Mark Dewing

Once #4131 goes in, I have a plan to keep the parameters as a total rotation from the initial coefficient matrix. The current rotated MO coefficients could be recomputed at...

I think using a generalized nonsymmetric solver is slower. There is code in QMCFixedSamplingLinearOptimize{Batched} to do so - the three parameter version of getLowestEigenvector uses dggev instead. Plus there's even...

Scalapack complications: To get the eigenvalues, the following routines get called * `dgehrd` - reduce a general matrix to upper Hessenberg form (Hessenberg form is upper triangular + one sub-diagonal)...

The LMY engine code does not work with the batched driver. At the very least, ``QMCCostFunctionBatched::engine_checkConfigurations`` is not implemented. If the solver in those papers is the SPAM (Subspace Projected...

See also #2509

It looks like there are three issues being addressed in the PR: 1. There are timers in the batched version that are not present in the legacy version (QMCCostFunction) 2....

The fillOverlapHamiltonianMatrices timer should be moved to QMCCostFunction to match the batched version. That would fix the one case where the same timer is implemented differently between legacy and batched...

Test this please

The issue is ConstantSizeMatrix can only be resized along dimensions independently. Even if there is enough underlying storage to accommodate the resized matrix, it still fails. One solution is to...

Updated with some fixes so it works with splines SPO's and single determinants. Replaces the in-place multiplication of the rotation matrix and the coefficient matrix. Note on the gradient test:...