srcarroll wrote: > I'm -1 on using `tensor.reshape` op. IMO, we should only use > tensor.expand/collapse_shape; they work much better with existing > transformations. > > Out of curiosity, what use case do you have in mind? Why do we lower fully > dynamic pack op? If it is at high level graph level, we can just use > `tensor.pack` which carries more meaningful information. If it is at low > level stage (e.g., around vectorization), I think the inner tile sizes should > already be resolved to static values? In this context, we can still use > `tensor.expand_shape`. It supports the case where one dynamic extent can be > expanded into a single dynamic extent and other static extents (e.g., `? -> > ?x4`).
I'll admit I dont know the use cases here. I worked on the `lower_unpack` transform to support dynamic sizes because someone in discord said they needed it. And then i saw the NYI comment for `lower_pack` so thought I'd do it ``` "non-static shape NYI, needs a more powerful tensor.expand_shape op" ``` If you never want to support dynamic tiles, then fine by me. But this shouldn't be a NYI comment if you never intend to support it. https://github.com/llvm/llvm-project/pull/76003 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
