Not getting all() quicker using Rcpp

107 Views Asked by At

As I'm a bit new to Rcpp, I might be missing a trick here.

Let's create two matrices:

library(Rcpp)
library(microbenchmark)

P <- matrix(0, 200,500)
for(i in 1:500) P[,i] <- rep(rep(sample(0:1), 2), 25)
Parent_Check <- matrix(0, nrow(P), nrow(P))

I now want the following done:

Test1 <- function(){
  for (i in 1:nrow(P)) {
    Parent_Check[i,] <- apply(P, 1, function(x) all(x == P[i,]))
  }

}
Test1()

I then created a Rcpp version for all() hoping to improve speed, defined as:

Rcpp::cppFunction(

'bool all_C(LogicalVector x) {
  // Note the use of is_true to return a bool type.
  return is_true(all(x == TRUE));
 }
'  
)

Checking the speeds using all_C, it proves to be slower:

Test2 <- function(){
  for (i in 1:nrow(P)) {
    Parent_Check[i,] <- apply(P, 1, function(x) all_C(x == P[i,]))
  }
  Parent_Check
}

microbenchmark::microbenchmark(Test1(), Test2(), times = 10)
    expr      min       lq     mean   median       uq      max neval
 Test1() 467.9671 471.1590 488.1784 479.4830 485.4755 578.5338    10
 Test2() 544.6561 552.7025 587.8888 570.4416 641.1202 657.7581    10

Trouble is, all_C() is slower than all(), so I suspect the slow speed for Test2() requires a better all_C call as well as a way of avoid apply in the above example.

I tried rewriting apply in Rcpp using this answer, but using this Rcpp apply function makes it even slower.

Any ideas on how to improve the speed of Test1() using Rcpp?

1

There are 1 best solutions below

1
On BEST ANSWER

As has been mentioned in the comments, trying to get a faster all() is not likely to help here. Rather, you would want to move the loops into C++. Doing that will also give you more control: for example, you can avoid always comparing all elements of the rows and instead short-circuit on the first element that isn’t equal.

Here’s my stab at what a faster solution could look like:

Rcpp::cppFunction('
// For all rows, check if it is equal to all other rows
LogicalMatrix f2(const NumericMatrix& x) {
  size_t n = x.rows();
  size_t p = x.cols();
  LogicalMatrix result(n, n);

  for (size_t i = 0; i < n; i++) {
    for (size_t j = 0; j < i; j++) {
      bool rows_equal = true;

      for (size_t k = 0; k < p; k++) {
        if (x(i, k) != x(j, k)) {
          rows_equal = false;
          break;
        }
      }

      result(i, j) = rows_equal;
      result(j, i) = rows_equal;
    }
    result(i, i) = true;
  }

  return result;
}
')

Original implementation:

set.seed(4)

P <- matrix(0, 200,500)
for(i in 1:500) P[,i] <- rep(rep(sample(0:1), 2), 25)

f1 <- function(P) {
  Parent_Check <- matrix(0, nrow(P), nrow(P))
  for (i in 1:nrow(P)) {
    Parent_Check[i,] <- apply(P, 1, function(x) all(x == P[i,]))
  }
  Parent_Check
}

And results:

bench::mark(f1(P), f2(P) * 1)
#> Warning: Some expressions had a GC in every iteration; so filtering is
#> disabled.
#> # A tibble: 2 x 6
#>   expression      min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr> <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 f1(P)      736.18ms 736.18ms      1.36     697MB    27.2 
#> 2 f2(P) * 1    6.37ms   6.95ms    134.       471KB     1.96