junrushao1994 commented on pull request #7630: URL: https://github.com/apache/tvm/pull/7630#issuecomment-800821686
Having well-documented syntax for the TensorIR TVM script is the key prerequisite of usability that we should push for - without the proper documents, nobody other than the developers will be able to really understand what exactly the syntax is and use the script well. Here is what we propose to do for the TensorIR TVM script (per discussion with Tianqi and Tristan): - prepare a stub file for TensorIR TVM script, like what numpy does - In the stub file, we document the syntax for all elements allowed in the TensorIR, including like `tir.alloc_buffer`, `ty.handle`, etc - Specifically, the `Block` signature needs more clearly explanation, like what is `bind`, what are the read/write regions, etc Besides the syntax, it is also desirable to have a centralized document explaining high-level ideas and motivations, to deliver deeper understanding to users who are interested. Note that the functionality of TensorIR is not complete until M1s and M2s are done, which means it is hard to have users before M1s/M2s are fully upstreamed. We would love to move fast and get this series of PRs quickly merged with proper unittests (and possible some of the integration test) - opening possibilities for developers to write demos and do codegen with TensorIR is the most urgent actionable item. After those PRs are merged, we should then move fast and write really good specs on how it works to attract interested users. It doesn't mean we should lower the bar of documents. Within the scope of this PR, I think it will be good enough to document each user-facing method clearly, like what we have been continuously pushing for. In the next PR, which includes a set of key data structures, I will provide a list of glossaries in the documents so that readers could better understand. The stub file and the spec can be done after M1s and M2s are fully upstreamed, when we started to have alphas for interested users :-) ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected]
