Lubomir Litchev

@MaheshRavishankar Thank you so much for the thoughtfull comments! I really appreciate it! I have had discussed these with Diego and following are some of the highlights of the conversations...

@MaheshRavishankar Thanks for the comments! First, this turned out longer than I expected. Sorry! We can discuss it more in Mai-Tai... 1. Sorry, my understanding (and Diego's from what I...

@MaheshRavishankar Reverted the files. The flag controlling the use of TopK vectorization was originally in these files, but limitations to the current implementation required moving it in KernelDispatch. I thought...

@hanhanW No worries! I know it is a lot of code... I have a PR request for making the createReadOrMaskedRead a utility function in MLIR - https://github.com/llvm/llvm-project/pull/89119 Adding LitTests at...

I would add the LitTest and make the PR official as well. They will provide the before and after code... I'll paste it here for a small testcase of 1x32xf32...

The codegen basically replaces the topk operation with the codegen I described above... // -----// IR Dump Before GeneralizeLinalgNamedOps (iree-global-opt-generalize-linalg-named-ops) //----- // util.func public @custom_call_topk_test_32() -> (!hal.buffer_view, !hal.buffer_view) attributes {iree.abi.stub,...

RFC issue: https://github.com/iree-org/iree/issues/17143

@benvanik Thanks! Yeah, I noticed after pushing it and fixing the diff (checkout before the rebase). It was a surprise that the rebase did what it did... I did update...

Made changes to the original PR by: 1. Moved the SCF lowering of TopK out of the GenericVectorize to a new pass TopkLowering. a. A new StringAttr was introduced to...

@MaheshRavishankar Whenever you get a chance. Thank you very much!!