func.func @exp(%arg0: tensor<1x1x1x96xf32>) -> tensor<1x1x1x96xf32> {
  %0 = "vhlo.exponential_v1" (%arg0) : (tensor<1x1x1x96xf32>) -> tensor<1x1x1x96xf32>
  func.return %0 : tensor<1x1x1x96xf32>
}

//CHECK:func.func private @exp(%arg0: tensor<1x1x1x96xf32>) -> tensor<1x1x1x96xf32> {
//CHECK-NEXT: %0 = "vhlo.exponential_v1"(%arg0) : (tensor<1x1x1x96xf32>) -> tensor<1x1x1x96xf32>
//CHECK-NEXT: return %0 : tensor<1x1x1x96xf32>
//CHECK-NEXT:}

  return %2 : tensor<1x1x1x4xf32>
}
func.func private @XlaCallModule_aten.avg_pool2d.default.impl_3(%arg0: tensor<1x1x1x9xf32>) -> tensor<1x1x1x4xf32>

// CHECK-LABEL: avg_pool2d_4
// CHECK: %cst = arith.constant dense<[0, 2, 3, 1]> : tensor<4xi32>
// CHECK: %0 = "tfl.transpose"(%arg0, %cst) : (tensor<1x1x1x9xf32>, tensor<4xi32>) -> tensor<1x1x9x1xf32>

    %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.

// 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>

// CHECK:           %[[VAL_8:.*]] = "tfl.reshape"(%[[VAL_7]], %[[VAL_3]]) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<1x1x1x2xf32>, tensor<1xi32>) -> tensor<2xf32>
// CHECK:           %[[VAL_9:.*]] = "tfl.reshape"(%[[VAL_8]], %[[VAL_4]]) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<2xf32>, tensor<2xi32>) -> tensor<2x1xf32>

  %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>

  %3 = "tfl.quantize"(%2) {qtype = tensor<1x1x1x3x!quant.uniform<i8:f32, 0.1>>, volatile} : (tensor<1x1x1x3xf32>) -> tensor<1x1x1x3x!quant.uniform<i8:f32, 0.1>>
  %4 = "tfl.dequantize"(%3) : (tensor<1x1x1x3x!quant.uniform<i8:f32, 0.1>>) -> tensor<1x1x1x3xf32>
  %5 = "tfl.add"(%1, %4) {fused_activation_function = "NONE"} : (tensor<1x5x5x3xf32>, tensor<1x1x1x3xf32>) -> tensor<1x5x5x3xf32>

  %cst_0 = "tf.Const"() {value = dense<0.400000e+00> : tensor<1x1x1x2xf32>} : () -> tensor<1x1x1x2xf32>
  %cst_1 = "tf.Const"() {value = dense<0.200000e+00> : tensor<1x1x1x2xf32>} : () -> tensor<1x1x1x2xf32>
  %cst_2 = "tf.Const"() {value = dense<0.300000e+00> : tensor<1x1x1x2xf32>} : () -> tensor<1x1x1x2xf32>

// 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>