Stack imbalance with RcppParallel

Question

I wrote the following code to train myself to use RcppParallel. It is just a toy example.

// [[Rcpp::depends(RcppParallel)]]
#include <Rcpp.h>
#include <RcppParallel.h>
#include <iostream>
using namespace Rcpp;
using namespace RcppParallel;


struct Lapin : public Worker {
  // input pars
  const NumericVector input;
  const size_t dim;

  // outputs a matrix
  NumericMatrix output;

  // two constructors
  Lapin(const NumericVector input, const int dim) : input(input), dim(dim), output(NumericMatrix(dim,dim)) {}

  Lapin(const Lapin & jeannot, Split) : input(jeannot.input), dim(jeannot.dim), output(NumericMatrix(dim,dim)) {}

  // the working operator
  void operator()(size_t begin, size_t end) {
    for(size_t k = begin; k < end; k++) {
      for(size_t i = 0; i < dim; i++) {
        for(size_t j = 0; j < dim; j++) {
          output(i,j) += input(k)+i+j;
        }
      }
    }
  }

  // the join
  void join(const Lapin & peter) {
    output += peter.output;
  }
};

// [[Rcpp::export]]
NumericMatrix f(NumericVector A, size_t dim) {
  Lapin groumf(A, dim);
  parallelReduce(0, A.length(), groumf);
  return groumf.output;
}

Here is what happens in R, after sourceCpp-ing it:

> f(rep(1,1100), 5)
     [,1] [,2] [,3] [,4] [,5]
[1,] 1100 2200 3300 4400 5500
[2,] 2200 3300 4400 5500 6600
[3,] 3300 4400 5500 6600 7700
[4,] 4400 5500 6600 7700 8800
[5,] 5500 6600 7700 8800 9900
> sourceCpp("parallel-matrix-reduce.cpp")
> f(rep(1,1100), 5)
Warning: stack imbalance in '.Call', 6 then 11
     [,1] [,2] [,3] [,4] [,5]
[1,] 1100 2200 3300 4400 5500
[2,] 2200 3300 4400 5500 6600
[3,] 3300 4400 5500 6600 7700
[4,] 4400 5500 6600 7700 8800
[5,] 5500 6600 7700 8800 9900

Note that the behavior is eratic: sometimes, I have no warning at all, sometimes it is at the first run... I guess my session info can be useful here:

> sessionInfo()
R version 3.1.2 (2014-10-31)
Platform: x86_64-redhat-linux-gnu (64-bit)

locale:
 [1] LC_CTYPE=en_US.UTF-8       LC_NUMERIC=C              
 [3] LC_TIME=fr_FR.UTF-8        LC_COLLATE=en_US.UTF-8    
 [5] LC_MONETARY=fr_FR.UTF-8    LC_MESSAGES=en_US.UTF-8   
 [7] LC_PAPER=fr_FR.UTF-8       LC_NAME=C                 
 [9] LC_ADDRESS=C               LC_TELEPHONE=C            
[11] LC_MEASUREMENT=fr_FR.UTF-8 LC_IDENTIFICATION=C       

attached base packages:
[1] stats     graphics  grDevices utils     datasets  methods   base     

other attached packages:
[1] Rcpp_0.11.3

loaded via a namespace (and not attached):
[1] RcppParallel_4.3.3 tools_3.1.2       

I thank you all in advance for your answers and comments.

EDIT As Dirk explains below, this is due to the use of R types in the Worker, which confuses the garbage collector. I settled the issue by using Armadillo matrices instead (I was a bit confused by RMatrix). Here is the corrected code:

// [[Rcpp::depends(RcppParallel)]]
// [[Rcpp::depends(RcppArmadillo)]]
#include <RcppArmadillo.h>
#include <RcppParallel.h>
#include <iostream>
using namespace Rcpp;
using namespace RcppParallel;


struct Lapin : public Worker {
  // input pars
  const arma::vec input;
  const size_t dim;

  // outputs a matrix
  arma::mat output;

  // two constructors
  Lapin(const arma::vec input, const int dim) : input(input), dim(dim), output(arma::mat(dim,dim)) {
    output.zeros();
  }

  Lapin(const Lapin & jeannot, Split) : input(jeannot.input), dim(jeannot.dim), output(arma::mat(dim,dim)) {
    output.zeros();
  }

  // the working operator
  void operator()(size_t begin, size_t end) {
    for(size_t k = begin; k < end; k++) { 
      for(size_t i = 0; i < dim; i++) {
        for(size_t j = 0; j < dim; j++) {
          output(i,j) += input(k)+i+j;
        } 
      }
    }
  }
  // the join
  void join(const Lapin & peter) {
    output += peter.output;
  }
};

// [[Rcpp::export]]
arma::mat f(arma::vec & A, size_t dim) {
  Lapin groumf(A, dim);
  parallelReduce(0, A.size(), groumf);
  return groumf.output;
}

Answer 1

Look more closely at an example such as the parallel distance from the Rcpp Gallery.

It does not use NumericMatrix but rather RMatrix<double>. I would do the same here, and have generally advocated not to rely on contact with R types while running parallel segments.