[Mlir-commits] [mlir] [mlir][linalg] Vectorize directly to a named contraction (PR #147296)

[Andrzej Warzyński]

================
@@ -2445,6 +2445,8 @@ def VectorizeOp : Op<Transform_Dialect, "structured.vectorize",
                        DefaultValuedOptionalAttr<DenseI64ArrayAttr, "{}">:
                           $static_vector_sizes,
                        OptionalAttr<UnitAttr>:$vectorize_nd_extract,
+                       OptionalAttr<UnitAttr>:$flatten1D_depthwise_conv,
----------------
banach-space wrote:

Let me add a bit more context.

I added the option to flatten the depthwise convs as an optimisation for convs with low channel dim count. While great for NEON (i.e. fixed width vectors), it's something that's tricky to generalise to scalable vectors. So I deliberately avoided extending the support ( I am waiting to see whether others find it useful).

I am fine with extending this Op, but if we do, we should also add more tests. Your call :)

https://github.com/llvm/llvm-project/pull/147296