[Mlir-commits] [mlir] [mlir][ArmSME] Add initial SME vector legalization pass (PR #79152)

[Benjamin Maxwell]

================
@@ -122,4 +122,23 @@ def TileAllocation
   let dependentDialects = ["func::FuncDialect"];
 }
 
+def VectorLegalization
+  : Pass<"arm-sme-vector-legalization", "mlir::ModuleOp"> {
+  let summary = "Legalize vectors for ArmSME";
+  let description = [{
+    This pass legalizes vector operations so that they can be lowered to ArmSME.
+    This includes decomposing operations that operate on vector types larger
+    than a single SME tile (e.g. `vector<[8]x[8]xf32>`) into multiple SME
----------------
MacDue wrote:

That matches the size of ZA. There's infinite bigger vector types (this pass does not care how many tiles the resulting type uses), and a few smaller sizes. There may also be cases when you want sizes smaller than ZA (but more than a tile), i.e. a matmul with a rather small width/height but a large reduction dimension. This pass allows the tiling to be fairly flexible.  

https://github.com/llvm/llvm-project/pull/79152