diff --git a/src/internals/psi_dswaptran.f90 b/src/internals/psi_dswaptran.f90 index f5dd92bb..104b9c92 100644 --- a/src/internals/psi_dswaptran.f90 +++ b/src/internals/psi_dswaptran.f90 @@ -7,8 +7,8 @@ subroutine psi_dswaptranm(flag,n,beta,y,desc_a,work,info,data) integer, intent(in) :: flag, n integer, intent(out) :: info - real(kind(1.d0)) :: y(:,:), beta - real(kind(1.d0)), target :: work(:) + integer :: y(:,:), beta + integer, target :: work(:) type(psb_desc_type) :: desc_a integer, optional :: data @@ -25,7 +25,7 @@ subroutine psi_dswaptranm(flag,n,beta,y,desc_a,work,info,data) integer :: int_err(5) integer :: blacs_pnum, krecvid, ksendid logical :: swap_mpi, swap_sync, swap_send, swap_recv - real(kind(1.d0)), pointer, dimension(:) :: sndbuf, rcvbuf + integer, pointer, dimension(:) :: sndbuf, rcvbuf character(len=20) :: name, ch_err interface psi_gth @@ -151,14 +151,14 @@ subroutine psi_dswaptranm(flag,n,beta,y,desc_a,work,info,data) ! Case SWAP_MPI if(swap_mpi) then - + ! gather elements into sendbuffer for swapping point_to_proc = 1 proc_to_comm = d_idx(point_to_proc+psb_proc_id_) do while (proc_to_comm .ne. -1) nerv = d_idx(point_to_proc+psb_n_elem_recv_) nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) - + idx_pt = point_to_proc+psb_elem_recv_ rcv_pt = brvidx(proc_to_comm) @@ -171,8 +171,8 @@ subroutine psi_dswaptranm(flag,n,beta,y,desc_a,work,info,data) ! swap elements using mpi_alltoallv call mpi_alltoallv(rcvbuf,rvsz,brvidx,& - & mpi_double_precision,sndbuf,sdsz,& - & bsdidx,mpi_double_precision,icomm,iret) + & mpi_integer,sndbuf,sdsz,& + & bsdidx,mpi_integer,icomm,iret) if(iret.ne.mpi_success) then int_err(1) = iret info=400 @@ -186,7 +186,7 @@ subroutine psi_dswaptranm(flag,n,beta,y,desc_a,work,info,data) do while (proc_to_comm .ne. -1) nerv = d_idx(point_to_proc+psb_n_elem_recv_) nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) - + idx_pt = point_to_proc+nerv+psb_elem_send_ snd_pt = bsdidx(proc_to_comm) call psi_sct(nesd,n,d_idx(idx_pt:idx_pt+nesd-1),& @@ -209,20 +209,20 @@ subroutine psi_dswaptranm(flag,n,beta,y,desc_a,work,info,data) rcv_pt = brvidx(proc_to_comm) call psi_gth(nerv,n,d_idx(idx_pt:idx_pt+nerv-1),& & y,rcvbuf(rcv_pt:rcv_pt+nerv*n-1)) - call dgesd2d(icontxt,nerv,n,rcvbuf(rcv_pt),nerv,proc_to_comm,0) + call igesd2d(icontxt,nerv,n,rcvbuf(rcv_pt),nerv,proc_to_comm,0) ! Then I receive snd_pt = brvidx(proc_to_comm) - call dgerv2d(icontxt,nesd,n,sndbuf(snd_pt),nesd,proc_to_comm,0) + call igerv2d(icontxt,nesd,n,sndbuf(snd_pt),nesd,proc_to_comm,0) else if (proc_to_comm .gt. myrow) then ! First I receive snd_pt = bsdidx(proc_to_comm) - call dgerv2d(icontxt,nesd,n,sndbuf(snd_pt),nesd,proc_to_comm,0) + call igerv2d(icontxt,nesd,n,sndbuf(snd_pt),nesd,proc_to_comm,0) ! Then I send idx_pt = point_to_proc+psb_elem_recv_ rcv_pt = brvidx(proc_to_comm) call psi_gth(nerv,n,d_idx(idx_pt:idx_pt+nerv-1),& & y,rcvbuf(rcv_pt:rcv_pt+nerv*n-1)) - call dgesd2d(icontxt,nerv,n,rcvbuf(rcv_pt),nerv,proc_to_comm,0) + call igesd2d(icontxt,nerv,n,rcvbuf(rcv_pt),nerv,proc_to_comm,0) else if (proc_to_comm .eq. myrow) then ! I send to myself idx_pt = point_to_proc+psb_elem_recv_ @@ -270,7 +270,7 @@ subroutine psi_dswaptranm(flag,n,beta,y,desc_a,work,info,data) p2ptag = krecvid(icontxt,proc_to_comm,myrow) snd_pt = brvidx(proc_to_comm) call mpi_irecv(sndbuf(rcv_pt),sdsz(proc_to_comm),& - & mpi_double_precision,prcid(proc_to_comm),& + & mpi_integer,prcid(proc_to_comm),& & p2ptag, icomm,rvhd(proc_to_comm),iret) if(iret.ne.mpi_success) then int_err(1) = iret @@ -290,16 +290,16 @@ subroutine psi_dswaptranm(flag,n,beta,y,desc_a,work,info,data) do while (proc_to_comm .ne. -1) nerv = d_idx(point_to_proc+psb_n_elem_recv_) nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) - + idx_pt = point_to_proc+psb_elem_recv_ rcv_pt = brvidx(proc_to_comm) call psi_gth(nerv,n,d_idx(idx_pt:idx_pt+nerv-1),& & y,rcvbuf(rcv_pt:rcv_pt+nerv*n-1)) - + if(proc_to_comm .ne. myrow) then p2ptag=ksendid(icontxt,proc_to_comm,myrow) call mpi_send(rcvbuf(rcv_pt),rvsz(proc_to_comm),& - & mpi_double_precision,prcid(proc_to_comm),& + & mpi_integer,prcid(proc_to_comm),& & p2ptag,icomm,iret) if(iret.ne.mpi_success) then int_err(1) = iret @@ -312,50 +312,85 @@ subroutine psi_dswaptranm(flag,n,beta,y,desc_a,work,info,data) proc_to_comm = d_idx(point_to_proc+psb_proc_id_) end do - do i=1, totxch - call mpi_waitany(nprow,rvhd,ixrec,p2pstat,iret) - if(iret.ne.mpi_success) then - int_err(1) = iret - info=400 - call psb_errpush(info,name,i_err=int_err) - goto 9999 - end if + if(.false.) then + do i=1, totxch + call mpi_waitany(nprow,rvhd,ixrec,p2pstat,iret) + if(iret.ne.mpi_success) then + int_err(1) = iret + info=400 + call psb_errpush(info,name,i_err=int_err) + goto 9999 + end if - if (ixrec .ne. mpi_undefined) then - ixrec=ixrec-1 ! mpi_waitany returns an 1 to nprow index - point_to_proc = ptp(ixrec) - proc_to_comm = d_idx(point_to_proc+psb_proc_id_) + if (ixrec .ne. mpi_undefined) then + ixrec=ixrec-1 ! mpi_waitany returns an 1 to nprow index + point_to_proc = ptp(ixrec) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) + + idx_pt = point_to_proc+nerv+psb_elem_send_ + snd_pt = bsdidx(proc_to_comm) + call psi_sct(nesd,n,d_idx(idx_pt:idx_pt+nesd-1),& + & sndbuf(snd_pt:snd_pt+n*nesd-1),beta,y) + else + int_err(1) = ixrec + info=400 + call psb_errpush(info,name,i_err=int_err) + goto 9999 + end if + end do + + point_to_proc = 1 + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) + do while (proc_to_comm .ne. -1) nerv = d_idx(point_to_proc+psb_n_elem_recv_) nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) - idx_pt = point_to_proc+nerv+psb_elem_send_ - snd_pt = bsdidx(proc_to_comm) - call psi_sct(nesd,n,d_idx(idx_pt:idx_pt+nesd-1),& - & sndbuf(snd_pt:snd_pt+n*nesd-1),beta,y) - else - int_err(1) = ixrec - info=400 - call psb_errpush(info,name,i_err=int_err) - goto 9999 - end if - end do - - point_to_proc = 1 - proc_to_comm = d_idx(point_to_proc+psb_proc_id_) - do while (proc_to_comm .ne. -1) - nerv = d_idx(point_to_proc+psb_n_elem_recv_) - nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) - - if(proc_to_comm .eq. myrow) then - idx_pt = point_to_proc+nerv+psb_elem_send_ - rcv_pt = brvidx(proc_to_comm) - call psi_sct(nesd,n,d_idx(idx_pt:idx_pt+nesd-1),& - & rcvbuf(rcv_pt:rcv_pt+n*nerv-1),beta,y) - end if + if(proc_to_comm .eq. myrow) then + idx_pt = point_to_proc+nerv+psb_elem_send_ + rcv_pt = brvidx(proc_to_comm) + call psi_sct(nesd,n,d_idx(idx_pt:idx_pt+nesd-1),& + & rcvbuf(rcv_pt:rcv_pt+n*nerv-1),beta,y) + end if - point_to_proc = point_to_proc+nerv+nesd+3 + point_to_proc = point_to_proc+nerv+nesd+3 + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) + end do + else + + point_to_proc = 1 proc_to_comm = d_idx(point_to_proc+psb_proc_id_) - end do + do while (proc_to_comm .ne. -1) + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) + + if(proc_to_comm.ne.myrow) then + call mpi_wait(rvhd(proc_to_comm),p2pstat,iret) + if(iret.ne.mpi_success) then + int_err(1) = iret + info=400 + call psb_errpush(info,name,i_err=int_err) + goto 9999 + end if + idx_pt = point_to_proc+nerv+psb_elem_send_ + snd_pt = bsdidx(proc_to_comm) + call psi_sct(nesd,n,d_idx(idx_pt:idx_pt+nesd-1),& + & sndbuf(snd_pt:snd_pt+n*nesd-1),beta,y) + else + idx_pt = point_to_proc+nerv+psb_elem_send_ + rcv_pt = brvidx(proc_to_comm) + call psi_sct(nesd,n,d_idx(idx_pt:idx_pt+nesd-1),& + & rcvbuf(rcv_pt:rcv_pt+n*nerv-1),beta,y) + + end if + + point_to_proc = point_to_proc+nerv+nesd+3 + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) + end do + + end if + else if (swap_send) then @@ -368,9 +403,9 @@ subroutine psi_dswaptranm(flag,n,beta,y,desc_a,work,info,data) idx_pt = point_to_proc+psb_elem_recv_ rcv_pt = brvidx(proc_to_comm) call psi_gth(nerv,n,d_idx(idx_pt:idx_pt+nerv-1),& - & y,rcvbuf(rcv_pt:rcv_pt+nerv*n-1)) - call dgesd2d(icontxt,nerv,n,rcvbuf(rcv_pt),nerv,proc_to_comm,0) - + & y,rcvbuf(rcv_pt:rcv_pt+nerv*n-1)) + call igesd2d(icontxt,nerv,n,rcvbuf(rcv_pt),nerv,proc_to_comm,0) + point_to_proc = point_to_proc+nerv+nesd+3 proc_to_comm = d_idx(point_to_proc+psb_proc_id_) end do @@ -385,7 +420,7 @@ subroutine psi_dswaptranm(flag,n,beta,y,desc_a,work,info,data) if(proc_to_comm.ne.myrow) then snd_pt = bsdidx(proc_to_comm) - call dgerv2d(icontxt,nesd,n,sndbuf(snd_pt),nesd,proc_to_comm,0) + call igerv2d(icontxt,nesd,n,sndbuf(snd_pt),nesd,proc_to_comm,0) idx_pt = point_to_proc+nerv+psb_elem_send_ call psi_sct(nesd,n,d_idx(idx_pt:idx_pt+nesd-1),& & sndbuf(snd_pt:snd_pt+n*nesd-1),beta,y) @@ -429,8 +464,8 @@ subroutine psi_dswaptranv(flag,beta,y,desc_a,work,info,data) integer, intent(in) :: flag integer, intent(out) :: info - real(kind(1.d0)) :: y(:), beta - real(kind(1.d0)), target :: work(:) + integer :: y(:), beta + integer, target :: work(:) type(psb_desc_type) :: desc_a integer, optional :: data @@ -447,7 +482,7 @@ subroutine psi_dswaptranv(flag,beta,y,desc_a,work,info,data) integer :: int_err(5) integer :: blacs_pnum, krecvid, ksendid logical :: swap_mpi, swap_sync, swap_send, swap_recv - real(kind(1.d0)), pointer, dimension(:) :: sndbuf, rcvbuf + integer, pointer, dimension(:) :: sndbuf, rcvbuf character(len=20) :: name, ch_err interface psi_gth @@ -574,14 +609,14 @@ subroutine psi_dswaptranv(flag,beta,y,desc_a,work,info,data) ! Case SWAP_MPI if(swap_mpi) then - + ! gather elements into sendbuffer for swapping point_to_proc = 1 proc_to_comm = d_idx(point_to_proc+psb_proc_id_) do while (proc_to_comm .ne. -1) nerv = d_idx(point_to_proc+psb_n_elem_recv_) nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) - + idx_pt = point_to_proc+psb_elem_recv_ rcv_pt = brvidx(proc_to_comm) call psi_gth(nerv,d_idx(idx_pt:idx_pt+nerv-1),& @@ -593,8 +628,8 @@ subroutine psi_dswaptranv(flag,beta,y,desc_a,work,info,data) ! swap elements using mpi_alltoallv call mpi_alltoallv(rcvbuf,rvsz,brvidx,& - & mpi_double_precision,sndbuf,sdsz,& - & bsdidx,mpi_double_precision,icomm,iret) + & mpi_integer,sndbuf,sdsz,& + & bsdidx,mpi_integer,icomm,iret) if(iret.ne.mpi_success) then int_err(1) = iret info=400 @@ -608,7 +643,7 @@ subroutine psi_dswaptranv(flag,beta,y,desc_a,work,info,data) do while (proc_to_comm .ne. -1) nerv = d_idx(point_to_proc+psb_n_elem_recv_) nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) - + idx_pt = point_to_proc+nerv+psb_elem_send_ snd_pt = bsdidx(proc_to_comm) call psi_sct(nesd,d_idx(idx_pt:idx_pt+nesd-1),& @@ -631,20 +666,20 @@ subroutine psi_dswaptranv(flag,beta,y,desc_a,work,info,data) rcv_pt = brvidx(proc_to_comm) call psi_gth(nerv,d_idx(idx_pt:idx_pt+nerv-1),& & y,rcvbuf(rcv_pt:rcv_pt+nerv-1)) - call dgesd2d(icontxt,nerv,1,rcvbuf(rcv_pt),nerv,proc_to_comm,0) + call igesd2d(icontxt,nerv,1,rcvbuf(rcv_pt),nerv,proc_to_comm,0) ! Then I receive snd_pt = brvidx(proc_to_comm) - call dgerv2d(icontxt,nesd,1,sndbuf(snd_pt),nesd,proc_to_comm,0) + call igerv2d(icontxt,nesd,1,sndbuf(snd_pt),nesd,proc_to_comm,0) else if (proc_to_comm .gt. myrow) then ! First I receive snd_pt = bsdidx(proc_to_comm) - call dgerv2d(icontxt,nesd,1,sndbuf(snd_pt),nesd,proc_to_comm,0) + call igerv2d(icontxt,nesd,1,sndbuf(snd_pt),nesd,proc_to_comm,0) ! Then I send idx_pt = point_to_proc+psb_elem_recv_ rcv_pt = brvidx(proc_to_comm) call psi_gth(nerv,d_idx(idx_pt:idx_pt+nerv-1),& & y,rcvbuf(rcv_pt:rcv_pt+nerv-1)) - call dgesd2d(icontxt,nerv,1,rcvbuf(rcv_pt),nerv,proc_to_comm,0) + call igesd2d(icontxt,nerv,1,rcvbuf(rcv_pt),nerv,proc_to_comm,0) else if (proc_to_comm .eq. myrow) then ! I send to myself idx_pt = point_to_proc+psb_elem_recv_ @@ -692,7 +727,7 @@ subroutine psi_dswaptranv(flag,beta,y,desc_a,work,info,data) p2ptag = krecvid(icontxt,proc_to_comm,myrow) snd_pt = brvidx(proc_to_comm) call mpi_irecv(sndbuf(snd_pt),sdsz(proc_to_comm),& - & mpi_double_precision,prcid(proc_to_comm),& + & mpi_integer,prcid(proc_to_comm),& & p2ptag, icomm,rvhd(proc_to_comm),iret) if(iret.ne.mpi_success) then int_err(1) = iret @@ -712,16 +747,16 @@ subroutine psi_dswaptranv(flag,beta,y,desc_a,work,info,data) do while (proc_to_comm .ne. -1) nerv = d_idx(point_to_proc+psb_n_elem_recv_) nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) - + idx_pt = point_to_proc+psb_elem_recv_ rcv_pt = brvidx(proc_to_comm) call psi_gth(nerv,d_idx(idx_pt:idx_pt+nerv-1),& & y,rcvbuf(rcv_pt:rcv_pt+nerv-1)) - + if(proc_to_comm .ne. myrow) then p2ptag=ksendid(icontxt,proc_to_comm,myrow) call mpi_send(rcvbuf(rcv_pt),rvsz(proc_to_comm),& - & mpi_double_precision,prcid(proc_to_comm),& + & mpi_integer,prcid(proc_to_comm),& & p2ptag,icomm,iret) if(iret.ne.mpi_success) then int_err(1) = iret @@ -734,50 +769,85 @@ subroutine psi_dswaptranv(flag,beta,y,desc_a,work,info,data) proc_to_comm = d_idx(point_to_proc+psb_proc_id_) end do - do i=1, totxch - call mpi_waitany(nprow,rvhd,ixrec,p2pstat,iret) - if(iret.ne.mpi_success) then - int_err(1) = iret - info=400 - call psb_errpush(info,name,i_err=int_err) - goto 9999 - end if + if(.false.) then + do i=1, totxch + call mpi_waitany(nprow,rvhd,ixrec,p2pstat,iret) + if(iret.ne.mpi_success) then + int_err(1) = iret + info=400 + call psb_errpush(info,name,i_err=int_err) + goto 9999 + end if - if (ixrec .ne. mpi_undefined) then - ixrec=ixrec-1 ! mpi_waitany returns an 1 to nprow index - point_to_proc = ptp(ixrec) - proc_to_comm = d_idx(point_to_proc+psb_proc_id_) + if (ixrec .ne. mpi_undefined) then + ixrec=ixrec-1 ! mpi_waitany returns an 1 to nprow index + point_to_proc = ptp(ixrec) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) + + idx_pt = point_to_proc+nerv+psb_elem_send_ + rcv_pt = bsdidx(proc_to_comm) + call psi_sct(nesd,d_idx(idx_pt:idx_pt+nesd-1),& + & rcvbuf(rcv_pt:rcv_pt+nerv-1),beta,y) + else + int_err(1) = ixrec + info=400 + call psb_errpush(info,name,i_err=int_err) + goto 9999 + end if + end do + + point_to_proc = 1 + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) + do while (proc_to_comm .ne. -1) nerv = d_idx(point_to_proc+psb_n_elem_recv_) nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) - idx_pt = point_to_proc+nerv+psb_elem_send_ - rcv_pt = bsdidx(proc_to_comm) - call psi_sct(nesd,d_idx(idx_pt:idx_pt+nesd-1),& - & rcvbuf(rcv_pt:rcv_pt+nerv-1),beta,y) - else - int_err(1) = ixrec - info=400 - call psb_errpush(info,name,i_err=int_err) - goto 9999 - end if - end do - - point_to_proc = 1 - proc_to_comm = d_idx(point_to_proc+psb_proc_id_) - do while (proc_to_comm .ne. -1) - nerv = d_idx(point_to_proc+psb_n_elem_recv_) - nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) - - if(proc_to_comm .eq. myrow) then - idx_pt = point_to_proc+nerv+psb_elem_send_ - rcv_pt = brvidx(proc_to_comm) - call psi_sct(nesd,d_idx(idx_pt:idx_pt+nesd-1),& - & rcvbuf(rcv_pt:rcv_pt+nerv-1),beta,y) - end if + if(proc_to_comm .eq. myrow) then + idx_pt = point_to_proc+nerv+psb_elem_send_ + rcv_pt = brvidx(proc_to_comm) + call psi_sct(nesd,d_idx(idx_pt:idx_pt+nesd-1),& + & rcvbuf(rcv_pt:rcv_pt+nerv-1),beta,y) + end if - point_to_proc = point_to_proc+nerv+nesd+3 + point_to_proc = point_to_proc+nerv+nesd+3 + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) + end do + + else + + point_to_proc = 1 proc_to_comm = d_idx(point_to_proc+psb_proc_id_) - end do + do while (proc_to_comm .ne. -1) + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) + + if(proc_to_comm.ne.myrow) then + call mpi_wait(rvhd(proc_to_comm),p2pstat,iret) + if(iret.ne.mpi_success) then + int_err(1) = iret + info=400 + call psb_errpush(info,name,i_err=int_err) + goto 9999 + end if + idx_pt = point_to_proc+nerv+psb_elem_send_ + snd_pt = bsdidx(proc_to_comm) + call psi_sct(nesd,d_idx(idx_pt:idx_pt+nesd-1),& + & sndbuf(snd_pt:snd_pt+nesd-1),beta,y) + else + idx_pt = point_to_proc+nerv+psb_elem_send_ + rcv_pt = brvidx(proc_to_comm) + call psi_sct(nesd,d_idx(idx_pt:idx_pt+nesd-1),& + & rcvbuf(rcv_pt:rcv_pt+nerv-1),beta,y) + + end if + + point_to_proc = point_to_proc+nerv+nesd+3 + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) + end do + + end if else if (swap_send) then @@ -790,9 +860,9 @@ subroutine psi_dswaptranv(flag,beta,y,desc_a,work,info,data) idx_pt = point_to_proc+psb_elem_recv_ rcv_pt = brvidx(proc_to_comm) call psi_gth(nerv,d_idx(idx_pt:idx_pt+nerv-1),& - & y,rcvbuf(rcv_pt:rcv_pt+nerv-1)) - call dgesd2d(icontxt,nerv,1,rcvbuf(rcv_pt),nerv,proc_to_comm,0) - + & y,rcvbuf(rcv_pt:rcv_pt+nerv-1)) + call igesd2d(icontxt,nerv,1,rcvbuf(rcv_pt),nerv,proc_to_comm,0) + point_to_proc = point_to_proc+nerv+nesd+3 proc_to_comm = d_idx(point_to_proc+psb_proc_id_) end do @@ -807,7 +877,7 @@ subroutine psi_dswaptranv(flag,beta,y,desc_a,work,info,data) if(proc_to_comm.ne.myrow) then snd_pt = bsdidx(proc_to_comm) - call dgerv2d(icontxt,nesd,1,sndbuf(snd_pt),nesd,proc_to_comm,0) + call igerv2d(icontxt,nesd,1,sndbuf(snd_pt),nesd,proc_to_comm,0) idx_pt = point_to_proc+psb_elem_recv_ rcv_pt = brvidx(proc_to_comm) call psi_sct(nesd,d_idx(idx_pt:idx_pt+nesd-1),& diff --git a/src/internals/psi_iswapdata.f90 b/src/internals/psi_iswapdata.f90 index d4ce19d3..5d8568ff 100644 --- a/src/internals/psi_iswapdata.f90 +++ b/src/internals/psi_iswapdata.f90 @@ -1,4 +1,4 @@ -subroutine psi_iswapdatam(flag,n,beta,y,desc_a,work,info,data) +subroutine psi_dswapdatam(flag,n,beta,y,desc_a,work,info,data) use psb_error_mod use psb_descriptor_type @@ -8,7 +8,7 @@ subroutine psi_iswapdatam(flag,n,beta,y,desc_a,work,info,data) integer, intent(in) :: flag, n integer, intent(out) :: info integer :: y(:,:), beta - integer, target :: work(:) + integer, target :: work(:) type(psb_desc_type) :: desc_a integer, optional :: data @@ -150,17 +150,17 @@ subroutine psi_iswapdatam(flag,n,beta,y,desc_a,work,info,data) ! Case SWAP_MPI if(swap_mpi) then - + ! gather elements into sendbuffer for swapping point_to_proc = 1 proc_to_comm = d_idx(point_to_proc+psb_proc_id_) do while (proc_to_comm .ne. -1) nerv = d_idx(point_to_proc+psb_n_elem_recv_) nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) - + idx_pt = point_to_proc+nerv+psb_elem_send_ snd_pt = bsdidx(proc_to_comm) - + call psi_gth(nesd,n,d_idx(idx_pt:idx_pt+nesd-1),& & y,sndbuf(snd_pt:snd_pt+nesd*n-1)) @@ -185,7 +185,7 @@ subroutine psi_iswapdatam(flag,n,beta,y,desc_a,work,info,data) do while (proc_to_comm .ne. -1) nerv = d_idx(point_to_proc+psb_n_elem_recv_) nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) - + idx_pt = point_to_proc+psb_elem_recv_ rcv_pt = brvidx(proc_to_comm) call psi_sct(nerv,n,d_idx(idx_pt:idx_pt+nerv-1),& @@ -288,13 +288,13 @@ subroutine psi_iswapdatam(flag,n,beta,y,desc_a,work,info,data) do while (proc_to_comm .ne. -1) nerv = d_idx(point_to_proc+psb_n_elem_recv_) nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) - + idx_pt = point_to_proc+nerv+psb_elem_send_ snd_pt = bsdidx(proc_to_comm) - + call psi_gth(nesd,n,d_idx(idx_pt:idx_pt+nesd-1),& & y,sndbuf(snd_pt:snd_pt+nesd*n-1)) - + if(proc_to_comm .ne. myrow) then p2ptag=ksendid(icontxt,proc_to_comm,myrow) call mpi_send(sndbuf(snd_pt),sdsz(proc_to_comm),& @@ -311,51 +311,84 @@ subroutine psi_iswapdatam(flag,n,beta,y,desc_a,work,info,data) proc_to_comm = d_idx(point_to_proc+psb_proc_id_) end do - do i=1, totxch - call mpi_waitany(nprow,rvhd,ixrec,p2pstat,iret) - if(iret.ne.mpi_success) then - int_err(1) = iret - info=400 - call psb_errpush(info,name,i_err=int_err) - goto 9999 - end if + if(.false.) then + do i=1, totxch + call mpi_waitany(nprow,rvhd,ixrec,p2pstat,iret) + if(iret.ne.mpi_success) then + int_err(1) = iret + info=400 + call psb_errpush(info,name,i_err=int_err) + goto 9999 + end if - if (ixrec .ne. mpi_undefined) then - ixrec=ixrec-1 ! mpi_waitany returns an 1 to nprow index - point_to_proc = ptp(ixrec) - proc_to_comm = d_idx(point_to_proc+psb_proc_id_) + if (ixrec .ne. mpi_undefined) then + ixrec=ixrec-1 ! mpi_waitany returns an 1 to nprow index + point_to_proc = ptp(ixrec) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) + + idx_pt = point_to_proc+psb_elem_recv_ + rcv_pt = brvidx(proc_to_comm) + call psi_sct(nerv,n,d_idx(idx_pt:idx_pt+nerv-1),& + & rcvbuf(rcv_pt:rcv_pt+n*nerv-1),beta,y) + else + int_err(1) = ixrec + info=400 + call psb_errpush(info,name,i_err=int_err) + goto 9999 + end if + end do + + point_to_proc = 1 + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) + do while (proc_to_comm .ne. -1) nerv = d_idx(point_to_proc+psb_n_elem_recv_) nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) - idx_pt = point_to_proc+psb_elem_recv_ - rcv_pt = brvidx(proc_to_comm) - call psi_sct(nerv,n,d_idx(idx_pt:idx_pt+nerv-1),& - & rcvbuf(rcv_pt:rcv_pt+n*nerv-1),beta,y) - else - int_err(1) = ixrec - info=400 - call psb_errpush(info,name,i_err=int_err) - goto 9999 - end if - end do - - point_to_proc = 1 - proc_to_comm = d_idx(point_to_proc+psb_proc_id_) - do while (proc_to_comm .ne. -1) - nerv = d_idx(point_to_proc+psb_n_elem_recv_) - nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) - - if(proc_to_comm .eq. myrow) then - idx_pt = point_to_proc+psb_elem_recv_ - snd_pt = bsdidx(proc_to_comm) - call psi_sct(nerv,n,d_idx(idx_pt:idx_pt+nerv-1),& - & sndbuf(snd_pt:snd_pt+n*nesd-1),beta,y) - end if + if(proc_to_comm .eq. myrow) then + idx_pt = point_to_proc+psb_elem_recv_ + snd_pt = bsdidx(proc_to_comm) + call psi_sct(nerv,n,d_idx(idx_pt:idx_pt+nerv-1),& + & sndbuf(snd_pt:snd_pt+n*nesd-1),beta,y) + end if - point_to_proc = point_to_proc+nerv+nesd+3 + point_to_proc = point_to_proc+nerv+nesd+3 + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) + end do + else + + point_to_proc = 1 proc_to_comm = d_idx(point_to_proc+psb_proc_id_) - end do + do while (proc_to_comm .ne. -1) + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) + + if(proc_to_comm.ne.myrow) then + call mpi_wait(rvhd(proc_to_comm),p2pstat,iret) + if(iret.ne.mpi_success) then + int_err(1) = iret + info=400 + call psb_errpush(info,name,i_err=int_err) + goto 9999 + end if + idx_pt = point_to_proc+psb_elem_recv_ + rcv_pt = brvidx(proc_to_comm) + call psi_sct(nerv,n,d_idx(idx_pt:idx_pt+nerv-1),& + & rcvbuf(rcv_pt:rcv_pt+n*nerv-1),beta,y) + else + idx_pt = point_to_proc+psb_elem_recv_ + snd_pt = bsdidx(proc_to_comm) + call psi_sct(nerv,n,d_idx(idx_pt:idx_pt+nerv-1),& + & sndbuf(snd_pt:snd_pt+n*nesd-1),beta,y) + + end if + point_to_proc = point_to_proc+nerv+nesd+3 + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) + end do + + end if else if (swap_send) then @@ -369,9 +402,9 @@ subroutine psi_iswapdatam(flag,n,beta,y,desc_a,work,info,data) idx_pt = point_to_proc+nerv+psb_elem_send_ snd_pt = bsdidx(proc_to_comm) call psi_gth(nesd,n,d_idx(idx_pt:idx_pt+nesd-1),& - & y,sndbuf(snd_pt:snd_pt+nesd*n-1)) + & y,sndbuf(snd_pt:snd_pt+nesd*n-1)) call igesd2d(icontxt,nesd,n,sndbuf(snd_pt),nesd,proc_to_comm,0) - + point_to_proc = point_to_proc+nerv+nesd+3 proc_to_comm = d_idx(point_to_proc+psb_proc_id_) end do @@ -414,10 +447,10 @@ subroutine psi_iswapdatam(flag,n,beta,y,desc_a,work,info,data) return end if return -end subroutine psi_iswapdatam +end subroutine psi_dswapdatam -subroutine psi_iswapdatav(flag,beta,y,desc_a,work,info,data) +subroutine psi_dswapdatav(flag,beta,y,desc_a,work,info,data) use psb_error_mod use psb_descriptor_type @@ -426,8 +459,8 @@ subroutine psi_iswapdatav(flag,beta,y,desc_a,work,info,data) integer, intent(in) :: flag integer, intent(out) :: info - integer :: y(:), beta - integer, target :: work(:) + integer :: y(:), beta + integer, target :: work(:) type(psb_desc_type) :: desc_a integer, optional :: data @@ -486,7 +519,7 @@ subroutine psi_iswapdatav(flag,beta,y,desc_a,work,info,data) end interface info = 0 - name='psi_iswap_datav' + name='psi_dswap_datav' call psb_erractionsave(err_act) icontxt=desc_a%matrix_data(psb_ctxt_) @@ -572,14 +605,14 @@ subroutine psi_iswapdatav(flag,beta,y,desc_a,work,info,data) ! Case SWAP_MPI if(swap_mpi) then - + ! gather elements into sendbuffer for swapping point_to_proc = 1 proc_to_comm = d_idx(point_to_proc+psb_proc_id_) do while (proc_to_comm .ne. -1) nerv = d_idx(point_to_proc+psb_n_elem_recv_) nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) - + idx_pt = point_to_proc+nerv+psb_elem_send_ snd_pt = bsdidx(proc_to_comm) call psi_gth(nesd,d_idx(idx_pt:idx_pt+nesd-1),& @@ -606,7 +639,7 @@ subroutine psi_iswapdatav(flag,beta,y,desc_a,work,info,data) do while (proc_to_comm .ne. -1) nerv = d_idx(point_to_proc+psb_n_elem_recv_) nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) - + idx_pt = point_to_proc+psb_elem_recv_ rcv_pt = brvidx(proc_to_comm) call psi_sct(nerv,d_idx(idx_pt:idx_pt+nerv-1),& @@ -709,12 +742,12 @@ subroutine psi_iswapdatav(flag,beta,y,desc_a,work,info,data) do while (proc_to_comm .ne. -1) nerv = d_idx(point_to_proc+psb_n_elem_recv_) nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) - + idx_pt = point_to_proc+nerv+psb_elem_send_ snd_pt = bsdidx(proc_to_comm) call psi_gth(nesd,d_idx(idx_pt:idx_pt+nesd-1),& & y,sndbuf(snd_pt:snd_pt+nesd-1)) - + if(proc_to_comm .ne. myrow) then p2ptag=ksendid(icontxt,proc_to_comm,myrow) call mpi_send(sndbuf(snd_pt),sdsz(proc_to_comm),& @@ -731,49 +764,85 @@ subroutine psi_iswapdatav(flag,beta,y,desc_a,work,info,data) proc_to_comm = d_idx(point_to_proc+psb_proc_id_) end do - do i=1, totxch - call mpi_waitany(nprow,rvhd,ixrec,p2pstat,iret) - if(iret.ne.mpi_success) then - int_err(1) = iret - info=400 - call psb_errpush(info,name,i_err=int_err) - goto 9999 - end if - if (ixrec .ne. mpi_undefined) then - ixrec=ixrec-1 ! mpi_waitany returns an 1 to nprow index - point_to_proc = ptp(ixrec) - proc_to_comm = d_idx(point_to_proc+psb_proc_id_) + if(.false.) then + do i=1, totxch + call mpi_waitany(nprow,rvhd,ixrec,p2pstat,iret) + if(iret.ne.mpi_success) then + int_err(1) = iret + info=400 + call psb_errpush(info,name,i_err=int_err) + goto 9999 + end if + if (ixrec .ne. mpi_undefined) then + ixrec=ixrec-1 ! mpi_waitany returns an 1 to nprow index + point_to_proc = ptp(ixrec) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) + + idx_pt = point_to_proc+psb_elem_recv_ + rcv_pt = brvidx(proc_to_comm) + call psi_sct(nerv,d_idx(idx_pt:idx_pt+nerv-1),& + & rcvbuf(rcv_pt:rcv_pt+nerv-1),beta,y) + else + int_err(1) = ixrec + info=400 + call psb_errpush(info,name,i_err=int_err) + goto 9999 + end if + end do + + point_to_proc = 1 + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) + do while (proc_to_comm .ne. -1) nerv = d_idx(point_to_proc+psb_n_elem_recv_) nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) - idx_pt = point_to_proc+psb_elem_recv_ - rcv_pt = brvidx(proc_to_comm) - call psi_sct(nerv,d_idx(idx_pt:idx_pt+nerv-1),& - & rcvbuf(rcv_pt:rcv_pt+nerv-1),beta,y) - else - int_err(1) = ixrec - info=400 - call psb_errpush(info,name,i_err=int_err) - goto 9999 - end if - end do - - point_to_proc = 1 - proc_to_comm = d_idx(point_to_proc+psb_proc_id_) - do while (proc_to_comm .ne. -1) - nerv = d_idx(point_to_proc+psb_n_elem_recv_) - nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) - - if(proc_to_comm .eq. myrow) then - idx_pt = point_to_proc+psb_elem_recv_ - snd_pt = bsdidx(proc_to_comm) - call psi_sct(nerv,d_idx(idx_pt:idx_pt+nerv-1),& - & sndbuf(snd_pt:snd_pt+nesd-1),beta,y) - end if + if(proc_to_comm .eq. myrow) then + idx_pt = point_to_proc+psb_elem_recv_ + snd_pt = bsdidx(proc_to_comm) + call psi_sct(nerv,d_idx(idx_pt:idx_pt+nerv-1),& + & sndbuf(snd_pt:snd_pt+nesd-1),beta,y) + end if - point_to_proc = point_to_proc+nerv+nesd+3 + point_to_proc = point_to_proc+nerv+nesd+3 + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) + end do + + else + + point_to_proc = 1 proc_to_comm = d_idx(point_to_proc+psb_proc_id_) - end do + do while (proc_to_comm .ne. -1) + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) + + if(proc_to_comm.ne.myrow) then + call mpi_wait(rvhd(proc_to_comm),p2pstat,iret) + if(iret.ne.mpi_success) then + int_err(1) = iret + info=400 + call psb_errpush(info,name,i_err=int_err) + goto 9999 + end if + idx_pt = point_to_proc+psb_elem_recv_ + rcv_pt = brvidx(proc_to_comm) + call psi_sct(nerv,d_idx(idx_pt:idx_pt+nerv-1),& + & rcvbuf(rcv_pt:rcv_pt+nerv-1),beta,y) + else + idx_pt = point_to_proc+psb_elem_recv_ + snd_pt = bsdidx(proc_to_comm) + call psi_sct(nerv,d_idx(idx_pt:idx_pt+nerv-1),& + & sndbuf(snd_pt:snd_pt+nesd-1),beta,y) + + end if + + point_to_proc = point_to_proc+nerv+nesd+3 + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) + end do + + end if + else if (swap_send) then @@ -787,9 +856,9 @@ subroutine psi_iswapdatav(flag,beta,y,desc_a,work,info,data) idx_pt = point_to_proc+nerv+psb_elem_send_ snd_pt = bsdidx(proc_to_comm) call psi_gth(nesd,d_idx(idx_pt:idx_pt+nesd-1),& - & y,sndbuf(snd_pt:snd_pt+nesd-1)) + & y,sndbuf(snd_pt:snd_pt+nesd-1)) call igesd2d(icontxt,nesd,1,sndbuf(snd_pt),nesd,proc_to_comm,0) - + point_to_proc = point_to_proc+nerv+nesd+3 proc_to_comm = d_idx(point_to_proc+psb_proc_id_) end do @@ -832,4 +901,4 @@ subroutine psi_iswapdatav(flag,beta,y,desc_a,work,info,data) return end if return -end subroutine psi_iswapdatav +end subroutine psi_dswapdatav diff --git a/src/internals/psi_iswaptran.f90 b/src/internals/psi_iswaptran.f90 index 0d064bdc..6f64c147 100644 --- a/src/internals/psi_iswaptran.f90 +++ b/src/internals/psi_iswaptran.f90 @@ -1,30 +1,31 @@ -subroutine psi_iswaptranm(flag,n,beta,y,desc_a,work,info) +subroutine psi_dswaptranm(flag,n,beta,y,desc_a,work,info,data) use psb_error_mod use psb_descriptor_type + use mpi implicit none - include 'mpif.h' integer, intent(in) :: flag, n integer, intent(out) :: info - integer :: y(:,:), beta - integer, target :: work(:) + real(kind(1.d0)) :: y(:,:), beta + real(kind(1.d0)), target :: work(:) type(psb_desc_type) :: desc_a + integer, optional :: data ! locals integer :: icontxt, nprow, npcol, myrow,& & mycol, point_to_proc, nesd, nerv,& - & proc_to_comm, p2ptag, icomm, p2pstat,& + & proc_to_comm, p2ptag, icomm, p2pstat(mpi_status_size),& & idxs, idxr, iret, errlen, ifcomm, rank,& & err_act, totxch, ixrec, i, lw, idx_pt,& & snd_pt, rcv_pt integer, pointer, dimension(:) :: bsdidx, brvidx,& - & sdsz, rvsz, prcid, ptp, rvhd, h_idx + & sdsz, rvsz, prcid, ptp, rvhd, d_idx integer :: int_err(5) integer :: blacs_pnum, krecvid, ksendid logical :: swap_mpi, swap_sync, swap_send, swap_recv - integer, pointer, dimension(:) :: sndbuf, rcvbuf + real(kind(1.d0)), pointer, dimension(:) :: sndbuf, rcvbuf character(len=20) :: name, ch_err interface psi_gth @@ -96,7 +97,19 @@ subroutine psi_iswaptranm(flag,n,beta,y,desc_a,work,info) swap_sync = iand(flag,psb_swap_sync_).ne.0 swap_send = iand(flag,psb_swap_send_).ne.0 swap_recv = iand(flag,psb_swap_recv_).ne.0 - h_idx => desc_a%halo_index + + if(present(data)) then + if(data.eq.psb_comm_halo_) then + d_idx => desc_a%halo_index + else if(data.eq.psb_comm_ovr_) then + d_idx => desc_a%ovrlap_index + else + d_idx => desc_a%halo_index + end if + else + d_idx => desc_a%halo_index + end if + idxs = 0 idxr = 0 totxch = 0 @@ -104,11 +117,11 @@ subroutine psi_iswaptranm(flag,n,beta,y,desc_a,work,info) rvhd(:) = mpi_request_null ! prepare info for communications - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) do while (proc_to_comm.ne.-1) if(proc_to_comm .ne. myrow) totxch = totxch+1 - nerv = h_idx(point_to_proc+psb_n_elem_recv_) - nesd = h_idx(point_to_proc+nerv+psb_n_elem_send_) + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) prcid(proc_to_comm) = blacs_pnum(icontxt,proc_to_comm,mycol) ptp(proc_to_comm) = point_to_proc @@ -122,7 +135,7 @@ subroutine psi_iswaptranm(flag,n,beta,y,desc_a,work,info) idxs = idxs+sdsz(proc_to_comm) point_to_proc = point_to_proc+nerv+nesd+3 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) end do if((idxr+idxs).lt.size(work)) then @@ -138,27 +151,28 @@ subroutine psi_iswaptranm(flag,n,beta,y,desc_a,work,info) ! Case SWAP_MPI if(swap_mpi) then - + ! gather elements into sendbuffer for swapping point_to_proc = 1 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) do while (proc_to_comm .ne. -1) - nerv = h_idx(point_to_proc+psb_n_elem_recv_) - nesd = h_idx(point_to_proc+nerv+psb_n_elem_send_) - + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) + idx_pt = point_to_proc+psb_elem_recv_ rcv_pt = brvidx(proc_to_comm) - call psi_gth(nerv,n,h_idx(idx_pt:idx_pt+nerv-1),& + + call psi_gth(nerv,n,d_idx(idx_pt:idx_pt+nerv-1),& & y,rcvbuf(rcv_pt:rcv_pt+nerv*n-1)) point_to_proc = point_to_proc+nerv+nesd+3 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) end do ! swap elements using mpi_alltoallv call mpi_alltoallv(rcvbuf,rvsz,brvidx,& - & mpi_integer,sndbuf,sdsz,& - & bsdidx,mpi_integer,icomm,iret) + & mpi_double_precision,sndbuf,sdsz,& + & bsdidx,mpi_double_precision,icomm,iret) if(iret.ne.mpi_success) then int_err(1) = iret info=400 @@ -168,95 +182,95 @@ subroutine psi_iswaptranm(flag,n,beta,y,desc_a,work,info) ! scatter elements from receivebuffer after swapping point_to_proc = 1 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) do while (proc_to_comm .ne. -1) - nerv = h_idx(point_to_proc+psb_n_elem_recv_) - nesd = h_idx(point_to_proc+nerv+psb_n_elem_send_) - + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) + idx_pt = point_to_proc+nerv+psb_elem_send_ snd_pt = bsdidx(proc_to_comm) - call psi_sct(nesd,n,h_idx(idx_pt:idx_pt+nesd-1),& + call psi_sct(nesd,n,d_idx(idx_pt:idx_pt+nesd-1),& & sndbuf(snd_pt:snd_pt+n*nesd-1),beta,y) point_to_proc = point_to_proc+nerv+nesd+3 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) end do else if (swap_sync) then point_to_proc = 1 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) do while (proc_to_comm .ne. -1) - nerv = h_idx(point_to_proc+psb_n_elem_recv_) - nesd = h_idx(point_to_proc+nerv+psb_n_elem_send_) + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) if (proc_to_comm .lt. myrow) then ! First I send idx_pt = point_to_proc+psb_elem_recv_ rcv_pt = brvidx(proc_to_comm) - call psi_gth(nerv,n,h_idx(idx_pt:idx_pt+nerv-1),& + call psi_gth(nerv,n,d_idx(idx_pt:idx_pt+nerv-1),& & y,rcvbuf(rcv_pt:rcv_pt+nerv*n-1)) - call igesd2d(icontxt,nerv,n,rcvbuf(rcv_pt),nerv,proc_to_comm,0) + call dgesd2d(icontxt,nerv,n,rcvbuf(rcv_pt),nerv,proc_to_comm,0) ! Then I receive snd_pt = brvidx(proc_to_comm) - call igerv2d(icontxt,nesd,n,sndbuf(snd_pt),nesd,proc_to_comm,0) + call dgerv2d(icontxt,nesd,n,sndbuf(snd_pt),nesd,proc_to_comm,0) else if (proc_to_comm .gt. myrow) then ! First I receive snd_pt = bsdidx(proc_to_comm) - call igerv2d(icontxt,nesd,n,sndbuf(snd_pt),nesd,proc_to_comm,0) + call dgerv2d(icontxt,nesd,n,sndbuf(snd_pt),nesd,proc_to_comm,0) ! Then I send idx_pt = point_to_proc+psb_elem_recv_ rcv_pt = brvidx(proc_to_comm) - call psi_gth(nerv,n,h_idx(idx_pt:idx_pt+nerv-1),& + call psi_gth(nerv,n,d_idx(idx_pt:idx_pt+nerv-1),& & y,rcvbuf(rcv_pt:rcv_pt+nerv*n-1)) - call igesd2d(icontxt,nerv,n,rcvbuf(rcv_pt),nerv,proc_to_comm,0) + call dgesd2d(icontxt,nerv,n,rcvbuf(rcv_pt),nerv,proc_to_comm,0) else if (proc_to_comm .eq. myrow) then ! I send to myself idx_pt = point_to_proc+psb_elem_recv_ rcv_pt = bsdidx(proc_to_comm) - call psi_gth(nerv,n,h_idx(idx_pt:idx_pt+nerv-1),& + call psi_gth(nerv,n,d_idx(idx_pt:idx_pt+nerv-1),& & y,rcvbuf(rcv_pt:rcv_pt+nerv*n-1)) end if point_to_proc = point_to_proc+nerv+nesd+3 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) end do point_to_proc = 1 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) do while (proc_to_comm .ne. -1) - nerv = h_idx(point_to_proc+psb_n_elem_recv_) - nesd = h_idx(point_to_proc+nerv+psb_n_elem_send_) + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) if(proc_to_comm.ne.myrow) then idx_pt = point_to_proc+nerv+psb_elem_send_ snd_pt = bsdidx(proc_to_comm) - call psi_sct(nesd,n,h_idx(idx_pt:idx_pt+nesd-1),& + call psi_sct(nesd,n,d_idx(idx_pt:idx_pt+nesd-1),& & sndbuf(snd_pt:snd_pt+n*nesd-1),beta,y) else idx_pt = point_to_proc+nerv+psb_elem_send_ rcv_pt = brvidx(proc_to_comm) - call psi_sct(nesd,n,h_idx(idx_pt:idx_pt+nesd-1),& + call psi_sct(nesd,n,d_idx(idx_pt:idx_pt+nesd-1),& & rcvbuf(rcv_pt:rcv_pt+n*nerv-1),beta,y) end if point_to_proc = point_to_proc+nerv+nesd+3 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) end do else if (swap_send .and. swap_recv) then ! First I post all the non blocking receives point_to_proc = 1 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) do while (proc_to_comm .ne. -1) - nerv = h_idx(point_to_proc+psb_n_elem_recv_) - nesd = h_idx(point_to_proc+nerv+psb_n_elem_send_) + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) if(proc_to_comm.ne.myrow) then p2ptag = krecvid(icontxt,proc_to_comm,myrow) snd_pt = brvidx(proc_to_comm) call mpi_irecv(sndbuf(rcv_pt),sdsz(proc_to_comm),& - & mpi_integer,prcid(proc_to_comm),& + & mpi_double_precision,prcid(proc_to_comm),& & p2ptag, icomm,rvhd(proc_to_comm),iret) if(iret.ne.mpi_success) then int_err(1) = iret @@ -267,25 +281,25 @@ subroutine psi_iswaptranm(flag,n,beta,y,desc_a,work,info) end if point_to_proc = point_to_proc+nerv+nesd+3 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) end do ! Then I post all the blocking sends point_to_proc = 1 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) do while (proc_to_comm .ne. -1) - nerv = h_idx(point_to_proc+psb_n_elem_recv_) - nesd = h_idx(point_to_proc+nerv+psb_n_elem_send_) - + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) + idx_pt = point_to_proc+psb_elem_recv_ rcv_pt = brvidx(proc_to_comm) - call psi_gth(nerv,n,h_idx(idx_pt:idx_pt+nerv-1),& + call psi_gth(nerv,n,d_idx(idx_pt:idx_pt+nerv-1),& & y,rcvbuf(rcv_pt:rcv_pt+nerv*n-1)) - + if(proc_to_comm .ne. myrow) then p2ptag=ksendid(icontxt,proc_to_comm,myrow) call mpi_send(rcvbuf(rcv_pt),rvsz(proc_to_comm),& - & mpi_integer,prcid(proc_to_comm),& + & mpi_double_precision,prcid(proc_to_comm),& & p2ptag,icomm,iret) if(iret.ne.mpi_success) then int_err(1) = iret @@ -295,95 +309,130 @@ subroutine psi_iswaptranm(flag,n,beta,y,desc_a,work,info) end if end if point_to_proc = point_to_proc+nerv+nesd+3 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) end do - do i=1, totxch - call mpi_waitany(nprow,rvhd,ixrec,p2pstat,iret) - if(iret.ne.mpi_success) then - int_err(1) = iret - info=400 - call psb_errpush(info,name,i_err=int_err) - goto 9999 - end if + if(.false.) then + do i=1, totxch + call mpi_waitany(nprow,rvhd,ixrec,p2pstat,iret) + if(iret.ne.mpi_success) then + int_err(1) = iret + info=400 + call psb_errpush(info,name,i_err=int_err) + goto 9999 + end if + + if (ixrec .ne. mpi_undefined) then + ixrec=ixrec-1 ! mpi_waitany returns an 1 to nprow index + point_to_proc = ptp(ixrec) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) + + idx_pt = point_to_proc+nerv+psb_elem_send_ + snd_pt = bsdidx(proc_to_comm) + call psi_sct(nesd,n,d_idx(idx_pt:idx_pt+nesd-1),& + & sndbuf(snd_pt:snd_pt+n*nesd-1),beta,y) + else + int_err(1) = ixrec + info=400 + call psb_errpush(info,name,i_err=int_err) + goto 9999 + end if + end do + + point_to_proc = 1 + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) + do while (proc_to_comm .ne. -1) + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) + + if(proc_to_comm .eq. myrow) then + idx_pt = point_to_proc+nerv+psb_elem_send_ + rcv_pt = brvidx(proc_to_comm) + call psi_sct(nesd,n,d_idx(idx_pt:idx_pt+nesd-1),& + & rcvbuf(rcv_pt:rcv_pt+n*nerv-1),beta,y) + end if - if (ixrec .ne. mpi_undefined) then - ixrec=ixrec-1 ! mpi_waitany returns an 1 to nprow index - point_to_proc = ptp(ixrec) - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) - nerv = h_idx(point_to_proc+psb_n_elem_recv_) - nesd = h_idx(point_to_proc+nerv+psb_n_elem_send_) + point_to_proc = point_to_proc+nerv+nesd+3 + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) + end do + else + + point_to_proc = 1 + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) + do while (proc_to_comm .ne. -1) + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) + + if(proc_to_comm.ne.myrow) then + call mpi_wait(rvhd(proc_to_comm),p2pstat,iret) + if(iret.ne.mpi_success) then + int_err(1) = iret + info=400 + call psb_errpush(info,name,i_err=int_err) + goto 9999 + end if + idx_pt = point_to_proc+nerv+psb_elem_send_ + snd_pt = bsdidx(proc_to_comm) + call psi_sct(nesd,n,d_idx(idx_pt:idx_pt+nesd-1),& + & sndbuf(snd_pt:snd_pt+n*nesd-1),beta,y) + else + idx_pt = point_to_proc+nerv+psb_elem_send_ + rcv_pt = brvidx(proc_to_comm) + call psi_sct(nesd,n,d_idx(idx_pt:idx_pt+nesd-1),& + & rcvbuf(rcv_pt:rcv_pt+n*nerv-1),beta,y) - idx_pt = point_to_proc+nerv+psb_elem_send_ - snd_pt = bsdidx(proc_to_comm) - call psi_sct(nesd,n,h_idx(idx_pt:idx_pt+nesd-1),& - & sndbuf(snd_pt:snd_pt+n*nesd-1),beta,y) - else - int_err(1) = ixrec - info=400 - call psb_errpush(info,name,i_err=int_err) - goto 9999 - end if - end do - - point_to_proc = 1 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) - do while (proc_to_comm .ne. -1) - nerv = h_idx(point_to_proc+psb_n_elem_recv_) - nesd = h_idx(point_to_proc+nerv+psb_n_elem_send_) - - if(proc_to_comm .eq. myrow) then - idx_pt = point_to_proc+nerv+psb_elem_send_ - rcv_pt = brvidx(proc_to_comm) - call psi_sct(nesd,n,h_idx(idx_pt:idx_pt+nesd-1),& - & rcvbuf(rcv_pt:rcv_pt+n*nerv-1),beta,y) - end if + end if + + point_to_proc = point_to_proc+nerv+nesd+3 + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) + end do + + end if - point_to_proc = point_to_proc+nerv+nesd+3 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) - end do else if (swap_send) then point_to_proc = 1 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) do while (proc_to_comm .ne. -1) - nerv = h_idx(point_to_proc+psb_n_elem_recv_) - nesd = h_idx(point_to_proc+nerv+psb_n_elem_send_) + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) idx_pt = point_to_proc+psb_elem_recv_ rcv_pt = brvidx(proc_to_comm) - call psi_gth(nerv,n,h_idx(idx_pt:idx_pt+nerv-1),& - & y,rcvbuf(rcv_pt:rcv_pt+nerv*n-1)) - call igesd2d(icontxt,nerv,n,rcvbuf(rcv_pt),nerv,proc_to_comm,0) - + call psi_gth(nerv,n,d_idx(idx_pt:idx_pt+nerv-1),& + & y,rcvbuf(rcv_pt:rcv_pt+nerv*n-1)) + call dgesd2d(icontxt,nerv,n,rcvbuf(rcv_pt),nerv,proc_to_comm,0) + point_to_proc = point_to_proc+nerv+nesd+3 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) end do else if (swap_recv) then point_to_proc = 1 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) do while (proc_to_comm .ne. -1) - nerv = h_idx(point_to_proc+psb_n_elem_recv_) - nesd = h_idx(point_to_proc+nerv+psb_n_elem_send_) + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) if(proc_to_comm.ne.myrow) then snd_pt = bsdidx(proc_to_comm) - call igerv2d(icontxt,nesd,n,sndbuf(snd_pt),nesd,proc_to_comm,0) + call dgerv2d(icontxt,nesd,n,sndbuf(snd_pt),nesd,proc_to_comm,0) idx_pt = point_to_proc+nerv+psb_elem_send_ - call psi_sct(nesd,n,h_idx(idx_pt:idx_pt+nesd-1),& + call psi_sct(nesd,n,d_idx(idx_pt:idx_pt+nesd-1),& & sndbuf(snd_pt:snd_pt+n*nesd-1),beta,y) else idx_pt = point_to_proc+nerv+psb_elem_send_ rcv_pt = brvidx(proc_to_comm) - call psi_sct(nesd,n,h_idx(idx_pt:idx_pt+nesd-1),& + call psi_sct(nesd,n,d_idx(idx_pt:idx_pt+nesd-1),& & rcvbuf(rcv_pt:rcv_pt+n*nerv-1),beta,y) end if point_to_proc = point_to_proc+nerv+nesd+3 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) end do end if @@ -398,7 +447,7 @@ subroutine psi_iswaptranm(flag,n,beta,y,desc_a,work,info) return end if return -end subroutine psi_iswaptranm +end subroutine psi_dswaptranm @@ -406,33 +455,34 @@ end subroutine psi_iswaptranm -subroutine psi_iswaptranv(flag,beta,y,desc_a,work,info) +subroutine psi_dswaptranv(flag,beta,y,desc_a,work,info,data) use psb_error_mod use psb_descriptor_type + use mpi implicit none - include 'mpif.h' integer, intent(in) :: flag integer, intent(out) :: info - integer :: y(:), beta - integer, target :: work(:) + real(kind(1.d0)) :: y(:), beta + real(kind(1.d0)), target :: work(:) type(psb_desc_type) :: desc_a + integer, optional :: data ! locals integer :: icontxt, nprow, npcol, myrow,& & mycol, point_to_proc, nesd, nerv,& - & proc_to_comm, p2ptag, icomm, p2pstat,& + & proc_to_comm, p2ptag, icomm, p2pstat(mpi_status_size),& & idxs, idxr, iret, errlen, ifcomm, rank,& & err_act, totxch, ixrec, i, lw, idx_pt,& & snd_pt, rcv_pt, n integer, pointer, dimension(:) :: bsdidx, brvidx,& - & sdsz, rvsz, prcid, ptp, rvhd, h_idx + & sdsz, rvsz, prcid, ptp, rvhd, d_idx integer :: int_err(5) integer :: blacs_pnum, krecvid, ksendid logical :: swap_mpi, swap_sync, swap_send, swap_recv - integer, pointer, dimension(:) :: sndbuf, rcvbuf + real(kind(1.d0)), pointer, dimension(:) :: sndbuf, rcvbuf character(len=20) :: name, ch_err interface psi_gth @@ -504,7 +554,19 @@ subroutine psi_iswaptranv(flag,beta,y,desc_a,work,info) swap_sync = iand(flag,psb_swap_sync_).ne.0 swap_send = iand(flag,psb_swap_send_).ne.0 swap_recv = iand(flag,psb_swap_recv_).ne.0 - h_idx => desc_a%halo_index + + if(present(data)) then + if(data.eq.psb_comm_halo_) then + d_idx => desc_a%halo_index + else if(data.eq.psb_comm_ovr_) then + d_idx => desc_a%ovrlap_index + else + d_idx => desc_a%halo_index + end if + else + d_idx => desc_a%halo_index + end if + idxs = 0 idxr = 0 totxch = 0 @@ -513,11 +575,11 @@ subroutine psi_iswaptranv(flag,beta,y,desc_a,work,info) n=1 ! prepare info for communications - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) do while (proc_to_comm.ne.-1) if(proc_to_comm .ne. myrow) totxch = totxch+1 - nerv = h_idx(point_to_proc+psb_n_elem_recv_) - nesd = h_idx(point_to_proc+nerv+psb_n_elem_send_) + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) prcid(proc_to_comm) = blacs_pnum(icontxt,proc_to_comm,mycol) ptp(proc_to_comm) = point_to_proc @@ -531,7 +593,7 @@ subroutine psi_iswaptranv(flag,beta,y,desc_a,work,info) idxs = idxs+sdsz(proc_to_comm) point_to_proc = point_to_proc+nerv+nesd+3 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) end do if((idxr+idxs).lt.size(work)) then @@ -547,27 +609,27 @@ subroutine psi_iswaptranv(flag,beta,y,desc_a,work,info) ! Case SWAP_MPI if(swap_mpi) then - + ! gather elements into sendbuffer for swapping point_to_proc = 1 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) do while (proc_to_comm .ne. -1) - nerv = h_idx(point_to_proc+psb_n_elem_recv_) - nesd = h_idx(point_to_proc+nerv+psb_n_elem_send_) - + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) + idx_pt = point_to_proc+psb_elem_recv_ rcv_pt = brvidx(proc_to_comm) - call psi_gth(nerv,h_idx(idx_pt:idx_pt+nerv-1),& + call psi_gth(nerv,d_idx(idx_pt:idx_pt+nerv-1),& & y,rcvbuf(rcv_pt:rcv_pt+nerv-1)) point_to_proc = point_to_proc+nerv+nesd+3 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) end do ! swap elements using mpi_alltoallv call mpi_alltoallv(rcvbuf,rvsz,brvidx,& - & mpi_integer,sndbuf,sdsz,& - & bsdidx,mpi_integer,icomm,iret) + & mpi_double_precision,sndbuf,sdsz,& + & bsdidx,mpi_double_precision,icomm,iret) if(iret.ne.mpi_success) then int_err(1) = iret info=400 @@ -577,95 +639,95 @@ subroutine psi_iswaptranv(flag,beta,y,desc_a,work,info) ! scatter elements from receivebuffer after swapping point_to_proc = 1 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) do while (proc_to_comm .ne. -1) - nerv = h_idx(point_to_proc+psb_n_elem_recv_) - nesd = h_idx(point_to_proc+nerv+psb_n_elem_send_) - + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) + idx_pt = point_to_proc+nerv+psb_elem_send_ snd_pt = bsdidx(proc_to_comm) - call psi_sct(nesd,h_idx(idx_pt:idx_pt+nesd-1),& + call psi_sct(nesd,d_idx(idx_pt:idx_pt+nesd-1),& & sndbuf(snd_pt:snd_pt+nesd-1),beta,y) point_to_proc = point_to_proc+nerv+nesd+3 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) end do else if (swap_sync) then point_to_proc = 1 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) do while (proc_to_comm .ne. -1) - nerv = h_idx(point_to_proc+psb_n_elem_recv_) - nesd = h_idx(point_to_proc+nerv+psb_n_elem_send_) + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) if (proc_to_comm .lt. myrow) then ! First I send idx_pt = point_to_proc+psb_elem_recv_ rcv_pt = brvidx(proc_to_comm) - call psi_gth(nerv,h_idx(idx_pt:idx_pt+nerv-1),& + call psi_gth(nerv,d_idx(idx_pt:idx_pt+nerv-1),& & y,rcvbuf(rcv_pt:rcv_pt+nerv-1)) - call igesd2d(icontxt,nerv,1,rcvbuf(rcv_pt),nerv,proc_to_comm,0) + call dgesd2d(icontxt,nerv,1,rcvbuf(rcv_pt),nerv,proc_to_comm,0) ! Then I receive snd_pt = brvidx(proc_to_comm) - call igerv2d(icontxt,nesd,1,sndbuf(snd_pt),nesd,proc_to_comm,0) + call dgerv2d(icontxt,nesd,1,sndbuf(snd_pt),nesd,proc_to_comm,0) else if (proc_to_comm .gt. myrow) then ! First I receive snd_pt = bsdidx(proc_to_comm) - call igerv2d(icontxt,nesd,1,sndbuf(snd_pt),nesd,proc_to_comm,0) + call dgerv2d(icontxt,nesd,1,sndbuf(snd_pt),nesd,proc_to_comm,0) ! Then I send idx_pt = point_to_proc+psb_elem_recv_ rcv_pt = brvidx(proc_to_comm) - call psi_gth(nerv,h_idx(idx_pt:idx_pt+nerv-1),& + call psi_gth(nerv,d_idx(idx_pt:idx_pt+nerv-1),& & y,rcvbuf(rcv_pt:rcv_pt+nerv-1)) - call igesd2d(icontxt,nerv,1,rcvbuf(rcv_pt),nerv,proc_to_comm,0) + call dgesd2d(icontxt,nerv,1,rcvbuf(rcv_pt),nerv,proc_to_comm,0) else if (proc_to_comm .eq. myrow) then ! I send to myself idx_pt = point_to_proc+psb_elem_recv_ rcv_pt = bsdidx(proc_to_comm) - call psi_gth(nerv,h_idx(idx_pt:idx_pt+nerv-1),& + call psi_gth(nerv,d_idx(idx_pt:idx_pt+nerv-1),& & y,rcvbuf(rcv_pt:rcv_pt+nerv-1)) end if point_to_proc = point_to_proc+nerv+nesd+3 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) end do point_to_proc = 1 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) do while (proc_to_comm .ne. -1) - nerv = h_idx(point_to_proc+psb_n_elem_recv_) - nesd = h_idx(point_to_proc+nerv+psb_n_elem_send_) + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) if(proc_to_comm.ne.myrow) then idx_pt = point_to_proc+nerv+psb_elem_send_ snd_pt = bsdidx(proc_to_comm) - call psi_sct(nesd,h_idx(idx_pt:idx_pt+nesd-1),& + call psi_sct(nesd,d_idx(idx_pt:idx_pt+nesd-1),& & sndbuf(snd_pt:snd_pt+nesd-1),beta,y) else idx_pt = point_to_proc+nerv+psb_elem_send_ rcv_pt = brvidx(proc_to_comm) - call psi_sct(nesd,h_idx(idx_pt:idx_pt+nesd-1),& + call psi_sct(nesd,d_idx(idx_pt:idx_pt+nesd-1),& & rcvbuf(rcv_pt:rcv_pt+nerv-1),beta,y) end if point_to_proc = point_to_proc+nerv+nesd+3 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) end do else if (swap_send .and. swap_recv) then ! First I post all the non blocking receives point_to_proc = 1 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) do while (proc_to_comm .ne. -1) - nerv = h_idx(point_to_proc+psb_n_elem_recv_) - nesd = h_idx(point_to_proc+nerv+psb_n_elem_send_) + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) if(proc_to_comm.ne.myrow) then p2ptag = krecvid(icontxt,proc_to_comm,myrow) snd_pt = brvidx(proc_to_comm) call mpi_irecv(sndbuf(snd_pt),sdsz(proc_to_comm),& - & mpi_integer,prcid(proc_to_comm),& + & mpi_double_precision,prcid(proc_to_comm),& & p2ptag, icomm,rvhd(proc_to_comm),iret) if(iret.ne.mpi_success) then int_err(1) = iret @@ -676,25 +738,25 @@ subroutine psi_iswaptranv(flag,beta,y,desc_a,work,info) end if point_to_proc = point_to_proc+nerv+nesd+3 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) end do ! Then I post all the blocking sends point_to_proc = 1 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) do while (proc_to_comm .ne. -1) - nerv = h_idx(point_to_proc+psb_n_elem_recv_) - nesd = h_idx(point_to_proc+nerv+psb_n_elem_send_) - + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) + idx_pt = point_to_proc+psb_elem_recv_ rcv_pt = brvidx(proc_to_comm) - call psi_gth(nerv,h_idx(idx_pt:idx_pt+nerv-1),& + call psi_gth(nerv,d_idx(idx_pt:idx_pt+nerv-1),& & y,rcvbuf(rcv_pt:rcv_pt+nerv-1)) - + if(proc_to_comm .ne. myrow) then p2ptag=ksendid(icontxt,proc_to_comm,myrow) call mpi_send(rcvbuf(rcv_pt),rvsz(proc_to_comm),& - & mpi_integer,prcid(proc_to_comm),& + & mpi_double_precision,prcid(proc_to_comm),& & p2ptag,icomm,iret) if(iret.ne.mpi_success) then int_err(1) = iret @@ -704,96 +766,131 @@ subroutine psi_iswaptranv(flag,beta,y,desc_a,work,info) end if end if point_to_proc = point_to_proc+nerv+nesd+3 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) end do - do i=1, totxch - call mpi_waitany(nprow,rvhd,ixrec,p2pstat,iret) - if(iret.ne.mpi_success) then - int_err(1) = iret - info=400 - call psb_errpush(info,name,i_err=int_err) - goto 9999 - end if + if(.false.) then + do i=1, totxch + call mpi_waitany(nprow,rvhd,ixrec,p2pstat,iret) + if(iret.ne.mpi_success) then + int_err(1) = iret + info=400 + call psb_errpush(info,name,i_err=int_err) + goto 9999 + end if + + if (ixrec .ne. mpi_undefined) then + ixrec=ixrec-1 ! mpi_waitany returns an 1 to nprow index + point_to_proc = ptp(ixrec) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) + + idx_pt = point_to_proc+nerv+psb_elem_send_ + rcv_pt = bsdidx(proc_to_comm) + call psi_sct(nesd,d_idx(idx_pt:idx_pt+nesd-1),& + & rcvbuf(rcv_pt:rcv_pt+nerv-1),beta,y) + else + int_err(1) = ixrec + info=400 + call psb_errpush(info,name,i_err=int_err) + goto 9999 + end if + end do + + point_to_proc = 1 + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) + do while (proc_to_comm .ne. -1) + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) + + if(proc_to_comm .eq. myrow) then + idx_pt = point_to_proc+nerv+psb_elem_send_ + rcv_pt = brvidx(proc_to_comm) + call psi_sct(nesd,d_idx(idx_pt:idx_pt+nesd-1),& + & rcvbuf(rcv_pt:rcv_pt+nerv-1),beta,y) + end if - if (ixrec .ne. mpi_undefined) then - ixrec=ixrec-1 ! mpi_waitany returns an 1 to nprow index - point_to_proc = ptp(ixrec) - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) - nerv = h_idx(point_to_proc+psb_n_elem_recv_) - nesd = h_idx(point_to_proc+nerv+psb_n_elem_send_) + point_to_proc = point_to_proc+nerv+nesd+3 + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) + end do + + else + + point_to_proc = 1 + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) + do while (proc_to_comm .ne. -1) + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) + + if(proc_to_comm.ne.myrow) then + call mpi_wait(rvhd(proc_to_comm),p2pstat,iret) + if(iret.ne.mpi_success) then + int_err(1) = iret + info=400 + call psb_errpush(info,name,i_err=int_err) + goto 9999 + end if + idx_pt = point_to_proc+nerv+psb_elem_send_ + snd_pt = bsdidx(proc_to_comm) + call psi_sct(nesd,d_idx(idx_pt:idx_pt+nesd-1),& + & sndbuf(snd_pt:snd_pt+nesd-1),beta,y) + else + idx_pt = point_to_proc+nerv+psb_elem_send_ + rcv_pt = brvidx(proc_to_comm) + call psi_sct(nesd,d_idx(idx_pt:idx_pt+nesd-1),& + & rcvbuf(rcv_pt:rcv_pt+nerv-1),beta,y) - idx_pt = point_to_proc+nerv+psb_elem_send_ - rcv_pt = bsdidx(proc_to_comm) - call psi_sct(nesd,h_idx(idx_pt:idx_pt+nesd-1),& - & rcvbuf(rcv_pt:rcv_pt+nerv-1),beta,y) - else - int_err(1) = ixrec - info=400 - call psb_errpush(info,name,i_err=int_err) - goto 9999 - end if - end do - - point_to_proc = 1 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) - do while (proc_to_comm .ne. -1) - nerv = h_idx(point_to_proc+psb_n_elem_recv_) - nesd = h_idx(point_to_proc+nerv+psb_n_elem_send_) - - if(proc_to_comm .eq. myrow) then - idx_pt = point_to_proc+nerv+psb_elem_send_ - rcv_pt = brvidx(proc_to_comm) - call psi_sct(nesd,h_idx(idx_pt:idx_pt+nesd-1),& - & rcvbuf(rcv_pt:rcv_pt+nerv-1),beta,y) - end if + end if - point_to_proc = point_to_proc+nerv+nesd+3 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) - end do + point_to_proc = point_to_proc+nerv+nesd+3 + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) + end do + + end if else if (swap_send) then point_to_proc = 1 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) do while (proc_to_comm .ne. -1) - nerv = h_idx(point_to_proc+psb_n_elem_recv_) - nesd = h_idx(point_to_proc+nerv+psb_n_elem_send_) + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) idx_pt = point_to_proc+psb_elem_recv_ rcv_pt = brvidx(proc_to_comm) - call psi_gth(nerv,h_idx(idx_pt:idx_pt+nerv-1),& - & y,rcvbuf(rcv_pt:rcv_pt+nerv-1)) - call igesd2d(icontxt,nerv,1,rcvbuf(rcv_pt),nerv,proc_to_comm,0) - + call psi_gth(nerv,d_idx(idx_pt:idx_pt+nerv-1),& + & y,rcvbuf(rcv_pt:rcv_pt+nerv-1)) + call dgesd2d(icontxt,nerv,1,rcvbuf(rcv_pt),nerv,proc_to_comm,0) + point_to_proc = point_to_proc+nerv+nesd+3 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) end do else if (swap_recv) then point_to_proc = 1 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) do while (proc_to_comm .ne. -1) - nerv = h_idx(point_to_proc+psb_n_elem_recv_) - nesd = h_idx(point_to_proc+nerv+psb_n_elem_send_) + nerv = d_idx(point_to_proc+psb_n_elem_recv_) + nesd = d_idx(point_to_proc+nerv+psb_n_elem_send_) if(proc_to_comm.ne.myrow) then snd_pt = bsdidx(proc_to_comm) - call igerv2d(icontxt,nesd,1,sndbuf(snd_pt),nesd,proc_to_comm,0) + call dgerv2d(icontxt,nesd,1,sndbuf(snd_pt),nesd,proc_to_comm,0) idx_pt = point_to_proc+psb_elem_recv_ rcv_pt = brvidx(proc_to_comm) - call psi_sct(nesd,h_idx(idx_pt:idx_pt+nesd-1),& + call psi_sct(nesd,d_idx(idx_pt:idx_pt+nesd-1),& & sndbuf(snd_pt:snd_pt+nesd-1),beta,y) else idx_pt = point_to_proc+nerv+psb_elem_send_ rcv_pt = brvidx(proc_to_comm) - call psi_sct(nesd,h_idx(idx_pt:idx_pt+nesd-1),& + call psi_sct(nesd,d_idx(idx_pt:idx_pt+nesd-1),& & rcvbuf(rcv_pt:rcv_pt+nerv-1),beta,y) end if point_to_proc = point_to_proc+nerv+nesd+3 - proc_to_comm = h_idx(point_to_proc+psb_proc_id_) + proc_to_comm = d_idx(point_to_proc+psb_proc_id_) end do end if @@ -808,4 +905,4 @@ subroutine psi_iswaptranv(flag,beta,y,desc_a,work,info) return end if return -end subroutine psi_iswaptranv +end subroutine psi_dswaptranv