C code with openmp called from R gives inconsistent results - r

Question

Below is a piece of C code run from R used to compare each row of a matrix to a vector. The number of identical values is stored in the first column of a two-column matrix.
I know it can easily be done in R (as done to check the results), but this is a first step for a more complex use case.
When openmp is not used, it works ok. When openmp is used, it give correlated (0.99) but inconsistent results.
Question1: What am I doing wrong?
Question2: I use a double for loop to fill the output matrix (ret) with zeros. What would be a better solution?
Also, inconsistencies were observed when the code was used in a package. I tried to make the code reproducible using inline, but it does not recognize the openmp statements (I tried to include 'omp.h', in the parameters of cfunction, ...).
Question3: How can we make this code work with inline?
I'm (too?) far outside my comfort zone on this topic.
library(inline)
compare <- cfunction(c(x = "integer", vec = "integer"), "
const int I = nrows(x), J = ncols(x);
SEXP ret;
PROTECT(ret = allocMatrix(INTSXP, I, 2));
int *ptx = INTEGER(x), *ptvec = INTEGER(vec), *ptret = INTEGER(ret);
for (int i=0; i<I; i++)
for (int j=0; j<2; j++)
ptret[j * I + i] = 0;
int i, j;
#pragma omp parallel for default(none) shared(ptx, ptvec, ptret) private(i,j)
for (j=0; j<J; j++)
for (i=0; i<I; i++)
if (ptx[i + I * j] == ptvec[j]) {++ptret[i];}
UNPROTECT(1);
return ret;
")
N = 3e3
M = 1e4
m = matrix(sample(c(-1:1), N*M, replace = TRUE), nc = M)
v = sample(-1:1, M, replace = TRUE)
cc = compare(m, v)
cr = rowSums(t(t(m) == v))
all.equal(cc[,1], cr)

Answer 1

Thanks to the comments above, I reconsidered the data race issue.
IIUC, my loop was parallelized on j (the columns). Then, each thread had its own value of i (the rows), but possible identical values across threads, that were then trying to increment ptret[i] at the same time.
To avoid this, I now loop on i first, so that only a single thread will increment each row.
Then, I realized that I could move the zero-initialization of ptret within the first loop.
It seems to work. I get identical results, increased CPU usage, and 3-4x speedup on my laptop.
I guess that solves questions 1 and 2. I will have a closer look at the inline/openmp problem.
Code below, fwiw.
#include <omp.h>
#include <R.h>
#include <Rinternals.h>
#include <stdio.h>
SEXP c_compare(SEXP x, SEXP vec)
{
const int I = nrows(x), J = ncols(x);
SEXP ret;
PROTECT(ret = allocMatrix(INTSXP, I, 2));
int *ptx = INTEGER(x), *ptvec = INTEGER(vec), *ptret = INTEGER(ret);
int i, j;
#pragma omp parallel for default(none) shared(ptx, ptvec, ptret) private(i, j)
for (i = 0; i < I; i++) {
// init ptret to zero
ptret[i] = 0;
ptret[I + i] = 0;
for (j = 0; j < J; j++)
if (ptx[i + I * j] == ptvec[j]) {
++ptret[i];
}
}
UNPROTECT(1);
return ret;
}