为什么Fortran中的单变量Horner比NumPy更快,而双变量Horner则不然

时间:2013-09-09 22:23:01

标签: python arrays numpy fortran f2py

我想在Python中执行多项式计算。 polynomial中的numpy包对我来说不够快。因此,我决定在Fortran中重写几个函数,并使用f2py创建可以轻松导入Python的共享库。目前,我正在根据他们的numpy对应物进行单变量和双变量多项式评估的基准测试。

在单变量例程中,我使用Horner's methodnumpy.polynomial.polynomial.polyval一样。我观察到,随着多项式阶数的增加,Fortran例程比numpy对象更快的因子会增加。

在双变量例程中,我使用Horner的方法两次。首先是y,然后是x。不幸的是,我观察到,为了增加多项式阶数,numpy对应物会赶上并最终超过我的Fortran例程。由于numpy.polynomial.polynomial.polyval2d使用类似于我的方法,我认为第二个观察是奇怪的。

我希望这个结果源于我对Fortran和f2py的经验不足。有人可能有任何线索为什么单变量例程总是显得优越,而双变量例程仅优于低阶多项式?

修改 这是我最新的更新代码,基准脚本和性能图:

polynomial.f95

subroutine polyval(p, x, pval, nx)

    implicit none

    real(8), dimension(nx), intent(in) :: p
    real(8), intent(in) :: x
    real(8), intent(out) :: pval
    integer, intent(in) :: nx
    integer :: i

    pval = 0.0d0
    do i = nx, 1, -1
        pval = pval*x + p(i)
    end do

end subroutine polyval

subroutine polyval2(p, x, y, pval, nx, ny)

    implicit none

    real(8), dimension(nx, ny), intent(in) :: p
    real(8), intent(in) :: x, y
    real(8), intent(out) :: pval
    integer, intent(in) :: nx, ny
    real(8) :: tmp
    integer :: i, j

    pval = 0.0d0
    do j = ny, 1, -1
        tmp = 0.0d0
        do i = nx, 1, -1
            tmp = tmp*x + p(i, j)
        end do
        pval = pval*y + tmp
    end do

end subroutine polyval2

subroutine polyval3(p, x, y, z, pval, nx, ny, nz)

    implicit none

    real(8), dimension(nx, ny, nz), intent(in) :: p
    real(8), intent(in) :: x, y, z
    real(8), intent(out) :: pval
    integer, intent(in) :: nx, ny, nz
    real(8) :: tmp, tmp2
    integer :: i, j, k

    pval = 0.0d0
    do k = nz, 1, -1
        tmp2 = 0.0d0
        do j = ny, 1, -1
            tmp = 0.0d0
            do i = nx, 1, -1
                tmp = tmp*x + p(i, j, k)
            end do
            tmp2 = tmp2*y + tmp
        end do
        pval = pval*z + tmp2
    end do

end subroutine polyval3

benchmark.py (使用此脚本生成图表)

import time
import os

import numpy as np
import matplotlib.pyplot as plt

# Compile and import Fortran module
os.system('f2py -c polynomial.f95 --opt="-O3 -ffast-math" \
--f90exec="gfortran-4.8" -m polynomial')
import polynomial

# Create random x and y value
x = np.random.rand()
y = np.random.rand()
z = np.random.rand()

# Number of repetition
repetition = 10

# Number of times to loop over a function
run = 100

# Number of data points
points = 26

# Max number of coefficients for univariate case
n_uni_min = 4
n_uni_max = 100

# Max number of coefficients for bivariate case
n_bi_min = 4
n_bi_max = 100

# Max number of coefficients for trivariate case
n_tri_min = 4
n_tri_max = 100

# Case on/off switch
case_on = [1, 1, 1]

case_1_done = 0
case_2_done = 0
case_3_done = 0

#=================#
# UNIVARIATE CASE #
#=================#

if case_on[0]:

    # Array containing the polynomial order + 1 for several univariate polynomials
    n_uni = np.array([int(x) for x in np.linspace(n_uni_min, n_uni_max, points)])

    # Initialise arrays for storing timing results
    time_uni_numpy = np.zeros(n_uni.size)
    time_uni_fortran = np.zeros(n_uni.size)

    for i in xrange(len(n_uni)):
        # Create random univariate polynomial of order n - 1
        p = np.random.rand(n_uni[i])

        # Time evaluation of polynomial using NumPy
        dt = []
        for j in xrange(repetition):
            t1 = time.time()
            for r in xrange(run): np.polynomial.polynomial.polyval(x, p)
            t2 = time.time()
            dt.append(t2 - t1)
        time_uni_numpy[i] = np.average(dt[2::])

        # Time evaluation of polynomial using Fortran
        dt = []
        for j in xrange(repetition):
            t1 = time.time()
            for r in xrange(run): polynomial.polyval(p, x)
            t2 = time.time()
            dt.append(t2 - t1)
        time_uni_fortran[i] = np.average(dt[2::])

    # Speed-up factor
    factor_uni = time_uni_numpy / time_uni_fortran

    results_uni = np.zeros([len(n_uni), 4])
    results_uni[:, 0] = n_uni
    results_uni[:, 1] = factor_uni
    results_uni[:, 2] = time_uni_numpy
    results_uni[:, 3] = time_uni_fortran
    print results_uni, '\n'

    plt.figure()
    plt.plot(n_uni, factor_uni)
    plt.title('Univariate comparison')
    plt.xlabel('# coefficients')
    plt.ylabel('Speed-up factor')
    plt.xlim(n_uni[0], n_uni[-1])
    plt.ylim(0, max(factor_uni))
    plt.grid(aa=True)

case_1_done = 1

#================#
# BIVARIATE CASE #
#================#

if case_on[1]:

    # Array containing the polynomial order + 1 for several bivariate polynomials
    n_bi = np.array([int(x) for x in np.linspace(n_bi_min, n_bi_max, points)])

    # Initialise arrays for storing timing results
    time_bi_numpy = np.zeros(n_bi.size)
    time_bi_fortran = np.zeros(n_bi.size)

    for i in xrange(len(n_bi)):
        # Create random bivariate polynomial of order n - 1 in x and in y
        p = np.random.rand(n_bi[i], n_bi[i])

        # Time evaluation of polynomial using NumPy
        dt = []
        for j in xrange(repetition):
            t1 = time.time()
            for r in xrange(run): np.polynomial.polynomial.polyval2d(x, y, p)
            t2 = time.time()
            dt.append(t2 - t1)
        time_bi_numpy[i] = np.average(dt[2::])

        # Time evaluation of polynomial using Fortran
        p = np.asfortranarray(p)
        dt = []
        for j in xrange(repetition):
            t1 = time.time()
            for r in xrange(run): polynomial.polyval2(p, x, y)
            t2 = time.time()
            dt.append(t2 - t1)
        time_bi_fortran[i] = np.average(dt[2::])

    # Speed-up factor
    factor_bi = time_bi_numpy / time_bi_fortran

    results_bi = np.zeros([len(n_bi), 4])
    results_bi[:, 0] = n_bi
    results_bi[:, 1] = factor_bi
    results_bi[:, 2] = time_bi_numpy
    results_bi[:, 3] = time_bi_fortran
    print results_bi, '\n'

    plt.figure()
    plt.plot(n_bi, factor_bi)
    plt.title('Bivariate comparison')
    plt.xlabel('# coefficients')
    plt.ylabel('Speed-up factor')
    plt.xlim(n_bi[0], n_bi[-1])
    plt.ylim(0, max(factor_bi))
    plt.grid(aa=True)

case_2_done = 1

#=================#
# TRIVARIATE CASE #
#=================#

if case_on[2]:

    # Array containing the polynomial order + 1 for several bivariate polynomials
    n_tri = np.array([int(x) for x in np.linspace(n_tri_min, n_tri_max, points)])

    # Initialise arrays for storing timing results
    time_tri_numpy = np.zeros(n_tri.size)
    time_tri_fortran = np.zeros(n_tri.size)

    for i in xrange(len(n_tri)):
        # Create random bivariate polynomial of order n - 1 in x and in y
        p = np.random.rand(n_tri[i], n_tri[i])

        # Time evaluation of polynomial using NumPy
        dt = []
        for j in xrange(repetition):
            t1 = time.time()
            for r in xrange(run): np.polynomial.polynomial.polyval3d(x, y, z, p)
            t2 = time.time()
            dt.append(t2 - t1)
        time_tri_numpy[i] = np.average(dt[2::])

        # Time evaluation of polynomial using Fortran
        p = np.asfortranarray(p)
        dt = []
        for j in xrange(repetition):
            t1 = time.time()
            for r in xrange(run): polynomial.polyval3(p, x, y, z)
            t2 = time.time()
            dt.append(t2 - t1)
        time_tri_fortran[i] = np.average(dt[2::])

    # Speed-up factor
    factor_tri = time_tri_numpy / time_tri_fortran

    results_tri = np.zeros([len(n_tri), 4])
    results_tri[:, 0] = n_tri
    results_tri[:, 1] = factor_tri
    results_tri[:, 2] = time_tri_numpy
    results_tri[:, 3] = time_tri_fortran
    print results_tri

    plt.figure()
    plt.plot(n_bi, factor_bi)
    plt.title('Trivariate comparison')
    plt.xlabel('# coefficients')
    plt.ylabel('Speed-up factor')
    plt.xlim(n_tri[0], n_tri[-1])
    plt.ylim(0, max(factor_tri))
    plt.grid(aa=True)
    print '\n'

case_3_done = 1

#==============================================================================

plt.show()

结果 enter image description here enter image description here enter image description here

EDIT 对steabert提案的更正

subroutine polyval(p, x, pval, nx)

    implicit none

    real*8, dimension(nx), intent(in) :: p
    real*8, intent(in) :: x
    real*8, intent(out) :: pval
    integer, intent(in) :: nx

    integer, parameter :: simd = 8
    real*8 :: tmp(simd), xpower(simd), maxpower
    integer :: i, j, k

    xpower(1) = x
    do i = 2, simd
        xpower(i) = xpower(i-1)*x
    end do
    maxpower = xpower(simd)

    tmp = 0.0d0
    do i = nx+1, simd+2, -simd
        do j = 1, simd
            tmp(j) = tmp(j)*maxpower + p(i-j)*xpower(simd-j+1)
        end do
    end do

    k = mod(nx-1, simd)
    if (k == 0) then
        pval = sum(tmp) + p(1)
    else
        pval = sum(tmp) + p(k+1)
        do i = k, 1, -1
            pval = pval*x + p(i)
        end do
    end if

end subroutine polyval

编辑测试代码以验证上面的代码是否为x>提供了不良结果1

import polynomial as P
import numpy.polynomial.polynomial as PP

import numpy as np

for n in xrange(2,100):
    poly1n = np.random.rand(n)
    poly1f = np.asfortranarray(poly1n)

    x = 2

    print np.linalg.norm(P.polyval(poly1f, x) - PP.polyval(x, poly1n)), '\n'

4 个答案:

答案 0 :(得分:6)

在双变量情况下,p是一个二维数组。这意味着C vs fortran对数组的排序是不同的。默认情况下,numpy函数提供C排序,显然fortran例程使用fortran排序。

f2py非常聪明,可以处理这个问题,并自动在C和fortran格式数组之间进行转换。但是,这会导致一些开销,这是性能降低的可能原因之一。您可以通过在计时例程之外使用p手动将numpy.asfortranarray转换为fortran类型来检查这是否是原因。当然,为了使其有意义,在您的实际用例中,您需要确保输入数组符合fortran顺序。

f2py有一个选项-DF2PY_REPORT_ON_ARRAY_COPY,可以在任何时候复制数组时发出警告。

如果这不是原因,那么您需要考虑更深入的细节,例如您正在使用哪个fortran编译器,以及它正在应用哪种优化。可能会降低速度的一些例子包括在堆上而不是堆栈上分配数组(对malloc进行昂贵的调用),尽管我认为这样的效果对于较大的数组来说不那么重要。

最后,你应该考虑对于大的N,对于双变量拟合,numpy例程已经基本上处于最佳效率的可能性。在这种情况下,numpy例程可能花费大部分时间来运行优化的C例程,并且python代码的开销相比之下变得可以忽略不计。在这种情况下,您不希望您的fortran代码显示任何显着的加速。

答案 1 :(得分:3)

我猜,你的 tmp 数组太大,以至于它需要L2,L3甚至主内存访问而不是缓存。打破这些循环并立即处理它们的块(剥离挖掘)可能会更好。

答案 2 :(得分:1)

您的功能很短,因此通过内联polyval可以获得更好的效果。您还可以通过简单地反转循环来避免计算索引:

subroutine polyval2(p, x, y, pval, nx, ny)

    implicit none

    real(8), dimension(nx, ny), intent(in), target :: p
    real(8), intent(in) :: x, y
    real(8), intent(out) :: pval
    integer, intent(in) :: nx, ny
    real(8) :: tmp
    integer :: i, ii

    pval = 1.d0
    do i = ny, 1
        tmp = 1.d0
        do ii = nx, 1
            tmp = tmp*x + p(ii,i)
        end do
        pval = pval*y + tmp
    end do

end subroutine polyval2

使用此代码,与您发布的原始代码相比,大型阵列的执行时间缩短了约10%。 (我测试了一个纯Fortran程序,代码为Nx = Ny = 1000,gfortran -O3 -funroll-loops

我同意haraldkl的观点,当尺寸变得太大时,性能的急剧下降对于缓存/内存访问模式来说非常典型。条带采矿有帮助,但我不鼓励自己这样做。请改为使用编译器标志:-floop-strip-mine gfortran-O3 ifort -funroll-loops。另外,请尝试循环展开gfortran ifortf2py -c --f90flags="..."

您可以使用{{1}}指定这些标记。

答案 3 :(得分:1)

遵循其他建议,在我测试之前,在计时器之前使用p=np.asfortranarray(p)确实将性能与numpy相提并论。我将双变量工作台的范围扩展到n_bi = np.array([2**i for i in xrange(1, 15)]),这样p矩阵就会大于我的L3缓存大小。

为了进一步优化这一点,我不认为自动编译器选项会有很大帮助,因为内部循环具有依赖性。只有在您手动展开它时,ifort才会向内部循环进行矢量化。 gfortran需要-O3-ffast-math。对于受主存储器带宽限制的矩阵大小,这会使性能优势从numpy提高1到3倍。

更新:将此应用于单变量代码并使用f2py --opt='-O3 -ffast-math' -c -m polynomial polynomial.f90进行编译后,我得到以下有关benchmark.py的来源和结果:

subroutine polyval(p, x, pval, nx)

implicit none

real*8, dimension(nx), intent(in) :: p
real*8, intent(in) :: x
real*8, intent(out) :: pval
integer, intent(in) :: nx

integer, parameter :: simd = 8
real*8 :: tmp(simd), vecx(simd), xfactor
integer :: i, j, k

! precompute factors
do i = 1, simd
    vecx(i)=x**(i-1)
end do
xfactor = x**simd

tmp = 0.0d0
do i = 1, nx, simd
    do k = 1, simd
        tmp(k) = tmp(k)*xfactor + p(nx-(i+k-1)+1)*vecx(simd-k+1)
    end do
end do
pval = sum(tmp)


end subroutine polyval

subroutine polyval2(p, x, y, pval, nx, ny)

implicit none

real*8, dimension(nx, ny), intent(in) :: p
real*8, intent(in) :: x, y
real*8, intent(out) :: pval
integer, intent(in) :: nx, ny

integer, parameter :: simd = 8
real*8 :: tmp(simd), vecx(simd), xfactor
integer :: i, j, k

! precompute factors
do i = 1, simd
    vecx(i)=x**(i-1)
end do
xfactor = x**simd

! horner
pval=0.0d0
do i = 1, ny
    tmp = 0.0d0
    do j = 1, nx, simd
        ! inner vectorizable loop
        do k = 1, simd
            tmp(k) = tmp(k)*xfactor + p(nx-(j+k-1)+1,ny-i+1)*vecx(simd-k+1)
        end do
    end do
    pval = pval*y + sum(tmp)
end do

end subroutine polyval2

更新2 :正如所指出的,此代码不正确,至少在simd不能整除大小时。它只是展示了手动帮助编译器的概念,所以不要像这样使用它。如果大小不是2的幂,则小的余数循环必须处理悬空指数。这样做并不困难,这是单变量情况的正确程序,应该直接将其扩展到双变量:

subroutine polyval(p, x, pval, nx)
implicit none

real*8, dimension(nx), intent(in) :: p
real*8, intent(in) :: x
real*8, intent(out) :: pval
integer, intent(in) :: nx

integer, parameter :: simd = 4
real*8 :: tmp(simd), vecx(simd), xfactor
integer :: i, j, k, nr

! precompute factors
do i = 1, simd
    vecx(i)=x**(i-1)
end do
xfactor = x**simd

! check remainder
nr = mod(nx, simd)

! horner
tmp = 0.0d0
do i = 1, nx-nr, simd
    do k = 1, simd
        tmp(k) = tmp(k)*xfactor + p(nx-(i+k-1)+1)*vecx(simd-k+1)
    end do
end do
pval = sum(tmp)

! do remainder
pval = pval * x**nr
do i = 1, nr
    pval = pval + p(i) * vecx(i)
end do
end subroutine polyval

univariate

bivariate

此外,人们应该小心尺寸非常小,因为时间太小,无法获得准确的性能。此外,相对于numpy的相对时间可能是欺骗性的,因为numpy的绝对时间可能非常糟糕。以下是最大案例的时间安排:

对于nx = 2 20的单变量,numpy的时间为1.21 s,自定义fortran版本的时间为1.69e-3 s。对于nx ny = 2 20的双变量,numpy的时间为8e-3 s,自定义版本的时间为1.68e-3 s。当总nx ny大小相同时,单变量和双变量的时间是相同的这一事实是非常重要的,因为它支持代码在内存带宽限制附近执行的事实。

更新3 :使用适用于较小尺寸的新python脚本,simd=4我获得以下效果:

enter image description here

enter image description here

更新4 :至于正确性,结果在双精度精度内是相同的,如果为单变量示例运行此python代码,可以看到:

import polynomial as P
import numpy.polynomial.polynomial as PP

import numpy as np

for n in xrange(2,100):
    poly1n = np.random.rand(n)
    poly1f = np.asfortranarray(poly1n)

    x = 2

    print "%18.14e" % P.polyval(poly1f, x)
    print "%18.14e" % PP.polyval(x, poly1n)
    print (P.polyval(poly1f, x) - PP.polyval(x, poly1n))/PP.polyval(x,poly1n), '\n'