# Bring Your Own Codegen to Deep Learning Compiler

**Authors:** Zhi Chen, Cody Hao Yu, Trevor Morris, Jorn Tuyls, Yi-Hsiang Lai, Jared Roesch, Elliott Delaye, Vin Sharma, Yida Wang  
**arXiv:** [2105.03215](https://arxiv.org/abs/2105.03215)  
**Date:** 2021-05-03  
**Summary page:** https://www.summarizepaper.com/en/paper/2105.03215v1/

## Key points

- The paper addresses challenges faced by accelerator vendors in developing and maintaining software stacks for deep neural networks (DNNs)
- DNNs are widely used in various applications, and accelerators play a crucial role in supporting fast and efficient inference tasks
- Achieving high model coverage with high performance requires each vendor to develop a full compiler stack for DNNs
- Continuous updates to hardware and/or software are needed to keep up with the rapid evolution of DNN model architectures and operators
- The authors propose an open-source framework that allows users to focus on developing their proprietary code generation tools while reusing components from existing deep learning compilers
- The framework provides flexible and user-friendly interfaces for partitioning models into segments, which can be executed on suitable processors, taking advantage of accelerator capabilities
- A case study demonstrates the deployment of the framework in multiple commercial vendors' compiler stacks using only a few thousand lines of code
- The approach offers a practical solution for addressing complexities associated with developing and maintaining software stacks for DNN accelerators
- By providing a reusable framework and enabling users to concentrate on their proprietary code generation tools, high model coverage with high performance can be achieved in DNN applications.

## Summary

The paper titled "Bring Your Own Codegen to Deep Learning Compiler" addresses the challenges faced by accelerator vendors in developing and maintaining software stacks for deep neural networks (DNNs). DNNs have become widely used in various applications, and accelerators play a crucial role in supporting fast and efficient inference tasks. However, achieving high model coverage with high performance requires each vendor to develop a full compiler stack to handle the ingestion, optimization, and execution of DNNs. This process poses significant challenges as it involves continuous updates to hardware and/or software to keep up with the rapid evolution of DNN model architectures and operators. To overcome these challenges, the authors propose an open-source framework that allows users to focus on developing their proprietary code generation tools while reusing components from existing deep learning compilers. The framework provides flexible and user-friendly interfaces that enable users to partition their models into segments. These segments can then be executed on the most suitable processors, taking advantage of the powerful computation capabilities of accelerators. The authors present a case study demonstrating the deployment of their framework in multiple commercial vendors' compiler stacks using only a few thousand lines of code. Overall, this paper introduces an innovative approach which offers a practical solution for addressing the complexities associated with developing and maintaining software stacks for DNN accelerators. By providing a reusable framework and enabling users to concentrate on their proprietary code generation tools, this approach enables users to achieve high model coverage with high performance in DNN applications.