GitLab

- 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

* 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

* Ensure that neighbor list is valid before solving for charges

* Add test with neighbor list that needs to be resized

* Try another approach to skip interactions for neighbor list generation only

* Increase CG error tolerance for test

* Replace SimTK-containing file headers

* Update file headers for new Tinker reader files added

* Unified storage of global parameters

* Fixes to CUDA and HIP

* Store global parameters as real instead of float

* Use common API for kernels

* More code uses common interface

* Bug fixes

* Unified interface for sorting

* Simplified interface for FFT

* Use common event API for synchronization

* Minor changes to make code more consistent between platforms

* Common implementation of NonbondedForce

* Bug fixes

* Flag to enable list of single pairs

* CUDA and OpenCL use common implementation of NonbondedForce

* Fixed compilation error

* HIP uses common implementation of NonbondedForce

* Created DPDIntegrator class

* Reference implementation of DPDIntegrator

* Build neighbor list for DPDIntegrator

* Minor fixes

* Documentation for DPDIntegrator

* Python API for DPDIntegrator

* Preliminary OpenCL implementation of DPDIntegrator

* Enable USE_PERIODIC

* Use updated positions in DPD thermostat

* Working on neighbor list for OpenCL DPDIntegrator

* ReorderListener for particle types

* Serialization for DPDIntegrator

* CUDA implementation of DPDIntegrator

* HIP implementation of DPDIntegrator

* Fixed compile error in Python wrapper

* Fixed compile error in wrappers

* Fixed uninitialized memory in reference neighbor list

* Added DPDIntegrator to C++ API docs

* Fixed incorrect launch size

* Fixed nan in DPD random number generator

* Minor optimizations

* Improved load balancing

* Fixed an indexing error

* Neighbor list uses the maximum cutoff of any force

* Fixed HIP compilation error

* Fixed access to invalid memory

* Added test case for diffusion coefficient

* Try to debug segfaults on CI

* Debugging

* Debugging

* Debugging

* Debugging

* Debugging

* Debugging

* Possible fix

* Debugging

* Debugging

* Debugging

* Use correct block size on CPU OpenCL

* Workaround for bug in Intel's OpenCL for CPUs

* Removed an unnecessary define

* Removed debugging code

* Include Dart

* More Intel workarounds

* Workaround for error in NVIDIA OpenCL

Skip neighbor list for very small systems

    https://github.com/openmm/openmm/pull/4070

Store bounding box sizes in half precision

    https://github.com/openmm/openmm/commit/2ae50f9

Use large blocks to optimize building the neighbor list

    https://github.com/openmm/openmm/commit/3955033

Improved sorting of blocks when building neighbor list

    https://github.com/openmm/openmm/commit/796ffaa

Fixed bug in large blocks optimization with triclinic boxes

    https://github.com/openmm/openmm/commit/4c10732

Optimize sorting of non-uniformly distributed data

    https://github.com/openmm/openmm/commit/71d9bb1

Co-authored-by: bdenhollander <44237618+bdenhollander@users.noreply.github.com>

Use a small kernel for copying interactionCounts to host memory

    hipMemcpy's CopyDeviceToHost operation has higher latency.

Do not set stream and event blocking/spin related flags

    Let the runtime choose the best option because overriding does not
    improve performance in most cases.

Remove NULL streams and use nonblocking streams explicitly

Make HipContext::pushAsCurrent/popAsCurrent thread-safe as they can be
called simultaneously from different threads via ContextSelector.

Allow peer access to be enabled more than once (if there are multiple
simulations one after another, like in benchmark.py).

Create peerCopyStream on a corresponding device

Use two-speed load balancing for multi GPU runs

    First 100 steps do coarse balancing, next 100 - fine tuning.
    Also ignore the slowest device (usually 0) if its fraction has
    reached 0, (i.e. no work can be transfered to other devices) and
    balance other devices.

Do not download inteactionCounts in parallel nonbonded tasks

    This is not required because updateNeighborListSize has been called
    and valid flag changed.

Initialize tilesAfterReorder properly

    It may contain a garbage value, and if it is large then
    updateNeighborListSize does not force reorder atoms after 25 steps
    in extremal cases.

Use cuCtxPushCurrent() and cuCtxPopCurrent() for selecting CUDA context

    https://github.com/openmm/openmm/pull/3258

Fixed uninitialized memory access

    https://github.com/openmm/openmm/issues/3392
    https://github.com/openmm/openmm/pull/3399

Fixed potential invalid memory access

    See https://github.com/openmm/openmm/pull/3428

Improved temperature reporting for Drude particles

    https://github.com/openmm/openmm/pull/3486
    https://github.com/openmm/openmm/commit/a5e42f5

Fixed race condition with multiple GPUs

    https://github.com/openmm/openmm/commit/6fb1c8a41edff980862750bc086f6a204eb50941

Use blocking sync when creating events

    https://github.com/openmm/openmm/commit/fe21d5ee4f14673a4ea38b7244991772a64ceec2

Very minor optimizations

    https://github.com/openmm/openmm/commit/109f6b2535da4e0c0dd88007d6ca06b4add2ce81

Use PocketFFT

    https://github.com/openmm/openmm/commit/1dac981a63300a2a53a7925f570995914f7163ed

Improved logic for deciding when to reorder atoms

    https://github.com/openmm/openmm/commit/48664a1f1a4490a4dabc277757545ac070e7b898

Ensure valid atom order after loading a checkpoint

    https://github.com/openmm/openmm/commit/a056d5a3754e193105409afa12c9f0c9a2d972a2

Improve performance running on multiple GPUs

    https://github.com/openmm/openmm/commit/0c82c2647de98da5c6dab7bf7a7b8b19705aadc0

Fixed errors when running on multiple GPUs

    https://github.com/openmm/openmm/commit/ed9df876d43c037c08d4762721e73e5caae086d9

Optimized reducing energy

    https://github.com/openmm/openmm/commit/2975f44

* Unload all loaded modules in HipContext's destructor,
  HIP modules keep file desctriptors opened, but OpenMM never unloads
  modules leaking these file descriptors. This can cause crashinf of
  some scripts like test-openmm-platforms from openmmtools.
* ROCm 6.0 defines operator* for complex types (that are typedefs for
  float2 and double2), they conflict with operators defined for vectors.
  This is fixed in newer ROCm versions.
* Revert HIP_DYNAMIC_SHARED back to extern __shared__ (the macro is
  in the headers).
* Reduce the speed of the HIP platform if there are no HIP devices in
  the system.

Optimize findBlocksWithInteractions

* Replace volatile shared mem accesses with shuffles;
* Add NUM_TILES_IN_BATCH for processing block1 by multiple warps
  (for small systems);
* Cherry-pick missing changes from .cu;
* Tune MAX_BITS_FOR_PAIRS depending on device and the system size;
* Store single pairs immediately (if there are any), this allows not to
  store flags to shared memory and filter buffer and flagsBuffer after
  saving single pairs;
* Use fma explicitly and sign bit for better device code;
* Use CDNA's MFMA with singe/mixed precision;
* On CDNA the coarse grained stage processes warpSize blocks for
  one block1, the fine grained stage checks atoms of two block2 vs atoms
  of the same block1, singlePairs and interactingAtoms are also stored
  by warps, not half-warps;

Optimize findBlockBounds

* Use shuffles;
* Use executeKernelFlat;
* Process 2 tiles per warp 64 on CDNA;
* Use more uniformly distributed keys when sorting blocks;

Use compareInt2LargeSIMD when tile size < SIMD width

Fix exclusion tiles sorting on AMD CDNA (64 threads per wave)

    The nonbonded kernel uses USE_NEIGHBOR_LIST (useNeighborList)
    so host code also must check it instead of useCutoff.

    See also https://github.com/openmm/openmm/issues/3462

* All AMD GPUs support shuffle, double precision and 64-bit int atomics;
* Remove unused code: !ENABLE_SHUFFLE code paths in nonbonded.hip;
* Use intrinsics in single-precision;
* Use realToFixedPoint (faster float32-to-int64);
* Remove shared atomIndices, use shuffles;
* Check early if atoms are in the cutoff range, sometimes all lanes in
  a warp can skip computations, single pairs can also skip useless
  atomics with zero values;
* Remove volatile skipTiles access, use shuffles;
* Distribute work for warps in a strided order;
* Skip warps that may be still busy in the first loop;
* Unify conditions for excluded atoms with `includeInteraction`;
* Move multiprocessors to HipContext;
* Increase number of warps for computeNonbonded;
* Disable packed math for >=MI200 (it affects performance of some
  kernels like computeGKForces of amoebagk);
* Remove defaultOptimizationOptions and createModule's optimizationFlags
  as they are never used;
* Support -save-temps.

* Compile kernels with max block size of 256 threads:
  The default hipcc behavior since ROCm 4.2 is to compile kernels
  with 1024 threads unless __launch_bounds__ is specified. This
  significantly increases register pressure especially in heavy kernels
  (double precision, for example), requiring register spilling;
* Optimize computeRange by using multiple blocks for reduction;
* Use blocks of 1024 threads for computeBucketPositions - it is executed
  as a single work group so larger block size is faster;
* Sort up-to lenghtNextPow2 instead of blockDim.x (faster for short
  buckets);
* Optimize sortShortList2;
* Optimize sortBuckets with bit instructions;
* Decrease bucket size for non-uniform sorting: too many buckets may
  have sizes too large to sort in shared memory;
* Add more sizes in tests.

Port changes in CUDA backend to HIP

Fix a warning about arithmetic operations on void* in HipArray::uploadSubArray

Fix "Error Initializing context ROCm 5.3.0"

    https://github.com/StreamHPC/openmm-hip/issues/3


    hipDeviceSetCacheConfig returns hipErrorNotSupported on 5.3
Co-authored-by: Nick Curtis <nicholas.curtis@amd.com>

* Improved sorting of blocks when building neighbor list

* Improved block sorting for OpenCL

* Made sort keys more evenly distributed

* Use large blocks to optimize building the neighbor list

* Large blocks optimization for OpenCL

* Fix test failures

* Select whether to use large blocks based on system size

* Skip neighbor list for very small systems

* Fixed typos

* Don't skip box size check when not using neighbor list

* Made test larger to ensure it uses neighbor list

It may contain a garbage value, and if it is large then
updateNeighborListSize does not force reorder atoms after 25 steps
in extremal cases.

* Minor optimizations to computing single pairs

* Adjusted MAX_BITS_FOR_PAIRS on Ampere

* Fixed potential invalid memory access

* Fixed exception

* Optimized CudaSort for non-uniformly distributed data

* Optimized OpenCLSort for non-uniformly distributed data

* Further tuned distributing elements between buckets

* Copied optimizations over to OpenCL

* Began converting AMOEBA to common platform

* Beginning of OpenCL platform for AMOEBA

* Converted AmoebaVdwForce to common platform

* Cleaned up reference AMOEBA tests

* Began converting AmoebaMultipoleForce to common platform

* Continue converting AmoebaMultipoleForce to common platform

* Bug fixes

* Bug fix

* Continue converting AmoebaMultipoleForce to common platform

* Converting AmoebaMultipoleForce and AmoebaGeneralizedKirkwoodForce to common platform

* Converting AmoebaMultipoleForce and AmoebaGeneralizedKirkwoodForce to common platform

* Creating OpenCL version of AmoebaMultipoleForce and AmoebaGeneralizedKirkwoodForce

* Creating OpenCL version of AmoebaMultipoleForce and AmoebaGeneralizedKirkwoodForce

* Creating OpenCL version of AmoebaMultipoleForce and AmoebaGeneralizedKirkwoodForce

* Converted arrays from real3 to real

* Bug fix to OpenCL AmoebaGeneralizedKirkwoodForce

* Fixes for AMD GPUs

* Began converting HippoNonbondedForce to common platform

* Continuing to convert HippoNonbondedForce to common platform

* Continuing to convert HippoNonbondedForce to common platform

* Working on unifying PME kernels

* Fixed error on devices without 64 bit atomics

* Unified PME kernels

* Converted HippoNonbondedForce to common platform

* Creating OpenCL implementation of HippoNonbondedForce

* Continuing OpenCL implementation of HippoNonbondedForce

* Mostly finished OpenCL implementation of HippoNonbondedForce

* Eliminated three component vector types in host code

* Fix errors on CPU OpenCL

* Skip double precision tests for AMOEBA on OpenCL

* Bug fixes

* Bug fixes

* Fixed compilation error

* Fixed range overflow with very large numbers of atoms

* More fixes to overflow with large numbers of atoms

* Fix test failures

* Began creating common compute framework to unify code between CUDA and OpenCL

* Began OpenCL implementation of common compute framework

* Common implementation of CMMotionRemover

* CUDA implementation of common compute interface

* Converted HarmonicBondForce to common compute API

* Converted standard bonded forces to common compute API

* Converted ExpressionUtilities to common compute API

* Created ComputeParameterSet

* Converted custom bonded forces to common compute API

* Converted CustomCentroidBondForce to common compute API

* Converted CustomManyParticleForce to common compute API

* Moved lots of duplicate code from CudaContext and OpenCLContext to ComputeContext

* Converted GayBerneForce to common compute API

* Removed obsolete kernels

* Converted verlet integrators to common compute API

* Converted Langevin and Brownian integrators to common compute API

* Converted CustomIntegrator to common compute API

* Converted CustomNonbondedForce to common compute API

* Removed uses of a deprecated API

* Fixed failing test cases

* Converted GBSAOBCForce to common compute API

* Began converting CustomGBForce to common compute API

* Finished converting CustomGBForce to common compute API

* Merged duplicated code in CudaIntegrationUtilities and OpenCLIntegrationUtilities

* Converted RMSDForce and AndersenThermostat to common compute API

* Converted CustomHbondForce to common compute API

* Merged scripts for encoding kernel sources

* Converted Drude plugin to common compute API

* Fixed errors in CMake scripts

* Attempt at fixing errors on Windows

* Added discussion of common compute API to developer guide

* Added Windows export macro for common classes

* Fixed error in CMMotionRemover

* Ubdated travis to newer Ubuntu version

* Fixed errors on CPU OpenCL

* Fixed Windows linking errors

* Added missing pragma for 32 bit atomics

* Replaced long long with mm_long

* More fixes to Windows linking

* Bug fix

* Created API for HIPPO force field

* Beginning of reference implementation of HIPPO

* Continuing reference implementation of HIPPO

* Continuing reference implementation of HIPPO

* Continuing reference implementation of HIPPO

* Continuing reference implementation of HIPPO

* Continuing reference implementation of HIPPO

* Continuing reference implementation of HIPPO

* Continuing reference implementation of HIPPO

* Completed reference of HIPPO with no cutoff

* Beginning cutoffs/PME for reference implementation of HIPPO

* Continuing PME for reference implementation of HIPPO

* Continuing PME for reference implementation of HIPPO

* Continuing PME for reference implementation of HIPPO

* Completed reference implementation of HIPPO

* Cleanup and optimization to HIPPO reference

* Further cleanup to HIPPO

* Combined direct space interactions into a single loop

* Compute direct space interactions in quasi-internal frame

* Beginning of CUDA implementation of HIPPO

* Continuing CUDA implementation of HIPPO

* Continuing CUDA implementation of HIPPO

* Continuing CUDA implementation of HIPPO

* Continuing CUDA implementation of HIPPO

* Continuing CUDA implementation of HIPPO

* Continuing CUDA implementation of HIPPO

* Continuing CUDA implementation of HIPPO

* Continuing CUDA implementation of HIPPO

* Continuing CUDA implementation of HIPPO

* Continuing CUDA implementation of HIPPO

* Continuing CUDA implementation of HIPPO

* Continuing CUDA implementation of HIPPO

* Continuing CUDA implementation of HIPPO

* Finished CUDA implementation of HIPPO

* More features and test cases for HippoNonbondedForce

* Serialization and Python API for HippoNonbondedForce

* Fixed sign error in computing forces