Subgraph Mapping Optimization Methods for Domain-Specific Architectures.

Because the deep learning model does not require the existence of domain-specific architectures (DSA) operator types or fusion modes corresponding to subgraphs when the code is finally deployed, these subgraphs often cannot be deployed directly on the DSA. If the custom subgraph is directly converted into the inference model on the corresponding hardware without processing, the searched subgraph may be re-split into the pre-recombination structure, resulting in optimization failure. In order to enable the custom subgraph to reason correctly on the hardware and achieve the expected performance, this paper maps the subgraph as an operator in the layer, so that the back-end of the hardware framework will treat the custom sub graph as a whole to perform scheduling optimization when generating code. The method proposed in this paper reorganizes the sub graphs of Faster-rcnn and MobileViT, and deploys the optimized models to Ascend 310 platform, which can improve the overall inference performance by 1.19 x ~ 1.45 x and tuning efficiency by 1.41 x ~2.08 x.