[Mlir-commits] [mlir] [mlir][ArmNeon] Implements unrolling patterns for LowerContractionToSMMLAPattern (PR #84848)

[Diego Caballero]

================
@@ -36,19 +38,17 @@ static Type matchContainerType(Type element, Type container) {
   return element;
 }
 
-/// Lowering from a single vector::contractOp directly to the arm neon smmla
-/// intrinsic. The shapes of the contract and intrinsic must match.
+/// Lowering from a vector::contractOp arm neon smmla intrinsic. This up to an
+/// 8x8x8 vector contract that is tiled (up to 16) smmla instructions with
----------------
dcaballe wrote:

Could we make this support any cases above the minimum 2, 2, 8 as long as the dimensions are multiple of that base cases?

https://github.com/llvm/llvm-project/pull/84848