With gfortran11, this behavior is more severe than with gfortran12, but both optimized executables produce results that diverge from the debug executable. With gfortran11 it takes 253 timesteps instead of 165 to reach the configured simulation time and the resulting density is significantly different from the result with debug. With gfortran12 the changes are more subtle and the optimized executable still gets done in 165 time steps, but the deviation is notably large, though still below 0.1% for the density in the end of the simulation.
This is a simulation with L2P and Chebyshev-Lobatto nodes, oversampling by a factor of 3, positivity preserving stabilization, and SSP RK2 time integration. L2P is also used in other cases, SSP RK2 is also used in euler/2D/toro2_x.
That leaves the Chebyshev-Lobatto nodes and the positivity preserving filter as first candidates for further investigation.
With gfortran11, this behavior is more severe than with gfortran12, but both optimized executables produce results that diverge from the debug executable. With gfortran11 it takes 253 timesteps instead of 165 to reach the configured simulation time and the resulting density is significantly different from the result with debug. With gfortran12 the changes are more subtle and the optimized executable still gets done in 165 time steps, but the deviation is notably large, though still below 0.1% for the density in the end of the simulation.
This is a simulation with L2P and Chebyshev-Lobatto nodes, oversampling by a factor of 3, positivity preserving stabilization, and SSP RK2 time integration. L2P is also used in other cases, SSP RK2 is also used in
euler/2D/toro2_x.That leaves the Chebyshev-Lobatto nodes and the positivity preserving filter as first candidates for further investigation.