%9 = "quantfork.qcast"(%8) {volatile} : (tensor<2x3x1x3xf32>) -> tensor<2x3x1x3x!quant.uniform<i8:f32, 3.000000e-01:1>>
    %10 = "quantfork.dcast"(%9) : (tensor<2x3x1x3x!quant.uniform<i8:f32, 3.000000e-01:1>>) -> tensor<2x3x1x3xf32>
    %11 = stablehlo.reshape %10 : (tensor<2x3x1x3xf32>) -> tensor<2x3x3xf32>

    %0 = "tf.Div"(%arg0, %cst_3) {device = ""} : (tensor<2x3x4xf32>, tensor<f32>) -> tensor<2x3x4xf32>
    %1 = "tf.AddV2"(%0, %cst_1) {device = ""} : (tensor<2x3x4xf32>, tensor<f32>) -> tensor<2x3x4xf32>
    %2 = "tf.Maximum"(%1, %cst_1) {device = ""} : (tensor<2x3x4xf32>, tensor<f32>) -> tensor<2x3x4xf32>
    %3 = "tf.Minimum"(%2, %cst_5) {device = ""} : (tensor<2x3x4xf32>, tensor<f32>) -> tensor<2x3x4xf32>

# CHECK:           %[[VAL_9:.*]] = "tfl.transpose"(%[[VAL_1]], %[[VAL_2]]) : (tensor<3x7xf32>, tensor<2xi32>) -> tensor<7x3xf32>
# CHECK:           %[[VAL_10:.*]] = "tfl.fully_connected"(%[[VAL_7]]#0, %[[VAL_9]], %[[VAL_3]]) <{fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"}> : (tensor<1x5x3xf32>, tensor<7x3xf32>, none) -> tensor<5x7xf32>

// 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>

  // CHECK:  return
}

func.func @pow(%arg0: tensor<2x1x3xf32>, %arg1: tensor<2x1x1xf32>) -> tensor<2x1x3xf32> {
  %0 = "tf.Pow"(%arg0, %arg1) : (tensor<2x1x3xf32>, tensor<2x1x1xf32>) -> tensor<2x1x3xf32>
  func.return %0 : tensor<2x1x3xf32>

  // CHECK-LABEL: pow
  // CHECK:  %[[pow:.*]] = tfl.pow(%arg0, %arg1) : (tensor<2x1x3xf32>, tensor<2x1x1xf32>) -> tensor<2x1x3xf32>
  // CHECK:  return
}

  %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>

    %2 = "quantfork.stats"(%1) {layerStats = dense<[5.00000000e-6, 7.00000000e-1]> : tensor<2xf32>} : (tensor<2x2x3xf32>) -> tensor<2x2x3xf32>
    return %2 : tensor<2x2x3xf32>
  }
// CHECK: func.func private @quantize_dot_general_batch_per_tensor_quantized_fn(%[[ARG_0:.+]]: tensor<2x2x2xf32>) -> tensor<2x2x3xf32> attributes {tf._original_func_name = "main_0"}

    %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

}

// -----

func.func @testSlice(%arg0: tensor<2x3x5xf32>, %arg1: tensor<3xi32>, %arg2: tensor<3xi32>) -> tensor<?x3x5xf32> {
  %0 = "tfl.slice"(%arg0, %arg1, %arg2) : (tensor<2x3x5xf32>, tensor<3xi32>, tensor<3xi32>) -> tensor<?x3x5xf32>
  func.return %0 : tensor<?x3x5xf32>
}

// -----

func.func @testSliceBadBeginDimension(%arg0: tensor<2x3x5xf32>, %arg1: tensor<2xi32>, %arg2: tensor<3xi32>) -> tensor<?x3x5xf32> {

  // CHECK-SAME: ({{%.+}}: tensor<1x2x3xf32>)
  // CHECK-SAME: -> (tensor<1x8x3xf32>, tensor<1x8x3xf32>)
  func.func @while_shape_invariant_different_dims(%arg0: tensor<1x2x3xf32>) -> (tensor<1x8x3xf32>, tensor<1x8x3xf32>) {
    // CHECK: "tf.While"
    // CHECK-SAME: (tensor<1x2x3xf32>)
    // CHECK-SAME: -> tensor<1x8x3xf32>