%2 = "tf.Cast"(%1) {Truncate = false} : (tensor<1x1x2xbf16>) -> tensor<1x1x2xf32>
  %3 = "tf.IdentityN"(%2) {device = ""} : (tensor<1x1x2xf32>) -> tensor<1x1x2xf32>
  return %3 : tensor<1x1x2xf32>
}

// CHECK: func @cast_bf16_batch_matmul_v2_to_fp32
// CHECK-DAG: %[[cst:.*]] = "tf.Const"() <{value = dense<{{.*}}> : tensor<10x2xf32>}> : () -> tensor<10x2xf32>

  func.func @matmul(%arg0: tensor<2x12xf32>) -> (tensor<*xf32>) {
    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<12x2xf32>} : () -> tensor<12x2xf32>
    %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 @not_quantize_const() -> (tensor<2x1024xf32>) {
    // Nothing happens if not connected wiht quantizable op.
    %cst_0 = "tf.Const"() {value = dense<2.000000e+00> : tensor<2x1024xf32>} : () -> tensor<2x1024xf32>
    func.return %cst_0: tensor<2x1024xf32>
  }

// CHECK-LABEL: func @not_quantize_const
// CHECK-DAG: %[[W:.*]] = "tf.Const"() <{value = dense<2.000000e+00> : tensor<2x1024xf32>
// CHECK: return %[[W]] : tensor<2x1024xf32>
}

  func.func @matmul(%arg0: tensor<2x12xf32>) -> (tensor<*xf32>) {
    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<12x2xf32>} : () -> tensor<12x2xf32>
    %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>
  }

    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<2x1024xf32>} : () -> tensor<2x1024xf32>
    %1 = "tf.PartitionedCall"(%arg0, %cst_0) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_matmul_fn} : (tensor<1x2x2x3xf32>, tensor<2x1024xf32>) -> tensor<*xf32>
    func.return %1: tensor<*xf32>
  }

  module {
    func.func @main(%arg0: tensor<2x2x2xf32>, %arg1: tensor<2x2x2xf32>) -> tensor<2x2x2xf32> attributes {_from_xla_call_module} {
      %0 = stablehlo.dot_general %arg0, %arg1,
        batching_dims = [0] x [0],
        contracting_dims = [2] x [1],
        precision = [DEFAULT, DEFAULT]
      : (tensor<2x2x2xf32>, tensor<2x2x2xf32>) -> tensor<2x2x2xf32>
        return %0 : tensor<2x2x2xf32>
    }
  }
)mlir";

  %7 = "quantfork.stats"(%6) {layerStats = dense<[0.000000e+00, 33.0349121]> : tensor<2xf32>} : (tensor<1x2xf32>) -> tensor<1x2xf32>
  func.return %7 : tensor<1x2xf32>

// CHECK-LABEL: QuantizeReshapeOp
func.func @QuantizeReshapeOp(%arg0: tensor<1x1x3xf32>) -> (tensor<1x3xf32>) {
  %1 = "quantfork.stats"(%arg0) {layerStats = dense<[-1.0, 1.0]> : tensor<2xf32>} : (tensor<1x1x3xf32>) -> tensor<1x1x3xf32>
  %2 = "tfl.pseudo_const"() {value = dense<[-1, 3]> : tensor<2xi32>} : () -> tensor<2xi32>
  %3 = "tfl.reshape"(%1, %2) : (tensor<1x1x3xf32>, tensor<2xi32>) -> tensor<1x3xf32>

  %1 = "tfl.add"(%cst_2, %cst_1) {fused_activation_function = "NONE"} : (tensor<2x2x2xf32>, tensor<  2x2xf32>) -> tensor<2x2x2xf32>
  %2 = "tfl.add"(%cst_0, %cst_2) {fused_activation_function = "NONE"} : (tensor<    2xf32>, tensor<2x2x2xf32>) -> tensor<2x2x2xf32>

  func.return %0, %1, %2 : tensor<2x2xf32>, tensor<2x2x2xf32>, tensor<2x2x2xf32>

func.func @prepareAdd(%arg0: tensor<2x2xf32>) -> tensor<2x2xf32> {
  %cst = arith.constant dense<[[0.0, 1.0], [2.0, 255.0]]> : tensor<2x2xf32>
  %add = "tfl.add"(%arg0, %cst) {fused_activation_function="NONE"} : (tensor<2x2xf32>, tensor<2x2xf32>) -> tensor<2x2xf32>
  func.return %add : tensor<2x2xf32>

// CHECK: %[[cst:.*]] = arith.constant dense<[{{\[}}0.000000e+00, 1.000000e+00], [2.000000e+00, 2.550000e+02]]>