Code Generation for Near-Roofline Finite Element Actions on GPUs from Symbolic Variational Forms
基于符号变分形式的GPU近屋顶线有限元作用代码生成
专题命中 代码生成 :code generation(title)
AI总结 该研究针对GPU上的有限元方法变分形式,提出一种基于代码转换与启发式成本模型的并行化策略,在三代Nvidia GPU上测试,60%测试用例实现超50%屋顶线性能。
Comments Key changes: (i) Performance evaluation on the Nvidia H200 NVL. (ii) Sensitivity analysis of parameters in our transform space to the transformed kernel's throughput. (iii) Evaluation of our cost model's ranking quality by measuring Spearman's ranking correlation and performing a regret analysis by considering the Top-k subset for varying k