#if 0 ! This file is part of ELPA. ! ! The ELPA library was originally created by the ELPA consortium, ! consisting of the following organizations: ! ! - Max Planck Computing and Data Facility (MPCDF), fomerly known as ! Rechenzentrum Garching der Max-Planck-Gesellschaft (RZG), ! - Bergische Universität Wuppertal, Lehrstuhl für angewandte ! Informatik, ! - Technische Universität München, Lehrstuhl für Informatik mit ! Schwerpunkt Wissenschaftliches Rechnen , ! - Fritz-Haber-Institut, Berlin, Abt. Theorie, ! - Max-Plack-Institut für Mathematik in den Naturwissenschaften, ! Leipzig, Abt. Komplexe Strukutren in Biologie und Kognition, ! and ! - IBM Deutschland GmbH ! ! This particular source code file contains additions, changes and ! enhancements authored by Intel Corporation which is not part of ! the ELPA consortium. ! ! More information can be found here: ! http://elpa.mpcdf.mpg.de/ ! ! ELPA is free software: you can redistribute it and/or modify ! it under the terms of the version 3 of the license of the ! GNU Lesser General Public License as published by the Free ! Software Foundation. ! ! ELPA is distributed in the hope that it will be useful, ! but WITHOUT ANY WARRANTY; without even the implied warranty of ! MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the ! GNU Lesser General Public License for more details. ! ! You should have received a copy of the GNU Lesser General Public License ! along with ELPA. If not, see ! ! ELPA reflects a substantial effort on the part of the original ! ELPA consortium, and we ask you to respect the spirit of the ! license that we chose: i.e., please contribute any changes you ! may have back to the original ELPA library distribution, and keep ! any derivatives of ELPA under the same license that we chose for ! the original distribution, the GNU Lesser General Public License. ! ! ! ELPA1 -- Faster replacements for ScaLAPACK symmetric eigenvalue routines ! ! Copyright of the original code rests with the authors inside the ELPA ! consortium. The copyright of any additional modifications shall rest ! with their original authors, but shall adhere to the licensing terms ! distributed along with the original code in the file "COPYING". ! ELPA2 -- 2-stage solver for ELPA ! ! Copyright of the original code rests with the authors inside the ELPA ! consortium. The copyright of any additional modifications shall rest ! with their original authors, but shall adhere to the licensing terms ! distributed along with the original code in the file "COPYING". #endif subroutine bandred_& &MATH_DATATYPE& &_& &PRECISION & (obj, na, a_mat, a_dev, lda, nblk, nbw, matrixCols, numBlocks, mpi_comm_rows, mpi_comm_cols, tmat, & tmat_dev, wantDebug, useGPU, success, & #if REALCASE == 1 useQR, & #endif max_threads) !------------------------------------------------------------------------------- ! bandred_real/complex: Reduces a distributed symmetric matrix to band form ! ! Parameters ! ! na Order of matrix ! ! a_mat(lda,matrixCols) Distributed matrix which should be reduced. ! Distribution is like in Scalapack. ! Opposed to Scalapack, a_mat(:,:) must be set completely (upper and lower half) ! a_mat(:,:) is overwritten on exit with the band and the Householder vectors ! in the upper half. ! ! lda Leading dimension of a_mat ! matrixCols local columns of matrix a_mat ! ! nblk blocksize of cyclic distribution, must be the same in both directions! ! ! nbw semi bandwith of output matrix ! ! mpi_comm_rows ! mpi_comm_cols ! MPI-Communicators for rows/columns ! ! tmat(nbw,nbw,numBlocks) where numBlocks = (na-1)/nbw + 1 ! Factors for the Householder vectors (returned), needed for back transformation ! !------------------------------------------------------------------------------- use cuda_functions use iso_c_binding use elpa1_compute #ifdef WITH_OPENMP use omp_lib #endif use precision use elpa_abstract_impl implicit none #include "../general/precision_kinds.F90" class(elpa_abstract_impl_t), intent(inout) :: obj integer(kind=ik) :: na, lda, nblk, nbw, matrixCols, numBlocks, mpi_comm_rows, mpi_comm_cols #ifdef USE_ASSUMED_SIZE MATH_DATATYPE(kind=rck) :: a_mat(lda,*), tmat(nbw,nbw,*) #else MATH_DATATYPE(kind=rck) :: a_mat(lda,matrixCols), tmat(nbw,nbw,numBlocks) #endif #if REALCASE == 1 real(kind=rk) :: eps #endif logical, intent(in) :: useGPU character(20) :: gpuString integer(kind=ik) :: my_prow, my_pcol, np_rows, np_cols, mpierr integer(kind=ik) :: l_cols, l_rows #if REALCASE == 1 integer(kind=ik) :: vmrCols #endif #ifdef WITH_OPENMP integer(kind=ik) :: mynlc, lrs, transformChunkSize #endif integer(kind=ik) :: i, j, lcs, lce, lre, lc, lr, cur_pcol, n_cols, nrow integer(kind=ik) :: istep, ncol, lch, lcx, nlc integer(kind=ik) :: tile_size, l_rows_tile, l_cols_tile real(kind=rk) :: vnorm2 MATH_DATATYPE(kind=rck) :: xf, aux1(nbw), aux2(nbw), vrl, tau, vav(nbw,nbw) ! complex(kind=COMPLEX_DATATYPE), allocatable :: tmpCUDA(:,:), vmrCUDA(:,:), umcCUDA(:,:) ! note the different dimension in real case MATH_DATATYPE(kind=rck), allocatable :: tmpCUDA(:), vmrCUDA(:), umcCUDA(:) MATH_DATATYPE(kind=rck), allocatable :: tmpCPU(:,:), vmrCPU(:,:), umcCPU(:,:) MATH_DATATYPE(kind=rck), allocatable :: vr(:) #if REALCASE == 1 ! needed for blocked QR decomposition integer(kind=ik) :: PQRPARAM(11), work_size real(kind=rk) :: dwork_size(1) real(kind=rk), allocatable :: work_blocked(:), tauvector(:), blockheuristic(:) #endif ! a_dev is passed from bandred_real to trans_ev_band integer(kind=C_intptr_T) :: a_dev, vmr_dev, umc_dev, tmat_dev, vav_dev #ifdef WITH_MPI integer(kind=ik), external :: numroc #endif integer(kind=ik) :: ierr integer(kind=ik) :: cur_l_rows, cur_l_cols, vmr_size, umc_size integer(kind=c_intptr_t) :: lc_start, lc_end #if COMPLEXCASE == 1 integer(kind=c_intptr_t) :: lce_1, lcs_1, lre_1 #endif integer(kind=ik) :: lr_end integer(kind=ik) :: na_cols #if COMPLEXCASE == 1 integer(kind=ik) :: na_rows #endif logical, intent(in) :: wantDebug logical, intent(out) :: success logical :: successCUDA integer(kind=ik) :: istat character(200) :: errorMessage integer(kind=ik) :: min_tile_size, error #if REALCASE == 1 logical, intent(in) :: useQR #endif integer(kind=ik) :: mystart, myend, m_way, n_way, work_per_thread, m_id, n_id, n_threads, & ii, pp integer(kind=c_intptr_t), parameter :: size_of_datatype = size_of_& &PRECISION& &_& &MATH_DATATYPE logical :: useGPU_reduction_lower_block_to_tridiagonal integer(kind=ik), intent(in) :: max_threads if(useGPU) then gpuString = "_gpu" else gpuString = "" endif call obj%timer%start("bandred_& &MATH_DATATYPE& &" // & PRECISION_SUFFIX // & gpuString ) useGPU_reduction_lower_block_to_tridiagonal = .false. if (useGPU) then useGPU_reduction_lower_block_to_tridiagonal = .true. #if REALCASE == 1 if (useQR) then !in this case switch off GPU usage for step "reduce current block to lower triangular form" ! since this is done by QR decomposition useGPU_reduction_lower_block_to_tridiagonal = .false. endif #endif endif if (wantDebug) call obj%timer%start("mpi_communication") call mpi_comm_rank(mpi_comm_rows,my_prow,mpierr) call mpi_comm_size(mpi_comm_rows,np_rows,mpierr) call mpi_comm_rank(mpi_comm_cols,my_pcol,mpierr) call mpi_comm_size(mpi_comm_cols,np_cols,mpierr) if (wantDebug) call obj%timer%stop("mpi_communication") success = .true. ! Semibandwith nbw must be a multiple of blocksize nblk if (mod(nbw,nblk)/=0) then if (my_prow==0 .and. my_pcol==0) then if (wantDebug) then write(error_unit,*) 'ELPA2_bandred_& &MATH_DATATYPE& &: ERROR: nbw=',nbw,', nblk=',nblk write(error_unit,*) 'ELPA2_bandred_& &MATH_DATATYPE& &: ELPA2 works only for nbw==n*nblk' endif success = .false. return endif endif ! na_rows in used nowhere; only na_cols if (useGPU) then #ifdef WITH_MPI #if COMPLEXCASE == 1 na_rows = numroc(na, nblk, my_prow, 0, np_rows) #endif na_cols = numroc(na, nblk, my_pcol, 0, np_cols) #else #if COMPLEXCASE == 1 na_rows = na #endif na_cols = na #endif /* WITH_MPI */ ! Here we convert the regular host array into a pinned host array successCUDA = cuda_malloc(a_dev, lda*na_cols* size_of_datatype) if (.not.(successCUDA)) then print *,"bandred_& &MATH_DATATYPE& &: error in cudaMalloc a_dev 1" stop 1 endif successCUDA = cuda_malloc(tmat_dev, nbw*nbw* size_of_datatype) if (.not.(successCUDA)) then print *,"bandred_& &MATH_DATATYPE& &: error in cudaMalloc tmat_dev 1" stop 1 endif successCUDA = cuda_malloc(vav_dev, nbw*nbw* size_of_datatype) if (.not.(successCUDA)) then print *,"bandred_& &MATH_DATATYPE& &: error in cudaMalloc vav_dev 1" stop 1 endif endif ! useGPU ! Matrix is split into tiles; work is done only for tiles on the diagonal or above tile_size = nblk*least_common_multiple(np_rows,np_cols) ! minimum global tile size ! make tile_size a smallest possible multiple of previously defined tile size, such that it is ! larger or equal to min_tile_size ! min_tile_size has been originally hardcoded as 128 * max(np_rows, np_cols), so it is now the implicit value ! it can, however, be set by the user call obj%get("min_tile_size", min_tile_size ,error) if (error .ne. ELPA_OK) then print *,"Problem setting option. Aborting..." stop endif if(min_tile_size == 0) then ! not set by the user, use the default value min_tile_size = 128*max(np_rows, np_cols) endif tile_size = ((min_tile_size-1)/tile_size+1)*tile_size l_rows_tile = tile_size/np_rows ! local rows of a tile l_cols_tile = tile_size/np_cols ! local cols of a tile #if REALCASE == 1 if (useQR) then if (which_qr_decomposition == 1) then call qr_pqrparam_init(obj,pqrparam(1:11), nblk,'M',0, nblk,'M',0, nblk,'M',1,'s') allocate(tauvector(na), stat=istat, errmsg=errorMessage) if (istat .ne. 0) then print *,"bandred_real: error when allocating tauvector "//errorMessage stop 1 endif allocate(blockheuristic(nblk), stat=istat, errmsg=errorMessage) if (istat .ne. 0) then print *,"bandred_real: error when allocating blockheuristic "//errorMessage stop 1 endif l_rows = local_index(na, my_prow, np_rows, nblk, -1) allocate(vmrCPU(max(l_rows,1),na), stat=istat, errmsg=errorMessage) if (istat .ne. 0) then print *,"bandred_real: error when allocating vmrCPU "//errorMessage stop 1 endif vmrCols = na #ifdef USE_ASSUMED_SIZE_QR call qr_pdgeqrf_2dcomm_& &PRECISION& &(obj, a_mat, lda, matrixCols, vmrCPU, max(l_rows,1), vmrCols, tauvector(1), na, tmat(1,1,1), & nbw, nbw, dwork_size, 1, -1, na, nbw, nblk, nblk, na, na, 1, 0, PQRPARAM(1:11), & mpi_comm_rows, mpi_comm_cols, blockheuristic) #else call qr_pdgeqrf_2dcomm_& &PRECISION& &(obj, a_mat(1:lda,1:matrixCols), matrixCols, lda, vmrCPU(1:max(l_rows,1),1:vmrCols), max(l_rows,1), & vmrCols, tauvector(1:na), na, tmat(1:nbw,1:nbw,1), nbw, & nbw, dwork_size(1:1), 1, -1, na, nbw, nblk, nblk, na, na, 1, 0, PQRPARAM(1:11), & mpi_comm_rows, mpi_comm_cols, blockheuristic) #endif work_size = int(dwork_size(1)) allocate(work_blocked(work_size), stat=istat, errmsg=errorMessage) if (istat .ne. 0) then print *,"bandred_real: error when allocating work_blocked "//errorMessage stop 1 endif work_blocked = 0.0_rk deallocate(vmrCPU, stat=istat, errmsg=errorMessage) if (istat .ne. 0) then print *,"bandred_real: error when deallocating vmrCPU "//errorMessage stop 1 endif endif ! which_qr_decomposition endif ! useQr #endif /* REALCASE */ if (useGPU) then cur_l_rows = 0 cur_l_cols = 0 successCUDA = cuda_memcpy(a_dev, loc(a_mat(1,1)), (lda)*(na_cols)* size_of_datatype, cudaMemcpyHostToDevice) if (.not.(successCUDA)) then print *,"bandred_& &MATH_DATATYPE& &: error in cudaMemcpy a_dev 2" stop 1 endif endif ! useGPU do istep = (na-1)/nbw, 1, -1 n_cols = MIN(na,(istep+1)*nbw) - istep*nbw ! Number of columns in current step ! Number of local columns/rows of remaining matrix l_cols = local_index(istep*nbw, my_pcol, np_cols, nblk, -1) l_rows = local_index(istep*nbw, my_prow, np_rows, nblk, -1) ! Allocate vmr and umc to their exact sizes so that they can be used in bcasts and reduces if (useGPU) then cur_l_rows = max(l_rows, 1) cur_l_cols = max(l_cols, 1) vmr_size = cur_l_rows * 2 * n_cols umc_size = cur_l_cols * 2 * n_cols ! Allocate vmr and umc only if the inew size exceeds their current capacity ! Added for FORTRAN CALLS if ((.not. allocated(vr)) .or. (l_rows + 1 .gt. ubound(vr, dim=1))) then if (allocated(vr)) then deallocate(vr, stat=istat, errmsg=errorMessage) if (istat .ne. 0) then print *,"bandred_& &MATH_DATATYPE& &: error when deallocating vr "//errorMessage stop 1 endif endif allocate(vr(l_rows + 1), stat=istat, errmsg=errorMessage) if (istat .ne. 0) then print *,"bandred_& &MATH_DATATYPE& &: error when allocating vr "//errorMessage stop 1 endif endif if ((.not. allocated(vmrCUDA)) .or. (vmr_size .gt. ubound(vmrCUDA, dim=1))) then if (allocated(vmrCUDA)) then deallocate(vmrCUDA, stat=istat, errmsg=errorMessage) if (istat .ne. 0) then print *,"bandred_& &MATH_DATATYPE& &: error when allocating vmrCUDA "//errorMessage stop 1 endif successCUDA = cuda_free(vmr_dev) if (.not.(successCUDA)) then print *,"bandred_& &MATH_DATATYPE&: error in cuda_free vmr_dev 1" stop 1 endif endif allocate(vmrCUDA(vmr_size), stat=istat, errmsg=errorMessage) if (istat .ne. 0) then print *,"bandred_& &MATH_DATATYPE& &: error when allocating vmrCUDA "//errorMessage stop 1 endif successCUDA = cuda_malloc(vmr_dev, vmr_size* size_of_datatype) if (.not.(successCUDA)) then print *,"bandred_& &MATH_DATATYPE& &: error in cudaMalloc: vmr_dev2" stop 1 endif endif if ((.not. allocated(umcCUDA)) .or. (umc_size .gt. ubound(umcCUDA, dim=1))) then if (allocated(umcCUDA)) then deallocate(umcCUDA, stat=istat, errmsg=errorMessage) if (istat .ne. 0) then print *,"bandred_& &MATH_DATATYPE& &: error when deallocating umcCUDA "//errorMessage stop 1 endif successCUDA = cuda_free(umc_dev) if (.not.(successCUDA)) then print *,"bandred_& &MATH_DATATYPE& &: error in cudaFree umc_dev 1" stop 1 endif endif allocate(umcCUDA(umc_size), stat=istat, errmsg=errorMessage) if (istat .ne. 0) then print *,"bandred_& &MATH_DATATYPE& &: error when deallocating umcCUDA "//errorMessage stop 1 endif successCUDA = cuda_malloc(umc_dev, umc_size* size_of_datatype) if (.not.(successCUDA)) then print *,"bandred_& &MATH_DATATYPE& &: error in cudaMalloc umc_dev 2" stop 1 endif endif else ! GPU not used ! unify the the name vmr and vmrCPU, as well as vmrGPU ! the same for umcCPU and umcGPU ! Allocate vmr and umcCPU to their exact sizes so that they can be used in bcasts and reduces allocate(vmrCPU(max(l_rows,1),2*n_cols), stat=istat, errmsg=errorMessage) if (istat .ne. 0) then print *,"bandred_& &MATH_DATATYPE& &: error when allocating vmrCPU "//errorMessage stop 1 endif allocate(umcCPU(max(l_cols,1),2*n_cols), stat=istat, errmsg=errorMessage) if (istat .ne. 0) then print *,"bandred_& &MATH_DATATYPE& &: error when allocating umcCPU "//errorMessage stop 1 endif allocate(vr(l_rows+1), stat=istat, errmsg=errorMessage) if (istat .ne. 0) then print *,"bandred_& &MATH_DATATYPE& &: error when allocating vr "//errorMessage stop 1 endif endif ! use GPU if (useGPU) then vmrCUDA(1 : cur_l_rows * n_cols) = 0.0_rck else vmrCPU(1:l_rows,1:n_cols) = 0.0_rck endif ! useGPU vr(:) = 0.0_rck tmat(:,:,istep) = 0.0_rck if (useGPU) then #if REALCASE == 1 umcCUDA(1 : umc_size) = 0.0_rck #endif lc_start = local_index(istep*nbw+1, my_pcol, np_cols, nblk, -1) lc_end = local_index(istep*nbw+n_cols, my_pcol, np_cols, nblk, -1) lr_end = local_index((istep-1)*nbw + n_cols, my_prow, np_rows, nblk, -1) if (lc_start .le. 0) lc_start = 1 ! Here we assume that the processor grid and the block grid are aligned cur_pcol = pcol(istep*nbw+1, nblk, np_cols) if (my_pcol == cur_pcol) then successCUDA = cuda_memcpy2d(loc(a_mat(1, lc_start)), & int((lda*size_of_datatype),kind=c_intptr_t), & (a_dev + int( ( (lc_start-1) * lda*size_of_datatype),kind=c_intptr_t )), & int(lda*size_of_datatype,kind=c_intptr_t), & int(lr_end*size_of_datatype,kind=c_intptr_t), & int((lc_end - lc_start+1),kind=c_intptr_t),int(cudaMemcpyDeviceToHost,kind=c_int)) if (.not.(successCUDA)) then print *,"bandred_& &MATH_DATATYPE& &: error in cudaMemcpy2d" stop 1 endif endif endif ! useGPU ! Reduce current block to lower triangular form #if REALCASE == 1 if (useQR) then if (useGPU) then ! vmrCPU(1:cur_l_rows,1:n_cols) = vmrCUDA(1 : cur_l_rows * n_cols) endif if (which_qr_decomposition == 1) then vmrCols = 2*n_cols #ifdef USE_ASSUMED_SIZE_QR call qr_pdgeqrf_2dcomm_& &PRECISION& &(obj, a_mat, lda, matrixCols, vmrCPU, max(l_rows,1), vmrCols, tauvector(1), & na, tmat(1,1,istep), nbw, nbw, work_blocked, work_size, & work_size, na, n_cols, nblk, nblk, & istep*nbw+n_cols-nbw, istep*nbw+n_cols, 1,& 0, PQRPARAM(1:11), mpi_comm_rows, mpi_comm_cols,& blockheuristic) #else call qr_pdgeqrf_2dcomm_& &PRECISION& &(obj, a_mat(1:lda,1:matrixCols), lda, matrixCols, vmrCPU(1:max(l_rows,1),1:vmrCols) , & max(l_rows,1), vmrCols, tauvector(1:na), na, & tmat(1:nbw,1:nbw,istep), nbw, nbw, work_blocked(1:work_size), work_size, & work_size, na, n_cols, nblk, nblk, & istep*nbw+n_cols-nbw, istep*nbw+n_cols, 1,& 0, PQRPARAM(1:11), mpi_comm_rows, mpi_comm_cols,& blockheuristic) #endif endif else !useQR #endif /* REALCASE == 1 */ do lc = n_cols, 1, -1 ncol = istep*nbw + lc ! absolute column number of householder Vector nrow = ncol - nbw ! Absolute number of pivot row lr = local_index(nrow, my_prow, np_rows, nblk, -1) ! current row length lch = local_index(ncol, my_pcol, np_cols, nblk, -1) ! HV local column number tau = 0 if (nrow == 1) exit ! Nothing to do cur_pcol = pcol(ncol, nblk, np_cols) ! Processor column owning current block if (my_pcol==cur_pcol) then ! Get Vector to be transformed; distribute last element and norm of ! remaining elements to all procs in current column vr(1:lr) = a_mat(1:lr,lch) ! Vector to be transformed if (my_prow==prow(nrow, nblk, np_rows)) then aux1(1) = dot_product(vr(1:lr-1),vr(1:lr-1)) aux1(2) = vr(lr) else aux1(1) = dot_product(vr(1:lr),vr(1:lr)) aux1(2) = 0.0_rck endif #ifdef WITH_MPI if (wantDebug) call obj%timer%start("mpi_communication") call mpi_allreduce(aux1, aux2, 2, MPI_MATH_DATATYPE_PRECISION, & MPI_SUM, mpi_comm_rows, mpierr) if (wantDebug) call obj%timer%stop("mpi_communication") #else /* WITH_MPI */ aux2 = aux1 ! this should be optimized #endif #if REALCASE == 1 vnorm2 = aux2(1) #endif #if COMPLEXCASE == 1 vnorm2 = real(aux2(1),kind=rk) #endif vrl = aux2(2) ! Householder transformation call hh_transform_& &MATH_DATATYPE& &_& &PRECISION & (obj, vrl, vnorm2, xf, tau, wantDebug) ! Scale vr and store Householder Vector for back transformation vr(1:lr) = vr(1:lr) * xf if (my_prow==prow(nrow, nblk, np_rows)) then a_mat(1:lr-1,lch) = vr(1:lr-1) a_mat(lr,lch) = vrl vr(lr) = 1.0_rck else a_mat(1:lr,lch) = vr(1:lr) endif endif ! Broadcast Householder Vector and tau along columns vr(lr+1) = tau #ifdef WITH_MPI if (wantDebug) call obj%timer%start("mpi_communication") call MPI_Bcast(vr, lr+1, MPI_MATH_DATATYPE_PRECISION, & cur_pcol, mpi_comm_cols, mpierr) if (wantDebug) call obj%timer%stop("mpi_communication") #endif /* WITH_MPI */ if (useGPU_reduction_lower_block_to_tridiagonal) then vmrCUDA(cur_l_rows * (lc - 1) + 1 : cur_l_rows * (lc - 1) + lr) = vr(1:lr) else vmrCPU(1:lr,lc) = vr(1:lr) endif tau = vr(lr+1) #if REALCASE == 1 tmat(lc,lc,istep) = tau ! Store tau in diagonal of tmat #endif #if COMPLEXCASE == 1 tmat(lc,lc,istep) = conjg(tau) ! Store tau in diagonal of tmat #endif ! Transform remaining columns in current block with Householder Vector ! Local dot product aux1 = 0.0_rck #ifdef WITH_OPENMP #if 0 ! original complex implementation without openmp. check performance nlc = 0 ! number of local columns do j=1,lc-1 lcx = local_index(istep*nbw+j, my_pcol, np_cols, nblk, 0) if (lcx>0) then nlc = nlc+1 aux1(nlc) = dot_product(vr(1:lr),a_mat(1:lr,lcx)) endif enddo ! Get global dot products #ifdef WITH_MPI if (wantDebug) call obj%timer%start("mpi_communication") if (nlc>0) call mpi_allreduce(aux1, aux2, nlc, MPI_COMPLEX_PRECISION, MPI_SUM, mpi_comm_rows, mpierr) ! Transform nlc = 0 do j=1,lc-1 lcx = local_index(istep*nbw+j, my_pcol, np_cols, nblk, 0) if (lcx>0) then nlc = nlc+1 a_mat(1:lr,lcx) = a_mat(1:lr,lcx) - conjg(tau)*aux2(nlc)*vr(1:lr) endif enddo if (wantDebug) call obj%timer%stop("mpi_communication") #else /* WITH_MPI */ ! if (nlc>0) aux2=aux1 ! Transform nlc = 0 do j=1,lc-1 lcx = local_index(istep*nbw+j, my_pcol, np_cols, nblk, 0) if (lcx>0) then nlc = nlc+1 a_mat(1:lr,lcx) = a_mat(1:lr,lcx) - conjg(tau)*aux1(nlc)*vr(1:lr) endif enddo #endif /* WITH_MPI */ ! ! ! Transform ! ! nlc = 0 ! do j=1,lc-1 ! lcx = local_index(istep*nbw+j, my_pcol, np_cols, nblk, 0) ! if (lcx>0) then ! nlc = nlc+1 ! a_mat(1:lr,lcx) = a_mat(1:lr,lcx) - conjg(tau)*aux2(nlc)*vr(1:lr) ! endif ! enddo #endif /* if 0 */ !Open up one omp region to avoid paying openmp overhead. !This does not help performance due to the addition of two openmp barriers around the MPI call, !But in the future this may be beneficial if these barriers are replaced with a faster implementation !$omp parallel private(mynlc, j, lcx, ii, pp ) shared(aux1) mynlc = 0 ! number of local columns !This loop does not have independent iterations, !'mynlc' is incremented each iteration, and it is difficult to remove this dependency !Thus each thread executes every iteration of the loop, except it only does the work if it 'owns' that iteration !That is, a thread only executes the work associated with an iteration if its thread id is congruent to !the iteration number modulo the number of threads do j=1,lc-1 lcx = local_index(istep*nbw+j, my_pcol, np_cols, nblk, 0) if (lcx>0 ) then mynlc = mynlc+1 if ( mod((j-1), omp_get_num_threads()) .eq. omp_get_thread_num() ) then if (lr>0) aux1(mynlc) = dot_product(vr(1:lr),a_mat(1:lr,lcx)) endif endif enddo ! Get global dot products !$omp barrier !$omp single #ifdef WITH_MPI if (wantDebug) call obj%timer%start("mpi_communication") if (mynlc>0) call mpi_allreduce(aux1, aux2, mynlc, MPI_MATH_DATATYPE_PRECISION, & MPI_SUM, mpi_comm_rows, mpierr) if (wantDebug) call obj%timer%stop("mpi_communication") #else /* WITH_MPI */ if (mynlc>0) aux2 = aux1 #endif /* WITH_MPI */ !$omp end single !$omp barrier ! Transform transformChunkSize=32 mynlc = 0 do j=1,lc-1 lcx = local_index(istep*nbw+j, my_pcol, np_cols, nblk, 0) if (lcx>0) then mynlc = mynlc+1 !This loop could be parallelized with an openmp pragma with static scheduling and chunk size 32 !However, for some reason this is slower than doing it manually, so it is parallelized as below. do ii=omp_get_thread_num()*transformChunkSize,lr,omp_get_num_threads()*transformChunkSize do pp = 1,transformChunkSize if (pp + ii > lr) exit #if REALCASE == 1 a_mat(ii+pp,lcx) = a_mat(ii+pp,lcx) - tau*aux2(mynlc)*vr(ii+pp) #endif #if COMPLEXCASE == 1 a_mat(ii+pp,lcx) = a_mat(ii+pp,lcx) - conjg(tau)*aux2(mynlc)*vr(ii+pp) #endif enddo enddo endif enddo !$omp end parallel #else /* WITH_OPENMP */ nlc = 0 ! number of local columns do j=1,lc-1 lcx = local_index(istep*nbw+j, my_pcol, np_cols, nblk, 0) if (lcx>0) then nlc = nlc+1 if (lr>0) aux1(nlc) = dot_product(vr(1:lr),a_mat(1:lr,lcx)) endif enddo ! Get global dot products #ifdef WITH_MPI if (wantDebug) call obj%timer%start("mpi_communication") if (nlc>0) call mpi_allreduce(aux1, aux2, nlc, MPI_MATH_DATATYPE_PRECISION, & MPI_SUM, mpi_comm_rows, mpierr) if (wantDebug) call obj%timer%stop("mpi_communication") #else /* WITH_MPI */ if (nlc>0) aux2=aux1 #endif /* WITH_MPI */ ! Transform nlc = 0 do j=1,lc-1 lcx = local_index(istep*nbw+j, my_pcol, np_cols, nblk, 0) if (lcx>0) then nlc = nlc+1 #if REALCASE == 1 a_mat(1:lr,lcx) = a_mat(1:lr,lcx) - tau*aux2(nlc)*vr(1:lr) #endif #if COMPLEXCASE == 1 a_mat(1:lr,lcx) = a_mat(1:lr,lcx) - conjg(tau)*aux2(nlc)*vr(1:lr) #endif endif enddo #endif /* WITH_OPENMP */ enddo ! lc if (useGPU_reduction_lower_block_to_tridiagonal) then ! store column tiles back to GPU cur_pcol = pcol(istep*nbw+1, nblk, np_cols) if (my_pcol == cur_pcol) then successCUDA = cuda_memcpy2d((a_dev+ & int(((lc_start-1)*lda*size_of_datatype),kind=c_intptr_t)), & int(lda*size_of_datatype,kind=c_intptr_t), loc(a_mat(1,lc_start)), & int(lda*size_of_datatype,kind=c_intptr_t), & int(lr_end*size_of_datatype,kind=c_intptr_t), & int((lc_end - lc_start+1),kind=c_intptr_t), & int(cudaMemcpyHostToDevice,kind=c_int)) if (.not.(successCUDA)) then print *, "bandred_& &MATH_DATATYPE& &: cuda memcpy a_dev failed ", istat stop 1 endif endif endif ! Calculate scalar products of stored Householder vectors. ! This can be done in different ways, we use dsyrk vav = 0 call obj%timer%start("blas") if (useGPU_reduction_lower_block_to_tridiagonal) then if (l_rows>0) & #if REALCASE == 1 call PRECISION_SYRK('U', 'T', & #endif #if COMPLEXCASE == 1 call PRECISION_HERK('U', 'C', & #endif n_cols, l_rows, ONE, & vmrCUDA, cur_l_rows, & ZERO, vav, ubound(vav,dim=1)) else ! useGPU_reduction_to_tridiagonal if (l_rows>0) & #if REALCASE == 1 call PRECISION_SYRK('U', 'T', & #endif #if COMPLEXCASE == 1 call PRECISION_HERK('U', 'C', & #endif n_cols, l_rows, ONE, vmrCPU, ubound(vmrCPU,dim=1), ZERO, vav, ubound(vav,dim=1)) endif call obj%timer%stop("blas") #if REALCASE == 1 call symm_matrix_allreduce_& #endif #if COMPLEXCASE == 1 call herm_matrix_allreduce_& #endif &PRECISION & (obj, n_cols,vav, nbw, nbw,mpi_comm_rows) ! Calculate triangular matrix T for block Householder Transformation call obj%timer%start("blas") do lc=n_cols,1,-1 tau = tmat(lc,lc,istep) if (lc vmc (stored in umc, second half) if (useGPU) then call elpa_transpose_vectors_& &MATH_DATATYPE& &_& &PRECISION & (obj, vmrCUDA, cur_l_rows, mpi_comm_rows, & umcCUDA(cur_l_cols * n_cols + 1), cur_l_cols, & mpi_comm_cols, 1, istep*nbw, n_cols, nblk, max_threads) else ! useGPU call elpa_transpose_vectors_& &MATH_DATATYPE& &_& &PRECISION & (obj, vmrCPU, ubound(vmrCPU,dim=1), mpi_comm_rows, & umcCPU(1,n_cols+1), ubound(umcCPU,dim=1), mpi_comm_cols, & 1, istep*nbw, n_cols, nblk, max_threads) endif ! Calculate umc = A**T * vmr ! Note that the distributed A has to be transposed ! Opposed to direct tridiagonalization there is no need to use the cache locality ! of the tiles, so we can use strips of the matrix #if 0 ! original complex implemetation check for performance umcCPU(1:l_cols,1:n_cols) = 0.0_rck vmrCPU(1:l_rows,n_cols+1:2*n_cols) = 0.0_rck if (l_cols>0 .and. l_rows>0) then do i=0,(istep*nbw-1)/tile_size lcs = i*l_cols_tile+1 lce = min(l_cols,(i+1)*l_cols_tile) if (lce0 .and. l_rows>0) #endif /* if 0 */ !Code for Algorithm 4 ! n_way is actually a branch for the number of OpenMP threads n_way = 1 #ifdef WITH_OPENMP #if REALCASE == 1 n_way = max_threads !$omp parallel private( i,lcs,lce,lrs,lre) #endif if (n_way > 1) then #if REALCASE == 1 !$omp do #endif do i=1,min(l_cols_tile, l_cols) umcCPU(i,1:n_cols) = 0.0_rck enddo #if REALCASE == 1 !$omp do #endif do i=1,l_rows vmrCPU(i,n_cols+1:2*n_cols) = 0.0_rck enddo if (l_cols>0 .and. l_rows>0) then !SYMM variant 4 !Partitioned Matrix Expression: ! Ct = Atl Bt + Atr Bb ! Cb = Atr' Bt + Abl Bb ! !Loop invariant: ! Ct = Atl Bt + Atr Bb ! !Update: ! C1 = A10'B0 + A11B1 + A21 B2 ! !This algorithm chosen because in this algoirhtm, the loop around the dgemm calls !is easily parallelized, and regardless of choise of algorithm, !the startup cost for parallelizing the dgemms inside the loop is too great #if REALCASE == 1 !$omp do schedule(static,1) #endif do i=0,(istep*nbw-1)/tile_size lcs = i*l_cols_tile+1 ! local column start lce = min(l_cols, (i+1)*l_cols_tile) ! local column end lrs = i*l_rows_tile+1 ! local row start lre = min(l_rows, (i+1)*l_rows_tile) ! local row end !C1 += [A11 A12] [B1 ! B2] if ( lre > lrs .and. l_cols > lcs ) then call obj%timer%start("blas") call PRECISION_GEMM('N', 'N', lre-lrs+1, n_cols, l_cols-lcs+1, & ONE, a_mat(lrs,lcs), ubound(a_mat,dim=1), & umcCPU(lcs,n_cols+1), ubound(umcCPU,dim=1), & ZERO, vmrCPU(lrs,n_cols+1), ubound(vmrCPU,dim=1)) call obj%timer%stop("blas") endif ! C1 += A10' B0 if ( lce > lcs .and. i > 0 ) then call obj%timer%start("blas") call PRECISION_GEMM(BLAS_TRANS_OR_CONJ, 'N', & lce-lcs+1, n_cols, lrs-1, & ONE, a_mat(1,lcs), ubound(a_mat,dim=1), & vmrCPU(1,1), ubound(vmrCPU,dim=1), & ZERO, umcCPU(lcs,1), ubound(umcCPU,dim=1)) call obj%timer%stop("blas") endif enddo endif ! l_cols>0 .and. l_rows>0 else ! n_way > 1 #endif /* WITH_OPENMP */ if (useGPU) then umcCUDA(1 : l_cols * n_cols) = 0.0_rck vmrCUDA(cur_l_rows * n_cols + 1 : cur_l_rows * n_cols * 2) = 0.0_rck else ! useGPU umcCPU(1:l_cols,1:n_cols) = 0.0_rck vmrCPU(1:l_rows,n_cols+1:2*n_cols) = 0.0_rck endif ! useGPU if (l_cols>0 .and. l_rows>0) then if (useGPU) then successCUDA = cuda_memcpy(vmr_dev, & loc(vmrCUDA(1)),& vmr_size*size_of_datatype,cudaMemcpyHostToDevice) if (.not.(successCUDA)) then print *,"bandred_& &MATH_DATATYPE& &: error in cudaMemcpy vmr_dev 3" stop 1 endif successCUDA = cuda_memcpy(umc_dev, & loc(umcCUDA(1)), & umc_size*size_of_datatype,cudaMemcpyHostToDevice) if (.not.(successCUDA)) then print *,"bandred_& &MATH_DATATYPE& &: error in cudaMemcpy umc_dev 3" stop 1 endif endif ! useGPU do i=0,(istep*nbw-1)/tile_size lcs = i*l_cols_tile+1 lce = min(l_cols,(i+1)*l_cols_tile) if (lce0 .and. l_rows>0 #ifdef WITH_OPENMP endif ! n_way > 1 #if REALCASE == 1 !$omp end parallel #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 ! Or if we used the Algorithm 4 if (tile_size < istep*nbw .or. n_way > 1) then if (useGPU) then call elpa_reduce_add_vectors_& &MATH_DATATYPE& &_& &PRECISION & (obj, vmrCUDA(cur_l_rows * n_cols + 1),cur_l_rows, & mpi_comm_rows, umcCUDA, & cur_l_cols, mpi_comm_cols, istep*nbw, n_cols, nblk, max_threads) else ! useGPU call elpa_reduce_add_vectors_& &MATH_DATATYPE& &_& &PRECISION & (obj, vmrCPU(1,n_cols+1),ubound(vmrCPU,dim=1),mpi_comm_rows, & umcCPU, ubound(umcCPU,dim=1), mpi_comm_cols, & istep*nbw, n_cols, nblk, max_threads) endif ! useGPU endif ! tile_size < istep*nbw .or. n_way > 1 if (l_cols>0) then if (useGPU) then #ifdef WITH_MPI allocate(tmpCUDA(l_cols * n_cols), stat=istat, errmsg=errorMessage) if (istat .ne. 0) then print *,"bandred_& &MATH_DATATYPE& &: error when allocating tmpCUDA "//errorMessage stop 1 endif if (wantDebug) call obj%timer%start("mpi_communication") call mpi_allreduce(umcCUDA, tmpCUDA, l_cols*n_cols, MPI_MATH_DATATYPE_PRECISION, & MPI_SUM, mpi_comm_rows, ierr) umcCUDA(1 : l_cols * n_cols) = tmpCUDA(1 : l_cols * n_cols) if (wantDebug) call obj%timer%stop("mpi_communication") #else /* WITH_MPI */ ! tmpCUDA(1 : l_cols * n_cols) = umcCUDA(1 : l_cols * n_cols) #endif /* WITH_MPI */ if (allocated(tmpCUDA)) then deallocate(tmpCUDA, stat=istat, errmsg=errorMessage) if (istat .ne. 0) then print *,"bandred_& &MATH_DATATYPE& &: error when deallocating tmpCUDA "//errorMessage stop 1 endif endif else ! useGPU allocate(tmpCPU(l_cols,n_cols), stat=istat, errmsg=errorMessage) if (istat .ne. 0) then print *,"bandred_& &MATH_DATATYPE& &: error when allocating tmpCPU "//errorMessage stop 1 endif #ifdef WITH_MPI if (wantDebug) call obj%timer%start("mpi_communication") call mpi_allreduce(umcCPU, tmpCPU, l_cols*n_cols, MPI_MATH_DATATYPE_PRECISION, & MPI_SUM, mpi_comm_rows, mpierr) umcCPU(1:l_cols,1:n_cols) = tmpCPU(1:l_cols,1:n_cols) if (wantDebug) call obj%timer%stop("mpi_communication") #else /* WITH_MPI */ ! tmpCPU(1:l_cols,1:n_cols) = umcCPU(1:l_cols,1:n_cols) #endif /* WITH_MPI */ deallocate(tmpCPU, stat=istat, errmsg=errorMessage) if (istat .ne. 0) then print *,"bandred_& &MATH_DATATYPE& &: error when deallocating tmpCPU "//errorMessage stop 1 endif endif ! useGPU endif ! l_cols > 0 ! U = U * Tmat**T if (useGPU) then successCUDA = cuda_memcpy(umc_dev, & loc(umcCUDA(1)), & umc_size*size_of_datatype, cudaMemcpyHostToDevice) if (.not.(successCUDA)) then print *,"bandred_& &MATH_DATATYPE& &: error in cudaMemcpy umc_dev 5" stop 1 endif successCUDA = cuda_memcpy(tmat_dev,loc(tmat(1,1,istep)),nbw*nbw*size_of_datatype,cudaMemcpyHostToDevice) if (.not.(successCUDA)) then print *,"bandred_& &MATH_DATATYPE& &: error in cudaMemcpy tmat_dev 2" stop 1 endif call obj%timer%start("cublas") call cublas_PRECISION_TRMM('Right', 'Upper', BLAS_TRANS_OR_CONJ, 'Nonunit', & l_cols, n_cols, ONE, tmat_dev, nbw, umc_dev, cur_l_cols) call obj%timer%stop("cublas") ! VAV = Tmat * V**T * A * V * Tmat**T = (U*Tmat**T)**T * V * Tmat**T successCUDA = cuda_memcpy(vav_dev,loc(vav(1,1)), nbw*nbw*size_of_datatype,cudaMemcpyHostToDevice) if (.not.(successCUDA)) then print *,"bandred_& &MATH_DATATYPE& &: error in cudaMemcpy vav_dev 2" stop 1 endif call obj%timer%start("cublas") call cublas_PRECISION_GEMM(BLAS_TRANS_OR_CONJ, 'N', & n_cols, n_cols, l_cols, ONE, umc_dev, cur_l_cols, & (umc_dev+(cur_l_cols * n_cols )*size_of_datatype),cur_l_cols, & ZERO, vav_dev, nbw) call cublas_PRECISION_TRMM('Right', 'Upper', BLAS_TRANS_OR_CONJ, 'Nonunit', & n_cols, n_cols, ONE, tmat_dev, nbw, vav_dev, nbw) call obj%timer%stop("cublas") successCUDA = cuda_memcpy(loc(vav(1,1)), vav_dev, nbw*nbw*size_of_datatype, cudaMemcpyDeviceToHost) if (.not.(successCUDA)) then print *,"bandred_& &MATH_DATATYPE& &: error in cudaMemcpy vav_dev3" stop 1 endif else ! useGPU call obj%timer%start("blas") call PRECISION_TRMM('Right', 'Upper', BLAS_TRANS_OR_CONJ, 'Nonunit', & l_cols,n_cols, ONE, tmat(1,1,istep), ubound(tmat,dim=1), & umcCPU, ubound(umcCPU,dim=1)) ! VAV = Tmat * V**T * A * V * Tmat**T = (U*Tmat**T)**T * V * Tmat**T call PRECISION_GEMM(BLAS_TRANS_OR_CONJ, 'N', & n_cols, n_cols, l_cols, ONE, umcCPU, ubound(umcCPU,dim=1), umcCPU(1,n_cols+1), & ubound(umcCPU,dim=1), ZERO, vav, ubound(vav,dim=1)) call PRECISION_TRMM('Right', 'Upper', BLAS_TRANS_OR_CONJ, 'Nonunit', & n_cols, n_cols, ONE, tmat(1,1,istep), & ubound(tmat,dim=1), vav, ubound(vav,dim=1)) call obj%timer%stop("blas") endif ! useGPU #if REALCASE == 1 call symm_matrix_allreduce_& #endif #if COMPLEXCASE == 1 call herm_matrix_allreduce_& #endif &PRECISION & (obj, n_cols,vav, nbw, nbw ,mpi_comm_cols) if (useGPU) then successCUDA = cuda_memcpy(vav_dev, loc(vav(1,1)), nbw*nbw*size_of_datatype,cudaMemcpyHostToDevice) if (.not.(successCUDA)) then print *,"bandred_& &MATH_DATATYPE& &: error in cudaMemcpy vav_dev4" stop 1 endif endif ! U = U - 0.5 * V * VAV if (useGPU) then call obj%timer%start("cublas") call cublas_PRECISION_GEMM('N', 'N', l_cols, n_cols, n_cols,& #if REALCASE == 1 -0.5_rk, & #endif #if COMPLEXCASE == 1 (-0.5_rk, 0.0_rk), & #endif (umc_dev+(cur_l_cols * n_cols )* & size_of_datatype), & cur_l_cols, vav_dev,nbw, & ONE, umc_dev, cur_l_cols) call obj%timer%stop("cublas") successCUDA = cuda_memcpy( & loc(umcCUDA(1)), & umc_dev, umc_size*size_of_datatype, cudaMemcpyDeviceToHost) if (.not.(successCUDA)) then print *,"bandred_& &MATH_DATATYPE& &: error in cudaMemcpy umc_dev 6" stop 1 endif ! Transpose umc -> umr (stored in vmr, second half) call elpa_transpose_vectors_& &MATH_DATATYPE& &_& &PRECISION & (obj, umcCUDA, cur_l_cols, mpi_comm_cols, & vmrCUDA(cur_l_rows * n_cols + 1), cur_l_rows, mpi_comm_rows, & 1, istep*nbw, n_cols, nblk, max_threads) successCUDA = cuda_memcpy(vmr_dev, & loc(vmrCUDA(1)), & vmr_size*size_of_datatype, cudaMemcpyHostToDevice) if (.not.(successCUDA)) then print *,"bandred_& &MATH_DATATYPE& &: error in cudaMemcpy vmr_dev 5 " stop 1 endif successCUDA = cuda_memcpy(umc_dev, & loc(umcCUDA(1)), & umc_size*size_of_datatype, cudaMemcpyHostToDevice) if (.not.(successCUDA)) then print *,"bandred_& &MATH_DATATYPE& &: error in cudaMemcpy umc_dev 7" stop 1 endif else ! useGPU call obj%timer%start("blas") call PRECISION_GEMM('N', 'N', l_cols, n_cols, n_cols, & #if REALCASE == 1 -0.5_rk, & #endif #if COMPLEXCASE == 1 (-0.5_rk, 0.0_rk), & #endif umcCPU(1,n_cols+1), ubound(umcCPU,dim=1), vav, & ubound(vav,dim=1), ONE, umcCPU, ubound(umcCPU,dim=1)) call obj%timer%stop("blas") ! Transpose umc -> umr (stored in vmr, second half) call elpa_transpose_vectors_& &MATH_DATATYPE& &_& &PRECISION & (obj, umcCPU, ubound(umcCPU,dim=1), mpi_comm_cols, & vmrCPU(1,n_cols+1), ubound(vmrCPU,dim=1), mpi_comm_rows, & 1, istep*nbw, n_cols, nblk, max_threads) endif ! useGPU ! A = A - V*U**T - U*V**T #ifdef WITH_OPENMP !$omp parallel private( ii, i, lcs, lce, lre, n_way, m_way, m_id, n_id, work_per_thread, mystart, myend ) n_threads = omp_get_num_threads() if (mod(n_threads, 2) == 0) then n_way = 2 else n_way = 1 endif m_way = n_threads / n_way m_id = mod(omp_get_thread_num(), m_way) n_id = omp_get_thread_num() / m_way do ii=n_id*tile_size,(istep*nbw-1),tile_size*n_way i = ii / tile_size lcs = i*l_cols_tile+1 lce = min(l_cols,(i+1)*l_cols_tile) lre = min(l_rows,(i+1)*l_rows_tile) if (lce lre ) myend = lre if ( myend-mystart+1 < 1) cycle call obj%timer%start("blas") call PRECISION_GEMM('N', BLAS_TRANS_OR_CONJ, myend-mystart+1, lce-lcs+1, 2*n_cols, -ONE, & vmrCPU(mystart, 1), ubound(vmrCPU,1), umcCPU(lcs,1), ubound(umcCPU,1), & ONE, a_mat(mystart,lcs), ubound(a_mat,1)) call obj%timer%stop("blas") enddo !$omp end parallel !#if COMPLEXCASE == 1 ! do i=0,(istep*nbw-1)/tile_size ! lcs = i*l_cols_tile+1 ! lce = min(l_cols,(i+1)*l_cols_tile) ! lre = min(l_rows,(i+1)*l_rows_tile) ! if (lce