%w = arith.constant dense<127.0> : tensor<512x12xf32>
  %b = arith.constant dense<0.0> : tensor<512xf32>
  %fc = "tfl.fully_connected"(%0, %w, %b) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<1x224x224x3xf32>, tensor<512x12xf32>, tensor<512xf32>) -> tensor<1x112x112x512xf32>

func.func @mean_4d_no_keepdim(%arg0: tensor<1x48x48x512xf32>) -> tensor<1x512xf32> {
  %cst = arith.constant dense<[1, 2]> : tensor<2xi32>
  %0 = "tfl.mean"(%arg0, %cst) {keep_dims = false} : (tensor<1x48x48x512xf32>, tensor<2xi32>) -> tensor<1x512xf32>
  func.return %0 : tensor<1x512xf32>
}

// CHECK:       func @mean_4d_no_keepdim([[VAL_0:%.*]]: tensor<1x48x48x512xf32>) -> tensor<1x512xf32> {

  %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>

// CHECK:           %[[VAL_11:.*]] = "tfl.reshape"(%[[VAL_9]], %[[VAL_6]]) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<384x512xf32>, tensor<4xi32>) -> tensor<1x1x384x512xf32>
// CHECK:           %[[VAL_12:.*]] = "tfl.reshape"(%[[VAL_10]], %[[VAL_7]]) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<128x512xf32>, tensor<4xi32>) -> tensor<128x1x1x512xf32>

  %w = arith.constant dense<127.0> : tensor<512x12xf32>
  %b = arith.constant dense<0.0> : tensor<512xf32>
  %fc = "tfl.fully_connected"(%arg0, %w, %b) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<1x224x224x3xf32>, tensor<512x12xf32>, tensor<512xf32>) -> tensor<1x112x112x512xf32>
  func.return %fc : tensor<1x112x112x512xf32>

  %cst_add = arith.constant dense<2.0> : tensor<2x512xf32>  // Not 1D
  %cst_weights = arith.constant dense<3.0> : tensor<1024x512xf32>
  %cst_bias = arith.constant dense<5.0> : tensor<1024xf32>

  %0 = "tfl.add"(%arg, %cst_add) {fused_activation_function = "NONE"} : (tensor<2x512xf32>, tensor<2x512xf32>) -> tensor<2x512xf32>

  %cst_input = arith.constant dense<1.0> : tensor<2x512xf32>
  %cst_weights = arith.constant dense<2.0> : tensor<1024x512xf32>
  %cst_bias = arith.constant dense<4.0> : tensor<1024xf32>

  %0 = "tfl.fully_connected" (%cst_input, %cst_weights, %cst_bias) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<2x512xf32>, tensor<1024x512xf32>, tensor<1024xf32>) -> tensor<2x1024xf32>

  // CHECK-DAG: %[[SEL:.*]] = mhlo.select %[[COMP]], %[[RS]], %[[ZERO_MAT]] : tensor<12x12xi1>, tensor<12x12xf32>
  // CHECK-DAG: %[[RED:.*]] = mhlo.reduce(%[[SEL]] init: %[[ZERO]]) applies mhlo.add across dimensions = [0] : (tensor<12x12xf32>, tensor<f32>) -> tensor<12xf32>