dbms: Server: Avoid negative zeroes in all of the remaining cases. [#METR-15210]

This commit is contained in:
Alexey Arno 2015-06-03 20:01:30 +03:00
parent 1037433f3d
commit b86aa4785c
3 changed files with 207 additions and 157 deletions

View File

@ -257,92 +257,25 @@ namespace DB
}; };
template<typename T> template<typename T>
struct BaseFloatRoundingComputation; class BaseFloatRoundingComputation;
template<> template<>
struct BaseFloatRoundingComputation<Float32> class BaseFloatRoundingComputation<Float32>
{ {
public:
using Scale = __m128; using Scale = __m128;
static const size_t data_count = 4; static const size_t data_count = 4;
};
template<> protected:
struct BaseFloatRoundingComputation<Float64> /// Предотвратить появление отрицательных нолей определённых в стандарте IEEE-754.
{ static inline void normalize(__m128 & val, const __m128 & mask)
using Scale = __m128d;
static const size_t data_count = 2;
};
/** Реализация низкоуровневых функций округления для значений с плавающей точкой.
*/
template<typename T, int rounding_mode, ScaleMode scale_mode>
struct FloatRoundingComputation : public BaseFloatRoundingComputation<T>
{
};
template<int rounding_mode>
struct FloatRoundingComputation<Float32, rounding_mode, PositiveScale>
: public BaseFloatRoundingComputation<Float32>
{
static inline void prepare(size_t scale, Scale & mm_scale)
{ {
Float32 fscale = static_cast<Float32>(scale); __m128 mask1 = _mm_cmpeq_ps(val, getZero());
mm_scale = _mm_load1_ps(&fscale); __m128 mask2 = _mm_and_ps(mask, mask1);
} mask2 = _mm_cmpeq_ps(mask2, getZero());
mask2 = _mm_min_ps(mask2, getTwo());
static inline void compute(const Float32 * __restrict in, const Scale & scale, Float32 * __restrict out) mask2 = _mm_sub_ps(mask2, getOne());
{ val = _mm_mul_ps(val, mask2);
__m128 val = _mm_loadu_ps(in);
val = _mm_mul_ps(val, scale);
val = _mm_round_ps(val, rounding_mode);
val = _mm_div_ps(val, scale);
_mm_storeu_ps(out, val);
}
};
template<int rounding_mode>
struct FloatRoundingComputation<Float32, rounding_mode, NegativeScale>
: public BaseFloatRoundingComputation<Float32>
{
static inline void prepare(size_t scale, Scale & mm_scale)
{
Float32 fscale = static_cast<Float32>(scale);
mm_scale = _mm_load1_ps(&fscale);
}
static inline void compute(const Float32 * __restrict in, const Scale & scale, Float32 * __restrict out)
{
__m128 val = _mm_loadu_ps(in);
/// Превратить отрицательные значения в положительные.
__m128 factor = _mm_cmpge_ps(val, getZero());
factor = _mm_min_ps(factor, getTwo());
factor = _mm_sub_ps(factor, getOne());
val = _mm_mul_ps(val, factor);
/// Алгоритм округления.
val = _mm_div_ps(val, scale);
__m128 res = _mm_cmpge_ps(val, getOneTenth());
val = _mm_round_ps(val, rounding_mode);
val = _mm_mul_ps(val, scale);
val = _mm_and_ps(val, res);
/// Предотвратить появление отрицательных нолей определённых в стандарте IEEE-754.
__m128 check = _mm_cmpeq_ps(val, getZero());
check = _mm_min_ps(check, getOne());
factor = _mm_add_ps(factor, check);
/// Вернуть настоящие знаки всех значений.
val = _mm_mul_ps(val, factor);
_mm_storeu_ps(out, val);
}
private:
static inline const __m128 & getOneTenth()
{
static const __m128 one_tenth = _mm_set1_ps(0.1);
return one_tenth;
} }
static inline const __m128 & getZero() static inline const __m128 & getZero()
@ -364,85 +297,23 @@ namespace DB
} }
}; };
template<int rounding_mode> template<>
struct FloatRoundingComputation<Float32, rounding_mode, ZeroScale> class BaseFloatRoundingComputation<Float64>
: public BaseFloatRoundingComputation<Float32>
{ {
static inline void prepare(size_t scale, Scale & mm_scale) public:
using Scale = __m128d;
static const size_t data_count = 2;
protected:
/// Предотвратить появление отрицательных нолей определённых в стандарте IEEE-754.
static inline void normalize(__m128d & val, const __m128d & mask)
{ {
} __m128d mask1 = _mm_cmpeq_pd(val, getZero());
__m128d mask2 = _mm_and_pd(mask, mask1);
static inline void compute(const Float32 * __restrict in, const Scale & scale, Float32 * __restrict out) mask2 = _mm_cmpeq_pd(mask2, getZero());
{ mask2 = _mm_min_pd(mask2, getTwo());
__m128 val = _mm_loadu_ps(in); mask2 = _mm_sub_pd(mask2, getOne());
val = _mm_round_ps(val, rounding_mode); val = _mm_mul_pd(val, mask2);
_mm_storeu_ps(out, val);
}
};
template<int rounding_mode>
struct FloatRoundingComputation<Float64, rounding_mode, PositiveScale>
: public BaseFloatRoundingComputation<Float64>
{
static inline void prepare(size_t scale, Scale & mm_scale)
{
Float64 fscale = static_cast<Float64>(scale);
mm_scale = _mm_load1_pd(&fscale);
}
static inline void compute(const Float64 * __restrict in, const Scale & scale, Float64 * __restrict out)
{
__m128d val = _mm_loadu_pd(in);
val = _mm_mul_pd(val, scale);
val = _mm_round_pd(val, rounding_mode);
val = _mm_div_pd(val, scale);
_mm_storeu_pd(out, val);
}
};
template<int rounding_mode>
struct FloatRoundingComputation<Float64, rounding_mode, NegativeScale>
: public BaseFloatRoundingComputation<Float64>
{
static inline void prepare(size_t scale, Scale & mm_scale)
{
Float64 fscale = static_cast<Float64>(scale);
mm_scale = _mm_load1_pd(&fscale);
}
static inline void compute(const Float64 * __restrict in, const Scale & scale, Float64 * __restrict out)
{
__m128d val = _mm_loadu_pd(in);
/// Превратить отрицательные значения в положительные.
__m128d factor = _mm_cmpge_pd(val, getZero());
factor = _mm_min_pd(factor, getTwo());
factor = _mm_sub_pd(factor, getOne());
val = _mm_mul_pd(val, factor);
/// Алгоритм округления.
val = _mm_div_pd(val, scale);
__m128d res = _mm_cmpge_pd(val, getOneTenth());
val = _mm_round_pd(val, rounding_mode);
val = _mm_mul_pd(val, scale);
val = _mm_and_pd(val, res);
/// Предотвратить появление отрицательных нолей определённых в стандарте IEEE-754.
__m128d check = _mm_cmpeq_pd(val, getZero());
check = _mm_min_pd(check, getOne());
factor = _mm_add_pd(factor, check);
/// Вернуть настоящие знаки всех значений.
val = _mm_mul_pd(val, factor);
_mm_storeu_pd(out, val);
}
private:
static inline const __m128d & getOneTenth()
{
static const __m128d one_tenth = _mm_set1_pd(0.1);
return one_tenth;
} }
static inline const __m128d & getZero() static inline const __m128d & getZero()
@ -464,10 +335,177 @@ namespace DB
} }
}; };
/** Реализация низкоуровневых функций округления для значений с плавающей точкой.
*/
template<typename T, int rounding_mode, ScaleMode scale_mode>
class FloatRoundingComputation;
template<int rounding_mode> template<int rounding_mode>
struct FloatRoundingComputation<Float64, rounding_mode, ZeroScale> class FloatRoundingComputation<Float32, rounding_mode, PositiveScale>
: public BaseFloatRoundingComputation<Float32>
{
public:
static inline void prepare(size_t scale, Scale & mm_scale)
{
Float32 fscale = static_cast<Float32>(scale);
mm_scale = _mm_load1_ps(&fscale);
}
static inline void compute(const Float32 * __restrict in, const Scale & scale, Float32 * __restrict out)
{
__m128 val = _mm_loadu_ps(in);
__m128 mask = _mm_cmplt_ps(val, getZero());
/// Алгоритм округления.
val = _mm_mul_ps(val, scale);
val = _mm_round_ps(val, rounding_mode);
val = _mm_div_ps(val, scale);
normalize(val, mask);
_mm_storeu_ps(out, val);
}
};
template<int rounding_mode>
class FloatRoundingComputation<Float32, rounding_mode, NegativeScale>
: public BaseFloatRoundingComputation<Float32>
{
public:
static inline void prepare(size_t scale, Scale & mm_scale)
{
Float32 fscale = static_cast<Float32>(scale);
mm_scale = _mm_load1_ps(&fscale);
}
static inline void compute(const Float32 * __restrict in, const Scale & scale, Float32 * __restrict out)
{
__m128 val = _mm_loadu_ps(in);
__m128 mask = _mm_cmplt_ps(val, getZero());
/// Превратить отрицательные значения в положительные.
__m128 factor = _mm_cmpge_ps(val, getZero());
factor = _mm_min_ps(factor, getTwo());
factor = _mm_sub_ps(factor, getOne());
val = _mm_mul_ps(val, factor);
/// Алгоритм округления.
val = _mm_div_ps(val, scale);
__m128 res = _mm_cmpge_ps(val, getOneTenth());
val = _mm_round_ps(val, rounding_mode);
val = _mm_mul_ps(val, scale);
val = _mm_and_ps(val, res);
/// Вернуть настоящие знаки всех значений.
val = _mm_mul_ps(val, factor);
normalize(val, mask);
_mm_storeu_ps(out, val);
}
private:
static inline const __m128 & getOneTenth()
{
static const __m128 one_tenth = _mm_set1_ps(0.1);
return one_tenth;
}
};
template<int rounding_mode>
class FloatRoundingComputation<Float32, rounding_mode, ZeroScale>
: public BaseFloatRoundingComputation<Float32>
{
public:
static inline void prepare(size_t scale, Scale & mm_scale)
{
}
static inline void compute(const Float32 * __restrict in, const Scale & scale, Float32 * __restrict out)
{
__m128 val = _mm_loadu_ps(in);
__m128 mask = _mm_cmplt_ps(val, getZero());
val = _mm_round_ps(val, rounding_mode);
normalize(val, mask);
_mm_storeu_ps(out, val);
}
};
template<int rounding_mode>
class FloatRoundingComputation<Float64, rounding_mode, PositiveScale>
: public BaseFloatRoundingComputation<Float64> : public BaseFloatRoundingComputation<Float64>
{ {
public:
static inline void prepare(size_t scale, Scale & mm_scale)
{
Float64 fscale = static_cast<Float64>(scale);
mm_scale = _mm_load1_pd(&fscale);
}
static inline void compute(const Float64 * __restrict in, const Scale & scale, Float64 * __restrict out)
{
__m128d val = _mm_loadu_pd(in);
__m128d mask = _mm_cmplt_pd(val, getZero());
/// Алгоритм округления.
val = _mm_mul_pd(val, scale);
val = _mm_round_pd(val, rounding_mode);
val = _mm_div_pd(val, scale);
normalize(val, mask);
_mm_storeu_pd(out, val);
}
};
template<int rounding_mode>
class FloatRoundingComputation<Float64, rounding_mode, NegativeScale>
: public BaseFloatRoundingComputation<Float64>
{
public:
static inline void prepare(size_t scale, Scale & mm_scale)
{
Float64 fscale = static_cast<Float64>(scale);
mm_scale = _mm_load1_pd(&fscale);
}
static inline void compute(const Float64 * __restrict in, const Scale & scale, Float64 * __restrict out)
{
__m128d val = _mm_loadu_pd(in);
__m128d mask = _mm_cmplt_pd(val, getZero());
/// Превратить отрицательные значения в положительные.
__m128d factor = _mm_cmpge_pd(val, getZero());
factor = _mm_min_pd(factor, getTwo());
factor = _mm_sub_pd(factor, getOne());
val = _mm_mul_pd(val, factor);
/// Алгоритм округления.
val = _mm_div_pd(val, scale);
__m128d res = _mm_cmpge_pd(val, getOneTenth());
val = _mm_round_pd(val, rounding_mode);
val = _mm_mul_pd(val, scale);
val = _mm_and_pd(val, res);
/// Вернуть настоящие знаки всех значений.
val = _mm_mul_pd(val, factor);
normalize(val, mask);
_mm_storeu_pd(out, val);
}
private:
static inline const __m128d & getOneTenth()
{
static const __m128d one_tenth = _mm_set1_pd(0.1);
return one_tenth;
}
};
template<int rounding_mode>
class FloatRoundingComputation<Float64, rounding_mode, ZeroScale>
: public BaseFloatRoundingComputation<Float64>
{
public:
static inline void prepare(size_t scale, Scale & mm_scale) static inline void prepare(size_t scale, Scale & mm_scale)
{ {
} }
@ -475,7 +513,11 @@ namespace DB
static inline void compute(const Float64 * __restrict in, const Scale & scale, Float64 * __restrict 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);
__m128d mask = _mm_cmplt_pd(val, getZero());
val = _mm_round_pd(val, rounding_mode); val = _mm_round_pd(val, rounding_mode);
normalize(val, mask);
_mm_storeu_pd(out, val); _mm_storeu_pd(out, val);
} }
}; };

View File

@ -852,3 +852,6 @@
2.738 2.738
2.734 2.734
2.73 2.73
0
0
0

View File

@ -947,3 +947,8 @@ SELECT round(y,3) FROM (SELECT 2.718281828459045 + 1/(1+x*x) AS y ARRAY JOIN ran
SELECT round(y,3) FROM (SELECT 2.718281828459045 + 1/(1+x*x) AS y ARRAY JOIN range(9) AS x); SELECT round(y,3) FROM (SELECT 2.718281828459045 + 1/(1+x*x) AS y ARRAY JOIN range(9) AS x);
SELECT round(y,3) FROM (SELECT 2.718281828459045 + 1/(1+x*x) AS y ARRAY JOIN range(10) AS x); SELECT round(y,3) FROM (SELECT 2.718281828459045 + 1/(1+x*x) AS y ARRAY JOIN range(10) AS x);
/* Negative zeroes. */
SELECT round(-0.002);
SELECT round(-0.002, -1);
SELECT round(-0.002, 1);