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.

- 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

* Context deselection before energy evaluation

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

* Avoid host wait when synchronizing PME queue

* Remove comment

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.

* PythonForce can be restricted to a subset of particles

* Fix exception with CUDA

* optimization, switch to unordered map

* Remove emplace, reduce reserve allocation

* Cache coefficients for long range correction

* updateParametersInContext() clears cache

* Fixed issue that caused inefficient sorting when a block contained only one atom

* Add the fix to OpenCL and HIP

* Make reference/CPU minimizer into a kernel

* Add per-platform support for GPU minimization

* Initial implementation of GPU minimization

* Fixes

* Increase robustness when initial gradient is huge

* Handle overflow leading to non-finite values gracefully

* Handle large forces in single precision more robustly

* Optimize kernels

* Fix kernel launch size

* Update banner years

* Don't create MinimizeKernel until first minimization requested

* Make some compile-time constants into kernel arguments

* Consolidate scale calculation kernel

* Condense alpha/beta reduction kernels using atomics

* Condense line search dot kernels with reductions

* Remove a download, and download grad norm separately

* Asynchronously check lbfgs convergence condition

* Restructure line search to avoid download waiting

* Start line search preemptively in case CPU evaluation is not needed

* In rare cases, constraint error might not decrease after one optimization round

* Better handling of unsupported 64-bit atomics, use FLT_MAX

* Pick gradient mode based on GPU vs. CPU evaluation

* Rework getDiff/getScale reduction, remove reduceBuffer

* Older CUDA might not like float hex literals

* Fix error in a comment

* API for querying devices

* CUDA and HIP implementations of getDevices()

* Fix test failures

* Fix test failures

* CUDA returns correct devices even if no context has been created

* Return a single device for Reference and CPU

* Fix CI failure

* Remove std::enable_if, warpRotateLeft is always used with TILE_SIZE

* Do not use built-in warpSize in constexpr contexts

Starting from ROCm 7 warpSize is no longer constexpr.
findInteractingBlocks.hip uses it for sizes of __shared__ arrays.

* Check if hipHostMallocNumaUser is allowed before using it

adaptFrictionPart1 is executed first, it reads randomForce and
segmentVelocity buffers before other kernels write them so they can
contain garbage after allocation (cuMemAlloc/hipMalloc/clCreateBuffer
do not clear allocated memory).

* Basic LCPO support

* Add basic test for LCPO from a prmtop file

* API for LCPOForce

* Started LCPO reference implementation

* Finished reference forces & test cases

* Use other test for finite difference since grid might have discontinuous forces

* Reference platform formatting

* Initial implementation of CPU platform

* Bugfixes

* More vectorization and improve neighbor list query speed

* Parallelize part of neighbor search

* Check box size for LCPO with periodic boundary conditions

* Fixes for updating parameters in context

* GBSAOBCForce doesn't use first & last indices for updates, so no need for this optimization here

* Changes to neighbor checking and optimization

* Fixes and minor changes

* Add global surface tension parameter

* Only process half of the pairs in the neighbor list

* Remove unnecessary checks

* Initial version of common platform implementation

* Asynchronously download neighbor list size

* Debugging

* Do pair precomputation in copyPairsToNeighborList

* Recompute interactions instead of scanning neighbor list in inner loop

* Condense position array before computations

* Also make neighbor count download asynchronous on device

* Fixes for kernel launching

* Topology-based LCPO parameter assignment

* Fixes, and use test system for LCPO with nucleic acids

* Always raise instead of warn when LCPO parameters can't be assigned

* Use Amber convention for phosphates

* Minor optimization to PME

* Minor simplification

* Initial implementation of PythonForce

* Continuing implementation of PythonForce

* Tests for PythonForce

* Fix memory leaks

* Documentation for PythonForce

* Fixed incorrect return type

* Fix compilation error on Python older than 3.12

* Handle all dtypes

* Optimizations to PythonForce

* Optimized getPositions()

* Test all platforms

* Fix test failures