%1 = "tf.PartitionedCall"(%arg0, %cst_0) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_matmul_fn_1} : (tensor<2x12xf32>, tensor<12x2xf32>) -> tensor<*xf32>
    func.return %1: tensor<*xf32>
  }

module {
func.func @norm1(%arg0: tensor<1x128x128xf32>) -> tensor<1x128x128xf32> {
  %0 = "tfl.pseudo_const"() {value = dense<1.000000e+00> : tensor<128xf32>} : () -> tensor<128xf32>
  %1 = "tfl.add"(%arg0, %0) {tac.device = "GPU", tac.inference_type = "FLOAT", fused_activation_function = "NONE"} : (tensor<1x128x128xf32>, tensor<128xf32>) -> tensor<1x128x128xf32>
  %2 = "tfl.pseudo_const"() {value = dense<[128, 128]> : tensor<2xi32>} : () -> tensor<2xi32>

    %recurrent_stats = "quantfork.stats"(%recurrent_input) {layerStats = dense<[-2.0, 1.0]> : tensor<2xf32>} : (tensor<1x20xf32>) -> tensor<1x20xf32>
    %cell_input = arith.constant dense<1.0> : tensor<1x20xf32>
    %cell_stats = "quantfork.stats"(%cell_input) {layerStats = dense<[-2.73090601, 7.94872093]> : tensor<2xf32>} : (tensor<1x20xf32>) -> tensor<1x20xf32>

  // CHECK-NEXT:    "vhlo.return_v1"(%1) : (tensor<f32>) -> ()
  // CHECK-NEXT: }) : (tensor<1x16x16x320xf32>, tensor<f32>) -> tensor<1x320xf32>
  %0 = stablehlo.reduce(%arg0 init: %arg1) applies stablehlo.add across dimensions = [1, 2] : (tensor<1x16x16x320xf32>, tensor<f32>) -> tensor<1x320xf32>
  return %0 : tensor<1x320xf32>
}

// -----

// CHECK-LABEL: op_with_region_mixed_tfl_shlo_tfl

    %9:4 = "tf.CustomAggregator"(%6) {calibration_method = 1 : i32, id = "1", max_percentile = 0.000000e+00 : f32, min_percentile = 0.000000e+00 : f32, num_bins = 0 : i32} : (tensor<1x64xf32>) -> (tensor<1x64xf32>, tensor<f32>, tensor<f32>, tensor<*xi64>)
    return %9#0 : tensor<1x64xf32>
  }

  %0 = "tf_device.cluster_func"(%arg0, %arg1) {func = @_func, use_spmd_for_xla_partitioning = true, use_tpu = true, num_cores_per_replica = 2 : i64} : (tensor<2x4xf32>, tensor<4x2xf32>) -> tensor<2x2xf32>
  %1:2 = "tf.TPUPartitionedOutputV2"(%0) {device = "", partition_dims = [2, 1]} : (tensor<2x2xf32>) -> (tensor<1x2xf32>, tensor<1x2xf32>)

// CHECK-LABEL: func @fold_into_mean
func.func @fold_into_mean(%arg0: tensor<1x64x112x112xf32>) -> tensor<1x64xf32> {

  // CHECK: %[[RED_IDX:.*]] = "tf.Const"() <{value = dense<[2, 3]> : tensor<2xi32>}>
  // CHECK: %[[MEAN:[0-9]*]] = "tf.Mean"(%arg0, %[[RED_IDX]])
  // CHECK-SAME: (tensor<1x64x112x112xf32>, tensor<2xi32>) -> tensor<1x64xf32>
  // CHECK: return %[[MEAN]]

// CHECK-LABEL: func @testCase(%arg0: tensor<i32>, %arg1: tensor<!tf_type.resource<tensor<1x2x3xf32>>>)
func.func @testCase(%arg0: tensor<i32>, %arg1: tensor<!tf_type.resource<tensor<1x2x3xf32>>>) -> tensor<1x2x3xf32> {
  %0 = "tf.Case"(%arg0, %arg1) {branches = [@branch_0, @branch_1], is_stateless = false} : (tensor<i32>, tensor<!tf_type.resource<tensor<1x2x3xf32>>>) -> tensor<1x2x3xf32>

  // CHECK: [[Result0:%.*]] = "tf.CaseRegion"

// RUN: tf-quant-opt %s -split-input-file -quant-merge-duplicate-resource-ops | FileCheck %s

func.func @merge_duplicate_variable(%arg0: tensor<1x20xf32>, %arg1: tensor<!tf_type.string>) -> (tensor<20x4096xf32>) {
  %0 = tf_executor.graph {
    %outputs_5, %control_6 = tf_executor.island wraps "tf.Const"() {value = dense<""> : tensor<1x!tf_type.string>} : () -> tensor<1x!tf_type.string>

  func.return %24 : tensor<1x4xf32>
}