[Mlir-commits] [mlir] [mlir][linalg] Pack matmul pass (PR #89782)

[Adam Siemieniuk]

adam-smnk wrote:

This is a packing scheme aimed at CPUs to improve performance of GEMM computations by improving data access pattern.
The pass is primarily a driver of the existing upstream packing and tiling utilities with CPU-centric decisions.

The transformation has been tested within our downstream project ([TPP-MLIR](https://github.com/plaidml/tpp-mlir)).
This is the first PR which aims to bring our Linalg tensor-level transformations upstream. Further follow-up will include pack and unpack propagation to amortize the cost of packing and VNNI layout packing which allows targeting type-aware SIMD instructions.

https://github.com/llvm/llvm-project/pull/89782