func.return %add : tensor<*xf32>
}
// CHECK: func @decompose_batch_norm
// CHECK-DAG: %[[CONST:.*]] = "tf.Const"() <{value = dense<2.49743462E-5> : tensor<2xf32>}> : () -> tensor<2xf32>

      Targs = [tensor<32xf32>, tensor<64xf32>, tensor<8xf32>, tensor<2xf32>],
      Tresults = [tensor<32xf32>, tensor<64xf32>, tensor<8xf32>]}
      : (tensor<32xf32>, tensor<64xf32>, tensor<8xf32>, tensor<2xf32>, tensor<2x!tf_type.string>)
        -> (tensor<32xf32>, tensor<64xf32>, tensor<8xf32>)
    tf_device.return %0#0, %0#1, %0#2 : tensor<32xf32>, tensor<64xf32>, tensor<8xf32>

// CHECK-NEXT:  %[[quant:.*]] = "tfl.quantize"(%[[split]]#0) <{qtype = tensor<2x!quant.uniform<u8:f32, 1.000000e+00>>}> {volatile} : (tensor<2xf32>) -> tensor<2x!quant.uniform<u8:f32, 1.000000e+00>>
// CHECK-NEXT:  return %[[quant]] : tensor<2x!quant.uniform<u8:f32, 1.000000e+00>>

  %min = arith.constant dense<0.0> : tensor<16xf32>
  %max = arith.constant dense<15.0> : tensor<16xf32>
  %mini = "tf.Identity"(%min) : (tensor<16xf32>) -> tensor<16xf32>
  %maxi = "tf.Identity"(%max) : (tensor<16xf32>) -> tensor<16xf32>
  %fq = "tf.FakeQuantWithMinMaxVarsPerChannel"(%in, %mini, %maxi) {num_bits = 3, narrow_range = false} : (tensor<3x3x3x16xf32>, tensor<16xf32>, tensor<16xf32>) -> tensor<3x3x3x16xf32>

    %1 = "tf.BiasAdd"(%0, %arg2) {data_format = "NHWC", device = ""} : (tensor<*xf32>, tensor<2xf32>) -> tensor<*xf32>
    %2 = "tf.Relu6"(%1) : (tensor<*xf32>) -> tensor<*xf32>
    func.return %2 : tensor<*xf32>
  }

// CHECK-LABEL: func @conv_with_single_layer

// CHECK: %[[quantize:.*]] = "tf.PartitionedCall"(%arg0

  %min = arith.constant dense<0.0> : tensor<16xf32>
  %max = arith.constant dense<255.0> : tensor<16xf32>
  %mini = "tf.Identity"(%min) : (tensor<16xf32>) -> tensor<16xf32>
  %maxi = "tf.Identity"(%max) : (tensor<16xf32>) -> tensor<16xf32>
  %fq = "tf.FakeQuantWithMinMaxVarsPerChannel"(%in, %mini, %maxi) {num_bits = 5, narrow_range = false} : (tensor<3x3x3x16xf32>, tensor<16xf32>, tensor<16xf32>) -> tensor<3x3x3x16xf32>

"TANH", input_to_cell_intermediate = tensor<0xf32>, input_to_forget_intermediate = tensor<0xf32>, input_to_input_intermediate = tensor<0xf32>, input_to_output_intermediate = tensor<0xf32>, proj_clip = 0.000000e+00 : f32, time_major = false} : (tensor<?x?x?xf32>, 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<?xf32>,...

  } : (tensor<2x!tf_type.qint8>, tensor<f32>, tensor<i32>) -> tensor<2xf32>
  func.return %1 : tensor<2xf32>
}

// -----

// CHECK-LABEL: func @uniform_quantize_and_dequantize_per_axis
func.func @uniform_quantize_and_dequantize_per_axis(%arg0 : tensor<2x2xf32>) -> tensor<2x2xf32> {
  %scales = "tf.Const"() { value = dense<[1.0, 2.0]> : tensor<2xf32> } : () -> tensor<2xf32>

    %4 = "tf.BiasAdd"(%3, %cst) {data_format = "NHWC", device = ""} : (tensor<*xf32>, tensor<2xf32>) -> tensor<*xf32>
    %5 = "tf.Relu6"(%4) {device = ""} : (tensor<*xf32>) -> tensor<*xf32>

    func.return %2, %5 : tensor<*xf32>, tensor<*xf32>
  }

  func.func @float_conv_1(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> (tensor<*xf32>, tensor<*xf32>) {

  %12 = mhlo.clamp %9, %11, %1 : tensor<2xf32>
  %13 = mhlo.multiply %arg0, %12 : tensor<2xf32>
  %14 = mhlo.divide %13, %1 : tensor<2xf32>
  return %14 : tensor<2xf32>
}

// CHECK-LABEL:   func.func @hardswish(
// CHECK-SAME:                         %[[VAL_0:.*]]: tensor<2xf32>) -> tensor<*xf32> {
// CHECK:           %[[VAL_1:.*]] = "tfl.hard_swish"(%[[VAL_0]]) : (tensor<2xf32>) -> tensor<2xf32>