// CHECK-NEXT:  %[[SELU_VAL:.*]] = "tf.Mul"(%[[ELU_VAL]], %[[SCALED_ALPHA]]) : (tensor<1x4x4x3xf32>, tensor<f32>) -> tensor<1x4x4x3xf32>
    // CHECK-NEXT:  %[[RES:.*]] = "tf.SelectV2"(%[[PRED]], %[[SCALED_FEATURES]], %[[SELU_VAL]]) : (tensor<1x4x4x3xi1>, tensor<1x4x4x3xf32>, tensor<1x4x4x3xf32>) -> tensor<1x4x4x3xf32>
    // CHECK-NEXT:  return %[[RES]] : tensor<1x4x4x3xf32>

    return %5 : tensor<1x3x1x1xf32>
  }
  func.func private @composite_gather_fn_1(%arg0: tensor<1x3x1x1xf32>, %arg1: tensor<1xi32>, %arg2: tensor<i32>) -> tensor<1x3x1x1xf32> attributes {tf_quant.composite_function} {

// CustomOpNotWeightOnly-LABEL: QuantizeCustomOp
func.func @QuantizeCustomOp(%arg0: tensor<1x1x1x1xf32>) -> tensor<*xf32> attributes {tf.entry_function = {inputs = "input", outputs = "custom_op"}} {
  %0 = "quantfork.stats"(%arg0) {layerStats = dense<[0.000000e+00, 2.550000e+02]> : tensor<2xf32>} : (tensor<1x1x1x1xf32>) -> tensor<1x1x1x1xf32>
  %w = arith.constant dense<127.0> : tensor<1024x1x1x1xf32>

  return %2 : tensor<1x1x1x4xf32>
}
func.func private @XlaCallModule_aten.avg_pool2d.default.impl_2(%arg0: tensor<1x1x1x8xf32>) -> tensor<1x1x1x4xf32>

// CHECK-LABEL: avg_pool2d_3
// CHECK: %cst = arith.constant dense<[0, 2, 3, 1]> : tensor<4xi32>
// CHECK: %0 = "tfl.transpose"(%arg0, %cst) : (tensor<1x1x1x8xf32>, tensor<4xi32>) -> tensor<1x1x8x1xf32>

func.func @attribute_cast(%arg0: tensor<1x4x4x1xf32>) -> tensor<1x2x2x1xf32> {
  %0 = "tfr.cast"(%arg0) : (tensor<1x4x4x1xf32>) -> !tfr.tensor
  %stride_i32 = arith.constant 2 : i32
  %1 = tfr.call @tf__my_max_pool(%0, %stride_i32, %stride_i32) : (!tfr.tensor, i32, i32) -> !tfr.tensor
  %2 = "tfr.cast"(%1) : (!tfr.tensor) -> tensor<1x2x2x1xf32>
  func.return %2 : tensor<1x2x2x1xf32>
// CHECK: tf__max_pool
}

  // CHECK-NEXT: [[EXPAND:%.*]] = "tf.ExpandDims"([[INPUT]], [[AXIS]]) : (tensor<1x128x3xf32>, tensor<i32>) -> tensor<1x128x1x3xf32>
  // CHECK-NEXT: [[CONV:%.*]] = "tf.Conv2D"([[EXPAND]], [[FILTER]]) <{dilations = [1, 2, 1, 1], padding = "SAME", strides = [1, 1, 1, 1]}> : (tensor<1x128x1x3xf32>, tensor<5x1x3x8xf32>) -> tensor<1x128x1x8xf32>

func.func @FakeQuantWithMinMaxVarsPerChannel(tensor<1x2x3x8xf32>, tensor<8xf32>, tensor<8xf32>) -> tensor<1x2x3x8xf32> {
^bb0(%arg0: tensor<1x2x3x8xf32>, %arg1: tensor<8xf32>, %arg2: tensor<8xf32>):
  %0 = "tf.FakeQuantWithMinMaxVarsPerChannel"(%arg0, %arg1, %arg2) : (tensor<1x2x3x8xf32>, tensor<8xf32>, tensor<8xf32>) -> tensor<1x2x3x8xf32>
  func.return %0 : tensor<1x2x3x8xf32>
}

// -----

    %unshaped = "tf.Cast"(%arg1) : (tensor<1x2x3xf32>) -> tensor<*xf32>
    // CHECK: <{is_stateless = true}>
    %0 = "tf.IfRegion"(%arg0) <{is_stateless = true}> ({
      // CHECK: "tf.Add"{{.+}}(tensor<1x2x3xf32>, tensor<1x2x3xf32>) -> tensor<1x2x3xf32>
      // CHECK: "tf.Yield"{{.+}}(tensor<1x2x3xf32>) -> ()

  %0:2 = "tf.SplitV"(%arg0, %arg1, %arg2) : (tensor<1x4x3x3xf32>, tensor<2xi32>, tensor<i32>) -> (tensor<1x4x2x3xf32>, tensor<1x4x1x3xf32>)
  func.return %0#0 : tensor<1x4x2x3xf32>

  // CHECK-LABEL: splitv
  // CHECK: "tfl.split_v"(%arg0, %arg1, %arg2) <{num_splits = 2 : i32}> : (tensor<1x4x3x3xf32>, tensor<2xi32>, tensor<i32>) -> (tensor<1x4x2x3xf32>, tensor<1x4x1x3xf32>)
}

    quantization_axis = -1 : i64, quantization_min_val = -128 : i64, quantization_max_val = 127 : i64
  } : (tensor<1x2x2x1x!tf_type.qint32>, tensor<f32>, tensor<i32>) -> tensor<1x2x2x1xf32>

  return %output : tensor<1x2x2x1xi32>