database/doris
2
0
Fork 0
Code Issues Pull Requests Actions 3 Packages Projects Releases Wiki Activity

[Optimiaze] Optimize HyperLogLog (#6625)

Browse Source 
1. Replace std::max with a ternary expression, std::max is much heavier than the ternary operator
2. Replace std::set with arrays, std::set is based on red-black trees, traversal will follow the chain domain, and cache hits are not good
3. Optimize the serialize function, improve the calculation speed of num_non_zero_registers by reducing branches, and the serialization of _registers after optimization is faster
4. The test found that the performance improvement is more obvious
This commit is contained in:
thinker
2021-10-10 23:04:39 +08:00
committed by GitHub
parent 4232f787ad
commit 0941322dd6
3 changed files with 183 additions and 82 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
be/src/exprs/aggregate_functions.cpp
View File

@ -1112,7 +1112,7 @@ void AggregateFunctions::hll_update(FunctionContext* ctx, const T& src, StringVa
if (hash_value != 0) {
int idx = hash_value % dst->len;
uint8_t first_one_bit = __builtin_ctzl(hash_value >> HLL_COLUMN_PRECISION) + 1;
dst->ptr[idx] = std::max(dst->ptr[idx], first_one_bit);
dst->ptr[idx] = (dst->ptr[idx] < first_one_bit ? first_one_bit : dst->ptr[idx]);
}
}
@ -1122,8 +1122,10 @@ void AggregateFunctions::hll_merge(FunctionContext* ctx, const StringVal& src, S
DCHECK_EQ(dst->len, std::pow(2, HLL_COLUMN_PRECISION));
DCHECK_EQ(src.len, std::pow(2, HLL_COLUMN_PRECISION));
auto dp = dst->ptr;
auto sp = src.ptr;
for (int i = 0; i < src.len; ++i) {
dst->ptr[i] = std::max(dst->ptr[i], src.ptr[i]);
dp[i] = (dp[i] < sp[i] ? sp[i] : dp[i]);
}
}