(UpdateShapeWithAxis<-1> $qtype, $old_value))),
          [(CanUpdateShapeWithAxis<-1> $qtype, $old_value)]>;

// The axis is set to 0 because the transpose is from the legalization of
// tf.conv2d and the new channel axis is the first dimension.
def ReorderTransposeDequantQuantUsedByConv :
      Pat<(TF_TransposeOp:$old_value
              (TFL_DequantizeOp (TFL_QuantizeOp $input, $qtype)), $perm),

  }

  func.func private @composite_conv2d_fn_1(%arg0: tensor<1x2x2x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> tensor<*xf32> attributes {tf_quant.composite_function} {

  }

  func.func private @composite_conv2d_fn_1(%arg0: tensor<1x2x2x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> tensor<*xf32> attributes {tf_quant.composite_function} {

      if (!input_type || !input_type.hasStaticShape()) {
        return false;
      }
      total_count += input_type.getNumElements();
    }
    *count = total_count;
    return true;
  }

  // For conv2d/depthwise_conv/fully_connected ops.
  // This algorithm actually comes from TOCO tooling_util.cc
  static bool GetArithmeticCountForConvAndFullyconnectedOp(mlir::Operation* op,

      def body(self, x, w):
        z = nn_ops.conv2d(x, w, padding='SAME')
        return z, w

      @def_function.function(
          input_signature=[
              tensor_spec.TensorSpec(
                  shape=input_shape, dtype=dtypes.float32, name='input_tensor'
              )
          ]
      )
      def main(self, x):
        x1 = nn_ops.conv2d(x, self.w, padding='SAME')

                         %input : tensor<*xf32>, %weight : tensor<*x!tf_type.qint8>,
                         %weight_scale : tensor<*xf32>, %weight_zp : tensor<*xi32>) -> tensor<*xf32>
      attributes {tf_quant.quantized_ops = ["Conv2D"]} {

    %out = "tf.UniformQuantizedConvolutionHybrid"(%input, %weight,
                           %weight_scale, %weight_zp) {
        Tlhs = "tfdtype$DT_FLOAT",
        Trhs = "tfdtype$DT_QINT8",

  %1 = "tf.FakeQuantWithMinMaxArgs"(%arg0) {device = "", max = 2.000000e-01 : f32, min = -1.000000e-01 : f32, narrow_range = false, num_bits = 8 : i64} : (tensor<1x3x4x3xf32>) -> tensor<*xf32>
  %2 = "tf.Conv2D"(%1, %0) {data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true} : (tensor<*xf32>, tensor<*xf32>) -> tensor<*xf32>

    %1 = "quantfork.stats"(%arg0) {layerStats = dense<[1.27501142, 149.824783]> : tensor<2xf32>} : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xf32>

  key: "MobilenetV1/MobilenetV1/Conv2d_0/BatchNorm/FusedBatchNorm@"
  value {
    file_line_cols {
      file_index: 5
      line: 362
    }
  }
}
traces {
  key: "MobilenetV1/MobilenetV1/Conv2d_0/Conv2D@"
  value {
    file_line_cols {
      file_index: 2
      line: 27
    }
    file_line_cols {
      file_index: 3
      line: 28
    }
    file_line_cols {
      file_index: 4

  %dq_bias = "quantfork.dcast"(%q_bias) : (tensor<2x!quant.uniform<i32:f32, 0.044022349891595126>>) -> tensor<2xf32>