// Unsupported strides
  %2 = "tf.MaxPool"(%arg0) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", ksize = [1, 3, 6, 1], padding = "VALID", strides = [1, 3, 1, 3]} : (tensor<1x1x1x16xf32>) -> tensor<1x1x1x16xf32>

  %5 = arith.addf %0, %1 : tensor<1x1x1x16xf32>
  %6 = arith.addf %2, %5 : tensor<1x1x1x16xf32>
  func.return %6 : tensor<1x1x1x16xf32>

// CHECK:  [[VAL_1:%.*]] = "tfl.reshape"(%2, %[[CST]]) {tac.device = "GPU",  tac.inference_type = "FLOAT"} : (tensor<1xf32>, tensor<4xi32>) -> tensor<1x1x1x1xf32>
// CHECK:  [[VAL_2:%.*]] = "tfl.concatenation"([[VAL_0]], [[VAL_1]]) <{axis = 3 : i32, fused_activation_function = "NONE"}> {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<1x1x1x1xf32>, tensor<1x1x1x1xf32>) -> tensor<1x1x1x2xf32>

    %1 = stablehlo.constant dense<1.000000e+03> : tensor<1x1x1x1xf32>  // Input inverse scale.
    %2 = stablehlo.constant dense<-128> : tensor<1x1x1x1xi8>  // Input zero point.
    %3 = stablehlo.constant dense<1> : tensor<3x3x4x4xi8>  // Quantized filter tensor.
    %4 = stablehlo.constant dense<3.000000e+03> : tensor<1x1x1x4xf32>
    %5 = stablehlo.constant dense<4.000000e+03> : tensor<1x1x1x1xf32>  // Output inverse scale.

// input shape.

// CHECK-LABEL: reduce_window_max_activation_transpose
func.func @reduce_window_max_activation_transpose(%arg0: tensor<1x16x16x4xf32>) -> tensor<1x4x8x8xf32> {
  %0 = stablehlo.constant dense<0xFF800000> : tensor<f32>  // -inf
  %1 = stablehlo.transpose %arg0, dims = [0, 3, 1, 2] : (tensor<1x16x16x4xf32>) -> tensor<1x4x16x16xf32>
  %2 = "stablehlo.reduce_window"(%1, %0) ({
  ^bb0(%arg1: tensor<f32>, %arg2: tensor<f32>):

                                      interior_padding = #vhlo.tensor_v1<dense<0> : tensor<3xi64>>}> : (tensor<1x160x1xf32>, tensor<f32>) -> tensor<1x161x1xf32>
  return %0 : tensor<1x161x1xf32>
}

//CHECK:func.func private @pad(%arg0: tensor<1x160x1xf32>, %arg1: tensor<f32>) -> tensor<1x161x1xf32> {
//CHECK-NEXT: %0 = "vhlo.pad_v1"(%arg0, %arg1) <{edge_padding_high = #vhlo.tensor_v1<dense<0> : tensor<3xi64>>,

  return %2 : tensor<1x1x1x4xf32>
}
func.func private @XlaCallModule_aten.avg_pool2d.default.impl_2(%arg0: tensor<1x1x1x8xf32>) -> tensor<1x1x1x4xf32>

// CHECK-LABEL: avg_pool2d_3
// CHECK: %cst = arith.constant dense<[0, 2, 3, 1]> : tensor<4xi32>
// CHECK: %0 = "tfl.transpose"(%arg0, %cst) : (tensor<1x1x1x8xf32>, tensor<4xi32>) -> tensor<1x1x8x1xf32>

// CHECK:           %[[VAL_6:.*]] = "tfl.reshape"(%[[VAL_1]], %[[VAL_2]]) : (tensor<1xf32>, tensor<4xi32>) -> tensor<1x1x1x1xf32>
// CHECK:           %[[VAL_7:.*]] = "tfl.concatenation"(%[[VAL_5]], %[[VAL_6]]) <{axis = 3 : i32, fused_activation_function = "NONE"}> : (tensor<1x1x1x1xf32>, tensor<1x1x1x1xf32>) -> tensor<1x1x1x2xf32>

    %1 = "tfl.reshape"(%arg1, %cst) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<1xf32>, tensor<4xi32>) -> tensor<1x1x1x1xf32>
    %2 = "tfl.concatenation"(%0, %1) {axis = 3 : i32, fused_activation_function = "NONE", tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<1x1x1x1xf32>, tensor<1x1x1x1xf32>) -> tensor<1x1x1x2xf32>

// CHECK:           %[[VAL_8:.*]] = "tfl.reshape"(%[[VAL_7]], %[[VAL_3]]) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<1x1x1x2xf32>, tensor<1xi32>) -> tensor<2xf32>
// CHECK:           %[[VAL_9:.*]] = "tfl.reshape"(%[[VAL_8]], %[[VAL_4]]) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<2xf32>, tensor<2xi32>) -> tensor<2x1xf32>

  %custom_2 = "tfl.custom"(%arg0, %dq_w) {custom_code = "CustomTestOp", custom_option = #tfl<const_bytes : "0x">} : (tensor<1x1x1x1xf32>, tensor<1024x1x1x1xf32>) -> tensor<*xf32>
  %custom_3 = "tfl.custom"(%arg0, %dq_w) {custom_code = "CustomTestOp", custom_option = #tfl<const_bytes : "0x">} : (tensor<1x1x1x1xf32>, tensor<1024x1x1x1xf32>) -> tensor<*xf32>