文献阅读（245）Roller

• 题目：Roller: Fast and Efficient Tensor Compilation for Deep Learning
• 时间：2022
• 会议：OSDI
• 研究机构：微软

本篇论文的主要的motivation在于现在DNN的编译探索工作的时间比较长，特别是针对Nivida以外的硬件平台如AMD GPU和Graphcore IPU，所以这篇论文换了一个思路，采用构造的方式生成kernel，首先介绍基本概念：

• rTile：最基础的抽象层级，就是一个data tile，不过与计算和访存的基础尺寸对应
• rProgram：包括了load、store、compute的基于rTile的程序，可以满足GPU中的SM执行
• kernel：利用rProgram构造kernel

1. rTile is a new tile abstraction that encapsulates tensor shapes that align with the key features of the underlying accelerator, thus achieving efficient execution by limiting the shape choices.
2. rProgram: adopts a recursive rTile-based construction algorithm to gradually increase the size of the rTile shape to construct an rProgram that saturates a single execution unit of the accelerator (e.g., an SM, a streaming multi-processor in a NVIDIA GPU)
3. kernel: performs the scale-out process, which simply replicates the resulting rProgram to other parallel execution units