Python Cython 0.2:prange意外减慢代码速度

在两种情况下考虑优化cython代码：

    for j in xrange(8):
        for x in xrange(1, 600):
            tmp[j] = 0.0
            for y in xrange(1, 800):
                tmp[j] += mag[j, x - 1, y - 1]
                hgi_out[j, x, y] = tmp[j] + hgi_out[j, x - 1, y]       

以及：

要测试上述代码，请尝试：

mg2 = np.random.rand(9, 600, 800).astype(F64TYPE)
hg2 = np.zeros((9, mg2.shape[1] + 1, mg2.shape[2] + 1), dtype=F64TYPE)
print timeit(lambda:hog_integral_b(mg2, hgi_out=hg2), number=10)

更新：

好的，我仔细查看了setup.py编译器选项：

from distutils.core import setup
from distutils.extension import Extension
from Cython.Distutils import build_ext
import numpy as np

ext_modules = [Extension("hog_cy", ["hog_cy.pyx"],
                         #extra_compile_args = ["-O3", "-fopenmp", "-fno-strict-aliasing"],
                         extra_compile_args = ["-O3", "-fopenmp"],
                         extra_link_args=["-fopenmp"]
                        )]

setup (
    name = 'performance test app',
    cmdclass = {'build_ext': build_ext},
    include_dirs = [np.get_include()],
    ext_modules = ext_modules,
)

---

选项

-fno strict aliasing

似乎造成了问题，一旦关闭，我就没有加速，但也没有损失

你正在制造一场吉尔之战，因为普兰奇不在诺吉尔区块内。代码中没有并发性，只有多个线程争夺GIL所有权：

cimport cython
from cython.parallel cimport prange, parallel

cdef void native_hog_integral_b(F64TYPE_t [:, :, ::1] mag, 
                                F64TYPE_t [:, :, ::1] hgi_out):

    cdef DSIZE_t m, n, j, dims = mag.shape[0]
    cdef F64TYPE_t val = 0
    m, n = mag.shape[1] + 1, mag.shape[2] + 1
    with nogil, parallel():
        cdef DSIZE_t x, y
        cdef F64TYPE_t tmp
        for j in prange(dims):
            for x in range(1, m):
                tmp = 0.0
                for y in range(1, n):
                    tmp += mag[j, x - 1, y - 1]
                    hgi_out[j, x, y] = tmp + hgi_out[j, x - 1, y]

---

你能发布一个完整的例子吗？我希望问题在发布整个代码时变得更清楚。很抱歉，这么迂腐，但是你能添加一个函数来调用

hog\u integral\u b

，使用正确形状和数据类型的样本（随机）数据吗？好的，我会这样做，给我一些时间好的，在本机函数末尾用测试代码回答更新（native_hog_integral_b）在其签名的末尾用nogil声明，同样的声明模式也适用于其他函数（上面的代码中没有显示）。我想可能编译器没有将某些内部变量视为局部线程安全的（例如：tmp）但尚未尝试。函数的nogil声明不要求释放GIL，也不释放GIL，它允许在nogil块内使用该函数。您可以在按住GIL的同时使用nogil函数。在键入“with nogil:”之前，不会释放GIL

from distutils.core import setup
from distutils.extension import Extension
from Cython.Distutils import build_ext
import numpy as np

ext_modules = [Extension("hog_cy", ["hog_cy.pyx"],
                         #extra_compile_args = ["-O3", "-fopenmp", "-fno-strict-aliasing"],
                         extra_compile_args = ["-O3", "-fopenmp"],
                         extra_link_args=["-fopenmp"]
                        )]

setup (
    name = 'performance test app',
    cmdclass = {'build_ext': build_ext},
    include_dirs = [np.get_include()],
    ext_modules = ext_modules,
)

cimport cython
from cython.parallel cimport prange, parallel

cdef void native_hog_integral_b(F64TYPE_t [:, :, ::1] mag, 
                                F64TYPE_t [:, :, ::1] hgi_out):

    cdef DSIZE_t m, n, j, dims = mag.shape[0]
    cdef F64TYPE_t val = 0
    m, n = mag.shape[1] + 1, mag.shape[2] + 1
    with nogil, parallel():
        cdef DSIZE_t x, y
        cdef F64TYPE_t tmp
        for j in prange(dims):
            for x in range(1, m):
                tmp = 0.0
                for y in range(1, n):
                    tmp += mag[j, x - 1, y - 1]
                    hgi_out[j, x, y] = tmp + hgi_out[j, x - 1, y]

cimport cython
from cython.parallel cimport prange, parallel

cdef void native_hog_integral_b(F64TYPE_t [:, :, ::1] mag, 
                                F64TYPE_t [:, :, ::1] hgi_out):

    cdef DSIZE_t m, n, j, dims = mag.shape[0]
    cdef F64TYPE_t val = 0
    m, n = mag.shape[1] + 1, mag.shape[2] + 1
    with nogil, parallel():
        cdef DSIZE_t x, y
        cdef F64TYPE_t tmp
        for j in prange(dims):
            for x in range(1, m):
                tmp = 0.0
                for y in range(1, n):
                    tmp += mag[j, x - 1, y - 1]
                    hgi_out[j, x, y] = tmp + hgi_out[j, x - 1, y]