arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2212.03034cs.LGcs.PFcs.PL
收藏

将基于脉动阵列的硬件加速器集成到 DNN 算子自动调优框架中

Integration of a systolic array based hardware accelerator into a DNN operator auto-tuning framework

FZI Research Center for Information Technology(FZI信息技术研究中心);University of Tübingen(蒂宾根大学)

F. N. Peccia, O. Bringmann

AI总结:

本文提出将 TVM 代码生成框架与脉动阵列加速器 Gemmini 集成,通过通用调度将 GEMM 运算卸载到 Gemmini 上,并在 FPGA 上实现 46 GOPs 峰值吞吐量,超越默认手工调优调度。

摘要:

由于缺乏针对异构片上系统(SoC)与定制加速器耦合部署神经网络的端到端软件工具,这一任务具有很大的挑战性。此外,加速器开发者针对典型张量运算提供的底层调度和映射策略,并不一定在每个具体用例中都是最优的。因此,能够在特定硬件配置上自动测试生成代码性能的框架具有特别的研究价值。本文介绍了代码生成框架 TVM 与基于脉动阵列的加速器 Gemmini 之间的集成方案。文中详细描述了一种将通用矩阵乘法(GEMM)张量运算卸载到 Gemmini 上的通用调度方法,并通过在其上执行 AutoTVM 调优流程来验证其适用性。我们生成的代码在 Xilinx ZCU102 FPGA 上以 100 MHz 时钟频率运行时,达到了每秒 460 亿次运算(46 GOPs)的峰值吞吐量,优于先前的工作。此外,该集成方案生成的代码在实际工作负载中能够超越 Gemmini 开发者提供的默认手工调优调度。

英文摘要:

The deployment of neural networks on heterogeneous SoCs coupled with custom accelerators is a challenging task because of the lack of end-to-end software tools provided for these systems. Moreover, the already available low level schedules and mapping strategies provided by the accelerator developers for typical tensor operations are not necessarily the best possible ones for each particular use case. This is why frameworks which automatically test the performance of the generated code on a specific hardware configuration are of special interest. In this work, the integration between the code generation framework TVM and the systolic array-based accelerator Gemmini is presented. A generic schedule to offload the GEneral Matrix Multiply (GEMM) tensor operation onto Gemmini is detailed, and its suitability is tested by executing the AutoTVM tuning process on it. Our generated code achieves a peak throughput of 46 giga-operations per second (GOPs) under a 100 MHz clock on a Xilinx ZCU102 FPGA, outperforming previous work. Furthermore, the code generated by this integration was able to surpass the default hand-tuned schedules provided by the Gemmini developers in real-world workloads.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑