%1 = "tf.Transpose"(%arg0, %0) : (tensor<1x4x4x8xf32>, tensor<4xi32>) -> tensor<1x8x4x4xf32>
  %2 = "tf.Transpose"(%1, %0) : (tensor<1x8x4x4xf32>, tensor<4xi32>) -> tensor<1x4x8x4xf32>
  %3 = "tf.Transpose"(%arg1, %0) : (tensor<1x4x4x8xf32>, tensor<4xi32>) -> tensor<1x8x4x4xf32>
  %4 = "tf.Transpose"(%3, %0) : (tensor<1x8x4x4xf32>, tensor<4xi32>) -> tensor<1x4x8x4xf32>

module {
func.func @main(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>, %arg2: tensor<1xf32>, %arg3: tensor<1xf32>) -> tensor<2x1xf32> attributes {tf.entry_function = {inputs = "input0,input1,input2,input3", outputs = "output"}} {
  %0 = "tfl.add"(%arg0, %arg1) {fused_activation_function = "RELU6"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
  %1 = "tfl.mul"(%0, %arg2) {fused_activation_function = "RELU6"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

// -----

func.func @testAddReluPack(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>) {
   // CHECK: tac.device = "GPU", tac.inference_type = "FLOAT"
  %0 = "tfl.add"(%arg0, %arg1) {fused_activation_function = "RELU6"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
   // CHECK: tac.device = "GPU", tac.inference_type = "FLOAT"
  %1 = "tfl.add"(%arg0, %0) {fused_activation_function = "RELU"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

  %1 = "tfl.pseudo_const"() {value = dense<0> : tensor<3xi32>} : () -> tensor<3xi32>
  %2 = "tfl.pseudo_const"() {value = dense<[0, 2, 0]> : tensor<3xi32>} : () -> tensor<3xi32>
  %3 = "tfl.pseudo_const"() {value = dense<1> : tensor<3xi32>} : () -> tensor<3xi32>

  %4 = "tfl.var_handle"() {container = "", shared_name = "read_assign/states0"} : () -> tensor<!tf_type.resource>

  %0 = "tfr.cast"(%arg0) : (tensor<*xf32>) -> !tfr.tensor
  %1 = "tfr.cast"(%0) : (!tfr.tensor) -> tensor<*xi32>
  %2 = "tfr.cast"(%0) : (!tfr.tensor) -> tensor<2xi32>
  func.return %1, %2 : tensor<*xi32>, tensor<2xi32>

// CHECK: %[[tf_cast_unranked:.*]] = "tf.Cast"(%arg0) <{Truncate = false}> : (tensor<*xf32>) -> tensor<*xi32>

// PerChannel-DAG: %[[zp1:.*]] = "tf.Const"() <{value = dense<{{[0-9]+}}> : tensor<3xi32>}> : () -> tensor<3xi32>
// PerChannel-DAG: %[[zp2:.*]] = "tf.Const"() <{value = dense<{{[0-9]+}}> : tensor<6xi32>}> : () -> tensor<6xi32>
// PerChannel: %[[out_1:.*]] = "tf.PartitionedCall"(%arg0, %[[q_w1]], %[[scale1]], %[[zp1]]) <{config = "", config_proto = "", executor_type = "",

}

// CHECK-LABEL: add
func.func @add(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>) -> tensor<1xf32> {
  %0 = "tf.AddV2"(%arg0, %arg1) : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
  func.return %0: tensor<1xf32>
// CHECK: %[[ADD_0:.*]] = "tf.AddV2"(%arg0, %arg1) : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
// CHECK: return %[[ADD_0]] : tensor<1xf32>
}

// CHECK-LABEL: softmax

func.func private @simple_callee(%arg0: tensor<2xi32>) -> tensor<2xi32> {
  %graph = tf_executor.graph {
    tf_executor.fetch %arg0 : tensor<2xi32>
  }
  func.return %graph#0 : tensor<2xi32>
}

// Test with a TPUPartitionedCallOp that is not inlined.
func.func @dont_inline_tpu_partitioned_call(%arg0: tensor<2xi32> {tf._user_specified_name = "inputs_0"}, %arg1: tensor<2xi32> {tf._user_specified_name = "inputs_1"}) -> tensor<2xi32> {

// CHECK-NOT: tf.BatchFunction
// CHECK: %[[ADD0:.*]] = "tf.AddV2"(%[[ARG0]], %[[ARG1]])
// CHECK: return %[[ADD0]] : tensor<1xf32>

func.func private @batched_func(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>) -> tensor<1xf32> {
  %0 = "tf.Identity"(%arg0) : (tensor<1xf32>) -> tensor<1xf32>
  %1 = "tf.Identity"(%arg1) : (tensor<1xf32>) -> tensor<1xf32>

#loc3])\0Amodule {\0A  func.func @host_func(%arg0: tensor<3360x?xi32> loc(fused[#loc1, #loc2, #loc3]), %arg1: tensor<3xi32> loc(fused[#loc1, #loc2, #loc3]), %arg2: tensor<i32> loc(fused[#loc1, #loc2, #loc3]), %arg3: tensor<2xi32> loc(fused[#loc1, #loc2, #loc3])) -> (tensor<1x1120x?xi32>, tensor<1x1120x?xi32>, tensor<1120x?xi32>, tensor<2xi32>) {\0A    %0 = \22tf.Reshape\22(%arg0, %arg1) {_xla_outside_compilation = \220\22} : (tensor<3360x?xi32>, tensor<3xi32>) -> tensor<3x1120x?xi32> loc(#loc9)\0A    %1:3 = ...