In our presentation, we will introduce Apache TVM - tensor compiler for boosting execution of ML networks. Apache TVM is a powerful tool for optimizing DL models. On the example of Adreno GPU we will demonstrate how user can significantly improve performance of his model with TVM, by utilizing HW specific of a target platform. In 2021, initial research to enable Adreno textures in Apache TVM was prototyped based on Mobilenet v1 workload and demonstrated promising performance results, meanwhile other networks still worked slowly on Qualcomm Adreno GPU. In 2022, a big amount of optimization work was done, and significant performance boost was achieved. In our talk, we will demonstrate how users can use TVM to run models on their target platforms and what should they do to achieve the best performance.

Speakers:

Egor Churaev: Deelvin Solutions, Sr. Software Engineer, Sr. Software Engineer with 10 years of experience, worked on several projects at Intel, in particular: OpenCL CPU compiler and Intel OpenVINO, PhD student at HSE University with topic related to emotion recognition with deep learning algorithms. Currently work on Apache TVM project. Commiter of Apache TVM.