Thanks @Mousius for drafing this RFC!
First of all, I completely agree on the importance to handle `arch`-specific
checks. Use our experience as an example, on CUDA we might want to check if the
PTX intrinsic `cp.async.commit_group` is available on certain architecture
before tensorizing using
Hi @junrushao1994, thanks for the elaborate reply 😸 I don't want to debate our
personal principles but I appreciate you sharing them and will reference them
where I can.
> **Current `arch`-specifc checks.** Most of the 6 groups of `arch`-specific
> helper functions, mentioned in the "Motivatio
Thanks folks for discussions. I think they summarizes to the following points
- Q0: Subfield grouping (e.g. features) or simply leave as top-level attrs
- Q1: Folder structure: `target/preprocessors/cuda.cc` vs
`target/cuda/cuda_preprocessor.cc`
- Note that code-reuse is less likely going to
Thanks folks for discussions. I think they summarizes to the following points
- Q0: Subfield grouping (e.g. features) or simply leave as top-level attrs
- Q1: Folder structure: `target/preprocessors/cuda.cc` vs
`target/cuda/cuda_preprocessor.cc`
- Note that code-reuse is less likely going to
One more collage/uma overlap aspect: Collage distinguishes 'registered'
backends (ie just TargetKinds) from 'activated' backends (ie Target objects in
the provided build targets). I think though the proposal here is the act of
registration is also activation? I need help understanding how this w
@wrongtest I've thought about the option A3 vs A4. From the parsing /
translation from TVM script to TIR, it is acceptable to have `T.allocate`
translated to `Allocate + DeclBuffer` two nodes. But it will be tricky for
`TVMScriptPrinter`. We will need to find both `Allocate` and `DeclBuffer` nod
@Mousius Thank you so much for your response! This makes lots of sense to me!
Also, thanks for including my personal principles in the discussion! It's my
personal principles which are completely okay to disagree with :-)
> I'm not sure why we consider that pollution given it should have a posit
