对于16字节实数,MPI_AllReduce的奇怪结果

Vla*_*r F 7 fortran mpi openmpi quadruple-precision

编译器:gfortran-4.8.5

MPI库:OpenMPI-1.7.2(预装OpenSuSE 13.2)

这个程序:

  use mpi
  implicit none

  real*16 :: x
  integer :: ierr, irank, type16

  call MPI_Init(ierr)

  call MPI_Comm_Rank(MPI_Comm_World, irank, ierr)

  if (irank+1==1) x = 2.1
  if (irank+1==8) x = 2.8
  if (irank+1==7) x = 5.2
  if (irank+1==4) x = 6.7
  if (irank+1==6) x = 6.5
  if (irank+1==3) x = 5.7
  if (irank+1==2) x = 4.0
  if (irank+1==5) x = 6.8

  print '(a,i0,a,f3.1)', "rank+1: ",irank+1," x: ",x

  call MPI_AllReduce(MPI_IN_PLACE, x, 1, MPI_REAL16, MPI_MAX, MPI_Comm_World, ierr)

  if (irank==0) print '(i0,a,f3.1)', irank+1," max x: ", x

  call MPI_Finalize(ierr)
end
Run Code Online (Sandbox Code Playgroud)

我也试过real(16),real(kind(1.q0)).real(real128)实际上与real*10此编译器等效.

结果是:

> mpif90 reduce16.f90 
> mpirun -n 8 ./a.out 
rank+1: 1 x: 2.1
rank+1: 2 x: 4.0
rank+1: 3 x: 5.7
rank+1: 4 x: 6.7
rank+1: 5 x: 6.8
rank+1: 6 x: 6.5
rank+1: 7 x: 5.2
rank+1: 8 x: 2.8
1 max x: 2.8
Run Code Online (Sandbox Code Playgroud)

该程序找到real*10保持的真正最大值MPI_REAL16.MPI规范(3.1,第628和674页)如果MPI_REAL16对应real*16real(real128)不相同则不是很清楚.

此外,假设MPI_REAL16实际上real(real128)并尝试在程序中使用它会导致不同的问题:

Error: There is no specific subroutine for the generic 'mpi_recv' at (1)
Error: There is no specific subroutine for the generic 'mpi_send' at (1)
Run Code Online (Sandbox Code Playgroud)

这不会发生real*16.(无论那个人应该能够传递任何位模式,所以这个检查是多余的)

使用16字节实数的正确方法是什么?OpenMPI库是错误的吗?

Jef*_*eff 1

虽然这应该在每个 MPI 实现中都能正常工作,但一个简单的解决方法是为这种用 Fortran 编写的类型实现用户定义的归约,因此在 C 中实现它没有问题(这就是 MPICH 和 OpenMPI 尝试的方式)做所有事情,因此当 C 无法重现 Fortran 的行为时会出现问题)。

下面是实现这一点的尝试。这是 Fortran 中用户定义的缩减。我确信经验丰富的现代 Fortran 程序员可以做得更好。

  subroutine sum_real16(iv,iov,n)
    implicit none
    integer, intent(in) ::  n
    real*16, intent(in) :: iv(:)
    real*16, intent(inout) :: iov(:)
    integer :: i
    do i = 1,n
      iov(i) = iov(i) + iv(i)
    enddo
  end subroutine sum_real16
  subroutine reduce_sum_real16(iv, iov, n, dt)
    use, intrinsic ::  iso_c_binding, only : c_ptr
    use mpi_f08
    implicit none
    type(c_ptr), value ::  iv, iov
    integer ::  n
    type(MPI_Datatype) ::  dt
    if ( dt .eq. MPI_REAL16 ) then
        call sum_real16(iv,iov,n)
    endif
  end subroutine reduce_sum_real16
  program test_reduce_sum_real16
    use, intrinsic ::  iso_c_binding
    use mpi_f08
    implicit none
    integer, parameter ::  n = 10
    real*16 :: output(n)
    real*16 :: input(n)
    real*16 :: error
    integer :: me, np
    procedure(MPI_User_function) :: reduce_sum_real16
    type(MPI_Op) :: mysum
    integer :: i
    call MPI_Init()
    call MPI_Comm_rank(MPI_COMM_WORLD,me)
    call MPI_Comm_size(MPI_COMM_WORLD,np)
    output = 0.0
    input  = 1.0*me
    call MPI_Op_create(reduce_sum_real16,.true.,mysum)
    call MPI_Allreduce(input,output,n,MPI_REAL16,mysum,MPI_COMM_WORLD)
    error = 0.0
    do i = 1,n
      error = error + (output(i)-1.0*np)
    enddo
    if (error.gt.0.0) then
        print*,'SAD PANDA = ',error
        call MPI_Abort(MPI_COMM_SELF,1)
    endif
    call MPI_Op_free(mysum)
    call MPI_Finalize()
  end program test_reduce_sum_real16
Run Code Online (Sandbox Code Playgroud)

该程序使用 Intel 16 Fortran 编译器和 MPICH 3.2+ 正常返回。显然我没有正确使用 I/O,所以我对这个程序正确性的信心不如我可以将所有结果写入标准输出那样高。