[all-commits] [llvm/llvm-project] 833166: [mlir][xegpu] Distribute three more convert_layout...

[Sang Ik Lee via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 833166e22fe20c3513a77a02e34385c16128ea31
      https://github.com/llvm/llvm-project/commit/833166e22fe20c3513a77a02e34385c16128ea31
  Author: Sang Ik Lee <sang.ik.lee at intel.com>
  Date:   2026-09-24 (Thu, 24 Sep 2026)

  Changed paths:
    M mlir/lib/Dialect/XeGPU/Transforms/XeGPUSgToLaneDistribute.cpp
    A mlir/test/Dialect/XeGPU/sg-to-lane-distribute-convert-layout-slice.mlir
    M mlir/test/Integration/Dialect/XeGPU/WG/simple_mxfp_gemm_quantizeA_F4.mlir

  Log Message:
  -----------
  [mlir][xegpu] Distribute three more convert_layout cases to lanes (#222693)

`xegpu-sg-to-lane-distribute` could not lower an `xegpu.convert_layout`
whose effective `lane_layout` changes in a way that moves elements
between lanes. Add one pattern per case found in the workgroup-level
mx-fp GEMM quantize-A kernels, where `#xegpu.slice` layouts of a rank-3
layout meet the rank-2 layouts of `load_nd` and `dpas_mx`:

- fully broadcast source to one row per lane: `vector.extract`
- source broadcast over two lane groups to one row per lane:
`vector.extract` plus two `gpu.shuffle`
- fully broadcast source to one column per lane group:
`vector.deinterleave` plus a select on the lane group



To unsubscribe from these emails, change your notification settings at https://github.com/llvm/llvm-project/settings/notifications