// 262657 = 1 + 512 + 512 * 512
    // expected-remark@+1 {{Cost: 262657}}
    %2 = "tf.MatMul"(%arg, %1) {device = "/job:localhost/replica:0/task:0/device:CPU:0", transpose_a = false, transpose_b = false} : (tensor<?x512xf32>, tensor<512x512xf32>) -> tensor<?x512xf32>
    // expected-remark@+1 {{Cost: 512}}
    func.return %2 : tensor<?x512xf32>
}

func.func @lift_float_matmul(%arg0: tensor<1x12x12x512xf32>) -> (tensor<*xf32>, tensor<*xf32>) {
  %cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<512x512xf32>} : () -> tensor<512x512xf32>
  %out_1 = "tf.MatMul"(%arg0, %cst) {
    device = "", transpose_a = false, transpose_b = false
  } : (tensor<1x12x12x512xf32>, tensor<512x512xf32>) -> tensor<*xf32>
  %out_2 = "tf.MatMul"(%arg0, %arg0) {
    device = "", transpose_a = false, transpose_b = true

func.func @add_with_activation_transpose_rank_two(%arg0: tensor<1x2xf32>) -> tensor<2x1xf32> {
  %0 = stablehlo.constant dense<2.000000e+00> : tensor<2x1xf32>
  %1 = stablehlo.transpose %arg0, dims = [1, 0] : (tensor<1x2xf32>) -> tensor<2x1xf32>
  %2 = stablehlo.add %1, %0 : tensor<2x1xf32>
  return %2 : tensor<2x1xf32>
}
// CHECK: %[[TRANSPOSE_0:.+]] = stablehlo.transpose

    %2 = "tf.XlaCallModule"(%arg0, %1, %0) <{Sout = [#tf_type.shape<?x2>], module = "", version = 9 : i64}> {_entry_function = @composite_dot_general_with_bias_and_relu6_dynamic_fn_2, _original_entry_function = "composite_dot_general_with_bias_and_relu6_dynamic_fn_2", _tfl_quant_trait = "fully_quantizable"} : (tensor<?x2xf32>, tensor<2x2xf32>, tensor<2xf32>) -> tensor<?x2xf32>

    %cst = "tf.Const"() {device = "", value = dense<1.000000e+01> : tensor<512x512xf32>} : () -> tensor<512x512xf32>
    %cst_sharded = "tf.XlaSharding"(%cst) {_XlaSharding = "\08\03\1A\03\01\04\02\22\08\00\04\01\05\02\06\03\070\01", device = "", sharding = "\08\03\1A\03\01\04\02\22\08\00\04\01\05\02\06\03\070\01", unspecified_dims = []} : (tensor<512x512xf32>) -> tensor<512x512xf32>

  %2 = "tf.Reshape"(%0, %cst_0) : (tensor<1x2xf32>, tensor<2xi64>) -> tensor<2x1xf32>
  func.return %1, %2 : tensor<2x1xf32>, tensor<2x1xf32>

// CHECK:  %cst = arith.constant
// CHECK:  %[[FQ:.*]] = "tf.FakeQuantWithMinMaxVars"(%arg0, %arg1, %arg2)
// CHECK:  %[[R1:.*]] = "tf.Reshape"(%[[FQ]], %cst)
// CHECK-SAME: tensor<2x1xf32>
// CHECK:  %[[R2:.*]] = "tf.Reshape"(%[[FQ]], %cst)

  %2 = "tf.Reshape"(%0, %cst_0) : (tensor<1x2xf32>, tensor<2xi64>) -> tensor<2x1xf32>
  func.return %1, %2 : tensor<2x1xf32>, tensor<2x1xf32>

// CHECK:  %cst = arith.constant
// CHECK:  %[[FQ:.*]] = "tf.FakeQuantWithMinMaxVars"(%arg0, %arg1, %arg2)
// CHECK:  %[[R1:.*]] = "tf.Reshape"(%[[FQ]], %cst)
// CHECK-SAME: tensor<2x1xf32>
// CHECK:  %[[R2:.*]] = "tf.Reshape"(%[[FQ]], %cst)

      %2 = "tf.XlaCallModule"(%arg0, %1, %0) <{Sout = [#tf_type.shape<?x2>], module = "", version = 9 : i64}> {_entry_function = @composite_dot_general_fn_1, _original_entry_function = "composite_dot_general_fn_1", _tfl_quant_trait = "fully_quantizable", _quantization_method = "weight_only_ptq { }"} : (tensor<?x2xf32>, tensor<2x2xf32>, tensor<2xf32>) -> tensor<?x2xf32>
      return %2 : tensor<?x2xf32>

func.func @RemoveRedundantUnpackPack(%arg0: tensor<2x5xf32>) -> tensor<2x5xf32> {
  %0:2 = "tfl.unpack"(%arg0) {axis = 0 : i32, num = 2 : i32} : (tensor<2x5xf32>) -> (tensor<5xf32>, tensor<5xf32>)
  %1 = "tfl.pack"(%0#0, %0#1) {axis = 0 : i32, values_count = 2 : i32} : (tensor<5xf32>, tensor<5xf32>) -> (tensor<2x5xf32>)
  func.return %1: tensor<2x5xf32>
  // CHECK-NOT: pack
  // CHECK: return %arg0 : tensor<2x5xf32>
}

// -----

    %2 = func.call @func_2_CPU_FLOAT(%0, %1) {tac.interface_name = "func_2"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<2x1xf32>
    func.return %2 : tensor<2x1xf32>
  }

  func.func private @func_2_CPU_FLOAT(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>) -> tensor<2x1xf32> attributes {tac.device = "CPU", tac.inference_type = "FLOAT", tac.interface_name = "func_2"} {