feinsum

Feinsum, short for Fast-Einstein Summations, provides tools for managing database of code-transformations on Batched Einstein Summations. Primarily, it serves the following purposes:

1. Provides tools for performance engineers to record code-transformation spaces for a class of Batched Einstein Summations.

2. Automatically record facts (such as walltime, FLOPS) about the points in the transformation space specified by a performance engineer. (current implementation uses OpenTuner to traverse the search space)

3. Defines (and implements) a specification to translate the recorded code-transformations to a subset of loopy kernels.

The rest of this manual is structured as follows. We outline our design choices are elaborate in Sec “Design”. We provide tutorials in Sec “Tutorial: Expressing code-transformation space” an Sec “Tutorial: Using the transformation knowledge”. The API reference is generated in Sec. “API Reference”.

Contents:

• Installing feinsum
• Tutorial: Expressing code-transformation space
• Tutorial: Using the transformation knowledge
• Design
  • Dependencies
  • What is a “Batched Einstein Summation”?
  • Grammar of loopy kernels
  • Why keep a performance engineer in the loop?
• API Reference
  • Einsum Nodes
  • Routines for empirical evaluation
  • Helpers to build feinsum.einsum.FusedEinsum
  • Interacting with code-transformation databases
  • Some helper loopy transformations relevant for Einstein Summations
  • Lowering to loopy
  • Canonicalization
  • Miscellaneous Utilities
• Indices and tables
• References
• 🚀 Github
• 💾 Download Releases