改善Fortran矩阵的指数性能(Expokit比Matlab,Python慢​​)

时间:2018-07-26 13:29:25

标签: python matlab fortran gfortran

我正在进行一个仿真,其中瓶颈正在执行大量复杂的双精度矩阵指数,并且发现尽管Fortran(Expokit)在小型矩阵上运行良好,但对于较大的矩阵,其性能却不如Matlab或Python

尽管其中需要更大的矩阵来显示性能差异,但是我在下面包含了一个显示类似行为的模型程序。查看探查器和source code,似乎Expokit大部分时间都花在调用zgemm()上,所以我唯一的想法是我的BLAS安装有问题。否则我不明白为什么Fortran的性能会比Matlab或Python差。对于提高Fortran矩阵指数代码性能的任何见解,我将不胜感激。

  

10000个矩阵(4x4、8x8、30x30、60x60、80x80)的结果:

     

Matlab:0.91、0.97、2.36、5.45、8.69

     

Python(s):2.59、2.89、9.70、35.4、72.7

     

Fortran,Expokit:0.037、0.12、4.14、30.6、74.9

     

具有8个内核的Fortran,Expokit,OpenMP:0.0039、0.016、0.52、3.87,   9.53

Fortran代码:

    subroutine expokit_test()

    use omp_lib
    use iso_fortran_env
    implicit none

    integer, parameter :: wp = selected_real_kind(15, 307), size=80
    complex(wp), parameter :: i = (0, 1._wp)
    integer :: count, a, b
    real(wp) :: wtime
    complex(wp) :: mat_exp(size, size),  mat(size, size), val

    val = 1E-8_wp
    mat = 0._wp

    do a = 1, size
        do b = 1, size
            mat(a, b) = a * b
        end do
    end do

    call omp_set_num_threads(8)
    wtime = omp_get_wtime()
    !$omp parallel do default(private) &
    !$omp& shared(mat, val)
    do count = 1, int(1E4)
        mat_exp = expm_complex(-i * mat * val)
    end do
    !$omp end parallel do
    wtime = omp_get_wtime () - wtime
    write(6, *) 'expm_complex', sngl(wtime)

end subroutine expokit_test

function expm_complex(A) result(B)
        ! Calculate matrix exponential of complex matrix A using Expokit

        use iso_fortran_env
        implicit none

        integer, parameter :: wp = selected_real_kind(15, 307)
        complex(wp), dimension(:, :), intent(in) :: A
        complex(wp), dimension(size(A, 1), size(A, 2)) :: B

        integer, parameter :: ideg = 2 ! Pade approximation, 6 is reccomended but 2 appears to be stable
        complex(wp) :: t = 1._wp
        complex(wp), dimension(4 * size(A, 1) * size(A, 2) + ideg + 1) :: wsp
        integer, dimension(size(A, 1)) :: iwsp
        integer :: iexp, ns, iflag, n
        n = size(A, 1)

        call ZGPADM(ideg, n, t, A, n, wsp, size(wsp, 1), iwsp, iexp, ns, iflag)
        B = reshape(wsp(iexp : iexp + n * n - 1), [n, n])

    end function expm_complex

Matlab代码:

size = 80;

for a=1:size
    for b=1:size
        A(a, b) = 1E-9 + (a * b);
    end

end

tic
for test=1:1E4
    t=expm(-1i*A*1E-8);
end
toc

Python代码:

size = 80
mat = np.ones((size, size))

for a in range(0, size):
    for b in range(0, size):
        mat[a, b] = ((a+1) * (b+1))

mat = mat + 1E-9

start = time.time()
for loop in range(0, int(1E4)):
    test = la.expm(-1j * mat * 1E-8)


end = time.time() - start
print('time taken', end)

0 个答案:

没有答案