This is an automated email from the ASF dual-hosted git repository.
Mryange pushed a commit to branch master
in repository https://gitbox.apache.org/repos/asf/doris.git
The following commit(s) were added to refs/heads/master by this push:
new 7038bb70bf2 [opt](expr) Vectorize AVG serialized state merge (#65729)
7038bb70bf2 is described below
commit 7038bb70bf25d2b5d23600fc0b0dd7f02f63abd4
Author: Mryange <[email protected]>
AuthorDate: Thu Jul 23 17:37:59 2026 +0800
[opt](expr) Vectorize AVG serialized state merge (#65729)
### What problem does this PR solve?
Clang preserves the original floating-point accumulation order when
merging serialized AVG states,
preventing vectorization of this contiguous hot loop. This change adds a
reusable reassociation hint
and applies it to AVG range merge. Existing raw Clang pragmas in AVG,
SUM, and AVG_WEIGHTED are also
replaced with the shared macro.
The focused benchmark reduced CPU time by 35.6% for 4K states and 38.5%
for 64K states.
---
be/src/common/compiler_util.h | 10 +++++++++-
be/src/exprs/aggregate/aggregate_function_avg.h | 6 +++---
be/src/exprs/aggregate/aggregate_function_avg_weighted.h | 9 +++------
be/src/exprs/aggregate/aggregate_function_sum.h | 5 ++---
4 files changed, 17 insertions(+), 13 deletions(-)
diff --git a/be/src/common/compiler_util.h b/be/src/common/compiler_util.h
index 559fca903e6..1d2a25b7742 100644
--- a/be/src/common/compiler_util.h
+++ b/be/src/common/compiler_util.h
@@ -56,4 +56,12 @@
#define NO_SANITIZE_UNDEFINED __attribute__((__no_sanitize__("undefined")))
#else
#define NO_SANITIZE_UNDEFINED
-#endif
\ No newline at end of file
+#endif
+
+// Allow reassociation only when the evaluation order is not part of the
operation's contract.
+// This may change the least significant bits of floating-point results.
+#ifdef __clang__
+#define ALLOW_FP_REASSOCIATION _Pragma("clang fp reassociate(on)")
+#else
+#define ALLOW_FP_REASSOCIATION
+#endif
diff --git a/be/src/exprs/aggregate/aggregate_function_avg.h
b/be/src/exprs/aggregate/aggregate_function_avg.h
index 876bf1ede7b..795dce34ec3 100644
--- a/be/src/exprs/aggregate/aggregate_function_avg.h
+++ b/be/src/exprs/aggregate/aggregate_function_avg.h
@@ -29,6 +29,7 @@
#include <type_traits>
#include <vector>
+#include "common/compiler_util.h"
#include "core/assert_cast.h"
#include "core/column/column.h"
#include "core/column/column_fixed_length_object.h"
@@ -167,9 +168,7 @@ public:
template <bool is_add>
NO_SANITIZE_UNDEFINED void update_value(AggregateDataPtr __restrict place,
const IColumn** columns, ssize_t
row_num) const {
-#ifdef __clang__
-#pragma clang fp reassociate(on)
-#endif
+ ALLOW_FP_REASSOCIATION
const auto& column =
assert_cast<const ColVecType&,
TypeCheckOnRelease::DISABLE>(*columns[0]);
if constexpr (is_add) {
@@ -262,6 +261,7 @@ public:
NO_SANITIZE_UNDEFINED void deserialize_and_merge_from_column_range(
AggregateDataPtr __restrict place, const IColumn& column, size_t
begin, size_t end,
Arena&) const override {
+ ALLOW_FP_REASSOCIATION
DCHECK(end <= column.size() && begin <= end)
<< ", begin:" << begin << ", end:" << end << ",
column.size():" << column.size();
auto& col = assert_cast<const ColumnFixedLengthObject&>(column);
diff --git a/be/src/exprs/aggregate/aggregate_function_avg_weighted.h
b/be/src/exprs/aggregate/aggregate_function_avg_weighted.h
index 3984b4de41e..7385281d5b6 100644
--- a/be/src/exprs/aggregate/aggregate_function_avg_weighted.h
+++ b/be/src/exprs/aggregate/aggregate_function_avg_weighted.h
@@ -26,6 +26,7 @@
#include <memory>
#include <type_traits>
+#include "common/compiler_util.h"
#include "common/status.h"
#include "core/assert_cast.h"
#include "core/binary_cast.hpp"
@@ -47,9 +48,7 @@ template <PrimitiveType T>
struct AggregateFunctionAvgWeightedData {
using DataType = typename PrimitiveTypeTraits<T>::CppType;
void add(const DataType& data_val, double weight_val) {
-#ifdef __clang__
-#pragma clang fp reassociate(on)
-#endif
+ ALLOW_FP_REASSOCIATION
data_sum = data_sum + (double(data_val) * weight_val);
weight_sum = weight_sum + weight_val;
}
@@ -65,9 +64,7 @@ struct AggregateFunctionAvgWeightedData {
}
void merge(const AggregateFunctionAvgWeightedData& rhs) {
-#ifdef __clang__
-#pragma clang fp reassociate(on)
-#endif
+ ALLOW_FP_REASSOCIATION
data_sum = data_sum + rhs.data_sum;
weight_sum = weight_sum + rhs.weight_sum;
}
diff --git a/be/src/exprs/aggregate/aggregate_function_sum.h
b/be/src/exprs/aggregate/aggregate_function_sum.h
index ca7c1aaf557..8eddb893ed9 100644
--- a/be/src/exprs/aggregate/aggregate_function_sum.h
+++ b/be/src/exprs/aggregate/aggregate_function_sum.h
@@ -25,6 +25,7 @@
#include <memory>
#include <vector>
+#include "common/compiler_util.h"
#include "core/assert_cast.h"
#include "core/column/column.h"
#include "core/data_type/data_type.h"
@@ -51,9 +52,7 @@ struct AggregateFunctionSumData {
typename PrimitiveTypeTraits<T>::CppType sum {};
NO_SANITIZE_UNDEFINED void add(typename PrimitiveTypeTraits<T>::CppType
value) {
-#ifdef __clang__
-#pragma clang fp reassociate(on)
-#endif
+ ALLOW_FP_REASSOCIATION
sum += value;
}
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
