%cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<10xf32>} : () -> tensor<10xf32>
    %cst_0 = "tf.Const"() {value = dense<[-1, 10]> : tensor<2xi32>} : () -> tensor<2xi32>
    %1 = "tf.MatMul"(%arg0, %arg1) {
      transpose_a = false, transpose_b = false

// -----

// 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>

    %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>

    "vhlo.return_v1" (%421) : (tensor<1xf32>) -> ()
   }) : (tensor<1x16x16x320xf32>, tensor<f32>) -> tensor<1x320xf32>
  return %0 : tensor<1x320xf32>
}

//CHECK:func.func private @reduce(%arg0: tensor<1x16x16x320xf32>, %arg1: tensor<f32>) -> tensor<1x320xf32> {
//CHECK-NEXT:  %0 = "vhlo.reduce_v1"(%arg0, %arg1) <{dimensions = #vhlo.tensor_v1<dense<[1, 2]> : tensor<2xi64>>}> ({ 

// 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 %0#0 : tensor<1x96xf32>

// CHECK: return %[[identity]] : tensor<1x3x2x2xf32>

func.func @cast_bf16_matmul_to_fp32(%arg0: tensor<1x10xf32>) -> (tensor<1x2xf32>) {
  %cst = "tf.Const"() {device = "", value = dense<1.000000e+01> : tensor<10x2xbf16>} : () -> tensor<10x2xbf16>
  %0 = "tf.Cast"(%arg0) {Truncate = false, device = ""} : (tensor<1x10xf32>) -> tensor<1x10xbf16>

    }) {dimensions = dense<1> : tensor<1xi64>} : (tensor<1x256xf32>, tensor<f32>) -> tensor<1xf32>
    %5  = mhlo.add %3, %arg6 : tensor<1xf32>
    "mhlo.return"(%1, %arg3, %arg4, %arg5, %5) : (tensor<i32>, tensor<i32>, tensor<i32>, tensor<1x256xf32>, tensor<1xf32>) -> ()
  }) : (tensor<i32>, tensor<i32>, tensor<i32>, tensor<1x256xf32>, tensor<1xf32>) -> (tensor<i32>, tensor<i32>, tensor<i32>, tensor<1x256xf32>, tensor<1xf32>)

    %1196 = "tfl.reshape"(%arg0, %arg1) : (tensor<?x10x1xf32>, tensor<6xi32>) -> tensor<1x?x1x10x1x1xf32>
    %1197 = "tfl.broadcast_to"(%1196, %arg2) : (tensor<1x?x1x10x1x1xf32>, tensor<6xi32>) -> tensor<1x?x1x10x5x1xf32>
    %1198 = "tfl.reshape"(%1197, %arg3) : (tensor<1x?x1x10x5x1xf32>, tensor<2xi32>) -> tensor<?x50xf32>
    return %1198 : tensor<?x50xf32>

  %2 = "tf.Add"(%1, %1) {
    _xla_compile_device_type = "TPU", _replication_info = "cluster",
    device = "/task:0/device:TPU:0", dtype = f32
  } : (tensor<1x80xf32>, tensor<1x80xf32>) -> tensor<1x80xf32>
  %3 = "tf.ResourceGatherNd"(%arg0, %0) {
    Tindices = i32
  } : (tensor<*x!tf_type.resource<tensor<80xf32>>>, tensor<i32>) -> tensor<1x80xf32>