return %1 : tensor<1x3xf32>
  }
}

// -----

// Do not merge when function is not quantized function.

module attributes {tf_saved_model.semantics} {
  // CHECK-LABEL: func.func private @merge_relu_fusion
  func.func private @merge_relu_fusion(%arg0: tensor<1x4xf32>) -> tensor<1x3xf32> {

func.func @fakeQuantPerChannelForActivation(%arg0: tensor<8x4xf32>) -> (tensor<8x4xf32>) {
  %arg1 = arith.constant dense<[0.0, -1.0, 1.0, 0.0]> : tensor<4xf32>
  %arg2 = arith.constant dense<[255.0, 254.0, 256.0, 1.0e-9]> : tensor<4xf32>
  %0 = "tf.FakeQuantWithMinMaxVarsPerChannel"(%arg0, %arg1, %arg2) {num_bits = 5, narrow_range = false} : (tensor<8x4xf32>, tensor<4xf32>, tensor<4xf32>) -> tensor<8x4xf32>
  func.return %0 : tensor<8x4xf32>

        return %2 : tensor<?x2xf32>
      }
      func.func private @composite_fn_1(%arg0: tensor<?x2xf32>, %arg1: tensor<2x2xf32>, %arg2: tensor<2xf32>) -> tensor<?x2xf32> attributes {_from_xla_call_module, tf_quant.composite_function} {
        return %arg0 : tensor<?x2xf32>
      }
    }
  )mlir";

// CHECK-NEXT:        buffer: 4
// CHECK-NEXT:      } ],
// CHECK-NEXT:      signature_defs: [  ]
// CHECK-NEXT:    }

  %0 = "tfl.cast" (%arg0) : (tensor<4x5xbf16>) -> tensor<4x5xf32> loc("cast1")
  %1 = "tfl.cast" (%0) : (tensor<4x5xf32>) -> tensor<4x5xbf16> loc("cast2")
  func.return %1 : tensor<4x5xbf16>

func.func @fakeQuantPerChannelForActivation(%arg0: tensor<8x4xf32>) -> (tensor<8x4xf32>) {
  %arg1 = arith.constant dense<[0.0, -1.0, 1.0, 0.0]> : tensor<4xf32>
  %arg2 = arith.constant dense<[15.0, 14.0, 16.0, 1.0e-9]> : tensor<4xf32>
  %0 = "tf.FakeQuantWithMinMaxVarsPerChannel"(%arg0, %arg1, %arg2) {num_bits = 3, narrow_range = false} : (tensor<8x4xf32>, tensor<4xf32>, tensor<4xf32>) -> tensor<8x4xf32>
  func.return %0 : tensor<8x4xf32>

    ```mlir
      %0 = "tf.Const"() {value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
      %1 = "tf.Const"() {device = "", value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
      %2 = "tf.Const"() {device = "baz", value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
    ```

    then running this pass with 'default-device=foobar', we get:

    ```mlir

  %0 = "tf.BatchMatMulV2"(%arg0, %arg1) {T = f32, adj_x = false, adj_y = false, grad_x = false, grad_y = false, device = ""} : (tensor<1x4x2xf32>, tensor<3x2x4xf32>) -> tensor<3x4x4xf32>
  func.return %0 : tensor<3x4x4xf32>
}

}

func.func @main2(%arg0: tensor<2x4xf32>, %arg1: tensor<2x4xf32>) -> tensor<2x4xf32> {
  %0 = "tfl.quantize"(%arg0) {qtype = tensor<2x4x!quant.uniform<u8:f32, 0.49803921568627452>>} : (tensor<2x4xf32>) -> tensor<2x4x!quant.uniform<u8:f32, 0.49803921568627452>>
  %1 = "tfl.quantize"(%arg1) {qtype = tensor<2x4x!quant.uniform<u8:f32, 0.49803921568627452>>} : (tensor<2x4xf32>) -> tensor<2x4x!quant.uniform<u8:f32, 0.49803921568627452>>

// CHECK: %0 = "tfl.pseudo_const"() <{value = dense<2> : tensor<2xi32>}> : () -> tensor<2xi32>
// CHECK: %1 = "tfl.reshape"(%arg0, %0) : (tensor<1x4xf32>, tensor<2xi32>) -> tensor<2x2xf32>

{
  "version": 3,
  "operator_codes": [
    {
      "builtin_code": "RESHAPE"
    }
  ],
  "subgraphs": [
    {
      "tensors": [
        {
          "shape": [1, 4],

    %9 = stablehlo.convert %2 : (tensor<2x3xi8>) -> tensor<2x3xf32>
    %10 = stablehlo.dot_general %8, %9, contracting_dims = [1] x [0] : (tensor<1x2xf32>, tensor<2x3xf32>) -> tensor<1x3xf32>
    %11 = stablehlo.convert %3 : (tensor<1x3xi32>) -> tensor<1x3xf32>
    %12 = stablehlo.subtract %10, %11 : tensor<1x3xf32>  // q1 * q2 - z1 * q2
    %13 = stablehlo.multiply %12, %4 : tensor<1x3xf32>  // s1 * s2