database/mariadb-columnstore-engine
1
0
Fork 0
mirror of https://github.com/mariadb-corporation/mariadb-columnstore-engine.git synced 2025-07-29 08:21:15 +03:00
Code Activity

MCOL-4313 Very fragile but high speed approach with inline ASM

Browse Source 
GCC compiler uses aligned versions of SIMD instructions expecting
aligned memory blocks that is hard to implement now
This commit is contained in:
Roman Nozdrin
2020-10-13 12:43:16 +00:00
parent 7d3d828790
commit 844472d812
4 changed files with 65 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
primitives/primproc/columncommand.cpp
View File

@ -231,7 +231,11 @@ void ColumnCommand::loadData()
{
ByteStream::hexbyte h;
utils::getEmptyRowValue(colType.colDataType, colType.colWidth, (uint8_t*)&h);
hPtr[idx] = h;
__asm__ volatile("movups %1,%0"
:"=m" ( hPtr[idx] ) // output
:"v"( h ) // input
: "memory" // clobbered
);
}
}
@ -328,7 +332,18 @@ void ColumnCommand::process_OT_BOTH()
bpp->relRids[i] = *((uint16_t*) &bpp->outputMsg[pos]);
pos += 2;
wide128Values[i] = *((int128_t*) &bpp->outputMsg[pos]);
int128_t* int128Ptr = reinterpret_cast<int128_t*>(&bpp->outputMsg[pos]);
__asm__ volatile("movdqu %0,%%xmm0;"
:
:"m"( *int128Ptr ) // input
:"xmm0" // clobbered
);
__asm__ volatile("movups %%xmm0,%0;"
: "=m" (wide128Values[i])// output
: // input
: "memory", "xmm0" // clobbered
);
pos += 16;
}

35
utils/messageqcpp/bytestream.cpp
View File

@ -240,7 +240,12 @@ ByteStream& ByteStream::operator<<(const uint128_t& o)
if (fBuf == 0 || (fCurInPtr - fBuf + 16U > fMaxLen + ISSOverhead))
growBuf(fMaxLen + BlockSize);
*((uint128_t*) fCurInPtr) = o;
__asm__ volatile("movups %1,%0;"
:"=m" ( *fCurInPtr ) // output
:"v"( o ) // input
: "memory" // clobbered
);
fCurInPtr += 16;
return *this;
@ -251,7 +256,11 @@ ByteStream& ByteStream::operator<<(const int128_t& o)
if (fBuf == 0 || (fCurInPtr - fBuf + 16U > fMaxLen + ISSOverhead))
growBuf(fMaxLen + BlockSize);
*((int128_t*) fCurInPtr) = o;
__asm__ volatile("movups %1,%0;"
:"=m" ( *fCurInPtr ) // output
:"v"( o ) // input
: "memory" // clobbered
);
fCurInPtr += 16;
return *this;
@ -441,7 +450,16 @@ void ByteStream::peek(uint128_t& o) const
if (length() < 16)
throw underflow_error("ByteStream>uint128_t: not enough data in stream to fill datatype");
o = *((uint128_t*) fCurOutPtr);
__asm__ volatile("movdqu %0,%%xmm0;"
:
:"m"( *fCurOutPtr ) // input
:"xmm0" // clobbered
);
__asm__ volatile("movups %%xmm0,%0;"
: "=m" (o)// output
: // input
: "memory", "xmm0" // clobbered
);
}
void ByteStream::peek(int128_t& o) const
@ -450,7 +468,16 @@ void ByteStream::peek(int128_t& o) const
if (length() < 16)
throw underflow_error("ByteStream>int128_t: not enough data in stream to fill datatype");
o = *((int128_t*) fCurOutPtr);
__asm__ volatile("movdqu %0,%%xmm0;"
:
:"m"( *fCurOutPtr ) // input
:"xmm0" // clobbered
);
__asm__ volatile("movups %%xmm0,%0;"
: "=m" (o)// output
: // input
: "memory", "xmm0" // clobbered
);
}
void ByteStream::peek(string& s) const

1
utils/messageqcpp/bytestream.h
View File

@ -151,6 +151,7 @@ public:
* push an int128_t onto the end of the stream. The byte order is whatever the native byte order is.
*/
EXPORT ByteStream& operator<<(const int128_t& o);
/**
* push an uint128_t onto the end of the stream. The byte order is whatever the native byte order is.
*/

16
utils/rowgroup/rowgroup.h
View File

@ -848,6 +848,22 @@ inline void Row::setBinaryField_offset<uint8_t>(uint8_t* value, uint32_t width,
memcpy(&data[offset], value, width);
}
template<>
inline void Row::setBinaryField_offset<int128_t>(int128_t* value, uint32_t width, uint32_t offset)
{
int128_t *dst128Ptr = reinterpret_cast<int128_t*>(&data[offset]);
__asm__ volatile("movdqu %0,%%xmm0;"
:
:"m"( *value ) // input
:"xmm0" // clobbered
);
__asm__ volatile("movups %%xmm0,%0;"
: "=m" (*dst128Ptr)// output
: // input
: "memory", "xmm0" // clobbered
);
}
inline void Row::setStringField(const uint8_t* strdata, uint32_t length, uint32_t colIndex)
{
uint64_t offset;