func.func @Int64SliceBeginSize(%arg0: tensor<4x128x32xf32>) -> tensor<1x128x32xf32> {
  %0 = "tfl.pseudo_const"() {value = dense<0> : tensor<3xi64>} : () -> tensor<3xi64>
  %1 = "tfl.pseudo_const"() {value = dense<[1, 128, 32]> : tensor<3xi64>} : () -> tensor<3xi64>
  %2 = "tfl.slice"(%arg0, %0, %1) : (tensor<4x128x32xf32>, tensor<3xi64>, tensor<3xi64>) -> tensor<1x128x32xf32>
  func.return %2 : tensor<1x128x32xf32>

  // CHECK-NEXT: [[RESULT:%.*]] = "tf.BiasAdd"([[SQUEEZE]], [[BIAS]]) : (tensor<1x128x128xf32>, tensor<128xf32>) -> tensor<1x128x128xf32>
  // CHECK-NEXT: return [[RESULT]] : tensor<1x128x128xf32>
}

func.func @testDilatedDepthWiseConvWithExpandSqueeze1(%arg0: tensor<1x128x128xf32>, %arg1: tensor<5x5x1x1xf32>, %arg2: tensor<128xf32>) -> tensor<1x128x128xf32> {

  %1 = "tfl.fully_connected"(%0, %arg1, %arg2) {fused_activation_function = "NONE", keep_num_dims = true, weights_format = "DEFAULT"} : (tensor<1x128x64xf32>, tensor<32x64xf32>, tensor<32xf32>) -> tensor<1x128x32xf32>
  func.return %1 : tensor<1x128x32xf32>
  // CHECK-DAG: %[[CST:.*]] = arith.constant dense<[1, 128, 64]> : tensor<3xi32>

// CHECK:      %[[R1:.*]] = "mhlo.concatenate"(%arg1, %arg3, %arg5, %arg7) <{dimension = 0 : i64}> : (tensor<1x128x72xf32>, tensor<1x128x72xf32>, tensor<1x128x72xf32>, tensor<1x128x72xf32>) -> tensor<4x128x72xf32>
// CHECK:      %[[R2:.*]] = "mhlo.dot_general"(%[[R0]], %[[R1]]) <{
// CHECK-SAME:   dot_dimension_numbers = #mhlo.dot<
// CHECK-SAME:     lhs_batching_dimensions = [0],

// CHECK-NOT: "tfl.batch_matmul"
func.func @Batchmatmul2Fullyconnected(%arg0: tensor<4x128x2xf32>) -> (tensor<4x128x1xf32>) {
  %0 = arith.constant dense<[[1.0], [2.0]]> : tensor<2x1xf32>
  %1 = "tfl.batch_matmul"(%arg0, %0) {adj_x = false, adj_y = false, asymmetric_quantize_inputs = false} : (tensor<4x128x2xf32>, tensor<2x1xf32>) -> tensor<4x128x1xf32>
  func.return %1 : tensor<4x128x1xf32>

  %7 = "tfl.add"(%1, %6) {tac.device = "GPU", tac.inference_type = "FLOAT", fused_activation_function = "NONE"} : (tensor<1x128x128xf32>, tensor<1x128x128xf32>) -> tensor<1x128x128xf32>
  func.return %7 : tensor<1x128x128xf32>
}

// CHECK:   func @norm1(%[[VAL_0:.*]]: tensor<1x128x128xf32>) -> tensor<1x128x128xf32> {

func.func private @func_20_GPU_FLOAT(%arg0: tensor<128x128xf32>, %arg1: tensor<3xi32>) -> tensor<1x128x128xf32> attributes {tac.device = "GPU", tac.inference_type = "FLOAT", tac.interface_name = "func_20"} {
  %0 = "tfl.reshape"(%arg0, %arg1) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<128x128xf32>, tensor<3xi32>) -> tensor<1x128x128xf32>
  func.return %0 : tensor<1x128x128xf32>
}

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>

  %0 = "tf.GatherV2"(%cst_0, %arg0, %cst) {batch_dims = 0 : i64, device = ""} : (tensor<128x32xf32>, tensor<6xi64>, tensor<i32>) -> tensor<6x32xf32>
  return %0 : tensor<6x32xf32>

// CHECK-DAG: %[[CST:.*]] = "tf.Const"() {{.*}} : () -> tensor<i32>
// CHECK-DAG: %[[CST_1:.*]] = "tf.Const"() {{.*}} : () -> tensor<128x32xf32>

    func.func @main(%a : tensor<5x14x1xf32>, %b : tensor<1x14x32xf32>) -> tensor<?x?x?xf32> attributes {tf.entry_function = {control_outputs = "", inputs = "i,j", outputs = "identity_RetVal"}} {
      %c = "mhlo.einsum"(%a, %b) {einsum_config = "bji,bjk->bik"} : (tensor<5x14x1xf32>, tensor<1x14x32xf32>) -> tensor<?x?x?xf32>
      return %c : tensor<?x?x?xf32>
    }
  })";