thevar1able/ClickHouse
1
0
Fork 0
mirror of https://github.com/ClickHouse/ClickHouse.git synced 2024-09-20 00:30:49 +00:00
Code Issues Packages Projects Releases Wiki Activity

Revert "ColumnVector: optimize UInt8 index with AVX512VBMI (#41247)"

Browse Source 
This reverts commit 8de524cb73.
This commit is contained in:
Alexey Milovidov 2022-09-26 22:36:12 +02:00
parent 32eca72bff
commit 5407291191
1 changed files with 2 additions and 134 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

136
src/Columns/ColumnVector.h
View File

@ -7,15 +7,11 @@
#include <base/unaligned.h>
#include <Core/Field.h>
#include <Common/assert_cast.h>
#include <Common/TargetSpecific.h>
#include <Core/TypeId.h>
#include <base/TypeName.h>
#include "config_core.h"
#if USE_MULTITARGET_CODE
# include <immintrin.h>
#endif
namespace DB
{
@ -395,124 +391,6 @@ protected:
Container data;
};
DECLARE_DEFAULT_CODE(
template <typename Container, typename Type>
inline void vectorIndexImpl(const Container & data, const PaddedPODArray<Type> & indexes, size_t limit, Container & res_data)
{
for (size_t i = 0; i < limit; ++i)
res_data[i] = data[indexes[i]];
}
);
DECLARE_AVX512VBMI_SPECIFIC_CODE(
template <typename Container, typename Type>
inline void vectorIndexImpl(const Container & data, const PaddedPODArray<Type> & indexes, size_t limit, Container & res_data)
{
static constexpr UInt64 MASK64 = 0xffffffffffffffff;
const size_t limit64 = limit & ~63;
size_t pos = 0;
size_t data_size = data.size();
auto data_pos = reinterpret_cast<const UInt8 *>(data.data());
auto indexes_pos = reinterpret_cast<const UInt8 *>(indexes.data());
auto res_pos = reinterpret_cast<UInt8 *>(res_data.data());
if (data_size <= 64)
{
/// one single mask load for table size <= 64
__mmask64 last_mask = MASK64 >> (64 - data_size);
__m512i table1 = _mm512_maskz_loadu_epi8(last_mask, data_pos);
/// 64 bytes table lookup using one single permutexvar_epi8
while (pos < limit64)
{
__m512i vidx = _mm512_loadu_epi8(indexes_pos + pos);
__m512i out = _mm512_permutexvar_epi8(vidx, table1);
_mm512_storeu_epi8(res_pos + pos, out);
pos += 64;
}
/// tail handling
if (limit > limit64)
{
__mmask64 tail_mask = MASK64 >> (limit64 + 64 - limit);
__m512i vidx = _mm512_maskz_loadu_epi8(tail_mask, indexes_pos + pos);
__m512i out = _mm512_permutexvar_epi8(vidx, table1);
_mm512_mask_storeu_epi8(res_pos + pos, tail_mask, out);
}
}
else if (data_size <= 128)
{
/// table size (64, 128] requires 2 zmm load
__mmask64 last_mask = MASK64 >> (128 - data_size);
__m512i table1 = _mm512_loadu_epi8(data_pos);
__m512i table2 = _mm512_maskz_loadu_epi8(last_mask, data_pos + 64);
/// 128 bytes table lookup using one single permute2xvar_epi8
while (pos < limit64)
{
__m512i vidx = _mm512_loadu_epi8(indexes_pos + pos);
__m512i out = _mm512_permutex2var_epi8(table1, vidx, table2);
_mm512_storeu_epi8(res_pos + pos, out);
pos += 64;
}
if (limit > limit64)
{
__mmask64 tail_mask = MASK64 >> (limit64 + 64 - limit);
__m512i vidx = _mm512_maskz_loadu_epi8(tail_mask, indexes_pos + pos);
__m512i out = _mm512_permutex2var_epi8(table1, vidx, table2);
_mm512_mask_storeu_epi8(res_pos + pos, tail_mask, out);
}
}
else
{
if (data_size > 256)
{
/// byte index will not exceed 256 boundary.
data_size = 256;
}
__m512i table1 = _mm512_loadu_epi8(data_pos);
__m512i table2 = _mm512_loadu_epi8(data_pos + 64);
__m512i table3, table4;
if (data_size <= 192)
{
/// only 3 tables need to load if size <= 192
__mmask64 last_mask = MASK64 >> (192 - data_size);
table3 = _mm512_maskz_loadu_epi8(last_mask, data_pos + 128);
table4 = _mm512_setzero_si512();
}
else
{
__mmask64 last_mask = MASK64 >> (256 - data_size);
table3 = _mm512_loadu_epi8(data_pos + 128);
table4 = _mm512_maskz_loadu_epi8(last_mask, data_pos + 192);
}
/// 256 bytes table lookup can use: 2 permute2xvar_epi8 plus 1 blender with MSB
while (pos < limit64)
{
__m512i vidx = _mm512_loadu_epi8(indexes_pos + pos);
__m512i tmp1 = _mm512_permutex2var_epi8(table1, vidx, table2);
__m512i tmp2 = _mm512_permutex2var_epi8(table3, vidx, table4);
__mmask64 msb = _mm512_movepi8_mask(vidx);
__m512i out = _mm512_mask_blend_epi8(msb, tmp1, tmp2);
_mm512_storeu_epi8(res_pos + pos, out);
pos += 64;
}
if (limit > limit64)
{
__mmask64 tail_mask = MASK64 >> (limit64 + 64 - limit);
__m512i vidx = _mm512_maskz_loadu_epi8(tail_mask, indexes_pos + pos);
__m512i tmp1 = _mm512_permutex2var_epi8(table1, vidx, table2);
__m512i tmp2 = _mm512_permutex2var_epi8(table3, vidx, table4);
__mmask64 msb = _mm512_movepi8_mask(vidx);
__m512i out = _mm512_mask_blend_epi8(msb, tmp1, tmp2);
_mm512_mask_storeu_epi8(res_pos + pos, tail_mask, out);
}
}
}
);
template <typename T>
template <typename Type>
ColumnPtr ColumnVector<T>::indexImpl(const PaddedPODArray<Type> & indexes, size_t limit) const
@ -521,18 +399,8 @@ ColumnPtr ColumnVector<T>::indexImpl(const PaddedPODArray<Type> & indexes, size_
auto res = this->create(limit);
typename Self::Container & res_data = res->getData();
#if USE_MULTITARGET_CODE
if constexpr (sizeof(T) == 1 && sizeof(Type) == 1)
{
/// VBMI optimization only applicable for (U)Int8 types
if (isArchSupported(TargetArch::AVX512VBMI))
{
TargetSpecific::AVX512VBMI::vectorIndexImpl<Container, Type>(data, indexes, limit, res_data);
return res;
}
}
#endif
TargetSpecific::Default::vectorIndexImpl<Container, Type>(data, indexes, limit, res_data);
for (size_t i = 0; i < limit; ++i)
res_data[i] = data[indexes[i]];
return res;
}