%1 = "tfl.mul"(%0, %cst_2) {fused_activation_function = "NONE"} : (tensor<1x8x8x256xf32>, tensor<1x1x1x256xf32>) -> tensor<1x8x8x256xf32>
  func.return %1 : tensor<1x8x8x256xf32>

// CHECK-DAG: %[[CST_0:.*]] = arith.constant dense<2.800000e+00> : tensor<256x3x3x207xf32>

  %0:4 = "tfl.split"(%arg0, %arg1) {num_splits = 4 : i32, tac.device = "CPU"} : (tensor<i32>, tensor<1x8x8x1024xf32>) -> (tensor<1x8x8x256xf32>, tensor<1x8x8x256xf32>, tensor<1x8x8x256xf32>, tensor<1x8x8x256xf32>)
  func.return %0#0, %0#1, %0#3 : tensor<1x8x8x256xf32>, tensor<1x8x8x256xf32>, tensor<1x8x8x256xf32>
}

// Confirm a wide array of attribute survives the round-trip
func.func @main(tensor<1x6x6x16xf32>) -> tensor<1x1x1x16xf32> {
^bb0(%arg0: tensor<1x6x6x16xf32>):
  // CHECK: "tfl.average_pool_2d"(%{{.*}}) <{filter_height = 3 : i32, filter_width = 6 : i32, fused_activation_function = "NONE", padding = "VALID", stride_h = 3 : i32, stride_w = 1 : i32}> : (tensor<1x6x6x16xf32>) -> tensor<1x1x1x16xf32>

^bb0(%arg0: tensor<1x6x6x16x!quant.uniform<u8:f32, 0.1>>, %arg1: tensor<1x6x6x16x!quant.uniform<u8:f32, 0.1>>):
  %0 = "tfl.dequantize"(%arg0) : (tensor<1x6x6x16x!quant.uniform<u8:f32, 0.1>>) -> tensor<1x6x6x16xf32>
  %1 = "tfl.dequantize"(%arg1) : (tensor<1x6x6x16x!quant.uniform<u8:f32, 0.1>>) -> tensor<1x6x6x16xf32>
  %2 = "tfl.minimum"(%0, %1) : (tensor<1x6x6x16xf32>, tensor<1x6x6x16xf32>) -> tensor<1x6x6x16xf32>

// RUN: flatbuffer_translate -mlir-to-tflite-flatbuffer %s -o - | flatbuffer_to_string - | FileCheck %s

func.func @main(tensor<1x6x6x16xf32>) -> tensor<1x1x1x16xf32> {
^bb0(%arg0: tensor<1x6x6x16xf32>):
  // CHECK:      {
  // CHECK-NEXT:   version: 3,
  // CHECK-NEXT:   operator_codes: [ {
  // CHECK-NEXT:     deprecated_builtin_code: 1,
  // CHECK-NEXT:     version: 1,
  // CHECK-NEXT:     builtin_code: AVERAGE_POOL_2D
  // CHECK-NEXT:   } ],

                             %arg2: tensor<256xf32>,          // batch_norm args
                             %arg3: tensor<7x7x3x64xf32>,    // conv filter #0
                             %arg4: tensor<1x1x64x256xf32>   // conv filter #1
                            ) -> tensor<?x256xf32> {

  // This is a simplified ResNet layer that gets input in NHWC format, converts

^bb0(%arg0: tensor<1x6x6x16x!quant.uniform<u8:f32, 7.812500e-03:128>>):
  %0 = "tfl.dequantize"(%arg0) : (tensor<1x6x6x16x!quant.uniform<u8:f32, 7.812500e-03:128>>) -> tensor<1x6x6x16xf32>
  %1 = "tfl.softmax"(%0) {beta = 1.000000e+00 : f32} : (tensor<1x6x6x16xf32>) -> tensor<1x6x6x16xf32>
  func.return %1 : tensor<1x6x6x16xf32>

// CHECK: %[[sm:.*]] = "tfl.softmax"(%arg0)

// CHECK:           %[[VAL_2:.*]] = "tf.Conv2D"(%[[VAL_0]], %[[VAL_1]]) <{data_format = "NHWC", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "VALID", strides = [1, 1, 1, 1], use_cudnn_on_gpu = true}> : (tensor<1x8x8x207xf32>, tensor<3x3x207x16xf32>) -> tensor<1x6x6x16xf32>
// CHECK:           return %[[VAL_2]] : tensor<1x6x6x16xf32>
// CHECK:         }

// CHECK-NEXT:  return %[[reshape]] : tensor<*xf32>
}

func.func @avgPool2D(%arg0: tensor<1x6x6x16xf32>) -> tensor<1x1x1x16xf32> {
  // OK
  %0 = "tf.AvgPool"(%arg0) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", ksize = [1, 3, 6, 1], padding = "VALID", strides = [1, 3, 1, 1]} : (tensor<1x6x6x16xf32>) -> tensor<1x1x1x16xf32>
  // Unsupported ksize