Adding wheel factorization to an indefinite sieve
Asked Answered
W

2

7

I’m modifying an indefinite sieve of Eratosthenes from here so it uses wheel factorization to skip more composites than its current form of just checking all odds.

I’ve worked out how to generate the steps to take to reach all the gaps along the wheel. From there I figured I could just substitute the +2’s for these wheel steps but it’s causing the sieve to skip primes. Here's the code:

from itertools import count, cycle

def dvprm(end):
    "finds primes by trial division. returns a list"
    primes=[2]
    for i in range(3, end+1, 2):
        if all(map(lambda x:i%x, primes)):
            primes.append(i)
    return primes

def prod(seq, factor=1):
    "sequence -> product"
    for i in seq:factor*=i
    return factor

def wheelGaps(primes):
    """returns list of steps to each wheel gap
    that start from the last value in primes"""
    strtPt= primes.pop(-1)#where the wheel starts
    whlCirm= prod(primes)# wheel's circumference

    #spokes are every number that are divisible by primes (composites)
    gaps=[]#locate where the non-spokes are (gaps)
    for i in xrange(strtPt, strtPt+whlCirm+1, 2):
        if not all(map(lambda x:i%x,primes)):continue#spoke 
        else: gaps.append(i)#non-spoke

    #find the steps needed to jump to each gap (beginning from the start of the wheel)
    steps=[]#last step returns to start of wheel
    for i,j in enumerate(gaps):
        if i==0:continue
        steps.append(j - gaps[i-1])
    return steps

def wheel_setup(num):
    "builds initial data for sieve"
    initPrms=dvprm(num)#initial primes from the "roughing" pump
    gaps = wheelGaps(initPrms[:])#get the gaps
    c= initPrms.pop(-1)#prime that starts the wheel

    return initPrms, gaps, c

def wheel_psieve(lvl=0, initData=None):
    '''postponed prime generator with wheels
    Refs:  http://stackoverflow.com/a/10733621
           http://stackoverflow.com/a/19391111'''

    whlSize=11#wheel size, 1 higher prime than
    # 5 gives 2- 3 wheel      11 gives 2- 7 wheel 
    # 7 gives 2- 5 wheel      13 gives 2-11 wheel
    #set to 0 for no wheel

    if lvl:#no need to rebuild the gaps, just pass them down the levels
        initPrms, gaps, c = initData
    else:#but if its the top level then build the gaps
        if whlSize>4:
            initPrms, gaps, c = wheel_setup(whlSize) 
        else:
            initPrms, gaps, c= dvprm(7), [2], 9

    #toss out the initial primes
    for p in initPrms:
        yield p

    cgaps=cycle(gaps)
    compost = {}#found composites to skip

    ps=wheel_psieve(lvl+1, (initPrms, gaps, c))

    p=next(ps)#advance lower level to appropriate square
    while p*p < c:
        p=next(ps)
    psq=p*p

    while True:
        step1 = next(cgaps)#step to next value

        step2=compost.pop(c, 0)#step to next multiple

        if not step2:

            #see references for details
            if c < psq:
                yield c
                c += step1
                continue

            else:
                step2=2*p
                p=next(ps)
                psq=p*p

        d = c + step2
        while d in compost:
            d+= step2
        compost[d]= step2

        c += step1

I'm using this to check it:

def test(num=100):
    found=[]
    for i,p in enumerate(wheel_psieve(), 1):
        if i>num:break
        found.append(p)

    print sum(found)
    return found

When I set the wheel size to 0, I get the correct sum of 24133 for the first one hundred primes, but when I use any other wheel size, I end up with missing primes, incorrect sums and composites. Even a 2-3 wheel (which uses alternate steps of 2 and 4) makes the sieve miss primes. What am I doing wrong?

Wail answered 31/5, 2015 at 4:43 Comment(0)
A
10

The odds, i.e. 2-coprimes, are generated by "rolling the wheel" [2], i.e. by repeated additions of 2, starting from the initial value of 3 (similarly from 5, 7, 9, ...),

n=3; n+=2; n+=2; n+=2; ...           # wheel = [2]
  3     5     7     9

The 2-3-coprimes are generated by repeated additions of 2, then 4, and again 2, then 4, and so on:

n=5; n+=2; n+=4; n+=2; n+=4; ...     # wheel = [2,4]
  5     7    11    13    17

Here we do need to know where to start adding the differences from, 2 or 4, depending on the initial value. For 5, 11, 17, ..., it's 2 (i.e. 0-th element of the wheel); for 7, 13, 19, ..., it's 4 (i.e. 1-st element).

How can we know where to start? The point to the wheel optimization is that we work only on this sequence of coprimes (in this example, 2-3-coprimes). So in the part of the code where we get the recursively generated primes, we will also maintain the rolling wheel stream, and advance it until we see that next prime in it. The rolling sequence will need to produce two results - the value and the wheel position. Thus when we see the prime, we also get the corresponding wheel position, and we can start off the generation of its multiples starting from that position on the wheel. We multiply everything by p of course, to start from p*p:

for (i, p) # the (wheel position, summated value) 
           in enumerated roll of the wheel:
  when p is the next prime:
    multiples of p are m =  p*p;       # map (p*) (roll wheel-at-i from p)
                       m += p*wheel[i]; 
                       m += p*wheel[i+1];    ...

So each entry in the dict will have to maintain its current value, its base prime, and its current wheel position (wrapping around to 0 for circularity, when needed).

To produce the resulting primes, we roll another coprimes sequence, and keep only those elements of it that are not in the dict, just as in the reference code.


update: after a few iterations on codereview (big thanks to the contributors there!) I've arrived at this code, using itertools as much as possible, for speed:

from itertools import accumulate, chain, cycle, count
def wsieve():  # wheel-sieve, by Will Ness.    ideone.com/mqO25A

    wh11 = [ 2,4,2,4,6,2,6,4,2,4,6, 6,2,6,4,2,6,4,6,8,4,2, 4,
             2,4,8,6,4,6,2,4,6,2,6, 6,4,2,4,6,2,6,4,2,4,2, 10,2,10]
    cs = accumulate(chain([11], cycle(wh11)))    # roll the wheel from 11
    yield(next(cs))       # cf. ideone.com/WFv4f,
    ps = wsieve()         # codereview.stackexchange.com/q/92365/9064
    p = next(ps)          # 11
    psq = p**2            # 121
    D = dict(zip(accumulate(chain([0], wh11)), count(0)))  # wheel roll lookup dict
    mults = {}
    for c in cs:          # candidates, coprime with 210, from 11
        if c in mults:
            wheel = mults.pop(c)
        elif c < psq:
            yield c
            continue
        else:    # c==psq:  map (p*) (roll wh from p) = roll (wh*p) from (p*p)
            i = D[(p-11) % 210]                 # look up wheel roll starting point
            wheel = accumulate( chain( [psq], 
                             cycle( [p*d for d in wh11[i:] + wh11[:i]])))
            next(wheel)
            p = next(ps)
            psq = p**2
        for m in wheel:   # pop, save in m, and advance
            if m not in mults:
                break
        mults[m] = wheel  # mults[143] = wheel@187

def primes():
    yield from (2, 3, 5, 7)
    yield from wsieve()

Unlike the above description, this code directly calculates where to start rolling the wheel for each prime, to generate its multiples

Alexaalexander answered 1/6, 2015 at 0:0 Comment(2)
Note that the wsieve starts at 11. If you're using this to generate a list of primes, tack [2, 3, 5, 7] to the front of it.Wail
i permitted myself to edit your code and make it PEP8 compliant. furthermore i changed p*p to p**2 - no other changes, i promise (ok. added primes generator)! feel free to roll back! not that you need an erastosthenes sieve if you have such a nice primes generator, but if you are interested: here is an old answer of mine for a memory-efficient (and pretty fast) implementation of just that.Delmydeloach
W
2

This is the version that I had come up with. It's not as clean as Ness' but it works. I'm posting it so there's another example on how to use wheel factorization in case anyone comes by. I've left in the ability to choose what wheel size to use but it's easy to nail down a more permanent one - just generate the size you want and paste that into the code.

from itertools import count

def wpsieve():
    """prime number generator
    call this function instead of roughing or turbo"""
    whlSize = 11
    initPrms, gaps, c = wheel_setup(whlSize)

    for p in initPrms:
        yield p

    primes = turbo(0, (gaps, c))

    for p, x in primes:
        yield p

def prod(seq, factor=1):
    "sequence -> product"
    for i in seq: factor *= i
    return factor

def wheelGaps(primes):
    """returns list of steps to each wheel gap
    that start from the last value in primes"""
    strtPt = primes.pop(-1)  # where the wheel starts
    whlCirm = prod(primes)  # wheel's circumference

    # spokes are every number that are divisible by primes (composites)
    gaps = []  # locate where the non-spokes are (gaps)
    for i in xrange(strtPt, strtPt + whlCirm + 1, 2):
        if not all(map(lambda x: i%x, primes)): continue  # spoke 
        else: gaps.append(i)  # non-spoke

    # find the steps needed to jump to each gap (beginning from the start of the wheel)
    steps = []  # last step returns to start of wheel
    for i, j in enumerate(gaps):
        if i == 0: continue
        steps.append(int(j - gaps[i-1]))
    return steps

def wheel_setup(num):
    "builds initial data for sieve"
    initPrms = roughing(num)  # initial primes from the "roughing" pump
    gaps = wheelGaps(initPrms[:])  # get the gaps
    c = initPrms.pop(-1)  # prime that starts the wheel

    return initPrms, gaps, c

def roughing(end):
    "finds primes by trial division (roughing pump)"
    primes = [2]
    for i in range(3, end + 1, 2):
        if all(map(lambda x: i%x, primes)):
            primes.append(i)
    return primes

def turbo(lvl=0, initData=None):
    """postponed prime generator with wheels (turbo pump)
    Refs:  http://stackoverflow.com/a/10733621
           http://stackoverflow.com/a/19391111"""

    gaps, c = initData

    yield (c, 0)

    compost = {}  # found composites to skip
    # store as current value: (base prime, wheel index)

    ps = turbo(lvl + 1, (gaps, c))

    p, x = next(ps)
    psq = p*p
    gapS = len(gaps) - 1

    ix = jx = kx = 0  # indices for cycling the wheel

    def cyc(x): return 0 if x > gapS else x  # wheel cycler

    while True:
        c += gaps[ix]  # add next step on c's wheel
        ix = cyc(ix + 1)  # and advance c's index

        bp, jx = compost.pop(c, (0,0))  # get base prime and its wheel index

        if not bp:

            if c < psq:  # prime
                yield c, ix  # emit index for above recursive level
                continue
            else:
                jx = kx  # swap indices as a new prime comes up
                bp = p
                p, kx = next(ps)
                psq = p*p

        d = c + bp * gaps[jx]  # calc new multiple
        jx = cyc(jx + 1)

        while d in compost:
            step = bp * gaps[jx]
            jx = cyc(jx + 1)
            d += step

        compost[d] = (bp, jx)

leaving in the option for the wheel size also lets you see how quickly larger wheels don't do much. Below is testing code for how long it takes to generate the wheel of selected size and how fast the sieve is with that wheel.

import time
def speed_test(num, whlSize):

    print('-'*50)

    t1 = time.time()
    initPrms, gaps, c = wheel_setup(whlSize)
    t2 = time.time()

    print('2-{} wheel'.format(initPrms[-1]))
    print('setup time: {} sec.'.format(round(t2 - t1, 5)))

    t3 = time.time()      
    prm = initPrms[:]
    primes = turbo(0, (gaps, c))
    for p, x in primes:
        prm.append(p)
        if len(prm) > num:
            break
    t4 = time.time()

    print('run time  : {} sec.'.format(len(prm), round(t4 - t3, 5)))
    print('prime sum : {}'.format(sum(prm)))

for w in [5, 7, 11, 13, 17, 19, 23, 29]:
    speed_test(1e7-1, w)

Here's how it ran on my computer using PyPy (Python 2.7 compatible) when set to generate ten million primes:

2- 3 wheel
setup time:  0.0 sec.
run time  : 18.349 sec.
prime sum : 870530414842019
--------------------------------------------------
2- 5 wheel
setup time:  0.001 sec.
run time  : 13.993 sec.
prime sum : 870530414842019
--------------------------------------------------
2- 7 wheel
setup time:  0.001 sec.
run time  :  7.821 sec.
prime sum : 870530414842019
--------------------------------------------------
2- 11 wheel
setup time:  0.03 sec.
run time  :  6.224 sec.
prime sum : 870530414842019
--------------------------------------------------
2- 13 wheel
setup time:  0.011 sec.
run time  :  5.624 sec.
prime sum : 870530414842019
--------------------------------------------------
2- 17 wheel
setup time:  0.047 sec.
run time  :  5.262 sec.
prime sum : 870530414842019
--------------------------------------------------
2- 19 wheel
setup time:  1.043 sec.
run time  :  5.119 sec.
prime sum : 870530414842019
--------------------------------------------------
2- 23 wheel
setup time: 22.685 sec.
run time  :  4.634 sec.
prime sum : 870530414842019

Larger wheels are possible, but you can see they become rather long to set up. There's also the law of diminishing returns as the wheels get larger - not much point to go past the 2-13 wheel as they don't really make it that much faster. I also ended up running into a memory error past the 2-23 wheel (which had some 36 million numbers in its gaps list).

Wail answered 26/9, 2015 at 23:52 Comment(4)
this is really fast. :)Alexaalexander
in wheelGaps you can try testing for gcd(i, d)==1, where d = whlCirm/2; maybe it'll make the setup faster?Alexaalexander
@WillNess I'm not terribly worried about the speed of the wheel gaps as their time is insignificant to the prime generation when run for long periods. But it's something I'll look at. The next step though is to see if the generator can be parallelized so it can run on multiple cores at once for an even faster speed!Wail
with a primesFrom(int start) generator, you'd just break your region up into 4(8) parts, and start 4(8) processes, each summing up its primes; then sum the 4(8) results.Alexaalexander

© 2022 - 2024 — McMap. All rights reserved.