OK, so I'm trying to write a Haskell program which counts prime numbers extremely fast. Presumably I am not the first person to try to do this. (In particular, I'm damned sure I saw some prior art, but I can't find it now...)
Initially, I want to count the number of primes less than 10^11. Currently I've left my program running for about 15 minutes and it's not even half way there yet. Some rabid C++ programmer claims his program only takes 8 seconds minutes. So clearly I'm doing something horrifyingly wrong.
(In case it matters, my current implementation uses IOUArray Integer Bool and multiple threads to process independent subranges of the search space. Currently it takes several seconds to remove all the multiples of 2 from a 10MB array chunk...)
Note that 10^11 is too big for 32-bit arithmetic. Also, 10^11 bits = 12.5 GB, far too much data to fit into Haskell's 32-bit address space. So you can't have the entire bitmap in memory at once. Finally, note that the number of primes less than 10^11 is just a shade less than 2^32, so there's no way you can store the actual integers all at once either.
Edit: Apparently I misread the timing information. What the C++ guy actually claimed was:
Counting primes < 10^11 takes 8 minutes using just one core, and 56 seconds using 4 cores. (CPU type not specified.)
Counting primes < 10^10 takes 5 seconds. (Not sure how many cores that's using.)
Sorry about the mistake...
Edit: My source code can be found here: http://hpaste.org/72898
Using the package arithmoi by the excellent StackOverflow teacher Daniel Fischer:
import Math.NumberTheory.Primes.Counting
main = print $ primeCount (10^11)
-- $ time ./arith
-- 4118054813
--
-- real 0m0.209s
-- user 0m0.198s
-- sys 0m0.008s
Which is 40 times faster than whatever your 'rabid' C++ friend has written; maybe he can learn a thing or two looking at the Haskell source ... Here are the haddocks
Some rabid C++ programmer claims his program only takes 8 seconds.
Is that wall-clock time or CPU time?
If wall-clock, and the task is split across 100 CPUs, say, it's not very impressive (it's decent), if split across 1000, it's pitiful.
If it's CPU time:
I'm pretty sure that time is not reached by actually sieving up to 1011.
With a few more than 4×109 primes until then, assuming a somewhat normal 2-3GHz CPU, you'd have 4-6 cycles per prime.
You cannot achieve that with a sieve of Eratosthenes, nor with a sieve of Atkin. Each prime has to be inspected and counted, each composite marked as such and inspected. That gives a theoretical lower bound of two cycles per number in the sieve, not counting e.g. array initialisation, loop bound checking, loop variable updates, redundant markings. You're not going to come near that theoretical bound.
A few data points:
Daniel Bernstein's primegen (Sieve of Atkin), with the sieving blocks adjusted to take full advantage of my 32KB L1-cache, takes 90 seconds to sieve the primes to 1011 and count them (234 seconds with the default sieve-block size of 8K words) on my Core i5 2410M (2.3GHz). (It's much optimised for the range up to 232, but above that, it becomes noticeably slower, for the limit 109 the times are 0.49 resp 0.64 seconds.)
My segmented Sieve of Eratosthenes, using some not exposed internals to avoid list creation, sieves and counts to 1011 in 340 seconds (sniff :-/, but hey, for 109 it took 2.92 seconds - it's getting closer, and somewhere between 1012 and 1013 it overtakes primegen :) Using the exposed interface creating a list of primes roughly doubles the time taken, as does compiling it with a 32-bit GHC.
So I'd wager that the reported time of 8 seconds - if CPU time - is, if correct, for an algorithm counting the number of primes without actually sieving the whole way. As indicated by applicative's answer, that can be done much faster.
dafis#schwartz:~/Haskell/Repos/arithmoi> time tests/primeCount 100000000000
4118054813
real 0m0.145s
user 0m0.139s
sys 0m0.006s
Note that 10^11 is too big for 32-bit arithmetic. Also, 10^11 bits = 12.5 GB, far too much data to fit into Haskell's 32-bit address space. So you can't have the entire bitmap in memory at once.
To sieve that range, you have to use a segmented sieve. Even if you're not restricted by a 32-bit address space, using such a large array will yield abysmal performance due to frequent cache misses. Your programme will spend most of its time to wait for data being transferred from main memory. Sieve in chunks that fit in your L2-cache (I haven't succeeded in trying to make it faster by making the sieve fit in L1, I guess the overhead of the GHC runtime is too large to make it work).
Also, eliminate the multiples of some small primes from the sieve, that reduces the needed work, and additionally improves performance by making the sieve smaller. Eliminating even numbers is trivial, multiples of 3 easy, multiples of 5 not very difficult.
Finally, note that the number of primes less than 10^11 is just a shade less than 2^32, so there's no way you can store the actual integers all at once either.
If you store the sieve as a list of bit-arrays, withe multiples of 2, 3 and 5 removed, you need about 3.3GB to store the chunks, so if you really can have up to 4GB, it would fit. But you should rather let the chunks you don't need anymore be garbage-collected immediately.
(In case it matters, my current implementation uses IOUArray Integer Bool and multiple threads to process independent subranges of the search space. Currently it takes several seconds to remove all the multiples of 2 from a 10MB array chunk...)
It does matter.
Use Int for the indices and unsafeRead/unsafeWrite to read and modify the array. Integer computations are much slower than Int computations, and the bounds-checking you get with readArray/writeArray really hurts.
10MB chunks are far too large, you lose cache-locality. Use chunks of a few hundred KB at most (L2 cache minus some space for other things needed).
Still, it shouldn't take several seconds to remove multiples of 2 even with Integer indices, bounds-checking and 10MB chunks. Can we see your code?
Post-vacation update:
Eight minutes to sieve the primes up to 1011 is possible without deep wizardry. I don't see how going from one to four cores could yield an eightfold speedup, since there should be no cache-effects here, but whatever, it may be, without seeing the code, I can't investigate.
So let's take a look at your code.
First, an incorrectness:
vs <-
mapM
(\ start -> do
let block = (start, start + block_size)
v <- newEmptyMVar
writeChan queue $ do
say $ "New chunk " ++ show block
chunk <- chunk_new block
sieve_top base chunk
c <- chunk_count chunk
putMVar v c
return v
)
(takeWhile (< target) $ iterate (+block_size) base_max)
The numbers base_max + k*block_size appear in two chunks each, if any of them is prime, that prime is counted twice, also you should cap the upper bound at target.
Now to the performance aspect:
One thing that jumps out is that it's real chatty, so chatty that it's measurable once you have adjusted the block_size to the cache (I took 256KB blocks for a 512KB L2 cache) - then the threads are slowed down by fighting for stdout for the if prime < 100 then say $ "Sieve " ++ show prime else return () message.
Let's look at your (silenced) sieving loop:
chunk_sieve :: Chunk -> Integer -> IO ()
chunk_sieve array prime = do
(min, max) <- getBounds array
let n0 = min `mod` prime
let n1 = if n0 == 0 then min else min - n0 + prime
mapM_
(\ n -> writeArray array n (n == prime))
(takeWhile (<= max) $ iterate (+prime) n1)
One thing that costs time is that each index is compared to the prime whose multiples are marked off. Each single comparison is cheap (though considerably more expensive than an Int comparison), but the huge number of comparisons, only one of which may yield True, adds up. Unconditionally writing False and if necessary writing True at the prime's index after the loop yields a considerable speedup.
For timing purposes I've reduced the target to 109 and ran it on two cores. The original code took 155s (elapsed, 292s user), with the reduced block_size 148s, silenced 143s. Omitting the comparison,
mapM_
(\ n -> writeArray array n False)
(takeWhile (<= max) $ iterate (+prime) n1)
when (min <= prime && prime <= max) $ writeArray array prime True
it runs in 131s.
Now it's time for some bigger speedups. Did I already mention that bounds-checking costs a lot of time? Since the loop condition guarantees that no out-of-bounds access is attempted (and the primes are small enough that no Int-overflow can happen), we should really use the unchecked access:
chunk_sieve :: Chunk -> Integer -> IO ()
chunk_sieve array prime = do
(min, max) <- getBounds array
let n0 = min `mod` prime
n1 = if n0 == 0 then min else min - n0 + prime
n2 = fromInteger (n1 - min)
mx = fromInteger (max - min)
pr = fromInteger prime
mapM_
(\ n -> unsafeWrite array n False)
(takeWhile (<= mx) $ iterate (+pr) n2)
when (min <= prime && prime <= max) $ writeArray array prime True
which reduces the running time to 96s. Much better, but still abysmal. The culprit is
takeWhile (<= mx) $ iterate (+pr) n2
GHC can't fuse that composition well, and you get a list of boxed Ints that is traversed. Replace that with an arithmetic sequence, [n2, n2+pr .. mx] and GHC happily creates a loop using unboxed Int#s, 37 seconds.
Much much better, but still bad. The biggest time-consumer now is
chunk_count :: Chunk -> IO Integer
chunk_count array = do
(min, max) <- getBounds array
work min max 0
where
work i max count = do
b <- readArray array i
let count' = count + if b then 1 else 0
evaluate count'
let i' = i+1
if i' > max
then return count'
else work i' max count'
Again, the bounds-checking costs a lot of time. With
chunk_count :: Chunk -> IO Integer
chunk_count array = do
(min, max) <- getBounds array
work 0 (fromInteger (max-min)) 0
where
work i max count = do
b <- unsafeRead array i
let count' = count + if b then 1 else 0
evaluate count'
let i' = i+1
if i' > max
then return count'
else work i' max count'
we're down to 15 seconds. Now, evaluate count' is a somewhat expensive way to make work strict in count. Using else work i' max $! count' in the last line instead of evaluate reduces the running time to 13 seconds. Defining work in a more suitable (for GHC, at least) way,
chunk_count :: Chunk -> IO Integer
chunk_count array = do
(min, max) <- getBounds array
let mx = fromInteger (max-min)
work i !ct
| mx < i = return ct
| otherwise = do
b <- unsafeRead array i
work (i+1) (if b then ct+1 else ct)
work 0 0
brings the time down to 6.55 seconds. Now we're in a situation where say $ "New chunk " ++ show block makes a measurable difference, disabling that gets us down to 6.18 seconds.
However, counting set bits by reading a byte from the array, masking off the undesired bits and comparing to 0 for each individual bit is not the most efficient way. It's faster to read entire Words from the array (via castIOUArray) and use popCount, if "you know what you're doing...", that gets us down to 4.25 seconds; stopping the marking when the square of the prime becomes larger than the upper bound of the chunk
sieve_top :: Chunk -> Chunk -> IO ()
sieve_top base chunk = work 2
where
work prime = do
chunk_sieve chunk prime
mp <- chunk_next_prime base prime
case mp of
Nothing -> return ()
Just p' -> do
(_,mx) <- getBounds chunk
when (p'*p' <= mx) $ work p'
to 3.9 seconds. Still not spectacular, but considering where we started, not bad. Just to illustrate the importance of cache locality once other bad behaviour has been reduced: the same code with the original 10MB block size takes 8.5 seconds.
Another small problem in your code is that all threads use the same mutable array of small primes for sieving. Since it is mutable, access to that must be synchronised, which adds a bit of overhead. With only two threads, the overhead isn't too big, using an immutable copy to do the sieving only reduces the time to 3.75 seconds here, but I expect that the effect would be larger for more threads. (I have only two physical cores - with hyperthreading - so using more than two threads doing the same kind of work introduces a slowdown that may invalidate conclusions drawn from that, but using four threads, I get 4.55 seconds with the immutable array versus 5.3 seconds with the mutable array. That seems to corroborate the growing synchronisation overhead.)
There's still a bit to be gained by eliminating more Integer calculations and writing code for GHC's optimiser (more worker/wrapper transformations, some static argument transformations), but not very much, maybe 10-15%.
The next big improvement is to be obtained by eliminating even numbers from the sieve. That reduces the work, allocation and running time by more than half. No prime sieve should ever consider even numbers, really, that's just a pointless waste of work.
This is a strange question/answer in that the accepted answer doesn't match the question: The question asks for help improving the speed of a sieve (correctly choosing a Page-Segmented Sieve of Eratosthenes implementation) but the accepted answer doesn't use a sieve but rather a numerical analysis technique and is just a library. Although that is fine for finding the total number of primes up to a large range very quickly (and there are faster and broader versions for doing that in other languages such as Kim Walisch's primecount in C++, and also for quickly calculating the sums of primes over a range, a sieve is useful for doing particular types of analysis such as finding prime gaps, existence of prime doubles, triples, etc. (generally K-Tuple primes), etc. In fact, general numerical analysis techniques such as the Meissel–Lehmer algorithm, upon which most of these are based, require a source of "seed primes" to start, which is best produced by an optimized Sieve of Eratosthenes.
In fact, Kim Walisch's primecount as per the above link has a GHC/Haskell API already built for it and can easily by called through the Foreign Function Interface (FFI) so therefore would be a better answer than the arithmoi library as it is faster. It is so fast that it is currently the record holder in calculating the number of primes up to 1e28! If one must make such a value available to a Haskell program and doesn't care if they understand how it gets it, it calculates the number of primes to 1e11 in tens of milliseconds.
In a similar fashion, if a sieve is what is really required, then Kim Walisch's primesieve also has a GHC Haskell FFI and could also be called directly.
While using libraries gets the job done, one doesn't learn anything about how to implement them by just using them; thus, the reason for Daniel Fischer's (DF's) very good tutorial answer and this follow-on series to what he started. DF's answer shows how to improve the question's code, but there is nowhere a summary that shows what the code should look like after working through all of his suggestions; This is especially important as the original question code in an OP's hpaste has disappeared (mercifully, as it is only a good example of how not to do it, but perhaps the code should be embedded in the question for reference), and we can only reconstruct what it did through DF's comments in his answer. This series of answers seeks to rectify that in case someone has a need for such a sieve in pure GHC Haskell, starting with a summary of the code to which DF's teaching leads one and following that up with further staged improvements.
TLDR; Jump to the end of the last of my answers for posted Haskell code that actually is almost as fast or as fast as Kim Walisch's primesieve, which up to now is probably the fastest in the world, at least up to smaller ranges of ten to a hundred billion or so and isn't beat by much above that by anything other than an extremely optimized version of YAFU which may be up to about 5% faster for large ranges. DF's final code before wheel factorization is stated to be about 40 times faster than the original question code, I extend that to where my code is 30 to 32 times faster yet again for a total of about 1200 to 1280 times faster than the original question code!.
The original question code
This will be the only time I reference it since it is no longer available (and in my opintion isn't worth modifying anyway): The only thing I liked about that code was the implementation of a thread pool, but even that was flawed in that it used mapM to feed the entire job queue to be processed by the threads to a channel, which is a non-deferred function that thus could potentially push a huge amount of work onto the job channel consuming a lot of memory rather than just pushing enough work to keep all the threads busy, and then feeding one more job for evey one returned from the results Channel. I will correct that in my code at the bottom of this answer and follow-on codes. In fact, only a results MVar pool is necessary, as the GHC runtime forks new threads faster than it takes to pipe new work to a waiting pool.
One problem with both the original code and DF's improved code is neither of them used the "-fllvm" compiler flag to use the LLVM back end code generator. For tight loops as we are trying to write here, LLVM can reduce the time per loop by as much as about a factor of two. It didn't matter for the original code, which had loops so non-tight that LLVM couldn't help, but DF's code does have tight loops and would have benefited by reducing the loop time to about 60 per cent.
The other problem with the use of MVar's (and thus Chan's) is that they aren't particularly fast, with an overhead of about three milliseconds per set activation. We had evidence of this problem in DF's answer in his final analysis where he said "using four threads, I get 4.55 seconds with the immutable array versus 5.3 seconds with the mutable array" as compared to the 3.75 seconds using two threads. Now his machine had only two cores and the extra two threads were Hyper Threaded, sharing most of the same resources as the other two, so one doesn't expect much in better performance in using them, but one doesn't expect worse performance as here. The reason is that there is so much overhead that adding the inefficient cores actually adds extra work and slows the final result. I also see this in my four "real" thread/core machine after increasing efficiency by using the LLVM back end. I only get a reduction of time to about 55% in using all four codes, which is in line with that when I use only two cores the total execution time actually increases. Since `MVar's are the only way we can implement "wait for result" using GHC, the solution to this is to make the work slices much larger (more coarse grained multi-threading) so this overhead becomes a negligible fraction, which is part of my algorithmic improvements in my seconds answer.
There is also no need to use channels as in infinite depth places to receive work and return results once the overloading problem is fixed, so I eliminated them in favor of just a "round robin" array which has the number of elements as the number of processes in the pool.
The test environments
I'm not sure if DF still has his Sandy Bridge laptop, and although I have had a Sandy Bridge CPU, it is currently down for maintenance. However, the online IDE website, Wandbox uses a Broadwell CPU of about exactly the same rating as DF's machine used in his answer at 2.3 GHz with a turbo boost to 2.9 GHz for single threaded use with two cores/four threads (Hyper Threaded). This has the same performance as DF's machine as proven by my taking his referenced internals of the "arithmoi" library and forcing it to run for a range of a billion. this Wandbox link shows it running at almost exactly the same 2.92 seconds as he mentioned in his answer. I did not bother to count the result (which is only about 0.01 milliseconds using pop count) as that would not change the comparison, but only forced it to run over the range with a 128 Kilobyte buffer size as is the default.
So, in Wandbox we have a readily referenced comparable machine; however, it has the limitation that it does not support the use of the LLVM back end, of which use is important for optimization of the tight culling loops we will be using. Accordingly, I will be doing comparisons with and without LLVM on my own machine, which is a Intel Skylake i5-2500 at 3.2 GHz with a boost to 3.6 GHz for single threaded use. There is a slight limitation in this in that the results will not scale directly by clock speed used because Skylake has a further improvement in the architecture for better branch prediction and elision of branches to as low as zero time when they are correctly predicted; as the loops which we are developing spend almost all of their time in tight loops, this can make a ten to fifteen percent reduction in execution time for the sieve implementations.
The principles behind a fast Sieve algorithm
These principles are just two, as follows:
It is important to keep the total number of operations low for a given sieving range.
Each operation must take a small number of CPU clock cycles on average.
The end execution efficiency is then the product of these two.
The performance target
DF seems to think that Atkin and Bernstein's "primegen" is a "gold standard" in sieve performance. It is not for the major reason that it does not and can not take a small number of CPU clock cycles per operation on average (principle 2) and the number of cycles it consumes increases with range faster than what I regard to be the "gold standard" - Kim Walisch's primesieve as referenced in the TLDR. While it is absolutely true that this implementation of the Sieve of Atkin (SoA) passes principle 1) above as it quickly converges to a constant number of operations of 0.2587 times the range for ranges as low as about 100 and at a range of about 1e11 this is less than the best practical implementations of a Maximally Wheel Factorized Sieve of Eratosthenes as per the combo sieve here as estimated by the formulas above that point on the web page (0.2984 for 1e11, higher with increasing range), it does not live up to expectations as to efficiency. The comparison made by the SoA document is flawed as to its comparison to the Sieve of Eratosthenese (SoE): in the same download as the code for "primegen" is the code for "eratspeed", which is a reference version of the SoE that sieves over about a billion. However, they crippled that reference version, as they limited it to the same 2/3/5 wheel factorization as is baked into the SoA (which can't be increased) instead of using the Maximally Wheel Factorized combo sieve (for which there is evidence in the file they knew of). This made the number of operations over this range a little over 400 million operations as compared to the SoA's about 258.7 million. Next, it appears that they further crippled the reference SoE in making the sieve buffer smaller than that used for the SoA so as to increase the time per operation of the SoE so about that of the SoA, in spite of those operations being simpler than those of the SoA. In this way they claimed that the SoA waas about 40% faster than the SoA.
Berstein has done some hand optimizations to the tight inner operation loop for both of these in a similar way, perhaps due to the C compilers of the day not being able to fully optimize these loops, and in order for those compilers not to undo these hand optimizations he states in the notes that the compiler should be run with only the first level of optimization. That is no longer true for the "gcc" version of today as the performance of both are increased with "-O3" high level optimizations. If both are set to 8192 32-bit words (32 Kilobytes) with the above and compiled with the optimizations as above, they both sieve to a billion in about 0.49/0.50 seconds on the DF-type machine, indicating that the number of CPU cycles per cull is about 36 per cent less for "eratspeed". If the Maximum Wheel Factorization combo principles were applied, then it should be about yet another 40% faster, and this is before doing optimizations on the tight inner culling loop; these optimizations to the culling loop are not possible with the SoA because it must use a variable-span-per-operation loop as compared to the fixed span per loop of SoE. This reference "eratspeed" SoE implementation will be discussed further down the answer(s), as this is the approach that leads to my improved algoithm answer.
As a final note about the SoA "primegen", Bernstein seems to believe that the sieve buffer needed to be limited to less than the CPU L1 cache size. While that may have been true for the CPU's on which he developed this work, it is no longer true for modern CPU's whose L2 cache performance is much faster than the SoA and close to the reference SoE tight inner loop time. Thus, if one makes the sieve buffer equal to the CPU L2 cache size (256 Kiloytes in the case of these CPU's), the time to sieve to a billion changes almost not at all at about 0.50 seconds for "primegen", but the time to sieve to 100 billion almost scales linearly to about 52 seconds (as it should). This works because the buffer has been increased so the SoA isn't plagued so quickly by operation span overflows, but it doesn't fix the problem, it just moves it further up the range and the SoA still won't be faster than a maximally optimized SoE at even the highest practical ranges.
For reference, a "primesieve" type of algorithm sieves to a billion in about 0.18 seconds for a range of a billion and about 25 seconds to 100 billion when single threaded with both times reduced by a factor of about two using two threads on the Wandbox/DF range of CPU.
State of DF's final results of his answer and proposed further work
DF stated that his final answer code would sieve to a billion in about 7.5/3.75 seconds when single threaded/run on two threads, repectively. This represents about 2.5514 billion operations and at a CPU clock of 2.9 GHz represents about 9 CPU clocks per cull (CpC). This is not good, as a basic culling loop should take about 3.5 CpC. This is the result of not using the LLVM as discussed above.
He suggested that the first further improvement would be Wheel factorization by odds-only for an improvement in speed of about 2.5 times and that further improvements using more extended wheel factorization are reasonably easily possible. This is true. However, his attempt to do extended wheel factorization in the primes function of the "arithmoi" library is very much a failure: 2.92 seconds to do 404.8 million culls at 2.9 GHz is about 21 CpC, just as the 340 seconds to do the 46.07 billion culls to sieve to 100 billion is also abysmal at about the same clocks per cull. This is so slow that there is no reason for this extended 2/3/5 wheel factorization as the result will be the same or slower than if one just used odds-only even at 9 CpC. The reason for this terrible efficiency is that he uses some complex, and thus slow, mathemeatics to do the reduction in culls for wheel factorization, but those computations take a lot of machine time. There are Look Up Table ways of doing this that are about twice as fast at about 12 CPU clock cycles per cull, but they are still too slow for use in such small ranges; their use should be limited to augmenting the efficient range for very large ranges where the percentage of the time they take is a small part of the overall time.
To show how bad these results are, here is a reference Wandbox Javascript odds-only version sieving to a billion in about 2.14 seconds or about 6.15 CpC; this runs on my Skylake machine at 1.54 seconds, with the reduced time above the ratio in clock rates due to the improvements in architecture as mentioned above for 5.4 CpC.
Further, in Haskell, here is a odds-only version from my submission to RosettaCode, the second faster version which runs on the reference Wandbox CPU in 2.26 seconds sieving to a billion compiled without LLVM for about 6.4 CpC. On my Skylake machine , this runs at 1.83 seconds and 1.023 seconds without/with LLVM respectively (6.4/3.6 CpC, respectively) and sieves to 100 billion in about 210/127 seconds no LLVM/LLVM, respectively (6.7/4.0 CpC, respectively). Note that these are faster than "arithmoi" library wheel factorized version. This will form the basis for further algorithmic improvements in my second answer.
So, the following code is an odds-only algorithm which performs as per DF mentions as his final answer:
-- Multi-threaded Page-Segmented Bit-Packed Odds-Only Sieve of Eratosthenes...
-- "Running a modern CPU single threaded is like
-- running a race car on one cylinder" me ...
-- compile with "-threaded" to use maximum available cores and threads...
-- compile with "-fllvm" for highest speed by a factor of up to two times.
{-# LANGUAGE FlexibleContexts, ScopedTypeVariables #-} -- , BangPatterns, MagicHash, UnboxedTuples, Strict
{-# OPTIONS_GHC -O2 -fllvm #-} -- or -O3 -keep-s-files -fno-cse -rtsopts
import Data.Int ( Int32, Int64 )
import Data.Word ( Word32, Word64 )
import Data.Bits ( (.&.), (.|.), shiftL, shiftR, popCount )
import Data.Array.Base (
UArray(..), listArray, assocs, unsafeAt, elems,
STUArray(..), newArray,
unsafeRead, unsafeWrite,
unsafeThaw, unsafeFreezeSTUArray, castSTUArray )
import Data.Array.ST ( runSTUArray )
import Control.Monad.ST ( ST, runST )
import Data.Time.Clock.POSIX ( getPOSIXTime )
-- imports to do with multi-threading...
import Data.Array (Array)
import Control.Monad ( forever, when )
import GHC.Conc ( getNumProcessors )
import Control.Monad.Cont ( join )
import Control.Concurrent
( ThreadId,
forkIO,
getNumCapabilities,
myThreadId,
setNumCapabilities )
import Control.Concurrent.MVar ( MVar, newEmptyMVar, putMVar, takeMVar )
import System.IO.Unsafe ( unsafePerformIO )
type Prime = Word64
type PrimeNdx = Int64
type StartAddr = Int32
type StartAddrArr = UArray Int StartAddr
type BasePrimeRep = Word32
type BasePrimeRepArr = UArray Int BasePrimeRep
type SieveBuffer = UArray Int Bool -- no point to artificial index!
-- constants related to odds-only...
cWHLPRMS :: [Prime]
cWHLPRMS = [2] -- excludes even numbers other than 2
cFRSTSVPRM :: Prime
cFRSTSVPRM = 3 -- start at first prime past the wheel prime(s)
makeSieveBuffer :: Int -> SieveBuffer
{-# INLINE makeSieveBuffer #-}
makeSieveBuffer szbts = runSTUArray $ do
newArray (0, szbts - 1) False
-- count the remaining un-marked composite bits using very fast popcount...
{-# INLINE countSieveBuffer #-}
countSieveBuffer :: Int -> SieveBuffer -> Int
countSieveBuffer lstndx sb = runST $ do
cmpsts <- unsafeThaw sb -- :: ST s (STUArray s PrimeNdx Bool)
wrdcmpsts <-
(castSTUArray :: STUArray s Int Bool ->
ST s (STUArray s Int Word64)) cmpsts
let lstwrd = lstndx `shiftR` 6
let lstmsk = 0xFFFFFFFFFFFFFFFE `shiftL` (lstndx .&. 63)
let loopwi wi cnt =
if wi < lstwrd then do
v <- unsafeRead wrdcmpsts wi
case cnt - popCount v of
ncnt -> ncnt `seq` loopwi (wi + 1) ncnt
else do
v <- unsafeRead wrdcmpsts lstwrd
return $ fromIntegral (cnt - popCount (v .|. lstmsk))
loopwi 0 (lstwrd * 64 + 64)
cWHLPTRNLEN64 :: Int
cWHLPTRNLEN64 = 2048
cWHLPTRN :: SieveBuffer -- twice as big to allow for overflow...
cWHLPTRN = makeSieveBuffer (131072 + 131072)
-- could be faster using primitive copyByteArray#...
-- in preparation for filling with pre-cull pattern...
fillSieveBuffer :: PrimeNdx -> SieveBuffer -> SieveBuffer
fillSieveBuffer lwi sb#(UArray _ _ rng _) = runSTUArray $ do
ptrn <- unsafeThaw cWHLPTRN :: ST s (STUArray s Int Bool)
ptrnu64 <- (castSTUArray :: STUArray s Int Bool ->
ST s (STUArray s Int Word64)) ptrn
cmpsts <- unsafeThaw sb :: ST s (STUArray s Int Bool)
cmpstsu64 <- (castSTUArray :: STUArray s Int Bool ->
ST s (STUArray s Int Word64)) cmpsts
let lmt = rng `shiftR` 6
lwi64 = lwi `shiftR` 6
loop i | i >= lmt = return cmpsts
| otherwise =
let mdlo = fromIntegral $ lwi64 `mod` fromIntegral cWHLPTRNLEN64
sloop j
| j >= cWHLPTRNLEN64 = loop (i + cWHLPTRNLEN64)
| otherwise = do
v <- unsafeRead ptrnu64 (mdlo + j)
unsafeWrite cmpstsu64 (i + j) v; sloop (j + 1) in sloop 0
loop 0
cullSieveBuffer :: PrimeNdx -> [BasePrimeRepArr] -> SieveBuffer -> SieveBuffer
cullSieveBuffer lwi bpras sb#(UArray _ _ rng _) = runSTUArray $ do
cmpsts <- unsafeThaw sb :: ST s (STUArray s Int Bool)
let limi = lwi + fromIntegral rng - 1
loopbpras [] = return cmpsts -- stop warning incomplete pattern match!
loopbpras (bpra#(UArray _ _ bprrng _) : bprastl) =
let loopbpi bpi
| bpi >= bprrng = loopbpras bprastl
| otherwise =
let bp = unsafeAt bpra bpi
bpndx = (fromIntegral bp - cFRSTSVPRM) `shiftR` 1
rsqri = fromIntegral ((bpndx + bpndx) * (bpndx + cFRSTSVPRM)
+ cFRSTSVPRM) - lwi in
if rsqri >= fromIntegral rng then return cmpsts else
let bpint = fromIntegral bp
bppn = fromIntegral bp
cullbits c | c >= rng = loopbpi (bpi + 1)
| otherwise = do unsafeWrite cmpsts c True
cullbits (c + bpint)
s = if rsqri >= 0 then fromIntegral rsqri else
let r = fromIntegral (-rsqri `rem` bppn)
in if r == 0 then 0 else fromIntegral (bppn - r)
in cullbits s in loopbpi 0
loopbpras bpras
-- multithreading goes here...
{-# NOINLINE cNUMPROCS #-}
cNUMPROCS :: Int -- force to the maximum number of threads available
cNUMPROCS = -- 1
-- {-
unsafePerformIO $ do -- no side effects because global!
np <- getNumProcessors; setNumCapabilities np
getNumCapabilities
--}
-- list of culled soeve buffers from index with give bit size...
makePrimePagesFrom :: forall r. PrimeNdx -> Int ->
(PrimeNdx -> SieveBuffer -> r) -> Bool -> [r]
makePrimePagesFrom stwi szbts cnvrtrf thrdd =
-- great, we can make an extra thread pool whenever we might need more, and
-- it should die and be collected whenever this goes out of scope!
let bpras = makeBasePrimeRepArrs thrdd
jbparms() =
let loop lwi szb =
(lwi, szb) : loop (lwi + fromIntegral szb) szb
in loop stwi szbts in
if thrdd then
let
{-# NOINLINE strttsk #-}
strttsk lwi szbts bpras mvr = -- do some strict work but define it non-strictly,
forkIO $ do -- else it will run in forground before threading!
-- and return it using a MVar; force strict execution in thread...
putMVar mvr $! cnvrtrf lwi $ cullSieveBuffer lwi bpras
$ fillSieveBuffer lwi $ makeSieveBuffer szbts
-- start a result pool, initialized to start with the first tasks...
{-# NOINLINE rsltpool #-}
rsltpool :: Array Int (MVar r) = unsafePerformIO $! do
mvlst <- mapM (const newEmptyMVar) [ 1 .. cNUMPROCS ] -- unique copies
mapM_ (\ (mvr, (lwi, szb)) -> strttsk lwi szb bpras mvr)
$ zip mvlst $ jbparms()
return $! listArray (0, cNUMPROCS - 1) mvlst
-- lazily loop over the entire job list...
loop (fdhd : fdtl) =
let {-# NOINLINE getnxt #-}
getnxt ((lwi, szb), i) = unsafePerformIO $! do -- wait for and get result of next page
let mvr = unsafeAt rsltpool i
r <- takeMVar mvr -- recycle mvr for next
strttsk lwi szb bpras mvr; return $! r
in getnxt fdhd : loop fdtl
-- lazily cycle over the rest of the jobs forever...
in rsltpool `seq` loop $ zip (drop cNUMPROCS $ jbparms())
(cycle [ 0 .. cNUMPROCS - 1 ]) else
-- back to non multi-threaded functions...
let loop ((lwi, szb) : jbpmstl) =
(cnvrtrf lwi . cullSieveBuffer lwi bpras . fillSieveBuffer lwi .
makeSieveBuffer) szb : loop jbpmstl
in loop $ jbparms()
makeBasePrimeRepArrs :: Bool -> [BasePrimeRepArr]
makeBasePrimeRepArrs thrdd =
let sb2bpra :: PrimeNdx -> SieveBuffer -> BasePrimeRepArr
sb2bpra lwi sb#(UArray _ _ rng _) =
let len = countSieveBuffer (rng - 1) sb
bpbs = fromIntegral cFRSTSVPRM + fromIntegral (lwi + lwi) in
listArray (0, len - 1) [ bpbs + fromIntegral (i + i) |
(i, False) <- assocs sb ]
fkbpras = [ sb2bpra 0 $ makeSieveBuffer 512 ]
bpra0 = sb2bpra 0 $ cullSieveBuffer 0 fkbpras $ makeSieveBuffer 131072
in bpra0 : makePrimePagesFrom 131072 131072 sb2bpra thrdd
-- result functions are here...
-- prepends the wheel factorized initial primes to the sieved primes output...
-- some faster not useing higher-order-functions, but still slow so who cares?
primes :: Int -> Bool -> [Prime]
primes szbts thrdd = cWHLPRMS ++ concat prmslsts where
-- convert a list of sieve buffers to a UArray of primes...
sb2prmsa :: PrimeNdx -> SieveBuffer -> UArray Int Prime
sb2prmsa lwi sb#(UArray _ _ rng _) = -- bsprm `seq` loop 0 where
let bsprm = cFRSTSVPRM + fromIntegral (lwi + lwi)
len = countSieveBuffer (rng - 1) sb in
bsprm `seq` len `seq`
listArray (0, len - 1)
[ bsprm + fromIntegral (i + i) | (i, False) <- assocs sb ]
prmslsts = map elems $ makePrimePagesFrom 0 szbts sb2prmsa thrdd
-- count the primes from the sieved page list to the limit...
countPrimesTo :: Prime -> Int -> Bool -> Int64
countPrimesTo limit szbts thrdd =
let lmtndx = fromIntegral $ (limit - cFRSTSVPRM) `shiftR` 1 :: PrimeNdx
sb2cnt lwi sb#(UArray _ _ rng _) =
let nlwi = lwi + fromIntegral rng in
if nlwi < lmtndx then (countSieveBuffer (rng - 1) sb, nlwi)
else (countSieveBuffer (fromIntegral (lmtndx - lwi)) sb, nlwi)
loop [] cnt = cnt
loop ((cnt, nxtlwi) : cntstl) ocnt =
if nxtlwi > lmtndx then ocnt + fromIntegral cnt
else loop cntstl $ ocnt + fromIntegral cnt
in if limit < cFRSTSVPRM then
if limit < 2 then 0 else 1
else loop (makePrimePagesFrom 0 szbts sb2cnt thrdd) 1
-- test it...
main :: IO ()
main = do
let limit = 10^9 :: Prime
-- page segmentation sized for most efficiency;
-- fastest with CPU L1 cache size but more address calculation overhead;
-- a little slower with CPU L2 cache size but just about enough to
-- cancell out the gain from reduced page start address calculations...
let cSIEVEPGSZ = (2^18) * 8 :: Int -- CPU L2 cache size in bits
let threaded = True
putStrLn $ "There are " ++ show cNUMPROCS ++ " threads available."
strt <- getPOSIXTime
-- let answr = length $ takeWhile (<= limit) $ primes cSIEVEPGSZ threaded -- slow way
let answr = countPrimesTo limit cSIEVEPGSZ threaded -- fast way
stop <- answr `seq` getPOSIXTime -- force evaluation of answr b4 stop time!
let elpsd = round $ 1e3 * (stop - strt) :: Int64
putStr $ "Found " ++ show answr
putStr $ " primes up to " ++ show limit
putStrLn $ " in " ++ show elpsd ++ " milliseconds."
This has been refactored from my RosettaCode submission referenced above by making it possible to have different sieve buffer sizes for the main sieve loop and the secondary base prime feed loop, as well as adding multi-threading (improved above DF's as discussed above). It runs at about the same speed as DF mentions in CpC for his final answer on equivalent to his machine without LLVM (about 3/1.5 seconds, 9 CpC to one billion) and runs at one second to a billion/125 seconds to 100 billion (3.7/4.1 CpC, respectively) on my Skylake machine with LLVM single threaded, and about half of those times when multi-threaded due to the problem of not being "coarse-grained" enough as explained above.
This answer is only a factor of less than two faster than DF's code, mostly due to the recommended use of the LLVM back end.
This answer presumes that the questioner just wanted a count of the primes to a limit by the fastest means possible and didn't know that a direct Sieve is not the fastest way of doing this, which explains why the prime counting function from the "arithmoi" library was chosen as the accepted answer. However, as analytic techniques go, the "arithmoi" prime counting function isn't the best, so this answer makes it possible to access faster methods.
The fastest currently available library to count the primes to a limit is Kim Walisch's C++ primecount repository, which can count the primes to 1e11 in about 8 milliseconds single-threaded on an Intel i5-6500 (3.6 GHz single threaded boost clock rate), about 1.5 seconds to 1e15, and about 99 seconds to 1e18, all using the fastest Xavier Gourdon algorithm, with all of these times reduced by approximately a factor of four when run multi-threaded (four cores). At this rate, this library can calculate the number of primes to the 64-bit number range (18446744073709551615) in about 163 seconds multi-threaded.
If one needed to make this computational ability and results available to Haskell, one can call the primecount library from GHC Haskell using FFI as will be described in the remainder of this answer. The primecount library is much better than the counting function from "arithmoi" as provided in the accepted answer because it uses a much better algorithm (Xavier Gourdon's of about 2000 with corrections and tuning by Kim Walisch) a greatly reduced use of RAM memory that is much less than proportional to the square root of the counting range, is based on a much better base Sieve of Eratosthenes implementation, and is multi-threaded to effectively be able to use all cores of a given computer. It holds the word record in being able to count the primes to a range of 1e29 (currently).
So the steps to call Kim Walisch's primecount from GHC Haskell are as follows:
Download the source code from the GitHub repo for your selected version from the "releases" link in either .zip or .tar.gz format (I used version 7.4).
Install a version of gcc/gcc++/g++, cmake, and make onto your machine as per the detailed build instructions(using MSYS2 on Windows, which also needs to be downloaded and installed).
Decompress the above downloaded file to whatever location using the utilities available on your machine (install 7zip if you don't already have a decompression program).
Open a terminal inside the resulting outer uncompressed folder level ("primecount-7.4" in my case).
Type the following command into the terminal (or copy and paste from here) followed by a Enter/Carriage Return: cmake . -DBUILD_SHARED_LIBS=ON -DBUILD_STATIC_LIBS=OFF
Type the following command followed by an Enter/Carriage Return to compile and link the target files: make -j
Make a GHC Haskell source code folder in a location of your choice (I named mine PrimeCount).
Copy or move the following files from the above primecount folder to this GHC Haskell source code folder: libprimecount.so.7.4 (for example, with the last digits the version number) and libprimecount.so.7; on Windows this will have a .dll extension rather than .so and there may be no other shortcut files. Then create the soft sym link file in the terminal if necessary in the source folder by the following terminal command: ln -s ./libprimecount.so.7 libprimecount.so (Creating a Windows shortcut can be more graphical in just opening the source folder in File Explorer, right clicking the destination .dll file, selecting create shortcut, then renaming the shortcut to the name of the target dll without the version numbers, if the windows shared libraries have associated version numbers).
copy the file "primecount.h" from the subfolder "include" to your GHC Haskell source code folder.
Create a file called, say, PrimeCountFFI.hs inside the GHC Haskell source code folder with the following contents:
-- this shows how to call Kim Walisch's C++ primecount...
-- compile with "ghc PrimeCountFFI -lprimecount -L."
{-# OPTIONS_GHC -O2 #-}
import System.Environment ( getArgs )
import Data.Int ( Int64, Int32 )
import Data.Word ( Word32 )
import Foreign.C.String ( CString, withCString, peekCString )
-- Pseudo FFI CTypes
type CInt = Int32
type CWord = Word32
{- 128-bit prime counting function.
- Count the number of primes <= x using Xavier Gourdon's
- algorithm. Uses all CPU cores by default.
-
- #param x Null-terminated string integer e.g. "12345".
- Note that x must be <= primecount_get_max_x() which is
- 10^31 on 64-bit systems and 2^63-1 on 32-bit systems.
- #param res Result output buffer.
- #param len Length of the res buffer. The length must be sufficiently
- large to fit the result, 32 is always enough.
- #return Returns -1 if an error occurs, else returns the number
- of characters (>= 1) that have been written to the
- res buffer, not counting the terminating null character.
-
- Run time: O(x^(2/3) / (log x)^2)
- Memory usage: O(x^(1/3) * (log x)^3) -}
foreign import ccall unsafe "primecount.h primecount_pi_str"
primeCountStr :: CString -> CString -> CWord -> IO CInt
-- Get the currently set number of threads
foreign import ccall unsafe "primecount.h primecount_get_num_threads"
getNumThreads :: IO CInt
main :: IO ()
main = do
input <- head <$> getArgs :: IO String
(rslt, answr) <- withCString input $ \ ip ->
withCString "01234567890123456789012345678901" $ \ rslt -> do
rtn <- primeCountStr ip rslt 32
answrstr <- peekCString rslt
return (rtn, answrstr)
numthrds <- getNumThreads
if rslt < 0 then error "error in computation!!!" else do
putStrLn $ "There are " ++ answr ++ " primes up to "
++ input ++ " using "
++ show numthrds ++ " threads."
Open a terminal inside the GHC Haskell source code folder and compile the program with the following command: ghc -lprimecount -L. PrimeCountFFI.
For some UNIX-like Operating Systems, one may have to tell the system to look for the linked shared library in the current directory rather than only in the standard install locations; this can be done for the current terminal session by the following shell command followed by Enter/the return key: export LD_LIBRARY_PATH=$PWD, which search path will be valid as long as the terminal session is open. You can run the program from here with a command such as the following equivalent example calls: ./PrimeCountFFI 1000000000000 or ./PrimeCountFFI 10^12 or ./PrimeCountFFI 10**12 or ./PrimeCountFFi 1e12 to find the count of primes up to the given count, in this case returning the following:
There are 37607912018 primes up to 1e12 using 4 threads.
which will take just a few milliseconds (1e14 takes about 150 milliseconds).
Now let's think about how pointless this exercise has been in that we may as well be using the primecount console application directly in which case we have command line help and all kinds of options available as to tracking the time, status, and the prime counting algorithm used as well as being able to tweak the algorithms and turn multi-threading off. Being able to call the primecount library from Haskell as shown here would only be useful if one wanted to do other processing on the results in Haskell as in using arguments generated from some other algorithm and/or passing the results for further processing in yet another algorithm.
Also, in just calling a library function whether it be primecount or the arithmoi function, we don't learn anything about how these functions work or what it takes to code the various prime counting algorithms.
Given the relative speeds of this test machine (an Intel i5-6500 at 3.6 GHz with single threaded boost) as compared to Daniel Fischer's test machine and that of the accepted answer using arithmoi, the arithmoi prime counting function should take about 100 milliseconds to count the primes to 1e11 and that is about what it takes on this machine at about 116 milliseconds, increasing to 446 milliseconds for 1e12, 2111 milliseconds for 1e13, 10535 milliseconds for 1e14, and 52263 milliseconds for 1e15, which is about the maximum range that can be used due to the high memory use. This is as compared to 468 milliseconds on this machine using Kim Walisch's primecount and 1629 milliseconds even single threaded, making the arithmoi over 30 times slower plus of limited use due to the high memory use.
I think that the arithmoi function is so slow because it uses a not very efficient Sieve of Eratosthenes as its base and also uses multi-precision Integers in far more places than it needs to, especially given that this program would never be used to count the number of primes higher than the 64-bit number range due to the high memory use.
It seems to me that there are much easier ways to be able to count the primes to these ranges even within the memory limitations of the arithmoi counting function, which I will show in my next answer.
I think another answer is required as the currently accepted answer only uses a "black box" library function with little explanation or understanding how it works and my own answer calling Kim Walisch's primecount through FFI not much better as to explanation although about 30 times faster single-threaded and scaled by the number of effective CPU cores faster due to multi-threading; this answer seeks to teach how prime counting functions work using the simplest of fast algorithms which is faster than the accepted answer's use of the arithmoi library although obviously won't be as fast as the specialized algorithms of primecount.
First, the arithmoi prime counting function is not as fast nor as elegant as it could be for the following reasons:
It uses the Meissel prime counting function technique (not the Meissel-Lehmer as stated), which is highly dependent on the speed of sieving, yet the implementation of sieving isn't all that fast in using too much math in calculating internal addresses for the wheel factorization such that it is no faster than using an odds-only page-segmented Sieve of Eratosthens (SoE) and likely slower.
It uses a non-recursive implementation of the "Phi" calculation, which is fine, but unlike usual implementations that use the Meissel algorithm to reduce the memory use, it uses memory proportional to the square root of the counting range rather than the cube root of the counting range.
An elegant implementation does not store all of the "P2" quotients (requiring an additional array of about O(n^(1/2)/log (n^(1/2))) times eight bytes of storage but runs two sieving operations, one to produce the quotients and one to process them.
The essential parts of the arithmoi files that get used by the prime counting function span multiple files and total about 1000 Lines of Code (LoC), so this very obscure and complex code will be hard to comprehend for a programmer who isn't also a mathemetician in the prime counting function field. Now, I could translate the JavaScript implementation of the Legarias, Miller, and Odlyzko (LMO) algorithm in another answer and also add further sieving improvements as mentioned in that answer, it should run at about the same speed as the LMO mode of primecount at only a few times slower than the fastest algorithm (depending on the counting range), but again that will approach 1000 LoC and will be difficult for programmers not accustomed to the field to comprehend as a first prime counting project.
This answer seeks to teach how prime counting functions work using the simpler Legendre algorithm that predates the Meissel work which was a follow on to Legendre, and will be shown to be faster than the arithmoi prime counting function although it will still use RAM memory proportional to the square root of the counting range (which is the general characteristic for algorithms of the Legendre type, not the Meissel type).
This Legendre algorithm is well known to the competitive programming communities who use it to often be the fastest of prime counting functions up to this limit of 1e11 at which it is tested, although many competitive programmers miscall it the Meissel-Lehmer algorithm (just as the arithmoi prime counting function is miss-called). I have posted a Nim language contribution to [the RosettaCode task for Legendre prime counting](https://rosettacode.org/wiki/Legendre_prime_counting_function#Non-Memoized_Versions - the last of these versions) that explains how it works by using partial sieving and which version could be translated to other languages including Haskell (it was originally translated from C++); however, it still has very high memory use of about eight times the square root of the counting range in bytes so about 0.8 Gigabytes to count the number of primes to 1e16, which is quite high.
I highly recommend reading the text from the above linked RosettaCode "partial sieving" Legendre prime counting function implementation article if one wants to understand the technique. The following Haskell code is a translation of the Nim code from that article, with a modification to the algorithm as discussed below:
{-# OPTIONS_GHC -O2 -fllvm #-}
{-# LANGUAGE FlexibleContexts, BangPatterns #-}
import Data.Time.Clock.POSIX ( getPOSIXTime ) -- for timing
import Data.Int ( Int64, Int32 )
import Data.Bits ( shiftL, shiftR, (.&.), (.|.) )
import Control.Monad ( forM_, when )
import Control.Monad.ST (ST, runST)
import Data.Array.Base ( STUArray(..), unsafeAt,
castSTUArray, unsafeFreezeSTUArray,
MArray(unsafeNewArray_, unsafeRead, unsafeWrite) )
range :: Int64
range = 10^(11 :: Int)
primeCount :: Int64 -> Int64
primeCount n =
if n < 3 then (if n < 2 then 0 else 1) else
let
{-# INLINE divide #-}
divide :: Int64 -> Int64 -> Int
divide nm d = truncate $ (fromIntegral nm :: Double) / fromIntegral d
{-# INLINE half #-}
half :: Int -> Int
half x = (x - 1) `shiftR` 1
rtlmt = floor $ sqrt (fromIntegral n :: Double)
mxndx = (rtlmt - 1) `div` 2
(!nbps, !nrs, !smalls, !roughs, !larges) = runST $ do
-- becomes `smalls` LUT -> the current counts of odd primes to index...
mss <- unsafeNewArray_ (0, mxndx) :: ST s (STUArray s Int Int32)
let msscst =
castSTUArray :: STUArray s Int Int32 -> ST s (STUArray s Int Int64)
mdss <- msscst mss -- for use in adjing counts LUT
forM_ [ 0 .. mxndx ] $ \ i -> unsafeWrite mss i (fromIntegral i)
-- becomes `roughs` LUT -> the current "k-roughs" for base prime sieved...
mrs <- unsafeNewArray_ (0, mxndx) :: ST s (STUArray s Int Int32)
forM_ [ 0 .. mxndx ] $ \ i -> unsafeWrite mrs i (fromIntegral i * 2 + 1)
-- becomes `larges` LUT -> the current count of odd primes indexed for
-- the inverse of the current "k-roughs" in the table above...
mls <- unsafeNewArray_ (0, mxndx) :: ST s (STUArray s Int Int64)
forM_ [ 0 .. mxndx ] $ \ i ->
let d = fromIntegral (i + i + 1)
in unsafeWrite mls i (fromIntegral (divide n d - 1) `div` 2)
cmpsts <- unsafeNewArray_ (0, mxndx) :: ST s (STUArray s Int Bool)
-- partial sieves to quad root of counting range, adjusting and
-- accumulating LUT's so that the overall current results are
-- accumulated to the `mls`/`larges` array...
-- also outputs `cbpi`/`nbps` is the number of base prime sieved and
-- `rlmti`/`nrs` is the effective size of the "k-roughs" sized LUT's...
let loop i !cbpi !rlmti =
let sqri = (i + i) * (i + 1) in
if sqri > mxndx then do
fss <- unsafeFreezeSTUArray mss
frs <- unsafeFreezeSTUArray mrs
fls <- unsafeFreezeSTUArray mls
return (cbpi, rlmti + 1, fss, frs, fls)
else do
v <- unsafeRead cmpsts i
if v then loop (i + 1) cbpi rlmti else do
unsafeWrite cmpsts i True -- cull current bp so not a "k-rough"!
let bp = i + i + 1
-- partial cull by current base prime...
cull c = if c > mxndx then return () else do
unsafeWrite cmpsts c True; cull (c + bp)
-- adjust `mls` array for current partial sieve;
-- also adjusts effective sizes of `mrs` and `mls`...
part ri nri = -- old "rough" index to new one...
if ri > rlmti then return (nri - 1) else do
r <- unsafeRead mrs ri -- "rough" always odd!
t <- unsafeRead cmpsts (fromIntegral r `shiftR` 1)
if t then part (ri + 1) nri else do -- skip newly culled
olv <- unsafeRead mls ri
let m = fromIntegral r * fromIntegral bp
-- split -> when multiple <= square root:
-- quotient `n / m` will be less than square root so
-- `mls` index will be found from indexing `mss`
-- (adjusted by current number bp's not in `mls`)...
adjv <- if m <= fromIntegral rtlmt then do
let ndx = fromIntegral m `shiftR` 1
sv <- unsafeRead mss ndx
unsafeRead mls (fromIntegral sv - cbpi)
-- else quotient will be less than square root so
-- quotient can be directly indexed from `mss`...
else do
sv <- unsafeRead mss (half (divide n m))
return (fromIntegral sv)
-- move "rough" and new "large" values to new places:
-- adjv includes number base primes already in `olv`
unsafeWrite mls nri (olv - (adjv - fromIntegral cbpi))
unsafeWrite mrs nri r; part (ri + 1) (nri + 1)
!pm0 = ((rtlmt `div` bp) - 1) .|. 1 -- max base prime mult
-- adjust `mss` counting table for current partial sieve;
-- for array range to `lmti`; prime multiple to `pm`...
-- adjust 64-bits at a time where possible for speed...
adjc lmti pm =
if pm < bp then return () else do
c <- unsafeRead mss (pm `shiftR` 1)
let ac = c - fromIntegral cbpi -- correction
bi = (pm * bp) `shiftR` 1 -- start array index
adj si = if si > lmti then adjc (bi - 1) (pm - 2)
else do ov <- unsafeRead mss si
unsafeWrite mss si (ov - ac)
adj (si + 1)
ac64 = fromIntegral ac :: Int64
dac = (ac64 `shiftL` 32) .|. ac64
dbi = (bi + 1) `shiftR` 1
dlmti = (lmti - 1) `shiftR` 1
dadj dsi = if dsi > dlmti then return ()
else do dov <- unsafeRead mdss dsi
unsafeWrite mdss dsi (dov - dac)
dadj (dsi + 1)
when (bi .&. 1 /= 0) $ do
ov <- unsafeRead mss bi
unsafeWrite mss bi (ov - ac)
dadj dbi
when (lmti .&. 1 == 0) $ do
ov <- unsafeRead mss lmti
unsafeWrite mss lmti (ov - ac)
adjc (bi - 1) (pm - 2)
cull sqri; nrlmti <- part 0 0; adjc mxndx pm0
loop (i + 1) (cbpi + 1) nrlmti
loop 1 0 mxndx
!ans0 = unsafeAt larges 0 - -- combine all counts; each includes nbps...
sum [ unsafeAt larges i | i <- [ 1 .. nrs - 1 ] ]
-- adjust for all the base prime counts subracted above...
!adj = (nrs + 2 * (nbps - 1)) * (nrs - 1) `div` 2
!adjans0 = ans0 + fromIntegral adj
-- add counts for base primes above quad root counting range
-- to cube root counting range multiplied by rough primes above
-- the base prime as long as the quotient of `n` divided by the
-- multiple is greater than the base prime; counts of indexed by
-- the quotient as above...
-- since all `roughs` are now prime, the multiple will always be
-- just two primes so the compensation will always be added;
-- also, the product will always be > the square root of the range so
-- the quotient will always be less than the square root of the range and
-- only the `smalls` count LUT needs be used (second case from above loop).
loopr ri !acc =
if ri >= nrs then acc else
let r = fromIntegral (unsafeAt roughs ri)
q = n `div` r
lmtsi = half (fromIntegral (q `div` r))
lmti = fromIntegral (unsafeAt smalls lmtsi) - nbps
addcnt pi !ac =
if pi > lmti then ac else
let p = fromIntegral (unsafeAt roughs pi)
ci = half (fromIntegral (divide q p))
in addcnt (pi + 1) (ac + fromIntegral (unsafeAt smalls ci))
in if lmti <= ri then acc else
-- adjust for the `nbps`'s over added in the `smalls` counts...
let !adj = fromIntegral ((lmti - ri) * (nbps + ri - 1))
in loopr (ri + 1) (addcnt (ri + 1) acc - adj)
in loopr 1 adjans0 + 1 -- add one for only even prime of two!
main :: IO ()
main = do
strt <- getPOSIXTime
let rslt = primeCount range
stop <- rslt `seq` getPOSIXTime -- force evaluation of anrswr b4 stop time!
let elpsd = round $ 1e3 * (stop - strt) :: Int64
putStrLn $ "Found " ++ show rslt ++ " primes to " ++
show range ++ " in " ++ show elpsd ++ " milliseconds."
Unfortunately, even when using the LLVM GHC back-end, before the algorithm modification the above code ran about twenty to twenty-five percent slower than the Nim or C++ code from which it was translated even when the C/C++ is compiled with clang which also has a LLVM back-end, primarily due to the current (2022) GHC Haskell compiler not emitting LLVM code that optimizes to the forms of SIMD vector instructions as the GCC C/C++ compiler can do, but also likely due to some optimization problem with the LLVM back-end, as a equivalent translation to Rust which also uses a LLVM back-end is even slower. Also the memory use is very high just as mentioned above for the arithmoi function at about eight Gigabytes of RAM to count the primes to 1e18 (which would take about an hour to complete instead of under two minutes using the fastest algorithm of primecount single-threaded).
The code was modified to speed up the "smalls" count LUT adjustment by 64-bits per loop rather than 32-bits for a gain of about ten percent so that it is only about ten percent slower than the C++ code compiled with clang/LLVM. Manually using GHC SIMD vector operations was also tried but without any further gain, likely because the memory addressing of GHC Haskell SIMD vectors isn't as efficient as the optimizations made by the C++ compilers; it may be that using GHC Haskell SIMD address offset (GHC equivalent to pointers) operations would make it faster, but for an extra small benefit it was a lot of work to try as it would require allocating the "smalls" array with pinning.
The above code is better than the arithmoi prime counting function in several respects, as follows:
It is a small at only about 100 LoC for the counting function without comments and fairly easy to understand (with the explanation from the RosettaCode contribution).
It is almost five times faster than the arithmoi code while using about the same RAM memory and likely would be five times faster if manually using the GHC SIMD address offset primitives.
It is so much simpler than the arithmoi code as to the sieving algorithm because sieving is a negligible part of the overall execution time for a prime counting function of the Legendre type so can use a simple sieve as in the odds-only SoE used here.
There are some improvements that could be made to the algorithm as follows:
Page-segmentation of the culling buffer would mean that "counts" LUT wouldn't be one huge array but would proportional to the size of the page segments.
The memory use of the above code can be reduced by different use of LUT's: using the "splitting" technique from LMO would make the "roughs" and "larges" LUT's unnecessary although there would be a "primes" table of the primes up to the square root of the counting range, which would be about O(n^(1/2) / log g) in size - reduced by the extra log term; this array could be encoded as "delta" values from the previous prime to reduce the size by a further factor of four. The large size "smalls" LUT would be essentially reduced to just the size of each page segment when page segmented as mentioned above. However, an additional table of the "special leaves" roots would be necessary which would be the square root of the counting range in size unless compressed by the "log n" factor, but compressing would require a sort and cost some time and wouldn't save all that much space since it would need to include the base prime factors in each element. In short, it isn't easy to reduce the need for space for any algorithms of the Legendre type to below being proportional to the square root of the counting range, and not by all that much (a single "log n" factor plus some small constant factors).
Using page-segmentation would make it relatively easy to convert to a multi-threaded algorithm.
The above improvements will come at the cost of some extra code complexity, but not as much as the arithmoi code or LMO implementations because of the much simpler sieving being adequate. I have not included a version with these improvements, as if one were to proceed with this amount of work, one may as well implement LMO and enjoy the benefits of memory use proportional to the cube root of the counting range, which would make it practically useful to the 64-bit counting range and higher. The main potential advantage of LMO over this function isn't so much speed but is the reduced use of RAM as counting ranges increase, which is perhaps one of the main reasons LMO was invented due to the limited RAM available on computers of the time; the main disadvantage of LMO as compared to this function is that it won't be that much much faster if any (even for larger counting ranges) without the ultimate in sieving algorithms; with an "ordinary" odds-only page-segmented sieve implementation, it won't be any faster and perhaps slower than this Legendre implmementation.
I have used many recursive functions now but still have trouble getting my head around how such a function exactly works (i'm familiar with the second line (i.e. | n==0 = 1) but am not so familiar with the final line (i.e. | n>0 = fac (n-1) * n)).
fac :: Int -> Int
fac n
| n==0 = 1
| n>0 = fac (n-1) * n
Recursive algorithms are very closely linked to mathematical induction. Perhaps studying one will help you better understand the other.
You need to keep two key principles in mind when using recursion:
Base Case
Inductive Step
The Inductive Step is often the most difficult piece, because it assumes that everything it relies upon has already been computed correctly. Making this leap of faith can be difficult (at least it took me a while to get the hang of it), but it is only because we've got preconditions on our functions; those preconditions (in this case, that n is a non-negative integer) must be specified so that the inductive step and base case are always true.
The Base Case is also sometimes difficult: say, you know that the factorial N! is N * (N-1)!, but how exactly do you handle the first step on the ladder? (In this case, it is easy, define 0! := 1. This explicit definition provides you with a way to terminate the recursive application of your Inductive Step.)
You can see your type specification and guard patterns in this function are providing the preconditions that guarantee the Inductive Step can be used over and over again until it reaches the Base Case, n == 0. If the preconditions can't be met, recursive application of the Inductive Step would fail to reach the Base Case, and your computation would never terminate. (Well, it would when it runs out of memory. :)
One complicating factor, especially with functional programming languages, is the very strong desire to re-write all 'simple' recursive functions, as you have here, with variants that use Tail Calls or Tail Recursion.
Because this function calls itself, and then performs another operation on the result, you can build a call-chain like this:
fac 3 3 * fac 2
fac 2 2 * fac 1
fac 1 1 * fac 0
fac 0 1
fac 1 1
fac 2 2
fac 3 6
That deep call stack takes up memory; but a compiler that notices that a function doesn't change any state after making a recursive call can optimize away the recursive calls. These kinds of functions typically pass along an accumulator argument. A fellow stacker has a very nice example: Tail Recursion in Haskell
factorial 1 c = c
factorial k c = factorial (k-1) (c*k)
This very complicated change :) means that the previous call chain is turned into this:
fac 3 1 fac 2 3
fac 2 3 fac 1 6
fac 1 6 6
(The nesting is there just for show; the runtime system wouldn't actually store details of the execution on the stack.)
This runs in constant memory, regardless of the value of n, and thus this optimization can convert 'impossible' algorithms into 'possible' algorithms. You'll see this kind of technique used extensively in functional programming, much as you'd see char * frequently in C programming or yield frequently in Ruby programming.
When you write | condition = expression it introduces a guard. The guards are tried in order from top to bottom until a true condition is found, and the corresponding expression is the result of your function.
This means that if n is zero, the result is 1, otherwise if n > 0 the result is fac (n-1) * n. If n is negative you get an incomplete pattern match error.
Once you've determined which expression to use, it's just a matter of substituting in the recursive calls to see what's going on.
fac 4
(fac 3) * 4
((fac 2) * 3) * 4
(((fac 1) * 2) * 3) * 4
((((fac 0) * 1) * 2) * 3) * 4
(((1 * 1) * 2) * 3) * 4
((1 * 2) * 3) * 4
(2 * 3) * 4
6 * 4
24
Especially for more complicated cases of recursion, the trick to save mental health is not to follow recursive calls, but just assume that they "do the right thing". E.g. in your fac example, you want to compute fac n. Imagine you already have the result fac (n-1). Then it's trivial to calculate fac n: just multiply it by n. But the magic of induction is that this reasonig actually works (as long as you provide a proper base case in order to terminate recursion). So e.g. for Fibonacci numbers, just look what the base case is, and assume that you are able to calculate the function for all numbers smaller then n:
fib 0 = 0
fib 1 = 1
fib n = fib (n-1) + fib (n-2)
See? You want to calculate fib n. It's easy if you would know fib (n-1) and fib (n-2). But you can simply assume you are able to calculate them, and that the "deeper levels" of recursion do the "right thing". So just use them, it will work.
Note that there are much better ways to write this function, as currently many values are recalculated very often.
BTW: The "best" way to write fac would be fac n = product [1..n].
whats throwing you? maybe the guards (the |) are confusing things.
You can think of the guards loosely as a chain of ifs, or a switch statement (difference being only one can run, and it directly evaluates to a result. does NOt perform a series of tasks, and certainly no side-effects. just evaluates to a value)
To pan imperative-like seudo-code....
Fac n:
if n == 0: return 1
if n > 0: return n * (result of calling fac w/ n decreased by one)
The call tree by other poster looks like it could be helpful. do yourself a favor and really walk through it