mirror of
https://github.com/ClickHouse/ClickHouse.git
synced 2024-11-21 15:12:02 +00:00
dbms: Server: Performance improvements. [#METR-15210]
This commit is contained in:
Alexey Arno
parent
79bad96d98
commit
8125cad146
@ -295,13 +295,13 @@ namespace DB
|
|||||||
struct FloatRoundingComputation<Float32, rounding_mode, PositiveScale>
|
struct FloatRoundingComputation<Float32, rounding_mode, PositiveScale>
|
||||||
: public BaseFloatRoundingComputation<Float32>
|
: public BaseFloatRoundingComputation<Float32>
|
||||||
{
|
{
|
||||||
static inline void prepareScale(size_t scale, Scale & mm_scale)
|
static inline void prepare(size_t scale, Scale & mm_scale)
|
||||||
{
|
{
|
||||||
Float32 fscale = static_cast<Float32>(scale);
|
Float32 fscale = static_cast<Float32>(scale);
|
||||||
mm_scale = _mm_load1_ps(&fscale);
|
mm_scale = _mm_load1_ps(&fscale);
|
||||||
}
|
}
|
||||||
|
|
||||||
static inline void compute(const Float32 * in, const Scale & scale, Float32 * out)
|
static inline void compute(const Float32 * __restrict in, const Scale & scale, Float32 * __restrict out)
|
||||||
{
|
{
|
||||||
__m128 val = _mm_loadu_ps(in);
|
__m128 val = _mm_loadu_ps(in);
|
||||||
val = _mm_mul_ps(val, scale);
|
val = _mm_mul_ps(val, scale);
|
||||||
@ -315,13 +315,13 @@ namespace DB
|
|||||||
struct FloatRoundingComputation<Float32, rounding_mode, NegativeScale>
|
struct FloatRoundingComputation<Float32, rounding_mode, NegativeScale>
|
||||||
: public BaseFloatRoundingComputation<Float32>
|
: public BaseFloatRoundingComputation<Float32>
|
||||||
{
|
{
|
||||||
static inline void prepareScale(size_t scale, Scale & mm_scale)
|
static inline void prepare(size_t scale, Scale & mm_scale)
|
||||||
{
|
{
|
||||||
Float32 fscale = static_cast<Float32>(scale);
|
Float32 fscale = static_cast<Float32>(scale);
|
||||||
mm_scale = _mm_load1_ps(&fscale);
|
mm_scale = _mm_load1_ps(&fscale);
|
||||||
}
|
}
|
||||||
|
|
||||||
static inline void compute(const Float32 * in, const Scale & scale, Float32 * out)
|
static inline void compute(const Float32 * __restrict in, const Scale & scale, Float32 * __restrict out)
|
||||||
{
|
{
|
||||||
__m128 val = _mm_loadu_ps(in);
|
__m128 val = _mm_loadu_ps(in);
|
||||||
|
|
||||||
@ -379,11 +379,11 @@ namespace DB
|
|||||||
struct FloatRoundingComputation<Float32, rounding_mode, ZeroScale>
|
struct FloatRoundingComputation<Float32, rounding_mode, ZeroScale>
|
||||||
: public BaseFloatRoundingComputation<Float32>
|
: public BaseFloatRoundingComputation<Float32>
|
||||||
{
|
{
|
||||||
static inline void prepareScale(size_t scale, Scale & mm_scale)
|
static inline void prepare(size_t scale, Scale & mm_scale)
|
||||||
{
|
{
|
||||||
}
|
}
|
||||||
|
|
||||||
static inline void compute(const Float32 * in, const Scale & scale, Float32 * out)
|
static inline void compute(const Float32 * __restrict in, const Scale & scale, Float32 * __restrict out)
|
||||||
{
|
{
|
||||||
__m128 val = _mm_loadu_ps(in);
|
__m128 val = _mm_loadu_ps(in);
|
||||||
val = _mm_round_ps(val, rounding_mode);
|
val = _mm_round_ps(val, rounding_mode);
|
||||||
@ -395,13 +395,13 @@ namespace DB
|
|||||||
struct FloatRoundingComputation<Float64, rounding_mode, PositiveScale>
|
struct FloatRoundingComputation<Float64, rounding_mode, PositiveScale>
|
||||||
: public BaseFloatRoundingComputation<Float64>
|
: public BaseFloatRoundingComputation<Float64>
|
||||||
{
|
{
|
||||||
static inline void prepareScale(size_t scale, Scale & mm_scale)
|
static inline void prepare(size_t scale, Scale & mm_scale)
|
||||||
{
|
{
|
||||||
Float64 fscale = static_cast<Float64>(scale);
|
Float64 fscale = static_cast<Float64>(scale);
|
||||||
mm_scale = _mm_load1_pd(&fscale);
|
mm_scale = _mm_load1_pd(&fscale);
|
||||||
}
|
}
|
||||||
|
|
||||||
static inline void compute(const Float64 * in, const Scale & scale, Float64 * out)
|
static inline void compute(const Float64 * __restrict in, const Scale & scale, Float64 * __restrict out)
|
||||||
{
|
{
|
||||||
__m128d val = _mm_loadu_pd(in);
|
__m128d val = _mm_loadu_pd(in);
|
||||||
val = _mm_mul_pd(val, scale);
|
val = _mm_mul_pd(val, scale);
|
||||||
@ -415,13 +415,13 @@ namespace DB
|
|||||||
struct FloatRoundingComputation<Float64, rounding_mode, NegativeScale>
|
struct FloatRoundingComputation<Float64, rounding_mode, NegativeScale>
|
||||||
: public BaseFloatRoundingComputation<Float64>
|
: public BaseFloatRoundingComputation<Float64>
|
||||||
{
|
{
|
||||||
static inline void prepareScale(size_t scale, Scale & mm_scale)
|
static inline void prepare(size_t scale, Scale & mm_scale)
|
||||||
{
|
{
|
||||||
Float64 fscale = static_cast<Float64>(scale);
|
Float64 fscale = static_cast<Float64>(scale);
|
||||||
mm_scale = _mm_load1_pd(&fscale);
|
mm_scale = _mm_load1_pd(&fscale);
|
||||||
}
|
}
|
||||||
|
|
||||||
static inline void compute(const Float64 * in, const Scale & scale, Float64 * out)
|
static inline void compute(const Float64 * __restrict in, const Scale & scale, Float64 * __restrict out)
|
||||||
{
|
{
|
||||||
__m128d val = _mm_loadu_pd(in);
|
__m128d val = _mm_loadu_pd(in);
|
||||||
|
|
||||||
@ -479,11 +479,11 @@ namespace DB
|
|||||||
struct FloatRoundingComputation<Float64, rounding_mode, ZeroScale>
|
struct FloatRoundingComputation<Float64, rounding_mode, ZeroScale>
|
||||||
: public BaseFloatRoundingComputation<Float64>
|
: public BaseFloatRoundingComputation<Float64>
|
||||||
{
|
{
|
||||||
static inline void prepareScale(size_t scale, Scale & mm_scale)
|
static inline void prepare(size_t scale, Scale & mm_scale)
|
||||||
{
|
{
|
||||||
}
|
}
|
||||||
|
|
||||||
static inline void compute(const Float64 * in, const Scale & scale, Float64 * out)
|
static inline void compute(const Float64 * __restrict in, const Scale & scale, Float64 * __restrict out)
|
||||||
{
|
{
|
||||||
__m128d val = _mm_loadu_pd(in);
|
__m128d val = _mm_loadu_pd(in);
|
||||||
val = _mm_round_pd(val, rounding_mode);
|
val = _mm_round_pd(val, rounding_mode);
|
||||||
@ -512,11 +512,11 @@ namespace DB
|
|||||||
{
|
{
|
||||||
auto divisor = Op::prepare(scale);
|
auto divisor = Op::prepare(scale);
|
||||||
|
|
||||||
auto begin_in = &in[0];
|
const T* begin_in = &in[0];
|
||||||
auto end_in = begin_in + in.size();
|
const T* end_in = begin_in + in.size();
|
||||||
auto p_out = &out[0];
|
|
||||||
|
|
||||||
for (auto p_in = begin_in; p_in != end_in; ++p_in)
|
T* __restrict p_out = &out[0];
|
||||||
|
for (const T* __restrict p_in = begin_in; p_in != end_in; ++p_in)
|
||||||
{
|
{
|
||||||
*p_out = Op::compute(*p_in, divisor);
|
*p_out = Op::compute(*p_in, divisor);
|
||||||
++p_out;
|
++p_out;
|
||||||
@ -545,43 +545,45 @@ namespace DB
|
|||||||
static inline void apply(const PODArray<T> & in, size_t scale, typename ColumnVector<T>::Container_t & out)
|
static inline void apply(const PODArray<T> & in, size_t scale, typename ColumnVector<T>::Container_t & out)
|
||||||
{
|
{
|
||||||
Scale mm_scale;
|
Scale mm_scale;
|
||||||
Op::prepareScale(scale, mm_scale);
|
Op::prepare(scale, mm_scale);
|
||||||
|
|
||||||
const size_t data_count = std::tuple_size<Data>();
|
const size_t data_count = std::tuple_size<Data>();
|
||||||
|
|
||||||
auto begin_in = &in[0];
|
const T* begin_in = &in[0];
|
||||||
auto end_in = begin_in + in.size();
|
const T* end_in = begin_in + in.size();
|
||||||
auto limit = end_in - (data_count - 1);
|
|
||||||
|
|
||||||
auto begin_out = &out[0];
|
T* begin_out = &out[0];
|
||||||
auto end_out = begin_out + out.size();
|
const T* end_out = begin_out + out.size();
|
||||||
|
|
||||||
auto p_in = begin_in;
|
const T* limit = end_in - (data_count - 1);
|
||||||
auto p_out = begin_out;
|
|
||||||
|
const T* __restrict p_in = begin_in;
|
||||||
|
T* __restrict p_out = begin_out;
|
||||||
for (; p_in < limit; p_in += data_count)
|
for (; p_in < limit; p_in += data_count)
|
||||||
{
|
{
|
||||||
Op::compute(reinterpret_cast<const T *>(p_in), mm_scale, reinterpret_cast<T *>(p_out));
|
Op::compute(p_in, mm_scale, p_out);
|
||||||
p_out += data_count;
|
p_out += data_count;
|
||||||
}
|
}
|
||||||
|
|
||||||
if (p_in < end_in)
|
if (p_in < end_in)
|
||||||
{
|
{
|
||||||
Data tmp{0};
|
Data tmp{0};
|
||||||
auto begin_tmp = &tmp[0];
|
T* begin_tmp = &tmp[0];
|
||||||
auto end_tmp = begin_tmp + data_count;
|
const T* end_tmp = begin_tmp + data_count;
|
||||||
|
|
||||||
for (auto p_tmp = begin_tmp; (p_tmp != end_tmp) && (p_in != end_in); ++p_tmp)
|
for (T* __restrict p_tmp = begin_tmp; (p_tmp != end_tmp) && (p_in != end_in); ++p_tmp)
|
||||||
{
|
{
|
||||||
*p_tmp = *p_in;
|
*p_tmp = *p_in;
|
||||||
++p_in;
|
++p_in;
|
||||||
}
|
}
|
||||||
|
|
||||||
Data res;
|
Data res;
|
||||||
|
const T* begin_res = &res[0];
|
||||||
|
const T* end_res = begin_res + data_count;
|
||||||
|
|
||||||
Op::compute(reinterpret_cast<T *>(&tmp), mm_scale, reinterpret_cast<T *>(&res));
|
Op::compute(reinterpret_cast<T *>(&tmp), mm_scale, reinterpret_cast<T *>(&res));
|
||||||
|
|
||||||
auto begin_res = &res[0];
|
for (const T* __restrict p_res = begin_res; (p_res != end_res) && (p_out != end_out); ++p_res)
|
||||||
auto end_res = begin_res + data_count;
|
|
||||||
for (auto p_res = begin_res; (p_res != end_res) && (p_out != end_out); ++p_res)
|
|
||||||
{
|
{
|
||||||
*p_out = *p_res;
|
*p_out = *p_res;
|
||||||
++p_out;
|
++p_out;
|
||||||
@ -596,7 +598,7 @@ namespace DB
|
|||||||
else
|
else
|
||||||
{
|
{
|
||||||
Scale mm_scale;
|
Scale mm_scale;
|
||||||
Op::prepareScale(scale, mm_scale);
|
Op::prepare(scale, mm_scale);
|
||||||
|
|
||||||
Data tmp{0};
|
Data tmp{0};
|
||||||
tmp[0] = val;
|
tmp[0] = val;
|
||||||
|
|||||||
@ -458,16 +458,16 @@
|
|||||||
-13 -13 -13
|
-13 -13 -13
|
||||||
-13 -13 -13
|
-13 -13 -13
|
||||||
-13 -13 -13
|
-13 -13 -13
|
||||||
-16 -6 -16
|
-10 -20 -10
|
||||||
-16 -6 -16
|
-10 -20 -10
|
||||||
-16 -6 -16
|
-10 -20 -10
|
||||||
-16 -6 -16
|
-10 -20 -10
|
||||||
-13 -6 -16
|
-13 -20 -10
|
||||||
-13 -6 -16
|
-13 -20 -10
|
||||||
-16 84 -16
|
0 -100 0
|
||||||
-16 84 -16
|
0 -100 0
|
||||||
-16 84 -16
|
0 -100 0
|
||||||
-16 84 -16
|
0 -100 0
|
||||||
0 0 0
|
0 0 0
|
||||||
0 0 0
|
0 0 0
|
||||||
-13 -13 -13
|
-13 -13 -13
|
||||||
@ -499,18 +499,18 @@
|
|||||||
-13 -13 -13
|
-13 -13 -13
|
||||||
-13 -13 -13
|
-13 -13 -13
|
||||||
-13 -13 -13
|
-13 -13 -13
|
||||||
-16 -6 -16
|
-10 -20 -10
|
||||||
-16 -6 -16
|
-10 -20 -10
|
||||||
-16 -6 -16
|
-10 -20 -10
|
||||||
-16 -6 -16
|
-10 -20 -10
|
||||||
-13 -6 -16
|
-13 -20 -10
|
||||||
-13 -6 -16
|
-13 -20 -10
|
||||||
-16 84 -16
|
0 -100 0
|
||||||
-16 84 -16
|
0 -100 0
|
||||||
-16 84 -16
|
0 -100 0
|
||||||
-16 84 -16
|
0 -100 0
|
||||||
-16 84 -16
|
0 -100 0
|
||||||
-16 84 -16
|
0 -100 0
|
||||||
-13 -13 -13
|
-13 -13 -13
|
||||||
-13 -13 -13
|
-13 -13 -13
|
||||||
-13 -13 -13
|
-13 -13 -13
|
||||||
@ -540,18 +540,18 @@
|
|||||||
-13 -13 -13
|
-13 -13 -13
|
||||||
-13 -13 -13
|
-13 -13 -13
|
||||||
-13 -13 -13
|
-13 -13 -13
|
||||||
-16 -6 -16
|
-10 -20 -10
|
||||||
-16 -6 -16
|
-10 -20 -10
|
||||||
-16 -6 -16
|
-10 -20 -10
|
||||||
-16 -6 -16
|
-10 -20 -10
|
||||||
-13 -6 -16
|
-13 -20 -10
|
||||||
-13 -6 -16
|
-13 -20 -10
|
||||||
-16 84 -16
|
0 -100 0
|
||||||
-16 84 -16
|
0 -100 0
|
||||||
-16 84 -16
|
0 -100 0
|
||||||
-16 84 -16
|
0 -100 0
|
||||||
-16 84 -16
|
0 -100 0
|
||||||
-16 84 -16
|
0 -100 0
|
||||||
-13 -13 -13
|
-13 -13 -13
|
||||||
-13 -13 -13
|
-13 -13 -13
|
||||||
-13 -13 -13
|
-13 -13 -13
|
||||||
@ -581,18 +581,18 @@
|
|||||||
-13 -13 -13
|
-13 -13 -13
|
||||||
-13 -13 -13
|
-13 -13 -13
|
||||||
-13 -13 -13
|
-13 -13 -13
|
||||||
-16 -6 -16
|
-10 -20 -10
|
||||||
-16 -6 -16
|
-10 -20 -10
|
||||||
-16 -6 -16
|
-10 -20 -10
|
||||||
-16 -6 -16
|
-10 -20 -10
|
||||||
-13 -6 -16
|
-13 -20 -10
|
||||||
-13 -6 -16
|
-13 -20 -10
|
||||||
-16 84 -16
|
0 -100 0
|
||||||
-16 84 -16
|
0 -100 0
|
||||||
-16 84 -16
|
0 -100 0
|
||||||
-16 84 -16
|
0 -100 0
|
||||||
-16 84 -16
|
0 -100 0
|
||||||
-16 84 -16
|
0 -100 0
|
||||||
-13 -13 -13
|
-13 -13 -13
|
||||||
-13 -13 -13
|
-13 -13 -13
|
||||||
-13 -13 -13
|
-13 -13 -13
|
||||||
@ -622,18 +622,18 @@
|
|||||||
-13 -13 -13
|
-13 -13 -13
|
||||||
-13 -13 -13
|
-13 -13 -13
|
||||||
-13 -13 -13
|
-13 -13 -13
|
||||||
0 0 0
|
-10 -20 -10
|
||||||
0 0 0
|
-10 -20 -10
|
||||||
0 0 0
|
-10 -20 -10
|
||||||
0 0 0
|
-10 -20 -10
|
||||||
-13 0 0
|
-13 -20 -10
|
||||||
-13 0 0
|
-13 -20 -10
|
||||||
0 0 0
|
0 -100 0
|
||||||
0 0 0
|
0 -100 0
|
||||||
0 0 0
|
0 -100 0
|
||||||
0 0 0
|
0 -100 0
|
||||||
0 0 0
|
0 -100 0
|
||||||
0 0 0
|
0 -100 0
|
||||||
-13 -13 -13
|
-13 -13 -13
|
||||||
-13 -13 -13
|
-13 -13 -13
|
||||||
-13 -13 -13
|
-13 -13 -13
|
||||||
@ -663,18 +663,18 @@
|
|||||||
-13 -13 -13
|
-13 -13 -13
|
||||||
-13 -13 -13
|
-13 -13 -13
|
||||||
-13 -13 -13
|
-13 -13 -13
|
||||||
0 0 0
|
-10 -20 -10
|
||||||
0 0 0
|
-10 -20 -10
|
||||||
0 0 0
|
-10 -20 -10
|
||||||
0 0 0
|
-10 -20 -10
|
||||||
-13 0 0
|
-13 -20 -10
|
||||||
-13 0 0
|
-13 -20 -10
|
||||||
0 0 0
|
0 -100 0
|
||||||
0 0 0
|
0 -100 0
|
||||||
0 0 0
|
0 -100 0
|
||||||
0 0 0
|
0 -100 0
|
||||||
0 0 0
|
0 -100 0
|
||||||
0 0 0
|
0 -100 0
|
||||||
2.72 2.72 2.71
|
2.72 2.72 2.71
|
||||||
2.72 2.72 2.71
|
2.72 2.72 2.71
|
||||||
2.72 2.72 2.71
|
2.72 2.72 2.71
|
||||||
@ -745,12 +745,12 @@
|
|||||||
-3 -2 -3
|
-3 -2 -3
|
||||||
-3 -2 -3
|
-3 -2 -3
|
||||||
-3 -2 -3
|
-3 -2 -3
|
||||||
0 0 0
|
0 -10 0
|
||||||
0 0 0
|
0 -10 0
|
||||||
0 0 0
|
0 -10 0
|
||||||
0 0 0
|
0 -10 0
|
||||||
-2.7 0 0
|
-2.7 -10 0
|
||||||
-2.7 0 0
|
-2.7 -10 0
|
||||||
0 0 0
|
0 0 0
|
||||||
0 0 0
|
0 0 0
|
||||||
0 0 0
|
0 0 0
|
||||||
|
|||||||
Loading…
Reference in New Issue
Block a user