Skip to content
Permalink
Branch: master
Find file Copy path
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
128 lines (113 sloc) 6.94 KB

Introduction

primesieve generates primes using the segmented sieve of Eratosthenes with wheel factorization. This algorithm has a run time complexity of operations and uses memory. Furthermore primesieve uses the bucket sieve algorithm which improves the cache efficiency when generating primes > 2^32. primesieve uses 8 bytes per sieving prime, hence its memory usage is about bytes per thread.

Algorithm details

Segmentation is currently the best known practical improvement to the sieve of Eratosthenes. Instead of sieving the interval [2, n] at once one subdivides the sieve interval into a number of equal sized segments that are then sieved consecutively. Segmentation drops the memory requirement of the sieve of Eratosthenes from O(n) to O(sqrt(n)). The segment size is usually chosen to fit into the CPU's fast L1 or L2 cache memory which significantly speeds up sieving. A segmented version of the sieve of Eratosthenes was first published by Singleton in 1969 [1]. Here is a simple implementation of the segmented sieve of Eratosthenes.

Wheel factorization is used to skip multiples of small primes. If a k-th wheel is added to the sieve of Eratosthenes then only those multiples are crossed off that are coprime to the first k primes, i.e. multiples that are divisible by any of the first k primes are skipped. The 1st wheel considers only odd numbers, the 2nd wheel (modulo 6) skips multiples of 2 and 3, the 3rd wheel (modulo 30) skips multiples of 2, 3, 5 and so on. Pritchard has shown in [2] that the running time of the sieve of Eratosthenes can be reduced by a factor of log log n if the wheel size is sqrt(n) but for cache reasons the sieve of Eratosthenes usually performs best with a modulo 30 or 210 wheel. Sorenson explains wheels in [3].

Additionally primesieve uses Tomás Oliveira e Silva's cache-friendly bucket list algorithm if needed [4]. This algorithm is relatively new, it has been devised by Tomás Oliveira e Silva in 2001 in order to speed up the segmented sieve of Eratosthenes for prime numbers past 32 bits. The idea is to store the sieving primes into lists of buckets with each list being associated with a segment. A list of sieving primes related to a specific segment contains only those primes that have multiple occurrence(s) in that segment. Whilst sieving a segment only the primes of the related list are used for sieving and each prime is reassigned to the list responsible for its next multiple when processed. The benefit of this approach is that it is now possible to use segments (i.e. sieve arrays) smaller than sqrt(n) without deteriorating efficiency, this is important as only small segments that fit into the CPU's L1 or L2 cache provide fast memory access.

Implementation

primesieve is written entirely in C++ and does not depend on external libraries. It's speed is mainly due to the segmentation of the sieve of Eratosthenes which prevents cache misses when crossing off multiples in the sieve array and the use of a bit array instead of a boolean sieve array. primesieve reuses and improves ideas from other great sieve of Eratosthenes implementations, namely Achim Flammenkamp's prime_sieve.c, Tomás Oliveira e Silva's A1 implementation and the author's older ecprime all written in the late '90s and '00s. Furthermore primesieve contains new optimizations to increase instruction-level parallelism and more efficiently uses the larger number of registers in today's CPUs.

Optimizations used in primesieve

  • Uses a bit array with 8 flags each 30 numbers for sieving
  • Pre-sieves multiples of small primes ≤ 19
  • Compresses the sieving primes in order to improve cache efficiency [5]
  • Starts crossing off multiples at the square
  • Uses a modulo 210 wheel that skips multiples of 2, 3, 5 and 7
  • Uses specialized algorithms for small, medium and big sieving primes
  • Uses a custom memory pool (for big sieving primes)
  • Processes multiple sieving primes per loop iteration to increase instruction-level parallelism
  • Multi-threaded using C++11 std::async

Highly optimized inner loop

primesieve's inner sieving loop has been optimized using extreme loop unrolling, on average crossing off a multiple uses just 1.375 instructions on x64 CPUs. Below is the assembly GCC generates for primesieve's inner sieving loop, each andb instruction unsets a bit (crosses off a multiple) in the sieve array.

; primesieve inner sieving loop
; Uses only 11 instructions (x86-64) to cross-off the next 8 multiples
.L99:
    andb    $-3, (%rax)
    andb    $-9, (%rax,%r14)
    andb    $127, (%rax,%r12)
    andb    $-33, (%rax,%rbp)
    andb    $-2, (%rax,%r10)
    andb    $-65, (%rax,%rbx)
    andb    $-5, (%rax,%r9)
    andb    $-17, (%rax,%r11)
    addq    %r13, %rax
    cmpq    %rdi, %rax
    jb  .L99

References

  1. R. C. Singleton, "An efficient prime number generator",
    Communications of the ACM 12, 563-564, 1969.
  2. Paul Pritchard, "Fast compact prime number sieves (among others)",
    Journal of Algorithms 4 (1983), 332-344.
  3. Jonathan Sorenson, "An analysis of two prime number sieves",
    Computer Science Technical Report Vol. 1028, 1991.
  4. Tomás Oliveira e Silva, "Fast implementation of the segmented sieve of Eratosthenes", 2002.
  5. Actually not the sieving primes are compressed but their sieve and wheel indexes.
You can’t perform that action at this time.