Add checks for unknown workgroup count when inferring boundaries.

[hanhanW]

The workgroup count is set to zero when we're not able to infer the bounds.

Fixes #9244

[dcaballe]

Thanks for looking into this!

[MaheshRavishankar]

Awesome! Thanks!

[MaheshRavishankar]

I actually have a question that is maybe unrelated, but might help fix this better. It might be worht looking at the call sites of these methods. The workgroup_count setting to 0 itself is wrong as default.

[hanhanW]

I actually have a question that is maybe unrelated, but might help fix this better. It might be worht looking at the call sites of these methods. The workgroup_count setting to 0 itself is wrong as default.

I had the same idea in the beginning, and I decided to check zeros.

Below is the implementation of getting workgroup_count. We're able to distribute some loops, but the bound (or say workload_size) can't be inferred. It does not only happen in peeling case, but also in dynamic shapes. We'll need a thing about unknown. Setting them to zero means that the workgroup count is unknown to me.

Ideally, the coding style would suggest to use Optional. It let the type tell us that they are unknown. However, it increases memory usage and code complexity. Using zero in this case sounds reasonable to me. zero is an invalid setting and could be an unknown option here.

https://github.com/google/iree/blob/2187016fa6866437502987ce8eda17988b5f1134/compiler/src/iree/compiler/Codegen/Common/RemoveTrivialLoops.cpp#L94-L134

[MaheshRavishankar]

I actually have a question that is maybe unrelated, but might help fix this better. It might be worht looking at the call sites of these methods. The workgroup_count setting to 0 itself is wrong as default.

I had the same idea in the beginning, and I decided to check zeros.

Below is the implementation of getting workgroup_count. We're able to distribute some loops, but the bound (or say workload_size) can't be inferred. It does not only happen in peeling case, but also in dynamic shapes. We'll need a thing about unknown. Setting them to zero means that the workgroup count is unknown to me.

Ideally, the coding style would suggest to use Optional. It let the type tell us that they are unknown. However, it increases memory usage and code complexity. Using zero in this case sounds reasonable to me. zero is an invalid setting and could be an unknown option here.

https://github.com/google/iree/blob/2187016fa6866437502987ce8eda17988b5f1134/compiler/src/iree/compiler/Codegen/Common/RemoveTrivialLoops.cpp#L94-L134

(Its now a priority for me to kill this pass! Ill look into how to do this. This code is all just tech debt now).

Its not clear to me how the workgroup_size ends up as zero if it isnt distributed. This all seems like a footgun. I'm fine with landing this, but I think for any exploration this is causing issues with, we should just drop this pass. (FYI @dcaballe )

[dcaballe]

Its not clear to me how the workgroup_size ends up as zero if it isnt distributed. This all seems like a footgun. I'm fine with landing this, but I think for any exploration this is causing issues with, we should just drop this pass. (FYI @dcaballe )

I tried dropping the pass but that led to other issues. Perhaps there are some dependencies with this pass down the road? Let me see if this fixes the issue.