// CHECK-SAME: ([[ARG0:%.*]]: tensor<!tf_type.resource<tensor<10x3xf32>>>, [[ARG1:%.*]]: tensor<!tf_type.resource<tensor<10x3xf32>>>, [[ARG2:%.*]]: tensor<!tf_type.resource<tensor<10x3xf32>>>, [[ARG3:%.*]]: tensor<!tf_type.resource<tensor<10x3xf32>>>)

// WRAPPED-NEXT:   tf_executor.island wraps "tfl.custom_tf"
// WRAPPED-NEXT:     ^bb0(%arg1: tensor<*xf32>, %arg2: tensor<186xf32>, %arg3: tensor<186xf32>):
// WRAPPED-NEXT:   %[[fq:.*]] = "tf.FakeQuantWithMinMaxVarsPerChannel"(%arg1, %arg2, %arg3) <{narrow_range = true, num_bits = 8 : i64}> {device = ""} : (tensor<*xf32>, tensor<186xf32>, tensor<186xf32>) -> tensor<*xf32>

  // MLIR @main function corresponds to the TF function "main_original".
  OwningOpRef<ModuleOp> module_op = ParseModuleOpString(R"mlir(
    func.func private @main(%arg: tensor<1x2xf32>) -> (tensor<1x2xf32>) attributes {tf._original_func_name = "main_original"} {
      return %arg : tensor<1x2xf32>
    }
  )mlir");
  ASSERT_TRUE(module_op);

  absl::flat_hash_map<FunctionName, FunctionAlias> function_aliases;

  %5 = "tf.BiasAdd"(%4, %arg2) : (tensor<1x128x128xf32>, tensor<128xf32>) -> tensor<1x128x128xf32>
  func.return %5 : tensor<1x128x128xf32>

  // CHECK-LABEL: testDilatedConvWithExpandSqueeze1
  // CHECK-SAME: ([[INPUT:%.*]]: tensor<1x128x128xf32>, [[FILTER:%.*]]: tensor<5x5x1x1xf32>, [[BIAS:%.*]]: tensor<128xf32>)
  // CHECK-NEXT: [[AXIS:%.*]] = "tf.Const"() <{value = dense<3> : tensor<i32>}> : () -> tensor<i32>

  %min = arith.constant dense<0.0> : tensor<16xf32>
  %max = arith.constant dense<255.0> : tensor<16xf32>
  %mini = "tf.Identity"(%min) : (tensor<16xf32>) -> tensor<16xf32>
  %maxi = "tf.Identity"(%max) : (tensor<16xf32>) -> tensor<16xf32>
  %fq = "tf.FakeQuantWithMinMaxVarsPerChannel"(%in, %mini, %maxi) {num_bits = 5, narrow_range = false} : (tensor<3x3x3x16xf32>, tensor<16xf32>, tensor<16xf32>) -> tensor<3x3x3x16xf32>

  %min = arith.constant dense<0.0> : tensor<16xf32>
  %max = arith.constant dense<15.0> : tensor<16xf32>
  %mini = "tf.Identity"(%min) : (tensor<16xf32>) -> tensor<16xf32>
  %maxi = "tf.Identity"(%max) : (tensor<16xf32>) -> tensor<16xf32>
  %fq = "tf.FakeQuantWithMinMaxVarsPerChannel"(%in, %mini, %maxi) {num_bits = 3, narrow_range = false} : (tensor<3x3x3x16xf32>, tensor<16xf32>, tensor<16xf32>) -> tensor<3x3x3x16xf32>

      func.func @main(%arg: tensor<1x2xf32> {tf_saved_model.index_path = ["input_tensor:0"]}) -> (tensor<1x2xf32> {tf_saved_model.index_path = ["output_tensor:0"]}) attributes {tf.entry_function = {inputs = "input_tensor:0", outputs = "output_tensor:0"}, tf_saved_model.exported_names = ["main"]} {
        %0 = tf_executor.graph {
          tf_executor.fetch %arg : tensor<1x2xf32>
        }
        return %0 : tensor<1x2xf32>
      }

// CHECK:           %[[VAL_4:.*]]:2 = call @func_0_GPU_FLOAT(%[[VAL_0]], %[[VAL_1]], %[[VAL_2]], %[[VAL_3]]) {tac.device = "GPU", tac.inference_type = "FLOAT", tac.interface_name = "func_0"} : (tensor<1xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xf32>) -> (tensor<1xf32>, tensor<1xf32>)

func.func @bias_adjust_pertensor(%arg0: tensor<1x2xf32>) -> (tensor<1x2xf32>) {
  %0 = "quantfork.stats"(%arg0) {
    layerStats = dense<[-1.28e-5, 1.27e-5]> : tensor<2xf32>
  } : (tensor<1x2xf32>) -> tensor<1x2xf32>
  %w = arith.constant dense<[[0.0, 1.0], [1.0, 2.0]]> : tensor<2x2xf32>
  %b = arith.constant dense<[0.0, 2.1473647e6]> : tensor<2xf32>
  %fc = "tfl.fully_connected"(%0, %w, %b) {

  %cst_weights = arith.constant dense<2.0> : tensor<1024x512xf32>
  %cst_bias = arith.constant dense<4.0> : tensor<1024xf32>

  %0 = "tfl.fully_connected" (%cst_input, %cst_weights, %cst_bias) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<512xf32>, tensor<1024x512xf32>, tensor<1024xf32>) -> tensor<1024xf32>

  func.return %0 : tensor<1024xf32>

  // 1.0 * 2.0 * 512 + 4.0 = 1028.0