gez*_*eza 6 c++ floating-point x86 ieee-754
在数值计算中,通常需要将数字缩放到安全范围内.
例如,计算欧几里德距离:sqrt(a^2+b^2).在这里,如果幅度a还是b过小/大,则溢/上溢可能发生.
解决此问题的常用方法是将数字除以最大幅度数.但是,这个解决方案是:
所以我认为不是除以最大幅度数,而是将它乘以一个接近2次幂的倒数.这似乎是一个更好的解决方案,如:
所以,我想创建一个小实用程序函数,它有一个像这样的逻辑(通过^,我的意思是取幂):
void getScaler(double value, double &scaler, double &scalerReciprocal) {
int e = <exponent of value>;
if (e<-1022) { scaler=2^-1022; scalerReciprocal = 2^1022; }
} else if (e>1022) { scaler=2^1022; scalerReciprocal = 2^-1022; }
} else { scaler=2^e; scalerReciprocal = 2^(2046-e); }
}
这个函数应该返回一个标准化的scaler&scalerReciprocal,两个都是2的幂数,scaler它接近于value,并且scalerReciprocal是它的倒数scaler.
为最大允许指数scaler/ scaleReciprocal是-1022..1022(我不想用低于正常工作scaler,为次正规数可能会很慢).
什么是快速的方法呢？这可以通过纯浮点运算来完成吗？或者我应该从中提取指数value,并使用简单的ifs来执行逻辑？是否有某种技巧与( - )1022快速比较(因为范围是对称的)？
注意:scaler不需要是最接近的2次幂.如果某些逻辑需要它,scaler可以是距离最接近值的2的小功率.
功能s = get_scale(z)计算"2的近距离".由于分数位s
为零,因此倒数s只是一个(廉价)整数减法:见函数inv_of_scale.
在x86上get_scale并inv_of_scale使用clang编译为非常高效的程序集.编译器clang翻译三元运算符,minsd并maxsd参见Peter Cordes的评论.使用gcc,将这些函数转换为x86内在函数代码(get_scale_x86和inv_of_scale_x86)的效率稍高,请参阅Godbolt.
请注意,C显式允许通过联合进行类型惩罚,而C++(c ++ 11)没有这样的权限
虽然gcc 8.2和clang 7.0不会抱怨联合,但是你可以通过使用memcpy技巧而不是使用技巧来改进C++ .联盟技巧.代码的这种修改应该是微不足道的.代码应该正确处理子正规.
#include<stdio.h>
#include<stdint.h>
#include<immintrin.h>
/* gcc -Wall -m64 -O3 -march=sandybridge dbl_scale.c */
union dbl_int64{
double d;
uint64_t i;
};
double get_scale(double t){
union dbl_int64 x;
union dbl_int64 x_min;
union dbl_int64 x_max;
uint64_t mask_i;
/* 0xFEDCBA9876543210 */
x_min.i = 0x0010000000000000ull;
x_max.i = 0x7FD0000000000000ull;
mask_i = 0x7FF0000000000000ull;
x.d = t;
x.i = x.i & mask_i; /* Set fraction bits to zero, take absolute value */
x.d = (x.d < x_min.d) ? x_min.d : x.d; /* If subnormal: set exponent to 1 */
x.d = (x.d > x_max.d) ? x_max.d : x.d; /* If exponent is very large: set exponent to 7FD, otherwise the inverse is a subnormal */
return x.d;
}
double get_scale_x86(double t){
__m128d x = _mm_set_sd(t);
__m128d x_min = _mm_castsi128_pd(_mm_set1_epi64x(0x0010000000000000ull));
__m128d x_max = _mm_castsi128_pd(_mm_set1_epi64x(0x7FD0000000000000ull));
__m128d mask = _mm_castsi128_pd(_mm_set1_epi64x(0x7FF0000000000000ull));
x = _mm_and_pd(x, mask);
x = _mm_max_sd(x, x_min);
x = _mm_min_sd(x, x_max);
return _mm_cvtsd_f64(x);
}
/* Compute the inverse 1/t of a double t with all zero fraction bits */
/* and exponent between the limits of function get_scale */
/* A single integer subtraction is much less expensive than a */
/* floating point division. */
double inv_of_scale(double t){
union dbl_int64 x;
/* 0xFEDCBA9876543210 */
uint64_t inv_mask = 0x7FE0000000000000ull;
x.d = t;
x.i = inv_mask - x.i;
return x.d;
}
double inv_of_scale_x86(double t){
__m128i inv_mask = _mm_set1_epi64x(0x7FE0000000000000ull);
__m128d x = _mm_set_sd(t);
__m128i x_i = _mm_sub_epi64(inv_mask, _mm_castpd_si128(x));
return _mm_cvtsd_f64(_mm_castsi128_pd(x_i));
}
int main(){
int n = 14;
int i;
/* Several example values, 4.94e-324 is the smallest subnormal */
double y[14] = { 4.94e-324, 1.1e-320, 1.1e-300, 1.1e-5, 0.7, 1.7, 123.1, 1.1e300,
1.79e308, -1.1e-320, -0.7, -1.7, -123.1, -1.1e307};
double z, s, u;
printf("Portable code:\n");
printf(" x pow_of_2 inverse pow2*inv x*inverse \n");
for (i = 0; i < n; i++){
z = y[i];
s = get_scale(z);
u = inv_of_scale(s);
printf("%14e %14e %14e %14e %14e\n", z, s, u, s*u, z*u);
}
printf("\nx86 specific SSE code:\n");
printf(" x pow_of_2 inverse pow2*inv x*inverse \n");
for (i = 0; i < n; i++){
z = y[i];
s = get_scale_x86(z);
u = inv_of_scale_x86(s);
printf("%14e %14e %14e %14e %14e\n", z, s, u, s*u, z*u);
}
return 0;
}
输出看起来很好:
Portable code:
x pow_of_2 inverse pow2*inv x*inverse
4.940656e-324 2.225074e-308 4.494233e+307 1.000000e+00 2.220446e-16
1.099790e-320 2.225074e-308 4.494233e+307 1.000000e+00 4.942713e-13
1.100000e-300 7.466109e-301 1.339386e+300 1.000000e+00 1.473324e+00
1.100000e-05 7.629395e-06 1.310720e+05 1.000000e+00 1.441792e+00
7.000000e-01 5.000000e-01 2.000000e+00 1.000000e+00 1.400000e+00
1.700000e+00 1.000000e+00 1.000000e+00 1.000000e+00 1.700000e+00
1.231000e+02 6.400000e+01 1.562500e-02 1.000000e+00 1.923437e+00
1.100000e+300 6.696929e+299 1.493222e-300 1.000000e+00 1.642544e+00
1.790000e+308 4.494233e+307 2.225074e-308 1.000000e+00 3.982882e+00
-1.099790e-320 2.225074e-308 4.494233e+307 1.000000e+00 -4.942713e-13
-7.000000e-01 5.000000e-01 2.000000e+00 1.000000e+00 -1.400000e+00
-1.700000e+00 1.000000e+00 1.000000e+00 1.000000e+00 -1.700000e+00
-1.231000e+02 6.400000e+01 1.562500e-02 1.000000e+00 -1.923437e+00
-1.100000e+307 5.617791e+306 1.780059e-307 1.000000e+00 -1.958065e+00
x86 specific SSE code:
x pow_of_2 inverse pow2*inv x*inverse
4.940656e-324 2.225074e-308 4.494233e+307 1.000000e+00 2.220446e-16
1.099790e-320 2.225074e-308 4.494233e+307 1.000000e+00 4.942713e-13
1.100000e-300 7.466109e-301 1.339386e+300 1.000000e+00 1.473324e+00
1.100000e-05 7.629395e-06 1.310720e+05 1.000000e+00 1.441792e+00
7.000000e-01 5.000000e-01 2.000000e+00 1.000000e+00 1.400000e+00
1.700000e+00 1.000000e+00 1.000000e+00 1.000000e+00 1.700000e+00
1.231000e+02 6.400000e+01 1.562500e-02 1.000000e+00 1.923437e+00
1.100000e+300 6.696929e+299 1.493222e-300 1.000000e+00 1.642544e+00
1.790000e+308 4.494233e+307 2.225074e-308 1.000000e+00 3.982882e+00
-1.099790e-320 2.225074e-308 4.494233e+307 1.000000e+00 -4.942713e-13
-7.000000e-01 5.000000e-01 2.000000e+00 1.000000e+00 -1.400000e+00
-1.700000e+00 1.000000e+00 1.000000e+00 1.000000e+00 -1.700000e+00
-1.231000e+02 6.400000e+01 1.562500e-02 1.000000e+00 -1.923437e+00
-1.100000e+307 5.617791e+306 1.780059e-307 1.000000e+00 -1.958065e+00
矢量
函数get_scale应该使用支持自动向量化的编译器进行向量化.下面的代码片段与clang很好地矢量化(不需要编写SSE/AVX内在函数代码).
/* Test how well get_scale vectorizes: */
void get_scale_vec(double * __restrict__ t, double * __restrict__ x){
int n = 1024;
int i;
for (i = 0; i < n; i++){
x[i] = get_scale(t[i]);
}
}
不幸的是,gcc找不到vmaxpd和vminpd说明.
根据 wim 的回答，这是另一种解决方案，它可以更快，因为它少了一条指令。输出有点不同，但仍然满足要求。
这个想法是使用位运算来修复边界情况：将 a01放入指数的 lsb，无论其值是多少。所以，指数：
00）01，或减半（当 10->01 时）或 1/4（当 11->01 时）所以，这个修改后的例程是有效的（而且我认为只用2 个快速 asm 指令就可以解决这个问题，这是非常酷的）：
#include<stdio.h>
#include<stdint.h>
#include<immintrin.h>
/* gcc -Wall -m64 -O3 -march=sandybridge dbl_scale.c */
union dbl_int64{
double d;
uint64_t i;
};
double get_scale(double t){
union dbl_int64 x;
uint64_t and_i;
uint64_t or_i;
/* 0xFEDCBA9876543210 */
and_i = 0x7FD0000000000000ull;
or_i = 0x0010000000000000ull;
x.d = t;
x.i = (x.i & and_i)|or_i; /* Set fraction bits to zero, take absolute value */
return x.d;
}
double get_scale_x86(double t){
__m128d x = _mm_set_sd(t);
__m128d x_and = _mm_castsi128_pd(_mm_set1_epi64x(0x7FD0000000000000ull));
__m128d x_or = _mm_castsi128_pd(_mm_set1_epi64x(0x0010000000000000ull));
x = _mm_and_pd(x, x_and);
x = _mm_or_pd(x, x_or);
return _mm_cvtsd_f64(x);
}
/* Compute the inverse 1/t of a double t with all zero fraction bits */
/* and exponent between the limits of function get_scale */
/* A single integer subtraction is much less expensive than a */
/* floating point division. */
double inv_of_scale(double t){
union dbl_int64 x;
/* 0xFEDCBA9876543210 */
uint64_t inv_mask = 0x7FE0000000000000ull;
x.d = t;
x.i = inv_mask - x.i;
return x.d;
}
double inv_of_scale_x86(double t){
__m128i inv_mask = _mm_set1_epi64x(0x7FE0000000000000ull);
__m128d x = _mm_set_sd(t);
__m128i x_i = _mm_sub_epi64(inv_mask, _mm_castpd_si128(x));
return _mm_cvtsd_f64(_mm_castsi128_pd(x_i));
}
int main(){
int n = 14;
int i;
/* Several example values, 4.94e-324 is the smallest subnormal */
double y[14] = { 4.94e-324, 1.1e-320, 1.1e-300, 1.1e-5, 0.7, 1.7, 123.1, 1.1e300,
1.79e308, -1.1e-320, -0.7, -1.7, -123.1, -1.1e307};
double z, s, u;
printf("Portable code:\n");
printf(" x pow_of_2 inverse pow2*inv x*inverse \n");
for (i = 0; i < n; i++){
z = y[i];
s = get_scale(z);
u = inv_of_scale(s);
printf("%14e %14e %14e %14e %14e\n", z, s, u, s*u, z*u);
}
printf("\nx86 specific SSE code:\n");
printf(" x pow_of_2 inverse pow2*inv x*inverse \n");
for (i = 0; i < n; i++){
z = y[i];
s = get_scale_x86(z);
u = inv_of_scale_x86(s);
printf("%14e %14e %14e %14e %14e\n", z, s, u, s*u, z*u);
}
return 0;
}
| 归档时间： |
|
| 查看次数： |
310 次 |
| 最近记录： |