493 494 495
      tmp = d
      call mpi_allreduce(tmp,d,na,MPI_REAL8,MPI_SUM,mpi_comm_rows,mpierr)
      tmp = d
      call mpi_allreduce(tmp,d,na,MPI_REAL8,MPI_SUM,mpi_comm_cols,mpierr)
      tmp = e
      call mpi_allreduce(tmp,e,na,MPI_REAL8,MPI_SUM,mpi_comm_rows,mpierr)
      tmp = e
      call mpi_allreduce(tmp,e,na,MPI_REAL8,MPI_SUM,mpi_comm_cols,mpierr)
496
#endif
497 498 499 500 501 502
      deallocate(tmp,  stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
        print *,"tridiag_real: error when deallocating tmp "//errorMessage
        stop
      endif

503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545
#ifdef HAVE_DETAILED_TIMINGS
      call timer%stop("tridiag_real")
#endif

    end subroutine tridiag_real

    subroutine trans_ev_real(na, nqc, a, lda, tau, q, ldq, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols)

    !-------------------------------------------------------------------------------
    !  trans_ev_real: Transforms the eigenvectors of a tridiagonal matrix back
    !                 to the eigenvectors of the original matrix
    !                 (like Scalapack Routine PDORMTR)
    !
    !  Parameters
    !
    !  na          Order of matrix a, number of rows of matrix q
    !
    !  nqc         Number of columns of matrix q
    !
    !  a(lda,matrixCols)    Matrix containing the Householder vectors (i.e. matrix a after tridiag_real)
    !              Distribution is like in Scalapack.
    !
    !  lda         Leading dimension of a
    !  matrixCols  local columns of matrix a and q
    !
    !  tau(na)     Factors of the Householder vectors
    !
    !  q           On input: Eigenvectors of tridiagonal matrix
    !              On output: Transformed eigenvectors
    !              Distribution is like in Scalapack.
    !
    !  ldq         Leading dimension of q
    !
    !  nblk        blocksize of cyclic distribution, must be the same in both directions!
    !
    !  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
546
      use precision
547 548
      implicit none

Andreas Marek's avatar
Andreas Marek committed
549
      integer(kind=ik)           :: na, nqc, lda, ldq, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols
Andreas Marek's avatar
Andreas Marek committed
550
      real(kind=rk)              :: tau(na)
551
#ifdef USE_ASSUMED_SIZE
Andreas Marek's avatar
Andreas Marek committed
552 553 554 555
      real(kind=rk)              :: a(lda,*), q(ldq,*)
#else
      real(kind=rk)              :: a(lda,matrixCols), q(ldq,matrixCols)
#endif
556

Andreas Marek's avatar
Andreas Marek committed
557
      integer(kind=ik)           :: max_stored_rows
558

Andreas Marek's avatar
Andreas Marek committed
559 560 561 562
      integer(kind=ik)           :: my_prow, my_pcol, np_rows, np_cols, mpierr
      integer(kind=ik)           :: totalblocks, max_blocks_row, max_blocks_col, max_local_rows, max_local_cols
      integer(kind=ik)           :: l_cols, l_rows, l_colh, nstor
      integer(kind=ik)           :: istep, i, n, nc, ic, ics, ice, nb, cur_pcol
563

Andreas Marek's avatar
Andreas Marek committed
564 565
      real(kind=rk), allocatable :: tmp1(:), tmp2(:), hvb(:), hvm(:,:)
      real(kind=rk), allocatable :: tmat(:,:), h1(:), h2(:)
566 567
      integer(kind=ik)           :: istat
      character(200)             :: errorMessage
568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585
#ifdef HAVE_DETAILED_TIMINGS
      call timer%start("trans_ev_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)

      totalblocks = (na-1)/nblk + 1
      max_blocks_row = (totalblocks-1)/np_rows + 1
      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

      max_stored_rows = (63/nblk+1)*nblk

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

      allocate(h1(max_stored_rows*max_stored_rows), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
        print *,"trans_ev_real: error when allocating h1 "//errorMessage
        stop
      endif

      allocate(h2(max_stored_rows*max_stored_rows), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
        print *,"trans_ev_real: error when allocating h2 "//errorMessage
        stop
      endif

      allocate(tmp1(max_local_cols*max_stored_rows), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
        print *,"trans_ev_real: error when allocating tmp1 "//errorMessage
        stop
      endif

      allocate(tmp2(max_local_cols*max_stored_rows), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
        print *,"trans_ev_real: error when allocating tmp2 "//errorMessage
        stop
      endif

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

      allocate(hvm(max_local_rows,max_stored_rows), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
        print *,"trans_ev_real: error when allocating hvm "//errorMessage
        stop
      endif
627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659

      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

      nstor = 0

      do istep=1,na,nblk

        ics = MAX(istep,3)
        ice = MIN(istep+nblk-1,na)
        if (ice<ics) cycle

        cur_pcol = pcol(istep, nblk, np_cols)

        nb = 0
        do ic=ics,ice

          l_colh = local_index(ic  , my_pcol, np_cols, nblk, -1) ! Column of Householder vector
          l_rows = local_index(ic-1, my_prow, np_rows, nblk, -1) ! # rows of Householder vector


          if (my_pcol==cur_pcol) then
            hvb(nb+1:nb+l_rows) = a(1:l_rows,l_colh)
            if (my_prow==prow(ic-1, nblk, np_rows)) then
              hvb(nb+l_rows) = 1.
            endif
          endif

          nb = nb+l_rows
        enddo

660
#ifdef WITH_MPI
661 662
        if (nb>0) &
            call MPI_Bcast(hvb,nb,MPI_REAL8,cur_pcol,mpi_comm_cols,mpierr)
663
#endif
664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686
        nb = 0
        do ic=ics,ice
          l_rows = local_index(ic-1, my_prow, np_rows, nblk, -1) ! # rows of Householder vector
          hvm(1:l_rows,nstor+1) = hvb(nb+1:nb+l_rows)
          nstor = nstor+1
          nb = nb+l_rows
        enddo

        ! Please note: for smaller matix sizes (na/np_rows<=256), a value of 32 for nstor is enough!
        if (nstor+nblk>max_stored_rows .or. istep+nblk>na .or. (na/np_rows<=256 .and. nstor>=32)) then

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

          tmat = 0
          if (l_rows>0) &
               call dsyrk('U','T',nstor,l_rows,1.d0,hvm,ubound(hvm,dim=1),0.d0,tmat,max_stored_rows)

          nc = 0
          do n=1,nstor-1
            h1(nc+1:nc+n) = tmat(1:n,n+1)
            nc = nc+n
          enddo
687
#ifdef WITH_MPI
688
          if (nc>0) call mpi_allreduce(h1,h2,nc,MPI_REAL8,MPI_SUM,mpi_comm_rows,mpierr)
689 690 691
#else
          if (nc>0) h2 = h1
#endif
692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710
          ! Calculate triangular matrix T

          nc = 0
          tmat(1,1) = tau(ice-nstor+1)
          do n=1,nstor-1
            call dtrmv('L','T','N',n,tmat,max_stored_rows,h2(nc+1),1)
            tmat(n+1,1:n) = -h2(nc+1:nc+n)*tau(ice-nstor+n+1)
            tmat(n+1,n+1) = tau(ice-nstor+n+1)
            nc = nc+n
          enddo

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

          if (l_rows>0) then
            call dgemm('T','N',nstor,l_cols,l_rows,1.d0,hvm,ubound(hvm,dim=1), &
                          q,ldq,0.d0,tmp1,nstor)
          else
            tmp1(1:l_cols*nstor) = 0
          endif
711
#ifdef WITH_MPI
712
          call mpi_allreduce(tmp1,tmp2,nstor*l_cols,MPI_REAL8,MPI_SUM,mpi_comm_rows,mpierr)
713 714 715
#else
          tmp2 = tmp1
#endif
716 717 718 719 720 721 722 723 724 725
          if (l_rows>0) then
            call dtrmm('L','L','N','N',nstor,l_cols,1.0d0,tmat,max_stored_rows,tmp2,nstor)
            call dgemm('N','N',l_rows,l_cols,nstor,-1.d0,hvm,ubound(hvm,dim=1), &
                          tmp2,nstor,1.d0,q,ldq)
          endif
          nstor = 0
        endif

      enddo

726 727 728 729 730
      deallocate(tmat, h1, h2, tmp1, tmp2, hvb, hvm, stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
        print *,"trans_ev_real: error when deallocating hvm "//errorMessage
        stop
      endif
731 732 733 734 735 736 737

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

    end subroutine trans_ev_real

Andreas Marek's avatar
Andreas Marek committed
738
#define DATATYPE COMPLEX(kind=ck)
739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780
#define BYTESIZE 16
#define COMPLEXCASE 1
#include "elpa_transpose_vectors.X90"
#include "elpa_reduce_add_vectors.X90"
#undef DATATYPE
#undef BYTESIZE
#undef COMPLEXCASE

    subroutine tridiag_complex(na, a, lda, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols, d, e, tau)

    !-------------------------------------------------------------------------------
    !  tridiag_complex: Reduces a distributed hermitian matrix to tridiagonal form
    !                   (like Scalapack Routine PZHETRD)
    !
    !  Parameters
    !
    !  na          Order of matrix
    !
    !  a(lda,matrixCols)    Distributed matrix which should be reduced.
    !              Distribution is like in Scalapack.
    !              Opposed to PZHETRD, a(:,:) must be set completely (upper and lower half)
    !              a(:,:) is overwritten on exit with the Householder vectors
    !
    !  lda         Leading dimension of a
    !  matrixCols  local columns of matrix a
    !
    !  nblk        blocksize of cyclic distribution, must be the same in both directions!
    !
    !  mpi_comm_rows
    !  mpi_comm_cols
    !              MPI-Communicators for rows/columns
    !
    !  d(na)       Diagonal elements (returned), identical on all processors
    !
    !  e(na)       Off-Diagonal elements (returned), identical on all processors
    !
    !  tau(na)     Factors for the Householder vectors (returned), needed for back transformation
    !
    !-------------------------------------------------------------------------------
#ifdef HAVE_DETAILED_TIMINGS
      use timings
#endif
Andreas Marek's avatar
Andreas Marek committed
781
      use precision
782 783
      implicit none

Andreas Marek's avatar
Andreas Marek committed
784
      integer(kind=ik)              :: na, lda, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols
Andreas Marek's avatar
Andreas Marek committed
785
      complex(kind=ck)              :: tau(na)
786
#ifdef USE_ASSUMED_SIZE
Andreas Marek's avatar
Andreas Marek committed
787 788 789 790
      complex(kind=ck)              :: a(lda,*)
#else
      complex(kind=ck)              :: a(lda,matrixCols)
#endif
Andreas Marek's avatar
Andreas Marek committed
791
      real(kind=rk)                 :: d(na), e(na)
792

Andreas Marek's avatar
Andreas Marek committed
793
      integer(kind=ik), parameter   :: max_stored_rows = 32
794

Andreas Marek's avatar
Andreas Marek committed
795
      complex(kind=ck), parameter   :: CZERO = (0.d0,0.d0), CONE = (1.d0,0.d0)
796

Andreas Marek's avatar
Andreas Marek committed
797 798 799 800 801
      integer(kind=ik)              :: my_prow, my_pcol, np_rows, np_cols, mpierr
      integer(kind=ik)              :: totalblocks, max_blocks_row, max_blocks_col, max_local_rows, max_local_cols
      integer(kind=ik)              :: l_cols, l_rows, nstor
      integer(kind=ik)              :: istep, i, j, lcs, lce, lrs, lre
      integer(kind=ik)              :: tile_size, l_rows_tile, l_cols_tile
802 803

#ifdef WITH_OPENMP
Andreas Marek's avatar
Andreas Marek committed
804 805
      integer(kind=ik)              :: my_thread, n_threads, max_threads, n_iter
      integer(kind=ik)              :: omp_get_thread_num, omp_get_num_threads, omp_get_max_threads
806 807
#endif

Andreas Marek's avatar
Andreas Marek committed
808 809
      real(kind=rk)                 :: vnorm2
      complex(kind=ck)              :: vav, xc, aux(2*max_stored_rows),  aux1(2), aux2(2), vrl, xf
810

Andreas Marek's avatar
Andreas Marek committed
811
      complex(kind=ck), allocatable :: tmp(:), vr(:), vc(:), ur(:), uc(:), vur(:,:), uvc(:,:)
812
#ifdef WITH_OPENMP
Andreas Marek's avatar
Andreas Marek committed
813
      complex(kind=ck), allocatable :: ur_p(:,:), uc_p(:,:)
814
#endif
Andreas Marek's avatar
Andreas Marek committed
815
      real(kind=rk), allocatable    :: tmpr(:)
816 817
      integer(kind=ik)              :: istat
      character(200)                :: errorMessage
818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843

#ifdef HAVE_DETAILED_TIMINGS
      call timer%start("tridiag_complex")
#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)

      ! 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


      totalblocks = (na-1)/nblk + 1
      max_blocks_row = (totalblocks-1)/np_rows + 1
      max_blocks_col = (totalblocks-1)/np_cols + 1

      max_local_rows = max_blocks_row*nblk
      max_local_cols = max_blocks_col*nblk

844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872
      allocate(tmp(MAX(max_local_rows,max_local_cols)), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
       print *,"tridiag_complex: error when allocating tmp "//errorMessage
       stop
      endif

      allocate(vr(max_local_rows+1), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
       print *,"tridiag_complex: error when allocating vr "//errorMessage
       stop
      endif

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

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

      allocate(uc(max_local_cols), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
       print *,"tridiag_complex: error when allocating uc "//errorMessage
       stop
      endif
873 874 875 876

#ifdef WITH_OPENMP
      max_threads = omp_get_max_threads()

877 878 879 880 881 882 883 884 885 886 887
      allocate(ur_p(max_local_rows,0:max_threads-1), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
       print *,"tridiag_complex: error when allocating ur_p "//errorMessage
       stop
      endif

      allocate(uc_p(max_local_cols,0:max_threads-1), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
       print *,"tridiag_complex: error when allocating uc_p "//errorMessage
       stop
      endif
888 889 890 891 892 893 894 895
#endif

      tmp = 0
      vr = 0
      ur = 0
      vc = 0
      uc = 0

896 897 898 899 900 901 902 903 904 905 906
      allocate(vur(max_local_rows,2*max_stored_rows), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
       print *,"tridiag_complex: error when allocating vur "//errorMessage
       stop
      endif

      allocate(uvc(max_local_cols,2*max_stored_rows), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
       print *,"tridiag_complex: error when allocating uvc "//errorMessage
       stop
      endif
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

      d(:) = 0
      e(:) = 0
      tau(:) = 0

      nstor = 0

      l_rows = local_index(na, my_prow, np_rows, nblk, -1) ! Local rows of a
      l_cols = local_index(na, my_pcol, np_cols, nblk, -1) ! Local cols of a
      if (my_prow==prow(na, nblk, np_rows) .and. my_pcol==pcol(na, nblk, np_cols)) d(na) = a(l_rows,l_cols)

      do istep=na,3,-1

        ! Calculate number of local rows and columns of the still remaining matrix
        ! on the local processor

        l_rows = local_index(istep-1, my_prow, np_rows, nblk, -1)
        l_cols = local_index(istep-1, my_pcol, np_cols, nblk, -1)

        ! Calculate vector for Householder transformation on all procs
        ! owning column istep

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

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

          vr(1:l_rows) = a(1:l_rows,l_cols+1)
          if (nstor>0 .and. l_rows>0) then
            aux(1:2*nstor) = conjg(uvc(l_cols+1,1:2*nstor))
            call ZGEMV('N',l_rows,2*nstor,CONE,vur,ubound(vur,dim=1), &
                        aux,1,CONE,vr,1)
          endif

          if (my_prow==prow(istep-1, nblk, np_rows)) then
            aux1(1) = dot_product(vr(1:l_rows-1),vr(1:l_rows-1))
            aux1(2) = vr(l_rows)
          else
            aux1(1) = dot_product(vr(1:l_rows),vr(1:l_rows))
            aux1(2) = 0.
          endif
948
#ifdef WITH_MPI
949
          call mpi_allreduce(aux1,aux2,2,MPI_DOUBLE_COMPLEX,MPI_SUM,mpi_comm_rows,mpierr)
950 951 952
#else
          aux2 = aux1
#endif
953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973
          vnorm2 = aux2(1)
          vrl    = aux2(2)

          ! Householder transformation

          call hh_transform_complex(vrl, vnorm2, xf, tau(istep))

          ! Scale vr and store Householder vector for back transformation

          vr(1:l_rows) = vr(1:l_rows) * xf
          if (my_prow==prow(istep-1, nblk, np_rows)) then
            vr(l_rows) = 1.
            e(istep-1) = vrl
          endif
          a(1:l_rows,l_cols+1) = vr(1:l_rows) ! store Householder vector for back transformation

        endif

        ! Broadcast the Householder vector (and tau) along columns

        if (my_pcol==pcol(istep, nblk, np_cols)) vr(l_rows+1) = tau(istep)
974
#ifdef WITH_MPI
975
        call MPI_Bcast(vr,l_rows+1,MPI_DOUBLE_COMPLEX,pcol(istep, nblk, np_cols),mpi_comm_cols,mpierr)
976
#endif
977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068
        tau(istep) =  vr(l_rows+1)

        ! Transpose Householder vector vr -> vc

!        call elpa_transpose_vectors  (vr, 2*ubound(vr,dim=1), mpi_comm_rows, &
!                                      vc, 2*ubound(vc,dim=1), mpi_comm_cols, &
!                                      1, 2*(istep-1), 1, 2*nblk)

        call elpa_transpose_vectors_complex  (vr, ubound(vr,dim=1), mpi_comm_rows, &
                                              vc, ubound(vc,dim=1), mpi_comm_cols, &
                                              1, (istep-1), 1, nblk)
        ! Calculate u = (A + VU**T + UV**T)*v

        ! For cache efficiency, we use only the upper half of the matrix tiles for this,
        ! thus the result is partly in uc(:) and partly in ur(:)

        uc(1:l_cols) = 0
        ur(1:l_rows) = 0
        if (l_rows>0 .and. l_cols>0) then

#ifdef WITH_OPENMP

#ifdef HAVE_DETAILED_TIMINGS
          call timer%start("OpenMP parallel")
#endif

!$OMP PARALLEL PRIVATE(my_thread,n_threads,n_iter,i,lcs,lce,j,lrs,lre)

          my_thread = omp_get_thread_num()
          n_threads = omp_get_num_threads()

          n_iter = 0

          uc_p(1:l_cols,my_thread) = 0.
          ur_p(1:l_rows,my_thread) = 0.
#endif

          do i=0,(istep-2)/tile_size
            lcs = i*l_cols_tile+1
            lce = min(l_cols,(i+1)*l_cols_tile)
            if (lce<lcs) cycle
            do j=0,i
              lrs = j*l_rows_tile+1
              lre = min(l_rows,(j+1)*l_rows_tile)
              if (lre<lrs) cycle
#ifdef WITH_OPENMP
              if (mod(n_iter,n_threads) == my_thread) then
                call ZGEMV('C',lre-lrs+1,lce-lcs+1,CONE,a(lrs,lcs),lda,vr(lrs),1,CONE,uc_p(lcs,my_thread),1)
                if (i/=j) call ZGEMV('N',lre-lrs+1,lce-lcs+1,CONE,a(lrs,lcs),lda,vc(lcs),1,CONE,ur_p(lrs,my_thread),1)
              endif
              n_iter = n_iter+1
#else
              call ZGEMV('C',lre-lrs+1,lce-lcs+1,CONE,a(lrs,lcs),lda,vr(lrs),1,CONE,uc(lcs),1)
              if (i/=j) call ZGEMV('N',lre-lrs+1,lce-lcs+1,CONE,a(lrs,lcs),lda,vc(lcs),1,CONE,ur(lrs),1)
#endif
            enddo
          enddo

#ifdef WITH_OPENMP
!$OMP END PARALLEL
#ifdef HAVE_DETAILED_TIMINGS
          call timer%stop("OpenMP parallel")
#endif

          do i=0,max_threads-1
            uc(1:l_cols) = uc(1:l_cols) + uc_p(1:l_cols,i)
            ur(1:l_rows) = ur(1:l_rows) + ur_p(1:l_rows,i)
          enddo
#endif

          if (nstor>0) then
            call ZGEMV('C',l_rows,2*nstor,CONE,vur,ubound(vur,dim=1),vr,1,CZERO,aux,1)
            call ZGEMV('N',l_cols,2*nstor,CONE,uvc,ubound(uvc,dim=1),aux,1,CONE,uc,1)
          endif

        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

        if (tile_size < istep-1) then
          call elpa_reduce_add_vectors_COMPLEX  (ur, ubound(ur,dim=1), mpi_comm_rows, &
                                          uc, ubound(uc,dim=1), mpi_comm_cols, &
                                          (istep-1), 1, nblk)
        endif

        ! Sum up all the uc(:) parts, transpose uc -> ur

        if (l_cols>0) then
          tmp(1:l_cols) = uc(1:l_cols)
1069
#ifdef WITH_MPI
1070
          call mpi_allreduce(tmp,uc,l_cols,MPI_DOUBLE_COMPLEX,MPI_SUM,mpi_comm_rows,mpierr)
1071 1072 1073
#else
          uc = tmp
#endif
1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089
        endif

!        call elpa_transpose_vectors  (uc, 2*ubound(uc,dim=1), mpi_comm_cols, &
!                                      ur, 2*ubound(ur,dim=1), mpi_comm_rows, &
!                                      1, 2*(istep-1), 1, 2*nblk)

        call elpa_transpose_vectors_complex  (uc, ubound(uc,dim=1), mpi_comm_cols, &
                                              ur, ubound(ur,dim=1), mpi_comm_rows, &
                                              1, (istep-1), 1, nblk)



        ! calculate u**T * v (same as v**T * (A + VU**T + UV**T) * v )

        xc = 0
        if (l_cols>0) xc = dot_product(vc(1:l_cols),uc(1:l_cols))
1090
#ifdef WITH_MPI
1091
        call mpi_allreduce(xc,vav,1,MPI_DOUBLE_COMPLEX,MPI_SUM,mpi_comm_cols,mpierr)
1092 1093 1094
#else
        vav = xc
#endif
1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145
        ! store u and v in the matrices U and V
        ! these matrices are stored combined in one here

        do j=1,l_rows
          vur(j,2*nstor+1) = conjg(tau(istep))*vr(j)
          vur(j,2*nstor+2) = 0.5*conjg(tau(istep))*vav*vr(j) - ur(j)
        enddo
        do j=1,l_cols
          uvc(j,2*nstor+1) = 0.5*conjg(tau(istep))*vav*vc(j) - uc(j)
          uvc(j,2*nstor+2) = conjg(tau(istep))*vc(j)
        enddo

        nstor = nstor+1

        ! If the limit of max_stored_rows is reached, calculate A + VU**T + UV**T

        if (nstor==max_stored_rows .or. istep==3) then

          do i=0,(istep-2)/tile_size
            lcs = i*l_cols_tile+1
            lce = min(l_cols,(i+1)*l_cols_tile)
            lrs = 1
            lre = min(l_rows,(i+1)*l_rows_tile)
            if (lce<lcs .or. lre<lrs) cycle
            call ZGEMM('N','C',lre-lrs+1,lce-lcs+1,2*nstor,CONE, &
                         vur(lrs,1),ubound(vur,dim=1),uvc(lcs,1),ubound(uvc,dim=1), &
                         CONE,a(lrs,lcs),lda)
          enddo

          nstor = 0

        endif

        if (my_prow==prow(istep-1, nblk, np_rows) .and. my_pcol==pcol(istep-1, nblk, np_cols)) then
          if (nstor>0) a(l_rows,l_cols) = a(l_rows,l_cols) &
                          + dot_product(vur(l_rows,1:2*nstor),uvc(l_cols,1:2*nstor))
          d(istep-1) = a(l_rows,l_cols)
        endif

      enddo ! istep

      ! Store e(1) and d(1)

      if (my_pcol==pcol(2, nblk, np_cols)) then
        if (my_prow==prow(1, nblk, np_rows)) then
          ! We use last l_cols value of loop above
          vrl = a(1,l_cols)
          call hh_transform_complex(vrl, 0.d0, xf, tau(2))
          e(1) = vrl
          a(1,l_cols) = 1. ! for consistency only
        endif
1146
#ifdef WITH_MPI
1147
        call mpi_bcast(tau(2),1,MPI_DOUBLE_COMPLEX,prow(1, nblk, np_rows),mpi_comm_rows,mpierr)
1148
#endif
1149
      endif
1150
#ifdef WITH_MPI
1151
      call mpi_bcast(tau(2),1,MPI_DOUBLE_COMPLEX,pcol(2, nblk, np_cols),mpi_comm_cols,mpierr)
1152
#endif
1153 1154 1155

      if (my_prow==prow(1, nblk, np_rows) .and. my_pcol==pcol(1, nblk, np_cols)) d(1) = a(1,1)

1156 1157 1158 1159 1160
      deallocate(tmp, vr, ur, vc, uc, vur, uvc, stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
       print *,"tridiag_complex: error when deallocating tmp "//errorMessage
       stop
      endif
1161 1162
      ! distribute the arrays d and e to all processors

1163 1164 1165 1166 1167
      allocate(tmpr(na), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
       print *,"tridiag_complex: error when allocating tmpr "//errorMessage
       stop
      endif
1168
#ifdef WITH_MPI
1169 1170 1171 1172 1173 1174 1175 1176
      tmpr = d
      call mpi_allreduce(tmpr,d,na,MPI_REAL8,MPI_SUM,mpi_comm_rows,mpierr)
      tmpr = d
      call mpi_allreduce(tmpr,d,na,MPI_REAL8,MPI_SUM,mpi_comm_cols,mpierr)
      tmpr = e
      call mpi_allreduce(tmpr,e,na,MPI_REAL8,MPI_SUM,mpi_comm_rows,mpierr)
      tmpr = e
      call mpi_allreduce(tmpr,e,na,MPI_REAL8,MPI_SUM,mpi_comm_cols,mpierr)
1177
#endif
1178 1179 1180 1181 1182 1183
      deallocate(tmpr, stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
       print *,"tridiag_complex: error when deallocating tmpr "//errorMessage
       stop
      endif

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 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225
#ifdef HAVE_DETAILED_TIMINGS
      call timer%stop("tridiag_complex")
#endif

    end subroutine tridiag_complex

    subroutine trans_ev_complex(na, nqc, a, lda, tau, q, ldq, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols)

    !-------------------------------------------------------------------------------
    !  trans_ev_complex: Transforms the eigenvectors of a tridiagonal matrix back
    !                    to the eigenvectors of the original matrix
    !                    (like Scalapack Routine PZUNMTR)
    !
    !  Parameters
    !
    !  na          Order of matrix a, number of rows of matrix q
    !
    !  nqc         Number of columns of matrix q
    !
    !  a(lda,matrixCols)    Matrix containing the Householder vectors (i.e. matrix a after tridiag_complex)
    !              Distribution is like in Scalapack.
    !
    !  lda         Leading dimension of a
    !
    !  tau(na)     Factors of the Householder vectors
    !
    !  q           On input: Eigenvectors of tridiagonal matrix
    !              On output: Transformed eigenvectors
    !              Distribution is like in Scalapack.
    !
    !  ldq         Leading dimension of q
    !
    !  nblk        blocksize of cyclic distribution, must be the same in both directions!
    !
    !  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
1226
      use precision
1227 1228
      implicit none

Andreas Marek's avatar
Andreas Marek committed
1229
      integer(kind=ik)              ::  na, nqc, lda, ldq, nblk, matrixCols, mpi_comm_rows, mpi_comm_cols
Andreas Marek's avatar
Andreas Marek committed
1230
      complex(kind=ck)              ::  tau(na)
1231
#ifdef USE_ASSUMED_SIZE
Andreas Marek's avatar
Andreas Marek committed
1232 1233 1234 1235
      complex(kind=ck)              :: a(lda,*), q(ldq,*)
#else
      complex(kind=ck)              ::  a(lda,matrixCols), q(ldq,matrixCols)
#endif
Andreas Marek's avatar
Andreas Marek committed
1236
      integer(kind=ik)              :: max_stored_rows
1237

Andreas Marek's avatar
Andreas Marek committed
1238
      complex(kind=ck), parameter   :: CZERO = (0.d0,0.d0), CONE = (1.d0,0.d0)
1239

Andreas Marek's avatar
Andreas Marek committed
1240 1241 1242 1243
      integer(kind=ik)              :: my_prow, my_pcol, np_rows, np_cols, mpierr
      integer(kind=ik)              :: totalblocks, max_blocks_row, max_blocks_col, max_local_rows, max_local_cols
      integer(kind=ik)              :: l_cols, l_rows, l_colh, nstor
      integer(kind=ik)              :: istep, i, n, nc, ic, ics, ice, nb, cur_pcol
1244

Andreas Marek's avatar
Andreas Marek committed
1245 1246
      complex(kind=ck), allocatable :: tmp1(:), tmp2(:), hvb(:), hvm(:,:)
      complex(kind=ck), allocatable :: tmat(:,:), h1(:), h2(:)
1247 1248
      integer(kind=ik)              :: istat
      character(200)                :: errorMessage
1249 1250 1251
#ifdef HAVE_DETAILED_TIMINGS
      call timer%start("trans_ev_complex")
#endif
1252

1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266
      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)

      totalblocks = (na-1)/nblk + 1
      max_blocks_row = (totalblocks-1)/np_rows + 1
      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

      max_stored_rows = (63/nblk+1)*nblk

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

      allocate(h1(max_stored_rows*max_stored_rows), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
       print *,"trans_ev_complex: error when allocating h1 "//errorMessage
       stop
      endif

      allocate(h2(max_stored_rows*max_stored_rows), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
       print *,"trans_ev_complex: error when allocating h2 "//errorMessage
       stop
      endif

      allocate(tmp1(max_local_cols*max_stored_rows), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
       print *,"trans_ev_complex: error when allocating tmp1 "//errorMessage
       stop
      endif

      allocate(tmp2(max_local_cols*max_stored_rows), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
       print *,"trans_ev_complex: error when allocating tmp2 "//errorMessage
       stop
      endif

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

      allocate(hvm(max_local_rows,max_stored_rows), stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
       print *,"trans_ev_complex: error when allocating hvm "//errorMessage
       stop
      endif