auto y = GatherV2(scope_, x, {9, 8, 7, 6}, /*axis=*/3);
  TensorShape y_shape({2, 3, 7, 4});
  RunTest(x, shape, y, y_shape);
}

TEST_F(ArrayGradTest, GatherV2Grad_BatchDim) {
  TensorShape shape({2, 100, 3});
  auto x = Placeholder(scope_, DT_FLOAT, Placeholder::Shape(shape));
  GatherV2::Attrs attrs;
  attrs.batch_dims_ = 1;
  auto y =

// CHECK-SAME: value = dense<1> : tensor<1xi32>
// CHECK-DAG: %[[CST_1:.*]] = "tf.Const"()
// CHECK-SAME: value = dense<1> : tensor<1xi32>

// Check that the constants used for tf.GatherV2 and tf.AddV2 are different.
// CHECK: %[[GATHER_V2:.*]] = "tf.GatherV2"(%[[ARG_0]], %[[ARG_1]], %[[CST_1]])
// CHECK: %[[ADDV2:.*]] = "tf.AddV2"(%[[CST_0]], %[[CST_0]])

// -----

  std::set<string> constrained_kernel = {"Diag",
                                         "DiagPart",
                                         "GatherNd",
                                         "GatherV2",
                                         "QuantizeAndDequantizeV2",
                                         "QuantizeAndDequantizeV3",
                                         "QuantizeAndDequantizeV4",

  }
  func.func private @composite_gather_fn_1(%arg0: tensor<1024x3x4x3xf32>, %arg1: tensor<1xi32>, %arg2: tensor<i32>) -> tensor<1x3x4x3xf32> attributes {tf_quant.composite_function} {
    %0 = "tf.GatherV2"(%arg0, %arg1, %arg2) {attr_map = "0:batch_dims", batch_dims = 0 : i64, device = ""} : (tensor<1024x3x4x3xf32>, tensor<1xi32>, tensor<i32>) -> tensor<1x3x4x3xf32>
    return %0 : tensor<1x3x4x3xf32>
  }

  %cst = "tf.Const"() {device = "", value = dense<0> : tensor<i32>} : () -> tensor<i32>
  %cst_0 = "tf.Const"() {device = "", value = dense<1.0> : tensor<128x32xf32>} : () -> tensor<128x32xf32>
  %0 = "tf.GatherV2"(%cst_0, %arg0, %cst) {batch_dims = 0 : i64, device = ""} : (tensor<128x32xf32>, tensor<6xi64>, tensor<i32>) -> tensor<6x32xf32>
  return %0 : tensor<6x32xf32>

// CHECK-DAG: %[[CST:.*]] = "tf.Const"() {{.*}} : () -> tensor<i32>

                         %weight_scale : tensor<*xf32>, %weight_zp : tensor<*xi32>) -> tensor<*xf32>
      attributes {tf_quant.quantized_ops = ${quantized_ops}}
  {
    %accum_out = "tf.GatherV2"(%weight, %input, %axis) {
      batch_dims = 0 : i64, attr_map = "batch_dims:0"} : (tensor<*xi8>, tensor<*xi32>, tensor<i32>) -> tensor<*xi8>

    %accum_out_new = "tf.Cast"(%accum_out) : (tensor<*xi8>) -> tensor<*xf32>

    // 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>
    // expected-remark@+1 {{Cost: 40}}
    func.return %1 : tensor<1x3x5x20xf32>
}

  if (batch_dims < 0) {
    // TODO(bdodson): Figure out if we can find the param rank here, like the
    // python implementation does.
    return errors::InvalidArgument(
        "C++ GatherV2 gradient does not support negative batch_dims.");
  }

  // Handle axis by transposing the axis dimension to be the first non-batch
  // dimension, compute the gradient and transpose the result back.

                         %out_scale : tensor<*xf32>, %out_zp : tensor<*xi32>) -> tensor<*xi8>
      attributes {tf_quant.quantized_ops = ["Gather"]} {

    %out = "tf.GatherV2"(%weight, %input, %axis) {
      batch_dims = 0 : i64, attr_map = "batch_dims:0"} : (tensor<*xi8>, tensor<*xi32>, tensor<i32>) -> tensor<*xi8>

      (TF_AssignVariableOp $v_resource, $new_v, (CreateConstBoolAttrFalse))
    ]
  >;

// Pattern to decompose tf.ResourceGather into tf.ReadVariable and tf.GatherV2.
def DecomposeResourceGather : Pat<
  (TF_ResourceGatherOp:$old_result
    $resource, $indices, $batch_dims, $validate_indices),
  (TF_GatherV2Op:$dest
    (CreateTFReadVariableOp $old_result, $old_result, $resource),