Andrzej Warzyński

> It is %3[%c0, %8, %c2], while I think it perhaps should be %3[%c0, %8, %c0]. Looking at the original tensor.extract, the vectoriser seems to be doing the right thing...

Taking a quick look at the vectoriser and this particular case should be a gather load already (I think that the presence of arith.subi is the key bit in the...

> Any idea what I may be missing? @mariecwhite Could you check whether you have this change: * https://github.com/llvm/llvm-project/pull/70039 (see the update in [LowerVectorContract.cpp](https://github.com/llvm/llvm-project/pull/70039/files#diff-773c2c23e70424be255025cf7a01f7913f22763f79e56abd3c0f0e64c56a6666)). > We did find that using...

> I did not get a chance on ARM side. I’d offer to help, but am OOO. I am also worried about the SVE and SME lowerings. In particular, for...

> are we talking about global transposes on the LHS of matmuls being required on SVE? That should be similar to what we do elsewhere, so I guess local? (sorry,...

> You should be really using data tiling to do this. You can then pick the inner tile sizes appropraitetly for SME or SVE Sure, but from what I can...

Tests for `linalg.mmt4d` in MLIR: * https://github.com/llvm/llvm-project/pull/81422 * https://github.com/llvm/llvm-project/pull/81790 * https://github.com/llvm/llvm-project/pull/84964 Masked + scalable vectoristion of `linalg.mmt4d`: * https://github.com/llvm/llvm-project/pull/83827 * https://github.com/llvm/llvm-project/pull/81906

## Supporting DT in the context of scalable vectors (SVE) and matrices (SME) Key steps in Data Tiling that will require updating in order to support SVE or SME. We...

Thanks for the feedback 🙏🏻 For those of you who missed it, we discussed the `linalg.mmt4d` part of this proposal in the mai-tai call yesterday (March 26th). Based on that...

Thanks for taking a look @hanhanW ! > The `linalg.mmt4d` and `tensor.pack` will be formed in different dispatches, which means that they will be launched in different kernels. Do we...