elpa2_trans_ev_band_to_full_template.F90 35.6 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51
#if 0
!    This file is part of ELPA.
!
!    The ELPA library was originally created by the ELPA consortium,
!    consisting of the following organizations:
!
!    - Max Planck Computing and Data Facility (MPCDF), formerly known as
!      Rechenzentrum Garching der Max-Planck-Gesellschaft (RZG),
!    - Bergische Universität Wuppertal, Lehrstuhl für angewandte
!      Informatik,
!    - Technische Universität München, Lehrstuhl für Informatik mit
!      Schwerpunkt Wissenschaftliches Rechnen ,
!    - Fritz-Haber-Institut, Berlin, Abt. Theorie,
!    - Max-Plack-Institut für Mathematik in den Naturwissenschaften,
!      Leipzig, Abt. Komplexe Strukutren in Biologie und Kognition,
!      and
!    - IBM Deutschland GmbH
!
!    This particular source code file contains additions, changes and
!    enhancements authored by Intel Corporation which is not part of
!    the ELPA consortium.
!
!    More information can be found here:
!    http://elpa.mpcdf.mpg.de/
!
!    ELPA is free software: you can redistribute it and/or modify
!    it under the terms of the version 3 of the license of the
!    GNU Lesser General Public License as published by the Free
!    Software Foundation.
!
!    ELPA is distributed in the hope that it will be useful,
!    but WITHOUT ANY WARRANTY; without even the implied warranty of
!    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
!    GNU Lesser General Public License for more details.
!
!    You should have received a copy of the GNU Lesser General Public License
!    along with ELPA.  If not, see <http://www.gnu.org/licenses/>
!
!    ELPA reflects a substantial effort on the part of the original
!    ELPA consortium, and we ask you to respect the spirit of the
!    license that we chose: i.e., please contribute any changes you
!    may have back to the original ELPA library distribution, and keep
!    any derivatives of ELPA under the same license that we chose for
!    the original distribution, the GNU Lesser General Public License.
!
! Copyright of the original code rests with the authors inside the ELPA
! consortium. The copyright of any additional modifications shall rest
! with their original authors, but shall adhere to the licensing terms
! distributed along with the original code in the file "COPYING".
#endif

52
#include "../general/sanity.F90"
53 54 55 56 57

    subroutine trans_ev_band_to_full_&
    &MATH_DATATYPE&
    &_&
    &PRECISION &
58
    (obj, na, nqc, nblk, nbw, a, a_dev, lda, tmat, tmat_dev, q, &
Andreas Marek's avatar
Andreas Marek committed
59
     q_dev, ldq, matrixCols, numBlocks, mpi_comm_rows, mpi_comm_cols, useGPU &
60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102
#if REALCASE == 1
     ,useQr)
#endif
#if COMPLEXCASE == 1
     )
#endif

    !-------------------------------------------------------------------------------
    !  trans_ev_band_to_full_real/complex:
    !  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/complex)
    !              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/complex
    !
    !  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
    !
    !-------------------------------------------------------------------------------
      use precision
      use cuda_functions
      use iso_c_binding
103
      use elpa_abstract_impl
104 105
      implicit none

106
      class(elpa_abstract_impl_t), intent(inout) :: obj
107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168
      logical, intent(in)                    :: useGPU
#if REALCASE == 1
     logical, intent(in)                     :: useQR
#endif
      integer(kind=ik)                       :: na, nqc, lda, ldq, nblk, nbw, matrixCols, numBlocks, mpi_comm_rows, mpi_comm_cols
#if REALCASE == 1
#ifdef USE_ASSUMED_SIZE
      real(kind=REAL_DATATYPE)               :: a(lda,*), q(ldq,*), tmat(nbw,nbw,*)
      !real(kind=rk8)               :: a(lda,*), q(ldq,*), tmat(nbw,nbw,*)
#else
      real(kind=REAL_DATATYPE)               :: a(lda,matrixCols), q(ldq,matrixCols), tmat(nbw, nbw, numBlocks)
      !real(kind=rk8)               :: a(lda,matrixCols), q(ldq,matrixCols), tmat(nbw, nbw, numBlocks)
#endif
#endif
#if COMPLEXCASE == 1
#ifdef USE_ASSUMED_SIZE
      complex(kind=COMPLEX_DATATYPE)         :: a(lda,*), q(ldq,*), tmat(nbw,nbw,*)
      !complex(kind=ck8)         :: a(lda,*), q(ldq,*), tmat(nbw,nbw,*)
#else
      complex(kind=COMPLEX_DATATYPE)         :: a(lda,matrixCols), q(ldq,matrixCols), tmat(nbw, nbw, numBlocks)
      !complex(kind=ck8)         :: a(lda,matrixCols), q(ldq,matrixCols), tmat(nbw, nbw, numBlocks)
#endif
#endif

#if REALCASE == 1
#ifdef DOUBLE_PRECISION_COMPLEX
       real(kind=REAL_DATATYPE), parameter   :: ZERO = 0.0_rk8, ONE = 1.0_rk8
#else
       real(kind=REAL_DATATYPE), parameter   :: ZERO = 0.0_rk4, ONE = 1.0_rk4
#endif

#endif
#if COMPLEXCASE == 1
#ifdef DOUBLE_PRECISION_COMPLEX
       complex(kind=COMPLEX_DATATYPE), parameter   :: ZERO = (0.0_rk8,0.0_rk8), ONE = (1.0_rk8,0.0_rk8)
#else
       complex(kind=COMPLEX_DATATYPE), parameter   :: ZERO = (0.0_rk4,0.0_rk4), ONE = (1.0_rk4,0.0_rk4)
#endif
#endif

      integer(kind=C_intptr_T)               :: a_dev ! passed from bandred_real at the moment not used since copied in bandred_real

      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

#if REALCASE ==1
      real(kind=REAL_DATATYPE), allocatable  :: tmp1(:), tmp2(:), hvb(:), hvm(:,:)
#endif
#if COMPLEXCASE == 1
      complex(kind=COMPLEX_DATATYPE), allocatable :: tmp1(:), tmp2(:), hvb(:), hvm(:,:)
#endif
      ! hvm_dev is fist used and set in this routine
      ! q is changed in trans_ev_tridi on the host, copied to device and passed here. this can be adapted
      ! tmp_dev is first used in this routine
      ! tmat_dev is passed along from bandred_real
      integer(kind=C_intptr_T)               :: hvm_dev, q_dev, tmp_dev, tmat_dev

      integer(kind=ik)                       :: i

Andreas Marek's avatar
Andreas Marek committed
169
#ifdef BAND_TO_FULL_BLOCKING
170 171
#if REALCASE == 1
      real(kind=REAL_DATATYPE), allocatable  :: tmat_complete(:,:), t_tmp(:,:), t_tmp2(:,:)
Andreas Marek's avatar
Andreas Marek committed
172 173 174 175
#endif
#if COMPLEXCASE == 1
      complex(kind=COMPLEX_DATATYPE), allocatable  :: tmat_complete(:,:), t_tmp(:,:), t_tmp2(:,:)
#endif
176 177
      integer(kind=ik)                       :: cwy_blocking, t_blocking, t_cols, t_rows
#endif
Andreas Marek's avatar
Andreas Marek committed
178

179 180 181
      integer(kind=ik)                       :: istat
      character(200)                         :: errorMessage
      logical                                :: successCUDA
182
      integer(kind=c_intptr_t), parameter      :: size_of_datatype = size_of_&
183 184 185
                                                                   &PRECISION&
                                                                   &_&
                                                                   &MATH_DATATYPE
186
      call obj%timer%start("trans_ev_band_to_full_&
187
      &MATH_DATATYPE&
Andreas Marek's avatar
Andreas Marek committed
188
      &" // &
189 190
      &PRECISION_SUFFIX &
      )
191
      call obj%timer%start("mpi_communication")
192 193 194 195 196 197

      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)

198
      call obj%timer%stop("mpi_communication")
199 200 201 202 203 204 205 206 207 208 209 210 211 212

      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

      if (useGPU) then

#if REALCASE == 1
        ! here the GPU and CPU version diverged: the CPU version now always uses the useQR path which
        ! is not implemented in the GPU version
#endif

Andreas Marek's avatar
Andreas Marek committed
213 214 215
        ! the GPU version does not (yet) support blocking
        ! but the handling is the same for real/complex case

216 217 218
        allocate(tmp1(max_local_cols*nbw), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
219 220
    &MATH_DATATYPE&
    &: error when allocating tmp1 "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
221
          stop 1
222 223 224 225 226
        endif

        allocate(tmp2(max_local_cols*nbw), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
227 228
    &MATH_DATATYPE&
    &: error when allocating tmp2 "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
229
          stop 1
230 231 232 233 234
        endif

        allocate(hvb(max_local_rows*nbw), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
235 236
    &MATH_DATATYPE&
    &: error when allocating hvb "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
237
          stop 1
238 239 240 241 242
        endif

        allocate(hvm(max_local_rows,nbw), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
243 244
    &MATH_DATATYPE&
    &: error when allocating hvm "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
245
          stop 1
246 247
        endif

248
        successCUDA = cuda_malloc(hvm_dev, (max_local_rows)*nbw* size_of_datatype)
249 250
        if (.not.(successCUDA)) then
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
251 252
    &MATH_DATATYPE&
    &: error in cudaMalloc"
Andreas Marek's avatar
Andreas Marek committed
253
          stop 1
254 255
        endif

256
        successCUDA = cuda_malloc(tmp_dev, (max_local_cols)*nbw* size_of_datatype)
257 258
        if (.not.(successCUDA)) then
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
259 260
    &MATH_DATATYPE&
    &: error in cudaMalloc"
Andreas Marek's avatar
Andreas Marek committed
261
          stop 1
262 263 264
        endif

!#ifdef WITH_MPI
Andreas Marek's avatar
Andreas Marek committed
265 266 267 268
!! it should be possible to keep tmat dev on the device and not copy it around
!! already existent on GPU
!        successCUDA = cuda_malloc(tmat_dev, nbw*nbw* &
!#if REALCASE == 1
Andreas Marek's avatar
Retab  
Andreas Marek committed
269
!  size_of_PRECISION_real)
Andreas Marek's avatar
Andreas Marek committed
270 271 272 273 274 275 276
!#endif
!#if COMPLEXCASE == 1
!        size_of_PRECISION_complex)
!#endif
!
!        if (.not.(successCUDA)) then
!          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
277 278
!    &MATH_DATATYPE&
!    &: error in cudaMalloc"
Andreas Marek's avatar
Andreas Marek committed
279
!          stop 1
Andreas Marek's avatar
Andreas Marek committed
280
!        endif
281 282 283 284
!#endif

#if REALCASE == 1
! q_dev already living on device
285
!        successCUDA = cuda_malloc(q_dev, ldq*matrixCols*size_of_datatype)
286 287
!        if (.not.(successCUDA)) then
!          print *,"trans_ev_band_to_full_real: error in cudaMalloc"
Andreas Marek's avatar
Andreas Marek committed
288
!          stop 1
289 290 291 292 293 294 295 296
!        endif
  !      q_temp(:,:) = 0.0
  !      q_temp(1:ldq,1:na_cols) = q(1:ldq,1:na_cols)

!        ! copy q_dev to device, maybe this can be avoided if q_dev can be kept on device in trans_ev_tridi_to_band
!        successCUDA = cuda_memcpy(q_dev, loc(q), (ldq)*(matrixCols)*size_of_PRECISION_real, cudaMemcpyHostToDevice)
!        if (.not.(successCUDA)) then
!          print *,"trans_ev_band_to_full_real: error in cudaMalloc"
Andreas Marek's avatar
Andreas Marek committed
297
!          stop 1
298 299 300
!        endif
#endif
#if COMPLEXCASE == 1
Andreas Marek's avatar
Andreas Marek committed
301 302 303
!         successCUDA = cuda_malloc(q_dev, ldq*matrixCols*size_of_PRECISION_complex)
!         if (.not.(successCUDA)) then
!           print *,"trans_ev_band_to_full_complex: error in cudaMalloc"
Andreas Marek's avatar
Andreas Marek committed
304
!           stop 1
Andreas Marek's avatar
Andreas Marek committed
305 306 307 308 309
!         endif
!
!         successCUDA = cuda_memcpy(q_dev, loc(q),ldq*matrixCols*size_of_PRECISION_complex, cudaMemcpyHostToDevice)
!          if (.not.(successCUDA)) then
!            print *,"trans_ev_band_to_full_complex: error in cudaMemcpy"
Andreas Marek's avatar
Andreas Marek committed
310
!            stop 1
Andreas Marek's avatar
Andreas Marek committed
311
!          endif
312 313 314
#endif

        ! if MPI is NOT used the following steps could be done on the GPU and memory transfers could be avoided
315
        successCUDA = cuda_memset(hvm_dev, 0, (max_local_rows)*(nbw)* size_of_datatype)
316 317
        if (.not.(successCUDA)) then
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
318 319
    &MATH_DATATYPE&
    &: error in cudaMalloc"
Andreas Marek's avatar
Andreas Marek committed
320
          stop 1
321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342
        endif

#if REALCASE == 1
        hvm = CONST_0_0   ! Must be set to 0 !!!
        hvb = CONST_0_0   ! Safety only
#endif
#if COMPLEXCASE == 1
        hvm = CONST_COMPLEX_0_0   ! Must be set to 0 !!!
        hvb = CONST_COMPLEX_0_0   ! Safety only
#endif
        l_cols = local_index(nqc, my_pcol, np_cols, nblk, -1) ! Local columns of q

        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
343
            ncol = istep*nbw + lc ! absolute column number of householder Vector
344 345 346 347 348 349 350 351 352 353 354
            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
#ifdef WITH_MPI
355
              call obj%timer%start("mpi_communication")
356 357
              call MPI_Bcast(hvb(ns+1), nb-ns,  &
#if REALCASE == 1
Andreas Marek's avatar
Retab  
Andreas Marek committed
358
                       MPI_REAL_PRECISION,&
359 360 361 362
#endif
#if COMPLEXCASE == 1
                             MPI_COMPLEX_PRECISION, &
#endif
Andreas Marek's avatar
Retab  
Andreas Marek committed
363
           pcol(ncol, nblk, np_cols), mpi_comm_cols, mpierr)
364

365
              call obj%timer%stop("mpi_communication")
366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388

#endif /* WITH_MPI */
              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 REALCASE == 1
            if (my_prow==prow(nrow, nblk, np_rows)) hvm(l_rows+1,lc) = CONST_1_0
#endif
#if COMPLEXCASE == 1
            if (my_prow==prow(nrow, nblk, np_rows)) hvm(l_rows+1,lc) = 1.
#endif
            nb = nb+l_rows
          enddo

389
          successCUDA = cuda_memcpy(hvm_dev, loc(hvm), max_local_rows*nbw* size_of_datatype, cudaMemcpyHostToDevice)
390 391 392

          if (.not.(successCUDA)) then
            print *,"trans_ev_band_to_full_real: error in cudaMemcpy"
Andreas Marek's avatar
Andreas Marek committed
393
            stop 1
394 395 396 397 398 399 400 401

          endif

          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
Andreas Marek's avatar
Retab  
Andreas Marek committed
402
      call obj%timer%start("cublas")
403 404 405 406 407 408
#if REALCASE == 1
            call cublas_PRECISION_GEMM('T', 'N',        &
#endif
#if COMPLEXCASE == 1
            call cublas_PRECISION_GEMM('C', 'N',        &
#endif
Andreas Marek's avatar
Retab  
Andreas Marek committed
409
                                 n_cols, l_cols, l_rows, ONE, hvm_dev, max_local_rows, &
410
                                       q_dev, ldq , ZERO, tmp_dev, n_cols)
Andreas Marek's avatar
Retab  
Andreas Marek committed
411
      call obj%timer%stop("cublas")
412

Andreas Marek's avatar
Andreas Marek committed
413
#if REALCASE == 1
414 415 416 417
#ifdef WITH_MPI

            ! copy data from device to host for a later MPI_ALLREDUCE
            ! copy to host maybe this can be avoided this is needed if MPI is used (allreduce)
418
            successCUDA = cuda_memcpy(loc(tmp1), tmp_dev, l_cols*n_cols*size_of_datatype, cudaMemcpyDeviceToHost)
419 420
            if (.not.(successCUDA)) then
              print *,"trans_ev_band_to_full_real: error in cudaMemcpy"
Andreas Marek's avatar
Andreas Marek committed
421
              stop 1
422 423 424
            endif

#else /* WITH_MPI */
Andreas Marek's avatar
Andreas Marek committed
425 426 427 428
            ! in real case no copy needed. maybe also in complexcase ?
#endif /* WITH_MPI */

#endif /* REALCASE */
429 430

#if COMPLEXCASE == 1
431
            successCUDA = cuda_memcpy(loc(tmp1), tmp_dev, n_cols*l_cols*size_of_datatype, &
432 433 434 435
                                       cudaMemcpyDeviceToHost)

            if (.not.(successCUDA)) then
              print *,"trans_ev_band_to_full_complex: error in cudaMemcpy"
Andreas Marek's avatar
Andreas Marek committed
436
              stop 1
437 438 439 440 441 442 443 444 445 446 447 448 449 450 451
            endif
#endif

          else ! l_rows>0

#if REALCASE == 1
            tmp1(1:l_cols*n_cols) = 0
#endif
#if COMPLEXCASE == 1
            tmp1(1:l_cols*n_cols) = CONST_COMPLEX_0_0
#endif

          endif ! l_rows>0

          !#ifdef WITH_GPU_VERSION
452
          !       istat = cuda_memcpy(loc(tmp1), tmp_dev, max_local_cols*nbw*size_of_datatype,cudaMemcpyDeviceToHost)
453 454
          !       if (istat .ne. 0) then
          !         print *,"error in cudaMemcpy"
Andreas Marek's avatar
Andreas Marek committed
455
          !         stop 1
456 457 458 459
          !       endif
          !#endif

#ifdef WITH_MPI
460
          call obj%timer%start("mpi_communication")
461 462 463 464 465 466 467 468
          call mpi_allreduce(tmp1, tmp2, n_cols*l_cols, &
#if REALCASE == 1
                             MPI_REAL_PRECISION,        &
#endif
#if COMPLEXCASE == 1
                             MPI_COMPLEX_PRECISION,        &
#endif
                             MPI_SUM, mpi_comm_rows, mpierr)
469
          call obj%timer%stop("mpi_communication")
470 471 472 473 474 475

#else /* WITH_MPI */
!          tmp2(1:n_cols*l_cols) = tmp1(1:n_cols*l_cols)
#endif /* WITH_MPI */

           !#ifdef WITH_GPU_VERSION
476
          !       istat = cuda_memcpy(tmp_dev, loc(tmp2), max_local_cols*nbw*size_of_datatype,cudaMemcpyHostToDevice)
477 478
          !       if (istat .ne. 0) then
          !         print *,"error in cudaMemcpy"
Andreas Marek's avatar
Andreas Marek committed
479
          !         stop 1
480 481 482 483 484 485 486
          !       endif
          !#endif

          if (l_rows>0) then
#ifdef WITH_MPI
            ! after the mpi_allreduce we have to copy back to the device
            ! copy back to device
487
            successCUDA = cuda_memcpy(tmp_dev, loc(tmp2), n_cols*l_cols* size_of_datatype, &
Andreas Marek's avatar
Retab  
Andreas Marek committed
488
              cudaMemcpyHostToDevice)
489 490
            if (.not.(successCUDA)) then
              print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
491 492
        &MATH_DATATYPE&
        &: error in cudaMemcpy"
Andreas Marek's avatar
Andreas Marek committed
493
              stop 1
494 495 496 497 498
            endif
#else /* WITH_MPI */

#if COMPLEXCASE == 1
            ! check whether this could be avoided like in the real case
499
            successCUDA = cuda_memcpy(tmp_dev,loc(tmp1),l_cols*n_cols*size_of_datatype,cudaMemcpyHostToDevice)
500 501
            if (.not.(successCUDA)) then
              print *,"trans_ev_band_to_full_complex: error in cudaMemcpy"
Andreas Marek's avatar
Andreas Marek committed
502
              stop 1
503 504 505 506
            endif
#endif
#endif /* WITH_MPI */

507
!#ifdef WITH_MPI
508 509 510 511 512 513 514
           ! IMPORTANT: even though tmat_dev is transfered from the previous rutine, we have to copy from tmat again 
           ! tmat is 3-dimensional array, while tmat_dev contains only one 2-dimensional slice of it - and here we 
           ! need to upload another slice
           successCUDA = cuda_memcpy(tmat_dev, loc(tmat(1,1,istep)), nbw*nbw*size_of_datatype, cudaMemcpyHostToDevice)

           if (.not.(successCUDA)) then
             print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
515 516
        &MATH_DATATYPE&
        &: error in cudaMemcpy"
517 518
             stop 1
           endif
519
!#endif /* WITH_MPI */
520

Andreas Marek's avatar
Retab  
Andreas Marek committed
521
      call obj%timer%start("cublas")
522 523 524 525 526 527 528 529 530 531
#if REALCASE == 1
            call cublas_PRECISION_TRMM('L', 'U', 'T', 'N',       &
#endif
#if COMPLEXCASE == 1
            call cublas_PRECISION_TRMM('L', 'U', 'C', 'N',      &
#endif
                                        n_cols, l_cols, ONE, tmat_dev, nbw, tmp_dev, n_cols)

            call cublas_PRECISION_GEMM('N', 'N', l_rows, l_cols, n_cols, -ONE, hvm_dev, max_local_rows, &
                                       tmp_dev, n_cols, one, q_dev, ldq)
Andreas Marek's avatar
Retab  
Andreas Marek committed
532
      call obj%timer%stop("cublas")
533 534 535 536

#if REALCASE == 1
            ! copy to host maybe this can be avoided
            ! this is not necessary hvm is not used anymore
537
            successCUDA = cuda_memcpy(loc(hvm), hvm_dev, ((max_local_rows)*nbw*size_of_datatype),cudaMemcpyDeviceToHost)
538 539
            if (.not.(successCUDA)) then
              print *,"trans_ev_band_to_full_real: error in cudaMemcpy"
Andreas Marek's avatar
Andreas Marek committed
540
              stop 1
541 542 543 544 545
            endif
#endif
          endif ! l_rows > 0

          !#ifdef WITH_GPU_VERSION
546
          !       istat = cuda_memcpy(loc(hvm), hvm_dev, ((max_local_rows)*nbw*size_of_datatype),cudaMemcpyDeviceToHost)
547 548
          !       if (istat .ne. 0) then
          !         print *,"error in cudaMemcpy"
Andreas Marek's avatar
Andreas Marek committed
549
          !         stop 1
550 551 552 553 554 555 556 557
          !       endif
          !
          !#endif
        enddo ! istep



      else ! do not useGPU
Andreas Marek's avatar
Andreas Marek committed
558 559

#ifdef BAND_TO_FULL_BLOCKING
560 561 562 563 564 565 566 567 568
        ! 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), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
569
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
570 571
    &MATH_DATATYPE&
    &: error when allocating tmp1 "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
572
          stop 1
573 574 575 576
        endif

        allocate(tmp2(max_local_cols*cwy_blocking), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
577
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
578 579
    &MATH_DATATYPE&
    &: error when allocating tmp2 "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
580
          stop 1
581 582 583 584
        endif

        allocate(hvb(max_local_rows*cwy_blocking), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
585
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
586 587
    &MATH_DATATYPE&
    &: error when allocating hvb "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
588
          stop 1
589 590 591 592
        endif

        allocate(hvm(max_local_rows,cwy_blocking), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
593
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
594 595
    &MATH_DATATYPE&
    &: error when allocating hvm "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
596
          stop 1
597 598
        endif

Andreas Marek's avatar
Andreas Marek committed
599 600
#else /* BAND_TO_FULL_BLOCKING */

601 602
        allocate(tmp1(max_local_cols*nbw), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
603
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
604 605
    &MATH_DATATYPE&
    &: error when allocating tmp1 "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
606
          stop 1
607 608 609 610
        endif

        allocate(tmp2(max_local_cols*nbw), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
611
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
612
    &MATH_DATATYPE&: error when allocating tmp2 "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
613
          stop 1
614 615 616 617
        endif

        allocate(hvb(max_local_rows*nbw), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
618
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
619 620
    &MATH_DATATYPE&
    &: error when allocating hvb "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
621
          stop 1
622 623 624 625
        endif

        allocate(hvm(max_local_rows,nbw), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
626
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
627 628
    &MATH_DATATYPE&
    &: error when allocating hvm "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
629
          stop 1
630
        endif
Andreas Marek's avatar
Andreas Marek committed
631
#endif /* BAND_TO_FULL_BLOCKING */
632

Andreas Marek's avatar
Andreas Marek committed
633
#ifdef BAND_TO_FULL_BLOCKING
634 635
        allocate(tmat_complete(cwy_blocking,cwy_blocking), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
636
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
637 638
    &MATH_DATATYPE&
    &: error when allocating tmat_complete "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
639
          stop 1
640 641 642
        endif
        allocate(t_tmp(cwy_blocking,nbw), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
643
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
644 645
    &MATH_DATATYPE&
    &: error when allocating t_tmp "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
646
          stop 1
647 648 649
        endif
        allocate(t_tmp2(cwy_blocking,nbw), stat=istat, errmsg=errorMessage)
        if (istat .ne. 0) then
Andreas Marek's avatar
Andreas Marek committed
650
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
651 652
    &MATH_DATATYPE&
    &: error when allocating t_tmp2 "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
653
          stop 1
654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674
        endif
#endif
!        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

#if REALCASE == 1
        hvm = CONST_0_0   ! Must be set to 0 !!!
        hvb = CONST_0_0   ! Safety only
#endif
#if COMPLEXCASE == 1
        hvm = CONST_COMPLEX_0_0   ! Must be set to 0 !!!
        hvb = CONST_COMPLEX_0_0   ! Safety only
#endif
        l_cols = local_index(nqc, my_pcol, np_cols, nblk, -1) ! Local columns of q

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

Andreas Marek's avatar
Andreas Marek committed
675
#ifdef BAND_TO_FULL_BLOCKING
676
        do istep=1,((na-1)/nbw-1)/t_blocking + 1
Andreas Marek's avatar
Andreas Marek committed
677
#else
678 679 680
        do istep=1,(na-1)/nbw
#endif

Andreas Marek's avatar
Andreas Marek committed
681
#ifdef BAND_TO_FULL_BLOCKING
682 683 684 685 686 687 688 689 690 691 692
          ! 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
Andreas Marek's avatar
Andreas Marek committed
693
#else /* BAND_TO_FULL_BLOCKING */
694
          n_cols = MIN(na,(istep+1)*nbw) - istep*nbw ! Number of columns in current step
Andreas Marek's avatar
Andreas Marek committed
695
#endif /* BAND_TO_FULL_BLOCKING */
696 697 698 699 700 701
          ! Broadcast all Householder vectors for current step compressed in hvb

          nb = 0
          ns = 0

          do lc = 1, n_cols
Andreas Marek's avatar
Andreas Marek committed
702
#ifdef BAND_TO_FULL_BLOCKING
703
            ncol = (istep-1)*cwy_blocking + nbw + lc ! absolute column number of householder Vector
Andreas Marek's avatar
Andreas Marek committed
704
#else
705
            ncol = istep*nbw + lc ! absolute column number of householder Vector
706 707 708 709 710 711 712 713 714 715 716 717
#endif
            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
#ifdef WITH_MPI
718
              call obj%timer%start("mpi_communication")
719 720
              call MPI_Bcast(hvb(ns+1), nb-ns,      &
#if REALCASE == 1
Andreas Marek's avatar
Retab  
Andreas Marek committed
721
                       MPI_REAL_PRECISION,    &
722 723 724 725
#endif
#if COMPLEXCASE == 1
                             MPI_COMPLEX_PRECISION, &
#endif
Andreas Marek's avatar
Retab  
Andreas Marek committed
726
           pcol(ncol, nblk, np_cols), mpi_comm_cols, mpierr)
727

728
              call obj%timer%stop("mpi_communication")
729 730 731 732 733 734 735 736 737 738

#endif /* WITH_MPI */
              ns = nb
            endif
          enddo ! lc

          ! Expand compressed Householder vectors into matrix hvm

          nb = 0
          do lc = 1, n_cols
Andreas Marek's avatar
Andreas Marek committed
739
#ifdef BAND_TO_FULL_BLOCKING
740
            nrow = (istep-1)*cwy_blocking + lc ! absolute number of pivot row
Andreas Marek's avatar
Andreas Marek committed
741
#else
742 743 744 745 746 747 748 749 750 751 752 753 754 755
            nrow = (istep-1)*nbw+lc ! absolute number of pivot row
#endif
            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 REALCASE == 1
            if (my_prow==prow(nrow, nblk, np_rows)) hvm(l_rows+1,lc) = CONST_1_0
#endif
#if COMPLEXCASE == 1
            if (my_prow==prow(nrow, nblk, np_rows)) hvm(l_rows+1,lc) = 1.
#endif
            nb = nb+l_rows
          enddo

Andreas Marek's avatar
Andreas Marek committed
756
#ifdef BAND_TO_FULL_BLOCKING
757 758 759 760 761 762 763 764 765
          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)
Andreas Marek's avatar
Andreas Marek committed
766

767
            if (i > 1) then
Andreas Marek's avatar
Retab  
Andreas Marek committed
768
        call obj%timer%start("blas")
Andreas Marek's avatar
Andreas Marek committed
769 770 771 772 773 774
#if REALCASE == 1
              call PRECISION_GEMM('T', 'N',      &
#endif
#if COMPLEXCASE == 1
              call PRECISION_GEMM('C', 'N',      &
#endif
Andreas Marek's avatar
Retab  
Andreas Marek committed
775
                           t_rows, t_cols, l_rows, ONE, hvm(1,1), max_local_rows, hvm(1,(i-1)*nbw+1), &
Andreas Marek's avatar
Andreas Marek committed
776 777
                                 max_local_rows, ZERO, t_tmp, cwy_blocking)

778
            call obj%timer%stop("blas")
779
#ifdef WITH_MPI
780
              call obj%timer%start("mpi_communication")
781

Andreas Marek's avatar
Andreas Marek committed
782 783
              call mpi_allreduce(t_tmp, t_tmp2, cwy_blocking*nbw,     &
#if  REALCASE == 1
Andreas Marek's avatar
Retab  
Andreas Marek committed
784
                           MPI_REAL_PRECISION,    &
Andreas Marek's avatar
Andreas Marek committed
785 786 787 788
#endif
#if COMPLEXCASE == 1
                                 MPI_COMPLEX_PRECISION,       &
#endif
Andreas Marek's avatar
Retab  
Andreas Marek committed
789
         MPI_SUM, mpi_comm_rows, mpierr)
790
              call obj%timer%stop("mpi_communication")
Andreas Marek's avatar
Retab  
Andreas Marek committed
791
        call obj%timer%start("blas")
Andreas Marek's avatar
Andreas Marek committed
792 793
              call PRECISION_TRMM('L', 'U', 'N', 'N', t_rows, t_cols, ONE, tmat_complete, cwy_blocking, t_tmp2, cwy_blocking)
              call PRECISION_TRMM('R', 'U', 'N', 'N', t_rows, t_cols, -ONE, tmat_complete(t_rows+1,t_rows+1), cwy_blocking, &
794
                         t_tmp2, cwy_blocking)
Andreas Marek's avatar
Retab  
Andreas Marek committed
795
             call obj%timer%stop("blas")
796 797 798

              tmat_complete(1:t_rows,t_rows+1:t_rows+t_cols) = t_tmp2(1:t_rows,1:t_cols)

Andreas Marek's avatar
Andreas Marek committed
799
#else /* WITH_MPI */
800
!              t_tmp2(1:cwy_blocking,1:nbw) = t_tmp(1:cwy_blocking,1:nbw)
Andreas Marek's avatar
Retab  
Andreas Marek committed
801
        call obj%timer%start("blas")
Andreas Marek's avatar
Andreas Marek committed
802 803
              call PRECISION_TRMM('L', 'U', 'N', 'N', t_rows, t_cols, ONE, tmat_complete, cwy_blocking, t_tmp, cwy_blocking)
              call PRECISION_TRMM('R', 'U', 'N', 'N', t_rows, t_cols, -ONE, tmat_complete(t_rows+1,t_rows+1), cwy_blocking, &
804
                         t_tmp, cwy_blocking)
Andreas Marek's avatar
Retab  
Andreas Marek committed
805
        call obj%timer%stop("blas")
806 807 808

              tmat_complete(1:t_rows,t_rows+1:t_rows+t_cols) = t_tmp(1:t_rows,1:t_cols)

Andreas Marek's avatar
Andreas Marek committed
809
#endif /* WITH_MPI */
810 811 812 813 814 815 816 817

!              call PRECISION_TRMM('L', 'U', 'N', 'N', t_rows, t_cols, CONST_1_0, tmat_complete, cwy_blocking, t_tmp2, cwy_blocking)
!              call PRECISION_TRMM('R', 'U', 'N', 'N', t_rows, t_cols, -CONST_1_0, 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
Andreas Marek's avatar
Andreas Marek committed
818
#else /* BAND_TO_FULL_BLOCKING */
819 820 821 822 823 824
          l_rows = local_index(MIN(na,(istep+1)*nbw), my_prow, np_rows, nblk, -1)
#endif

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

          if (l_rows>0) then
825
            call obj%timer%start("blas")
826 827 828 829 830 831 832

#if REALCASE == 1
            call PRECISION_GEMM('T', 'N',         &
#endif
#if COMPLEXCASE == 1
            call PRECISION_GEMM('C', 'N',         &
#endif
Andreas Marek's avatar
Retab  
Andreas Marek committed
833
                          n_cols, l_cols, l_rows, ONE, hvm, ubound(hvm,dim=1), &
834
                                q, ldq, ZERO, tmp1, n_cols)
Andreas Marek's avatar
Retab  
Andreas Marek committed
835
      call obj%timer%stop("blas")
836 837 838 839 840 841 842 843 844 845 846 847

          else ! l_rows>0

#if REALCASE == 1
            tmp1(1:l_cols*n_cols) = CONST_0_0
#endif
#if COMPLEXCASE == 1
            tmp1(1:l_cols*n_cols) = CONST_COMPLEX_0_0
#endif
          endif ! l_rows>0

#ifdef WITH_MPI
848
          call obj%timer%start("mpi_communication")
849 850
          call mpi_allreduce(tmp1, tmp2, n_cols*l_cols,  &
#if REALCASE == 1
Andreas Marek's avatar
Retab  
Andreas Marek committed
851
                       MPI_REAL_PRECISION,         &
852 853
#endif
#if COMPLEXCASE == 1
Andreas Marek's avatar
Retab  
Andreas Marek committed
854
                       MPI_COMPLEX_PRECISION,       &
855 856

#endif
Andreas Marek's avatar
Retab  
Andreas Marek committed
857
           MPI_SUM, mpi_comm_rows ,mpierr)
858

859
          call obj%timer%stop("mpi_communication")
860

Andreas Marek's avatar
Retab  
Andreas Marek committed
861
    call obj%timer%start("blas")
862 863

          if (l_rows>0) then
Andreas Marek's avatar
Andreas Marek committed
864 865
#ifdef BAND_TO_FULL_BLOCKING

866
#if REALCASE == 1
Andreas Marek's avatar
Andreas Marek committed
867 868 869 870 871
            call PRECISION_TRMM('L', 'U', 'T', 'N',        &
#endif
#if COMPLEXCASE == 1
            call PRECISION_TRMM('L', 'U', 'C', 'N',        &
#endif
Andreas Marek's avatar
Retab  
Andreas Marek committed
872
                          n_cols, l_cols, ONE, tmat_complete, cwy_blocking, tmp2, n_cols)
873
            call PRECISION_GEMM('N', 'N', l_rows, l_cols, n_cols, -ONE, hvm, ubound(hvm,dim=1), tmp2, n_cols, ONE, q, ldq)
Andreas Marek's avatar
Andreas Marek committed
874 875 876 877 878

#else /* BAND_TO_FULL_BLOCKING */

#if REALCASE == 1
            call PRECISION_TRMM('L', 'U', 'T', 'N',        &
879 880
#endif
#if COMPLEXCASE == 1
Andreas Marek's avatar
Andreas Marek committed
881 882 883
            call PRECISION_TRMM('L', 'U', 'C', 'N',        &
#endif
                                n_cols, l_cols, ONE, tmat(1,1,istep), ubound(tmat,dim=1), tmp2, n_cols)
884 885
            call PRECISION_GEMM('N', 'N', l_rows, l_cols, n_cols, -ONE, hvm, ubound(hvm,dim=1), &
                                tmp2, n_cols, ONE, q, ldq)
Andreas Marek's avatar
Andreas Marek committed
886 887

#endif /* BAND_TO_FULL_BLOCKING */
888 889

          endif
Andreas Marek's avatar
Retab  
Andreas Marek committed
890
    call obj%timer%stop("blas")
891 892
#else /* WITH_MPI */
!          tmp2 = tmp1
Andreas Marek's avatar
Retab  
Andreas Marek committed
893
    call obj%timer%start("blas")
894
          if (l_rows>0) then
Andreas Marek's avatar
Andreas Marek committed
895 896
#ifdef BAND_TO_FULL_BLOCKING

897
#if REALCASE == 1
Andreas Marek's avatar
Andreas Marek committed
898 899 900 901 902 903
            call PRECISION_TRMM('L', 'U', 'T', 'N',        &
#endif
#if COMPLEXCASE == 1
            call PRECISION_TRMM('L', 'U', 'C', 'N',        &
#endif
                                n_cols, l_cols, ONE, tmat_complete, cwy_blocking, tmp1, n_cols)
904
            call PRECISION_GEMM('N', 'N', l_rows, l_cols, n_cols, -ONE, hvm, ubound(hvm,dim=1), tmp1, n_cols, ONE, q, ldq)
Andreas Marek's avatar
Andreas Marek committed
905 906 907 908 909 910


#else /* BAND_TO_FULL_BLOCKING */

#if REALCASE == 1
            call PRECISION_TRMM('L', 'U', 'T', 'N',        &
911 912
#endif
#if COMPLEXCASE == 1
Andreas Marek's avatar
Andreas Marek committed
913 914 915
            call PRECISION_TRMM('L', 'U', 'C', 'N',        &
#endif
                                n_cols, l_cols, ONE, tmat(1,1,istep), ubound(tmat,dim=1), tmp1, n_cols)
916 917
            call PRECISION_GEMM('N', 'N', l_rows, l_cols, n_cols, -ONE, hvm, ubound(hvm,dim=1), &
                                tmp1, n_cols, ONE, q, ldq)
Andreas Marek's avatar
Andreas Marek committed
918 919

#endif  /* BAND_TO_FULL_BLOCKING */
920
          endif
Andreas Marek's avatar
Retab  
Andreas Marek committed
921
    call obj%timer%stop("blas")
922 923 924 925 926 927 928 929 930 931 932 933 934 935
#endif /* WITH_MPI */

!          if (l_rows>0) then
!            call PRECISION_TRMM('L', 'U', 'T', 'N', n_cols, l_cols, CONST_1_0, tmat_complete, cwy_blocking, tmp2, n_cols)
!            call PRECISION_GEMM('N', 'N', l_rows, l_cols, n_cols, -CONST_1_0, hvm, ubound(hvm,dim=1), tmp2, n_cols, CONST_1_0, q, ldq)
!          endif

        enddo ! istep

      endif ! useGPU

      deallocate(tmp1, tmp2, hvb, stat=istat, errmsg=errorMessage)
      if (istat .ne. 0) then
        print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
936 937
  &MATH_DATATYPE&
  &: error when deallocating tmp1 tmp2 hvb "//errorMessage
Andreas Marek's avatar
Andreas Marek committed
938
        stop 1
939 940 941 942 943 944
      endif

      if (useGPU) then
        successCUDA = cuda_free(hvm_dev)
        if (.not.(successCUDA)) then
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
945
    &MATH_DATATYPE&
946
          &: error in cudaFree"
Andreas Marek's avatar
Andreas Marek committed
947
          stop 1
948 949 950 951 952
        endif

        successCUDA = cuda_free(tmp_dev)
        if (.not.(successCUDA)) then
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
953
    &MATH_DATATYPE&
954
          &: error in cudaFree"
Andreas Marek's avatar
Andreas Marek committed
955
          stop 1
956 957 958 959 960
        endif

        successCUDA = cuda_free(tmat_dev)
        if (.not.(successCUDA)) then
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
961
    &MATH_DATATYPE&
962
          &: error in cudaFree"
Andreas Marek's avatar
Andreas Marek committed
963
          stop 1
964 965 966
        endif

         ! final transfer of q_dev
967
         successCUDA = cuda_memcpy(loc(q), q_dev, ldq*matrixCols* size_of_datatype, cudaMemcpyDeviceToHost)
968 969 970

         if (.not.(successCUDA)) then
          print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
971
    &MATH_DATATYPE&
972
          &: error in cudamemcpu q_dev"
Andreas Marek's avatar
Andreas Marek committed
973
          stop 1
974 975 976 977 978 979 980
         endif

         !   q(1:ldq,1:na_cols) = q_temp(1:ldq,1:na_cols)

         successCUDA = cuda_free(q_dev)
         if (.not.(successCUDA)) then
           print *,"trans_ev_band_to_full_&
Andreas Marek's avatar
Retab  
Andreas Marek committed
981
     &MATH_DATATYPE&
982
           &: error in cudaFree"
Andreas Marek's avatar
Andreas Marek committed
983
           stop 1
984 985 986 987 988
         endif

         !   deallocate(q_temp, stat=istat, errmsg=errorMessage)
         !   if (istat .ne. 0) then
         !     print *,"error when deallocating q_temp "//errorMessage
Andreas Marek's avatar
Andreas Marek committed