func.return %6 : tensor<1x401408xf32>
}

// CHECK-LABEL: quantized_constant
func.func @quantized_constant(%arg0: tensor<1x2xf32>) -> tensor<2x2xf32> {
  %1 = "tfl.quantize"(%arg0) {qtype = tensor<1x2x!quant.uniform<u8:f32, 1.0>>, volatile} : (tensor<1x2xf32>) -> tensor<1x2x!quant.uniform<u8:f32, 1.0>>

// CHECK-DAG: %[[input_19:.*]] = "quantfork.stats"({{.*}}) <{layerStats = dense<[-2.000000e+00, 4.000000e+00]> : tensor<2xf32>}> : (tensor<1x2xf32>) -> tensor<1x2xf32>

// CHECK: "tfl.unidirectional_sequence_lstm"({{.*}}, %[[input_18]], %[[input_19]], %{{[0-9]+}}, %{{[0-9]+}}, %{{[0-9]+}}, %{{[0-9]+}})

// RUN:   -split-input-file | FileCheck %s --check-prefix=STATIC-RANGE-PTQ-TO-COMPUTE-HEAVY

// STATIC-RANGE-PTQ-TO-COMPUTE-HEAVY: @main
func.func @main(%arg0: tensor<1x2xf32>) -> tensor<1x2xf32> {
  %0 = stablehlo.add %arg0, %arg0 : tensor<1x2xf32>
  return %0 : tensor<1x2xf32>
}
// Tests that `composite_add_fn_1` does not quantize when quantizing
// only compute-heavy ops.

}

// CHECK-LABEL: QuantizeConcat
func.func @QuantizeConcat(tensor<1x2xf32>, tensor<1x2xf32>) -> tensor<2x2x!quant.uniform<u8:f32, 1.000000e-01:128>> {
^bb0(%arg0: tensor<1x2xf32>, %arg1: tensor<1x2xf32>):
  %0 = "tfl.concatenation"(%arg0, %arg1) {axis = 0 : i32, fused_activation_function = "NONE"} : (tensor<1x2xf32>, tensor<1x2xf32>) -> tensor<2x2xf32>

"/job:tpu_host_worker/replica:0/task:3/device:CPU:0"}, %arg15: tensor<*x!tf_type.resource<tensor<6x32xf32>>> {tf._user_specified_name = "927", tf.device = "/job:tpu_host_worker/replica:0/task:0/device:CPU:0"}, %arg16: tensor<*x!tf_type.resource<tensor<6x32xf32>>> {tf._user_specified_name = "929", tf.device = "/job:tpu_host_worker/replica:0/task:1/device:CPU:0"}, %arg17: tensor<*x!tf_type.resource<tensor<6x32xf32>>> {tf._user_specified_name = "931", tf.device = "/job:tpu_host_worker/replica:0/task:2/device:CPU:0"},...

}

// CHECK-LABEL: test_gather
func.func @test_gather(%arg0 : tensor<1x2x20xf32>, %arg1 : tensor<3x5xi32>) -> (tensor<1x3x5x20xf32>){
    // expected-remark@+1 {{Cost: 1}}
    %0 = "tf.Const"() { value = dense<[1]> : tensor<1xi32> } : () -> tensor<1xi32>
    // expected-remark@+1 {{Cost: 300}}
    %1 = "tf.GatherV2"(%arg0, %arg1, %0) : (tensor<1x2x20xf32>, tensor<3x5xi32>, tensor<1xi32>) -> tensor<1x3x5x20xf32>

  %0 = "tf_device.cluster_func"(%arg0, %arg1) {func = @_func, use_spmd_for_xla_partitioning = true, use_tpu = true, num_cores_per_replica = 2 : i64} : (tensor<2x4xf32>, tensor<4x2xf32>) -> tensor<2x2xf32>
  %1:2 = "tf.TPUPartitionedOutputV2"(%0) {device = "", partition_dims = [2, 1]} : (tensor<2x2xf32>) -> (tensor<1x2xf32>, tensor<1x2xf32>)

  // CHECK:  %cst = arith.constant dense<1> : tensor<1xi32>
  // CHECK:  %4 = "tfl.reduce_max"(%arg0, %cst) <{keep_dims = false}> : (tensor<1x32x1xf32>, tensor<1xi32>) -> tensor<1x1xf32>
  // CHECK:  %5 = "tfl.arg_max"(%arg0, %cst) : (tensor<1x32x1xf32>, tensor<1xi32>) -> tensor<1x1xi32>
  // CHECK:  return %4, %5 : tensor<1x1xf32>, tensor<1x1xi32>
}

// CHECK-LABEL:   func @convert_pytorch_argmax

// Tests for ops with canonicalization patterns.

// CHECK-LABEL: get_real_shape
func.func @get_real_shape(%arg0: tensor<1x2xf32>) -> tensor<2xindex> {
  %0 = "tfr.cast"(%arg0) : (tensor<1x2xf32>) -> !tfr.tensor
  %1 = tfr.get_shape %0 -> !shape.shape
  %2 = shape.to_extent_tensor %1 : !shape.shape -> tensor<2xindex>
  func.return %2 : tensor<2xindex>

// CHECK:           %[[VAL_11:.*]] = "tf.ArgMax"(%[[VAL_0]], %[[VAL_9]]) : (tensor<1x32x1xf32>, tensor<1xi32>) -> tensor<1x1xi32>
// CHECK:           return %[[VAL_10]], %[[VAL_11]] : tensor<1x1xf32>, tensor<1x1xi32>
// CHECK:         }
func.func @convert_argmax_with_reshaped_iota(%arg0: tensor<1x32x1xf32>) -> (tensor<1x1xf32>, tensor<1x1xi32>) {