trjhtr

Using the algorithm at the end of this research paper, I've written an extremely fast ¹ Ackermann function in Rust. Can the implementation be made faster?

Thoughts:

• Using a struct makes a lot of sense, but it's weird to have to Some(Box::new(v2)). Can this be avoided in some way?


• Are for loops the best way of doing the v transformations? Would (1..(i + 1)).fold(Record::new(1, 0, None), |v, k| a_use_p(k - 1, 1, v)) be more "Rustic"? It seems slightly faster, but that's hard to tell.


• Is there a way to parallelize this? Or any part of it?

Thanks.

extern crate time;

use time::PreciseTime;

fn main() 
 let m = 3;

 for n in 1..20 
 let s = PreciseTime::now();
 let res = fast_ackermann(m, n);
 let e = PreciseTime::now();
 println!("a_opt(3, ) -> took ", n, res, s.to(e))
 


#[derive(Clone)]
struct Record 
 result: u64,
 previous_result: u64,
 cache: Option<Box<Record>>,


impl Record 
 fn new(result: u64, previous_result: u64, cache: Option<Box<Record>>) -> Record 
 Record result, previous_result, cache 
 


fn fast_ackermann(m: u64, n: u64) -> u64 
 let mut cache = Record::new(1, 0, None);
 for m_builder in 0..m 
 cache = ack_with_incrementalization(m_builder, 1, cache)
 
 for n_builder in 1..(n + 1) 
 cache = ack_with_incrementalization(m, n_builder, cache)
 
 cache.result


fn ack_with_incrementalization(m: u64, n: u64, current_cache: Record) -> Record 
 if m == 0 
 Record::new(n + 1, 0, None)
 else if n == 0 
 let mut new_cache = Record::new(1, 0, None);
 for m_builder in 0..m 
 new_cache = ack_with_incrementalization(m_builder, 1, new_cache);
 
 new_cache
 else if n == 1 
 let cache_result = current_cache.result;
 let mut new_cache = current_cache;
 for n_builder in 2..(cache_result + 1) 
 new_cache = ack_with_incrementalization(m - 1, n_builder, new_cache);
 
 Record::new(new_cache.result, cache_result, Some(Box::new(new_cache)))
 else 
 let cache_result = current_cache.result;
 let mut new_cache = *current_cache.cache.unwrap();
 for n_builder in (current_cache.previous_result + 1)..(current_cache.result + 1) 
 new_cache = ack_with_incrementalization(m - 1, n_builder, new_cache);
 
 Record::new(new_cache.result, cache_result, Some(Box::new(new_cache)))
 

¹ With m = 3, n = 20, the code takes 0.95 seconds. With n = 30 it takes 16 minutes. In the paper, n = 20 takes 5.05 seconds and n =30 takes 87 minutes. A version with an intermediate cache (in Rust), n = 20 took 2.1 seconds and n = 25 (not 30 like the others) took 39 minutes.

I realized that the nested Records never go more than 3 deep (outermost, middle, innermost), which means it could be rewritten to use primitive arrays.

Comparison 1:

n: 20
fas_ack -> 8388605 took PT0.399960323S
arr_ack -> 8388605 took PT0.115302129S
n: 25
fas_ack -> 268435453 took PT11.283535003S
arr_ack -> 268435453 took PT3.553065491S
n: 30
fas_ack -> 8589934589 took PT757.605898139S
arr_ack -> 8589934589 took PT233.040801010S

Comparison 2:

n: 20
fas_ack -> 8388605 took PT0.807311828S
arr_ack -> 8388605 took PT0.442201425S
n: 25
fas_ack -> 268435453 took PT35.765848937S
arr_ack -> 268435453 took PT14.004015773S
n: 30
fas_ack -> 8589934589 took PT959.685955407S
arr_ack -> 8589934589 took PT141.390498532S

Updated code:

extern crate time;

use time::PreciseTime;

fn main() 
 let m = 3;

 for n in [20, 25, 30].iter() 
 println!("n: ", n);
 let s = PreciseTime::now();
 let res = arr_ack(m, *n);
 let e = PreciseTime::now();
 println!("arr_ack -> took ", res, s.to(e))
 


fn arr_ack(m: u64, n: u64) -> u64 
 let mut cache: [u64; 6] = [0; 6];
 cache[0] = 1;
 for m_builder in 0..m 
 cache = _arr_ack(m_builder, 1, cache)
 
 for n_builder in 1..(n + 1) 
 cache = _arr_ack(m, n_builder, cache)
 
 cache[0]


fn _arr_ack(m: u64, n: u64, mut cache: [u64; 6]) -> [u64; 6] 
 if m == 0 
 [n + 1, 0, 0, 0, 0, 0]
 else if m == 1 
 [n + 2, 0, 0, 0, 0, 0]
 else if n == 0 
 let mut new_cache = [1, 0, 0, 0, 0, 0];
 for m_builder in 0..m 
 new_cache = _arr_ack(m_builder, 1, new_cache);
 
 new_cache
 else if n == 1 
 let previous_result = cache[0];
 for n_builder in 2..(previous_result + 1) 
 cache = _arr_ack(m - 1, n_builder, cache);
 
 [cache[0], previous_result, cache[0], cache[1], cache[2], cache[3]]
 else 
 let previous_result = cache[0];
 let mut new_cache = [cache[2], cache[3], cache[4], cache[5], 0, 0];
 for n_builder in (cache[1] + 1)..(previous_result + 1) 
 new_cache = _arr_ack(m - 1, n_builder, new_cache);
 
 [new_cache[0], previous_result, new_cache[0], new_cache[1], new_cache[2], new_cache[3]]