Optimization of Light-Front Hamiltonian Generation for Domestic Heterogeneous Platforms

The Basis Light-Front Quantization (BLFQ) method is considered a promising approach for achieving first-principles calculations of nucleons, with its key challenge being the expansion of basis states to improve computational accuracy. However, increasing the number of basis states leads to an exponential growth in the size of the light-front Hamiltonian matrix. Traditional computational methods struggle with efficiency and scalability when handling such large-scale tasks, severely limiting the development of the BLFQ method.To address this issue, we propose an optimized Hamiltonian matrix generation scheme based on a domestic heterogeneous platform. This scheme integrates a two-level parallel strategy that combines MPI-based multi-process parallelism with GPU-like multithreaded parallelism, offloading intensive computations to acceleration cards for parallel processing. Additionally, optimizations such as atomic operations and hybrid sorting are applied to effectively reduce data copying within loops.Experimental results demonstrate that the optimized program achieves significant acceleration across various 5-Fock state test models with different basis truncation parameters, achieving speedup ratios ranging from 3.06 to 8.51. This work provides new technical support and a viable development path for realizing first-principles calculations of nucleons.