Fastest way to list all primes below N

Question

This is the best algorithm I could come up.

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

Can it be made even faster?

This code has a flaw: Since numbers is an unordered set, there is no guarantee that numbers.pop() will remove the lowest number from the set. Nevertheless, it works (at least for me) for some input numbers:

>>> 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

Answer 1

A very fast 1.9s< implementation using numba! I think for 100 million numbers this is the fastest! This is based on an idea I had for primes on a post long ago

import numpy
from numba import njit, prange

@njit(parallel=True)
def parallel_sieve(limit):
    sieve_size = (limit // 2) + 1
    sieve_array = np.ones(sieve_size, dtype=np.bool_)
    sieve_array[0] = False

    for i in prange(1, int(limit ** 0.5) // 2 + 1):
        if sieve_array[i]:
            p = 2 * i + 1
            start_index = (p * p) // 2
            sieve_array[start_index::p] = False

    primes = [2]
    for i in range(1, sieve_size):
        if sieve_array[i]:
            primes.append(2 * i + 1)
    
    return np.array(primes)


limit = 1_000_000_000 #1 billion
a = parallel_sieve(limit)
%timeit parallel_sieve(limit)

# 3.77 s ± 543 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)


In [13]: len(a)
Out[13]: 50847534

Answer 2

at this point of time, fermet method may be the fastest way of matching prime numbers. but there are cases when fermet may go wrong (carmichael numbers). in such cases you may wish to use the square root primes check method to do a recheck of whether it is prime or not.

i have not seen any person use this way of calculation at this time. well, say till july 2024, strange?

https://en.wikipedia.org/wiki/Fermat_number

https://en.wikipedia.org/wiki/Primality_test

# Fermet way of checking primes with base 2
# you can change the 2 into any other coprime number

def isPrimeFermet(num):
    if (num < 2): return False;
    if (num == 2 or num == 3 or num == 5 or num == 7 or num == 11):  return True
    if (2**(num - 1) % num == 1): return True;
    return False;

#
# optimised uses checks of primes for until 11 and its multiplicals 
#      before using the use of fermet method
# 
# this check reduces the possibilities of carmichael numbers which are 
#     derivatived from coprimes of 2,3,5,7,11 
# 
# you may add more of prime numbers to the checks to 
#     better the possibility until 100 or more stored primes
#     the function will have better probability to not
#     get any carmichael nonprime number
#

def isPrimeFermetOptimized(num):
    if (num < 2): return False;
    if (num == 2 or num == 3 or num == 5 or num == 7 or num == 11): 
        return True
    if ( (num == 1) or ( (num > 11) and ( ((num % 2) == 0) or ((num % 3) == 0) or ((num % 5) == 0) or ((num % 7) == 0) ) ) ):
        return False
    if (2**(num - 1) % num == 1): return True;
    return False;

get prime numbers between one (or alternatively start) to range of until r using the function isPrimeFermetRange(range) or isPrimeFermetRange(start, range)

def isPrimeFermetRange(r):
    primes = [2];
    for i in range(1, r):
        if (isPrimeFermet(i)):
            primes.append(i);
    return primes;


def isPrimeFermetRange(s, r):
    primes = [];
    for i in range(s, r):
        if (isPrimeFermet(i)):
            primes.append(i);
    return primes;

a confirmation for carmichael numbers can be done using the square root methods of checking primes.

# if you are unsure whether it is a carmical number, then
#      you may use the square root method to 
#      recheck and confirm the number to be a prime or carmical

def isPrimeSquareRoot(num):
    count = 0
    if (num < 2): return False;

    s = math.sqrt(num)
    s = int(s+1)
    for i in range(2, s):
        count += 1
        if (num % i == 0):
            return False
    return True

#
# optimised uses checks of primes for until 11 and its multiplicals 
#      before using the use of fermet method
# 
# this check reduces the possibilities of carmichael numbers which are 
#     derivatived from coprimes of 2,3,5,7,11 
# 
# you may add more of prime numbers to the checks to 
#     better the possibility until 100 or more stored primes
#     the function will have better probability to not
#     get any carmichael nonprime number
#

def isPrimeSquareRootOptimised(num):
    count = 0
    if (num < 2): return False;
    if (num == 2 or num == 3 or num == 5 or num == 7 or num == 11): 
        return True
    if ( (num == 1) or ( (num > 11) and ( ((num % 2) == 0) or ((num % 3) == 0) or ((num % 5) == 0) or ((num % 7) == 0) ) ) ):
        return False
    s = math.sqrt(num)
    s = int(s+1)
    for i in range(2, s):
        count += 1
        if (num % i == 0):
            return False
    return True

Answer 3

I have a faster sieve_of_eratosthenes(n) It is still slower than primesfrom2to(n) but is faster than the original sieve of Eratosthenes. It is a bit slower than from2 because it does not compress the primes list by 3.

def sieve_of_eratosthenes(n):
   """Create a sieve of Eratosthenes up to the given size."""
   p =np.ones(n,dtype=bool)  # One byte per value
   p[0:2] = False   # Clear zero and one
   p[4::2] = False  # Clear multiples of 2
   p[9::3] = False  # Clear multiples of 3

   for i in range(5,isqrt(n)+1,6):  # Process only numbers of the form 6k-1 and 6k+1
       if p[i]:  # If 6k-1 is prime
           p[i*i::2*i] = False  # Clear multiples of 6k-1
       h = i + 2  # 6k+1
       if p[h]:  # If 6k+1 is prime
           p[h*h::2*h] = False  # Clear multiples of 6k+1
   #return p      
   return np.r_[np.flatnonzero(p)]

Output looks like this against primesfrom2to(n):

Primes from 2 to ..
100: 25 time: 0.0001073
1,000: 168 time: 0.0002327
10,000: 1,229 time: 0.0003760
100,000: 9,592 time: 0.0007730
1,000,000: 78,498 time: 0.0033276
10,000,000: 664,579 time: 0.0242670
100,000,000: 5,761,455 time: 0.3516040
1,000,000,000: 50,847,534 time: 4.2745979

Primes bitarray sieve ..
100: 25 time: 0.0011809
1,000: 168 time: 0.0012677
10,000: 1,229 time: 0.0013816
100,000: 9,592 time: 0.0017211
1,000,000: 78,498 time: 0.0042264
10,000,000: 664,579 time: 0.0379338
100,000,000: 5,761,455 time: 0.5743241
1,000,000,000: 50,847,534 time: 6.5816255

Answer 4

You can save a lot of space by first pre-filtering down all 8 scenarios of the mod 30 wheel collapsed down to this simple formula :

((x % 30) - 15)^4 % 30 == 16

or simply

( x % 30 - 15 )^4 % 30 == 16

Then each byte can represent exactly 1 spin of the wheel, which is already 73.3% space savings right off the bat compared to 1 bit for every natural number.

(I usually prefer starting the wheel at 11,13,...,31,37 instead of 1,7,...,23,29)

Answer 5

A deterministic implementation of Miller-Rabin's Primality test on the assumption that N < 9,080,191

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)

According to the article on Wikipedia (http://en.wikipedia.org/wiki/Miller–Rabin_primality_test) testing N < 9,080,191 for a = 31 and 73 is enough to decide whether N is composite or not.

And I adapted the source code from the probabilistic implementation of original Miller-Rabin's test found here: https://www.literateprograms.org/miller-rabin_primality_test__python_.html

Answer 6

def find_primes(n):
    # Create a boolean list to track prime numbers
    primes = [True] * (n+1)
    primes[0] = primes[1] = False

    # Mark all multiples of 2 as composite (except 2 itself)
    for i in range(4, n+1, 2):
        primes[i] = False

    # Iterate over all odd numbers and mark their multiples as composite
    for i in range(3, int(n**0.5)+1, 2):
        if primes[i]:
            for j in range(i*i, n+1, 2*i):
                primes[j] = False

    # Return a list of prime numbers
    return [i for i in range(n+1) if primes[I]]
primes = find_primes(1000000)
print(primes)

The find_primes function takes an integer n as input and returns a list of all prime numbers from 1 to n. The code first creates a boolean list primes of length n+1 to track prime numbers. It then marks all multiples of 2 as composite (except 2 itself), since all other even numbers are not prime. Next, it iterates over all odd numbers from 3 to the square root of n and marks their multiples as composite. Finally, it returns a list of all prime numbers by filtering out the composite numbers from the boolean list.

Note that the Sieve of Eratosthenes algorithm has a time complexity of O(n log log n), which makes it much more efficient than testing each number for primality individually.

Answer 7

I may have been slow to join the party, but I make up for it with the fastest code (at least it is on my machine, as of writing). This approach uses both numpy and bitarray, and is inspired by primesfrom2to from this answer.

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)))

Here's a comparison with primesfrom2to, which was previously found to be the fastest solution in unutbu's comparison:

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

For finding primes under 1 million, bit_primes was slightly faster. For larger values of n, the difference can be more significant. In some cases, bit_primes was over twice as fast:

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

For reference, here's the minimally modified (to work in Python 3) version of primesfrom2to I compared with:

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)]

Answer 8

I'm surprised nobody mentioned numba yet.

This version gets to the 1M mark in 2.47 ms ± 36.5 µs.

Years ago, pseudo-code for a version of Atkin's sieve was given on the Wikipedia page Prime number. This isn't there anymore, and a reference to the Sieve of Atkin seems to be a different algorithm. A 2007/03/01 version of the Wikipedia page, Primer number as of 2007-03-01, shows the pseudo-code I used as reference.

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)

Timing

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)

Answer 9

Starting from this 2021 answer, I haven't found bitarray approach to be beneficial for primes below 1 billion.

But I was able to speedup primesfrom2to almost x2 with several tricks:

• use numexpr library to convert numpy expressions to tight loops with less allocations
• replace np.ones with a faster alternative
• manipulate first 9 elements of a sieve in a way, so there is no need to change array shape later

All in all, time for primes <1 billion went from 25s to 14.5s on my machine

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

Answer 10

I've updated much of the code for Python 3 and threw it at perfplot (a project of mine) to see which is actually fastest. Turns out that, for large n, primesfrom{2,3}to takes the cake:

---

Code to reproduce the plot:

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()

Answer 11

Faster & more memory-wise pure Python code:

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]]

or starting with half sieve

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]]

Faster & more memory-wise numpy code:

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

a faster variation starting with a third of a sieve:

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)]

A (hard-to-code) pure-python version of the above code would be:

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]]

Unfortunately pure-python don't adopt the simpler and faster numpy way of doing assignment, and calling len() inside the loop as in [False]*len(sieve[((k*k)//3)::2*k]) is too slow. So I had to improvise to correct input (& avoid more math) and do some extreme (& painful) math-magic.

Personally I think it is a shame that numpy (which is so widely used) is not part of Python standard library, and that the improvements in syntax and speed seem to be completely overlooked by Python developers.

Answer 12

You have a faster code and also the simplest code generating primes. but for the higher number, it's not working for n=10000, 10000000 it failed maybe it's .pop() method

Consider: is N prime?

• case 1:

  You got some factors of N,

  for i in range(2, N):
  

  If N is prime loop is performed for ~(N-2) times. else less number of times

• case 2:

  for i in range(2, int(math.sqrt(N)): 
  

  Loop is performed for almost ~(sqrt(N)-2) times if N is prime else will break somewhere

• case 3:

  Better We Divide N With Only number of primes<=sqrt(N)

  Where loop is performed for only π(sqrt(N)) times

  π(sqrt(N)) << sqrt(N) as N increases

  from math import sqrt
  from time import *
  prime_list = [2]
  n = int(input())
  s = time()
  for n0 in range(2,n+1):
      for i0 in prime_list:
          if n0%i0==0:
              break
          elif i0>=int(sqrt(n0)):
              prime_list.append(n0)
              break
  e = time()
  print(e-s)
  #print(prime_list); print(f'pi({n})={len(prime_list)}')
  print(f'{n}: {len(prime_list)}, time: {e-s}')
  

• Output

  100: 25, time: 0.00010275840759277344
  1000: 168, time: 0.0008606910705566406
  10000: 1229, time: 0.015588521957397461
  100000: 9592, time: 0.023436546325683594
  1000000: 78498, time: 4.1965954303741455
  10000000: 664579, time: 109.24591708183289
  100000000: 5761455, time: 2289.130858898163
  

For less than 1000 seems slow but then for <10^6 I think it's faster.

Though, I couldn't understand the time complexity.

Answer 13

Here is an interesting technique to generate prime numbers (yet not the most efficient) using python's list comprehensions:

noprimes = [j for i in range(2, 8) for j in range(i*2, 50, i)]
primes = [x for x in range(2, 50) if x not in noprimes]

Answer 14

This is a variation of the solution in the question that should be faster than what's in the question. It uses a static sieve of Eratosthenes with no other optimizations.

from typing import List

def list_primes(limit: int) -> List[int]:
    primes = set(range(2, limit + 1))
    for i in range(2, limit + 1):
        if i in primes:
            primes.difference_update(set(list(range(i, limit + 1, i))[1:]))
    return sorted(primes)

>>> list_primes(100)
[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]

Answer 15

I found a pure Python 2 prime generator here , in a comment by Willy Good, that is faster than rwh2_primes.

def primes235(limit):
yield 2; yield 3; yield 5
if limit < 7: return
modPrms = [7,11,13,17,19,23,29,31]
gaps = [4,2,4,2,4,6,2,6,4,2,4,2,4,6,2,6] # 2 loops for overflow
ndxs = [0,0,0,0,1,1,2,2,2,2,3,3,4,4,4,4,5,5,5,5,5,5,6,6,7,7,7,7,7,7]
lmtbf = (limit + 23) // 30 * 8 - 1 # integral number of wheels rounded up
lmtsqrt = (int(limit ** 0.5) - 7)
lmtsqrt = lmtsqrt // 30 * 8 + ndxs[lmtsqrt % 30] # round down on the wheel
buf = [True] * (lmtbf + 1)
for i in xrange(lmtsqrt + 1):
    if buf[i]:
        ci = i & 7; p = 30 * (i >> 3) + modPrms[ci]
        s = p * p - 7; p8 = p << 3
        for j in range(8):
            c = s // 30 * 8 + ndxs[s % 30]
            buf[c::p8] = [False] * ((lmtbf - c) // p8 + 1)
            s += p * gaps[ci]; ci += 1
for i in xrange(lmtbf - 6 + (ndxs[(limit - 7) % 30])): # adjust for extras
    if buf[i]: yield (30 * (i >> 3) + modPrms[i & 7])

My results:

$ time ./prime_rwh2.py 1e8
5761455 primes found < 1e8

real    0m3.201s
user    0m2.609s
sys     0m0.578s
$ time ./prime_wheel.py 1e8
5761455 primes found < 1e8

real    0m2.710s
user    0m2.469s
sys     0m0.219s

...on my recent midrange laptop (i5 8265U 1.6GHz) running Ubuntu on Win 10.

This is a mod 30 wheel sieve which skips multiples of 2, 3, and 5. It works great for me up to about 2.5e9, when my laptop starts running out of 8G RAM and swapping a lot.

I like mod 30 since it has only 8 remainders that aren't multiples of 2, 3, or 5. That enables using shifts and "&" for multiplication, division and mod, and should allow packing results for one mod 30 wheel into a byte. I have morphed Willy's code into a segmented mod 30 wheel sieve to eliminate the thrashing for big N and posted it here.

There is an even faster Javascript version which is segmented and uses a mod 210 wheel (no multiples of 2, 3, 5, or 7) by @GordonBGood with an in depth explanation that is useful to me.

Answer 16

The simplest way I've found of doing this is:

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

Answer 17

Here is a numpy version of Sieve of Eratosthenes having both good complexity (lower than sorting an array of length n) and vectorization. Compared to @unutbu times this just as fast as the packages with 46 microsecons to find all primes below a million.

import numpy as np 
def generate_primes(n):
    is_prime = np.ones(n+1,dtype=bool)
    is_prime[0:2] = False
    for i in range(int(n**0.5)+1):
        if is_prime[i]:
            is_prime[i**2::i]=False
    return np.where(is_prime)[0]

Timings:

import time    
for i in range(2,10):
    timer =time.time()
    generate_primes(10**i)
    print('n = 10^',i,' time =', round(time.time()-timer,6))

>> n = 10^ 2  time = 5.6e-05
>> n = 10^ 3  time = 6.4e-05
>> n = 10^ 4  time = 0.000114
>> n = 10^ 5  time = 0.000593
>> n = 10^ 6  time = 0.00467
>> n = 10^ 7  time = 0.177758
>> n = 10^ 8  time = 1.701312
>> n = 10^ 9  time = 19.322478

Answer 18

Fastest prime sieve in Pure Python:

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

I optimised Sieve of Eratosthenes for speed and memory.

Benchmark

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)

Output

>>> 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

Answer 19

Here is two updated (pure Python 3.6) versions of one of the fastest functions,

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:])]

Answer 20

For 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]]

Answer 21

I tested some unutbu's functions, i computed it with hungred millions number

The winners are the functions that use numpy library,

Note: It would also interesting make a memory utilization test :)

Sample code

Complete code on my github repository

#!/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()

Answer 22

It's all written and tested. So there is no need to reinvent the wheel.

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

gives us a record breaking 12.2 msec!

10 loops, best of 10: 12.2 msec per loop

If this is not fast enough, you can try PyPy:

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

which results in:

10 loops, best of 10: 2.03 msec per loop

The answer with 247 up-votes lists 15.9 ms for the best solution. Compare this!!!

Answer 23

If you accept itertools but not numpy, here is an adaptation of rwh_primes2 for Python 3 that runs about twice as fast on my machine. The only substantial change is using a bytearray instead of a list for the boolean, and using compress instead of a list comprehension to build the final list. (I'd add this as a comment like moarningsun if I were able.)

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

Comparisons:

>>> 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

and

>>> 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

Answer 24

Warning: timeit results may vary due to differences in hardware or version of Python.

Below is a script which compares a number of implementations:

• ambi_sieve_plain,
• rwh_primes,
• rwh_primes1,
• rwh_primes2,
• sieveOfAtkin,
• sieveOfEratosthenes,
• sundaram3,
• sieve_wheel_30,
• ambi_sieve (requires numpy)
• primesfrom3to (requires numpy)
• primesfrom2to (requires numpy)

Many thanks to stephan for bringing sieve_wheel_30 to my attention. Credit goes to Robert William Hanks for primesfrom2to, primesfrom3to, rwh_primes, rwh_primes1, and rwh_primes2.

Of the plain Python methods tested, with psyco, for n=1000000, rwh_primes1 was the fastest tested.

+---------------------+-------+
| 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 |
+---------------------+-------+

Of the plain Python methods tested, without psyco, for n=1000000, rwh_primes2 was the fastest.

+---------------------+-------+
| 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 |
+---------------------+-------+

Of all the methods tested, allowing numpy, for n=1000000, primesfrom2to was the fastest tested.

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

Timings were measured using the command:

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

with {method} replaced by each of the method names.

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)

Running the script tests that all implementations give the same result.

Answer 25

It's instructive to write your own prime finding code, but it's also useful to have a fast reliable library at hand. I wrote a wrapper around the C++ library primesieve, named it primesieve-python

Try it pip install primesieve

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

I'd be curious to see the speed compared.

Answer 26

If you don't want to reinvent the wheel, you can install the symbolic maths library sympy (yes it's Python 3 compatible)

pip install sympy

And use the primerange function

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

Answer 27

This is the way you can compare with others.

# You have to list primes upto n
nums = xrange(2, n)
for i in range(2, 10):
    nums = filter(lambda s: s==i or s%i, nums)
print nums

So simple...

Answer 28

There's a pretty neat sample from the Python Cookbook here -- the fastest version proposed on that URL is:

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

so that would give

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

Measuring at the shell prompt (as I prefer to do) with this code in pri.py, I observe:

$ 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

so it looks like the Cookbook solution is over twice as fast.

Answer 29

I'm slow responding to this question but it seemed like a fun exercise. I'm using numpy which might be cheating and I doubt this method is the fastest but it should be clear. It sieves a Boolean array referring to its indices only and elicits prime numbers from the indices of all True values. No modulo needed.

import numpy as np
def ajs_primes3a(upto):
    mat = np.ones((upto), dtype=bool)
    mat[0] = False
    mat[1] = False
    mat[4::2] = False
    for idx in range(3, int(upto ** 0.5)+1, 2):
        mat[idx*2::idx] = False
    return np.where(mat == True)[0]

Answer 30

I collected several prime number sieves over time. The fastest on my computer is this:

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