%10 = stablehlo.convert %3 : (tensor<1x1x3xi32>) -> tensor<1x1x3xf32>
    %11 = stablehlo.broadcast_in_dim %10, dims = [0, 1, 2] : (tensor<1x1x3xf32>) -> tensor<1x4x3xf32>  // Optional
    %12 = stablehlo.subtract %9, %11 : tensor<1x4x3xf32>  // Precalculated zp_neg.
    %13 = stablehlo.broadcast_in_dim %4, dims = [0, 1, 2] : (tensor<1x1x3xf32>) -> tensor<1x4x3xf32>  // Optional
    %14 = stablehlo.multiply %12, %13 : tensor<1x4x3xf32>  // s1 * s2

// CHECK-LABEL: QuantizeReshapeOp
func.func @QuantizeReshapeOp(%arg0: tensor<1x1x3xf32>) -> (tensor<1x3xf32>) {
  %1 = "quantfork.stats"(%arg0) {layerStats = dense<[-1.0, 1.0]> : tensor<2xf32>} : (tensor<1x1x3xf32>) -> tensor<1x1x3xf32>
  %2 = "tfl.pseudo_const"() {value = dense<[-1, 3]> : tensor<2xi32>} : () -> tensor<2xi32>
  %3 = "tfl.reshape"(%1, %2) : (tensor<1x1x3xf32>, tensor<2xi32>) -> tensor<1x3xf32>

  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>

// 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>

    %output_0, %min_1, %max_2, %histogram_3 = "tf.CustomAggregator"(%0) <{calibration_method = 1 : i32, id = "keeping_id", max_percentile = 0.000000e+00 : f32, min_percentile = 0.000000e+00 : f32, num_bins = 0 : i32}> : (tensor<10x1x3xf32>) -> (tensor<10x1x3xf32>, tensor<f32>, tensor<f32>, tensor<0xi64>)
    return %output_0 : tensor<10x1x3xf32>
  }
  // CHECK-LABEL: @main

  %4 = "quantfork.stats"(%3) {
    layerStats = dense<[0.0, 4.0]> : tensor<2xf32>
  } : (tensor<?x1x1x3xf32>) -> tensor<?x1x1x3xf32>
  %5 = "tfl.sqrt"(%4) : (tensor<?x1x1x3xf32>) -> tensor<?x1x1x3xf32>
  %6 = "quantfork.stats"(%5) {
    layerStats = dense<[0.0, 2.0]> : tensor<2xf32>
  } : (tensor<?x1x1x3xf32>) -> tensor<?x1x1x3xf32>

  %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>

module {
func.func @tfInplaceUpdate(%arg0: tensor<2x1x2xf32>) -> tensor<2x1x2xf32> {
  %1 = arith.constant dense<1> : tensor<1xi32>
  %2 = arith.constant dense<2.0> : tensor<1x1x2xf32>
  %3 = "tf.InplaceUpdate"(%arg0, %1, %2) {device = ""}
    : (tensor<2x1x2xf32>, tensor<1xi32>, tensor<1x1x2xf32>) -> tensor<2x1x2xf32>
  func.return %3 : tensor<2x1x2xf32>
}
}

//CHECK: module {

      return %0 : tensor<1x1x4xf32>
    }
  )mlir";

  const OwningOpRef<ModuleOp> module_op =
      ParseModuleOpString(kXlaCallModuleOpWithQuantizationMethodAttr);
  ASSERT_TRUE(module_op);

// CHECK-SAME:                 %[[ARG_0:.*]]: tensor<1x1x167xf32>
func.func @dot_general_fn(%arg0: tensor<1x1x167xf32>) -> tensor<1x1x64xf32> {
  %0 = stablehlo.constant dense<2.000000e+00> : tensor<167x64xf32>
  %1 = stablehlo.dot_general %arg0, %0, contracting_dims = [2] x [0], precision = [DEFAULT, DEFAULT] : (tensor<1x1x167xf32>, tensor<167x64xf32>) -> tensor<1x1x64xf32>
  return %1 : tensor<1x1x64xf32>
}