%3 = "tfl.reshape"(%1, %cst) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<100xf32>, tensor<4xi32>) -> tensor<1x1x1x100xf32>
    %4 = "tfl.concatenation"(%2, %3) {axis = 3 : i32, fused_activation_function = "NONE", tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<1x1x1x100xf32>, tensor<1x1x1x100xf32>) -> tensor<1x1x1x200xf32>

  %cst = arith.constant dense<2.0> : tensor<1x1x1x1x2xf32>
  %shape = arith.constant dense<[1, 1, 1, 1, 2]> : tensor<5xi32>
  %1 = "tfl.reshape"(%arg0, %shape) : (tensor<2x1x1x1x1xf32>, tensor<5xi32>) -> tensor<1x1x1x1x2xf32>
  %2 = "tfl.add"(%1, %cst) {fused_activation_function = "NONE"} : (tensor<1x1x1x1x2xf32>, tensor<1x1x1x1x2xf32>) -> tensor<1x1x1x1x2xf32>

  %custom_2 = "tfl.custom"(%arg0, %dq_w) {custom_code = "CustomTestOp", custom_option = #tfl<const_bytes : "0x">} : (tensor<1x1x1x1xf32>, tensor<1024x1x1x1xf32>) -> tensor<*xf32>
  %custom_3 = "tfl.custom"(%arg0, %dq_w) {custom_code = "CustomTestOp", custom_option = #tfl<const_bytes : "0x">} : (tensor<1x1x1x1xf32>, tensor<1024x1x1x1xf32>) -> tensor<*xf32>

func.func @main(tensor<1x6x6x16xf32>) -> tensor<1x1x1x16xf32> {
^bb0(%arg0: tensor<1x6x6x16xf32>):
  // CHECK: "tfl.average_pool_2d"(%{{.*}}) <{filter_height = 3 : i32, filter_width = 6 : i32, fused_activation_function = "NONE", padding = "VALID", stride_h = 3 : i32, stride_w = 1 : i32}> : (tensor<1x6x6x16xf32>) -> tensor<1x1x1x16xf32>

// CHECK:  [[VAL_1:%.*]] = "tfl.reshape"(%2, %[[CST]]) {tac.device = "GPU",  tac.inference_type = "FLOAT"} : (tensor<1xf32>, tensor<4xi32>) -> tensor<1x1x1x1xf32>
// CHECK:  [[VAL_2:%.*]] = "tfl.concatenation"([[VAL_0]], [[VAL_1]]) <{axis = 3 : i32, fused_activation_function = "NONE"}> {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<1x1x1x1xf32>, tensor<1x1x1x1xf32>) -> tensor<1x1x1x2xf32>

  func.return %0 : tensor<1x1x1x16xf32>

    %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.

  %0 = "tfl.dequantize"(%arg0) : (tensor<1000x1000x!quant.uniform<i8:f32, 7.812500e-03>>) -> tensor<1000x1000xf32>
  %1 = "tfl.gather"(%0, %arg1) {axis = 0 : i32, batch_dims = 0 : i32}: (tensor<1000x1000xf32>, tensor<1x1xi32>) -> tensor<1x1x1000xf32>
  func.return %1 : tensor<1x1x1000xf32>

// CHECK-NEXT: tfl.gather
// CHECK-NEXT: tfl.dequantize
}

// 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>

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