%1 = stablehlo.constant dense<1.000000e+03> : tensor<1x1x1x1xf32>  // Input inverse scale.
    %2 = stablehlo.constant dense<-128> : tensor<1x1x1x1xi8>  // Input zero point.
    %3 = stablehlo.constant dense<1> : tensor<3x3x4x4xi8>  // Quantized filter tensor.
    %4 = stablehlo.constant dense<3.000000e+03> : tensor<1x1x1x4xf32>
    %5 = stablehlo.constant dense<4.000000e+03> : tensor<1x1x1x1xf32>  // Output inverse scale.

func.func @add_with_i32_five_dim_broadcasting(tensor<1x1x1x1x1xi32>, tensor<1xi32>) -> tensor<1x1x1x1x1xi32> {
^bb0(%arg0: tensor<1x1x1x1x1xi32>, %arg1: tensor<1xi32>):
  // CHECK: tfl.add(%arg0, %arg1) <{fused_activation_function = "RELU6"}>
  %0 = "tfl.add"(%arg0, %arg1) {fused_activation_function = "RELU6"} : (tensor<1x1x1x1x1xi32>, tensor<1xi32>) -> tensor<1x1x1x1x1xi32>
  func.return %0#0 : tensor<1x1x1x1x1xi32>
}

// -----

// MinElement-LABEL: QuantizeCustomOp
func.func @QuantizeCustomOp(%arg0: tensor<1x1x1x1xf32>) -> (tensor<*xf32>, tensor<*xf32>, 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_1 = arith.constant dense<127.0> : tensor<4096x1x1x1xf32>

  func.return %1 : tensor<1x1x1x16xf32>

// CHECK: %[[avgp:.*]] = "tfl.average_pool_2d"(%arg0)
// CHECK: %[[dq:.*]] = "tfl.dequantize"(%[[avgp]]) : (tensor<1x1x1x16x!quant.uniform<u8:f32, 7.812500e-03:128>>) -> tensor<1x1x1x16xf32>
// CHECK: return %[[dq]] : tensor<1x1x1x16xf32>
}

// CHECK-LABEL: QuantizeReshape2D

  %0 = mhlo.constant dense<[[[[0, 1, 2, 3]]]]> : tensor<1x1x1x4xi32>
  %1 = mhlo.constant dense<[[[[0, 1, 2, 3]], [[0, 1, 2, 3]]]]> : tensor<1x2x1x4xi32>
  %2 = "mhlo.broadcast_in_dim"(%0) <{broadcast_dimensions = dense<[0, 1, 2, 3]> : tensor<4xi64>}> : (tensor<1x1x1x4xi32>) -> tensor<1x2x1x4xi32>
  %3 = mhlo.multiply %1, %2 : tensor<1x2x1x4xi32>

    return %15 : tensor<1x2240x1120x512xi8>
  }

// CHECK-LABEL: func @conv_with_filter_larger_than_1GB
// CHECK-DAG: %[[CONST:.*]] = "tf.Const"() <{value = dense<-237772800> : tensor<1x1x1x512xi32>}> : () -> tensor<1x1x1x512xi32>
// CHECK: %[[PADV2_0:.*]] = "tf.PadV2"
// CHECK: %[[XLACONVV2_0:.*]] = "tf.XlaConvV2"(%[[PADV2_0]]
// CHECK: %[[SUB_0:.*]] = "tf.Sub"(%[[XLACONVV2_0]], %[[CONST]])

  %7 = "tf.FloorDiv"(%arg1, %5) {device = ""} : (tensor<1x1x2x1xi32>, tensor<i32>) -> tensor<1x1x2x1xi32>
  %8 = "tf.FloorMod"(%7, %4) {device = ""} : (tensor<1x1x2x1xi32>, tensor<i32>) -> tensor<1x1x2x1xi32>
  %9 = "tf.FloorDiv"(%arg1, %4) {device = ""} : (tensor<1x1x2x1xi32>, tensor<i32>) -> tensor<1x1x2x1xi32>

// CHECK:           %[[VAL_6:.*]] = "tfl.reshape"(%[[VAL_1]], %[[VAL_2]]) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<1xf32>, tensor<4xi32>) -> tensor<1x1x1x1xf32>

// CHECK:           %[[VAL_6:.*]] = "tfl.reshape"(%[[VAL_1]], %[[VAL_2]]) : (tensor<1xf32>, tensor<4xi32>) -> tensor<1x1x1x1xf32>
// CHECK:           %[[VAL_7:.*]] = "tfl.concatenation"(%[[VAL_5]], %[[VAL_6]]) <{axis = 3 : i32, fused_activation_function = "NONE"}> : (tensor<1x1x1x1xf32>, tensor<1x1x1x1xf32>) -> tensor<1x1x1x2xf32>

    %1 = "tfl.reshape"(%arg1, %cst) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<1xf32>, tensor<4xi32>) -> tensor<1x1x1x1xf32>
    %2 = "tfl.concatenation"(%0, %1) {axis = 3 : i32, fused_activation_function = "NONE", tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<1x1x1x1xf32>, tensor<1x1x1x1xf32>) -> tensor<1x1x1x2xf32>