Rcpp: my distance matrix program is slower than the function in package

Question

I would like to calculate the pairwise euclidean distance matrix. I wrote Rcpp programs by the suggestion of Dirk Eddelbuettel as follows

NumericMatrix calcPWD1 (NumericMatrix x){
  int outrows = x.nrow();
  double d;
  NumericMatrix out(outrows,outrows);

  for (int i = 0 ; i < outrows - 1; i++){
    for (int j = i + 1  ; j < outrows ; j ++){
      NumericVector v1= x.row(i);
      NumericVector v2= x.row(j);
      NumericVector v3=v1-v2;
      d = sqrt(sum(pow(v3,2)));
      out(j,i)=d;
      out(i,j)=d;
    }
  }
  return (out) ;
}

But I find my program is slower than dist function.

> benchmark(as.matrix(dist(b)),calcPWD1(b))
                test replications elapsed relative user.self sys.self user.child sys.child
1 as.matrix(dist(b))          100  24.831    1.000    24.679    0.010          0         0
2        calcPWD1(b)          100  27.362    1.102    27.346    0.007          0         0

Do you guys have any suggestion? My matrix is very simple. There is no column names or row names, just plain matrix (for example like b=matrix(c(rnorm(1000*10)),1000,10)). Here is the program of dist

> dist
function (x, method = "euclidean", diag = FALSE, upper = FALSE, 
    p = 2) 
{
    if (!is.na(pmatch(method, "euclidian"))) 
        method <- "euclidean"
    METHODS <- c("euclidean", "maximum", "manhattan", "canberra", 
        "binary", "minkowski")
    method <- pmatch(method, METHODS)
    if (is.na(method)) 
        stop("invalid distance method")
    if (method == -1) 
        stop("ambiguous distance method")
    x <- as.matrix(x)
    N <- nrow(x)
    attrs <- if (method == 6L) 
        list(Size = N, Labels = dimnames(x)[[1L]], Diag = diag, 
            Upper = upper, method = METHODS[method], p = p, call = match.call(), 
            class = "dist")
    else list(Size = N, Labels = dimnames(x)[[1L]], Diag = diag, 
        Upper = upper, method = METHODS[method], call = match.call(), 
        class = "dist")
    .Call(C_Cdist, x, method, attrs, p)
}
<bytecode: 0x56b0d40>
<environment: namespace:stats>

I expect my program is faster than dist since in dist, there are too many thing to need to be checked (like method, diag).

Answer 1

You were almost there. But your inner loop body tried to do too much in one line. Template programming is hard enough as it is, and sometimes it is just better to spread instructions out a little to give the compiler a better chance. So I just made it five statements, and built immediatelt.

New code:

#include <Rcpp.h>

using namespace Rcpp;

double dist1 (NumericVector x, NumericVector y){
  int n = y.length();
  double total = 0;
  for (int i = 0; i < n ; ++i) {
    total += pow(x(i)-y(i),2.0);
  }
  total = sqrt(total);
  return total;
}

// [[Rcpp::export]]
NumericMatrix calcPWD (NumericMatrix x){
  int outrows = x.nrow();
  int outcols = x.nrow();
  NumericMatrix out(outrows,outcols);

  for (int i = 0 ; i < outrows - 1; i++){
    for (int j = i + 1  ; j < outcols ; j ++) {
      NumericVector v1 = x.row(i);
      NumericVector v2 = x.row(j-1);
      double d = dist1(v1, v2);
      out(j-1,i) = d;
      out(i,j-1)= d;
    }
  }
  return (out) ;
}

/*** R
M <- matrix(log(1:9), 3, 3)
calcPWD(M)
*/

Running it:

R> sourceCpp("/tmp/mikebrown.cpp")

R> M <- matrix(log(1:9), 3, 3)

R> calcPWD(M)
         [,1]     [,2] [,3]
[1,] 0.000000 0.740322    0
[2,] 0.740322 0.000000    0
[3,] 0.000000 0.000000    0
R> 

You may want to check your indexing logic though. Looks like you missed more comparisons.

Edit: For kicks, here is a more compact version of your distance function:

// [[Rcpp::export]]
double dist2(NumericVector x, NumericVector y){
  double d = sqrt( sum( pow(x - y, 2) ) );
  return d;
}

Answer 2

Rcpp vs. Internal R Functions (C/Fortran)

First of all, just because you are writing the algorithm using Rcpp does not necessarily mean it will beat out the R equivalent, especially if the R function calls a C or Fortran routine to perform the bulk of the computations. In other cases where the function is written purely in R, there is a high probability that transforming it in Rcpp will yield the desired speed gain.

Remember, when rewriting internal functions, one is going up against the R Core team of absolutely insane C programmers most likely will win out.

Base Implementation of dist()

Secondly, the distance calculation R uses is done in C as indicated by:

.Call(C_Cdist, x, method, attrs, p)

, which is the last line of the dist() function's R source. This gives it a slight advantage vs. C++ as it more granular instead of templated.

Furthermore, the C implementation uses OpenMP when available to parallelize the computation.

Proposed modification

Thirdly, by changing the subset order slightly and avoiding creating an additional variable, the timings between versions decrease.

#include <Rcpp.h>

// [[Rcpp::export]]
Rcpp::NumericMatrix calcPWD1 (const Rcpp::NumericMatrix & x){
  unsigned int outrows = x.nrow(), i = 0, j = 0;
  double d;
  Rcpp::NumericMatrix out(outrows,outrows);

  for (i = 0; i < outrows - 1; i++){
    Rcpp::NumericVector v1 = x.row(i);
    for (j = i + 1; j < outrows ; j ++){
      d = sqrt(sum(pow(v1-x.row(j), 2.0)));
      out(j,i)=d;
      out(i,j)=d;
    }
  }

  return out;
}