[Mlir-commits] [mlir] [mlir][x86vector] Shuffle BF16 vector.contract output for Flat layout. (PR #174590)

[Arun Thangamani]

================
@@ -258,14 +258,84 @@ func.func @matmul_outer_product_to_bf16dp_bcst_B(
   return %0 : !vecC
 }
 
-// CHECK-LABEL: @matmul_outer_product_to_bf16dp_bcst_B
+// CHECK-LABEL: @matmul_outer_product_to_int8dp
 // CHECK: vector.broadcast
+// CHECK: x86vector.avx.dot.i8
+
+module attributes {transform.with_named_sequence} {
+  transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
+    %func = transform.structured.match ops{["func.func"]} in %arg1 : (!transform.any_op) -> !transform.any_op
+    transform.apply_patterns to %func {
+      transform.apply_patterns.x86vector.vector_contract_to_packed_type_dot_product
+    } : !transform.any_op
+    transform.yield
+  }
+}
+
+// -----
+
+!vecA = vector<1x2xbf16>
+!vecB = vector<2x16xbf16>
----------------
arun-thmn wrote:

No, this transform should be called along with pass to `BF16 dp lowering` or `vector.fma` lowering.

https://github.com/llvm/llvm-project/pull/174590