//
module {
  func.func @serving_default() -> tensor<1x3xbf16> {
    %cst = "tf.Const"() <{value = dense<[1]> : tensor<1xi32>}> : () -> tensor<1xi32>
    %0 = "tf.VarHandleOp"() <{container = "", shared_name = "Variable"}> : () -> tensor<!tf_type.resource<tensor<2x3xbf16>>>
    %1 = "tf.ResourceGather"(%0, %cst) <{batch_dims = 0 : i64, validate_indices = true}> : (tensor<!tf_type.resource<tensor<2x3xbf16>>>, tensor<1xi32>) -> tensor<1x3xbf16>

  %0 = "tf.BroadcastTo"(%arg0, %arg1) : (tensor<3xbf16>, tensor<2xi64>) -> tensor<3x3xbf16>
  func.return %0: tensor<3x3xbf16>

// CHECK-LABEL: broadcast_to_bf16
// CHECK:  [[CST:%.*]] = arith.constant dense<1.000000e+00> : tensor<3x3xbf16>
// CHECK:  [[MUL:%.*]] = tfl.mul(%arg0, [[CST]]) <{fused_activation_function = "NONE"}> : (tensor<3xbf16>, tensor<3x3xbf16>) -> tensor<3x3xbf16>

// CHECK:  return %[[V1]] : tensor<32x1x256xbf16>
}

// CHECK-LABEL: Relu_bf16
func.func @Relu_bf16(%arg0: tensor<2x3xbf16>) -> tensor<2x3xbf16> {
  %cst = arith.constant dense<0.0> : tensor<2x3xbf16>
  %0 = "tfl.maximum"(%arg0, %cst) : (tensor<2x3xbf16>, tensor<2x3xbf16>) -> tensor<2x3xbf16>
  func.return %0 : tensor<2x3xbf16>

  // CHECK: %[[RESULT:.*]] = "tfl.relu"(%arg0)
  // CHECK: return %[[RESULT]]
}

  // CHECK-DAG:  %[[CST_1:.*]] = "tfl.pseudo_const"() <{value = dense<0.000000e+00> : tensor<4x4xf16>}> : () -> tensor<4x4xf16>
  // CHECK-DAG:  %[[DQ_0:.*]] = "tfl.dequantize"(%[[CST_0]]) : (tensor<4x4xf16>) -> tensor<4x4xf32>
  // CHECK-DAG:  %[[DQ_1:.*]] = "tfl.dequantize"(%[[CST_1]]) : (tensor<4x4xf16>) -> tensor<4x4xf32>

    %1 = "tf.Cast"(%0) <{truncate = false}> : (tensor<3x1xf32>) -> tensor<3x1xbf16>
    %2 = "tf.VarHandleOp"() <{container = "", shared_name = "y"}> : () -> tensor<!tf_type.resource<tensor<3x1xbf16>>>
    "tf.AssignVariableOp"(%2, %1) : (tensor<!tf_type.resource<tensor<3x1xbf16>>>, tensor<3x1xbf16>) -> ()
    return
  }

  // CHECK: tfrt_fallback_async.const_dense_tensor  dense<1.000000e+00> : tensor<1xbf16>
  %1 = "tf.Const"() {device = "/device:CPU:0", value = dense<[1.0]> : tensor<1xbf16>} : () -> tensor<4xbf16>
  // CHECK: corert.executeop({{.*}}) "tf.Const"() {dtype = ui64, value = dense<[1, 2, 3, 4]> : tensor<4xui64>} : 1
  %2 = "tf.Const"() {device = "/device:GPU:0", value = dense<[1, 2, 3, 4]> : tensor<4xui64>} : () -> tensor<4xui64>