The attached program is a straightforward implementation of matrix inversion
via the Gauss-Jordan algorithm. The explicitly parallelized version (not
attached) produces exactly the same result for each run independent of the
number of threads. The version parallelized via OpenMP produces incorrect
results for sufficiently large matrices and two or more threads. An example:
$ OMP_NUM_THREADS=1 ./matinv_openmp 180
error = 4.75689e-14; epsilon = 2.22045e-16; error / (epsilon * n) = 1.19018
Error within bounds.
$ OMP_NUM_THREADS=2 ./matinv_openmp 180
error = nan; epsilon = 2.22045e-16; error / (epsilon * n) = nan
Error out of bounds.
I compiled this program as follows:
$ ~/gcc-4.2.3/bin/gcc -Wall -fopenmp -g matinv_openmp.c -static -o
matinv_openmp -lm
$ ~/gcc-4.2.3/bin/gcc -v
Using built-in specs.
Target: x86_64-unknown-linux-gnu
Configured with: /home/bart/software/gcc-4.2.3/configure --disable-nls
--enable-threads=posix --enable-tls --prefix=/home/bart/gcc-4.2.3
Thread model: posix
gcc version 4.2.3
--
Summary: OpenMP: Incorrect result when run with two or more
threads
Product: gcc
Version: 4.2.3
Status: UNCONFIRMED
Severity: major
Priority: P3
Component: libgomp
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: bart dot vanassche at gmail dot com
GCC build triplet: x86_64-unknown-linux-gnu
GCC host triplet: x86_64-unknown-linux-gnu
GCC target triplet: x86_64-unknown-linux-gnu
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35517
