使用MPI_Gatherv for Fortran

时间:2011-06-28 18:31:14

标签: mpi parallel-processing fortran90

此问题遵循 MPI_type_create_subarray和MPI_Gather 上的现有主题。我的目标是使用MPI_Type_Create_Subarray和Fortran 90中的MPI_Gatherv将所有从属进程(数量为4个)的更大数组的子数组收集到主进程(rank = 0)上的更大数组中。这有助于我理解MPI_Gatherv用于我的其他项目。以下是我的示例代码:

    program main
    implicit none
    include "mpif.h"
    integer :: ierr, myRank, nProcs
    integer :: sendsubarray, recvsubarray, resizedrecvsubarray
    integer, dimension(2) :: starts,sizes,subsizes
    integer, dimension(:), allocatable :: counts, disps
    integer, parameter :: nx_glb=10, ny_glb=10, nx=5, ny=5
    integer, dimension(:,:), target, allocatable :: mat, matG
    integer, pointer :: sendPtr(:,:), recvPtr(:,:)
    integer :: i, j

    call mpi_init(ierr)
    call mpi_comm_rank(mpi_comm_world, myRank, ierr)
    call mpi_comm_size(mpi_comm_world, nProcs, ierr)

    sizes(1)=nx+2; sizes(2)=ny+2
    subsizes(1)=nx; subsizes(2)=ny
    starts(1)=2; starts(2)=2
    call mpi_type_create_subarray(2, sizes, subsizes, starts, mpi_order_fortran, &
                                mpi_integer, sendsubarray, ierr)
    call mpi_type_commit(sendsubarray,ierr)

    allocate(mat(1:nx+2,1:ny+2))
    do j=1, ny+2
     do i=1, nx+2
      if(i.eq.1 .or. i.eq.nx+2 .or. j.eq.1 .or. j.eq.ny+2) then
       mat(i,j)=1000
      else
       mat(i,j) = myRank
      end if
     end do
    end do

    sendPtr=>mat
    if(myRank.eq.0) then
     allocate(matG(nx_glb,ny_glb))
     matG=1000
     sizes(1)=nx_glb; sizes(2)=ny_glb
     subsizes(1)=nx; subsizes(2)=ny
     starts(1)=1; starts(2)=1
     call mpi_type_create_subarray(2, sizes, subsizes, starts, mpi_order_fortran, &
                                   mpi_integer, recvsubarray, ierr)
     call mpi_type_commit(recvsubarray, ierr)
     call mpi_type_create_resized(recvsubarray, 1, sizeof(i), resizedrecvsubarray, ierr)
     call mpi_type_commit(resizedrecvsubarray,ierr)
     recvPtr=>matG
    end if

    counts(1:4) = (/1, 1, 1, 1/)
    disps(1:4) = (/0, 5, 50, 55/)
    call mpi_gatherv(sendPtr,1,sendsubarray,recvPtr,counts,disps,resizedrecvsubarray, &
                     0,mpi_comm_world,ierr)

    if(myRank.eq.0) then
     do i=1, nx_glb
      write(1000,*) (matG(i,j),j=1, ny_glb)
     end do
    end if

    call mpi_finalize(ierr)

    end program main

但是,执行此代码会产生forrtl: severe(174): SIGSEGV, segmentation fault occurred

似乎我试图指向一个在收集时未初始化或声明的数组的变量/位置。我试图以多种方式进行调试,但徒劳无功。

非常感谢提前。

1 个答案:

答案 0 :(得分:1)

当你看到这里的主要问题时,你会踢自己;你没有分配计数或分配。

顺便说一句,我强烈建议您使用use mpi而不是include mpif.h; use语句(在隐式none之前)引入了具有更好的类型检查的F90接口。当您这样做时,您还会看到,对于您的类型创建已调整大小,您需要kind mpi_address_kind的整数。

更新

好的,所以对于如何进行gatherv这个更大的问题,你的事情基本上是对的,但是你是对的,启动,disps等必须是零索引,而不是1,因为实际的MPI库是从C的角度来看,即使使用FORTRAN绑定也是如此。所以对于sentubarray,start必须是[1,1];对于recv子阵列,它必须是[0,0],并且resize,start必须为0且extent必须是sizeof(type)(并且这两者必须是mpi_address_kind类型的整数)。

我正在使用这些更新附加您的代码版本,并且底层数组具有类型字符,因此更容易打印出诊断并查看正在发生的事情:

program main
    use mpi
    implicit none
    integer :: ierr, myRank, nProcs
    integer :: sendsubarray, recvsubarray, resizedrecvsubarray
    integer, dimension(2) :: starts,sizes,subsizes
    integer, dimension(:), allocatable :: counts, disps
    integer, parameter :: nx_glb=10, ny_glb=10, nx=5, ny=5
    character, dimension(:,:), target, allocatable :: mat, matG
    character :: c
    integer :: i, j, p
    integer(kind=mpi_address_kind) :: start, extent

    call mpi_init(ierr)
    call mpi_comm_rank(mpi_comm_world, myRank, ierr)
    call mpi_comm_size(mpi_comm_world, nProcs, ierr)

    sizes(1)=nx+2; sizes(2)=ny+2
    subsizes(1)=nx; subsizes(2)=ny
    starts(1)=1; starts(2)=1
    call mpi_type_create_subarray(2, sizes, subsizes, starts, mpi_order_fortran, &
                                mpi_character, sendsubarray, ierr)
    call mpi_type_commit(sendsubarray,ierr)

    allocate(mat(1:nx+2,1:ny+2))
    mat='.'
    forall (i=2:nx+1,j=2:ny+1) mat(i,j)=ACHAR(ICHAR('0')+myRank)

    if(myRank.eq.0) then
     allocate(matG(nx_glb,ny_glb))
     matG='.'
     sizes(1)=nx_glb; sizes(2)=ny_glb
     subsizes(1)=nx; subsizes(2)=ny
     starts(1)=0; starts(2)=0
     call mpi_type_create_subarray(2, sizes, subsizes, starts, mpi_order_fortran, &
                                   mpi_character, recvsubarray, ierr)
     call mpi_type_commit(recvsubarray, ierr)
     extent = sizeof(c)
     start = 0
     call mpi_type_create_resized(recvsubarray, start, extent, resizedrecvsubarray, ierr)
     call mpi_type_commit(resizedrecvsubarray,ierr)
    end if

    allocate(counts(4),disps(4))
    counts(1:4) = (/1, 1, 1, 1/)
    disps(1:4) = (/0, 5, 50, 55/)
    call mpi_gatherv(mat,1,sendsubarray,matG,counts,disps,resizedrecvsubarray, &
                     0,mpi_comm_world,ierr)

    do p=0,nProcs
      if (myRank == p) then
         print *, 'Local array for rank ', myRank
         do i=1, nx+2
          print *, (mat(i,j),j=1,ny+2)
         end do
      endif
      call MPI_Barrier(MPI_COMM_WORLD,ierr)
    enddo
    if(myRank.eq.0) then
     print *, 'Global array: '
     do i=1, nx_glb
      print *, (matG(i,j),j=1, ny_glb)
     end do
    end if

    call mpi_finalize(ierr)

end program main

输出:

 Local array for rank            0
 .......
 .00000.
 .00000.
 .00000.
 .00000.
 .00000.
 .......
 Local array for rank            1
 .......
 .11111.
 .11111.
 .11111.
 .11111.
 .11111.
 .......
 Local array for rank            2
 .......
 .22222.
 .22222.
 .22222.
 .22222.
 .22222.
 .......
 Local array for rank            3
 .......
 .33333.
 .33333.
 .33333.
 .33333.
 .33333.
 .......
 Global array: 
 0000022222
 0000022222
 0000022222
 0000022222
 0000022222
 1111133333
 1111133333
 1111133333
 1111133333
 1111133333

......有意义吗?这非常类似于这个问题的C版本,这里已经回答了问题(MPI_Type_create_subarray and MPI_Gather),但你已经把事情大部分都解决了......

哦,是的,还有一件事 - 你实际上并不需要在Fortran中设置指向send / recv数据的指针。在C中,您需要显式地将指针传递给数据数组;在fortran中,你可以传递数组(并且它们已经通过引用传递,例如相当于C传递指向变量的指针)。所以你可以传递数组。