database/mariadb-columnstore-engine
1
0
Fork 0
mirror of https://github.com/mariadb-corporation/mariadb-columnstore-engine.git synced 2025-07-30 19:23:07 +03:00
Code Activity

MCOL-4809 This patch introduces vectorized scanning/filtering for short CHAR/VARCHAR columns

Browse Source 
Short CHAR/VARCHAR column values contain integer-encoded strings.
    After certain manipulations(orderSwap(strnxfrm(str))) the values
    become integers that preserve original strings order relation
    according to a certain translation rules(collation). Prepared
    values are ready to be SIMD-processed.
This commit is contained in:
Roman Nozdrin
2022-02-17 22:29:14 +00:00
parent f7e4cefda3
commit 7cdc914b4e
4 changed files with 602 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files

140
primitives/linux-port/column.cpp
View File

@ -1,5 +1,5 @@
/* Copyright (C) 2014 InfiniDB, Inc.
Copyright (C) 2016-2021 MariaDB Corporation
Copyright (C) 2016-2022 MariaDB Corporation
This program is free software; you can redistribute it and/or
modify it under the terms of the GNU General Public License
@ -22,6 +22,7 @@
#include <cassert>
#include <cmath>
#include <functional>
#include <type_traits>
#ifndef _MSC_VER
#include <pthread.h>
#else
@ -62,6 +63,55 @@ inline uint64_t order_swap(uint64_t x)
return ret;
}
// Dummy template
template<typename T,
typename std::enable_if<sizeof(T) >= sizeof(uint128_t), T>::type* = nullptr>
inline T orderSwap(T x)
{
return x;
}
template<typename T,
typename std::enable_if<sizeof(T) == sizeof(int64_t), T>::type* = nullptr>
inline T orderSwap(T x)
{
T ret = (x >> 56) |
((x << 40) & 0x00FF000000000000ULL) |
((x << 24) & 0x0000FF0000000000ULL) |
((x << 8) & 0x000000FF00000000ULL) |
((x >> 8) & 0x00000000FF000000ULL) |
((x >> 24) & 0x0000000000FF0000ULL) |
((x >> 40) & 0x000000000000FF00ULL) |
(x << 56);
return ret;
}
template<typename T,
typename std::enable_if<sizeof(T) == sizeof(int32_t), T>::type* = nullptr>
inline T orderSwap(T x)
{
T ret = (x >> 24) |
((x << 8) & 0x00FF0000U) |
((x >> 8) & 0x0000FF00U) |
(x << 24);
return ret;
}
template<typename T,
typename std::enable_if<sizeof(T) == sizeof(int16_t), T>::type* = nullptr>
inline T orderSwap(T x)
{
T ret = (x >> 8) | (x <<8);
return ret;
}
template<typename T,
typename std::enable_if<sizeof(T) == sizeof(uint8_t), T>::type* = nullptr>
inline T orderSwap(T x)
{
return x;
}
template <class T>
inline int compareBlock(const void* a, const void* b)
{
@ -107,8 +157,11 @@ inline bool colCompare_(const T& val1, const T& val2, uint8_t COP)
}
}
inline bool colCompareStr(const ColRequestHeaderDataType& type, uint8_t COP, const utils::ConstString& val1,
const utils::ConstString& val2)
inline bool colCompareStr(const ColRequestHeaderDataType &type,
uint8_t COP,
const utils::ConstString &val1,
const utils::ConstString &val2,
const bool printOut = false)
{
int error = 0;
bool rc = primitives::StringComparator(type).op(&error, COP, val1, val2);
@ -1179,7 +1232,7 @@ void scalarFiltering(
#if defined(__x86_64__)
template <typename VT, typename SIMD_WRAPPER_TYPE, bool HAS_INPUT_RIDS, typename T,
typename std::enable_if<HAS_INPUT_RIDS == false, T>::type* = nullptr>
inline SIMD_WRAPPER_TYPE simdDataLoadTemplate(VT& processor, const T* srcArray, const T* origSrcArray,
inline SIMD_WRAPPER_TYPE simdDataLoad(VT& processor, const T* srcArray, const T* origSrcArray,
const primitives::RIDType* ridArray, const uint16_t iter)
{
return {processor.loadFrom(reinterpret_cast<const char*>(srcArray))};
@ -1189,7 +1242,7 @@ inline SIMD_WRAPPER_TYPE simdDataLoadTemplate(VT& processor, const T* srcArray,
// TODO Move the logic into simd namespace class methods and use intrinsics
template <typename VT, typename SIMD_WRAPPER_TYPE, bool HAS_INPUT_RIDS, typename T,
typename std::enable_if<HAS_INPUT_RIDS == true, T>::type* = nullptr>
inline SIMD_WRAPPER_TYPE simdDataLoadTemplate(VT& processor, const T* srcArray, const T* origSrcArray,
inline SIMD_WRAPPER_TYPE simdDataLoad(VT& processor, const T* srcArray, const T* origSrcArray,
const primitives::RIDType* ridArray, const uint16_t iter)
{
constexpr const uint16_t WIDTH = sizeof(T);
@ -1205,6 +1258,32 @@ inline SIMD_WRAPPER_TYPE simdDataLoadTemplate(VT& processor, const T* srcArray,
return {result};
}
template <ENUM_KIND KIND, typename VT,typename SIMD_WRAPPER_TYPE, typename T,
typename std::enable_if<KIND != KIND_TEXT, T>::type* = nullptr>
inline SIMD_WRAPPER_TYPE simdSwapedOrderDataLoad(const ColRequestHeaderDataType &type, VT& processor, typename VT::SimdType& dataVector)
{
return {dataVector};
}
template <ENUM_KIND KIND, typename VT,typename SIMD_WRAPPER_TYPE, typename T,
typename std::enable_if<KIND == KIND_TEXT, T>::type* = nullptr>
inline SIMD_WRAPPER_TYPE simdSwapedOrderDataLoad(const ColRequestHeaderDataType &type,
VT& processor, typename VT::SimdType& dataVector)
{
constexpr const uint16_t WIDTH = sizeof(T);
constexpr const uint16_t VECTOR_SIZE = VT::vecByteSize / WIDTH;
using SimdType = typename VT::SimdType;
SimdType result;
T* resultTypedPtr = reinterpret_cast<T*>(&result);
T* srcTypedPtr = reinterpret_cast<T*>(&dataVector);
for (uint32_t i = 0; i < VECTOR_SIZE; ++i)
{
utils::ConstString s{reinterpret_cast<const char*>(&srcTypedPtr[i]), WIDTH};
resultTypedPtr[i] = orderSwap(type.strnxfrm<T>(s.rtrimZero()));
}
return {result};
}
// This routine filters input block in a vectorized manner.
// It supports all output types, all input types.
// It doesn't support KIND==TEXT so upper layers filters this KIND out beforehand.
@ -1214,8 +1293,8 @@ inline SIMD_WRAPPER_TYPE simdDataLoadTemplate(VT& processor, const T* srcArray,
// to glue the masks produced by actual filters.
// Then it takes a vector of data, run filters and logical function using pointers.
// See the corresponding dispatcher to get more details on vector processing class.
template <typename T, typename VT, bool HAS_INPUT_RIDS, int OUTPUT_TYPE, ENUM_KIND KIND, typename FT,
typename ST>
template<typename T, typename VT, bool HAS_INPUT_RIDS, int OUTPUT_TYPE,
ENUM_KIND KIND, typename FT, typename ST>
void vectorizedFiltering(NewColRequestHeader* in, ColResultHeader* out, const T* srcArray,
const uint32_t srcSize, primitives::RIDType* ridArray, const uint16_t ridSize,
ParsedColumnFilter* parsedColumnFilter, const bool validMinMax, const T emptyValue,
@ -1225,8 +1304,12 @@ void vectorizedFiltering(NewColRequestHeader* in, ColResultHeader* out, const T*
using SimdType = typename VT::SimdType;
using SimdWrapperType = typename VT::SimdWrapperType;
using FilterType = typename VT::FilterType;
using UT = typename std::conditional<std::is_unsigned<FilterType>::value || datatypes::is_uint128_t<FilterType>::value || std::is_same<double, FilterType>::value,
FilterType, typename datatypes::make_unsigned<FilterType>::type>::type;
VT simdProcessor;
SimdType dataVec;
[[maybe_unused]] SimdType swapedOrderDataVec;
[[maybe_unused]] auto typeHolder = in->colType;
SimdType emptyFilterArgVec = simdProcessor.emptyNullLoadValue(emptyValue);
SimdType nullFilterArgVec = simdProcessor.emptyNullLoadValue(nullValue);
MT writeMask, nonEmptyMask, nonNullMask, nonNullOrEmptyMask;
@ -1292,11 +1375,27 @@ void vectorizedFiltering(NewColRequestHeader* in, ColResultHeader* out, const T*
for (uint32_t j = 0; j < filterCount; ++j)
{
// Preload filter argument values only once.
filterArgsVectors.push_back(simdProcessor.loadValue(*((FilterType*)&filterValues[j])));
if constexpr (KIND == KIND_TEXT)
{
// Preload filter argument values only once.
// First cast filter value as the corresponding unsigned int value
UT filterValue = *((UT*)&filterValues[j]);
// Cast to ConstString to preprocess the string
utils::ConstString s{reinterpret_cast<const char*>(&filterValue), sizeof(UT)};
// Strip all 0 bytes on the right, convert byte into collation weights array
// and swap bytes order.
UT bigEndianFilterWeights = orderSwap(typeHolder.strnxfrm<UT>(s.rtrimZero()));
filterArgsVectors.push_back(simdProcessor.loadValue(bigEndianFilterWeights));
}
else
{
FilterType filterValue = *((FilterType*)&filterValues[j]);
filterArgsVectors.push_back(simdProcessor.loadValue(filterValue));
}
switch (filterCOPs[j])
{
case (COMPARE_EQ):
// Skipping extra filter pass generated by IS NULL
// Filter against NULL value
if (memcmp(&filterValues[j], &nullValue, sizeof(nullValue)) == 0)
copFunctorVec.push_back(std::mem_fn(&VT::nullEmptyCmpEq));
else
@ -1329,9 +1428,10 @@ void vectorizedFiltering(NewColRequestHeader* in, ColResultHeader* out, const T*
{
primitives::RIDType ridOffset = i * VECTOR_SIZE;
assert(!HAS_INPUT_RIDS || (HAS_INPUT_RIDS && ridSize >= ridOffset));
dataVec = simdDataLoadTemplate<VT, SimdWrapperType, HAS_INPUT_RIDS, T>(simdProcessor, srcArray,
origSrcArray, ridArray, i)
.v;
dataVec = simdDataLoad<VT, SimdWrapperType, HAS_INPUT_RIDS, T>(simdProcessor, srcArray,
origSrcArray, ridArray, i).v;
if constexpr(KIND==KIND_TEXT)
swapedOrderDataVec = simdSwapedOrderDataLoad<KIND, VT, SimdWrapperType, T>(typeHolder, simdProcessor, dataVec).v;
nonEmptyMask = simdProcessor.nullEmptyCmpNe(dataVec, emptyFilterArgVec);
writeMask = nonEmptyMask;
// NULL check
@ -1346,7 +1446,11 @@ void vectorizedFiltering(NewColRequestHeader* in, ColResultHeader* out, const T*
for (uint32_t j = 0; j < filterCount; ++j)
{
// filter using compiled filter and preloaded filter argument
filterMask = copFunctorVec[j](simdProcessor, dataVec, filterArgsVectors[j]);
if constexpr(KIND==KIND_TEXT)
filterMask = copFunctorVec[j](simdProcessor, swapedOrderDataVec, filterArgsVectors[j]);
else
filterMask = copFunctorVec[j](simdProcessor, dataVec, filterArgsVectors[j]);
filterMask = bopFunctor(prevFilterMask, filterMask);
prevFilterMask = filterMask;
}
@ -1389,7 +1493,6 @@ void vectorizedFiltering(NewColRequestHeader* in, ColResultHeader* out, const T*
out->Min = Min;
out->Max = Max;
}
// process the tail. scalarFiltering changes out contents, e.g. Min/Max, NVALS, RIDs and values array
// This tail also sets out::Min/Max, out::validMinMax if validMinMax is set.
uint32_t processedSoFar = rid;
@ -1526,7 +1629,8 @@ void filterColumnData(NewColRequestHeader* in, ColResultHeader* out, uint16_t* r
#if defined(__x86_64__)
// Don't use vectorized filtering for text based data types.
if (KIND <= KIND_FLOAT && WIDTH < 16)
if (WIDTH < 16 &&
(KIND != KIND_TEXT || (KIND == KIND_TEXT && in->colType.strnxfrmIsValid()) ))
{
bool canUseFastFiltering = true;
for (uint32_t i = 0; i < filterCount; ++i)
@ -1672,6 +1776,8 @@ template <typename T,
void PrimitiveProcessor::_scanAndFilterTypeDispatcher(NewColRequestHeader* in, ColResultHeader* out)
{
constexpr int W = sizeof(T);
using UT = typename std::conditional<std::is_unsigned<T>::value || datatypes::is_uint128_t<T>::value, T,
typename datatypes::make_unsigned<T>::type>::type;
const uint16_t ridSize = in->NVALS;
uint16_t* ridArray = in->getRIDArrayPtr(W);
const uint32_t itemsPerBlock = logicalBlockMode ? BLOCK_SIZE : BLOCK_SIZE / W;
@ -1682,16 +1788,12 @@ void PrimitiveProcessor::_scanAndFilterTypeDispatcher(NewColRequestHeader* in, C
dataType == execplan::CalpontSystemCatalog::TEXT) &&
!isDictTokenScan(in))
{
using UT = typename std::conditional<std::is_unsigned<T>::value, T,
typename datatypes::make_unsigned<T>::type>::type;
filterColumnData<UT, KIND_TEXT>(in, out, ridArray, ridSize, block, itemsPerBlock, parsedColumnFilter);
return;
}
if (datatypes::isUnsigned(dataType))
{
using UT = typename std::conditional<std::is_unsigned<T>::value || datatypes::is_uint128_t<T>::value, T,
typename datatypes::make_unsigned<T>::type>::type;
filterColumnData<UT, KIND_UNSIGNED>(in, out, ridArray, ridSize, block, itemsPerBlock, parsedColumnFilter);
return;
}