Open kostrzewa opened 2 years ago
For reasons of simplicity it makes sense to run the inversions and get back the solution. This avoids having to duplicate all the logic that we have in place for differentiating between the different types of monomials in QUDA.
For reasons of performance it makes sense to work entirely in QUDA from that point onwards, although there are some complications related to the different types of monomials.
This is the major performance bottleneck left for a reasonably efficient (>70% GPU utilisation) HMC
computeCloverForceQuda
computeCloverForceQuda
computeCloverForceQuda