3

我遇到了两者的奇怪行为，gfortran并且ifort在调查为什么reshape比使用循环的幼稚实现慢得多的过程中：

my_reshape我为两个整形函数（reshape3Dto1D和reshape1Dto3D）定义了一个接口。当直接调用接口函数而不是成员函数时，我注意到加速了 10-40%！更改调用顺序、级别优化，甚至编译器都没有改变这种行为。

我犯了一个错误，还是有人对此有解释？

这是（简化的）代码：

reshapetest_simple.F90：

module test
  interface my_reshape
    module procedure :: reshape3Dto1D
  end interface
contains
  function reshape3Dto1D( mat, dims )
    use, intrinsic :: iso_fortran_env
    implicit none
    integer, parameter  :: cp = REAL64
    real(cp),intent(in) :: mat(:,:,:)
    integer,intent(in)  :: dims(1)
    real(cp)            :: reshape3Dto1D(dims(1))
    integer             :: x,y,z, i

    i=0
    do z=1,size(mat,3)
      do y=1,size(mat,2)
        do x=1,size(mat,1)
          i=i+1
          reshape3Dto1D(i) = mat(x,y,z)
        enddo ! y
      enddo ! y
    enddo !z
  end function
end module

program changeDim
  use test
  use omp_lib
  use, intrinsic :: iso_fortran_env
  implicit none
  integer, parameter :: cp = REAL64

  real(REAL64) :: t1, t2, t3, t4
  integer,parameter   :: dimX=100, dimY=100, dimZ=100
  integer,parameter   :: dimProduct = dimX*dimY*dimZ
  integer             :: stat
  real(cp),pointer,contiguous    :: matrix3d(:,:,:), matrix1d(:)

  allocate( matrix3d(dimX,dimY,dimZ), matrix1d(dimProduct), stat=stat )
  if (stat/=0) stop 'Cannot allocate memory';
  call random_number(matrix3d)
  matrix1d = 0._cp

  ! (1) Naive copy using a function
  t1 = omp_get_wtime()
  matrix1d = reshape3Dto1D( matrix3d, [ dimProduct ] )
  t2 = omp_get_wtime()
  ! (2) Reshape
  matrix1d = reshape( matrix3d, [ dimProduct ] )
  t3 = omp_get_wtime()
  ! (3) Same as (1), but using the interface
  matrix1d = my_reshape( matrix3d, [ dimProduct ] )
  t4 = omp_get_wtime()

  write(*,*) 'Reshape:             ',t3-t2
  write(*,*) 'Naive fct direct:    ',t2-t1
  write(*,*) 'Naive fct interface: ',t4-t3

  deallocate( matrix3d, matrix1d )
end program

我用过gfortran 4.8.1，和ifort 13.1.3。二进制文件是使用编译的

ifort -o reshape-ifort -openmp reshapetest_simple.F90 -O3
gfortran -o reshape-gfortran -fopenmp reshapetest_simple.F90 -O3

并给出了以下结果：

OMP_NUM_THREADS=1 ./reshape-gfortran 
 Reshape:                6.8527370000310839E-003
 Naive fct direct:       5.0175579999631736E-003
 Naive fct interface:    4.6131109999123510E-003
OMP_NUM_THREADS=1 ./reshape-ifort 
 Reshape:               3.495931625366211E-003
 Naive fct direct:      5.089998245239258E-003
 Naive fct interface:   3.136873245239258E-003

顺便说一句：我知道对于这种重塑最好使用指针来避免复制数组......

4

1 回答 1

1

影响因素

此行为与问题大小 (100x100x100) 和real,pointer,contiguous.

实验

代替

real(cp),pointer,contiguous :: matrix3d(:,:,:), matrix1d(:)

和

real(cp),  allocatable      :: matrix3d(:,:,:), matrix1d(:)

在我的测试中，随着 dimX、dimY 和 dimZ 增加到 200，这种行为或多或少会消失。

使用 GNU Fortran 4.7.3 计时

指针，暗淡 = 100

Reshape:                5.5749416351318359E-003
Naive fct direct:       6.2539577484130859E-003
Naive fct interface:    2.8791427612304688E-003

可分配，暗淡 = 100

Reshape:                4.2719841003417969E-003
Naive fct direct:       1.4619827270507813E-003
Naive fct interface:    1.3799667358398438E-003

指针，暗淡 = 200

Reshape:                4.2979001998901367E-002
Naive fct direct:       5.7554006576538086E-002
Naive fct interface:    3.6303043365478516E-002

可分配，暗淡 = 200

Reshape:                4.3957948684692383E-002
Naive fct direct:       1.1255979537963867E-002
Naive fct interface:    1.1703014373779297E-002
于 2013-10-08T18:27:02.277 回答