%3 = "tf.Multinomial"(%0, %1) {device = "/job:localhost/replica:0/task:0/device:CPU:0", seed = 1 : i64, seed2 = 1 : i64} : (tensor<*xf32>, tensor<*xi32>) -> tensor<*xi64>
  %4 = "tf.Multinomial"(%0, %1) {device = "/job:localhost/replica:0/task:0/device:CPU:0", seed = 0 : i64, seed2 = 0 : i64} : (tensor<*xf32>, tensor<*xi32>) -> tensor<*xi64>

        quantization_axis = -1 : i64, quantization_min_val = -128 : i64, quantization_max_val = 127 : i64
      } : (tensor<1xf32>, tensor<f32>, tensor<i32>) -> tensor<1x!tf_type.qint8>
      %1 = "tf.UniformDequantize"(%0, %scales, %zps) {
        quantization_axis = -1 : i64, quantization_min_val = -128 : i64, quantization_max_val = 127 : i64
      } : (tensor<1x!tf_type.qint8>, tensor<f32>, tensor<i32>) -> tensor<1xf32>

  // CHECK-NEXT: %[[RESULT_0:.*]] = "tf.Const"() <{value = dense<1> : tensor<i64>}> : () -> tensor<i64>
  // CHECK-NEXT: %[[RESULT_1:.*]] = "tf.AddV2"(%[[RESULT_0]], %[[RESULT_0]]) : (tensor<i64>, tensor<i64>) -> tensor<i64>
  %0 = "tf.Const"() {value = dense<1> : tensor<i64>} : () -> tensor<i64>
  %1 = "tf.AddV2"(%0, %0) : (tensor<i64>, tensor<i64>) -> tensor<i64>
  func.return

  "tf.TPUReplicateMetadata"() {num_cores_per_replica = 2 : i64, num_replicas = 2 : i64} : () -> ()
  %pi_0 = "tf.TPUPartitionedInput"(%arg0, %arg1) {_XlaSharding = "123", partition_dim = 0 : i64} : (tensor<!tf_type.resource<tensor<10x3xf32>>>, tensor<!tf_type.resource<tensor<10x3xf32>>>) -> tensor<!tf_type.resource<tensor<20x3xf32>>>

  "tf.InitializeTableFromTextFileV2"(%0, %cst) {delimiter = " ", device = "", key_index = -2 : i64, value_index = -1 : i64, vocab_size = -1 : i64} : (tensor<!tf_type.resource>, tensor<!tf_type.string>) -> ()
  func.return
  // CHECK-LABEL: func @init_all_tables

{_xla_inferred_shapes = [#tf_type.shape<*>, #tf_type.shape<*>], allowed_batch_sizes = [64, 128, 256], batch_timeout_micros = 5000 : i64, batching_queue = "", container = "", device = "/job:localhost/replica:0/task:0/device:CPU:0", enable_large_batch_splitting = true, f = @compute_0, max_batch_size = 256 : i64, max_enqueued_batches = 10000 : i64, num_batch_threads = 2 : i64, operandSegmentSizes = array<i32: 1, 1>, shared_name = "computation"} : (tensor<?x?xi32>, tensor<?x?xi32>) -> (tensor<*xi32>, tensor<*xi32>)...

        quantization_axis = -1 : i64, quantization_min_val = -128 : i64, quantization_max_val = 127 : i64
      } : (tensor<?xf32>, tensor<f32>, tensor<i32>) -> tensor<?x!tf_type.qint8>
      %1 = "tf.UniformDequantize"(%0, %scales, %zps) {
        quantization_axis = -1 : i64, quantization_min_val = -128 : i64, quantization_max_val = 127 : i64
      } : (tensor<?x!tf_type.qint8>, tensor<f32>, tensor<i32>) -> tensor<?xf32>

  "tf.InitializeTableFromTextFileV2"(%0, %cst) {delimiter = " ", device = "", key_index = -2 : i64, value_index = -1 : i64, vocab_size = -1 : i64} : (tensor<!tf_type.resource>, tensor<!tf_type.string>) -> ()
  func.return
  // CHECK-LABEL: func @init_all_tables

  // CHECK-SAME: allowed_batch_sizes = [6]
  // CHECK-SAME: batch_timeout_micros = 100000 : i64
  // CHECK-SAME: batching_queue = ""
  // CHECK-SAME: container = ""
  // CHECK-SAME: enable_large_batch_splitting = false
  // CHECK-SAME: max_batch_size = 6 : i64
  // CHECK-SAME: max_enqueued_batches = 10 : i64
  // CHECK-SAME: num_batch_threads = 1 : i64
  // CHECK-SAME: shared_name = "batch/"

	chki8(int8(i32), ci32&0xff-1<<8)
	chki8(int8(i64), ci64&0xff)
	chki8(int8(u8), cu8&0xff-1<<8)
	chki8(int8(u16), cu16&0xff)
	chki8(int8(u32), cu32&0xff)
	chki8(int8(u64), cu64&0xff)
	//	chki8(int8(f32), 0)
	//	chki8(int8(f64), 0)

	chki16(int16(i8), ci8&0xffff-1<<16)
	chki16(int16(i16), ci16&0xffff-1<<16)
	chki16(int16(i32), ci32&0xffff-1<<16)
	chki16(int16(i64), ci64&0xffff-1<<16)
	chki16(int16(u8), cu8&0xffff)