Goal

• 使用特定硬件上 stoa 的计算库自动生成可部署且高性能的代码
• 通过自动生成代码，考虑更大的解决搜索空间，自动选择更高效的算子即算子排布、组合等进行编译优化

Related Work

• XLA: Tensorflow compiler
  • Constant shape dimension
  • Data layout is specific
  • Operations are low level tensor primitives
    • Map, Broadcast, Reduce, Convolution, ReduceWindow,...

• TensorRT: Rule based Fusion

  

  • Simple Graph-based Element-wise Kernel Generator

    

• Halide: Image processing language

• Loopy: python based kernel genrator

• TACO: sparse tensor code generator

• Tensor Comprehension

Motivation

• 目前的深度学习框架都基于计算图的表示方式
  • 需要针对不同的硬件构建和优化算子 → Tensor Index Expression

Challenges and Goals

Leveraging Specific Hardware Features and Abstractions

• inputs
  • multi-dimensional, with fixed variable length
  • different data layouts
  • special requirements for memory hierarchy
• accelerator designs
  • favor leaner control and offload most scheduling complexity to compiler stack
  • needs to explicitly control pipeline dependencies to hide memory access latency

Large Search Space for Optimization

• Producing efficient code without manually tuning operators
• Huge configuration space
  • memory access, threading pattern, novel hardware primitives

Others