A collection of high performance functions implemented in C++ with Rcpp for solving problems in combinatorics and computational mathematics. Utilizes the library RcppThread where multithreading is needed. We also make use of the RMatrix.h header file from RcppParallel for thread safe accessors for Rcpp matrices. Featured functions:

`primeSieve`

- Generates all primes less than a**billion in under 0.5 seconds.**

- It is also efficient for large numbers by using the cache friendly improvements originally developed by Tomás Oliveira.

`primeCount`

- Counts the number of primes below a**1e12 in ~130 milliseconds**and**1e15 in under 50 seconds.**

`comboGeneral`

/`permuteGeneral`

- Generate all combinations/permutations of a vector (including multisets) meeting specific criteria.- Produce results in parallel using the
`Parallel`

argument. You can also apply each of the five compiled functions given by the argument`constraintFun`

in parallel as well. E.g. Obtaining the row sums of all combinations:`comboGeneral(20, 10, constraintFun = "sum", Parallel = TRUE)`

- Alternatively, the arguments
`lower`

and`upper`

make it possible to generate combinations/permutations in chunks allowing for parallelization via the package`parallel`

. This is convenient when you want to apply a custom function to the output in parallel as well (see this stackoverflow post for a use case). - GMP support allows for exploration of combinations/permutations of vectors with many elements.

- Produce results in parallel using the
`comboSample`

/`permuteSample`

- Easily generate random samples of combinations/permutations in parallel.- You can pass a vector of specific indices or rely on the internal sampling functions. We call
`sample`

when the total number of results is small and for larger cases, the sampling is done in a very similar fashion to`urand.bigz`

from the`gmp`

package.

- You can pass a vector of specific indices or rely on the internal sampling functions. We call

The `primeSieve`

function and the `primeCount`

function are both based off of the excellent work by Kim Walisch. The respective repos can be found here: kimwalisch/primesieve; kimwalisch/primecount

Additionally, many of the sieving functions make use of the fast integer division library libdivide by ridiculousfish.

```
install.packages("RcppAlgos")
## Or install the development version
devtools::install_github("jwood000/RcppAlgos")
```

Easily executed with a very simple interface. Output is in lexicographical order.

```
## Find all 3-tuples combinations without
## repetition of the numbers c(1, 2, 3, 4).
comboGeneral(4, 3)
[,1] [,2] [,3]
[1,] 1 2 3
[2,] 1 2 4
[3,] 1 3 4
[4,] 2 3 4
## Find all 3-tuples permutations without
## repetition of the numbers c(1, 2, 3, 4).
permuteGeneral(4, 3)
[,1] [,2] [,3]
[1,] 1 2 3
[2,] 1 2 4
[3,] 1 3 2
[4,] 1 3 4
. . . .
. . . .
[21,] 4 2 1
[22,] 4 2 3
[23,] 4 3 1
[24,] 4 3 2
## For combinations/permutations with repetition, simply
## set the repetition argument to TRUE
comboGeneral(4, 3, repetition = TRUE)
[,1] [,2] [,3]
[1,] 1 1 1
[2,] 1 1 2
[3,] 1 1 3
. . . .
. . . .
## They are very efficient
system.time(comboGeneral(25, 13))
user system elapsed
0.116 0.062 0.178
system.time(comboGeneral(25, 13, nThreads = 8))
user system elapsed
0.192 0.228 0.058
nrow(comboGeneral(25,13))
[1] 5200300
```

Oftentimes, one needs to find combinations/permutations that meet certain requirements.

```
## Generate some random data
set.seed(101)
s <- sample(500, 20)
## Find all 5-tuples permutations without repetition
## of s (defined above) such that the sum is equal to 1176.
p <- permuteGeneral(v = s,
m = 5,
constraintFun = "sum",
comparisonFun = "==",
limitConstraints = 1176,
keepResults = TRUE)
tail(p)
[,1] [,2] [,3] [,4] [,5] [,6]
[3955,] 354 287 149 187 199 1176
[3956,] 354 287 149 199 187 1176
[3957,] 354 287 187 149 199 1176
[3958,] 354 287 187 199 149 1176
[3959,] 354 287 199 149 187 1176
[3960,] 354 287 199 187 149 1176
## Get combinations such that the product is between
## 3600 and 4000 (including 3600 but not 4000)
comboGeneral(5, 7, TRUE, constraintFun = "prod",
comparisonFun = c(">=","<"),
limitConstraints = c(3600, 4000),
keepResults = TRUE)
[,1] [,2] [,3] [,4] [,5] [,6] [,7] [,8]
[1,] 1 2 3 5 5 5 5 3750
[2,] 1 3 3 4 4 5 5 3600
[3,] 1 3 4 4 4 4 5 3840
[4,] 2 2 3 3 4 5 5 3600
[5,] 2 2 3 4 4 4 5 3840
[6,] 3 3 3 3 3 3 5 3645
[7,] 3 3 3 3 3 4 4 3888
```

Sometimes, the standard combination/permutation functions don’t quite get us to our desired goals. For example, one may need all permutations of a vector with some of the elements repeated a specific amount of times (i.e. a multiset). Consider the following vector `a <- c(1,1,1,1,2,2,2,7,7,7,7,7)`

and one would like to find permutations of `a`

of length 6. Using traditional methods, we would need to generate all permutations, then eliminate duplicate values. Even considering that `permuteGeneral`

is very efficient, this approach is clunky and not as fast as it could be. Observe:

```
getPermsWithSpecificRepetition <- function(z, n) {
b <- permuteGeneral(z, n)
myDupes <- duplicated(apply(b, 1, paste, collapse=""))
b[!myDupes, ]
}
a <- c(1,1,1,1,2,2,2,7,7,7,7,7)
system.time(test <- getPermsWithSpecificRepetition(a, 6))
user system elapsed
4.168 0.046 4.230
```

Situations like this call for the use of the `freqs`

argument. Simply, enter the number of times each unique element is repeated and Voila!

```
system.time(test2 <- permuteGeneral(unique(a), 6, freqs = rle(a)$lengths))
user system elapsed
0 0 0
identical(test, test2)
[1] TRUE
```

Here are some more general examples with multisets:

```
## Generate all permutations of a vector with specific
## length of repetition for each element (i.e. multiset)
permuteGeneral(3, freqs = c(1,2,2))
[,1] [,2] [,3] [,4] [,5]
[1,] 1 2 2 3 3
[2,] 1 2 3 2 3
[3,] 1 2 3 3 2
[4,] 1 3 2 2 3
[5,] 1 3 2 3 2
. . . . . .
. . . . . .
[26,] 3 2 3 1 2
[27,] 3 2 3 2 1
[28,] 3 3 1 2 2
[29,] 3 3 2 1 2
[30,] 3 3 2 2 1
## or combinations of a certain length
comboGeneral(3, 2, freqs = c(1,2,2), constraintFun = "prod")
[,1] [,2] [,3]
[1,] 1 2 2
[2,] 1 3 3
[3,] 2 2 4
[4,] 2 3 6
[5,] 3 3 9
```

```
facPerms <- permuteGeneral(factor(c("low", "med", "high"),
levels = c("low", "med", "high"),
ordered = TRUE),
freqs = c(1,4,2))
> facPerms[1:5, ]
[,1] [,2] [,3] [,4] [,5] [,6] [,7]
[1,] low med med med med high high
[2,] low med med med high med high
[3,] low med med med high high med
[4,] low med med high med med high
[5,] low med med high med high med
Levels: low < med < high
```

Using the parameter `Parallel`

, we can easily generate combinations/permutations with great efficiency.

```
## RcppAlgos uses the number of cores available minus one
RcppAlgos::stdThreadMax()
[1] 8
identical(comboGeneral(20, 10, freqs = rep(1:4, 5)),
comboGeneral(20, 10, freqs = rep(1:4, 5), Parallel = TRUE))
[1] TRUE
## Using 7 cores
library(microbenchmark)
microbenchmark(serial = comboGeneral(20, 10, freqs = rep(1:4, 5)),
parallel = comboGeneral(20, 10, freqs = rep(1:4, 5), Parallel = TRUE))
Unit: milliseconds
expr min lq mean median uq max neval
serial 224.77650 239.09760 243.50147 241.41096 244.00927 292.3136 100
parallel 80.21005 82.16222 84.74438 84.42145 85.58884 123.5970 100
```

And applying any of the constraint functions in parallel is highly efficient as well. Consider obtaining the row sums of all combinations:

```
## base R using combn and FUN
combnSum <- combn(20, 10, sum)
algosSum <- comboGeneral(20, 10, constraintFun = "sum")
identical(as.integer(combnSum), algosSum[,11])
[1] TRUE
## Using parallel
paralSum <- comboGeneral(20, 10, constraintFun = "sum", Parallel = TRUE)
identical(paralSum, algosSum)
[1] TRUE
microbenchmark(serial = comboGeneral(20, 10, constraintFun = "sum"),
parallel = comboGeneral(20, 10, constraintFun = "sum", Parallel = TRUE),
combnSum = combn(20, 10, sum))
Unit: milliseconds
expr min lq mean median uq max neval
serial 3.108999 3.470917 4.394282 3.798261 4.345005 8.038390 100
parallel 1.186217 1.377346 1.786400 1.454697 1.588850 3.564296 100
combnSum 197.533911 213.208303 220.711443 217.775698 226.261102 266.828690 100
```

`rowSums`

and `rowMeans`

In fact, finding row sums or row means is even faster than simply applying the highly efficient `rowSums`

/`rowMeans`

after the combinations have already been generated:

```
## Pre-generate combinations
combs <- comboGeneral(25, 10)
## Testing rowSums alone against generating combinations as well as summing
microbenchmark(serial = comboGeneral(25, 10, constraintFun = "sum"),
parallel = comboGeneral(25, 10, constraintFun = "sum", Parallel = TRUE),
rowsums = rowSums(combs))
Unit: milliseconds
expr min lq mean median uq max neval
serial 113.05467 117.46261 127.46829 124.99978 135.67496 176.68228 100
parallel 39.97370 42.79960 49.36836 44.65376 56.82488 67.95255 100
rowsums 97.18933 98.62043 107.64972 104.80272 113.32359 144.27974 100
all.equal(rowSums(combs),
comboGeneral(25, 10,
constraintFun = "sum",
Parallel = TRUE)[,11])
[1] TRUE
## Testing rowMeans alone against generating combinations as well as obtain row means
microbenchmark(serial = comboGeneral(25, 10, constraintFun = "mean"),
parallel = comboGeneral(25, 10, constraintFun = "mean", Parallel = TRUE),
rowmeans = rowMeans(combs))
Unit: milliseconds
expr min lq mean median uq max neval
serial 171.22701 182.2575 198.14229 198.29388 213.14787 240.1016 100
parallel 54.41817 59.2010 73.12819 74.61183 77.93679 119.8481 100
rowmeans 102.34084 103.2962 115.09863 107.12878 120.19177 174.8596 100
all.equal(rowMeans(combs),
comboGeneral(25, 10,
constraintFun = "mean",
Parallel = TRUE)[,11])
[1] TRUE
```

In both cases above, `RcppAlgos`

is doing double the work nearly twice as fast!!!

`lower`

and `upper`

There are arguments `lower`

and `upper`

that can be utilized to generate chunks of combinations/permutations without having to generate all of them followed by subsetting. As the output is in lexicographical order, these arguments specify where to start and stop generating. For example, `comboGeneral(5, 3)`

outputs 10 combinations of the vector `1:5`

chosen 3 at a time. We can set `lower`

to 5 in order to start generation from the 5^{th} lexicographical combination. Similarly, we can set `upper`

to 4 in order only generate the first 4 combinations. We can also use them together to produce only a certain chunk of combinations. For example, setting `lower`

to 4 and `upper`

to 6 only produces the 4^{th}, 5^{th}, and 6^{th} lexicographical combinations. Observe:

```
comboGeneral(5, 3, lower = 4, upper = 6)
[,1] [,2] [,3]
[1,] 1 3 4
[2,] 1 3 5
[3,] 1 4 5
## is equivalent to the following:
comboGeneral(5, 3)[4:6, ]
[,1] [,2] [,3]
[1,] 1 3 4
[2,] 1 3 5
[3,] 1 4 5
```

In addition to being useful by avoiding the unnecessary overhead of generating all combination/permutations followed by subsetting just to see a few specific results, `lower`

and `upper`

can be utilized to generate large number of combinations/permutations in parallel. Observe:

```
## Over 3 billion results
comboCount(35, 15)
[1] 3247943160
## 10086780 evenly divides 3247943160
system.time(lapply(seq(1, 3247943160, 10086780), function(x) {
temp <- comboGeneral(35, 15, lower = x, upper = x + 10086779)
## do something
x
}))
user system elapsed
99.495 49.993 149.467
## Enter parallel
library(parallel)
system.time(mclapply(seq(1, 3247943160, 10086780), function(x) {
temp <- comboGeneral(35, 15, lower = x, upper = x + 10086779)
## do something
x
}, mc.cores = 6))
user system elapsed
125.361 85.916 35.641
```

These arguments are also useful when one needs to explore combinations/permutations of really large vectors:

```
set.seed(222)
myVec <- rnorm(1000)
## HUGE number of combinations
comboCount(myVec, 50, repetition = TRUE)
Big Integer ('bigz') :
[1] 109740941767310814894854141592555528130828577427079559745647393417766593803205094888320
## Let's look at one hundred thousand combinations in the range (1e15 + 1, 1e15 + 1e5)
system.time(b <- comboGeneral(myVec, 50, TRUE,
lower = 1e15 + 1,
upper = 1e15 + 1e5))
user system elapsed
0.008 0.001 0.010
b[1:5, 45:50]
[,1] [,2] [,3] [,4] [,5] [,6]
[1,] 0.5454861 0.4787456 0.7797122 2.004614 -1.257629 -0.7740501
[2,] 0.5454861 0.4787456 0.7797122 2.004614 -1.257629 0.1224679
[3,] 0.5454861 0.4787456 0.7797122 2.004614 -1.257629 -0.2033493
[4,] 0.5454861 0.4787456 0.7797122 2.004614 -1.257629 1.5511027
[5,] 0.5454861 0.4787456 0.7797122 2.004614 -1.257629 1.0792094
```

We can also produce random samples of combinations/permutations with `comboSample`

and `permuteSample`

. This is really useful when we need a reproducible set of random combinations/permutations. Many of the traditional ways of doing this involved relying on heavy use of `sample`

and hoping that we don’t generate duplicate results. Both functions have a similar interface to their respective `General`

functions. Observe:

```
comboSample(10, 8, TRUE, n = 5, seed = 84)
[,1] [,2] [,3] [,4] [,5] [,6] [,7] [,8]
[1,] 3 3 3 6 6 10 10 10
[2,] 1 3 3 4 4 7 9 10
[3,] 3 7 7 7 9 10 10 10
[4,] 3 3 3 9 10 10 10 10
[5,] 1 2 2 3 3 4 4 7
## We can also use sampleVec to generate specific results
## E.g. the below generates the 1st, 5th, 25th, 125th, and
## 625th lexicographical combinations
comboSample(10, 8, TRUE, sampleVec = c(1, 5, 25, 125, 625))
[,1] [,2] [,3] [,4] [,5] [,6] [,7] [,8]
[1,] 1 1 1 1 1 1 1 1
[2,] 1 1 1 1 1 1 1 5
[3,] 1 1 1 1 1 1 3 8
[4,] 1 1 1 1 1 3 6 9
[5,] 1 1 1 1 5 6 10 10
## Is the same as:
comboGeneral(10, 8, TRUE)[5^(0:4), ]
[,1] [,2] [,3] [,4] [,5] [,6] [,7] [,8]
[1,] 1 1 1 1 1 1 1 1
[2,] 1 1 1 1 1 1 1 5
[3,] 1 1 1 1 1 1 3 8
[4,] 1 1 1 1 1 3 6 9
[5,] 1 1 1 1 5 6 10 10
```

Just like the `General`

counterparts (i.e. `combo/permuteGeneral`

), we can easily explore combinations/permutations of large vectors where the total number of results is enormous in parallel (using `Parallel`

or `nThreads`

).

```
## Uses min(stdThreadMax() - 1, 5) thread (in this case)
permuteSample(500, 10, TRUE, n = 5, seed = 123, Parallel = TRUE)
[,1] [,2] [,3] [,4] [,5] [,6] [,7] [,8] [,9] [,10]
[1,] 55 435 274 324 200 152 6 313 121 377
[2,] 196 166 331 154 443 329 155 233 354 442
[3,] 235 325 94 27 370 117 302 86 229 126
[4,] 284 104 464 104 207 127 117 9 390 414
[5,] 456 76 381 456 219 23 376 187 11 123
permuteSample(factor(state.abb), 15, n = 3, seed = 50, nThreads = 3)
[,1] [,2] [,3] [,4] [,5] [,6] [,7] [,8] [,9] [,10] [,11] [,12] [,13] [,14] [,15]
[1,] ME FL DE OK ND CA PA AL ID MO NM HI KY MT NJ
[2,] AZ CA AL CT ME SD ID SC OK NH HI TN ND IA MT
[3,] MD MO NC MT NH AL VA MA VT WV NJ NE MN MS MI
50 Levels: AK AL AR AZ CA CO CT DE FL GA HI IA ID IL IN KS KY LA MA MD ME MI MN ... WY
permuteCount(factor(state.abb), 15)
Big Integer ('bigz') :
[1] 2943352142120754524160000
```

You can also pass user defined functions by utilizing the argument `FUN`

. This feature’s main purpose is for convenience, however it is somewhat more efficient than generating all combinations/permutations and then using a function from the `apply`

family (N.B. the argument `Parallel`

has no effect when `FUN`

is employed).

```
funCustomComb = function(n, r) {
combs = comboGeneral(n, r)
lapply(1:nrow(combs), function(x) cumprod(combs[x,]))
}
identical(funCustomComb(15, 8), comboGeneral(15, 8, FUN = cumprod))
[1] TRUE
library(microbenchmark)
microbenchmark(f1 = funCustomComb(15, 8),
f2 = comboGeneral(15, 8, FUN = cumprod), unit = "relative")
unit: relative
expr min lq mean median uq max neval
f1 6.946481 6.891553 6.334866 6.821221 6.934111 2.686777 100
f2 1.000000 1.000000 1.000000 1.000000 1.000000 1.000000 100
comboGeneral(15, 8, FUN = cumprod, upper = 3)
[[1]]
[1] 1 2 6 24 120 720 5040 40320
[[2]]
[1] 1 2 6 24 120 720 5040 45360
[[3]]
[1] 1 2 6 24 120 720 5040 50400
## Works with the sampling functions as well
permuteSample(5000, 1000, n = 3, seed = 101, FUN = sd)
[[1]]
[1] 1431.949
[[2]]
[1] 1446.859
[[3]]
[1] 1449.272
```

`RcppAlgos`

comes equipped with several functions for quickly generating essential components for problems common in computational mathematics. All functions below can be executed in parallel by using the argument `nThreads`

.

The following sieving functions (`primeFactorizeSieve`

, `divisorsSieve`

, `numDivisorSieve`

, & `eulerPhiSieve`

) are very useful and flexible. Generate components up to a number or between two bounds.

```
## get the number of divisors for every number from 1 to n
numDivisorSieve(20)
[1] 1 2 2 3 2 4 2 4 3 4 2 6 2 4 4 5 2 6 2 6
## If you want the complete factorization from 1 to n, use divisorsList
system.time(allFacs <- divisorsSieve(10^5, namedList = TRUE))
user system elapsed
0.040 0.003 0.043
allFacs[c(4339, 15613, 22080)]
$`4339`
[1] 1 4339
$`15613`
[1] 1 13 1201 15613
$`22080`
[1] 1 2 3 4 5 6 8 10 12 15
[11] 16 20 23 24 30 32 40 46 48 60
[21] 64 69 80 92 96 115 120 138 160 184
[31] 192 230 240 276 320 345 368 460 480 552
[41] 690 736 920 960 1104 1380 1472 1840 2208 2760
[51] 3680 4416 5520 7360 11040 22080
## Between two bounds
primeFactorizeSieve(10^12, 10^12 + 5)
[[1]]
[1] 2 2 2 2 2 2 2 2 2 2 2 2 5 5 5 5 5 5 5 5 5 5 5 5
[[2]]
[1] 73 137 99990001
[[3]]
[1] 2 3 166666666667
[[4]]
[1] 61 14221 1152763
[[5]]
[1] 2 2 17 149 197 501001
[[6]]
[1] 3 5 66666666667
## Creating a named object
eulerPhiSieve(20, namedVector = TRUE)
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
1 1 2 2 4 2 6 4 6 4 10 4 12 6 8 8 16 6 18 8
system.time(a <- eulerPhiSieve(1e12, 1e12 + 1e7))
user system elapsed
1.049 0.041 1.108
## Using nThreads for greater efficiency
system.time(b <- eulerPhiSieve(1e12, 1e12 + 1e7, nThreads = 8))
user system elapsed
3.731 0.013 0.502
identical(a, b)
[1] TRUE
```

There are three very fast vectorized functions for general factoring (e.g. all divisors of number), primality testing, as well as prime factoring (`divisorsRcpp`

, `isPrimeRcpp`

, `primeFactorize`

)

```
## get result for individual numbers
primeFactorize(123456789)
[1] 3 3 3607 3803
## or for an entire vector
set.seed(100)
myVec <- sample(-100000000:100000000, 5)
divisorsRcpp(myVec, namedList = TRUE)
$`-38446778`
[1] -38446778 -19223389 -2 -1 1
[6] 2 19223389 38446778
$`-48465500`
[1] -48465500 -24232750 -12116375 -9693100 -4846550
[6] -2423275 -1938620 -969310 -484655 -387724
[11] -193862 -96931 -500 -250 -125
[16] -100 -50 -25 -20 -10
[21] -5 -4 -2 -1 1
[26] 2 4 5 10 20
[31] 25 50 100 125 250
[36] 500 96931 193862 387724 484655
[41] 969310 1938620 2423275 4846550 9693100
[46] 12116375 24232750 48465500
$`10464487`
[1] 1 11 317 3001 3487 33011
[7] 951317 10464487
$`-88723370`
[1] -88723370 -44361685 -17744674 -8872337 -10
[6] -5 -2 -1 1 2
[11] 5 10 8872337 17744674 44361685
[16] 88723370
$`-6290143`
[1] -6290143 -1 1 6290143
## Creating a named object
isPrimeRcpp(995:1000, namedVector = TRUE)
995 996 997 998 999 1000
FALSE FALSE TRUE FALSE FALSE FALSE
system.time(a <- primeFactorize(1e12:(1e12 + 3e4)))
user system elapsed
1.202 0.002 1.207
## Using nThreads for greater efficiency
system.time(b <- primeFactorize(1e12:(1e12 + 3e4), nThreads = 8))
user system elapsed
1.633 0.001 0.217
identical(a, b)
[1] TRUE
```

Both of these functions are based on the excellent algorithms developed by Kim Walisch. First `primeSieve`

:

```
## Quickly generate large primes over small interval
options(scipen = 50)
system.time(myPs <- primeSieve(10^13+10^3, 10^13))
user system elapsed
0.017 0.006 0.023
myPs
[1] 10000000000037 10000000000051 10000000000099 10000000000129
[5] 10000000000183 10000000000259 10000000000267 10000000000273
[9] 10000000000279 10000000000283 10000000000313 10000000000343
[13] 10000000000391 10000000000411 10000000000433 10000000000453
[17] 10000000000591 10000000000609 10000000000643 10000000000649
[21] 10000000000657 10000000000687 10000000000691 10000000000717
[25] 10000000000729 10000000000751 10000000000759 10000000000777
[29] 10000000000853 10000000000883 10000000000943 10000000000957
[33] 10000000000987 10000000000993
## Object created is small
object.size(myPs)
320 bytes
## primes under a billion!!!
system.time(a <- primeSieve(10^9))
user system elapsed
1.263 0.095 1.368
## Using nThreads
system.time(b <- primeSieve(10^9, nThreads = 8))
user system elapsed
2.339 0.047 0.418
```

Since version `2.3.0`

, we are implementing the cache-friendly improvements for larger primes originally developed by Tomás Oliveira.

```
## Version <= 2.2.0.. i.e. older versions
system.time(old <- RcppAlgos2.2::primeSieve(1e15, 1e15 + 1e9))
user system elapsed
7.615 0.140 7.792
## v2.3.0 is over 3x faster!
system.time(a <- primeSieve(1e15, 1e15 + 1e9))
user system elapsed
2.452 0.208 2.676
## And using nThreads we are ~8x faster
system.time(b <- primeSieve(1e15, 1e15 + 1e9, nThreads = 8))
user system elapsed
5.054 0.754 0.977
identical(a, b)
[1] TRUE
identical(a, old)
[1] TRUE
```

The library by Kim Walisch relies on OpenMP for parallel computation with Legendre’s Formula. Currently, the default compiler on `macOS`

is `clang`

, which does not support `OpenMP`

. James Balamuta (a.k.a. TheCoatlessProfessor… well at least we think so) has written a great article on this topic, which you can find here: https://thecoatlessprofessor.com/programming/openmp-in-r-on-os-x/. One of the goals of `RcppAlgos`

is to be accessible by all users. With this in mind, we set out to count primes in parallel without `OpenMP`

.

At first glance, this seems trivial as we have a function in `Primes.cpp`

called `phiWorker`

that counts the primes up to `x`

. If you look in phi.cpp in the `primecount`

library by Kim Walisch, we see that `OpenMP`

does its magic on a for loop that makes repeated calls to `phi`

(which is what `phiWorker`

is based on). All we need to do is break this loop into *n* intervals where *n* is the number of threads. Simple, right?

We can certainly do this, but what you will find is that *n - 1* threads will complete very quickly and the *n ^{th}* thread will be left with a heavy computation. In order to alleviate this unbalanced load, we take advantage of thread pooling provided by

`RcppThread`

which allows us to reuse threads efficiently as well as breaking up the loop mentioned above into smaller intervals. The idea is to completely calculate `phi`

up to a limit `m`

using all With this is mind, here are some results:

```
## Enumerate the number of primes below trillion
system.time(underOneTrillion <- primeCount(10^12))
user system elapsed
0.478 0.000 0.480
underOneTrillion
[1] 37607912018
## Enumerate the number of primes below a billion in 2 milliseconds
library(microbenchmark)
microbenchmark(primeCount(10^9))
Unit: milliseconds
expr min lq mean median uq max neval
primeCount(10^9) 1.93462 1.938516 2.044785 1.957809 2.090828 3.584245 100
primeCount(10^9)
[1] 50847534
system.time(underOneHundredTrillion <- primeCount(1e14, nThreads = 8))
user system elapsed
49.894 0.102 6.774
underOneHundredTrillion
[1] 3204941750802
## From Kim Walisch's primecount library:
## Josephs-MBP:primecount-4 josephwood$ ./primecount 1e14 --legendre --time
## 3204941750802
## Seconds: 4.441
```

I welcome any and all feedback. If you would like to report a bug, have a question, or have suggestions for possible improvements, please contact me here: jwood000@gmail.com