modified: src/force/neighbor.cu

[YixinDeng]

Summary
This PR adds local size_t casts to large NEP array size calculations and flattened neighbor-list/descriptor offsets that can exceed 32-bit integer range in large-system GPU runs.

Modification

• Cast large GPU_Vector::resize() products to size_t in NEP::NEP():

  • nep_data.NL_radial
  • nep_data.Fp
  • nep_data.sum_fxyz

• Cast per-GPU GPU_Vector::resize() products to size_t in NEP_MULTIGPU::allocate_memory():

  • nep_data[gpu].NL_radial
  • nep_data[gpu].Fp
  • nep_data[gpu].sum_fxyz

• Cast flattened NEP large-box offsets to size_t for:

  • global neighbor-list reads before building NEP radial lists
  • NL_radial writes and reads
  • Fp writes and reads
  • sum_fxyz writes and reads

• Cast the global neighbor-list allocation size in Neighbor::initialize().

• Cast flattened global/local neighbor-list offsets in non-ILP global neighbor-list kernels.

• Cast flattened read/write offsets in the ordinary neighbor-list sort kernel in neighbor.cuh.

Validation

The x-axis is the number of atoms, shown on a logarithmic scale. The speed panel reports throughput in units of 10^7 atom step s^-1. The memory panel reports peak GPU memory used in GiB. Blue circles represent the old version; orange squares represent the new version. Dashed vertical lines in the memory panel mark the first failed point for each version. The A100 speed panel includes an inset that zooms in on the high-atom-count region of the new version.

Compared with the old version, the new version reaches about 3.62x more atoms on A100 and about 1.47x more atoms on 2V100. The corresponding peak GPU memory usage increases from 22.50 to 77.69 GiB on A100, and from 39.54 to 57.22 GiB on 2V100.