// CHECK: %[[v1:.*]] = "tf.Transpose"(%arg1, %[[cst_0]]) : (tensor<5x3x2xf32>, tensor<3xi32>) -> tensor<5x2x3xf32>
  // CHECK: %[[v2:.*]] = "tf.BatchMatMulV2"(%[[v0]], %[[v1]]) <{adj_x = false, adj_y = false}> : (tensor<5x7x2xf32>, tensor<5x2x3xf32>) -> tensor<5x7x3xf32>
  // CHECK: %[[v3:.*]] = "tf.Transpose"(%[[v2]], %[[cst_0]]) : (tensor<5x7x3xf32>, tensor<3xi32>) -> tensor<5x3x7xf32>
}

// 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:           %[[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>

  %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 @lift_float_batch_matmul(%arg0: tensor<4x4x3xf32>) -> (tensor<4x4x3xf32>) {
  %cst = "tf.Const"() {device = "", value = dense<1.0> : tensor<4x3x3xf32>} : () -> tensor<4x3x3xf32>
  %0 = "tf.BatchMatMulV2"(%arg0, %cst) {adj_x = false, adj_y = false, device = ""} : (tensor<4x4x3xf32>, tensor<4x3x3xf32>) -> tensor<4x4x3xf32>
  return %0 : tensor<4x4x3xf32>

// 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],

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

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>

  %0:3 = "tf.Split"(%arg0, %arg1) : (tensor<i32>, tensor<1x4x3x3xf32>) -> (tensor<1x4x3xf32>, tensor<1x4x3xf32>, tensor<1x4x3xf32>)
  func.return %0#0 : tensor<1x4x3xf32>

  // CHECK-LABEL: split
  // CHECK: "tfl.split"(%arg0, %arg1) <{num_splits = 3 : i32}> : (tensor<i32>, tensor<1x4x3x3xf32>) -> (tensor<1x4x3xf32>, tensor<1x4x3xf32>, tensor<1x4x3xf32>)
}

  // 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>