// CHECK: return %[[v2]] : tensor<3x4x6x2xf32>
}

func.func @einsum_reduceddim(%arg0: tensor<2x5x7xf32>, %arg1: tensor<2x5x7x3xf32>) -> tensor<2x5x3xf32> {
  %0 = "tf.Einsum"(%arg0, %arg1) {T = "tfdtype$DT_FLOAT", equation = "bin,binj->bij"}: (tensor<2x5x7xf32>, tensor<2x5x7x3xf32>) -> tensor<2x5x3xf32>
  func.return %0 : tensor<2x5x3xf32>
  // CHECK-LABEL: einsum_reduceddim

    %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

// 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>

func.func @my_add_n_multiple_inputs(%arg0: tensor<2x3xf32>, %arg1: tensor<2x3xf32>, %arg2: tensor<2x3xf32>) -> tensor<2x3xf32> {
  %0 = "tf.MyAddN"(%arg0, %arg1, %arg2) {N=3:i32} : (tensor<2x3xf32>, tensor<2x3xf32>, tensor<2x3xf32>) -> tensor<2x3xf32>
  func.return %0 : tensor<2x3xf32>

// CHECK-NEXT: %[[ADD0:.*]] = "tf.RiscAdd"(%arg0, %arg1) : (tensor<2x3xf32>, tensor<2x3xf32>) -> tensor<*xf32>

  %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>

// 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],

    // CHECK: %[[Q1:.*]] = "quantfork.qcast"(%[[ARG0]]) {volatile} : (tensor<1x2xf32>) -> tensor<1x2x!quant.uniform<i8:f32, 5.000000e-03>>
    // CHECK: %[[Q2:.*]] = "quantfork.qcast"(%[[ARG1]]) {volatile} : (tensor<2x3xf32>) -> tensor<2x3x!quant.uniform<i8<-127:127>:f32:1, {6.000000e-03,6.000000e-03,6.000000e-03}>>
    // CHECK: %[[CALL:.*]] = call @quantized_dot_general_fn_1(%[[Q1]], %[[Q2]])

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>

    >} : (
      tensor<2x5x6xi8>, tensor<6x8x2xi8>
    ) -> tensor<2x5x8xi32>
  %1 = chlo.broadcast_add %0, %zp_offset : (
      tensor<2x5x8xi32>, tensor<2x5x8xi32>) -> tensor<2x5x8xi32>
  %2 = chlo.broadcast_add %1, %bias : (
      tensor<2x5x8xi32>, tensor<2x5x8xi32>) -> tensor<2x5x8xi32>
  return %2 : tensor<2x5x8xi32>

    %15 = "quantfork.qcast"(%14) {volatile} : (tensor<2x3x3xf32>) -> tensor<2x3x3x!quant.uniform<i8:f32, 3.000000e-01:1>>
    %16 = "quantfork.dcast"(%15) : (tensor<2x3x3x!quant.uniform<i8:f32, 3.000000e-01:1>>) -> tensor<2x3x3xf32>
    return %16 : tensor<2x3x3xf32>
  }

  // CHECK: quantized_dot_general_fn_1