A High-Performance Implementation of Atomistic Spin Dynamics Simulations on x86 CPUs

Atomistic spin dynamics simulations provide valuable information about the energy spectrum of magnetic materials in different phases, allowing one to identify instabilities and nature their excitations. However, time cost evaluating dynamical correlation function $S(\mathbf{q}, t)$ increases quadratically as number spins $N$, leading significant computational effort, making simulation large systems very challenging. In this work, we propose use a highly optimized general matrix multiply (GEMM) subroutine calculate spin-spin that can achieve near-optimal hardware utilization. Furthermore, fuse element-wise operations calculation into in-house GEMM kernel, which results further performance improvements 44\% - 71\% on several relatively lattice sizes when compared implementation uses OpenBLAS, is state-of-the-art open source library for Basic Linear Algebra Subroutine (BLAS).