// CHECK-NEXT:    "vhlo.return_v1"(%1) : (tensor<f32>) -> ()
  // CHECK-NEXT: }) : (tensor<1x16x16x320xf32>, tensor<f32>) -> tensor<1x320xf32>
  %0 = stablehlo.reduce(%arg0 init: %arg1) applies stablehlo.add across dimensions = [1, 2] : (tensor<1x16x16x320xf32>, tensor<f32>) -> tensor<1x320xf32>
  return %0 : tensor<1x320xf32>
}

// -----

// CHECK-LABEL: op_with_region_mixed_tfl_shlo_tfl

  %4 = "tfl.relu"(%3) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<128x128xf32>) -> tensor<128x128xf32>
  %5 = "tfl.pseudo_const"() {value = dense<1.000000e+00> : tensor<128x128xf32>} : () -> tensor<128x128xf32>
  %6 = "tfl.pseudo_const"() {value = dense<1.000000e+00> : tensor<128xf32>} : () -> tensor<128xf32>

    return %1 : tensor<1x3xf32>
  }

  func.func private @composite_dot_general_fn(%arg0: tensor<1x2xf32>, %arg1: tensor<2x3xf32>) -> tensor<1x3xf32> attributes {_from_xla_call_module} {
    %0 = stablehlo.dot_general %arg0, %arg1, contracting_dims = [1] x [0] : (tensor<1x2xf32>, tensor<2x3xf32>) -> tensor<1x3xf32>
    return %0 : tensor<1x3xf32>

  %1 = "tf.BiasAdd"(%0, %arg0) {data_format = "NHWC"} : (tensor<*xf32>, tensor<128xf32>) -> tensor<*xf32>
  %2 = "tf.Identity"(%1) : (tensor<*xf32>) -> tensor<*xf32>
  func.return %2 : tensor<*xf32>
}

// CHECK-LABEL: conv2DBiasAdd_reluActivation
func.func @conv2DBiasAdd_reluActivation(%arg0: tensor<128xf32>, %arg1: tensor<1x1x3x128xf32>, %arg2: tensor<8x32x32x3xf32>) -> (tensor<*xf32>) {

func.func @single_state_single_dataset_type_multiple_arguments(
      %arg0: tensor<!tf_type.variant>,
      %arg1: tensor<i64>,
      %arg2: tensor<!tf_type.resource<tensor<64xf32>>>,
      %arg3: tensor<!tf_type.resource<tensor<128xf32>>>
    ) {
    // CHECK:      %[[ANON_ITER:[0-9]*]] = "tf.AnonymousIteratorV3"

  // CHECK-NEXT: [[RESULT:%.*]] = "tf.BiasAdd"([[SQUEEZE]], [[BIAS]]) : (tensor<1x128x8xf32>, tensor<8xf32>) -> tensor<1x128x8xf32>
  // CHECK-NEXT: return [[RESULT]] : tensor<1x128x8xf32>
}

func.func @testDilatedConv1DExpandW(%arg0: tensor<1x128x3xf32>, %arg1: tensor<5x1x3x8xf32>) -> tensor<1x128x8xf32> {
  %cst = "tf.Const"() {value = dense<0> : tensor<1x2xi32>} : () -> tensor<1x2xi32>

  %flow_1 = "tf.TensorArrayWriteV3"(%handle, %index, %value, %flow) {device = "/job:localhost/replica:0/task:0/device:CPU:0"} : (tensor<2x!tf_type.resource<tensor<1x512xf32>>>, tensor<i32>, tensor<1x512xf32>, tensor<f32>) -> tensor<f32>

  %prelu = "tfl.prelu"(%arg0, %cst) : (tensor<1x10x10x3xf32>, tensor<1x1x3xf32>) -> tensor<1x10x10x3xf32>
  func.return %prelu : tensor<1x10x10x3xf32>

// CHECK: %[[cst:.*]] = arith.constant dense<[{{\[}}[1.66394591, 3.61694336, 2.0382936]]]> : tensor<1x1x3xf32>

    %3 = stablehlo.concatenate %2, %1, dim = 0 : (tensor<1x3xf32>, tensor<1x3xf32>) -> tensor<2x3xf32>
    return %3 : tensor<2x3xf32>
  }
  func.func private @composite_dot_general_fn_1(%arg0: tensor<1x2xf32>, %arg1: tensor<2x3xf32>) -> tensor<1x3xf32> attributes {_from_xla_call_module, tf_quant.composite_function} {

}

// -----

module {
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>
}