[Mlir-commits] [mlir] [mlir][ArmSME] Add support for vector.transpose (PR #66760)

[Cullen Rhodes]

================
@@ -215,3 +219,121 @@ func.func @broadcast_vec2d_from_vec1d(%arg0: vector<[8]xi16>) {
   "prevent.dce"(%0) : (vector<[8]x[8]xi16>) -> ()
   return
 }
+
+// =============================================================================
+// vector.transpose
+// =============================================================================
+
+// -----
+
+// CHECK-LABEL:   func.func @transpose_i8(
+// CHECK-SAME:                            %[[TILE:.*]]: vector<[16]x[16]xi8>)
+// CHECK:           %[[C16:.*]] = arith.constant 16 : index
+// CHECK:           %[[C0:.*]] = arith.constant 0 : index
+// CHECK:           %[[VSCALE:.*]] = vector.vscale
+// CHECK:           %[[MIN_TILE_SLICES:.*]] = arith.muli %[[VSCALE]], %[[C16]] : index
+// CHECK:           %[[NUM_TILE_SLICES:.*]] = memref.alloca(%[[MIN_TILE_SLICES]], %[[MIN_TILE_SLICES]]) : memref<?x?xi8>
+// CHECK:           arm_sme.tile_store %[[TILE]], <hor>, %[[NUM_TILE_SLICES]]{{\[}}%[[C0]], %[[C0]]] : memref<?x?xi8>, vector<[16]x[16]xi8>
+// CHECK:           arm_sme.tile_load <ver>, %[[NUM_TILE_SLICES]]{{\[}}%[[C0]], %[[C0]]] : memref<?x?xi8>, vector<[16]x[16]xi8>
+func.func @transpose_i8(%arg0: vector<[16]x[16]xi8>) {
+  %0 = vector.transpose %arg0, [1, 0] : vector<[16]x[16]xi8> to vector<[16]x[16]xi8>
+  "prevent.dce"(%0) : (vector<[16]x[16]xi8>) -> ()
+  return
+}
+
+// -----
+
+// CHECK-LABEL: @transpose_i16
+// CHECK: arith.constant 8
+// CHECK: arm_sme.tile_store
+// CHECK: arm_sme.tile_load
----------------
c-rhodes wrote:

Done

https://github.com/llvm/llvm-project/pull/66760