!!$ !!$ Parallel Sparse BLAS v2.0 !!$ (C) Copyright 2006 Salvatore Filippone University of Rome Tor Vergata !!$ Alfredo Buttari University of Rome Tor Vergata !!$ !!$ Redistribution and use in source and binary forms, with or without !!$ modification, are permitted provided that the following conditions !!$ are met: !!$ 1. Redistributions of source code must retain the above copyright !!$ notice, this list of conditions and the following disclaimer. !!$ 2. Redistributions in binary form must reproduce the above copyright !!$ notice, this list of conditions, and the following disclaimer in the !!$ documentation and/or other materials provided with the distribution. !!$ 3. The name of the PSBLAS group or the names of its contributors may !!$ not be used to endorse or promote products derived from this !!$ software without specific written permission. !!$ !!$ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS !!$ ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED !!$ TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR !!$ PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE PSBLAS GROUP OR ITS CONTRIBUTORS !!$ BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR !!$ CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF !!$ SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS !!$ INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN !!$ CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) !!$ ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE !!$ POSSIBILITY OF SUCH DAMAGE. !!$ !!$ ! ! File: psi_dswapdata.F90 ! ! Subroutine: psi_dswapdatam ! Does the data exchange among processes. Essentially this is doing ! a variable all-to-all data exchange (ALLTOALLV in MPI parlance), but ! it is capable of pruning empty exchanges, which are very likely in out ! application environment. All the variants have the same structure ! In all these subroutines X may be: I Integer ! D real(kind(1.d0)) ! Z complex(kind(1.d0)) ! Basically the operation is as follows: on each process, we identify ! sections SND(Y) and RCV(Y); then we do a SEND(PACK(SND(Y))); ! then we receive, and we do an update with Y = UNPACK(RCV(Y)) + BETA * Y ! but only on the elements involved in the UNPACK operation. ! Thus: for halo data exchange, the receive section is confined in the ! halo indices, and BETA=0, whereas for overlap exchange the receive section ! is scattered in the owned indices, and BETA=1. ! ! Arguments: ! flag - integer Choose the algorithm for data exchange: ! this is chosen through bit fields. ! swap_mpi = iand(flag,psb_swap_mpi_) /= 0 ! swap_sync = iand(flag,psb_swap_sync_) /= 0 ! swap_send = iand(flag,psb_swap_send_) /= 0 ! swap_recv = iand(flag,psb_swap_recv_) /= 0 ! if (swap_mpi): use underlying MPI_ALLTOALLV. ! if (swap_sync): use PSB_SND and PSB_RCV in ! synchronized pairs ! if (swap_send .and. swap_recv): use mpi_irecv ! and mpi_send ! if (swap_send): use psb_snd (but need another ! call with swap_recv to complete) ! if (swap_recv): use psb_rcv (completing a ! previous call with swap_send) ! ! ! n - integer Number of columns in Y ! beta - X Choose overwrite or sum. ! y(:,:) - X The data area ! desc_a - type(psb_desc_type). The communication descriptor. ! work(:) - X Buffer space. If not sufficient, will do ! our own internal allocation. ! info - integer. return code. ! data - integer which list is to be used to exchange data ! default psb_comm_halo_ ! psb_comm_halo_ use halo_index ! psb_comm_ext_ use ext_index ! psb_comm_ovrl_ use ovrl_index ! ! ! subroutine psi_dswapdatam(flag,n,beta,y,desc_a,work,info,data) use psi_mod, psb_protect_name => psi_dswapdatam use psb_error_mod use psb_descriptor_type use psb_penv_mod !!$ use psi_gthsct_mod #ifdef MPI_MOD use mpi #endif implicit none #ifdef MPI_H include 'mpif.h' #endif integer, intent(in) :: flag, n integer, intent(out) :: info real(kind(1.d0)) :: y(:,:), beta real(kind(1.d0)), target :: work(:) type(psb_desc_type),target :: desc_a integer, optional :: data ! locals integer :: ictxt, np, me, nesd, nerv,& & proc_to_comm, p2ptag, icomm, p2pstat(mpi_status_size),& & idxs, idxr, iret, err_act, totxch, i, idx_pt,& & snd_pt, rcv_pt, pnti, data_ integer, allocatable, dimension(:) :: bsdidx, brvidx,& & sdsz, rvsz, prcid, rvhd, sdhd integer, pointer :: d_idx(:) integer :: int_err(5) logical :: swap_mpi, swap_sync, swap_send, swap_recv,& & albf,do_send,do_recv logical, parameter :: usersend=.false. real(kind(1.d0)), pointer, dimension(:) :: sndbuf, rcvbuf character(len=20) :: name info = 0 name='psi_swap_data' call psb_erractionsave(err_act) ictxt = psb_cd_get_context(desc_a) icomm = psb_cd_get_mpic(desc_a) call psb_info(ictxt,me,np) if (np == -1) then info = 2010 call psb_errpush(info,name) goto 9999 endif if (.not.psb_is_asb_desc(desc_a)) then info = 1122 call psb_errpush(info,name) goto 9999 endif swap_mpi = iand(flag,psb_swap_mpi_) /= 0 swap_sync = iand(flag,psb_swap_sync_) /= 0 swap_send = iand(flag,psb_swap_send_) /= 0 swap_recv = iand(flag,psb_swap_recv_) /= 0 do_send = swap_mpi .or. swap_sync .or. swap_send do_recv = swap_mpi .or. swap_sync .or. swap_recv if(present(data)) then data_ = data else data_ = psb_comm_halo_ end if select case(data_) case(psb_comm_halo_) d_idx => desc_a%halo_index totxch = desc_a%matrix_data(psb_thal_xch_) idxr = desc_a%matrix_data(psb_thal_rcv_) idxs = desc_a%matrix_data(psb_thal_snd_) case(psb_comm_ovr_) d_idx => desc_a%ovrlap_index totxch = desc_a%matrix_data(psb_tovr_xch_) idxr = desc_a%matrix_data(psb_tovr_rcv_) idxs = desc_a%matrix_data(psb_tovr_snd_) case(psb_comm_ext_) d_idx => desc_a%ext_index totxch = desc_a%matrix_data(psb_text_xch_) idxr = desc_a%matrix_data(psb_text_rcv_) idxs = desc_a%matrix_data(psb_text_snd_) case default call psb_errpush(4010,name,a_err='wrong Data selector') goto 9999 end select idxr = idxr * n idxs = idxs * n if (swap_mpi) then allocate(sdsz(0:np-1), rvsz(0:np-1), bsdidx(0:np-1),& & brvidx(0:np-1), rvhd(0:np-1), sdhd(0:np-1), prcid(0:np-1),& & stat=info) if(info /= 0) then call psb_errpush(4000,name) goto 9999 end if rvhd(:) = mpi_request_null sdsz(:) = 0 rvsz(:) = 0 ! prepare info for communications pnti = 1 snd_pt = 1 rcv_pt = 1 do i=1, totxch proc_to_comm = d_idx(pnti+psb_proc_id_) nerv = d_idx(pnti+psb_n_elem_recv_) nesd = d_idx(pnti+nerv+psb_n_elem_send_) call psb_get_rank(prcid(proc_to_comm),ictxt,proc_to_comm) brvidx(proc_to_comm) = rcv_pt rvsz(proc_to_comm) = n*nerv bsdidx(proc_to_comm) = snd_pt sdsz(proc_to_comm) = n*nesd rcv_pt = rcv_pt + n*nerv snd_pt = snd_pt + n*nesd pnti = pnti + nerv + nesd + 3 end do else allocate(rvhd(totxch),prcid(totxch),stat=info) if(info /= 0) then call psb_errpush(4000,name) goto 9999 end if end if idxr = max(idxr,1) idxs = max(idxs,1) if((idxr+idxs) < size(work)) then sndbuf => work(1:idxs) rcvbuf => work(idxs+1:idxs+idxr) albf=.false. else allocate(sndbuf(idxs),rcvbuf(idxr), stat=info) if(info /= 0) then call psb_errpush(4000,name) goto 9999 end if albf=.true. end if if (do_send) then ! Pack send buffers pnti = 1 snd_pt = 1 do i=1, totxch nerv = d_idx(pnti+psb_n_elem_recv_) nesd = d_idx(pnti+nerv+psb_n_elem_send_) idx_pt = 1+pnti+nerv+psb_n_elem_send_ call psi_gth(nesd,n,d_idx(idx_pt:idx_pt+nesd-1),& & y,sndbuf(snd_pt:snd_pt+n*nesd-1)) snd_pt = snd_pt + n*nesd pnti = pnti + nerv + nesd + 3 end do end if ! Case SWAP_MPI if (swap_mpi) then ! swap elements using mpi_alltoallv call mpi_alltoallv(sndbuf,sdsz,bsdidx,& & mpi_double_precision,rcvbuf,rvsz,& & brvidx,mpi_double_precision,icomm,iret) if(iret /= mpi_success) then int_err(1) = iret info=400 call psb_errpush(info,name,i_err=int_err) goto 9999 end if else if (swap_sync) then pnti = 1 snd_pt = 1 rcv_pt = 1 do i=1, totxch proc_to_comm = d_idx(pnti+psb_proc_id_) nerv = d_idx(pnti+psb_n_elem_recv_) nesd = d_idx(pnti+nerv+psb_n_elem_send_) if (proc_to_comm < me) then if (nesd>0) call psb_snd(ictxt,& & sndbuf(snd_pt:snd_pt+n*nesd-1), proc_to_comm) if (nerv>0) call psb_rcv(ictxt,& & rcvbuf(rcv_pt:rcv_pt+n*nerv-1), proc_to_comm) else if (proc_to_comm > me) then if (nerv>0) call psb_rcv(ictxt,& & rcvbuf(rcv_pt:rcv_pt+n*nerv-1), proc_to_comm) if (nesd>0) call psb_snd(ictxt,& & sndbuf(snd_pt:snd_pt+n*nesd-1), proc_to_comm) end if rcv_pt = rcv_pt + n*nerv snd_pt = snd_pt + n*nesd pnti = pnti + nerv + nesd + 3 end do else if (swap_send .and. swap_recv) then ! First I post all the non blocking receives pnti = 1 snd_pt = 1 rcv_pt = 1 do i=1, totxch proc_to_comm = d_idx(pnti+psb_proc_id_) nerv = d_idx(pnti+psb_n_elem_recv_) nesd = d_idx(pnti+nerv+psb_n_elem_send_) call psb_get_rank(prcid(i),ictxt,proc_to_comm) if (nerv>0) then p2ptag = krecvid(ictxt,proc_to_comm,me) call mpi_irecv(rcvbuf(rcv_pt),n*nerv,& & mpi_double_precision,prcid(i),& & p2ptag, icomm,rvhd(i),iret) end if rcv_pt = rcv_pt + n*nerv snd_pt = snd_pt + n*nesd pnti = pnti + nerv + nesd + 3 end do ! Then I post all the blocking sends if (usersend) call mpi_barrier(icomm,info) pnti = 1 snd_pt = 1 rcv_pt = 1 do i=1, totxch proc_to_comm = d_idx(pnti+psb_proc_id_) nerv = d_idx(pnti+psb_n_elem_recv_) nesd = d_idx(pnti+nerv+psb_n_elem_send_) p2ptag=ksendid(ictxt,proc_to_comm,me) if (nesd>0) then if (usersend) then call mpi_rsend(sndbuf(snd_pt),n*nesd,& & mpi_double_precision,prcid(i),& & p2ptag,icomm,iret) else call mpi_send(sndbuf(snd_pt),n*nesd,& & mpi_double_precision,prcid(i),& & p2ptag,icomm,iret) end if if(iret /= mpi_success) then int_err(1) = iret info=400 call psb_errpush(info,name,i_err=int_err) goto 9999 end if end if rcv_pt = rcv_pt + n*nerv snd_pt = snd_pt + n*nesd pnti = pnti + nerv + nesd + 3 end do pnti = 1 do i=1, totxch proc_to_comm = d_idx(pnti+psb_proc_id_) nerv = d_idx(pnti+psb_n_elem_recv_) nesd = d_idx(pnti+nerv+psb_n_elem_send_) p2ptag = krecvid(ictxt,proc_to_comm,me) if ((proc_to_comm /= me).and.(nerv>0)) then call mpi_wait(rvhd(i),p2pstat,iret) if(iret /= mpi_success) then int_err(1) = iret info=400 call psb_errpush(info,name,i_err=int_err) goto 9999 end if end if pnti = pnti + nerv + nesd + 3 end do else if (swap_send) then pnti = 1 snd_pt = 1 rcv_pt = 1 do i=1, totxch proc_to_comm = d_idx(pnti+psb_proc_id_) nerv = d_idx(pnti+psb_n_elem_recv_) nesd = d_idx(pnti+nerv+psb_n_elem_send_) if (nesd>0) call psb_snd(ictxt,& & sndbuf(snd_pt:snd_pt+n*nesd-1), proc_to_comm) rcv_pt = rcv_pt + n*nerv snd_pt = snd_pt + n*nesd pnti = pnti + nerv + nesd + 3 end do else if (swap_recv) then pnti = 1 snd_pt = 1 rcv_pt = 1 do i=1, totxch proc_to_comm = d_idx(pnti+psb_proc_id_) nerv = d_idx(pnti+psb_n_elem_recv_) nesd = d_idx(pnti+nerv+psb_n_elem_send_) if (nerv>0) call psb_rcv(ictxt,& & rcvbuf(rcv_pt:rcv_pt+n*nerv-1), proc_to_comm) rcv_pt = rcv_pt + n*nerv snd_pt = snd_pt + n*nesd pnti = pnti + nerv + nesd + 3 end do end if if (do_recv) then pnti = 1 snd_pt = 1 rcv_pt = 1 do i=1, totxch proc_to_comm = d_idx(pnti+psb_proc_id_) nerv = d_idx(pnti+psb_n_elem_recv_) nesd = d_idx(pnti+nerv+psb_n_elem_send_) idx_pt = 1+pnti+psb_n_elem_recv_ call psi_sct(nerv,n,d_idx(idx_pt:idx_pt+nerv-1),& & rcvbuf(rcv_pt:rcv_pt+n*nerv-1),beta,y) rcv_pt = rcv_pt + n*nerv snd_pt = snd_pt + n*nesd pnti = pnti + nerv + nesd + 3 end do end if if (swap_mpi) then deallocate(sdsz,rvsz,bsdidx,brvidx,rvhd,prcid,sdhd,& & stat=info) else deallocate(rvhd,prcid,stat=info) end if if(info /= 0) then call psb_errpush(4000,name) goto 9999 end if if(albf) deallocate(sndbuf,rcvbuf,stat=info) if(info /= 0) then call psb_errpush(4000,name) goto 9999 end if call psb_erractionrestore(err_act) return 9999 continue call psb_erractionrestore(err_act) if (err_act == psb_act_abort_) then call psb_error(ictxt) return end if return end subroutine psi_dswapdatam !!$ !!$ Parallel Sparse BLAS v2.0 !!$ (C) Copyright 2006 Salvatore Filippone University of Rome Tor Vergata !!$ Alfredo Buttari University of Rome Tor Vergata !!$ !!$ Redistribution and use in source and binary forms, with or without !!$ modification, are permitted provided that the following conditions !!$ are met: !!$ 1. Redistributions of source code must retain the above copyright !!$ notice, this list of conditions and the following disclaimer. !!$ 2. Redistributions in binary form must reproduce the above copyright !!$ notice, this list of conditions, and the following disclaimer in the !!$ documentation and/or other materials provided with the distribution. !!$ 3. The name of the PSBLAS group or the names of its contributors may !!$ not be used to endorse or promote products derived from this !!$ software without specific written permission. !!$ !!$ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS !!$ ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED !!$ TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR !!$ PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE PSBLAS GROUP OR ITS CONTRIBUTORS !!$ BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR !!$ CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF !!$ SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS !!$ INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN !!$ CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) !!$ ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE !!$ POSSIBILITY OF SUCH DAMAGE. !!$ !!$ ! ! ! Subroutine: psi_dswapdatav ! Does the data exchange among processes. Essentially this is doing ! a variable all-to-all data exchange (ALLTOALLV in MPI parlance), but ! it is capable of pruning empty exchanges, which are very likely in out ! application environment. All the variants have the same structure ! In all these subroutines X may be: I Integer ! D real(kind(1.d0)) ! Z complex(kind(1.d0)) ! Basically the operation is as follows: on each process, we identify ! sections SND(Y) and RCV(Y); then we do a SEND(PACK(SND(Y))); ! then we receive, and we do an update with Y = UNPACK(RCV(Y)) + BETA * Y ! but only on the elements involved in the UNPACK operation. ! Thus: for halo data exchange, the receive section is confined in the ! halo indices, and BETA=0, whereas for overlap exchange the receive section ! is scattered in the owned indices, and BETA=1. ! ! Arguments: ! flag - integer Choose the algorithm for data exchange: ! this is chosen through bit fields. ! swap_mpi = iand(flag,psb_swap_mpi_) /= 0 ! swap_sync = iand(flag,psb_swap_sync_) /= 0 ! swap_send = iand(flag,psb_swap_send_) /= 0 ! swap_recv = iand(flag,psb_swap_recv_) /= 0 ! if (swap_mpi): use underlying MPI_ALLTOALLV. ! if (swap_sync): use PSB_SND and PSB_RCV in ! synchronized pairs ! if (swap_send .and. swap_recv): use mpi_irecv ! and mpi_send ! if (swap_send): use psb_snd (but need another ! call with swap_recv to complete) ! if (swap_recv): use psb_rcv (completing a ! previous call with swap_send) ! ! ! n - integer Number of columns in Y ! beta - X Choose overwrite or sum. ! y(:) - X The data area ! desc_a - type(psb_desc_type). The communication descriptor. ! work(:) - X Buffer space. If not sufficient, will do ! our own internal allocation. ! info - integer. return code. ! data - integer which list is to be used to exchange data ! default psb_comm_halo_ ! psb_comm_halo_ use halo_index ! psb_comm_ext_ use ext_index ! psb_comm_ovrl_ use ovrl_index ! ! subroutine psi_dswapdatav(flag,beta,y,desc_a,work,info,data) use psi_mod, psb_protect_name => psi_dswapdatav use psb_error_mod use psb_descriptor_type use psb_penv_mod !!$ use psi_gthsct_mod #ifdef MPI_MOD use mpi #endif implicit none #ifdef MPI_H include 'mpif.h' #endif integer, intent(in) :: flag integer, intent(out) :: info real(kind(1.d0)) :: y(:), beta real(kind(1.d0)), target :: work(:) type(psb_desc_type),target :: desc_a integer, optional :: data ! locals integer :: ictxt, np, me, nesd, nerv,& & proc_to_comm, p2ptag, icomm, p2pstat(mpi_status_size),& & idxs, idxr, iret, err_act, totxch, i, & & idx_pt, snd_pt, rcv_pt, n, pnti, data_ integer, allocatable, dimension(:) :: bsdidx, brvidx,& & sdsz, rvsz, prcid, rvhd, sdhd integer, pointer :: d_idx(:) integer :: int_err(5) logical :: swap_mpi, swap_sync, swap_send, swap_recv,& & albf,do_send,do_recv logical, parameter :: usersend=.false. real(kind(1.d0)), pointer, dimension(:) :: sndbuf, rcvbuf character(len=20) :: name info = 0 name='psi_swap_datav' call psb_erractionsave(err_act) ictxt=psb_cd_get_context(desc_a) call psb_info(ictxt,me,np) if (np == -1) then info = 2010 call psb_errpush(info,name) goto 9999 endif if (.not.psb_is_asb_desc(desc_a)) then info = 1122 call psb_errpush(info,name) goto 9999 endif icomm = desc_a%matrix_data(psb_mpi_c_) n=1 swap_mpi = iand(flag,psb_swap_mpi_) /= 0 swap_sync = iand(flag,psb_swap_sync_) /= 0 swap_send = iand(flag,psb_swap_send_) /= 0 swap_recv = iand(flag,psb_swap_recv_) /= 0 do_send = swap_mpi .or. swap_sync .or. swap_send do_recv = swap_mpi .or. swap_sync .or. swap_recv if(present(data)) then data_ = data else data_ = psb_comm_halo_ end if select case(data_) case(psb_comm_halo_) d_idx => desc_a%halo_index totxch = desc_a%matrix_data(psb_thal_xch_) idxr = desc_a%matrix_data(psb_thal_rcv_) idxs = desc_a%matrix_data(psb_thal_snd_) case(psb_comm_ovr_) d_idx => desc_a%ovrlap_index totxch = desc_a%matrix_data(psb_tovr_xch_) idxr = desc_a%matrix_data(psb_tovr_rcv_) idxs = desc_a%matrix_data(psb_tovr_snd_) case(psb_comm_ext_) d_idx => desc_a%ext_index totxch = desc_a%matrix_data(psb_text_xch_) idxr = desc_a%matrix_data(psb_text_rcv_) idxs = desc_a%matrix_data(psb_text_snd_) case default call psb_errpush(4010,name,a_err='wrong Data selector') goto 9999 end select idxr = idxr * n idxs = idxs * n if (swap_mpi) then allocate(sdsz(0:np-1), rvsz(0:np-1), bsdidx(0:np-1),& & brvidx(0:np-1), rvhd(0:np-1), sdhd(0:np-1), prcid(0:np-1),& & stat=info) if(info /= 0) then call psb_errpush(4000,name) goto 9999 end if rvhd(:) = mpi_request_null sdsz(:) = 0 rvsz(:) = 0 ! prepare info for communications pnti = 1 snd_pt = 1 rcv_pt = 1 do i=1, totxch proc_to_comm = d_idx(pnti+psb_proc_id_) nerv = d_idx(pnti+psb_n_elem_recv_) nesd = d_idx(pnti+nerv+psb_n_elem_send_) call psb_get_rank(prcid(proc_to_comm),ictxt,proc_to_comm) brvidx(proc_to_comm) = rcv_pt rvsz(proc_to_comm) = nerv bsdidx(proc_to_comm) = snd_pt sdsz(proc_to_comm) = nesd rcv_pt = rcv_pt + nerv snd_pt = snd_pt + nesd pnti = pnti + nerv + nesd + 3 end do else allocate(rvhd(totxch),prcid(totxch),stat=info) if(info /= 0) then call psb_errpush(4000,name) goto 9999 end if end if idxr = max(idxr,1) idxs = max(idxs,1) if((idxr+idxs) < size(work)) then sndbuf => work(1:idxs) rcvbuf => work(idxs+1:idxs+idxr) albf=.false. else allocate(sndbuf(idxs),rcvbuf(idxr), stat=info) if(info /= 0) then call psb_errpush(4000,name) goto 9999 end if albf=.true. end if if (do_send) then ! Pack send buffers pnti = 1 snd_pt = 1 do i=1, totxch nerv = d_idx(pnti+psb_n_elem_recv_) nesd = d_idx(pnti+nerv+psb_n_elem_send_) idx_pt = 1+pnti+nerv+psb_n_elem_send_ call psi_gth(nesd,d_idx(idx_pt:idx_pt+nesd-1),& & y,sndbuf(snd_pt:snd_pt+nesd-1)) snd_pt = snd_pt + nesd pnti = pnti + nerv + nesd + 3 end do end if ! Case SWAP_MPI if (swap_mpi) then ! swap elements using mpi_alltoallv call mpi_alltoallv(sndbuf,sdsz,bsdidx,& & mpi_double_precision,rcvbuf,rvsz,& & brvidx,mpi_double_precision,icomm,iret) if(iret /= mpi_success) then int_err(1) = iret info=400 call psb_errpush(info,name,i_err=int_err) goto 9999 end if else if (swap_sync) then pnti = 1 snd_pt = 1 rcv_pt = 1 do i=1, totxch proc_to_comm = d_idx(pnti+psb_proc_id_) nerv = d_idx(pnti+psb_n_elem_recv_) nesd = d_idx(pnti+nerv+psb_n_elem_send_) if (proc_to_comm < me) then if (nesd>0) call psb_snd(ictxt,& & sndbuf(snd_pt:snd_pt+nesd-1), proc_to_comm) if (nerv>0) call psb_rcv(ictxt,& & rcvbuf(rcv_pt:rcv_pt+nerv-1), proc_to_comm) else if (proc_to_comm > me) then if (nerv>0) call psb_rcv(ictxt,& & rcvbuf(rcv_pt:rcv_pt+nerv-1), proc_to_comm) if (nesd>0) call psb_snd(ictxt,& & sndbuf(snd_pt:snd_pt+nesd-1), proc_to_comm) end if rcv_pt = rcv_pt + nerv snd_pt = snd_pt + nesd pnti = pnti + nerv + nesd + 3 end do else if (swap_send .and. swap_recv) then ! First I post all the non blocking receives pnti = 1 snd_pt = 1 rcv_pt = 1 do i=1, totxch proc_to_comm = d_idx(pnti+psb_proc_id_) nerv = d_idx(pnti+psb_n_elem_recv_) nesd = d_idx(pnti+nerv+psb_n_elem_send_) call psb_get_rank(prcid(i),ictxt,proc_to_comm) if (nerv>0) then p2ptag = krecvid(ictxt,proc_to_comm,me) call mpi_irecv(rcvbuf(rcv_pt),nerv,& & mpi_double_precision,prcid(i),& & p2ptag, icomm,rvhd(i),iret) end if rcv_pt = rcv_pt + nerv snd_pt = snd_pt + nesd pnti = pnti + nerv + nesd + 3 end do ! Then I post all the blocking sends if (usersend) call mpi_barrier(icomm,info) pnti = 1 snd_pt = 1 rcv_pt = 1 do i=1, totxch proc_to_comm = d_idx(pnti+psb_proc_id_) nerv = d_idx(pnti+psb_n_elem_recv_) nesd = d_idx(pnti+nerv+psb_n_elem_send_) p2ptag=ksendid(ictxt,proc_to_comm,me) if (nesd>0) then if (usersend) then call mpi_rsend(sndbuf(snd_pt),nesd,& & mpi_double_precision,prcid(i),& & p2ptag,icomm,iret) else call mpi_send(sndbuf(snd_pt),nesd,& & mpi_double_precision,prcid(i),& & p2ptag,icomm,iret) end if if(iret /= mpi_success) then int_err(1) = iret info=400 call psb_errpush(info,name,i_err=int_err) goto 9999 end if end if rcv_pt = rcv_pt + nerv snd_pt = snd_pt + nesd pnti = pnti + nerv + nesd + 3 end do pnti = 1 do i=1, totxch proc_to_comm = d_idx(pnti+psb_proc_id_) nerv = d_idx(pnti+psb_n_elem_recv_) nesd = d_idx(pnti+nerv+psb_n_elem_send_) p2ptag = krecvid(ictxt,proc_to_comm,me) if ((proc_to_comm /= me).and.(nerv>0)) then call mpi_wait(rvhd(i),p2pstat,iret) if(iret /= mpi_success) then int_err(1) = iret info=400 call psb_errpush(info,name,i_err=int_err) goto 9999 end if end if pnti = pnti + nerv + nesd + 3 end do else if (swap_send) then pnti = 1 snd_pt = 1 rcv_pt = 1 do i=1, totxch proc_to_comm = d_idx(pnti+psb_proc_id_) nerv = d_idx(pnti+psb_n_elem_recv_) nesd = d_idx(pnti+nerv+psb_n_elem_send_) if (nesd>0) call psb_snd(ictxt,& & sndbuf(snd_pt:snd_pt+nesd-1), proc_to_comm) rcv_pt = rcv_pt + nerv snd_pt = snd_pt + nesd pnti = pnti + nerv + nesd + 3 end do else if (swap_recv) then pnti = 1 snd_pt = 1 rcv_pt = 1 do i=1, totxch proc_to_comm = d_idx(pnti+psb_proc_id_) nerv = d_idx(pnti+psb_n_elem_recv_) nesd = d_idx(pnti+nerv+psb_n_elem_send_) if (nerv>0) call psb_rcv(ictxt,& & rcvbuf(rcv_pt:rcv_pt+nerv-1), proc_to_comm) rcv_pt = rcv_pt + nerv snd_pt = snd_pt + nesd pnti = pnti + nerv + nesd + 3 end do end if if (do_recv) then pnti = 1 snd_pt = 1 rcv_pt = 1 do i=1, totxch proc_to_comm = d_idx(pnti+psb_proc_id_) nerv = d_idx(pnti+psb_n_elem_recv_) nesd = d_idx(pnti+nerv+psb_n_elem_send_) idx_pt = 1+pnti+psb_n_elem_recv_ call psi_sct(nerv,d_idx(idx_pt:idx_pt+nerv-1),& & rcvbuf(rcv_pt:rcv_pt+nerv-1),beta,y) rcv_pt = rcv_pt + nerv snd_pt = snd_pt + nesd pnti = pnti + nerv + nesd + 3 end do end if if (swap_mpi) then deallocate(sdsz,rvsz,bsdidx,brvidx,rvhd,prcid,sdhd,& & stat=info) else deallocate(rvhd,prcid,stat=info) end if if(info /= 0) then call psb_errpush(4000,name) goto 9999 end if if(albf) deallocate(sndbuf,rcvbuf,stat=info) if(info /= 0) then call psb_errpush(4000,name) goto 9999 end if call psb_erractionrestore(err_act) return 9999 continue call psb_erractionrestore(err_act) if (err_act == psb_act_abort_) then call psb_error(ictxt) return end if return end subroutine psi_dswapdatav