This is an automated email from the ASF dual-hosted git repository.
lihaopeng pushed a commit to branch vectorized
in repository https://gitbox.apache.org/repos/asf/incubator-doris.git
commit f3ce1cba69aa27f8ae0af63451bab41d5048c367
Author: HappenLee <happen...@hotmail.com>
AuthorDate: Wed Jan 12 13:20:48 2022 +0800
[Vectorized][Enhancement] use simd to speed up coalesce and if_not_null
function (#7722)
Co-authored-by: lihaopeng <lihaop...@baidu.com>
---
be/src/vec/functions/function_coalesce.cpp | 210 ++++++++++++++++++++-----
be/src/vec/functions/is_not_null.cpp | 4 +-
be/src/vec/functions/simple_function_factory.h | 5 +-
be/test/vec/function/function_string_test.cpp | 42 +++++
4 files changed, 214 insertions(+), 47 deletions(-)
diff --git a/be/src/vec/functions/function_coalesce.cpp
b/be/src/vec/functions/function_coalesce.cpp
index 65d544c..99b6110 100644
--- a/be/src/vec/functions/function_coalesce.cpp
+++ b/be/src/vec/functions/function_coalesce.cpp
@@ -28,6 +28,8 @@ class FunctionCoalesce : public IFunction {
public:
static constexpr auto name = "coalesce";
+ mutable FunctionBasePtr func_is_not_null;
+
static FunctionPtr create() { return std::make_shared<FunctionCoalesce>();
}
String get_name() const override { return name; }
@@ -41,47 +43,70 @@ public:
size_t get_number_of_arguments() const override { return 0; }
DataTypePtr get_return_type_impl(const DataTypes& arguments) const
override {
+ DataTypePtr res;
for (const auto& arg : arguments) {
if (!arg->is_nullable()) {
- return arg;
+ res = arg;
+ break;
}
}
- return arguments[0];
+
+ res = res ? res : arguments[0];
+
+ const ColumnsWithTypeAndName is_not_null_col{
+ {nullptr, make_nullable(res), ""}
+ };
+ func_is_not_null = SimpleFunctionFactory::instance().
+ get_function("is_not_null_pred", is_not_null_col,
std::make_shared<DataTypeUInt8>());
+
+ return res;
}
Status execute_impl(FunctionContext* context, Block& block, const
ColumnNumbers& arguments,
size_t result, size_t input_rows_count) override {
DCHECK_GE(arguments.size(), 1);
+ DataTypePtr result_type = block.get_by_position(result).type;
ColumnNumbers filtered_args;
filtered_args.reserve(arguments.size());
- for (const auto& arg : arguments) {
- const auto& type = block.get_by_position(arg).type;
- if (type->only_null()) {
- continue;
- }
- filtered_args.push_back(arg);
- if (!type->is_nullable()) {
- break;
+
+ for (size_t i = 0; i < arguments.size(); ++i) {
+ const auto& arg_type = block.get_by_position(arguments[i]).type;
+ filtered_args.push_back(arguments[i]);
+ if (!arg_type->is_nullable()) {
+ if (i == 0) { //if the first column not null, return it's
directly
+ block.get_by_position(result).column =
block.get_by_position(arguments[0]).column;
+ return Status::OK();
+ } else {
+ break;
+ }
}
}
size_t remaining_rows = input_rows_count;
size_t argument_size = filtered_args.size();
- std::vector<int> record_idx(input_rows_count, -1); //used to save
column idx
+ std::vector<uint32_t> record_idx(input_rows_count, 0); //used to save
column idx, record the result data of each row from which column
+ std::vector<uint8_t> filled_flags(input_rows_count, 0); //used to save
filled flag, in order to check current row whether have filled data
+
MutableColumnPtr result_column;
+ if (!result_type->is_nullable()) {
+ result_column = result_type->create_column();
+ } else {
+ result_column = remove_nullable(result_type)->create_column();
+ }
- DataTypePtr type = block.get_by_position(result).type;
- if (!type->is_nullable()) {
- result_column = type->create_column();
+ // because now the string types does not support random position
writing,
+ // so insert into result data have two methods, one is for string
types, one is for others type remaining
+ bool is_string_result = result_column->is_column_string();
+ if (is_string_result) {
+ result_column->reserve(input_rows_count);
} else {
- result_column = remove_nullable(type)->create_column();
+ result_column->resize(input_rows_count);
}
- result_column->reserve(input_rows_count);
auto return_type = std::make_shared<DataTypeUInt8>();
- auto null_map = ColumnUInt8::create(input_rows_count, 1);
- auto& null_map_data = null_map->get_data();
- ColumnPtr argument_columns[argument_size];
+ auto null_map = ColumnUInt8::create(input_rows_count, 1); //if
null_map_data==1, the current row should be null
+ auto* __restrict null_map_data = null_map->get_data().data();
+ ColumnPtr argument_columns[argument_size]; //use to save nested_column
if is nullable column
for (size_t i = 0; i < argument_size; ++i) {
block.get_by_position(filtered_args[i]).column =
@@ -93,40 +118,69 @@ public:
}
}
+ Block temporary_block {
+ ColumnsWithTypeAndName {
+ block.get_by_position(filtered_args[0]),
+ {nullptr, std::make_shared<DataTypeUInt8>(), ""}
+ }
+ };
+
for (size_t i = 0; i < argument_size && remaining_rows; ++i) {
- const ColumnsWithTypeAndName is_not_null_col
{block.get_by_position(filtered_args[i])};
- Block temporary_block(is_not_null_col);
- temporary_block.insert(ColumnWithTypeAndName {nullptr,
return_type, ""});
- auto func_is_not_null =
SimpleFunctionFactory::instance().get_function("is_not_null_pred",
is_not_null_col, return_type);
+ temporary_block.get_by_position(0).column =
block.get_by_position(filtered_args[i]).column;
func_is_not_null->execute(context, temporary_block, {0}, 1,
input_rows_count);
- auto res_column =
std::move(*temporary_block.get_by_position(1).column->convert_to_full_column_if_const()).mutate();
+ auto res_column =
(*temporary_block.get_by_position(1).column->convert_to_full_column_if_const()).mutate();
auto& res_map =
assert_cast<ColumnVector<UInt8>*>(res_column.get())->get_data();
- auto* res = res_map.data();
-
- //TODO: if need to imporve performance in the feature, here it's
could SIMD
- for (size_t j = 0; j < input_rows_count && remaining_rows; ++j) {
- if (res[j] && null_map_data[j]) {
- null_map_data[j] = 0;
- remaining_rows--;
- record_idx[j] = i;
+ auto* __restrict res = res_map.data();
+
+ // Here it's SIMD thought the compiler automatically
+ // true: res[j]==1 && null_map_data[j]==1, false: others
+ // if true: remaining_rows--; record_idx[j]=column_idx;
null_map_data[j]=0, so the current row could fill result
+ for (size_t j = 0; j < input_rows_count; ++j) {
+ remaining_rows -= (res[j] & null_map_data[j]);
+ record_idx[j] += (res[j] & null_map_data[j]) * i;
+ null_map_data[j] -= (res[j] & null_map_data[j]);
+ }
+
+ if (remaining_rows == 0) {
+ //check whether all result data from the same column
+ size_t is_same_column_count = 0;
+ const auto data = record_idx[0];
+ for (size_t row = 0; row < input_rows_count; ++row) {
+ is_same_column_count += (record_idx[row] == data);
+ }
+
+ if (is_same_column_count == input_rows_count) {
+ if (result_type->is_nullable()) {
+ block.get_by_position(result).column =
make_nullable(argument_columns[i], false);
+ } else {
+ block.get_by_position(result).column =
argument_columns[i];
+ }
+ return Status::OK();
}
}
+
+ if (!is_string_result) {
+ //if not string type, could check one column firstly,
+ //and then fill the not null value in result column,
+ //this method may result in higher CPU cache
+ filled_result_column(result_type, result_column,
argument_columns[i], null_map_data,
+ filled_flags.data(), input_rows_count);
+ }
}
- //TODO: According to the record results, fill in result one by one,
- //that's fill in result use different methods for different types,
- //because now the string types does not support random position
writing,
- //But other type could, so could check one column, and fill the not
null value in result column,
- //and then next column, this method may result in higher CPU cache
- for (int row = 0; row < input_rows_count; ++row) {
- if (record_idx[row] == -1) {
- result_column->insert_default();
- } else {
-
result_column->insert_from(*argument_columns[record_idx[row]].get(), row);
+
+ if (is_string_result) {
+ //if string type, should according to the record results, fill in
result one by one,
+ for (size_t row = 0; row < input_rows_count; ++row) {
+ if (null_map_data[row]) { //should be null
+ result_column->insert_default();
+ } else {
+
result_column->insert_from(*argument_columns[record_idx[row]].get(), row);
+ }
}
}
- if (type->is_nullable()) {
+ if (result_type->is_nullable()) {
block.replace_by_position(result,
ColumnNullable::create(std::move(result_column), std::move(null_map)));
} else {
block.replace_by_position(result, std::move(result_column));
@@ -134,6 +188,76 @@ public:
return Status::OK();
}
+
+ template <typename ColumnType>
+ Status insert_result_data(MutableColumnPtr& result_column, ColumnPtr&
argument_column,
+ const UInt8* __restrict null_map_data, UInt8*
__restrict filled_flag,
+ const size_t input_rows_count) {
+ auto* __restrict result_raw_data =
+
reinterpret_cast<ColumnType*>(result_column.get())->get_data().data();
+ auto* __restrict column_raw_data =
+ reinterpret_cast<const
ColumnType*>(argument_column.get())->get_data().data();
+
+
+ // Here it's SIMD thought the compiler automatically also
+ // true: null_map_data[row]==0 && filled_idx[row]==0
+ // if true, could filled current row data into result column
+ for (size_t row = 0; row < input_rows_count; ++row) {
+ result_raw_data[row] += (!(null_map_data[row] | filled_flag[row]))
* column_raw_data[row];
+ filled_flag[row] += (!(null_map_data[row] | filled_flag[row]));
+ }
+ return Status::OK();
+ }
+
+ //TODO: this function is same as case when, should be replaced by macro
+ Status filled_result_column(const DataTypePtr& data_type,
MutableColumnPtr& result_column,
+ ColumnPtr& argument_column, UInt8* __restrict
null_map_data,
+ UInt8* __restrict filled_flag, const size_t
input_rows_count) {
+ WhichDataType which(data_type->is_nullable()
+ ? reinterpret_cast<const
DataTypeNullable*>(data_type.get())
+ ->get_nested_type()
+ : data_type);
+ if (which.is_uint8()) {
+ return insert_result_data<ColumnUInt8>(result_column,
argument_column, null_map_data,
+ filled_flag,
input_rows_count);
+ } else if (which.is_int16()) {
+ return insert_result_data<ColumnInt16>(result_column,
argument_column, null_map_data,
+ filled_flag,
input_rows_count);
+ } else if (which.is_uint32()) {
+ return insert_result_data<ColumnUInt32>(result_column,
argument_column, null_map_data,
+ filled_flag,
input_rows_count);
+ } else if (which.is_uint64()) {
+ return insert_result_data<ColumnUInt64>(result_column,
argument_column, null_map_data,
+ filled_flag,
input_rows_count);
+ } else if (which.is_int8()) {
+ return insert_result_data<ColumnInt8>(result_column,
argument_column, null_map_data,
+ filled_flag,
input_rows_count);
+ } else if (which.is_int16()) {
+ return insert_result_data<ColumnInt16>(result_column,
argument_column, null_map_data,
+ filled_flag,
input_rows_count);
+ } else if (which.is_int32()) {
+ return insert_result_data<ColumnInt32>(result_column,
argument_column, null_map_data,
+ filled_flag,
input_rows_count);
+ } else if (which.is_int64()) {
+ return insert_result_data<ColumnInt64>(result_column,
argument_column, null_map_data,
+ filled_flag,
input_rows_count);
+ } else if (which.is_date_or_datetime()) {
+ return insert_result_data<ColumnVector<DateTime>>(
+ result_column, argument_column, null_map_data,
filled_flag, input_rows_count);
+ } else if (which.is_float32()) {
+ return insert_result_data<ColumnFloat32>(result_column,
argument_column, null_map_data,
+ filled_flag,
input_rows_count);
+ } else if (which.is_float64()) {
+ return insert_result_data<ColumnFloat64>(result_column,
argument_column, null_map_data,
+ filled_flag,
input_rows_count);
+ } else if (which.is_decimal()) {
+ return insert_result_data<ColumnDecimal<Decimal128>>(
+ result_column, argument_column, null_map_data,
filled_flag, input_rows_count);
+ } else {
+ return Status::NotSupported(fmt::format("Unexpected type {} of
argument of function {}",
+ data_type->get_name(),
get_name()));
+ }
+ }
};
void register_function_coalesce(SimpleFunctionFactory& factory) {
diff --git a/be/src/vec/functions/is_not_null.cpp
b/be/src/vec/functions/is_not_null.cpp
index 1bfeaa6..346dac1 100644
--- a/be/src/vec/functions/is_not_null.cpp
+++ b/be/src/vec/functions/is_not_null.cpp
@@ -52,8 +52,8 @@ public:
if (auto* nullable =
check_and_get_column<ColumnNullable>(*elem.column)) {
/// Return the negated null map.
auto res_column = ColumnUInt8::create(input_rows_count);
- const auto& src_data = nullable->get_null_map_data();
- auto& res_data = assert_cast<ColumnUInt8&>(*res_column).get_data();
+ const auto* __restrict src_data =
nullable->get_null_map_data().data();
+ auto* __restrict res_data =
assert_cast<ColumnUInt8&>(*res_column).get_data().data();
for (size_t i = 0; i < input_rows_count; ++i) {
res_data[i] = !src_data[i];
diff --git a/be/src/vec/functions/simple_function_factory.h
b/be/src/vec/functions/simple_function_factory.h
index 1d064bb..d757920 100644
--- a/be/src/vec/functions/simple_function_factory.h
+++ b/be/src/vec/functions/simple_function_factory.h
@@ -65,10 +65,11 @@ void register_function_like(SimpleFunctionFactory& factory);
void register_function_regexp(SimpleFunctionFactory& factory);
void register_function_random(SimpleFunctionFactory& factory);
void register_function_coalesce(SimpleFunctionFactory& factory);
+
class SimpleFunctionFactory {
using Creator = std::function<FunctionBuilderPtr()>;
- using FunctionCreators = std::unordered_map<std::string, Creator>;
- using FunctionIsVariadic = std::unordered_set<std::string>;
+ using FunctionCreators = phmap::flat_hash_map<std::string, Creator>;
+ using FunctionIsVariadic = phmap::flat_hash_set<std::string>;
public:
void register_function(const std::string& name, Creator ptr) {
diff --git a/be/test/vec/function/function_string_test.cpp
b/be/test/vec/function/function_string_test.cpp
index bd2f9a7..9f1f5d7 100644
--- a/be/test/vec/function/function_string_test.cpp
+++ b/be/test/vec/function/function_string_test.cpp
@@ -681,6 +681,48 @@ TEST(function_string_test, function_unhex_test) {
vectorized::check_function<vectorized::DataTypeString, true>(func_name,
input_types, data_set);
}
+TEST(function_string_test, function_coalesce_test) {
+ std::string func_name = "coalesce";
+ {
+ std::vector<std::any> input_types = {vectorized::TypeIndex::Int32,
+ vectorized::TypeIndex::Int32,
+ vectorized::TypeIndex::Int32};
+ DataSet data_set = {{{Null(), Null(), (int32_t)1}, {(int32_t)1}},
+ {{Null(), Null(), (int32_t)2}, {(int32_t)2}},
+ {{Null(), Null(), (int32_t)3}, {(int32_t)3}},
+ {{Null(), Null(), (int32_t)4}, {(int32_t)4}}};
+ vectorized::check_function<vectorized::DataTypeInt32, true>(func_name,
input_types,
+ data_set);
+ }
+
+ {
+ std::vector<std::any> input_types = {vectorized::TypeIndex::String,
+ vectorized::TypeIndex::String,
+ vectorized::TypeIndex::Int32};
+ DataSet data_set = {
+ {{std::string("qwer"), Null(), (int32_t)1},
{std::string("qwer")}},
+ {{std::string("asdf"), Null(), (int32_t)2},
{std::string("asdf")}},
+ {{std::string("zxcv"), Null(), (int32_t)3},
{std::string("zxcv")}},
+ {{std::string("vbnm"), Null(), (int32_t)4},
{std::string("vbnm")}},
+ };
+ vectorized::check_function<vectorized::DataTypeString,
true>(func_name, input_types,
+ data_set);
+ }
+
+ {
+ std::vector<std::any> input_types = {vectorized::TypeIndex::String,
+ vectorized::TypeIndex::String,
+ vectorized::TypeIndex::String};
+ DataSet data_set = {
+ {{Null(), std::string("abc"), std::string("hij")},
{std::string("abc")}},
+ {{Null(), std::string("def"), std::string("klm")},
{std::string("def")}},
+ {{Null(), std::string(""), std::string("xyz")},
{std::string("")}},
+ {{Null(), Null(), std::string("uvw")}, {std::string("uvw")}}};
+ vectorized::check_function<vectorized::DataTypeString,
true>(func_name, input_types,
+ data_set);
+ }
+}
+
} // namespace doris
int main(int argc, char** argv) {
---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscr...@doris.apache.org
For additional commands, e-mail: commits-h...@doris.apache.org
