%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
    %14 = stablehlo.multiply %12, %13 : tensor<1x4x3xf32>  // s1 * s2

  %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>

      %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>

      return %0 : tensor<1x1x4xf32>
    }
  )mlir";

  const OwningOpRef<ModuleOp> module_op =
      ParseModuleOpString(kXlaCallModuleOpWithQuantizationMethodAttr);
  ASSERT_TRUE(module_op);

  // 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

  %2 = "mhlo.dot_general"(%arg2, %arg3) {
    dot_dimension_numbers = #mhlo.dot<
      lhs_contracting_dimensions = [2],
      rhs_contracting_dimensions = [0]
  >} : (tensor<3x1x512xf32>, tensor<512x13xf32>) -> tensor<3x1x13xf32>
  %r = "mhlo.concatenate"(%0, %1, %2) <{dimension = 0 : i64}> : (tensor<1x1x13xf32>, tensor<2x1x13xf32>, tensor<3x1x13xf32>) -> tensor<6x1x13xf32>

    %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>

    func.return %2 : tensor<2x100xf32>
  }
  func.func @func_2_CPU_FLOAT(%arg0: tensor<100xf32>, %arg1: tensor<100xf32>) -> tensor<2x100xf32> attributes {tac.cost = 2.000000e+01 : f32, tac.device = "CPU", tac.inference_type = "FLOAT", tac.interface_name = "func_2"} {

  %prelu = "tfl.prelu"(%arg0, %cst) : (tensor<1x10x10x3xf32>, tensor<1x1x3xf32>) -> tensor<1x10x10x3xf32>
  func.return %prelu : tensor<1x10x10x3xf32>

// CHECK: %[[cst:.*]] = arith.constant dense<[{{\[}}[1.66394591, 3.61694336, 2.0382936]]]> : tensor<1x1x3xf32>