func.func @lift_float_matmul(%arg0: tensor<1x12x12x512xf32>) -> (tensor<*xf32>, tensor<*xf32>) {
  %cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<512x512xf32>} : () -> tensor<512x512xf32>
  %out_1 = "tf.MatMul"(%arg0, %cst) {
    device = "", transpose_a = false, transpose_b = false
  } : (tensor<1x12x12x512xf32>, tensor<512x512xf32>) -> tensor<*xf32>
  %out_2 = "tf.MatMul"(%arg0, %arg0) {
    device = "", transpose_a = false, transpose_b = true

func.func @lift_float_matmul(%arg0: tensor<1x12x12x512xf32>) -> (tensor<*xf32>, tensor<*xf32>) {
  %cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<512x512xf32>} : () -> tensor<512x512xf32>
  %out_1 = "tf.MatMul"(%arg0, %cst) {
    device = "", transpose_a = false, transpose_b = false
  } : (tensor<1x12x12x512xf32>, tensor<512x512xf32>) -> tensor<*xf32>
  %out_2 = "tf.MatMul"(%arg0, %arg0) {
    device = "", transpose_a = false, transpose_b = true

    %cst = "tf.Const"() {device = "", value = dense<1.000000e+01> : tensor<512x512xf32>} : () -> tensor<512x512xf32>
    %cst_sharded = "tf.XlaSharding"(%cst) {_XlaSharding = "\08\03\1A\03\01\04\02\22\08\00\04\01\05\02\06\03\070\01", device = "", sharding = "\08\03\1A\03\01\04\02\22\08\00\04\01\05\02\06\03\070\01", unspecified_dims = []} : (tensor<512x512xf32>) -> tensor<512x512xf32>

    // expected-remark@+1 {{Cost: 2}}
    %1 = "tf.ReadVariableOp"(%0) {device = "/job:localhost/replica:0/task:0/device:CPU:0"} : (tensor<!tf_type.resource<tensor<512x512xf32>>>) -> tensor<512x512xf32>
    // 262657 = 1 + 512 + 512 * 512
    // expected-remark@+1 {{Cost: 262657}}

// CHECK-LABEL: testDotToDotGeneralMatrixMatrix
func.func @testDotToDotGeneralMatrixMatrix(%arg0: tensor<2x3072xf32>, %arg1: tensor<3072x512xf32>) -> tensor<2x512xf32> {
  %0 = "mhlo.dot"(%arg0, %arg1) : (tensor<2x3072xf32>, tensor<3072x512xf32>) -> tensor<2x512xf32>
  func.return %0 : tensor<2x512xf32>

// CHECK:      %[[RES:.*]] = "mhlo.dot_general"(%arg0, %arg1) <{
// CHECK-SAME:   dot_dimension_numbers = #mhlo.dot<

  %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>