| 向春阳* **,李振东***,贾伟乐* **.量子化学密度矩阵重正化群关键算子并行优化[J].高技术通讯(中文),2026,36(5):441~454 |
| 量子化学密度矩阵重正化群关键算子并行优化 |
| Parallel optimization of key operators in density matrix renormalization group algorithm for quantum chemistry |
| |
| DOI：10. 3772 / j. issn. 1002 - 0470. 2026. 05. 001 |
| 中文关键词: 密度矩阵重正化群； 高性能计算； 异构计算 |
| 英文关键词: density matrix renormalization group, high-performance computing, heterogeneous computing |
| 基金项目: |
| 作者 | 单位 | | 向春阳* ** | （*处理器芯片全国重点实验室，中国科学院计算技术研究所北京 100190）
（**中国科学院大学北京 100049）
（***北京师范大学化学学院北京 100875） | | 李振东*** | | | 贾伟乐* ** | |
|
| 摘要点击次数: 29 |
| 全文下载次数: 13 |
| 中文摘要: |
| 多核过渡金属化合物（如固氮酶中的铁硫簇）因其复杂的电子结构，对当代量子化学方法构成了严峻挑战。为此，本文提出了一种基于中央处理器-图形处理器（central processing unit-graphics processing unit，CPU-GPU）异构计算平台的优化方案，旨在提升密度矩阵重正化群（density matrix renormalization group，DMRG）算法的计算效率。其核心理念在于，通过异构计算优化DMRG中最耗时的关键算符，即对角化（diagonalization）和重正化（renormalization），从而显著提升整体性能。这一挑战的关键不仅在于计算速度的提升，更在于如何巧妙分解这些算符的计算负载并实现高效的资源分配。为此，本文设计了一套多层次优化策略：利用GPU上的层次化批量矩阵乘法，辅以高级精简指令集机器CPU上针对不规则矩阵乘法的并行优化，充分挖掘了CPU-GPU异构系统的计算潜能。实验结果显示，相较于我们先前的研究，DMRG的综合性能实现了11%~83%的提升。这一进步不仅验证了CPU-GPU协同计算在处理DMRG算法时的优越性能，也为量子化学领域的高性能计算开辟了新的路径。 |
| 英文摘要: |
| Polynuclear transition metal compounds, such as the iron-sulfur clusters in nitrogenase, present a formidable challenge to modern quantum chemistry methods due to their intricate electronic structures. To tackle this, we introduce an optimization framework based on a central processing unit-graphics processing unit (CPU-GPU) heterogeneous computing platform, designed to enhance the computational efficiency of the density matrix renormalization group (DMRG) algorithm. The central concept hinges on leveraging heterogeneous computing to optimize the most time-intensive key operators in DMRG—specifically, diagonalization and renormalization—thereby markedly improving overall performance. The crux of this challenge lies not merely in boosting computational speed, but in the astute decomposition of these operators’ workloads and the efficient allocation of resources. To this end, we devise a multi-tiered optimization strategy: employing hierarchical batched matrix-matrix operations on GPUs, complemented by CPU-based parallel optimization for irregular matrices, to fully harness the computational potential of the CPU-GPU heterogeneous system. Experimental results reveal that, compared to our previous studies, the comprehensive performance of DMRG achieves a 11% to 83% improvement. This advancement not only underscores the superior capability of CPU-GPU synergistic computing in addressing DMRG computations but also paves new avenues for high-performance computing in quantum chemistry. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
