I am using OpenMP for some basic multiprocessing computations and a very simple code is producing data races when compiled with sanitizers but seems to work correctly with any number of threads once the sanitizer is removed. I reduce it to a MWE:
#include <omp.h>
#include <iostream>
#include <cmath>
#include <cstdlib>
void fill(double * array, int size);
double suma(const double * array, int size);
int main(int argc, char *argv[])
{
const int N = std::atoi(argv[1]);
double * data = new double [N];
fill(data, N);
std::cout << data[0] << "\n";
double total = suma(data, N);
std::cout << total/N << "\n";
delete [] data;
return 0;
}
void fill(double * array, int size)
{
#pragma omp parallel for
for(int ii = 0; ii < size; ii++) {
array[ii] = 2*ii*std::sin(std::sqrt(ii/56.7)) +
std::cos(std::pow(1.2*ii, 0.3));
}
}
double suma(const double * array, int size)
{
double val = 0.0;
#pragma omp parallel for reduction(+:val)
for(int ii = 0; ii < size; ii++) {
val += array[ii];
}
return val;
}
I compiled it as
g++ -g -fopenmp -fsanitize=thread mwe.cpp
with g++ version 11.2.0, and then execute it as
OMP_NUM_THREADS=2 ./a.out 10000000 1
and then I get sanitizers errors like:
==================
WARNING: ThreadSanitizer: data race (pid=26332)
Write of size 4 at 0x7ffd2ecd3e90 by main thread:
#0 suma(double const*, int) mwe.cpp:37 (a.out+0x4014ac)
#1 main mwe.cpp:17 (a.out+0x401380)
Previous read of size 8 at 0x7ffd2ecd3e90 by thread T1:
[failed to restore the stack]
Location is stack of main thread.
Location is global '<null>' at 0x000000000000 ([stack]+0x00000001ee90)
Thread T1 (tid=26334, running) created by main thread at:
#0 pthread_create <null> (libtsan.so.0+0x5fdc5)
#1 <null> <null> (libgomp.so.1+0x1d99b)
#2 main mwe.cpp:14 (a.out+0x401340)
SUMMARY: ThreadSanitizer: data race mwe.cpp:37 in suma(double const*, int)
==================
==================
WARNING: ThreadSanitizer: data race (pid=26332)
Read of size 8 at 0x7ffd2ecd3e80 by thread T1:
#0 suma(double const*, int) [clone ._omp_fn.0] mwe.cpp:37 (a.out+0x401766)
#1 <null> <null> (libgomp.so.1+0x1d385)
Previous write of size 8 at 0x7ffd2ecd3e80 by main thread:
#0 suma(double const*, int) mwe.cpp:37 (a.out+0x4014be)
#1 main mwe.cpp:17 (a.out+0x401380)
Location is stack of main thread.
Location is global '<null>' at 0x000000000000 ([stack]+0x00000001ee80)
Thread T1 (tid=26334, running) created by main thread at:
#0 pthread_create <null> (libtsan.so.0+0x5fdc5)
#1 <null> <null> (libgomp.so.1+0x1d99b)
#2 main mwe.cpp:14 (a.out+0x401340)
SUMMARY: ThreadSanitizer: data race mwe.cpp:37 in suma(double const*, int) [clone ._omp_fn.0]
==================
I cannot see the source of this data race to fix it. Actually, I was using std::vector first, with the same result. What could be wrong with the code? I am pretty sure that there could be alternative ways to implement what I want to do, but I want to learn how to fix the current one. Or is this a false positive? Thanks in advance.
Edit: title with correct sanitizer option