CODEWITHSUNDEEP
c++rmultithreadingrcpp
emdroll
emdroll

Reputation: 73

Is Rcpp vector thread safe to use with OpenMP?

I'm using Rcpp with OpenMP for parallel computation in C++. I read Is accessing/setting elements in a Rcpp::List thread safe? and it seems that using Rcpp inside a threaded code block is dangerous. But it is quite common to read or write on Rcpp NumericVectors in a for loop such as https://gallery.rcpp.org/articles/hierarchical-risk-parity/.

Below is what I wrote. The task is straightforward. I first declare a NumericVector with a large size n outside a (parallel) for loop. I do some calculation and fill each element of the NumericVector using index i. Once done, I compute a desired quantile of it.

#include <Rcpp.h>
#include <omp.h>

// [[Rcpp::export]]
double foo(const int n, const int nthreads) {

  Rcpp::NumericVector out(n);
  #pragma omp parallel for num_threads(nthreads)
  for (int i = 0; i < n; ++i) {
    out[i] = doSomething();
  }

  Rcpp::Environment stats("package:stats");
  Rcpp::Function quantile = stats["quantile"];
  return Rcpp::as<double>(quantile(out, Rcpp::Named("probs") = 0.95));
}

I have two questions.

  1. Is there any reason to use other containers (e.g. std::vector<double>) instead of NumericVector for thread safety? I tested the above code with std::vector<double> and got identical result. I heard that conversion between the two involves deep copy. If I use std::vector<double>, then does deep copying occur when quntile() is called?

  2. If using NumericVector is safe, would it be better to just return NumericVector and apply quntile() in R in terms of performance?

Upvotes: 5

Views: 400

Answers (0)

Related Questions