// CHECK: mhlo.reduce
    // CHECK: mhlo.dot_general
    // CHECK: mhlo.transpose
    %0 = "tf.BatchMatMulV2"(%arg0, %arg1) {T = f32, adj_x = false, adj_y = false, grad_x = false, grad_y = false, device = ""} : (tensor<1x4x2xf32>, tensor<3x2x4xf32>) -> tensor<3x4x4xf32>
    func.return %0 : tensor<3x4x4xf32>
  }

  // CHECK-LABEL: approx_topk

// CHECK-NEXT:    return %2 : tensor<3x2xf32>
}

func.func @transpose_3d(%arg0: tensor<1x2x3xf32>) -> tensor<3x2x1xf32> {
  %0 = "mhlo.transpose"(%arg0) <{permutation = dense<[2, 1, 0]> : tensor<3xi64>}> : (tensor<1x2x3xf32>) -> tensor<3x2x1xf32>
  func.return %0 : tensor<3x2x1xf32>

// CHECK-LABEL:   transpose_3d

  %lookup0 = "tfl.embedding_lookup"(%index0, %value0) : (tensor<5xi32>, tensor<3x2xf32>) -> tensor<5x2xf32>
  %lookup1 = "tfl.embedding_lookup"(%index1, %value1) : (tensor<3xi32>, tensor<2x2x2xf32>) -> tensor<3x2x2xf32>
  func.return %lookup0, %lookup1 : tensor<5x2xf32>, tensor<3x2x2xf32>

        none, none, none, none) -> tensor<1x2x3xf32>
  %17 = "quantfork.stats"(%16) {
    layerStats = dense<[0.0, 1.0]> : tensor<2xf32>,
    axisStats = dense<[
      [-1.0, 1.0],
      [-8.0, 8.0],
      [-0.5, 0.5]
    ]> : tensor<3x2xf32>, axis = 2 : i64
  } : (tensor<1x2x3xf32>) -> tensor<1x2x3xf32>
  func.return %17 : tensor<1x2x3xf32>

  
  %40 = "tfl.read_variable"(%4) : (tensor<!tf_type.resource>) -> tensor<1x2x3xf32>
  %41 = "quantfork.stats"(%40) {layerStats = dense<[0.0, 1.0]> : tensor<2xf32>} : (tensor<1x2x3xf32>) -> tensor<1x2x3xf32>
  %42 = "tfl.concatenation"(%41, %0) {axis = 1 : i32, fused_activation_function = "NONE"} : (tensor<1x2x3xf32>, tensor<1x2x3xf32>) -> tensor<1x4x3xf32>

  module {
    func.func @constant_add() -> (tensor<3x2xf32>) {
      %cst1 = stablehlo.constant dense<2.4> : tensor<3x2xf32>
      %cst2 = stablehlo.constant dense<5.7> : tensor<3x2xf32>
      %add = stablehlo.add %cst1, %cst2 : (tensor<3x2xf32>, tensor<3x2xf32>) -> tensor<3x2xf32>
      func.return %add : tensor<3x2xf32>
    }
  }
)mlir";

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

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

  module {
    func.func @main(%arg0: tensor<2x2x2xf32>, %arg1: tensor<2x2x2xf32>) -> tensor<2x2x2xf32> attributes {_from_xla_call_module} {
      %0 = stablehlo.dot_general %arg0, %arg1,
        batching_dims = [0] x [0],
        contracting_dims = [2] x [1],
        precision = [DEFAULT, DEFAULT]
      : (tensor<2x2x2xf32>, tensor<2x2x2xf32>) -> tensor<2x2x2xf32>
        return %0 : tensor<2x2x2xf32>
    }
  }
)mlir";

    func.func @main() -> (tensor<1x4x4xf32>) {
      %%arg0 = "tf.Const"() {value = dense<-3.0> : tensor<1x4x2xf32>} : () -> tensor<1x4x2xf32>
      %%arg1 = "tf.Const"() {value = dense<-3.0> : tensor<1x2x4xf32>} : () -> tensor<1x2x4xf32>

      %%1 = "tf.%s"(%%arg0, %%arg1) {T = f32, adj_x = false, adj_y = false, grad_x = false, grad_y = false, device = ""} : (tensor<1x4x2xf32>, tensor<1x2x4xf32>) -> tensor<1x4x4xf32>

  %0 = "tfl.var_handle"() {container = "", shared_name = "states"} : () -> tensor<!tf_type.resource<tensor<1x2x3xf32>>>
  %1 = "tfl.read_variable"(%0) : (tensor<!tf_type.resource<tensor<1x2x3xf32>>>) -> tensor<1x2x3xf32>
  %2 = "tfl.quantize"(%1) {qtype = tensor<1x2x3x!quant.uniform<u8<1:255>:f32, 0.047244094488188976:128>>} : (tensor<1x2x3xf32>) -> tensor<1x2x3x!quant.uniform<u8<1:255>:f32, 0.047244094488188976:128>>