%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: float_matmul
func.func @float_matmul(
  %arg0: tensor<1x10xf32>, %arg1: tensor<10x10xf32>) -> (tensor<*xf32>, tensor<*xf32>, tensor<*xf32>) {
  %cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<10xf32>} : () -> tensor<10xf32>
  %0 = "tf.MatMul"(%arg0, %arg1) {
    transpose_a = false, transpose_b = false
  } : (tensor<1x10xf32>, tensor<10x10xf32>) -> tensor<*xf32>

// CHECK:  "tfl.cos"(%arg0) : (tensor<f32>) -> tensor<f32>
}

func.func @elu(%arg0: tensor<11x16xf32>) -> tensor<11x16xf32> {
  %0 = "tf.Elu"(%arg0) : (tensor<11x16xf32>) -> tensor<11x16xf32>
  func.return %0 : tensor<11x16xf32>

// CHECK-LABEL:elu
// CHECK:  "tfl.elu"(%arg0) : (tensor<11x16xf32>) -> tensor<11x16xf32>
}

func.func @expandDims(%arg0: tensor<2x2xf32>, %arg1: tensor<i32>) -> tensor<1x2x2xf32> {

  "tfl.custom"(%0, %2) {custom_code = "FlexAssignVariableOp", custom_option = #tfl<const_bytes : "0x1041737369676E5661726961626C654F70003B121041737369676E5661726961626C654F701A001A002A0B0A056474797065120230012A140A0E76616C69646174655F736861706512022800320000024F3E1414042801">} : (tensor<!tf_type.resource<tensor<1x10xf32>>>, tensor<1x10xf32>) -> ()

  %1 = "tf.MatMul"(%arg0, %arg0) {device = "", transpose_a = false, transpose_b = false} : (tensor<10x10xi32>, tensor<10x10xi32>) -> tensor<10x10xi32>
  %2 = "tf.PreventGradient"(%0) : (tensor<10x10xi32>) -> tensor<10x10xi32>
  %3 = "tf.PreventGradient"(%1) : (tensor<10x10xi32>) -> tensor<10x10xi32>
  %4 = "tf.AddV2"(%2, %3) {device = ""} : (tensor<10x10xi32>, tensor<10x10xi32>) -> tensor<10x10xi32>

  // CHECK: %[[ADD:.*]] = tfl.add %[[VAR_VAL]], %arg0 {fused_activation_function = "NONE"} : tensor<1x10xf32>
  // CHECK: "tfl.assign_variable"(%[[RESOURCE]], %[[ADD]]) : (tensor<!tf_type.resource<tensor<1x10xf32>>>, tensor<1x10xf32>) -> ()
  // CHECK: %[[RESULT:.*]] = "tfl.read_variable"(%[[RESOURCE]]) : (tensor<!tf_type.resource<tensor<1x10xf32>>>) -> tensor<1x10xf32>

  %0 = arith.constant dense<-1> : tensor<1xi32>
  %1 = "tfl.reduce_max"(%arg0, %0) {keep_dims = true} : (tensor<10x10xf32>, tensor<1xi32>) -> tensor<10x1xf32>
  %2 = tfl.sub(%arg0, %1) {fused_activation_function = "NONE"} : (tensor<10x10xf32>, tensor<10x1xf32>) -> tensor<10x10xf32>
  %3 = "tfl.exp"(%2) : (tensor<10x10xf32>) -> tensor<10x10xf32>

  %2 = "tfl.pseudo_const"() {value = dense<[[0.2]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
  %3 = "tfl.pseudo_const"() {value = dense<[[0.3]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
  %4 = "tfl.pseudo_const"() {value = dense<[[0.4]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
  %5 = "tfl.pseudo_const"() {value = dense<[[0.5]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>

  // 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

    %5 = stablehlo.constant dense<2.000000e+02> : tensor<1x1xf32>  // Output inverse scale.
    %6 = stablehlo.constant dense<2> : tensor<1x1xi8>  // Output zero point.
    %7 = call @uniform_quantize_0(%arg0, %0, %1) : (tensor<1x2xf32>, tensor<1x1xf32>, tensor<1x1xi8>) -> tensor<1x2xi8>
    %8 = stablehlo.convert %7 : (tensor<1x2xi8>) -> tensor<1x2xf32>