%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>
  }

  %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>

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>

// 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],

    return %3 : tensor<1x3xf32>
  }
  func.func private @composite_matmul_with_bias_fn_2(%arg0: tensor<1x4xf32>, %arg1: tensor<4x3xf32>, %arg2: tensor<3xf32>) -> tensor<1x3xf32> attributes {tf_quant.composite_function} {
    %0 = "tf.MatMul"(%arg0, %arg1) <{grad_a = false, grad_b = false, transpose_a = false, transpose_b = false}> {attr_map = "0:transpose_a,1:transpose_b", device = ""} : (tensor<1x4xf32>, tensor<4x3xf32>) -> tensor<1x3xf32>

    func.func private @composite_dot_general_fn_1(%arg0: tensor<1x1024xf32>, %arg1: tensor<1024x3xf32>) -> tensor<1x3xf32> attributes {_from_xla_call_module, tf_quant.composite_function} {
      %0 = stablehlo.dot_general %arg0, %arg1, contracting_dims = [1] x [0], precision = [DEFAULT, DEFAULT] : (tensor<1x1024xf32>, tensor<1024x3xf32>) -> tensor<1x3xf32>
      return %0 : tensor<1x3xf32>
    }
  }
)mlir";

    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>

    %0 = stablehlo.dot_general %arg0, %arg1, contracting_dims = [1] x [0], precision = [DEFAULT, DEFAULT] : (tensor<1x1024xf32>, tensor<1024x3xf32>) -> tensor<1x3xf32>
    %1 = stablehlo.add %0, %arg2 : tensor<1x3xf32>
    return %0 : tensor<1x3xf32>
  }

func.func @convert_dot_general_dynamic_batch_dim(%arg0: tensor<2x?x2x3xf32>, %arg1: tensor<2x?x4x3xf32>) -> tensor<2x?x2x4xf32> {
%0 = "mhlo.dot_general"(%arg0, %arg1) {
  dot_dimension_numbers = #mhlo.dot<
    lhs_batching_dimensions = [0, 1],
    rhs_batching_dimensions = [0, 1],
    lhs_contracting_dimensions = [3],
    rhs_contracting_dimensions = [3]
  >} : (tensor<2x?x2x3xf32>, tensor<2x?x4x3xf32>) -> tensor<2x?x2x4xf32>

      %1 = "quantfork.qcast"(%0) {volatile} : (tensor<1x3xf32>) -> tensor<1x3x!quant.uniform<i8:f32, 0.13170163023705575:-1>>
      %2 = "quantfork.dcast"(%1) : (tensor<1x3x!quant.uniform<i8:f32, 0.13170163023705575:-1>>) -> tensor<1x3xf32>
      %3 = stablehlo.reshape %2 : (tensor<1x3xf32>) -> tensor<3x1xf32>