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

* 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

* 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

* Add a workaround for infinite loop in computeNonbonded (HIP)

computeNonbonded hangs in some tests (without neighbor list).
Reproducible on ROCm 6.4 and 6.4.1 (maybe on older versions too) on various architectures (both CDNA and RDNA).
Affected tests: TestHipATMForce, TestHipMonteCarloBarostat, TestHipNonbondedForce, TestHipVirtualSites.

Disassembly shows that the compiler splits branches of `if (skipBase+tgx < NUM_TILES_WITH_EXCLUSIONS)` and does
`SHFL(skipTiles, TILE_SIZE-1) < pos` checks in them separately, even though `__builtin_amdgcn_ds_bpermute`
is a convergent function. Apparently in this case not all lanes participate in each call.

* Simplify includeTile check using ballot

* PME_ORDER threads process one atom;
* PME_ORDER threads access consecutive addresses;
* No need to permute z indices with zindexTable;
* finishSpreadCharge is needed only with fixed point charge spreading;

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

* VkFFT-based 3D FFT;
* Caching of compiled VkFFT kernels;
* Extend FFT tests with more sizes.

* 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.

* Compile with -munsafe-fp-atomics to enable fast hardware f32 atomic
  add on global memory on pre-MI100 GPUs;
* Use fixed point charge spreading on other GPUs, otherwise float atomic
  add will be compiled as a slow CAS loop;
* Tune block sizes, use executeKernelFlat;
* Tune launch bounds of PME grid-related kernels: force the compiler to
  use all registers by limiting max waves per EU to 1.

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>