diff --git a/cpp/src/join/join_common_utils.hpp b/cpp/src/join/join_common_utils.hpp index 86402a0e7de..573101cefd9 100644 --- a/cpp/src/join/join_common_utils.hpp +++ b/cpp/src/join/join_common_utils.hpp @@ -22,7 +22,6 @@ #include #include -#include #include #include @@ -51,11 +50,6 @@ using mixed_multimap_type = cudf::detail::cuco_allocator, cuco::legacy::double_hashing<1, hash_type, hash_type>>; -using semi_map_type = cuco::legacy::static_map>; - using row_hash_legacy = cudf::row_hasher; diff --git a/cpp/src/join/mixed_join_common_utils.cuh b/cpp/src/join/mixed_join_common_utils.cuh index 19701816867..4a52cfe098a 100644 --- a/cpp/src/join/mixed_join_common_utils.cuh +++ b/cpp/src/join/mixed_join_common_utils.cuh @@ -25,6 +25,7 @@ #include #include +#include namespace cudf { namespace detail { @@ -160,6 +161,39 @@ struct pair_expression_equality : public expression_equality { } }; +/** + * @brief Equality comparator that composes two row_equality comparators. + */ +struct double_row_equality_comparator { + row_equality const equality_comparator; + row_equality const conditional_comparator; + + __device__ bool operator()(size_type lhs_row_index, size_type rhs_row_index) const noexcept + { + using experimental::row::lhs_index_type; + using experimental::row::rhs_index_type; + + return equality_comparator(lhs_index_type{lhs_row_index}, rhs_index_type{rhs_row_index}) && + conditional_comparator(lhs_index_type{lhs_row_index}, rhs_index_type{rhs_row_index}); + } +}; + +// A CUDA Cooperative Group of 1 thread for the hash set for mixed semi. +auto constexpr DEFAULT_MIXED_SEMI_JOIN_CG_SIZE = 1; + +// The hash set type used by mixed_semi_join with the build_table. +using hash_set_type = + cuco::static_set, + cuda::thread_scope_device, + double_row_equality_comparator, + cuco::linear_probing, + cudf::detail::cuco_allocator, + cuco::storage<1>>; + +// The hash_set_ref_type used by mixed_semi_join kerenels for probing. +using hash_set_ref_type = hash_set_type::ref_type; + } // namespace detail } // namespace cudf diff --git a/cpp/src/join/mixed_join_kernels_semi.cu b/cpp/src/join/mixed_join_kernels_semi.cu index 7459ac3e99c..bd8c80652a0 100644 --- a/cpp/src/join/mixed_join_kernels_semi.cu +++ b/cpp/src/join/mixed_join_kernels_semi.cu @@ -38,38 +38,48 @@ CUDF_KERNEL void __launch_bounds__(block_size) table_device_view right_table, table_device_view probe, table_device_view build, - row_hash const hash_probe, row_equality const equality_probe, - cudf::detail::semi_map_type::device_view hash_table_view, + hash_set_ref_type set_ref, cudf::device_span left_table_keep_mask, cudf::ast::detail::expression_device_view device_expression_data) { + auto constexpr cg_size = hash_set_ref_type::cg_size; + + auto const tile = cg::tiled_partition(cg::this_thread_block()); + // Normally the casting of a shared memory array is used to create multiple // arrays of different types from the shared memory buffer, but here it is // used to circumvent conflicts between arrays of different types between // different template instantiations due to the extern specifier. extern __shared__ char raw_intermediate_storage[]; - cudf::ast::detail::IntermediateDataType* intermediate_storage = + auto intermediate_storage = reinterpret_cast*>(raw_intermediate_storage); auto thread_intermediate_storage = - &intermediate_storage[threadIdx.x * device_expression_data.num_intermediates]; + intermediate_storage + (tile.meta_group_rank() * device_expression_data.num_intermediates); - cudf::size_type const left_num_rows = left_table.num_rows(); - cudf::size_type const right_num_rows = right_table.num_rows(); - auto const outer_num_rows = left_num_rows; + // Equality evaluator to use + auto const evaluator = cudf::ast::detail::expression_evaluator( + left_table, right_table, device_expression_data); - cudf::size_type outer_row_index = threadIdx.x + blockIdx.x * block_size; + // Make sure to swap_tables here as hash_set will use probe table as the left one + auto constexpr swap_tables = true; + auto const equality = single_expression_equality{ + evaluator, thread_intermediate_storage, swap_tables, equality_probe}; - auto evaluator = cudf::ast::detail::expression_evaluator( - left_table, right_table, device_expression_data); + // Create set ref with the new equality comparator + auto const set_ref_equality = set_ref.with_key_eq(equality); - if (outer_row_index < outer_num_rows) { - // Figure out the number of elements for this key. - auto equality = single_expression_equality{ - evaluator, thread_intermediate_storage, false, equality_probe}; + // Total number of rows to query the set + auto const outer_num_rows = left_table.num_rows(); + // Grid stride for the tile + auto const cg_grid_stride = cudf::detail::grid_1d::grid_stride() / cg_size; - left_table_keep_mask[outer_row_index] = - hash_table_view.contains(outer_row_index, hash_probe, equality); + // Find all the rows in the left table that are in the hash table + for (auto outer_row_index = cudf::detail::grid_1d::global_thread_id() / cg_size; + outer_row_index < outer_num_rows; + outer_row_index += cg_grid_stride) { + auto const result = set_ref_equality.contains(tile, outer_row_index); + if (tile.thread_rank() == 0) { left_table_keep_mask[outer_row_index] = result; } } } @@ -78,9 +88,8 @@ void launch_mixed_join_semi(bool has_nulls, table_device_view right_table, table_device_view probe, table_device_view build, - row_hash const hash_probe, row_equality const equality_probe, - cudf::detail::semi_map_type::device_view hash_table_view, + hash_set_ref_type set_ref, cudf::device_span left_table_keep_mask, cudf::ast::detail::expression_device_view device_expression_data, detail::grid_1d const config, @@ -94,9 +103,8 @@ void launch_mixed_join_semi(bool has_nulls, right_table, probe, build, - hash_probe, equality_probe, - hash_table_view, + set_ref, left_table_keep_mask, device_expression_data); } else { @@ -106,9 +114,8 @@ void launch_mixed_join_semi(bool has_nulls, right_table, probe, build, - hash_probe, equality_probe, - hash_table_view, + set_ref, left_table_keep_mask, device_expression_data); } diff --git a/cpp/src/join/mixed_join_kernels_semi.cuh b/cpp/src/join/mixed_join_kernels_semi.cuh index 43714ffb36a..b08298e64e4 100644 --- a/cpp/src/join/mixed_join_kernels_semi.cuh +++ b/cpp/src/join/mixed_join_kernels_semi.cuh @@ -45,9 +45,8 @@ namespace detail { * @param[in] right_table The right table * @param[in] probe The table with which to probe the hash table for matches. * @param[in] build The table with which the hash table was built. - * @param[in] hash_probe The hasher used for the probe table. * @param[in] equality_probe The equality comparator used when probing the hash table. - * @param[in] hash_table_view The hash table built from `build`. + * @param[in] set_ref The hash table device view built from `build`. * @param[out] left_table_keep_mask The result of the join operation with "true" element indicating * the corresponding index from left table is present in output * @param[in] device_expression_data Container of device data required to evaluate the desired @@ -58,9 +57,8 @@ void launch_mixed_join_semi(bool has_nulls, table_device_view right_table, table_device_view probe, table_device_view build, - row_hash const hash_probe, row_equality const equality_probe, - cudf::detail::semi_map_type::device_view hash_table_view, + hash_set_ref_type set_ref, cudf::device_span left_table_keep_mask, cudf::ast::detail::expression_device_view device_expression_data, detail::grid_1d const config, diff --git a/cpp/src/join/mixed_join_semi.cu b/cpp/src/join/mixed_join_semi.cu index aa4fa281159..83a55eca50f 100644 --- a/cpp/src/join/mixed_join_semi.cu +++ b/cpp/src/join/mixed_join_semi.cu @@ -45,45 +45,6 @@ namespace cudf { namespace detail { -namespace { -/** - * @brief Device functor to create a pair of hash value and index for a given row. - */ -struct make_pair_function_semi { - __device__ __forceinline__ cudf::detail::pair_type operator()(size_type i) const noexcept - { - // The value is irrelevant since we only ever use the hash map to check for - // membership of a particular row index. - return cuco::make_pair(static_cast(i), 0); - } -}; - -/** - * @brief Equality comparator that composes two row_equality comparators. - */ -class double_row_equality { - public: - double_row_equality(row_equality equality_comparator, row_equality conditional_comparator) - : _equality_comparator{equality_comparator}, _conditional_comparator{conditional_comparator} - { - } - - __device__ bool operator()(size_type lhs_row_index, size_type rhs_row_index) const noexcept - { - using experimental::row::lhs_index_type; - using experimental::row::rhs_index_type; - - return _equality_comparator(lhs_index_type{lhs_row_index}, rhs_index_type{rhs_row_index}) && - _conditional_comparator(lhs_index_type{lhs_row_index}, rhs_index_type{rhs_row_index}); - } - - private: - row_equality _equality_comparator; - row_equality _conditional_comparator; -}; - -} // namespace - std::unique_ptr> mixed_join_semi( table_view const& left_equality, table_view const& right_equality, @@ -95,7 +56,7 @@ std::unique_ptr> mixed_join_semi( rmm::cuda_stream_view stream, rmm::device_async_resource_ref mr) { - CUDF_EXPECTS((join_type != join_kind::INNER_JOIN) && (join_type != join_kind::LEFT_JOIN) && + CUDF_EXPECTS((join_type != join_kind::INNER_JOIN) and (join_type != join_kind::LEFT_JOIN) and (join_type != join_kind::FULL_JOIN), "Inner, left, and full joins should use mixed_join."); @@ -136,7 +97,7 @@ std::unique_ptr> mixed_join_semi( // output column and follow the null-supporting expression evaluation code // path. auto const has_nulls = cudf::nullate::DYNAMIC{ - cudf::has_nulls(left_equality) || cudf::has_nulls(right_equality) || + cudf::has_nulls(left_equality) or cudf::has_nulls(right_equality) or binary_predicate.may_evaluate_null(left_conditional, right_conditional, stream)}; auto const parser = ast::detail::expression_parser{ @@ -155,27 +116,20 @@ std::unique_ptr> mixed_join_semi( auto right_conditional_view = table_device_view::create(right_conditional, stream); auto const preprocessed_build = - experimental::row::equality::preprocessed_table::create(build, stream); + cudf::experimental::row::equality::preprocessed_table::create(build, stream); auto const preprocessed_probe = - experimental::row::equality::preprocessed_table::create(probe, stream); + cudf::experimental::row::equality::preprocessed_table::create(probe, stream); auto const row_comparator = - cudf::experimental::row::equality::two_table_comparator{preprocessed_probe, preprocessed_build}; + cudf::experimental::row::equality::two_table_comparator{preprocessed_build, preprocessed_probe}; auto const equality_probe = row_comparator.equal_to(has_nulls, compare_nulls); - semi_map_type hash_table{ - compute_hash_table_size(build.num_rows()), - cuco::empty_key{std::numeric_limits::max()}, - cuco::empty_value{cudf::detail::JoinNoneValue}, - cudf::detail::cuco_allocator{rmm::mr::polymorphic_allocator{}, stream}, - stream.value()}; - // Create hash table containing all keys found in right table // TODO: To add support for nested columns we will need to flatten in many // places. However, this probably isn't worth adding any time soon since we // won't be able to support AST conditions for those types anyway. auto const build_nulls = cudf::nullate::DYNAMIC{cudf::has_nulls(build)}; auto const row_hash_build = cudf::experimental::row::hash::row_hasher{preprocessed_build}; - auto const hash_build = row_hash_build.device_hasher(build_nulls); + // Since we may see multiple rows that are identical in the equality tables // but differ in the conditional tables, the equality comparator used for // insertion must account for both sets of tables. An alternative solution @@ -190,20 +144,28 @@ std::unique_ptr> mixed_join_semi( auto const equality_build_equality = row_comparator_build.equal_to(build_nulls, compare_nulls); auto const preprocessed_build_condtional = - experimental::row::equality::preprocessed_table::create(right_conditional, stream); + cudf::experimental::row::equality::preprocessed_table::create(right_conditional, stream); auto const row_comparator_conditional_build = cudf::experimental::row::equality::two_table_comparator{preprocessed_build_condtional, preprocessed_build_condtional}; auto const equality_build_conditional = row_comparator_conditional_build.equal_to(build_nulls, compare_nulls); - double_row_equality equality_build{equality_build_equality, equality_build_conditional}; - make_pair_function_semi pair_func_build{}; - auto iter = cudf::detail::make_counting_transform_iterator(0, pair_func_build); + hash_set_type row_set{ + {compute_hash_table_size(build.num_rows())}, + cuco::empty_key{JoinNoneValue}, + {equality_build_equality, equality_build_conditional}, + {row_hash_build.device_hasher(build_nulls)}, + {}, + {}, + cudf::detail::cuco_allocator{rmm::mr::polymorphic_allocator{}, stream}, + {stream.value()}}; + + auto iter = thrust::make_counting_iterator(0); // skip rows that are null here. if ((compare_nulls == null_equality::EQUAL) or (not nullable(build))) { - hash_table.insert(iter, iter + right_num_rows, hash_build, equality_build, stream.value()); + row_set.insert_async(iter, iter + right_num_rows, stream.value()); } else { thrust::counting_iterator stencil(0); auto const [row_bitmask, _] = @@ -211,18 +173,19 @@ std::unique_ptr> mixed_join_semi( row_is_valid pred{static_cast(row_bitmask.data())}; // insert valid rows - hash_table.insert_if( - iter, iter + right_num_rows, stencil, pred, hash_build, equality_build, stream.value()); + row_set.insert_if_async(iter, iter + right_num_rows, stencil, pred, stream.value()); } - auto hash_table_view = hash_table.get_device_view(); - - detail::grid_1d const config(outer_num_rows, DEFAULT_JOIN_BLOCK_SIZE); - auto const shmem_size_per_block = parser.shmem_per_thread * config.num_threads_per_block; + detail::grid_1d const config(outer_num_rows * hash_set_type::cg_size, DEFAULT_JOIN_BLOCK_SIZE); + auto const shmem_size_per_block = + parser.shmem_per_thread * + cuco::detail::int_div_ceil(config.num_threads_per_block, hash_set_type::cg_size); auto const row_hash = cudf::experimental::row::hash::row_hasher{preprocessed_probe}; auto const hash_probe = row_hash.device_hasher(has_nulls); + hash_set_ref_type const row_set_ref = row_set.ref(cuco::contains).with_hash_function(hash_probe); + // Vector used to indicate indices from left/probe table which are present in output auto left_table_keep_mask = rmm::device_uvector(probe.num_rows(), stream); @@ -231,9 +194,8 @@ std::unique_ptr> mixed_join_semi( *right_conditional_view, *probe_view, *build_view, - hash_probe, equality_probe, - hash_table_view, + row_set_ref, cudf::device_span(left_table_keep_mask), parser.device_expression_data, config, diff --git a/cpp/tests/join/mixed_join_tests.cu b/cpp/tests/join/mixed_join_tests.cu index 6c147c8a128..9041969bec7 100644 --- a/cpp/tests/join/mixed_join_tests.cu +++ b/cpp/tests/join/mixed_join_tests.cu @@ -778,6 +778,138 @@ TYPED_TEST(MixedLeftSemiJoinTest, BasicEquality) {1}); } +TYPED_TEST(MixedLeftSemiJoinTest, MixedLeftSemiJoinGatherMap) +{ + auto const col_ref_left_1 = cudf::ast::column_reference(0, cudf::ast::table_reference::LEFT); + auto const col_ref_right_1 = cudf::ast::column_reference(0, cudf::ast::table_reference::RIGHT); + auto left_one_greater_right_one = + cudf::ast::operation(cudf::ast::ast_operator::GREATER, col_ref_left_1, col_ref_right_1); + + this->test({{2, 3, 9, 0, 1, 7, 4, 6, 5, 8}, {1, 2, 3, 4, 5, 6, 7, 8, 9, 0}}, + {{6, 5, 9, 8, 10, 32}, {0, 1, 2, 3, 4, 5}, {7, 8, 9, 0, 1, 2}}, + {0}, + {1}, + left_one_greater_right_one, + {2, 7, 8}); +} + +TYPED_TEST(MixedLeftSemiJoinTest, MixedLeftSemiJoinGatherMapLarge) +{ + using T1 = double; + + // Number of rows in each column + auto constexpr N = 10000; + + // Generate column data for left and right tables + auto const [left_col0, right_col0] = gen_random_nullable_repeated_columns(N, 200); + auto const [left_col1, right_col1] = gen_random_nullable_repeated_columns(N, 100); + + // Setup data and nulls for the left table + std::vector, std::vector>> lefts = { + {left_col0.first, left_col0.second}, {left_col1.first, left_col1.second}}; + std::vector> left_wrappers; + std::vector left_columns; + for (auto [data, valids] : lefts) { + left_wrappers.emplace_back( + cudf::test::fixed_width_column_wrapper(data.begin(), data.end(), valids.begin())); + left_columns.emplace_back(left_wrappers.back()); + }; + + // Setup data and nulls for the right table + std::vector, std::vector>> rights = { + {right_col0.first, right_col0.second}, {right_col1.first, right_col1.second}}; + std::vector> right_wrappers; + std::vector right_columns; + for (auto [data, valids] : rights) { + right_wrappers.emplace_back( + cudf::test::fixed_width_column_wrapper(data.begin(), data.end(), valids.begin())); + right_columns.emplace_back(left_wrappers.back()); + }; + + // Left and right table views. + auto const left_table = cudf::table_view{left_columns}; + auto const right_table = cudf::table_view{right_columns}; + + // Using the zeroth column for equality. + auto const left_equality = left_table.select({0}); + auto const right_equality = right_table.select({0}); + + // Column references for equality column. + auto const col_ref_left_0 = cudf::ast::column_reference(0, cudf::ast::table_reference::LEFT); + auto const col_ref_right_0 = cudf::ast::column_reference(0, cudf::ast::table_reference::RIGHT); + auto left_zero_eq_right_zero = + cudf::ast::operation(cudf::ast::ast_operator::EQUAL, col_ref_left_0, col_ref_right_0); + + // Mixed semi join with zeroth column equality + { + // Expected left_semi_join result + auto const expected_mixed_semi_join = + cudf::conditional_left_semi_join(left_table, right_table, left_zero_eq_right_zero); + + // Actual mixed_left_semi_join result + auto const mixed_semi_join = cudf::mixed_left_semi_join(left_equality, + right_equality, + left_table, + right_table, + left_zero_eq_right_zero, + cudf::null_equality::UNEQUAL); + + // Copy data back to host for comparisons + auto expected_indices = cudf::detail::make_std_vector_async( + cudf::device_span(*expected_mixed_semi_join), cudf::get_default_stream()); + auto result_indices = cudf::detail::make_std_vector_sync( + cudf::device_span(*mixed_semi_join), cudf::get_default_stream()); + + // Sort the indices for 1-1 comparison + std::sort(expected_indices.begin(), expected_indices.end()); + std::sort(result_indices.begin(), result_indices.end()); + + // Expected and actual vectors must match. + EXPECT_EQ(expected_mixed_semi_join->size(), mixed_semi_join->size()); + EXPECT_TRUE( + std::equal(expected_indices.begin(), expected_indices.end(), result_indices.begin())); + } + + // Mixed semi join with zeroth column equality and first column GREATER conditional + { + // Column references for conditional column. + auto const col_ref_left_1 = cudf::ast::column_reference(1, cudf::ast::table_reference::LEFT); + auto const col_ref_right_1 = cudf::ast::column_reference(1, cudf::ast::table_reference::RIGHT); + auto left_one_gt_right_one = + cudf::ast::operation(cudf::ast::ast_operator::GREATER, col_ref_left_1, col_ref_right_1); + + // Expected left_semi_join result + auto const expected_mixed_semi_join = cudf::conditional_left_semi_join( + left_table, + right_table, + cudf::ast::operation( + cudf::ast::ast_operator::LOGICAL_AND, left_zero_eq_right_zero, left_one_gt_right_one)); + + // Actual left_semi_join result + auto const mixed_semi_join = cudf::mixed_left_semi_join(left_equality, + right_equality, + left_table, + right_table, + left_one_gt_right_one, + cudf::null_equality::UNEQUAL); + + // Copy data back to host for comparisons + auto expected_indices = cudf::detail::make_std_vector_async( + cudf::device_span(*expected_mixed_semi_join), cudf::get_default_stream()); + auto result_indices = cudf::detail::make_std_vector_sync( + cudf::device_span(*mixed_semi_join), cudf::get_default_stream()); + + // Sort the indices for 1-1 comparison + std::sort(expected_indices.begin(), expected_indices.end()); + std::sort(result_indices.begin(), result_indices.end()); + + // Expected and actual vectors must match. + EXPECT_EQ(expected_mixed_semi_join->size(), mixed_semi_join->size()); + EXPECT_TRUE( + std::equal(expected_indices.begin(), expected_indices.end(), result_indices.begin())); + } +} + TYPED_TEST(MixedLeftSemiJoinTest, BasicEqualityDuplicates) { this->test({{0, 1, 2, 1}, {3, 4, 5, 6}, {10, 20, 30, 40}}, @@ -900,3 +1032,18 @@ TYPED_TEST(MixedLeftAntiJoinTest, AsymmetricLeftLargerEquality) left_zero_eq_right_zero, {0, 1, 3}); } + +TYPED_TEST(MixedLeftAntiJoinTest, MixedLeftAntiJoinGatherMap) +{ + auto const col_ref_left_1 = cudf::ast::column_reference(0, cudf::ast::table_reference::LEFT); + auto const col_ref_right_1 = cudf::ast::column_reference(0, cudf::ast::table_reference::RIGHT); + auto left_one_greater_right_one = + cudf::ast::operation(cudf::ast::ast_operator::GREATER, col_ref_left_1, col_ref_right_1); + + this->test({{2, 3, 9, 0, 1, 7, 4, 6, 5, 8}, {1, 2, 3, 4, 5, 6, 7, 8, 9, 0}}, + {{6, 5, 9, 8, 10, 32}, {0, 1, 2, 3, 4, 5}, {7, 8, 9, 0, 1, 2}}, + {0}, + {1}, + left_one_greater_right_one, + {0, 1, 3, 4, 5, 6, 9}); +} diff --git a/python/cudf/cudf/_lib/aggregation.pyx b/python/cudf/cudf/_lib/aggregation.pyx index 7c91533cf93..3c96b90f0a1 100644 --- a/python/cudf/cudf/_lib/aggregation.pyx +++ b/python/cudf/cudf/_lib/aggregation.pyx @@ -78,8 +78,11 @@ class Aggregation: ) @classmethod - def nunique(cls): - return cls(pylibcudf.aggregation.nunique(pylibcudf.types.NullPolicy.EXCLUDE)) + def nunique(cls, dropna=True): + return cls(pylibcudf.aggregation.nunique( + pylibcudf.types.NullPolicy.EXCLUDE + if dropna else pylibcudf.types.NullPolicy.INCLUDE + )) @classmethod def nth(cls, size): diff --git a/python/cudf/cudf/core/groupby/groupby.py b/python/cudf/cudf/core/groupby/groupby.py index cb8cd0cd28b..be05075a2cd 100644 --- a/python/cudf/cudf/core/groupby/groupby.py +++ b/python/cudf/cudf/core/groupby/groupby.py @@ -2232,6 +2232,22 @@ def func(x): return self.agg(func) + @_performance_tracking + def nunique(self, dropna: bool = True): + """ + Return number of unique elements in the group. + + Parameters + ---------- + dropna : bool, default True + Don't include NaN in the counts. + """ + + def func(x): + return getattr(x, "nunique")(dropna=dropna) + + return self.agg(func) + @_performance_tracking def std( self, diff --git a/python/cudf/cudf/tests/test_groupby.py b/python/cudf/cudf/tests/test_groupby.py index 848bc259e7b..14ba9894fd3 100644 --- a/python/cudf/cudf/tests/test_groupby.py +++ b/python/cudf/cudf/tests/test_groupby.py @@ -1940,6 +1940,23 @@ def test_groupby_nunique(agg, by): assert_groupby_results_equal(expect, got, check_dtype=False) +@pytest.mark.parametrize("dropna", [True, False]) +def test_nunique_dropna(dropna): + gdf = cudf.DataFrame( + { + "a": [1, 1, 2], + "b": [4, None, 5], + "c": [None, None, 7], + "d": [1, 1, 3], + } + ) + pdf = gdf.to_pandas() + + result = gdf.groupby("a")["b"].nunique(dropna=dropna) + expected = pdf.groupby("a")["b"].nunique(dropna=dropna) + assert_groupby_results_equal(result, expected, check_dtype=False) + + @pytest.mark.parametrize( "n", [0, 1, 2, 10],