%2 = "tf.Cast"(%1) {Truncate = false} : (tensor<1x2xbf16>) -> tensor<1x2xf32>
  %3 = "tf.IdentityN"(%2) {device = ""} : (tensor<1x2xf32>) -> tensor<1x2xf32>
  return %3 : tensor<1x2xf32>
}

// CHECK: func @cast_bf16_matmul_to_fp32
// CHECK-DAG: %[[cst:.*]] = "tf.Const"() <{value = dense<{{.*}}> : tensor<10x2xf32>}> : () -> tensor<10x2xf32>
// CHECK: %[[matmul:.*]] = "tf.MatMul"(%arg0, %[[cst]])

    %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>

^bb0(%arg0: tensor<1x384xf32>, %arg1: tensor<1x96xf32>, %arg2: tensor<384x480xf32>, %arg3: tensor<384xf32>, %arg4: tensor<1x96xf32>):
  %0:4 = "tfl.basic_lstm"(%arg0, %arg1, %arg2, %arg3, %arg4) {fused_activation_function = "RELU", cell_clip = 1.0 : f32, proj_clip = 2.0 : f32} : (tensor<1x384xf32>, tensor<1x96xf32>, tensor<384x480xf32>, tensor<384xf32>, tensor<1x96xf32>) -> (tensor<1x96xf32>, tensor<1x96xf32>, tensor<1x480xf32>, tensor<1x384xf32>)

  %8 = "tfl.concatenation"(%2, %0) {axis = -1 : i32, fused_activation_function = "NONE"} : (tensor<1x1xf32>, tensor<1x1xf32>) -> tensor<1x2xf32>
  %9 = "quantfork.stats"(%8) {layerStats = dense<[-0.488159984, 0.189515018]> : tensor<2xf32>} : (tensor<1x2xf32>) -> tensor<1x2xf32>
  %10 = "tfl.concatenation"(%9, %7) {axis = -1 : i32, fused_activation_function = "NONE"} : (tensor<1x2xf32>, tensor<1x2xf32>) -> tensor<1x4xf32>

    func.func @main() -> (tensor<1x4x4xf32>) {
      %%arg0 = "tf.Const"() {value = dense<-3.0> : tensor<1x4x2xf32>} : () -> tensor<1x4x2xf32>
      %%arg1 = "tf.Const"() {value = dense<-3.0> : tensor<1x2x4xf32>} : () -> tensor<1x2x4xf32>

      %%1 = "tf.%s"(%%arg0, %%arg1) {T = f32, adj_x = false, adj_y = false, grad_x = false, grad_y = false, device = ""} : (tensor<1x4x2xf32>, tensor<1x2x4xf32>) -> tensor<1x4x4xf32>

//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:}

  %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>

  } : (tensor<9x10x!tf_type.qint8>, tensor<f32>, tensor<i32>) -> tensor<9x10xf32>
  %0 = "tf.MatMul"(%input, %filter_new) {
  } : (tensor<8x9xf32>, tensor<9x10xf32>) -> tensor<8x10xf32>
  return %0 : tensor<8x10xf32>
})mlir";
  constexpr absl::string_view kProgram = R"mlir(
func.func @main(%input: tensor<8x9xf32>, %filter: tensor<9x10xi8>) -> tensor<8x10xf32> {
  %filter_scale = "tf.Const"() { value = dense<0.0235> : tensor<f32> } : ()

%arg17, %arg18, %arg19, %arg20, %arg21, %arg22, %arg23) <{fused_activation_function = "NONE", time_major = false}> : (tensor<600x?x20xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<40xf32>, tensor<?x?xf32>, tensor<?xf32>, tensor<600x40xf32>, tensor<600x40xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>)...

  %4 = "tf.MatMul"(%arg0, %3) {device = "", transpose_a = false, transpose_b = false} : (tensor<2x3xf32>, tensor<3x4xf32>) -> tensor<2x4xf32>
  %5 = "tf.Identity"(%4) {device = ""} : (tensor<2x4xf32>) -> tensor<2x4xf32>
  %6 = "tf.Identity"(%5) {device = ""} : (tensor<2x4xf32>) -> tensor<2x4xf32>
  func.return %6 : tensor<2x4xf32>