An operational model of schedule primitives in Tensor Expression

[YuanLin]

Hi there,

Yongfeng Gu (@maplegu) and I are working on an operational model of the schedule primitives in Tensor Expression. This document aims to make it easier to understand

• the interactions between different schedule primitives,
• the impact of the schedule primitives on the final code generation.

The model is inspired by Chapter 7 of Jonathan Ragan-Kelley’s PhD thesis and an early version of the TVM paper.

The very initial draft (rev 0.1) is available here. We welcome you all to comment on it, correct it, answer open questions, or help improve it in any form you like.

The document is currently in Google doc. Please let us know if there is a better way to collaborate on the working document.

We hope it can be included as part of the TVM Design and Developer Guide document when ready.

We will also give a lightning talk about this at the coming TVM conference. Come and ask us questions

[pub] An Operational Model of Schedules in Tensor Expression

[broune]

I think it’s great to be more precise and complete with the definitions/documentation of what a schedule does, especially if that can help ramp-up time on learning to write (good) schedules or root out any disagreements on what something should mean. So this is already great for documentation.

I’m curious if you have thoughts on using something like this in relation to applying formal methods to schedules, or a schedule debugger/exploration tool based on this model and visualization, or other follow-on uses?

(I, for one, like Google docs, the collaborate and commenting features there are quite good.)

[YuanLin]

@broune, great chatting with you at the last TVM conference. This is just to share what we discussed on the forum. We have thought about all the use cases you mentioned. First need to get the documentation more complete We are hoping to get the visualization tool we are working on out soon.