Benoit Jacob

Maybe as a first step disable on PR runs as discussed above https://github.com/iree-org/iree/issues/16856#issuecomment-2012396348. Then after a month of that, ask whether to drop altogether.

https://github.com/iree-org/iree/pull/17115 is merged now.

The first problem here is that we did materialize encoding for the `[f32, i8, f32]` combination of element types, which is not meant to be supported. I can see the...

> 1. We should decouple DT from UK. I don't see a reason to not apply DT to f32, i8 -> f32 matmul at least when UK is not enabled....

I saw it but it's different in two ways. 1. It's about < 8 bit as opposed to (any non-power-of-two) bits, so for example #16250 is concerned with `i4` while...

> Is this trying to avoid that rematerialization? Trying to avoid all stack buffers (that are not tiny static sized).

> IIUC there were raising patterns added to force the fusion that Flow should be doing by default. That's what I'm questioning here: I don't think we should ever have...

Oh... the whole concept of rematerialization is news to me here. Is this a mechanism whereby a local intermediate tensor inside a dispatch function, which would otherwise become a stack...

> No, that's just within the dispatches - it does not create new dispatches or global tensors. Ah... ok but then I need help understanding @qedawkins above: > For the...

> Luckily I don't see a reason to believe this is the case on CPU yet. On the contrary, it's very clear on CPU that the exp computation is essentially...