基于TVM深度学习编译器的算子融合规则研究OA
RESEARCH ON OPERATOR FUSION RULES BASED ON TVM DEEP LEARNING COMPILER
TVM深度学习编译器针对不同类型的算子提供了三种通用的算子融合规则.为了进一步增大 TVM深度学习编译器的算子融合粒度,提出两种新型算子融合规则,一种针对只包含降维运算的计算结构,一种针对同时包含降维运算和逐元素运算的计算结构.实验结果表明,应用新型算子融合规则后,TVM推理 GoogleNet、DenseNet、BVLC AlexNet、Mobilenetv2 等模型的时间减少了 17.3%～32.2%.
TVM deep learning compiler provides three general operator fusion rules for different types of operators.In order to further increase the granularity of operator fusion in the TVMdeep learning compiler,two novel operator fusion rules are proposed,one for the computational structure that contains only the dimensionality reduction operation,and one for the computational structure that contains both the dimensionality reduction operation and the element-by-element operation.The experimental results show that after applying the novel operator fusion rules,the time of TVM reasoning about models such as GoogleNet,DenseNet,BVLC AlexNet,and Mobilenetv2 is reduced by 17.3%～32.2%.
赵薇;李颖颖;韩林
郑州大学计算机与人工智能学院 河南 郑州 450001信息工程大学数学工程与先进计算国家重点实验室 河南 郑州 450001国家超算郑州中心 河南 郑州 450001
信息技术与安全科学
算子融合TVM深度学习编译器深度学习编译优化模型推理
Operator fusionTVM deep learning compilerDeep learning compilation optimizationModel inference
《计算机应用与软件》 2026 (5)
18-22,62,6
评论