%14 = stablehlo.multiply %12, %13 : tensor<1x4x3xf32>  // s1 * s2
    %15 = call @uniform_quantize_1(%14, %5, %6) : (tensor<1x4x3xf32>, tensor<1x1x1xf32>, tensor<1x1x1xi8>) -> tensor<1x4x3xi8>
    %16 = call @uniform_dequantize_0(%15, %5, %6) : (tensor<1x4x3xi8>, tensor<1x1x1xf32>, tensor<1x1x1xi8>) -> tensor<1x4x3xf32>
    return %16 : tensor<1x4x3xf32>
  }

// CHECK-LABEL: QuantizeUnidirectionalLstmFullPerTensor
func.func @QuantizeUnidirectionalLstmFullPerTensor(%arg0: tensor<1x2x3xf32>) -> (tensor<1x2x3xf32>) {
  %input = "quantfork.stats"(%arg0) {layerStats = dense<[0.0, 1.0]> : tensor<2xf32>} : (tensor<1x2x3xf32>) -> tensor<1x2x3xf32>
  %1 = "tfl.pseudo_const"() {value = dense<[[0.1]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>

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

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

    func.func @main() -> (tensor<1x4x4xf32>) {
      %%arg0 = "tf.Const"() {value = dense<-3.0> : tensor<1x4x2xf32>} : () -> tensor<1x4x2xf32>
      %%arg1 = "tf.Const"() {value = dense<-3.0> : tensor<1x2x4xf32>} : () -> tensor<1x2x4xf32>

      %%1 = "tf.%s"(%%arg0, %%arg1) {T = f32, adj_x = false, adj_y = false, grad_x = false, grad_y = false, device = ""} : (tensor<1x4x2xf32>, tensor<1x2x4xf32>) -> tensor<1x4x4xf32>

func.func @reshape_removeIdentity(tensor<4x4x4xf32>) -> tensor<4x4x4xf32> {
^bb0(%arg0: tensor<4x4x4xf32>) :
  %cst = arith.constant dense<[4, 4, 4]> : tensor<3xi32>
  %0 = "tfl.reshape"(%arg0, %cst) : (tensor<4x4x4xf32>, tensor<3xi32>) -> tensor<4x4x4xf32>
  func.return %0 : tensor<4x4x4xf32>

// CHECK-LABEL: func @reshape_removeIdentity
// CHECK:  return %arg0 : tensor<4x4x4xf32>
}

  %50 = "tfl.read_variable"(%5) : (tensor<!tf_type.resource>) -> tensor<1x2x3xf32>
  %51 = "quantfork.stats"(%50) {layerStats = dense<[0.0, 1.0]> : tensor<2xf32>} : (tensor<1x2x3xf32>) -> tensor<1x2x3xf32>
  %52 = "tfl.concatenation"(%51, %0) {axis = 1 : i32, fused_activation_function = "NONE"} : (tensor<1x2x3xf32>, tensor<1x2x3xf32>) -> tensor<1x4x3xf32>

  %3 = "tfl.dequantize"(%2) : (tensor<1x3x3x!quant.uniform<i8:f32, 0.004:-128>>) -> tensor<1x3x3xf32>
  %4 = "tfl.div"(%arg0, %3) {fused_activation_function = "NONE"} : (tensor<1x3x3xf32>, tensor<1x3x3xf32>) -> tensor<1x3x3xf32>
  func.return %4 : tensor<1x3x3xf32>
//  CHECK: %[[logistic:.*]] = "tfl.logistic"
//  CHECK: %[[dq:.*]] = "tfl.dequantize"(%[[logistic]])

    %5 = "quantfork.qcast"(%4) {volatile} : (tensor<3x4x2xf32>) -> tensor<3x4x2x!quant.uniform<i8:f32, 0.13170163023705575:-1>>
    %6 = "quantfork.dcast"(%5) : (tensor<3x4x2x!quant.uniform<i8:f32, 0.13170163023705575:-1>>) -> tensor<3x4x2xf32>
    %7 = "stablehlo.gather"(%6, %arg2) {
    dimension_numbers = #stablehlo.gather<

// CHECK-NEXT: return %[[dq]] : tensor<2x2xf32>
}

// CHECK-LABEL: prepareStatistics
func.func @prepareStatistics(%arg0: tensor<8x4x3xf32>) -> tensor<8x4x3xf32> {
  %0 = "quantfork.stats"(%arg0) {
    layerStats = dense<[-1.0, 1.0]> : tensor<2xf32>
  } : (tensor<8x4x3xf32>) -> tensor<8x4x3xf32>
  %1 = "quantfork.stats"(%0) {
    layerStats = dense<[-1.0, 1.0]> : tensor<2xf32>,
    axisStats = dense<[
      [-1.0, 1.0],

  %0 = "tfl.var_handle"() {container = "", shared_name = "states"} : () -> tensor<!tf_type.resource<tensor<1x2x3xf32>>>
  %1 = "tfl.read_variable"(%0) : (tensor<!tf_type.resource<tensor<1x2x3xf32>>>) -> tensor<1x2x3xf32>
  %2 = "tfl.quantize"(%1) {qtype = tensor<1x2x3x!quant.uniform<u8<1:255>:f32, 0.047244094488188976:128>>} : (tensor<1x2x3xf32>) -> tensor<1x2x3x!quant.uniform<u8<1:255>:f32, 0.047244094488188976:128>>