RcppParallel Parallelizing distance computation: segfault - c++

Question

I have a matrix, for which I want to compute the distance (let's say Euclidean) between the ith row and every other row(i.e. I want the ith row of the pairwise distance matrix).
#include <Rcpp.h>
#include <cmath>
#include <algorithm>
#include <RcppParallel.h>
//#include <RcppArmadillo.h>
#include <queue>
using namespace std;
using namespace Rcpp;
using namespace RcppParallel;
// [[Rcpp::export]]
double dist_fun(NumericVector row1, NumericVector row2){
double rval = 0;
for (int i = 0; i < row1.length(); i++){
rval += (row1[i] - row2[i]) * (row1[i] - row2[i]);
}
return rval;
}
// [[Rcpp::export]]
NumericVector dist_row(NumericMatrix mat, int i){
NumericVector row(mat.nrow());
NumericMatrix::Row row1 = mat.row(i - 1);
for (int j = 0; j < mat.nrow(); j++){
NumericMatrix::Row row2 = mat.row(j);
row(j) = dist_fun(row1, row2);
}
return row;
}
// [[Rcpp::depends(RcppParallel)]]
struct JsDistance: public Worker {
// input matrix to read from
const NumericMatrix mat;
int i;
// output vector to write to
NumericVector output;
// initialize from Rcpp input and output matrixes (the RMatrix class
// can be automatically converted to from the Rcpp matrix type)
JsDistance(const NumericMatrix mat, int i, NumericVector output)
: mat(mat), i(i), output(output) {}
// function call operator that work for the specified range (begin/end)
void operator()(std::size_t begin, std::size_t end) {
NumericVector row1 = mat.row(i);
for (std::size_t j = begin; j < end; j++) {
NumericVector row2 = mat.row(j);
output[j] = dist_fun(row1, row2);
}
}
};
// [[Rcpp::export]]
NumericVector parallel_dist_row(NumericMatrix mat, int i) {
// allocate the matrix we will return
NumericVector output(mat.nrow());
// create the worker
JsDistance JsDistance(mat, i, output);
// call it with parallelFor
parallelFor(0, mat.nrow(), JsDistance);
return output;
}
The sequential way using Rcpp is the function 'row_dist' as written above. Yet the matrix I want to work with is very large so I want to parallelize it. But then I will run into a segfault error which I don't quite understand why. To trigger the error you can run the following code:
library(Rcpp)
library(RcppParallel)
setThreadOptions(numThreads = 20)
set.seed(42)
X = matrix(rnorm(10000 * 400), 10000, 400)
sourceCpp("question.cpp")
start1 = proc.time()
print(dist_row(X, 2)[1:30])
print(proc.time() - start1)
start2 = proc.time()
print(parallel_dist_row(X, 2)[1:30])
print(proc.time() - start2)
Can someone give me some hint about what I did wrong? Thanks in advance for your time!
=======================================================================
Edit:
inline double d(double a, double b){
return fabs(a - b);
}
// [[Rcpp::depends(RcppParallel)]
struct dtwDistance: public Worker {
// Input matrix to read from must be of the RMatrix<T> form
// if using Rcpp objects
const RMatrix<double> mat;
int i;
// Output vector to write to must be of the RVector<T> form
// if using Rcpp objects
RVector<double> output;
// initialize from Rcpp input and output matrixes (the RMatrix class
// can be automatically converted to from the Rcpp matrix type)
dtwDistance(const NumericMatrix mat, int i, NumericVector output)
: mat(mat), i(i - 1), output(output) {}
// Note the -1 ^^^^ to match results from prior function
// Function call operator to iterate over a specified range (begin/end)
void operator()(std::size_t begin, std::size_t end) {
RMatrix<double>::Row row1 = mat.row(i);
for (std::size_t j = begin; j < end; ++j) {
RMatrix<double>::Row row2 = mat.row(j);
size_t n = row1.length();
size_t m = row2.length();
NumericMatrix cost(n + 1, m + 1);
for (int ii = 1; ii <= n; ii++){
cost(i, 0) = numeric_limits<double>::infinity();
}
for (int jj = 1; jj <= m; jj++){
cost(0, j) = numeric_limits<double>::infinity();
}
for (int ii = 1; ii <= n; ii++){
for (int jj = 1; jj <= m; jj++){
double dist = d(row1[ii - 1], row2[jj - 1]);
cost(ii, jj) = dist + min(min(cost(ii - 1, jj), cost(ii, jj - 1)), cost(ii - 1, jj - 1));
//cout << ii << ", " << jj << ", " << cost(ii, jj) << "\n";
}
}
output[j] = cost(n, m);
}
}
};
// [[Rcpp::export]]
NumericVector parallel_dist_row_dtw(NumericMatrix mat, int i) {
// allocate the matrix we will return
//RMatrix<double> input(mat);
NumericVector y(mat.nrow());
//RVector<double> output(y);
// create the worker
dtwDistance dtwDistance(mat, i, y);
// call it with parallelFor
parallelFor(0, mat.nrow(), dtwDistance);
return y;
}
The distance I needed to calculate is the dynamic time warping distance. I implemented it as above. Yet when running, it will give a 'stack imbalance' warning. And there will be a segfault after several runs. I'm wondering what is the problem now.
To trigger the problem, I did:
library(Rcpp)
library(RcppParallel)
setThreadOptions(numThreads = 4)
sourceCpp("scripts/chisq_dtw.cpp")
set.seed(42)
X = matrix(rnorm(1000), 100, 10)
parallel_dist_row_dtw(X, 1)
parallel_dist_row_dtw(X, 2)
parallel_dist_row_dtw(X, 3)
parallel_dist_row_dtw(X, 4)
parallel_dist_row_dtw(X, 5)

Answer 1

The issue is you are not using the thread-safe wrapper around R objects via RMatrix<T> and RVector<T>. These classes are important because of the parallelization being executed on a background thread, which is an area that is not safe to call R or Rcpp APIs. The official documentation emphasizes this in the Safe Accessors section.
In particular, we have:
To provide safe and convenient access to the arrays underlying R vectors and matrices RcppParallel introduces several accessor classes:
RVector<T> — Wrap R vectors of various types
RMatrix<T> — Wrap R matrices of various types (also includes Row and Column classes)
To create a thread safe accessor for an Rcpp vector or matrix just construct an instance of RVector or RMatrix with it.
Code Fix
So, your work can be fixed by switching *Matrix to RMatrix<T> and *Vector to RVector<T>.
struct JsDistance: public Worker {
// Input matrix to read from must be of the RMatrix<T> form
// if using Rcpp objects
const RMatrix<double> mat;
int i;
// Output vector to write to must be of the RVector<T> form
// if using Rcpp objects
RVector<double> output;
// initialize from Rcpp input and output matrixes (the RMatrix class
// can be automatically converted to from the Rcpp matrix type)
JsDistance(const NumericMatrix mat, int i, NumericVector output)
: mat(mat), i(i - 1), output(output) {}
// Note the -1 ^^^^ to match results from prior function
// Function call operator to iterate over a specified range (begin/end)
void operator()(std::size_t begin, std::size_t end) {
RMatrix<double>::Row row1 = mat.row(i);
for (std::size_t j = begin; j < end; ++j) {
RMatrix<double>::Row row2 = mat.row(j);
double rval = 0;
for (unsigned int k = 0; k < row1.length(); ++k) {
rval += (row1[k] - row2[k]) * (row1[k] - row2[k]);
}
output[j] = rval;
}
}
};
In particular, the data types used here are of the form RMatrix<double> even for accessing the matrix.
Also, within the parallelized version there is a missing i-1 statement. To remedy this, I've opted to have it taken care of in the constructor of JSDistance.
Test
set.seed(42)
X = matrix(rnorm(10000 * 400), 10000, 400)
start1 = proc.time()
print(dist_row(X, 2)[1:30])
# [1] 811.8873 0.0000 799.8153 810.1442 720.3232 730.6083 797.8441 781.8066 827.1511 834.1863 842.9392 850.2476 724.5842 673.1428 775.0994
# [16] 805.5752 804.9281 774.9770 799.7669 870.3187 815.1129 934.7581 726.1554 804.2097 758.4943 772.8931 806.6026 715.8257 847.8980 831.7555
print(proc.time() - start1)
# user system elapsed
# 0.22 0.00 0.23
start2 = proc.time()
print(parallel_dist_row(X, 2)[1:30])
# [1] 811.8873 0.0000 799.8153 810.1442 720.3232 730.6083 797.8441 781.8066 827.1511 834.1863 842.9392 850.2476 724.5842 673.1428 775.0994
# [16] 805.5752 804.9281 774.9770 799.7669 870.3187 815.1129 934.7581 726.1554 804.2097 758.4943 772.8931 806.6026 715.8257 847.8980 831.7555
print(proc.time() - start2)
# user system elapsed
# 0.28 0.00 0.06
all.equal(parallel_dist_row(X, 2), dist_row(X, 2))
# [1] TRUE