从2021年的答案开始，我还没有发现二进制数组方法对10亿以下的质数有利。

但我可以用几个技巧将质数从2加速到接近x2:

使用numexpr库将numpy表达式转换为分配较少的紧循环 取代np。有更快的选择 以某种方式操作筛选的前9个元素，因此不需要改变数组的形状

总之，在我的机器上，质数< 10亿的时间从25秒变成了14.5秒

import numexpr as ne
import numpy as np

def primesfrom2to_numexpr(n):
    # https://stackoverflow.com/questions/2068372/fastest-way-to-list-all-primes-below-n-in-python/3035188#3035188
    """ Input n>=24, Returns a array of primes, 2 <= p < n + a few over"""
    sieve = np.zeros((n // 3 + (n % 6 == 2))//4+1, dtype=np.int32)
    ne.evaluate('sieve + 0x01010101', out=sieve)
    sieve = sieve.view('int8')
    #sieve = np.ones(n // 3 + (n % 6 == 2), dtype=np.bool_)
    sieve[0] = 0
    for i in np.arange(int(n ** 0.5) // 3 + 1):
        if sieve[i]:
            k = 3 * i + 1 | 1
            sieve[((k * k) // 3)::2 * k] = 0
            sieve[(k * k + 4 * k - 2 * k * (i & 1)) // 3::2 * k] = 0
    sieve[[0,8]] = 1
    result = np.flatnonzero(sieve)
    ne.evaluate('result * 3 + 1 + result%2', out=result)
    result[:9] = [2,3,5,7,11,13,17,19,23]
    return result

对于足够大的N，真正最快的解决方案是下载一个预先计算的质数列表，将其存储为元组，并执行如下操作:

for pos,i in enumerate(primes):
    if i > N:
        print primes[:pos]

如果只有N >个质数[-1]，则计算更多的质数并将新列表保存在代码中，以便下次同样快。

要跳出思维定势。

随着时间的推移，我收集了几个质数筛子。我电脑上最快的是这样的:

from time import time
# 175 ms for all the primes up to the value 10**6
def primes_sieve(limit):
    a = [True] * limit
    a[0] = a[1] = False
    #a[2] = True
    for n in xrange(4, limit, 2):
        a[n] = False
    root_limit = int(limit**.5)+1
    for i in xrange(3,root_limit):
        if a[i]:
            for n in xrange(i*i, limit, 2*i):
                a[n] = False
    return a

LIMIT = 10**6
s=time()
primes = primes_sieve(LIMIT)
print time()-s

对于最快的代码，numpy解决方案是最好的。不过，出于纯粹的学术原因，我发布了我的纯python版本，它比上面发布的食谱版本快不到50%。由于我将整个列表放在内存中，所以需要足够的空间来容纳所有内容，但它的可伸缩性似乎相当好。

def daniel_sieve_2(maxNumber):
    """
    Given a number, returns all numbers less than or equal to
    that number which are prime.
    """
    allNumbers = range(3, maxNumber+1, 2)
    for mIndex, number in enumerate(xrange(3, maxNumber+1, 2)):
        if allNumbers[mIndex] == 0:
            continue
        # now set all multiples to 0
        for index in xrange(mIndex+number, (maxNumber-3)/2+1, number):
            allNumbers[index] = 0
    return [2] + filter(lambda n: n!=0, allNumbers)

结果是:

>>>mine = timeit.Timer("daniel_sieve_2(1000000)",
...                    "from sieves import daniel_sieve_2")
>>>prev = timeit.Timer("get_primes_erat(1000000)",
...                    "from sieves import get_primes_erat")
>>>print "Mine: {0:0.4f} ms".format(min(mine.repeat(3, 1))*1000)
Mine: 428.9446 ms
>>>print "Previous Best {0:0.4f} ms".format(min(prev.repeat(3, 1))*1000)
Previous Best 621.3581 ms

我测试了一些unutbu的功能，我用饥饿的百万数字计算它

获胜者是使用numpy库的函数，

注意:做一个内存利用率测试也很有趣:)

示例代码

完整的代码在我的github存储库

#!/usr/bin/env python

import lib
import timeit
import sys
import math
import datetime

import prettyplotlib as ppl
import numpy as np

import matplotlib.pyplot as plt
from prettyplotlib import brewer2mpl

primenumbers_gen = [
    'sieveOfEratosthenes',
    'ambi_sieve',
    'ambi_sieve_plain',
    'sundaram3',
    'sieve_wheel_30',
    'primesfrom3to',
    'primesfrom2to',
    'rwh_primes',
    'rwh_primes1',
    'rwh_primes2',
]

def human_format(num):
    # https://stackoverflow.com/questions/579310/formatting-long-numbers-as-strings-in-python?answertab=active#tab-top
    magnitude = 0
    while abs(num) >= 1000:
        magnitude += 1
        num /= 1000.0
    # add more suffixes if you need them
    return '%.2f%s' % (num, ['', 'K', 'M', 'G', 'T', 'P'][magnitude])


if __name__=='__main__':

    # Vars
    n = 10000000 # number itereration generator
    nbcol = 5 # For decompose prime number generator
    nb_benchloop = 3 # Eliminate false positive value during the test (bench average time)
    datetimeformat = '%Y-%m-%d %H:%M:%S.%f'
    config = 'from __main__ import n; import lib'
    primenumbers_gen = {
        'sieveOfEratosthenes': {'color': 'b'},
        'ambi_sieve': {'color': 'b'},
        'ambi_sieve_plain': {'color': 'b'},
         'sundaram3': {'color': 'b'},
        'sieve_wheel_30': {'color': 'b'},
# # #        'primesfrom2to': {'color': 'b'},
        'primesfrom3to': {'color': 'b'},
        # 'rwh_primes': {'color': 'b'},
        # 'rwh_primes1': {'color': 'b'},
        'rwh_primes2': {'color': 'b'},
    }


    # Get n in command line
    if len(sys.argv)>1:
        n = int(sys.argv[1])

    step = int(math.ceil(n / float(nbcol)))
    nbs = np.array([i * step for i in range(1, int(nbcol) + 1)])
    set2 = brewer2mpl.get_map('Paired', 'qualitative', 12).mpl_colors

    print datetime.datetime.now().strftime(datetimeformat)
    print("Compute prime number to %(n)s" % locals())
    print("")

    results = dict()
    for pgen in primenumbers_gen:
        results[pgen] = dict()
        benchtimes = list()
        for n in nbs:
            t = timeit.Timer("lib.%(pgen)s(n)" % locals(), setup=config)
            execute_times = t.repeat(repeat=nb_benchloop,number=1)
            benchtime = np.mean(execute_times)
            benchtimes.append(benchtime)
        results[pgen] = {'benchtimes':np.array(benchtimes)}

fig, ax = plt.subplots(1)
plt.ylabel('Computation time (in second)')
plt.xlabel('Numbers computed')
i = 0
for pgen in primenumbers_gen:

    bench = results[pgen]['benchtimes']
    avgs = np.divide(bench,nbs)
    avg = np.average(bench, weights=nbs)

    # Compute linear regression
    A = np.vstack([nbs, np.ones(len(nbs))]).T
    a, b = np.linalg.lstsq(A, nbs*avgs)[0]

    # Plot
    i += 1
    #label="%(pgen)s" % locals()
    #ppl.plot(nbs, nbs*avgs, label=label, lw=1, linestyle='--', color=set2[i % 12])
    label="%(pgen)s avg" % locals()
    ppl.plot(nbs, a * nbs + b, label=label, lw=2, color=set2[i % 12])
print datetime.datetime.now().strftime(datetimeformat)

ppl.legend(ax, loc='upper left', ncol=4)

# Change x axis label
ax.get_xaxis().get_major_formatter().set_scientific(False)
fig.canvas.draw()
labels = [human_format(int(item.get_text())) for item in ax.get_xticklabels()]

ax.set_xticklabels(labels)
ax = plt.gca()

plt.show()

我可能迟到了，但必须为此添加自己的代码。它使用大约n/2的空间，因为我们不需要存储偶数，我还使用bitarray python模块，进一步大幅减少内存消耗，并允许计算所有高达1,000,000,000的质数

from bitarray import bitarray
def primes_to(n):
    size = n//2
    sieve = bitarray(size)
    sieve.setall(1)
    limit = int(n**0.5)
    for i in range(1,limit):
        if sieve[i]:
            val = 2*i+1
            sieve[(i+i*val)::val] = 0
    return [2] + [2*i+1 for i, v in enumerate(sieve) if v and i > 0]

python -m timeit -n10 -s "import euler" "euler.primes_to(1000000000)"
10 loops, best of 3: 46.5 sec per loop

这是在64bit 2.4GHZ MAC OSX 10.8.3上运行的