// CHECK-LABEL: testFoldEnsureShapeOp
func.func @testFoldEnsureShapeOp(%arg0: tensor<10x20xf32>) -> (tensor<10x20xf32>, tensor<10x20xf32>, tensor<20x10xf32>) {
  %0 = "tf.EnsureShape"(%arg0) {shape = #tf_type.shape<10x20>} : (tensor<10x20xf32>) -> tensor<10x20xf32>
  %1 = "tf.EnsureShape"(%arg0) {shape = #tf_type.shape<?x20>} : (tensor<10x20xf32>) -> tensor<10x20xf32>
  // Failing case which should not be folded.

  %3 = "tf.Add"(%2, %arg1) : (tensor<?x8x10xf32>, tensor<8x10xf32>) -> tensor<?x8x10xf32>
  %4 = "tf.Add"(%2, %arg2) : (tensor<?x8x10xf32>, tensor<8x10xf32>) -> tensor<?x8x10xf32>
  %5 = "tf.Add"(%arg1, %arg2) : (tensor<8x10xf32>, tensor<8x10xf32>) -> tensor<8x10xf32>

  // CHECK-LABEL: func @parallel_execute_with_tiled_input
  // CHECK-SAME: (%[[ARG_0:[a-z0-9]*]]: tensor<128x10xf32>, %[[ARG_1:[a-z0-9]*]]: tensor<128x10xf32>, %[[ARG_2:[a-z0-9]*]]: tensor<*xi32>, %[[ARG_3:[a-z0-9]*]]: tensor<*xi32>)
  func.func @parallel_execute_with_tiled_input(%arg0: tensor<128x10xf32>, %arg1: tensor<128x10xf32>, %arg2: tensor<*xi32>, %arg3: tensor<*xi32>) -> (tensor<*xi32>, tensor<*xi1>) {
    // CHECK: tf_device.replicate

}

// -----

// CHECK-LABEL: func @testBatchMatMulV2NoBatchDimension
func.func @testBatchMatMulV2NoBatchDimension(%lhs: tensor<5x10xf32>, %rhs: tensor<10x10xf32>) -> (tensor<5x10xf32>) {
  %0 = "tf.BatchMatMulV2"(%lhs, %rhs) : (tensor<5x10xf32>, tensor<10x10xf32>) -> tensor<5x10xf32>
  func.return %0 : tensor<5x10xf32>
}

// -----

}

func.func @matmul_batch(%arg0: tensor<10x15xf32>, %arg1: tensor<15x17xf32>) -> tensor<10x17xf32> {
  %0 = "tf.BatchMatMul"(%arg0, %arg1) {T = "tfdtype$DT_FLOAT", device = "/device:CPU:0", name = "MatMul", adj_x = false, adj_y = false} :
(tensor<10x15xf32>, tensor<15x17xf32>) -> tensor<10x17xf32>
  func.return %0 : tensor<10x17xf32>
// CHECK-LABEL: matmul_batch

}

// -----

func.func @fully_connected_no_bias(%arg0: tensor<2x2x10xf32>, %arg1: tensor<40x40xf32>, %arg2: none) -> tensor<1x40xf32> {
  %0 = "tfl.fully_connected"(%arg0, %arg1, %arg2) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<2x2x10xf32>, tensor<40x40xf32>, none) -> tensor<1x40xf32>
  func.return %0 : tensor<1x40xf32>
}

// -----

    }
  }
  return new_arg_type;
}

// Combination of value producer and port of value produced (e.g.,
//   <value result output>:<value in output tensor>,
// so for tf.Const -> tensor<10x20xf32>, [0,2,18] would point to a unique output
// scalar value).
struct ValuePort {
  PointerUnion<Operation*, BlockArgument> producer;
  SmallVector<unsigned int, 2> port;

    func.return %2 : tensor<1xi32>
  }

  // CHECK-LABEL: func @tensor_cast_partial_infer
  func.func @tensor_cast_partial_infer(%arg0: tensor<?x10xi32>) -> tensor<10x?xi32> {
   // CHECK: return {{.*}} : tensor<10x10xi32>
    %2 = tensor.cast %arg0 : tensor<?x10xi32> to tensor<10x?xi32>
    func.return %2 : tensor<10x?xi32>
  }

  // CHECK-LABEL: operand_pack_unranked