Open mark-hobbs opened 2 years ago
Sounds like a good feature - so will the Numba part will be for calculating the force vector, rather than doing the ODE (time integration)?
For the non-OpenCL ODE solve, could we use an existing integrator for the time integration parts? e.g., from scipy.integrate
. Sometimes the advice given is "never write your own ODE solver".
The Cython implementation of the Euler integrator is 30x slower than equivalent vectorised Matlab code. It is also extremely limited and only allows for a linear damage model with no damping. It is important that there is a reasonably fast solver that does not require OpenCL. I propose to use Numba (jit compiler) instead of Cython. Initial testing shows that it is faster and does not require C like syntax that some users will be unfamiliar with.