Han-Chung Wang
Han-Chung Wang
I generated an example using cpu=znver4. In your example, I think we will get something like for fusion (when we enable unpack propagation): ```mlir #map = affine_map (d0, d1, d2,...
The https://github.com/openxla/iree/issues/16025 is more for functionality -- i.e., making sure things are all controlled by ourselves. I think Benoit has more ideas about how it should work in terms of...
Yes, that's what I think in my mind. I don't want to enable the fusion while we are doing bad control on tile size selection, or say that's unexpected. And...
@pashu123 please help the further triaging. We dumped the inputs and outputs and verified that there are NANs.
Some related read: https://discourse.llvm.org/t/rfc-fix-floating-point-max-and-min-operations-in-mlir/72671 I recall the time that we split the `min(max)` to `minimum/minnum(maximum/maxnum)`. We could miss it in softmax because it was not on my radar.
The upstream change is landed to IREE: https://github.com/iree-org/iree/pull/18033 closing the issue.
It is no longer an issue with our encoding changes. :)
@pashu123 if this PR fixes the issue, you can add `Fixes https://github.com/iree-org/iree/issues/21107` in the PR description. The issue will be closed when you land the PR.
I wrote down what should happen step by step here, and I'm using the below MLIR input as example: ```mlir util.func public @broadcasting_dequant_op(%arg0 : tensor
status update, I have a change that sets encodings for matmul, see below snippet. The next step is bubble up the encodings op across generic ops, set the bcast_map if...