摘要：问题描述：这是我能想到的最好的算法。def get_primes(n): numbers = set(range(n, 1, -1)) primes = [] while numbers: p = numbers.pop() primes.append(...

问题描述：

这是我能想到的最好的算法。

def get_primes(n):
    numbers = set(range(n, 1, -1))
    primes = []
    while numbers:
        p = numbers.pop()
        primes.append(p)
        numbers.difference_update(set(range(p*2, n+1, p)))
    return primes

>>> timeit.Timer(stmt='get_primes.get_primes(1000000)', setup='import   get_primes').timeit(1)
1.1499958793645562

还能变得更快吗？

这段代码有一个缺陷：由于numbers是无序集合，因此无法保证会numbers.pop()从集合中删除最小数字。不过，对于某些输入数字，它还是有效的（至少对我来说是这样）：

>>> sum(get_primes(2000000))
142913828922L
#That's the correct sum of all numbers below 2 million
>>> 529 in get_primes(1000)
False
>>> 529 in get_primes(530)
True

解决方案 1：

警告： timeit由于硬件或 Python 版本的差异，结果可能会有所不同。

下面是比较多种实现的脚本：

ambi_sieve_plain，
rwh_primes，
rwh_primes1，
rwh_primes2，
阿特金筛法
埃拉托斯特尼
筛法
sundaram3，
筛子轮30，
ambi_sieve（需要 numpy）
primesfrom3to（需要 numpy）
primesfrom2to（需要 numpy）

非常感谢stephan让我注意到 sieve_wheel_30。感谢Robert William Hanks提供的 primesfrom2to、primesfrom3to、rwh_primes、rwh_primes1 和 rwh_primes2。

在使用 psyco测试的普通 Python 方法中，对于 n=1000000，
rwh_primes1是测试速度最快的。

+---------------------+-------+
| Method              | ms    |
+---------------------+-------+
| rwh_primes1         | 43.0  |
| sieveOfAtkin        | 46.4  |
| rwh_primes          | 57.4  |
| sieve_wheel_30      | 63.0  |
| rwh_primes2         | 67.8  |    
| sieveOfEratosthenes | 147.0 |
| ambi_sieve_plain    | 152.0 |
| sundaram3           | 194.0 |
+---------------------+-------+

在所测试的没有 psyco 的普通 Python 方法中，对于 n=1000000，
rwh_primes2是最快的。

+---------------------+-------+
| Method              | ms    |
+---------------------+-------+
| rwh_primes2         | 68.1  |
| rwh_primes1         | 93.7  |
| rwh_primes          | 94.6  |
| sieve_wheel_30      | 97.4  |
| sieveOfEratosthenes | 178.0 |
| ambi_sieve_plain    | 286.0 |
| sieveOfAtkin        | 314.0 |
| sundaram3           | 416.0 |
+---------------------+-------+

在所有测试方法中，允许 numpy，对于 n=1000000，
primesfrom2to是测试速度最快的。

+---------------------+-------+
| Method              | ms    |
+---------------------+-------+
| primesfrom2to       | 15.9  |
| primesfrom3to       | 18.4  |
| ambi_sieve          | 29.3  |
+---------------------+-------+

使用以下命令测量时间：

python -mtimeit -s"import primes" "primes.{method}(1000000)"

用{method}每个方法名称替换。

primes.py：

#!/usr/bin/env python
import psyco; psyco.full()
from math import sqrt, ceil
import numpy as np

def rwh_primes(n):
    # https://stackoverflow.com/questions/2068372/fastest-way-to-list-all-primes-below-n-in-python/3035188#3035188
    """ Returns  a list of primes < n """
    sieve = [True] * n
    for i in xrange(3,int(n**0.5)+1,2):
        if sieve[i]:
            sieve[i*i::2*i]=[False]*((n-i*i-1)/(2*i)+1)
    return [2] + [i for i in xrange(3,n,2) if sieve[i]]

def rwh_primes1(n):
    # https://stackoverflow.com/questions/2068372/fastest-way-to-list-all-primes-below-n-in-python/3035188#3035188
    """ Returns  a list of primes < n """
    sieve = [True] * (n/2)
    for i in xrange(3,int(n**0.5)+1,2):
        if sieve[i/2]:
            sieve[i*i/2::i] = [False] * ((n-i*i-1)/(2*i)+1)
    return [2] + [2*i+1 for i in xrange(1,n/2) if sieve[i]]

def rwh_primes2(n):
    # https://stackoverflow.com/questions/2068372/fastest-way-to-list-all-primes-below-n-in-python/3035188#3035188
    """ Input n>=6, Returns a list of primes, 2 <= p < n """
    correction = (n%6>1)
    n = {0:n,1:n-1,2:n+4,3:n+3,4:n+2,5:n+1}[n%6]
    sieve = [True] * (n/3)
    sieve[0] = False
    for i in xrange(int(n**0.5)/3+1):
      if sieve[i]:
        k=3*i+1|1
        sieve[      ((k*k)/3)      ::2*k]=[False]*((n/6-(k*k)/6-1)/k+1)
        sieve[(k*k+4*k-2*k*(i&1))/3::2*k]=[False]*((n/6-(k*k+4*k-2*k*(i&1))/6-1)/k+1)
    return [2,3] + [3*i+1|1 for i in xrange(1,n/3-correction) if sieve[i]]

def sieve_wheel_30(N):
    # http://zerovolt.com/?p=88
    ''' Returns a list of primes <= N using wheel criterion 2*3*5 = 30

Copyright 2009 by zerovolt.com
This code is free for non-commercial purposes, in which case you can just leave this comment as a credit for my work.
If you need this code for commercial purposes, please contact me by sending an email to: info [at] zerovolt [dot] com.'''
    __smallp = ( 2, 3, 5, 7, 11, 13, 17, 19, 23, 29, 31, 37, 41, 43, 47, 53, 59,
    61, 67, 71, 73, 79, 83, 89, 97, 101, 103, 107, 109, 113, 127, 131, 137, 139,
    149, 151, 157, 163, 167, 173, 179, 181, 191, 193, 197, 199, 211, 223, 227,
    229, 233, 239, 241, 251, 257, 263, 269, 271, 277, 281, 283, 293, 307, 311,
    313, 317, 331, 337, 347, 349, 353, 359, 367, 373, 379, 383, 389, 397, 401,
    409, 419, 421, 431, 433, 439, 443, 449, 457, 461, 463, 467, 479, 487, 491,
    499, 503, 509, 521, 523, 541, 547, 557, 563, 569, 571, 577, 587, 593, 599,
    601, 607, 613, 617, 619, 631, 641, 643, 647, 653, 659, 661, 673, 677, 683,
    691, 701, 709, 719, 727, 733, 739, 743, 751, 757, 761, 769, 773, 787, 797,
    809, 811, 821, 823, 827, 829, 839, 853, 857, 859, 863, 877, 881, 883, 887,
    907, 911, 919, 929, 937, 941, 947, 953, 967, 971, 977, 983, 991, 997)

    wheel = (2, 3, 5)
    const = 30
    if N < 2:
        return []
    if N <= const:
        pos = 0
        while __smallp[pos] <= N:
            pos += 1
        return list(__smallp[:pos])
    # make the offsets list
    offsets = (7, 11, 13, 17, 19, 23, 29, 1)
    # prepare the list
    p = [2, 3, 5]
    dim = 2 + N // const
    tk1  = [True] * dim
    tk7  = [True] * dim
    tk11 = [True] * dim
    tk13 = [True] * dim
    tk17 = [True] * dim
    tk19 = [True] * dim
    tk23 = [True] * dim
    tk29 = [True] * dim
    tk1[0] = False
    # help dictionary d
    # d[a , b] = c  ==> if I want to find the smallest useful multiple of (30*pos)+a
    # on tkc, then I need the index given by the product of [(30*pos)+a][(30*pos)+b]
    # in general. If b < a, I need [(30*pos)+a][(30*(pos+1))+b]
    d = {}
    for x in offsets:
        for y in offsets:
            res = (x*y) % const
            if res in offsets:
                d[(x, res)] = y
    # another help dictionary: gives tkx calling tmptk[x]
    tmptk = {1:tk1, 7:tk7, 11:tk11, 13:tk13, 17:tk17, 19:tk19, 23:tk23, 29:tk29}
    pos, prime, lastadded, stop = 0, 0, 0, int(ceil(sqrt(N)))
    # inner functions definition
    def del_mult(tk, start, step):
        for k in xrange(start, len(tk), step):
            tk[k] = False
    # end of inner functions definition
    cpos = const * pos
    while prime < stop:
        # 30k + 7
        if tk7[pos]:
            prime = cpos + 7
            p.append(prime)
            lastadded = 7
            for off in offsets:
                tmp = d[(7, off)]
                start = (pos + prime) if off == 7 else (prime * (const * (pos + 1 if tmp < 7 else 0) + tmp) )//const
                del_mult(tmptk[off], start, prime)
        # 30k + 11
        if tk11[pos]:
            prime = cpos + 11
            p.append(prime)
            lastadded = 11
            for off in offsets:
                tmp = d[(11, off)]
                start = (pos + prime) if off == 11 else (prime * (const * (pos + 1 if tmp < 11 else 0) + tmp) )//const
                del_mult(tmptk[off], start, prime)
        # 30k + 13
        if tk13[pos]:
            prime = cpos + 13
            p.append(prime)
            lastadded = 13
            for off in offsets:
                tmp = d[(13, off)]
                start = (pos + prime) if off == 13 else (prime * (const * (pos + 1 if tmp < 13 else 0) + tmp) )//const
                del_mult(tmptk[off], start, prime)
        # 30k + 17
        if tk17[pos]:
            prime = cpos + 17
            p.append(prime)
            lastadded = 17
            for off in offsets:
                tmp = d[(17, off)]
                start = (pos + prime) if off == 17 else (prime * (const * (pos + 1 if tmp < 17 else 0) + tmp) )//const
                del_mult(tmptk[off], start, prime)
        # 30k + 19
        if tk19[pos]:
            prime = cpos + 19
            p.append(prime)
            lastadded = 19
            for off in offsets:
                tmp = d[(19, off)]
                start = (pos + prime) if off == 19 else (prime * (const * (pos + 1 if tmp < 19 else 0) + tmp) )//const
                del_mult(tmptk[off], start, prime)
        # 30k + 23
        if tk23[pos]:
            prime = cpos + 23
            p.append(prime)
            lastadded = 23
            for off in offsets:
                tmp = d[(23, off)]
                start = (pos + prime) if off == 23 else (prime * (const * (pos + 1 if tmp < 23 else 0) + tmp) )//const
                del_mult(tmptk[off], start, prime)
        # 30k + 29
        if tk29[pos]:
            prime = cpos + 29
            p.append(prime)
            lastadded = 29
            for off in offsets:
                tmp = d[(29, off)]
                start = (pos + prime) if off == 29 else (prime * (const * (pos + 1 if tmp < 29 else 0) + tmp) )//const
                del_mult(tmptk[off], start, prime)
        # now we go back to top tk1, so we need to increase pos by 1
        pos += 1
        cpos = const * pos
        # 30k + 1
        if tk1[pos]:
            prime = cpos + 1
            p.append(prime)
            lastadded = 1
            for off in offsets:
                tmp = d[(1, off)]
                start = (pos + prime) if off == 1 else (prime * (const * pos + tmp) )//const
                del_mult(tmptk[off], start, prime)
    # time to add remaining primes
    # if lastadded == 1, remove last element and start adding them from tk1
    # this way we don't need an "if" within the last while
    if lastadded == 1:
        p.pop()
    # now complete for every other possible prime
    while pos < len(tk1):
        cpos = const * pos
        if tk1[pos]: p.append(cpos + 1)
        if tk7[pos]: p.append(cpos + 7)
        if tk11[pos]: p.append(cpos + 11)
        if tk13[pos]: p.append(cpos + 13)
        if tk17[pos]: p.append(cpos + 17)
        if tk19[pos]: p.append(cpos + 19)
        if tk23[pos]: p.append(cpos + 23)
        if tk29[pos]: p.append(cpos + 29)
        pos += 1
    # remove exceeding if present
    pos = len(p) - 1
    while p[pos] > N:
        pos -= 1
    if pos < len(p) - 1:
        del p[pos+1:]
    # return p list
    return p

def sieveOfEratosthenes(n):
    """sieveOfEratosthenes(n): return the list of the primes < n."""
    # Code from: <dickinsm@gmail.com>, Nov 30 2006
    # http://groups.google.com/group/comp.lang.python/msg/f1f10ced88c68c2d
    if n <= 2:
        return []
    sieve = range(3, n, 2)
    top = len(sieve)
    for si in sieve:
        if si:
            bottom = (si*si - 3) // 2
            if bottom >= top:
                break
            sieve[bottom::si] = [0] * -((bottom - top) // si)
    return [2] + [el for el in sieve if el]

def sieveOfAtkin(end):
    """sieveOfAtkin(end): return a list of all the prime numbers <end
    using the Sieve of Atkin."""
    # Code by Steve Krenzel, <Sgk284@gmail.com>, improved
    # Code: https://web.archive.org/web/20080324064651/http://krenzel.info/?p=83
    # Info: http://en.wikipedia.org/wiki/Sieve_of_Atkin
    assert end > 0
    lng = ((end-1) // 2)
    sieve = [False] * (lng + 1)

    x_max, x2, xd = int(sqrt((end-1)/4.0)), 0, 4
    for xd in xrange(4, 8*x_max + 2, 8):
        x2 += xd
        y_max = int(sqrt(end-x2))
        n, n_diff = x2 + y_max*y_max, (y_max << 1) - 1
        if not (n & 1):
            n -= n_diff
            n_diff -= 2
        for d in xrange((n_diff - 1) << 1, -1, -8):
            m = n % 12
            if m == 1 or m == 5:
                m = n >> 1
                sieve[m] = not sieve[m]
            n -= d

    x_max, x2, xd = int(sqrt((end-1) / 3.0)), 0, 3
    for xd in xrange(3, 6 * x_max + 2, 6):
        x2 += xd
        y_max = int(sqrt(end-x2))
        n, n_diff = x2 + y_max*y_max, (y_max << 1) - 1
        if not(n & 1):
            n -= n_diff
            n_diff -= 2
        for d in xrange((n_diff - 1) << 1, -1, -8):
            if n % 12 == 7:
                m = n >> 1
                sieve[m] = not sieve[m]
            n -= d

    x_max, y_min, x2, xd = int((2 + sqrt(4-8*(1-end)))/4), -1, 0, 3
    for x in xrange(1, x_max + 1):
        x2 += xd
        xd += 6
        if x2 >= end: y_min = (((int(ceil(sqrt(x2 - end))) - 1) << 1) - 2) << 1
        n, n_diff = ((x*x + x) << 1) - 1, (((x-1) << 1) - 2) << 1
        for d in xrange(n_diff, y_min, -8):
            if n % 12 == 11:
                m = n >> 1
                sieve[m] = not sieve[m]
            n += d

    primes = [2, 3]
    if end <= 3:
        return primes[:max(0,end-2)]

    for n in xrange(5 >> 1, (int(sqrt(end))+1) >> 1):
        if sieve[n]:
            primes.append((n << 1) + 1)
            aux = (n << 1) + 1
            aux *= aux
            for k in xrange(aux, end, 2 * aux):
                sieve[k >> 1] = False

    s  = int(sqrt(end)) + 1
    if s  % 2 == 0:
        s += 1
    primes.extend([i for i in xrange(s, end, 2) if sieve[i >> 1]])

    return primes

def ambi_sieve_plain(n):
    s = range(3, n, 2)
    for m in xrange(3, int(n**0.5)+1, 2): 
        if s[(m-3)/2]: 
            for t in xrange((m*m-3)/2,(n>>1)-1,m):
                s[t]=0
    return [2]+[t for t in s if t>0]

def sundaram3(max_n):
    # https://stackoverflow.com/questions/2068372/fastest-way-to-list-all-primes-below-n-in-python/2073279#2073279
    numbers = range(3, max_n+1, 2)
    half = (max_n)//2
    initial = 4

    for step in xrange(3, max_n+1, 2):
        for i in xrange(initial, half, step):
            numbers[i-1] = 0
        initial += 2*(step+1)

        if initial > half:
            return [2] + filter(None, numbers)

################################################################################
# Using Numpy:
def ambi_sieve(n):
    # http://tommih.blogspot.com/2009/04/fast-prime-number-generator.html
    s = np.arange(3, n, 2)
    for m in xrange(3, int(n ** 0.5)+1, 2): 
        if s[(m-3)/2]: 
            s[(m*m-3)/2::m]=0
    return np.r_[2, s[s>0]]

def primesfrom3to(n):
    # https://stackoverflow.com/questions/2068372/fastest-way-to-list-all-primes-below-n-in-python/3035188#3035188
    """ Returns a array of primes, p < n """
    assert n>=2
    sieve = np.ones(n/2, dtype=np.bool)
    for i in xrange(3,int(n**0.5)+1,2):
        if sieve[i/2]:
            sieve[i*i/2::i] = False
    return np.r_[2, 2*np.nonzero(sieve)[0][1::]+1]    

def primesfrom2to(n):
    # https://stackoverflow.com/questions/2068372/fastest-way-to-list-all-primes-below-n-in-python/3035188#3035188
    """ Input n>=6, Returns a array of primes, 2 <= p < n """
    sieve = np.ones(n/3 + (n%6==2), dtype=np.bool)
    sieve[0] = False
    for i in xrange(int(n**0.5)/3+1):
        if sieve[i]:
            k=3*i+1|1
            sieve[      ((k*k)/3)      ::2*k] = False
            sieve[(k*k+4*k-2*k*(i&1))/3::2*k] = False
    return np.r_[2,3,((3*np.nonzero(sieve)[0]+1)|1)]

if __name__=='__main__':
    import itertools
    import sys

    def test(f1,f2,num):
        print('Testing {f1} and {f2} return same results'.format(
            f1=f1.func_name,
            f2=f2.func_name))
        if not all([a==b for a,b in itertools.izip_longest(f1(num),f2(num))]):
            sys.exit("Error: %s(%s) != %s(%s)"%(f1.func_name,num,f2.func_name,num))

    n=1000000
    test(sieveOfAtkin,sieveOfEratosthenes,n)
    test(sieveOfAtkin,ambi_sieve,n)
    test(sieveOfAtkin,ambi_sieve_plain,n) 
    test(sieveOfAtkin,sundaram3,n)
    test(sieveOfAtkin,sieve_wheel_30,n)
    test(sieveOfAtkin,primesfrom3to,n)
    test(sieveOfAtkin,primesfrom2to,n)
    test(sieveOfAtkin,rwh_primes,n)
    test(sieveOfAtkin,rwh_primes1,n)         
    test(sieveOfAtkin,rwh_primes2,n)

运行脚本测试所有实现是否都产生相同的结果。

解决方案 2：

速度更快、内存占用更少的纯 Python 代码：

def primes(n):
    """ Returns  a list of primes < n """
    sieve = [True] * n
    for i in range(3,int(n**0.5)+1,2):
        if sieve[i]:
            sieve[i*i::2*i]=[False]*((n-i*i-1)//(2*i)+1)
    return [2] + [i for i in range(3,n,2) if sieve[i]]

或从半筛开始

def primes1(n):
    """ Returns  a list of primes < n """
    sieve = [True] * (n//2)
    for i in range(3,int(n**0.5)+1,2):
        if sieve[i//2]:
            sieve[i*i//2::i] = [False] * ((n-i*i-1)//(2*i)+1)
    return [2] + [2*i+1 for i in range(1,n//2) if sieve[i]]

更快、更节省内存的 numpy 代码：

import numpy
def primesfrom3to(n):
    """ Returns a array of primes, 3 <= p < n """
    sieve = numpy.ones(n//2, dtype=bool)
    for i in range(3,int(n**0.5)+1,2):
        if sieve[i//2]:
            sieve[i*i//2::i] = False
    return 2*numpy.nonzero(sieve)[0][1::]+1

一种更快的变化，从三分之一筛子开始：

import numpy
def primesfrom2to(n):
    """ Input n>=6, Returns a array of primes, 2 <= p < n """
    sieve = numpy.ones(n//3 + (n%6==2), dtype=bool)
    for i in range(1,int(n**0.5)//3+1):
        if sieve[i]:
            k=3*i+1|1
            sieve[       k*k//3     ::2*k] = False
            sieve[k*(k-2*(i&1)+4)//3::2*k] = False
    return numpy.r_[2,3,((3*numpy.nonzero(sieve)[0][1:]+1)|1)]

上述代码的纯 Python 版本（难以编码）如下：

def primes2(n):
    """ Input n>=6, Returns a list of primes, 2 <= p < n """
    n, correction = n-n%6+6, 2-(n%6>1)
    sieve = [True] * (n//3)
    for i in range(1,int(n**0.5)//3+1):
      if sieve[i]:
        k=3*i+1|1
        sieve[      k*k//3      ::2*k] = [False] * ((n//6-k*k//6-1)//k+1)
        sieve[k*(k-2*(i&1)+4)//3::2*k] = [False] * ((n//6-k*(k-2*(i&1)+4)//6-1)//k+1)
    return [2,3] + [3*i+1|1 for i in range(1,n//3-correction) if sieve[i]]

不幸的是，纯 Python 不采用更简单、更快速的 numpy 赋值方式，而且len()在循环内调用[False]*len(sieve[((k*k)//3)::2*k])太慢了。所以我不得不即兴发挥以纠正输入（并避免更多数学运算）并做一些极端（且痛苦）的数学魔法。

我个人认为，numpy（被广泛使用）不是 Python 标准库的一部分，这很遗憾，而且语法和速度的改进似乎被 Python 开发人员完全忽视了。

解决方案 3：

这里有一个来自 Python Cookbook 的非常简洁的示例—— 该 URL 上提出的最快版本是：

import itertools
def erat2( ):
    D = {  }
    yield 2
    for q in itertools.islice(itertools.count(3), 0, None, 2):
        p = D.pop(q, None)
        if p is None:
            D[q*q] = q
            yield q
        else:
            x = p + q
            while x in D or not (x&1):
                x += p
            D[x] = p

这样就可以

def get_primes_erat(n):
  return list(itertools.takewhile(lambda p: p<n, erat2()))

使用 pri.py 中的这段代码在 shell 提示符下进行测量（我更喜欢这样做），我观察到：

$ python2.5 -mtimeit -s'import pri' 'pri.get_primes(1000000)'
10 loops, best of 3: 1.69 sec per loop
$ python2.5 -mtimeit -s'import pri' 'pri.get_primes_erat(1000000)'
10 loops, best of 3: 673 msec per loop

因此看起来 Cookbook 解决方案的速度快两倍多。

解决方案 4：

使用Sundaram 的 Sieve，我认为我打破了纯 Python 的记录：

def sundaram3(max_n):
    numbers = range(3, max_n+1, 2)
    half = (max_n)//2
    initial = 4

    for step in xrange(3, max_n+1, 2):
        for i in xrange(initial, half, step):
            numbers[i-1] = 0
        initial += 2*(step+1)

        if initial > half:
            return [2] + filter(None, numbers)

比较：

C:USERS>python -m timeit -n10 -s "import get_primes" "get_primes.get_primes_erat(1000000)"
10 loops, best of 3: 710 msec per loop

C:USERS>python -m timeit -n10 -s "import get_primes" "get_primes.daniel_sieve_2(1000000)"
10 loops, best of 3: 435 msec per loop

C:USERS>python -m timeit -n10 -s "import get_primes" "get_primes.sundaram3(1000000)"
10 loops, best of 3: 327 msec per loop

解决方案 5：

该算法速度很快，但是有一个严重的缺陷：

>>> sorted(get_primes(530))
[2, 3, 5, 7, 11, 13, 17, 19, 23, 29, 31, 37, 41, 43, 47, 53, 59, 61, 67, 71, 73,
79, 83, 89, 97, 101, 103, 107, 109, 113, 127, 131, 137, 139, 149, 151, 157, 163,
167, 173, 179, 181, 191, 193, 197, 199, 211, 223, 227, 229, 233, 239, 241, 251,
257, 263, 269, 271, 277, 281, 283, 293, 307, 311, 313, 317, 331, 337, 347, 349,
353, 359, 367, 373, 379, 383, 389, 397, 401, 409, 419, 421, 431, 433, 439, 443,
449, 457, 461, 463, 467, 479, 487, 491, 499, 503, 509, 521, 523, 527, 529]
>>> 17*31
527
>>> 23*23
529

您假设这numbers.pop()将返回集合中的最小数字，但这根本无法保证。集合是无序的，pop()会删除并返回任意元素，因此不能使用它从剩余数字中选择下一个素数。

解决方案 6：

对于具有足够大 N 的真正最快的解决方案，将下载预先计算的素数列表，将其存储为元组并执行以下操作：

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

如果N > primes[-1] 只计算更多的素数并在代码中保存新列表，那么下一次就会同样快。

总是跳出固有的思维框架。

解决方案 7：

如果你不想重新发明轮子，你可以安装符号数学库sympy（是的，它与 Python 3 兼容）

pip install sympy

并使用primerange函数

from sympy import sieve
primes = list(sieve.primerange(1, 10**6))

解决方案 8：

如果您接受 itertools 但不接受 numpy，这里有一个适用于 Python 3 的 rwh_primes2 改编版，在我的计算机上运行速度大约是它的两倍。唯一实质性的变化是使用字节数组而不是列表来表示布尔值，并使用压缩而不是列表推导来构建最终列表。（如果可以的话，我会像 moarningsun 一样将其添加为评论。）

import itertools
izip = itertools.zip_longest
chain = itertools.chain.from_iterable
compress = itertools.compress
def rwh_primes2_python3(n):
    """ Input n>=6, Returns a list of primes, 2 <= p < n """
    zero = bytearray([False])
    size = n//3 + (n % 6 == 2)
    sieve = bytearray([True]) * size
    sieve[0] = False
    for i in range(int(n**0.5)//3+1):
      if sieve[i]:
        k=3*i+1|1
        start = (k*k+4*k-2*k*(i&1))//3
        sieve[(k*k)//3::2*k]=zero*((size - (k*k)//3 - 1) // (2 * k) + 1)
        sieve[  start ::2*k]=zero*((size -   start  - 1) // (2 * k) + 1)
    ans = [2,3]
    poss = chain(izip(*[range(i, n, 6) for i in (1,5)]))
    ans.extend(compress(poss, sieve))
    return ans

比较：

>>> timeit.timeit('primes.rwh_primes2(10**6)', setup='import primes', number=1)
0.0652179726976101
>>> timeit.timeit('primes.rwh_primes2_python3(10**6)', setup='import primes', number=1)
0.03267321276325674

和

>>> timeit.timeit('primes.rwh_primes2(10**8)', setup='import primes', number=1)
6.394284538007014
>>> timeit.timeit('primes.rwh_primes2_python3(10**8)', setup='import primes', number=1)
3.833829450302801

解决方案 9：

以下是速度最快的函数之一的两个更新版本（纯 Python 3.6），

from itertools import compress

def rwh_primes1v1(n):
    """ Returns  a list of primes < n for n > 2 """
    sieve = bytearray([True]) * (n//2)
    for i in range(3,int(n**0.5)+1,2):
        if sieve[i//2]:
            sieve[i*i//2::i] = bytearray((n-i*i-1)//(2*i)+1)
    return [2,*compress(range(3,n,2), sieve[1:])]

def rwh_primes1v2(n):
    """ Returns a list of primes < n for n > 2 """
    sieve = bytearray([True]) * (n//2+1)
    for i in range(1,int(n**0.5)//2+1):
        if sieve[i]:
            sieve[2*i*(i+1)::2*i+1] = bytearray((n//2-2*i*(i+1))//(2*i+1)+1)
    return [2,*compress(range(3,n,2), sieve[1:])]

解决方案 10：

我更新了 Python 3 的大部分代码，并将其放在perfplot（我的一个项目）上，看看哪个实际上是最快的。结果发现，对于较大的n，primesfrom{2,3}to它是最好的：

在此处输入图片描述

重现情节的代码：

import perfplot
from math import sqrt, ceil
import numpy as np
import sympy


def rwh_primes(n):
    # https://stackoverflow.com/questions/2068372/fastest-way-to-list-all-primes-below-n-in-python/3035188#3035188
    """ Returns  a list of primes < n """
    sieve = [True] * n
    for i in range(3, int(n ** 0.5) + 1, 2):
        if sieve[i]:
            sieve[i * i::2 * i] = [False] * ((n - i * i - 1) // (2 * i) + 1)
    return [2] + [i for i in range(3, n, 2) if sieve[i]]


def rwh_primes1(n):
    # https://stackoverflow.com/questions/2068372/fastest-way-to-list-all-primes-below-n-in-python/3035188#3035188
    """ Returns  a list of primes < n """
    sieve = [True] * (n // 2)
    for i in range(3, int(n ** 0.5) + 1, 2):
        if sieve[i // 2]:
            sieve[i * i // 2::i] = [False] * ((n - i * i - 1) // (2 * i) + 1)
    return [2] + [2 * i + 1 for i in range(1, n // 2) if sieve[i]]


def rwh_primes2(n):
    # https://stackoverflow.com/questions/2068372/fastest-way-to-list-all-primes-below-n-in-python/3035188#3035188
    """Input n>=6, Returns a list of primes, 2 <= p < n"""
    assert n >= 6
    correction = n % 6 > 1
    n = {0: n, 1: n - 1, 2: n + 4, 3: n + 3, 4: n + 2, 5: n + 1}[n % 6]
    sieve = [True] * (n // 3)
    sieve[0] = False
    for i in range(int(n ** 0.5) // 3 + 1):
        if sieve[i]:
            k = 3 * i + 1 | 1
            sieve[((k * k) // 3)::2 * k] = [False] * (
                (n // 6 - (k * k) // 6 - 1) // k + 1
            )
            sieve[(k * k + 4 * k - 2 * k * (i & 1)) // 3::2 * k] = [False] * (
                (n // 6 - (k * k + 4 * k - 2 * k * (i & 1)) // 6 - 1) // k + 1
            )
    return [2, 3] + [3 * i + 1 | 1 for i in range(1, n // 3 - correction) if sieve[i]]


def sieve_wheel_30(N):
    # http://zerovolt.com/?p=88
    """ Returns a list of primes <= N using wheel criterion 2*3*5 = 30

Copyright 2009 by zerovolt.com
This code is free for non-commercial purposes, in which case you can just leave this comment as a credit for my work.
If you need this code for commercial purposes, please contact me by sending an email to: info [at] zerovolt [dot] com."""
    __smallp = (
        2,
        3,
        5,
        7,
        11,
        13,
        17,
        19,
        23,
        29,
        31,
        37,
        41,
        43,
        47,
        53,
        59,
        61,
        67,
        71,
        73,
        79,
        83,
        89,
        97,
        101,
        103,
        107,
        109,
        113,
        127,
        131,
        137,
        139,
        149,
        151,
        157,
        163,
        167,
        173,
        179,
        181,
        191,
        193,
        197,
        199,
        211,
        223,
        227,
        229,
        233,
        239,
        241,
        251,
        257,
        263,
        269,
        271,
        277,
        281,
        283,
        293,
        307,
        311,
        313,
        317,
        331,
        337,
        347,
        349,
        353,
        359,
        367,
        373,
        379,
        383,
        389,
        397,
        401,
        409,
        419,
        421,
        431,
        433,
        439,
        443,
        449,
        457,
        461,
        463,
        467,
        479,
        487,
        491,
        499,
        503,
        509,
        521,
        523,
        541,
        547,
        557,
        563,
        569,
        571,
        577,
        587,
        593,
        599,
        601,
        607,
        613,
        617,
        619,
        631,
        641,
        643,
        647,
        653,
        659,
        661,
        673,
        677,
        683,
        691,
        701,
        709,
        719,
        727,
        733,
        739,
        743,
        751,
        757,
        761,
        769,
        773,
        787,
        797,
        809,
        811,
        821,
        823,
        827,
        829,
        839,
        853,
        857,
        859,
        863,
        877,
        881,
        883,
        887,
        907,
        911,
        919,
        929,
        937,
        941,
        947,
        953,
        967,
        971,
        977,
        983,
        991,
        997,
    )
    # wheel = (2, 3, 5)
    const = 30
    if N < 2:
        return []
    if N <= const:
        pos = 0
        while __smallp[pos] <= N:
            pos += 1
        return list(__smallp[:pos])
    # make the offsets list
    offsets = (7, 11, 13, 17, 19, 23, 29, 1)
    # prepare the list
    p = [2, 3, 5]
    dim = 2 + N // const
    tk1 = [True] * dim
    tk7 = [True] * dim
    tk11 = [True] * dim
    tk13 = [True] * dim
    tk17 = [True] * dim
    tk19 = [True] * dim
    tk23 = [True] * dim
    tk29 = [True] * dim
    tk1[0] = False
    # help dictionary d
    # d[a , b] = c  ==> if I want to find the smallest useful multiple of (30*pos)+a
    # on tkc, then I need the index given by the product of [(30*pos)+a][(30*pos)+b]
    # in general. If b < a, I need [(30*pos)+a][(30*(pos+1))+b]
    d = {}
    for x in offsets:
        for y in offsets:
            res = (x * y) % const
            if res in offsets:
                d[(x, res)] = y
    # another help dictionary: gives tkx calling tmptk[x]
    tmptk = {1: tk1, 7: tk7, 11: tk11, 13: tk13, 17: tk17, 19: tk19, 23: tk23, 29: tk29}
    pos, prime, lastadded, stop = 0, 0, 0, int(ceil(sqrt(N)))

    # inner functions definition
    def del_mult(tk, start, step):
        for k in range(start, len(tk), step):
            tk[k] = False

    # end of inner functions definition
    cpos = const * pos
    while prime < stop:
        # 30k + 7
        if tk7[pos]:
            prime = cpos + 7
            p.append(prime)
            lastadded = 7
            for off in offsets:
                tmp = d[(7, off)]
                start = (
                    (pos + prime)
                    if off == 7
                    else (prime * (const * (pos + 1 if tmp < 7 else 0) + tmp)) // const
                )
                del_mult(tmptk[off], start, prime)
        # 30k + 11
        if tk11[pos]:
            prime = cpos + 11
            p.append(prime)
            lastadded = 11
            for off in offsets:
                tmp = d[(11, off)]
                start = (
                    (pos + prime)
                    if off == 11
                    else (prime * (const * (pos + 1 if tmp < 11 else 0) + tmp)) // const
                )
                del_mult(tmptk[off], start, prime)
        # 30k + 13
        if tk13[pos]:
            prime = cpos + 13
            p.append(prime)
            lastadded = 13
            for off in offsets:
                tmp = d[(13, off)]
                start = (
                    (pos + prime)
                    if off == 13
                    else (prime * (const * (pos + 1 if tmp < 13 else 0) + tmp)) // const
                )
                del_mult(tmptk[off], start, prime)
        # 30k + 17
        if tk17[pos]:
            prime = cpos + 17
            p.append(prime)
            lastadded = 17
            for off in offsets:
                tmp = d[(17, off)]
                start = (
                    (pos + prime)
                    if off == 17
                    else (prime * (const * (pos + 1 if tmp < 17 else 0) + tmp)) // const
                )
                del_mult(tmptk[off], start, prime)
        # 30k + 19
        if tk19[pos]:
            prime = cpos + 19
            p.append(prime)
            lastadded = 19
            for off in offsets:
                tmp = d[(19, off)]
                start = (
                    (pos + prime)
                    if off == 19
                    else (prime * (const * (pos + 1 if tmp < 19 else 0) + tmp)) // const
                )
                del_mult(tmptk[off], start, prime)
        # 30k + 23
        if tk23[pos]:
            prime = cpos + 23
            p.append(prime)
            lastadded = 23
            for off in offsets:
                tmp = d[(23, off)]
                start = (
                    (pos + prime)
                    if off == 23
                    else (prime * (const * (pos + 1 if tmp < 23 else 0) + tmp)) // const
                )
                del_mult(tmptk[off], start, prime)
        # 30k + 29
        if tk29[pos]:
            prime = cpos + 29
            p.append(prime)
            lastadded = 29
            for off in offsets:
                tmp = d[(29, off)]
                start = (
                    (pos + prime)
                    if off == 29
                    else (prime * (const * (pos + 1 if tmp < 29 else 0) + tmp)) // const
                )
                del_mult(tmptk[off], start, prime)
        # now we go back to top tk1, so we need to increase pos by 1
        pos += 1
        cpos = const * pos
        # 30k + 1
        if tk1[pos]:
            prime = cpos + 1
            p.append(prime)
            lastadded = 1
            for off in offsets:
                tmp = d[(1, off)]
                start = (
                    (pos + prime)
                    if off == 1
                    else (prime * (const * pos + tmp)) // const
                )
                del_mult(tmptk[off], start, prime)
    # time to add remaining primes
    # if lastadded == 1, remove last element and start adding them from tk1
    # this way we don't need an "if" within the last while
    if lastadded == 1:
        p.pop()
    # now complete for every other possible prime
    while pos < len(tk1):
        cpos = const * pos
        if tk1[pos]:
            p.append(cpos + 1)
        if tk7[pos]:
            p.append(cpos + 7)
        if tk11[pos]:
            p.append(cpos + 11)
        if tk13[pos]:
            p.append(cpos + 13)
        if tk17[pos]:
            p.append(cpos + 17)
        if tk19[pos]:
            p.append(cpos + 19)
        if tk23[pos]:
            p.append(cpos + 23)
        if tk29[pos]:
            p.append(cpos + 29)
        pos += 1
    # remove exceeding if present
    pos = len(p) - 1
    while p[pos] > N:
        pos -= 1
    if pos < len(p) - 1:
        del p[pos + 1 :]
    # return p list
    return p


def sieve_of_eratosthenes(n):
    """sieveOfEratosthenes(n): return the list of the primes < n."""
    # Code from: <dickinsm@gmail.com>, Nov 30 2006
    # http://groups.google.com/group/comp.lang.python/msg/f1f10ced88c68c2d
    if n <= 2:
        return []
    sieve = list(range(3, n, 2))
    top = len(sieve)
    for si in sieve:
        if si:
            bottom = (si * si - 3) // 2
            if bottom >= top:
                break
            sieve[bottom::si] = [0] * -((bottom - top) // si)
    return [2] + [el for el in sieve if el]


def sieve_of_atkin(end):
    """return a list of all the prime numbers <end using the Sieve of Atkin."""
    # Code by Steve Krenzel, <Sgk284@gmail.com>, improved
    # Code: https://web.archive.org/web/20080324064651/http://krenzel.info/?p=83
    # Info: http://en.wikipedia.org/wiki/Sieve_of_Atkin
    assert end > 0
    lng = (end - 1) // 2
    sieve = [False] * (lng + 1)

    x_max, x2, xd = int(sqrt((end - 1) / 4.0)), 0, 4
    for xd in range(4, 8 * x_max + 2, 8):
        x2 += xd
        y_max = int(sqrt(end - x2))
        n, n_diff = x2 + y_max * y_max, (y_max << 1) - 1
        if not (n & 1):
            n -= n_diff
            n_diff -= 2
        for d in range((n_diff - 1) << 1, -1, -8):
            m = n % 12
            if m == 1 or m == 5:
                m = n >> 1
                sieve[m] = not sieve[m]
            n -= d

    x_max, x2, xd = int(sqrt((end - 1) / 3.0)), 0, 3
    for xd in range(3, 6 * x_max + 2, 6):
        x2 += xd
        y_max = int(sqrt(end - x2))
        n, n_diff = x2 + y_max * y_max, (y_max << 1) - 1
        if not (n & 1):
            n -= n_diff
            n_diff -= 2
        for d in range((n_diff - 1) << 1, -1, -8):
            if n % 12 == 7:
                m = n >> 1
                sieve[m] = not sieve[m]
            n -= d

    x_max, y_min, x2, xd = int((2 + sqrt(4 - 8 * (1 - end))) / 4), -1, 0, 3
    for x in range(1, x_max + 1):
        x2 += xd
        xd += 6
        if x2 >= end:
            y_min = (((int(ceil(sqrt(x2 - end))) - 1) << 1) - 2) << 1
        n, n_diff = ((x * x + x) << 1) - 1, (((x - 1) << 1) - 2) << 1
        for d in range(n_diff, y_min, -8):
            if n % 12 == 11:
                m = n >> 1
                sieve[m] = not sieve[m]
            n += d

    primes = [2, 3]
    if end <= 3:
        return primes[: max(0, end - 2)]

    for n in range(5 >> 1, (int(sqrt(end)) + 1) >> 1):
        if sieve[n]:
            primes.append((n << 1) + 1)
            aux = (n << 1) + 1
            aux *= aux
            for k in range(aux, end, 2 * aux):
                sieve[k >> 1] = False

    s = int(sqrt(end)) + 1
    if s % 2 == 0:
        s += 1
    primes.extend([i for i in range(s, end, 2) if sieve[i >> 1]])

    return primes


def ambi_sieve_plain(n):
    s = list(range(3, n, 2))
    for m in range(3, int(n ** 0.5) + 1, 2):
        if s[(m - 3) // 2]:
            for t in range((m * m - 3) // 2, (n >> 1) - 1, m):
                s[t] = 0
    return [2] + [t for t in s if t > 0]


def sundaram3(max_n):
    # https://stackoverflow.com/questions/2068372/fastest-way-to-list-all-primes-below-n-in-python/2073279#2073279
    numbers = range(3, max_n + 1, 2)
    half = (max_n) // 2
    initial = 4

    for step in range(3, max_n + 1, 2):
        for i in range(initial, half, step):
            numbers[i - 1] = 0
        initial += 2 * (step + 1)

        if initial > half:
            return [2] + filter(None, numbers)


# Using Numpy:
def ambi_sieve(n):
    # http://tommih.blogspot.com/2009/04/fast-prime-number-generator.html
    s = np.arange(3, n, 2)
    for m in range(3, int(n ** 0.5) + 1, 2):
        if s[(m - 3) // 2]:
            s[(m * m - 3) // 2::m] = 0
    return np.r_[2, s[s > 0]]


def primesfrom3to(n):
    # https://stackoverflow.com/questions/2068372/fastest-way-to-list-all-primes-below-n-in-python/3035188#3035188
    """ Returns an array of primes, p < n """
    assert n >= 2
    sieve = np.ones(n // 2, dtype=bool)
    for i in range(3, int(n ** 0.5) + 1, 2):
        if sieve[i // 2]:
            sieve[i * i // 2::i] = False
    return np.r_[2, 2 * np.nonzero(sieve)[0][1::] + 1]


def primesfrom2to(n):
    # https://stackoverflow.com/questions/2068372/fastest-way-to-list-all-primes-below-n-in-python/3035188#3035188
    """ Input n>=6, Returns an array of primes, 2 <= p < n """
    assert n >= 6
    sieve = np.ones(n // 3 + (n % 6 == 2), dtype=bool)
    sieve[0] = False
    for i in range(int(n ** 0.5) // 3 + 1):
        if sieve[i]:
            k = 3 * i + 1 | 1
            sieve[((k * k) // 3)::2 * k] = False
            sieve[(k * k + 4 * k - 2 * k * (i & 1)) // 3::2 * k] = False
    return np.r_[2, 3, ((3 * np.nonzero(sieve)[0] + 1) | 1)]


def sympy_sieve(n):
    return list(sympy.sieve.primerange(1, n))


b = perfplot.bench(
    setup=lambda n: n,
    kernels=[
        rwh_primes,
        rwh_primes1,
        rwh_primes2,
        sieve_wheel_30,
        sieve_of_eratosthenes,
        sieve_of_atkin,
        # ambi_sieve_plain,
        # sundaram3,
        ambi_sieve,
        primesfrom3to,
        primesfrom2to,
        sympy_sieve,
    ],
    n_range=[2 ** k for k in range(3, 25)],
    xlabel="n",
)
b.save("out.png")
b.show()

解决方案 11：

编写自己的素数查找代码很有指导意义，但手头有一个快速可靠的库也很有用。我围绕C++ 库 primesieve编写了一个包装器，将其命名为primesieve-python

尝试一下pip install primesieve

import primesieve
primes = primesieve.generate_primes(10**8)

我很好奇想看看速度上的比较。

解决方案 12：

如果你能控制 N，列出所有素数的最快方法就是预先计算它们。说真的。预先计算是一种被忽视的优化方法。

解决方案 13：

下面是我在 Python 中通常用来生成素数的代码：

$ python -mtimeit -s'import sieve' 'sieve.sieve(1000000)' 
10 loops, best of 3: 445 msec per loop
$ cat sieve.py
from math import sqrt

def sieve(size):
 prime=[True]*size
 rng=xrange
 limit=int(sqrt(size))

 for i in rng(3,limit+1,+2):
  if prime[i]:
   prime[i*i::+i]=[False]*len(prime[i*i::+i])

 return [2]+[i for i in rng(3,size,+2) if prime[i]]

if __name__=='__main__':
 print sieve(100)

它无法与这里发布的更快的解决方案竞争，但至少它是纯 Python。

谢谢你提出这个问题。我今天确实学到了很多东西。

解决方案 14：

使用 Numpy 实现的半筛略有不同：

http://rebrained.com/?p=458

导入数学
导入 numpy
def prime6（最多）：
    primes=numpy.arange(3,upto+1,2)
    isprime=numpy.ones((upto-1)/2,dtype=bool)
    对于素数中的因子[：int（math.sqrt（upto））]：
        如果 isprime[(factor-2)/2]: isprime[(factor*3-2)/2:(upto-1)/2:factor]=0
    返回 numpy.insert(primes[isprime],0,2)

有人能把这个与其他时间进行比较吗？在我的计算机上，它看起来与其他 Numpy 半筛相当。

解决方案 15：

一切都已编写并测试过。因此无需重新发明轮子。

python -m timeit -r10 -s"from sympy import sieve" "primes = list(sieve.primerange(1, 10**6))"

创下了12.2 毫秒的记录！

10 loops, best of 10: 12.2 msec per loop

如果这还不够快，你可以尝试 PyPy：

pypy -m timeit -r10 -s"from sympy import sieve" "primes = list(sieve.primerange(1, 10**6))"

结果是：

10 loops, best of 10: 2.03 msec per loop

获得 247 个赞成票的答案列出了最佳解决方案的 15.9 毫秒。比较一下！！！

解决方案 16：

纯 Python中最快的素数筛：

from itertools import compress

def half_sieve(n):
    """
    Returns a list of prime numbers less than `n`.
    """
    if n <= 2:
        return []
    sieve = bytearray([True]) * (n // 2)
    for i in range(3, int(n ** 0.5) + 1, 2):
        if sieve[i // 2]:
            sieve[i * i // 2::i] = bytearray((n - i * i - 1) // (2 * i) + 1)
    primes = list(compress(range(1, n, 2), sieve))
    primes[0] = 2
    return primes

我对埃拉托斯特尼筛选法进行了速度和内存优化。

基准

from time import clock
import platform

def benchmark(iterations, limit):
    start = clock()
    for x in range(iterations):
        half_sieve(limit)
    end = clock() - start
    print(f'{end/iterations:.4f} seconds for primes < {limit}')

if __name__ == '__main__':
    print(platform.python_version())
    print(platform.platform())
    print(platform.processor())
    it = 10
    for pw in range(4, 9):
        benchmark(it, 10**pw)

输出

>>> 3.6.7
>>> Windows-10-10.0.17763-SP0
>>> Intel64 Family 6 Model 78 Stepping 3, GenuineIntel
>>> 0.0003 seconds for primes < 10000
>>> 0.0021 seconds for primes < 100000
>>> 0.0204 seconds for primes < 1000000
>>> 0.2389 seconds for primes < 10000000
>>> 2.6702 seconds for primes < 100000000

解决方案 17：

假设 N < 9,080,191，米勒-拉宾素数测试的确定性实现

import sys

def miller_rabin_pass(a, n):
    d = n - 1
    s = 0
    while d % 2 == 0:
        d >>= 1
        s += 1

    a_to_power = pow(a, d, n)
    if a_to_power == 1:
        return True
    for i in range(s-1):
        if a_to_power == n - 1:
            return True
        a_to_power = (a_to_power * a_to_power) % n
    return a_to_power == n - 1


def miller_rabin(n):
    if n <= 2:
        return n == 2

    if n < 2_047:
        return miller_rabin_pass(2, n)

    return all(miller_rabin_pass(a, n) for a in (31, 73))


n = int(sys.argv[1])
primes = [2]
for p in range(3,n,2):
  if miller_rabin(p):
    primes.append(p)
print len(primes)

根据维基百科上的文章（http://en.wikipedia.org/wiki/Miller–Rabin_primality_test），测试 N < 9,080,191，其中 a = 31 和 73 足以判断 N 是否为合数。

我改编了原始 Miller-Rabin 测试的概率实现的源代码，该源代码位于此处：https ://www.literateprograms.org/miller-rabin_primality_test__python_.html

解决方案 18：

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，因此您需要足够的空间来容纳所有内容，但它似乎扩展得相当好。

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)

结果如下：

>>>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

解决方案 19：

我知道这项比赛已经结束几年了。……

尽管如此，这是我对纯 Python 素数筛选的建议，基于在向前处理筛选时使用适当的步骤省略 2、3 和 5 的倍数。尽管如此，对于 N<10^9，它实际上比 @Robert William Hanks 的优越解决方案 rwh_primes2 和 rwh_primes1 更慢。通过使用高于 1.5* 10^8 的 ctypes.c_ushort 筛选数组，它可以以某种方式适应内存限制。

10^6

$ python -mtimeit -s"import primeSieveSpeedComp" "primeSieveSpeedComp.primeSieveSeq(1000000)" 10 次循环，3 次中最佳：每次循环 46.7 毫秒

比较：$ python -mtimeit -s"import primeSieveSpeedComp" "primeSieveSpeedComp.rwh_primes1(1000000)" 10 次循环，3 次中最好的一次：每次循环 43.2 毫秒比较：$ python -m timeit -s"import primeSieveSpeedComp" "primeSieveSpeedComp.rwh_primes2(1000000)" 10 次循环，3 次中最好的一次：每次循环 34.5 毫秒

10^7

$ python -mtimeit -s"import primeSieveSpeedComp" "primeSieveSpeedComp.primeSieveSeq(10000000)" 10 次循环，3 次中最好：每次循环 530 毫秒

比较：$ python -mtimeit -s"import primeSieveSpeedComp" "primeSieveSpeedComp.rwh_primes1(10000000)" 10 次循环，3 次中最好的一次：每次循环 494 毫秒比较：$ python -m timeit -s"import primeSieveSpeedComp" "primeSieveSpeedComp.rwh_primes2(10000000)" 10 次循环，3 次中最好的一次：每次循环 375 毫秒

10^8

$ python -mtimeit -s"import primeSieveSpeedComp" "primeSieveSpeedComp.primeSieveSeq(100000000)" 10 次循环，3 次中最好：每次循环 5.55 秒

比较： $ python -mtimeit -s"import primeSieveSpeedComp" "primeSieveSpeedComp.rwh_primes1(100000000)" 10 次循环，3 次中最好的一次：每次循环 5.33 秒比较： $ python -m timeit -s"import primeSieveSpeedComp" "primeSieveSpeedComp.rwh_primes2(100000000)" 10 次循环，3 次中最好的一次：每次循环 3.95 秒

10^9

$ python -mtimeit -s"import primeSieveSpeedComp" "primeSieveSpeedComp.primeSieveSeq(1000000000)" 10 次循环，3 次中最好：每次循环61.2秒

比较：$ python -mtimeit -n 3 -s"import primeSieveSpeedComp" "primeSieveSpeedComp.rwh_primes1(1000000000)" 3 次循环，3 次中最好的一次：每次循环97.8
秒
比较：$ python -m timeit -s"import primeSieveSpeedComp" "primeSieveSpeedComp.rwh_primes2(1000000000)" 10 次循环，3 次中最佳：每次循环 41.9 秒

您可以将下面的代码复制到 ubuntus primeSieveSpeedComp 中来查看这些测试。

def primeSieveSeq(MAX_Int):
    if MAX_Int > 5*10**8:
        import ctypes
        int16Array = ctypes.c_ushort * (MAX_Int >> 1)
        sieve = int16Array()
        #print 'uses ctypes "unsigned short int Array"'
    else:
        sieve = (MAX_Int >> 1) * [False]
        #print 'uses python list() of long long int'
    if MAX_Int < 10**8:
        sieve[4::3] = [True]*((MAX_Int - 8)/6+1)
        sieve[12::5] = [True]*((MAX_Int - 24)/10+1)
    r = [2, 3, 5]
    n = 0
    for i in xrange(int(MAX_Int**0.5)/30+1):
        n += 3
        if not sieve[n]:
            n2 = (n << 1) + 1
            r.append(n2)
            n2q = (n2**2) >> 1
            sieve[n2q::n2] = [True]*(((MAX_Int >> 1) - n2q - 1) / n2 + 1)
        n += 2
        if not sieve[n]:
            n2 = (n << 1) + 1
            r.append(n2)
            n2q = (n2**2) >> 1
            sieve[n2q::n2] = [True]*(((MAX_Int >> 1) - n2q - 1) / n2 + 1)
        n += 1
        if not sieve[n]:
            n2 = (n << 1) + 1
            r.append(n2)
            n2q = (n2**2) >> 1
            sieve[n2q::n2] = [True]*(((MAX_Int >> 1) - n2q - 1) / n2 + 1)
        n += 2
        if not sieve[n]:
            n2 = (n << 1) + 1
            r.append(n2)
            n2q = (n2**2) >> 1
            sieve[n2q::n2] = [True]*(((MAX_Int >> 1) - n2q - 1) / n2 + 1)
        n += 1
        if not sieve[n]:
            n2 = (n << 1) + 1
            r.append(n2)
            n2q = (n2**2) >> 1
            sieve[n2q::n2] = [True]*(((MAX_Int >> 1) - n2q - 1) / n2 + 1)
        n += 2
        if not sieve[n]:
            n2 = (n << 1) + 1
            r.append(n2)
            n2q = (n2**2) >> 1
            sieve[n2q::n2] = [True]*(((MAX_Int >> 1) - n2q - 1) / n2 + 1)
        n += 3
        if not sieve[n]:
            n2 = (n << 1) + 1
            r.append(n2)
            n2q = (n2**2) >> 1
            sieve[n2q::n2] = [True]*(((MAX_Int >> 1) - n2q - 1) / n2 + 1)
        n += 1
        if not sieve[n]:
            n2 = (n << 1) + 1
            r.append(n2)
            n2q = (n2**2) >> 1
            sieve[n2q::n2] = [True]*(((MAX_Int >> 1) - n2q - 1) / n2 + 1)
    if MAX_Int < 10**8:
        return [2, 3, 5]+[(p << 1) + 1 for p in [n for n in xrange(3, MAX_Int >> 1) if not sieve[n]]]
    n = n >> 1
    try:
        for i in xrange((MAX_Int-2*n)/30 + 1):
            n += 3
            if not sieve[n]:
                r.append((n << 1) + 1)
            n += 2
            if not sieve[n]:
                r.append((n << 1) + 1)
            n += 1
            if not sieve[n]:
                r.append((n << 1) + 1)
            n += 2
            if not sieve[n]:
                r.append((n << 1) + 1)
            n += 1
            if not sieve[n]:
                r.append((n << 1) + 1)
            n += 2
            if not sieve[n]:
                r.append((n << 1) + 1)
            n += 3
            if not sieve[n]:
                r.append((n << 1) + 1)
            n += 1
            if not sieve[n]:
                r.append((n << 1) + 1)
    except:
        pass
    return r

解决方案 20：

我测试了一些unutbu 的功能，我用百万数字来计算

获胜者是使用 numpy 库的函数，

注意：进行内存利用率测试也很有趣:)

计算时间结果

示例代码

完整代码在我的 github 仓库中

#!/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()

解决方案 21：

对于 Python 3

def rwh_primes2(n):
    correction = (n%6>1)
    n = {0:n,1:n-1,2:n+4,3:n+3,4:n+2,5:n+1}[n%6]
    sieve = [True] * (n//3)
    sieve[0] = False
    for i in range(int(n**0.5)//3+1):
      if sieve[i]:
        k=3*i+1|1
        sieve[      ((k*k)//3)      ::2*k]=[False]*((n//6-(k*k)//6-1)//k+1)
        sieve[(k*k+4*k-2*k*(i&1))//3::2*k]=[False]*((n//6-(k*k+4*k-2*k*(i&1))//6-1)//k+1)
    return [2,3] + [3*i+1|1 for i in range(1,n//3-correction) if sieve[i]]

解决方案 22：

我发现最简单的方法是：

primes = []
for n in range(low, high + 1):
    if all(n % i for i in primes):
        primes.append(n)

解决方案 23：

我很惊讶numba还没有人提及。

此版本在 2.47 ms ± 36.5 µs 内达到 1M 标记。

几年前，维基百科页面Prime number上给出了 Atkin 筛法的一个版本的伪代码。现在这个页面已经不存在了，对Atkin 筛法的引用似乎是一种不同的算法。2007/03/01 版本的维基百科页面，截至 2007-03-01 的 Primer number，显示了我用作参考的伪代码。

import numpy as np
from numba import njit

@njit
def nb_primes(n):
    # Generates prime numbers 2 <= p <= n
    # Atkin's sieve -- see https://en.wikipedia.org/w/index.php?title=Prime_number&oldid=111775466
    sqrt_n = int(np.sqrt(n)) + 1

    # initialize the sieve
    s = np.full(n + 1, -1, dtype=np.int8)
    s[2] = 1
    s[3] = 1

    # put in candidate primes:
    # integers which have an odd number of
    # representations by certain quadratic forms
    for x in range(1, sqrt_n):
        x2 = x * x
        for y in range(1, sqrt_n):
            y2 = y * y
            k = 4 * x2 + y2
            if k <= n and (k % 12 == 1 or k % 12 == 5): s[k] *= -1
            k = 3 * x2 + y2
            if k <= n and (k % 12 == 7): s[k] *= -1
            k = 3 * x2 - y2
            if k <= n and x > y and k % 12 == 11: s[k] *= -1

    # eliminate composites by sieving
    for k in range(5, sqrt_n):
        if s[k]:
            k2 = k*k
            # k is prime, omit multiples of its square; this is sufficient because
            # composites which managed to get on the list cannot be square-free
            for i in range(1, n // k2 + 1):
                j = i * k2 # j ∈ {k², 2k², 3k², ..., n}
                s[j] = -1
    return np.nonzero(s>0)[0]

# initial run for "compilation" 
nb_primes(10)

定时

In[10]:
%timeit nb_primes(1_000_000)

Out[10]:
2.47 ms ± 36.5 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

In[11]:
%timeit nb_primes(10_000_000)

Out[11]:
33.4 ms ± 373 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

In[12]:
%timeit nb_primes(100_000_000)

Out[12]:
828 ms ± 5.64 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

解决方案 24：

我可能加入派对的速度很慢，但我用最快的代码弥补了这一点（至少在我写这篇文章的时候，它在我的计算机上是这样的）。这种方法同时使用了 numpy 和 bitarray，灵感来自primesfrom2to这个答案。

import numpy as np
from bitarray import bitarray


def bit_primes(n):
    bit_sieve = bitarray(n // 3 + (n % 6 == 2))
    bit_sieve.setall(1)
    bit_sieve[0] = False

    for i in range(int(n ** 0.5) // 3 + 1):
        if bit_sieve[i]:
            k = 3 * i + 1 | 1
            bit_sieve[k * k // 3::2 * k] = False
            bit_sieve[(k * k + 4 * k - 2 * k * (i & 1)) // 3::2 * k] = False

    np_sieve = np.unpackbits(np.frombuffer(bit_sieve.tobytes(), dtype=np.uint8)).view(bool)
    return np.concatenate(((2, 3), ((3 * np.flatnonzero(np_sieve) + 1) | 1)))

以下是与的比较，之前在unutbu 的比较primesfrom2to中发现它是最快的解决方案：

python3 -m timeit -s "import fast_primes" "fast_primes.bit_primes(1000000)"
200 loops, best of 5: 1.19 msec per loop

python3 -m timeit -s "import fast_primes" "fast_primes.primesfrom2to(1000000)"
200 loops, best of 5: 1.23 msec per loop

对于查找 100 万以下的素数，bit_primes稍快一些。对于较大的值n，差异可能更明显。在某些情况下，bit_primes快两倍以上：

python3 -m timeit -s "import fast_primes" "fast_primes.bit_primes(500_000_000)"
1 loop, best of 5: 540 msec per loop

python3 -m timeit -s "import fast_primes" "fast_primes.primesfrom2to(500_000_000)"
1 loop, best of 5: 1.15 sec per loop

作为参考，下面是经过最少修改（可在 Python 3 中使用）的 I 版本，primesfrom2to与之进行比较：

def primesfrom2to(n):
    # https://stackoverflow.com/questions/2068372/fastest-way-to-list-all-primes-below-n-in-python/3035188#3035188
    """ Input n>=6, Returns a array of primes, 2 <= p < n"""
    sieve = np.ones(n // 3 + (n % 6 == 2), dtype=np.bool)
    sieve[0] = False
    for i in range(int(n ** 0.5) // 3 + 1):
        if sieve[i]:
            k = 3 * i + 1 | 1
            sieve[((k * k) // 3)::2 * k] = False
            sieve[(k * k + 4 * k - 2 * k * (i & 1)) // 3::2 * k] = False
    return np.r_[2, 3, ((3 * np.nonzero(sieve)[0] + 1) | 1)]

解决方案 25：

第一次使用 Python，因此我使用的一些方法可能看起来有点麻烦。我直接将我的 C++ 代码转换为 Python，这就是我所得到的（尽管在 Python 中速度有点慢）

#!/usr/bin/env python
import time

def GetPrimes(n):

    Sieve = [1 for x in xrange(n)]

    Done = False
    w = 3

    while not Done:

        for q in xrange (3, n, 2):
            Prod = w*q
            if Prod < n:
                Sieve[Prod] = 0
            else:
                break

        if w > (n/2):
            Done = True
        w += 2

    return Sieve



start = time.clock()

d = 10000000
Primes = GetPrimes(d)

count = 1 #This is for 2

for x in xrange (3, d, 2):
    if Primes[x]:
        count+=1

elapsed = (time.clock() - start)
print "
Found", count, "primes in", elapsed, "seconds!
"

pythonw Primes.py
在 12.799119 秒内找到 664579 个素数！

#!/usr/bin/env python
import time

def GetPrimes2(n):

    Sieve = [1 for x in xrange(n)]

    for q in xrange (3, n, 2):
        k = q
        for y in xrange(k*3, n, k*2):
            Sieve[y] = 0

    return Sieve



start = time.clock()

d = 10000000
Primes = GetPrimes2(d)

count = 1 #This is for 2

for x in xrange (3, d, 2):
    if Primes[x]:
        count+=1

elapsed = (time.clock() - start)
print "
Found", count, "primes in", elapsed, "seconds!
"

pythonw Primes2.py
在 10.230172 秒内找到 664579 个素数！

#!/usr/bin/env python
import time

def GetPrimes3(n):

    Sieve = [1 for x in xrange(n)]

    for q in xrange (3, n, 2):
        k = q
        for y in xrange(k*k, n, k << 1):
            Sieve[y] = 0

    return Sieve



start = time.clock()

d = 10000000
Primes = GetPrimes3(d)

count = 1 #This is for 2

for x in xrange (3, d, 2):
    if Primes[x]:
        count+=1

elapsed = (time.clock() - start)
print "
Found", count, "primes in", elapsed, "seconds!
"

python Primes2.py
在 7.113776 秒内找到 664579 个素数！

解决方案 26：

我的猜测是，最快的方法就是在代码中对素数进行硬编码。

那么为什么不编写一个缓慢的脚本来生成另一个包含所有硬连线数字的源文件，然后在运行实际程序时导入该源文件呢？

当然，这只有当您在编译时知道 N 的上限时才有效，但对于（几乎）所有项目欧拉问题来说都是如此。

附言： 我可能是错的，如果解析带有硬连线素数的源代码比首先计算它们要慢，但据我所知，Python 是从编译.pyc文件运行的，因此在这种情况下读取包含所有素数（最多 N 个）的二进制数组应该非常快。

解决方案 27：

抱歉打扰，但 erat2() 算法存在严重缺陷。

在寻找下一个合数时，我们只需要测试奇数。如果 q、p 都是奇数，那么 q+p 就是偶数，不需要测试，但 q+2*p 总是奇数。这样就消除了 while 循环条件中的“如果偶数”测试，并节省了大约 30% 的运行时间。

既然我们已经做到了这一点：不要使用优雅的“D.pop(q,None)”获取和删除方法，而是使用“if q in D: p=D[q],del D[q]”，这样速度会快一倍！至少在我的计算机上（P3-1Ghz）是这样的。所以我建议使用这个聪明的算法来实现：

def erat3( ):
    from itertools import islice, count

    # q is the running integer that's checked for primeness.
    # yield 2 and no other even number thereafter
    yield 2
    D = {}
    # no need to mark D[4] as we will test odd numbers only
    for q in islice(count(3),0,None,2):
        if q in D:                  #  is composite
            p = D[q]
            del D[q]
            # q is composite. p=D[q] is the first prime that
            # divides it. Since we've reached q, we no longer
            # need it in the map, but we'll mark the next
            # multiple of its witnesses to prepare for larger
            # numbers.
            x = q + p+p        # next odd(!) multiple
            while x in D:      # skip composites
                x += p+p
            D[x] = p
        else:                  # is prime
            # q is a new prime.
            # Yield it and mark its first multiple that isn't
            # already marked in previous iterations.
            D[q*q] = q
            yield q

解决方案 28：

我可能迟到了，但必须为此添加自己的代码。它

列出 N 以下所有素数的最快方法