// FB-EMPTY:
  // FB-NEXT:             }
  // FB-NEXT:       } ],

  // CHECK-LABEL: func @main(%arg0: tensor<40x37xf32>, %arg1: tensor<40x37xf32>)
  // CHECK-SAME:      -> tensor<40x40xf32>
  // CHECK:         %[[stat:.*]] = "quantfork.stats"(%arg0) <{layerStats = dense
  // CHECK-SAME:      <[-1.000000e+00, 1.000000e+00]> : tensor<2xf32>}>
  // CHECK-SAME:      : (tensor<40x37xf32>) -> tensor<40x37xf32>

func.func @testReadVariableOpOfCast(%arg0: tensor<!tf_type.resource<tensor<8x40xf32>>>) -> tensor<8x40xf32> {
  %0 = "tf.Cast"(%arg0) : (tensor<!tf_type.resource<tensor<8x40xf32>>>) -> tensor<*x!tf_type.resource>
  %1 = "tf.ReadVariableOp"(%0) {device = "/job:localhost/replica:0/task:0/device:GPU:0"} : (tensor<*x!tf_type.resource>) -> tensor<8x40xf32>
  func.return %1: tensor<8x40xf32>

// CHECK-SAME: (%[[ARG0:.*]]: tensor<3x4xf32>)
func.func @expm1(%arg0: tensor<3x4xf32>) -> tensor<3x4xf32> {
  %0 = "tf.Expm1"(%arg0) : (tensor<3x4xf32>) -> tensor<3x4xf32>
  func.return %0 : tensor<3x4xf32>
  // CHECK: %[[ONE:.*]] = "tf.Const"() <{value = dense<1.000000e+00> : tensor<f32>}> : () -> tensor<f32>
  // CHECK: %[[EXP:.*]] = "tf.Exp"(%[[ARG0]]) : (tensor<3x4xf32>) -> tensor<3x4xf32>

func.func @RemoveFcZeroBias(%arg0: tensor<1x37xf32>, %arg1: tensor<40x37xf32>) -> tensor<1x40xf32> {
  %0 = "tfl.pseudo_const"() {value = dense<0.0> : tensor<40xf32>} : () -> tensor<40xf32>
  %1 = "tfl.fully_connected"(%arg0, %arg1, %0) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<1x37xf32>, tensor<40x37xf32>, tensor<40xf32>) -> tensor<1x40xf32>
// CHECK: "tfl.fully_connected"

    %5 = "quantfork.qcast"(%4) {volatile} : (tensor<3x4xf32>) -> tensor<3x4x!quant.uniform<i8:f32, 0.13170163023705575:-1>>
    %6 = "quantfork.dcast"(%5) : (tensor<3x4x!quant.uniform<i8:f32, 0.13170163023705575:-1>>) -> tensor<3x4xf32>
    %7 = stablehlo.slice %6 [1:3, 2:4] : (tensor<3x4xf32>) -> tensor<2x2xf32>

//CHECK-NEXT:}

func.func @iota() -> tensor<3x4xf32> {
 %0 = "vhlo.iota_v1" () <{iota_dimension = #vhlo.integer_v1<0 : i64>}> : () -> tensor<3x4xf32>
 return %0 : tensor<3x4xf32>
}

//CHECK:func.func private @iota() -> tensor<3x4xf32> {
//CHECK-NEXT: %0 = "vhlo.iota_v1"() <{iota_dimension = #vhlo.integer_v1<0 : i64>}> : () -> tensor<3x4xf32>
//CHECK-NEXT: return %0 : tensor<3x4xf32>
//CHECK-NEXT:}

  %cst = "tf.Const"() {value = dense<1.000000e+00> : tensor<3x4xf32>} : () -> tensor<3x4xf32>
  %cst_0 = "tf.Const"() {value = dense<0> : tensor<i32>} : () -> tensor<i32>
  %0 = "tf.Abs"(%cst) {device = ""} : (tensor<3x4xf32>) -> tensor<3x4xf32>
  %1 = "tf.Max"(%0, %cst_0) {device = "", keep_dims = false} : (tensor<3x4xf32>, tensor<i32>) -> tensor<4xf32>
  %2 = "tf.Neg"(%1) {device = ""} : (tensor<4xf32>) -> tensor<4xf32>

  %0 = "tfl.unidirectional_sequence_lstm"(%arg0, %arg1, %arg2, %arg3, %arg4, %arg5, %arg6, %arg7, %arg8, %arg9, %arg10, %arg11, %arg12, %arg13, %arg14, %arg15, %arg16, %arg17, %arg18, %arg19, %arg20, %arg21, %arg22, %arg23) {fused_activation_function...

  func.func @non_max_suppression_v4(%arg0: tensor<3x4xf32>, %arg1: tensor<3xf32>, %arg2: tensor<f32>, %arg3: tensor<f32>) -> tensor<2xi32> {
    %max_size = mhlo.constant dense<2> : tensor<i32>
    // CHECK-NOT: tf.NonMaxSuppressionV4
    %0:2 = "tf.NonMaxSuppressionV4"(%arg0, %arg1, %max_size, %arg2, %arg3) {pad_to_max_output_size = true}: (tensor<3x4xf32>, tensor<3xf32>, tensor<i32>, tensor<f32>, tensor<f32>) -> (tensor<2xi32>, tensor<i32>)

    // CHECK-NOT: tf.TPUPartitionedOutputV2
    %partitioned_output:2 = "tf.TPUPartitionedOutputV2"(%computation) {_XlaSharding = "\08\03\1A\02\01\02\22\02\00\01", partition_dims = [1, 2]} : (tensor<3x4xf32>) -> (tensor<3x2xf32>, tensor<3x2xf32>)
    // CHECK: "tf.AssignVariableOp"(%arg0, %[[PARALLEL_EXECUTE_OUTPUT]]#0)