Rcpp rowMaxs vs. matrixStats rowMaxs

Question

I am trying to efficiently compute rowMaxs in Rcpp. A very simple implementation is

arma::mat RcppRowmaxs(arma::mat x){  

  int N = x.n_rows;
  arma::mat rm(N,1);

  for(int nn = 0; nn < N; nn++){
      rm(nn) = max(x.row(nn));
  }

  return(rm);
}

which works perfectly fine. However, comparing this function to other packages, it turned out that other implementations are by far more efficient. Specifically, Rfast::rowMaxs is more than 6 times faster than the simple Rcpp implementation!

Naturally, I tried to mimic the behavior of Rfast. However, as a beginner in Rcpp, I only tried to load Rfast::rowMaxs directly in Rcpp as described e.g. here. Unfortunately, using an Rcpp script to load an R function that again calls an Rcpp script seems pretty slow following my benchmark (see row "RfastinRcpp"):

m = matrix(rnorm(1000*1000),1000,1000)

microbenchmark::microbenchmark(

  matrixStats    = matrixStats::rowMaxs(m),
  Rfast          = Rfast::rowMaxs(m,value=T),
  Rcpp           = RcppRowmaxs(m),
  RfastinRcpp    = RfastRcpp(m),
  apply          = apply(m,1,max)

)

Unit: microseconds
        expr       min         lq       mean     median        uq        max neval cld
 matrixStats  1929.570  2042.8975  2232.1980  2086.5180  2175.470   4025.923   100 a  
       Rfast   666.711   727.2245   842.5578   795.2215   891.443   1477.969   100 a  
        Rcpp  5552.216  5825.4855  6186.9850  5997.8295  6373.737   8568.878   100  b 
 RfastinRcpp  7495.042  7931.2480  9471.8453  8382.6350 10659.672  19968.817   100  b 
       apply 12281.758 15145.7495 22015.2798 17202.9730 20310.939 136844.591   100   c

Any tips on how to improve performance in the function I've provided above? I've looked at the source code from Rfast and believe that this is the correct file. However, so far I did not manage to locate the important parts of the code.

Edit: Changed the post to focus on Rfast now, following the answer of Michail.

Answer 1

I just did some experiments on my laptop. I have an 5 year old HP with 2 intel i5 cores at 2.3 GHz. Attached is an picture with my results. Rfast's implementation is way faster than matrixStats' implementation, always and as the matrix gets bigger the time difference increases.