我很惊讶居然没人提到numba。

该版本在2.47 ms±36.5µs内达到1M标记。

几年前，维基百科页面上出现了一个阿特金筛子的伪代码。这已经不存在了，参考阿特金筛似乎是一个不同的算法。一个2007/03/01版本的维基百科页面(Primer number as 2007-03-01)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)

时机

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)

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。

到目前为止，我尝试过的最快的方法是基于Python烹饪书erat2函数:

import itertools as it
def erat2a( ):
    D = {  }
    yield 2
    for q in it.islice(it.count(3), 0, None, 2):
        p = D.pop(q, None)
        if p is None:
            D[q*q] = q
            yield q
        else:
            x = q + 2*p
            while x in D:
                x += 2*p
            D[x] = p

关于加速的解释，请看下面的答案。

我知道比赛已经结束好几年了。.．.

尽管如此，这是我对纯python质数筛子的建议，基于在向前处理筛子时使用适当的步骤省略2、3和5的倍数。尽管如此，在N<10^9时，它实际上比@Robert William Hanks的优解rwh_primes2和rwh_primes1要慢。通过使用大于1.5* 10^8的ctypes.c_ushort筛分数组，可以在某种程度上适应内存限制。

10^6

$ python -mtimeit -s"import primeSieveSpeedComp" "primeSieveSpeedComp. primesieveseq (1000000)" 10个循环，最好的3:46.7毫秒每循环

import primeSieveSpeedComp (primeSieveSpeedComp) “primeSieveSpeedComp.rwh_primes1(1000000)”10个循环，最好的3:43.2 每回路Msec $ python -m timeit -s"import primeSieveSpeedComp" “primeSieveSpeedComp.rwh_primes2(1000000)”10圈，最好成绩是3:34.5 每回路Msec

10^7

$ python -mtimeit -s"import primeSieveSpeedComp" "primeSieveSpeedComp. primesieveseq (10000000)" 10个循环，最好是3:530毫秒每循环

import primeSieveSpeedComp (primeSieveSpeedComp) “primeSieveSpeedComp.rwh_primes1(10000000)”10圈，3:494的最佳成绩 每回路Msec $ python -m timeit -s"import primeSieveSpeedComp" “primeSieveSpeedComp.rwh_primes2(10000000)”10圈，最好的3:375 每回路Msec

10^8

$ python -mtimeit -s"import primeSieveSpeedComp" "primeSieveSpeedComp. primesieveseq (100000000)" 10圈，最好的3:5.55秒每圈

import primeSieveSpeedComp (primeSieveSpeedComp) “primeSieveSpeedComp.rwh_primes1(100000000)”10圈，最好成绩是3:5.33 秒/循环 $ python -m timeit -s"import primeSieveSpeedComp" “primeSieveSpeedComp.rwh_primes2(100000000)”10圈，最好的3:3.95 秒/循环

10^9

$ python -mtimeit -s"import primeSieveSpeedComp" "primeSieveSpeedComp. primesieveseq (1000000000)" 10圈，最好的3圈:每圈61.2秒

$ python -mtimeit -n 3 -s"import primeSieveSpeedComp" “primeSieveSpeedComp.rwh_primes1(1000000000)”3圈，最好的3:97.8 秒/循环 $ python -m timeit -s"import primeSieveSpeedComp" “primeSieveSpeedComp.rwh_primes2(1000000000)”10个循环，3个最好: 每循环41.9秒

您可以将下面的代码复制到ubuntu primeSieveSpeedComp中以查看此测试。

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

在写这篇文章的时候，这是最快的工作解决方案(至少在我的机器上是这样)。它同时使用numpy和bitarray，并受到这个答案的primesfrom2to的启发。

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)))

下面是与素数from2to的比较，它之前被发现是unutbu比较中最快的解:

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

对于寻找100万以下的质数，bit_primes稍微快一些。 n值越大，差异就越大。在某些情况下，bit_primes的速度是原来的两倍多:

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，以下是primesfrom2to I的最小修改版本(适用于Python 3):

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)]

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