// Use other output
  %3:6 = "tf.FusedBatchNormV3"( %2#0, %arg1, %arg2, %arg3, %arg4) {T = "tfdtype$DT_FLOAT", U = "tfdtype$DT_FLOAT", data_format = "NHWC", epsilon = 0.001 : f32, is_training = false} : (tensor<8x8x8x8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>) -> (tensor<8x8x8x8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>)

  %5 = shape.shape_of %4 : tensor<?x10xf32> -> tensor<2xindex>
  %6 = stablehlo.dynamic_broadcast_in_dim %1, %5, dims = [1] : (tensor<10xf32>, tensor<2xindex>) -> tensor<?x10xf32>
  %7 = stablehlo.add %4, %6 : tensor<?x10xf32>
  %8 = stablehlo.clamp %2, %7, %3 : (tensor<f32>, tensor<?x10xf32>, tensor<f32>) -> tensor<?x10xf32>
  func.return %8: tensor<?x10xf32>
}

// CHECK:           %[[VAL_4:.*]]:2 = call @func_0_GPU_FLOAT(%[[VAL_0]], %[[VAL_1]], %[[VAL_2]], %[[VAL_3]]) {tac.device = "GPU", tac.inference_type = "FLOAT", tac.interface_name = "func_0"} : (tensor<1xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xf32>) -> (tensor<1xf32>, tensor<1xf32>)

}

// CHECK-LABEL: func @add_f32
func.func @add_f32(%arg0: tensor<3xf32>, %arg1: tensor<3xf32>) -> tensor<3xf32> {
  // CHECK-DAG: [[ADD:%.+]] = "tf.AddV2"(%arg0, %arg1) : (tensor<3xf32>, tensor<3xf32>) -> tensor<3xf32>
  %add = "tf.Add"(%arg0, %arg1) : (tensor<3xf32>, tensor<3xf32>) -> tensor<3xf32>
  func.return %add : tensor<3xf32>
}

// CHECK-LABEL: func @BiasAddGrad_NHWC

    %1 = "tf.AddV2"(%0, %cst_0) {device = ""} : (tensor<?x?x?x?x3xf32>, tensor<f32>) -> tensor<?x?x?x?x3xf32>
    %2 = "tf.Floor"(%1) {device = ""} : (tensor<?x?x?x?x3xf32>) -> tensor<?x?x?x?x3xf32>
    %3 = "tf.ClipByValue"(%2, %cst_5, %cst_6) {device = ""} : (tensor<?x?x?x?x3xf32>, tensor<f32>, tensor<f32>) -> tensor<?x?x?x?x3xf32>

  %outputs_50 =  "tf.Mul"(%outputs_30, %outputs_48) {device = ""} : (tensor<*xf32>, tensor<*xf32>) -> tensor<*xf32>
  %outputs_52 =  "tf.Reshape"(%outputs_50, %outputs_0) {device = ""} : (tensor<*xf32>, tensor<2xi32>) -> tensor<*xf32>
  %outputs_54 =  "tf.MatMul"(%outputs_40, %outputs_52) {device = "", transpose_a = false, transpose_b = false} : (tensor<*xf32>, tensor<*xf32>) -> tensor<*xf32>

  // CHECK: %[[IND_SLICE0:.*]] = "tf.Slice"(%[[INDS]], %[[IND_SLICE0_START]], %[[IND_SLICE0_SIZE]]) : (tensor<2xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<1xi32>
  // CHECK: %[[SLICE0_START:.*]] = "tf.ConcatV2"(%[[IND_SLICE0]],
  // CHECK: %[[OLD_SLICE0:.*]] = "tf.Slice"(%[[READ2]], %[[SLICE0_START]],
  // CHECK-SAME: (tensor<5x3xf32>, tensor<2xi32>, tensor<2xi32>) -> tensor<1x3xf32>

    // CHECK-DAG: %[[SIGN:.*]] = "tf.Sign"(%[[PROX]]) : (tensor<4xf32>) -> tensor<4xf32>
    // CHECK-DAG: %[[ABS:.*]] = "tf.Abs"(%[[PROX]]) : (tensor<4xf32>) -> tensor<4xf32>
    // CHECK-DAG: %[[SCALED_L1:.*]] = "tf.Mul"(%[[ADAGRAD_LR]], %[[L1]]) : (tensor<4xf32>, tensor<f32>) -> tensor<4xf32>
    // CHECK-DAG: %[[PROX_NEW:.*]] = "tf.Sub"(%[[ABS]], %[[SCALED_L1]]) : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>