// -----

func.func @pack_CPU(%arg0: tensor<100xf32>, %arg1: tensor<100xf32>) -> tensor<2x100xf32> attributes {tac.device = "CPU", tac.interface_name = "func_2"} {
  // CHECK: tac.cost = 1.000000e+02
  %0 = "tfl.pack"(%arg0, %arg1) {axis = 0 : i32, tac.device = "CPU", values_count = 2 : i32} : (tensor<100xf32>, tensor<100xf32>) -> tensor<2x100xf32>
  func.return %0 : tensor<2x100xf32>
}

func.func @pack(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>) -> tensor<2x1xf32> {
  %0 = "tfl.pack"(%arg0, %arg1) {axis = 0 : i32, values_count = 2 : i32} : (tensor<1xf32>, tensor<1xf32>) -> tensor<2x1xf32>
  func.return %0 : tensor<2x1xf32>
}

// CHECK:   func @pack(%[[VAL_0:.*]]: tensor<1xf32>, %[[VAL_1:.*]]: tensor<1xf32>) -> tensor<2x1xf32> {
// CHECK-DAG:       %[[VAL_2:.*]] = "tfl.pseudo_const"{{.*}}dense<1> : tensor<4xi32>

    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<2x1024xf32>} : () -> tensor<2x1024xf32>
    %1 = "tf.PartitionedCall"(%arg0, %cst_0) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_matmul_fn} : (tensor<1x2x2x3xf32>, tensor<2x1024xf32>) -> tensor<*xf32>
    func.return %1: tensor<*xf32>
  }

// CHECK: %[[v2:.*]] = "tf.Reshape"(%[[v1]], %[[cst_1]]) : (tensor<2x1x2xf32>, tensor<4xi64>) -> tensor<2x1x1x2xf32>
// CHECK: return %[[v2]] : tensor<2x1x1x2xf32>
}

func.func @einsum_with_runtime_outputshape1(%arg0 : tensor<?x36x32xf32>, %arg1 : tensor<?x36x?x32xf32>) -> tensor<?x36x?xf32> {

func.func @mean_4d_no_keepdim(%arg0: tensor<1x48x48x512xf32>) -> tensor<1x512xf32> {
  %cst = arith.constant dense<[1, 2]> : tensor<2xi32>
  %0 = "tfl.mean"(%arg0, %cst) {keep_dims = false} : (tensor<1x48x48x512xf32>, tensor<2xi32>) -> tensor<1x512xf32>
  func.return %0 : tensor<1x512xf32>
}

// CHECK:       func @mean_4d_no_keepdim([[VAL_0:%.*]]: tensor<1x48x48x512xf32>) -> tensor<1x512xf32> {

func.func @TfBatchMatMulV2(%arg0: tensor<4x128x2xf32>, %arg1:  tensor<2x1xf32>) -> tensor<4x128x1xf32> {
  %0 = "tfl.custom"(%arg0, %arg1) {
    custom_code = "FlexBatchMatMulV2",
    custom_option = #tfl<const_bytes : "0x0D42617463684D61744D756C56320038120D42617463684D61744D756C56321A001A002A070A0154120230012A0B0A0561646A5F78120228002A0B0A0561646A5F791202280032000002493B1414042801">
  } : (tensor<4x128x2xf32>, tensor<2x1xf32>) -> tensor<4x128x1xf32>

/// (realValue: FloatAttr, quantizedElementType: UniformQuantizedType[i8:f32])
///   -> (IntegerAttr, outConvertedType: i8)
/// 2. realValue is an elements attribute:
/// (realValue: DenseElementsAttr[tensor<2x2xf32>],
///  quantizedElementType: UniformQuantizedType[i8:f32])
///   -> (DenseElementsAttr[tensor<2x2xi8>], outConvertedType: tensor<2x2xi8>)
Attribute quantizeAttr(Attribute realValue,

  %3 = "tfl.pack"(%1, %2) {axis = 0 : i32, per_device_costs = {CPU = 2.0 : f32, GPU = -1.0 : f32}, values_count = 2 : i32, tac.device = "CPU"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<2x1xf32>
  func.return %3 : tensor<2x1xf32>
})";
  const std::string kExpectedFB = CreateRuntimeMetadata();
  mlir::DialectRegistry registry;
  registry.insert<mlir::TFL::TensorFlowLiteDialect, mlir::arith::ArithDialect,

// CHECK: %0 = "tfl.pseudo_const"() <{value = dense<2> : tensor<2xi32>}> : () -> tensor<2xi32>
// CHECK: %1 = "tfl.reshape"(%arg0, %0) : (tensor<1x4xf32>, tensor<2xi32>) -> tensor<2x2xf32>

{
  "version": 3,
  "operator_codes": [
    {
      "builtin_code": "RESHAPE"
    }
  ],
  "subgraphs": [
    {
      "tensors": [
        {
          "shape": [1, 4],

    %4:4 = "tf.CustomAggregator"(%3) {calibration_method = 1 : i32, id = "1", max_percentile = 0.000000e+00 : f32, min_percentile = 0.000000e+00 : f32, num_bins = 0 : i32} : (tensor<3x11xf32>) -> (tensor<3x11xf32>, tensor<f32>, tensor<f32>, tensor<*xi64>)
    %5 = stablehlo.add %4#0, %1 : tensor<3x11xf32>
    %6 = stablehlo.multiply %5, %1 : tensor<3x11xf32>