hcq/oceanbase
hcq/oceanbase
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

idispatch SIMD function by cpuinfo for PFoR codec

Browse Source
This commit is contained in:
saltonz 2024-03-11 02:46:06 +00:00 committed by ob-robot
parent b11a5ec7fd
commit 18ddde9250
1 changed files with 59 additions and 66 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

125
deps/oblib/src/lib/codec/ob_simd_fixed_pfor.h vendored
View File

@ -16,6 +16,7 @@
#include "ob_codecs.h"
#include "ob_bp_util.h"
#include "ob_generated_unalign_simd_bp_func.h"
#include "common/ob_target_specific.h"
namespace oceanbase
{
@ -110,7 +111,7 @@ public:
static OB_INLINE void inner_do_encode(
const UIntT *__restrict in,
uint32_t n,
char *__restrict out,
char *out,
const uint32_t b,
const uint32_t bx,
const uint64_t out_buf_len,
@ -124,20 +125,7 @@ public:
char *orig_out = out;
if (bx == 0) { // no exception
if (4 == sizeof(UIntT)) {
// uint32_t simd packing
uSIMD_fastpackwithoutmask_128_32((uint32_t *)in, reinterpret_cast<__m128i *>(out), b);
out += BlockSize * b / 8;
} else if (2 == sizeof(UIntT)) {
// uint16_t simd packing
uSIMD_fastpackwithoutmask_128_16((const uint16_t *)in, reinterpret_cast<__m128i *>(out), b);
out += BlockSize * b / 8;
} else {
// uint64_t, uint8_t use scalar packing
uint64_t out_pos = 0;
scalar_bit_packing<UIntT>(in, BlockSize, b, out, out_buf_len, out_pos);
out += out_pos;
}
inner_bit_packing<UIntT>(in, out, b, out_buf_len);
} else {
uint64_t i = 0;
uint64_t xn = 0;// the count of exceptions
@ -174,18 +162,7 @@ public:
}
// packing data
if (4 == sizeof(UIntT)) {
uSIMD_fastpackwithoutmask_128_32((uint32_t *)_in, reinterpret_cast<__m128i *>(out), b);
out += BlockSize * b / 8;
} else if (2 == sizeof(UIntT)) {
uSIMD_fastpackwithoutmask_128_16((const uint16_t *)_in, reinterpret_cast<__m128i *>(out), b);
out += BlockSize * b / 8;
} else {
// uint8 & uint16 does not support simd packing
uint64_t out_pos = 0;
scalar_bit_packing<UIntT>(_in, BlockSize, b, out, out_buf_len, out_pos);
out += out_pos;
}
inner_bit_packing<UIntT>(_in, out, b, out_buf_len);
}
len = (uint32_t)(out - orig_out);
}
@ -307,24 +284,7 @@ public:
b = *in++; // bit width
if (0 == (b & 0x80)) { // no exception value, direct unpack
if (4 == sizeof(UIntT)) {
uSIMD_fastunpack_128_32(reinterpret_cast<const __m128i *>(in), reinterpret_cast<uint32_t *>(out), b);
in += BlockSize * b / 8; // convert to byte;
} else if (2 == sizeof(UIntT)) {
uSIMD_fastunpack_128_16(reinterpret_cast<const __m128i *>(in), reinterpret_cast<uint16_t *>(out), b);
in += BlockSize * b / 8;
} else {
// uint8 & uint64 does not support simd packing
uint64_t in_pos = in - _in;
uint64_t in_len = length;
uint64_t tmp_out_pos = 0;
uint64_t tmp_out_buf_len = BlockSize;
UIntT *tmp_out = (UIntT *)out;
scalar_bit_unpacking<UIntT>(_in, in_len, in_pos, BlockSize, b,
tmp_out, tmp_out_buf_len, tmp_out_pos);
in = _in + in_pos;
}
out += BlockSize * sizeof(UIntT);
inner_bit_unpacking<UIntT>(in, _in, length, out, b);
} else {
b &= (0x80 - 1); // get normal bit width
bx = *in++; // get exception bit width
@ -346,27 +306,9 @@ public:
in = _in + in_pos;
}
if (4 == sizeof(UIntT)) {
// unpacking data
uSIMD_fastunpack_128_32(reinterpret_cast<const __m128i *>(in), reinterpret_cast<uint32_t *>(out), b);
in += BlockSize * b / 8;
} else if (2 == sizeof(UIntT)) {
uSIMD_fastunpack_128_16(reinterpret_cast<const __m128i *>(in), reinterpret_cast<uint16_t *>(out), b);
in += BlockSize * b / 8;
} else {
// uint8 & uint64 does not support simd packing
uint64_t in_pos = in - _in;
uint64_t in_len = length;
uint64_t tmp_out_pos = 0;
uint64_t tmp_out_buf_len = BlockSize;
UIntT *tmp_out = (UIntT *)out;
scalar_bit_unpacking<UIntT>(_in, in_len, in_pos, BlockSize, b,
tmp_out, tmp_out_buf_len, tmp_out_pos);
in = _in + in_pos;
}
UIntT *out_arr = reinterpret_cast<UIntT *>(out);
out += BlockSize * sizeof(UIntT);
// unpacking data
inner_bit_unpacking<UIntT>(in, _in, length, out, b);
UIntT *out_arr = reinterpret_cast<UIntT *>(_out + out_pos);
// patch exception, TODO, oushen, optimize later
int64_t ex_idx = 0;
@ -440,6 +382,57 @@ public:
}
virtual const char * name() const override { return "ObSIMDFixedPFor128"; }
private:
template<typename UIntT>
static OB_INLINE void inner_bit_packing(
const UIntT *__restrict in,
char *&out,
const uint32_t bit,
const uint64_t out_buf_len)
{
if (4 == sizeof(UIntT) && common::is_arch_supported(ObTargetArch::AVX2)) {
// uint32_t simd packing
uSIMD_fastpackwithoutmask_128_32((uint32_t *)in, reinterpret_cast<__m128i *>(out), bit);
out += BlockSize * bit / 8;
} else if (2 == sizeof(UIntT) && common::is_arch_supported(ObTargetArch::AVX2)) {
// uint16_t simd packing
uSIMD_fastpackwithoutmask_128_16((const uint16_t *)in, reinterpret_cast<__m128i *>(out), bit);
out += BlockSize * bit / 8;
} else {
// uint64_t, uint8_t use scalar packing
uint64_t out_pos = 0;
scalar_bit_packing<UIntT>(in, BlockSize, bit, out, out_buf_len, out_pos);
out += out_pos;
}
}
template<typename UIntT>
static OB_INLINE void inner_bit_unpacking(
const char *&in,
const char *_in,
const uint64_t length,
char *&out,
const uint32_t bit)
{
if (4 == sizeof(UIntT) && common::is_arch_supported(ObTargetArch::AVX2)) {
uSIMD_fastunpack_128_32(reinterpret_cast<const __m128i *>(in), reinterpret_cast<uint32_t *>(out), bit);
in += BlockSize * bit / 8; // convert to byte;
} else if (2 == sizeof(UIntT) && common::is_arch_supported(ObTargetArch::AVX2)) {
uSIMD_fastunpack_128_16(reinterpret_cast<const __m128i *>(in), reinterpret_cast<uint16_t *>(out), bit);
in += BlockSize * bit / 8;
} else {
// uint8 & uint64 does not support simd packing
uint64_t in_pos = in - _in;
uint64_t in_len = length;
uint64_t tmp_out_pos = 0;
uint64_t tmp_out_buf_len = BlockSize;
UIntT *tmp_out = (UIntT *)out;
scalar_bit_unpacking<UIntT>(_in, in_len, in_pos, BlockSize, bit,
tmp_out, tmp_out_buf_len, tmp_out_pos);
in = _in + in_pos;
}
out += BlockSize * sizeof(UIntT);
}
};
} // namespace common