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
还能做得更快吗?
这段代码有一个缺陷:由于numbers是一个无序集，不能保证numbers.pop()将从集合中移除最低的数字。尽管如此，它还是适用于(至少对我来说)一些输入数字:
>>> 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
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)]
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)]
使用Sundaram的Sieve，我想我打破了pure-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)
Comparasion:
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
这里有一个来自Python Cookbook的非常简洁的示例——该URL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
这就给出了
def get_primes_erat(n):
return list(itertools.takewhile(lambda p: p<n, erat2()))
在shell提示符(正如我喜欢做的那样)中测量这段代码在pri.py中，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
所以看起来食谱解决方案的速度是原来的两倍多。
在Pure Python中最快的质数筛分:
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
我优化了埃拉托色尼筛子的速度和内存。
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
假设N < 9,080,191, Miller-Rabin's Primality检验的确定性实现
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(http://en.wikipedia.org/wiki/Miller -Rabin_primality_test)上的文章，对于a = 37和73，测试N < 9,080,191足以判断N是否为合数。
我从原始米勒-拉宾测试的概率实现中改编了源代码:https://www.literateprograms.org/miller-rabin_primality_test__python_.html
