device = ""
      } : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x3x4x2xf32>
    return %0 : tensor<1x3x4x2xf32>
  }

  // CHECK: func.func private @qdq_for_conv_weight_per_channel_default(%[[ARG0:.+]]: tensor<1x3x4x3xf32>)
  // CHECK: %[[CST:.+]] = "tf.Const"() <{value = dense<3.000000e-01> : tensor<2x3x3x2xf32>}> : () -> tensor<2x3x3x2xf32>

        version = 5 : i64
      } : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>, tensor<2xf32>) -> tensor<1x3x4x2xf32>
    %2 = "quantfork.stats"(%1) {layerStats = dense<[5.00000000e-6, 7.00000000e-1]> : tensor<2xf32>} : (tensor<1x3x4x2xf32>) -> tensor<1x3x4x2xf32>
    return %2 : tensor<1x3x4x2xf32>
  }
// CHECK: func.func private @quantize_conv_with_bias_1d_fn(%[[ARG_0:.+]]: tensor<1x3x4x3xf32>) -> tensor<1x3x4x2xf32> attributes {tf._original_func_name = "main_0"}

    return %1 : tensor<1x3x4x2xf32>
  }

  func.func private @composite_conv_fn(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> tensor<1x3x4x2xf32> attributes {_from_xla_call_module} {

  func.return %dequant : tensor<1x3x4x2xf32>
}

// -----

// CHECK-LABEL: @dont_merge_quantization_followed_by_quantization
// CHECK-SAME: %[[ARG_0:.*]]: tensor<1x3x4x3xf32>
func.func @dont_merge_quantization_followed_by_quantization(%arg0: tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xf32> {
  // CHECK: %[[QUANT_ARG_0:.*]] = stablehlo.uniform_quantize %[[ARG_0]]

module {
  func.func @depthwise_conv(%arg0: tensor<1x3x4x3xf32>) -> (tensor<*xf32>, tensor<*xf32>) {
    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<3xf32>} : () -> tensor<3xf32>
    %cst_1 = "tf.Const"() {value = dense<3.000000e+00> : tensor<2x3x3x1xf32>} : () -> tensor<2x3x3x1xf32>
    %cst_2 = "tf.Const"() {value = dense<3.000000e+00> : tensor<2x3x3x2xf32>} : () -> tensor<2x3x3x2xf32>

    %0 = "tf.FakeQuantWithMinMaxArgs"(%arg0) {device = "", max = 2.000000e-01 : f32, min = -1.000000e-01 : f32, narrow_range = false, num_bits = 8 : i64} : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xf32>

  // CHECK: return %[[RESULT]] : tensor<2x3x4x6xf32>
}

// -----

func.func @batchMatMulTwoDimAdjXY(%arg0: tensor<2x3x5x4xf32>, %arg1: tensor<2x3x6x5xf32>) -> tensor<2x3x4x6xf32> {
  %0 = "tf.BatchMatMul"(%arg0, %arg1) {adj_x = true, adj_y = true} : (tensor<2x3x5x4xf32>, tensor<2x3x6x5xf32>) -> tensor<2x3x4x6xf32>
  func.return %0 : tensor<2x3x4x6xf32>

    %5 = "tf.Cast"(%4) {device = ""} : (tensor<2x6x4x5xf32>) -> tensor<2x6x4x5xi8>
    %6 = "tf.Div"(%arg1, %cst_2) {device = ""} : (tensor<2x3x4x5xf32>, tensor<f32>) -> tensor<2x3x4x5xf32>
    %7 = "tf.AddV2"(%6, %cst_1) {device = ""} : (tensor<2x3x4x5xf32>, tensor<f32>) -> tensor<2x3x4x5xf32>
    %8 = "tf.Maximum"(%7, %cst_1) {device = ""} : (tensor<2x3x4x5xf32>, tensor<f32>) -> tensor<2x3x4x5xf32>

func.func @ConvertSliceToIdentityI32(%arg0: tensor<2x3x4x5xf32>) -> tensor<2x3x4x5xf32> {
  %begin = arith.constant dense<0> : tensor<4xi32>
  %shape = arith.constant dense<[2,3,4,5]> : tensor<4xi32>
  %0 = "tfl.slice"(%arg0, %begin, %shape) : (tensor<2x3x4x5xf32>, tensor<4xi32>, tensor<4xi32>) -> tensor<2x3x4x5xf32>
  func.return %0 : tensor<2x3x4x5xf32>
  // CHECK: return %arg0
}

// CHECK: %1 = "tfl.transpose"(%0, %cst) : (tensor<2x3x4x5xf32>, tensor<4xi32>) -> tensor<5x2x3x4xf32>

// -----

// CHECK-LABEL: pushTposeBcastScalarCstInput
func.func @pushTposeBcastScalarCstInput(%arg0: tensor<2x3x4x5xf32>) -> tensor<5x2x3x4xf32> {
  %perm = arith.constant dense<[3, 0, 1, 2]> : tensor<4xi32>