[Mlir-commits] [mlir] [mlir][x86] Lower vector.contract to packed type tiled dot-product. (PR #182810)

[Arun Thangamani]

arun-thmn wrote:

> High-level question, how does this functionality compare to the `convert-vector-to-amx` pass? Is there any reuse or overlap between the two? This set of rewrites seems more comprehensive, perhaps the older pass should be retired altogether?

Our lowering pattern does more than the `convert-vector-to-amx`, we support lowering of `batch/batch-reduce` matmul, where as `convert-vector-to-amx` is restricted only to `matmul`.

Also, we do optimized lowering of `vector.contract` if it is enclosed within `reduction` loops, `convert-vector-to-amx` focused on individual contracts ends-up in completed code.

https://github.com/llvm/llvm-project/pull/182810