This PR adds a slower complex kernel (approx. 2x slower, but not always and this will be slightly improved once I remove determinant checks) which correctly uses per-correlation weights. This solver is also much closer to the maths as written and can be used a basis for parameterized solvers.
Other changes include fixes to net gain computation (note that this needs further checks in the distributed case as coerce_literal can confuse the distributed scheduler).
This PR adds a slower complex kernel (approx. 2x slower, but not always and this will be slightly improved once I remove determinant checks) which correctly uses per-correlation weights. This solver is also much closer to the maths as written and can be used a basis for parameterized solvers.
Other changes include fixes to net gain computation (note that this needs further checks in the distributed case as
coerce_literalcan confuse the distributed scheduler).Correlation mode now just uses an integer value.