GitLab

Run Coulomb and dispersion reciprocal PME work on separate HIP queues for
LJPME when PME streams are enabled.  Use separate grids, sorters, events, and
energy buffers so the two reciprocal branches can overlap safely.

Keep the behavior HIP-only based on RTX4090 CUDA profiling, where the same
split increased PME spread/list contention and regressed apoa1ljpme.

This reverts commit 4e7070c2.

The split-pme-stream optimization to be implemented will put
LJ-PME reciprocal kernels on a non-critical path.

* More barostats support scaling particles independently

* Documentation update

* MonteCarloMembraneBarostat can scale particles independently

Use explicit 128-thread launches for the three LangevinMiddle integration kernels to improve HIP throughput while
preserving the existing PME launch heuristics.

Use explicit 128-thread block launches for selected HIP PME kernels that
benefit from larger blocks.  Keep the platform default block size unchanged,
and leave small-system grid indexing and charge spreading on the existing
default launch configuration.

The heuristic applies 128-thread launches to finishSpreadCharge on HIP, and
uses 128-thread launches for findAtomGridIndex and gridSpreadCharge only for
larger systems.  Coulomb PME and LJPME dispersion paths are handled in
parallel, while interpolation and energy evaluation remain unchanged.

Apply heuristics for HIP neighbor-list construction:
use fewer nonbonded force blocks for small neighbor-list systems, use two
tiles per batch for larger atom-block counts, and increase the
findBlocksWithInteractions thread block size for small atom-block counts.

Standard concurrent validation shows no clear per-case regression and a
small geomean throughput improvement over the current blocksPerCU baseline.

* Try to get docs build working

* Update changed link

* Try switching a URL to https

- Use bitwise prefix accounting when storing sparse interactions as single pairs in the HIP pair-list kernel. This reduces the number of ballot operations needed to compute per-lane single-pair offsets.
- For HIP CDNA single precision, raise MAX_BITS_FOR_PAIRS to 8 so more sparse interactions are emitted as single pairs instead of full tiles. Keep the existing double precision and RDNA thresholds unchanged.
- Also simplify the HIP LJPME direct correction by computing alpha^2*r2

* Created ExpandedEnsembleSampler

* Attempt at fixing test errors on Windows

* Another attempt at fixing test errors on Windows

* More output options

* Minor fixes

* Still trying to fix Windows errors

* Debugging

* Just skip the test on Windows

* Fix error on older Python

* Context deselection before energy evaluation

* Check that the correct context is popped by popAsCurrent()

* Avoid host wait when synchronizing PME queue

* Remove comment

* Created ReplicaExchangeReporter

* Documentation for ReplicaExchangeReporter

Avoid forcing Coulomb PME to re-sort whenever LJ-PME is enabled, and give dispersion PME its own atom-grid index and sort state so the performance impact can be measured independently.

* Created ReplicaExchangeSampler

* Improvements to ReplicaExchangeSampler

* PythonForce can be restricted to a subset of particles

* Fix exception with CUDA

* Infrastructure for multistate sampling

* Added computeRelativeEnergies()

* optimization, switch to unordered map

* Remove emplace, reduce reserve allocation

* Cache coefficients for long range correction

* updateParametersInContext() clears cache

* fix 5234 and add test

* clean up docstring and standardize test name