thevar1able/ClickHouse
1
0
Fork 0
mirror of https://github.com/ClickHouse/ClickHouse.git synced 2024-12-04 05:22:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

Disable vectorization for uint64 --> float32 cast

Browse Source
This commit is contained in:
Robert Schulze 2022-07-22 10:59:56 +00:00
parent cad0e7a62c
commit 6a631426b7
No known key found for this signature in database
GPG Key ID: 26703B55FB13728A
1 changed files with 21 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

19
src/Functions/FunctionsConversion.h
View File

@ -324,11 +324,30 @@ struct ConvertImpl
} }
else else
{ {
if constexpr (std::is_same_v<FromDataType, DataTypeUInt64> && std::is_same_v<ToDataType, DataTypeFloat32>)
{
/// Turns out that when ClickHouse is compiled with AVX1 or AVX2 instructions, Clang's autovectorizer produces
/// code for UInt64-to-Float23 conversion which is only ~50% as fast as scalar code. Interestingly, scalar code
/// is equally fast than code compiled for SSE4.2, so we might as well disable vectorization. This situation
/// may change with AVX512 which has a dediated instruction for that usecase (_mm512_cvtepi64_ps).
#if defined(__x86_64__)
# ifdef __clang__
# pragma clang loop vectorize(disable) interleave(disable)
# endif
#endif
for (size_t i = 0; i < input_rows_count; ++i) for (size_t i = 0; i < input_rows_count; ++i)
{ {
vec_to[i] = static_cast<ToFieldType>(vec_from[i]); vec_to[i] = static_cast<ToFieldType>(vec_from[i]);
} }
} }
else
{
for (size_t i = 0; i < input_rows_count; ++i)
{
vec_to[i] = static_cast<ToFieldType>(vec_from[i]);
}
}
}
} }
} }