函数s = get_scale(z)
计算"最接近2的幂次方"。由于s
的小数位为零,因此s
的倒数只需执行一个(廉价的)整数减法:参见函数inv_of_scale
。
在x86上,get_scale
和inv_of_scale
编译成相当高效的汇编代码与clang。编译器clang将三元运算符转换为minsd
和maxsd
,另请参阅Peter Cordes的评论。
对于gcc而言,将这些函数转换为x86内置代码(get_scale_x86
和inv_of_scale_x86
)稍微更有效一些,具体请看Godbolt。
注意,C明确允许通过联合类型(wording)进行类型强制转换,但C++(c++11)不允许。虽然gcc8.2和clang7.0不会抱怨联合,但你可以通过使用memcpy
技巧而不是联合技巧来提高C++的移植性。修改代码应该很简单。代码应正确处理子规范数。
#include<stdio.h>
#include<stdint.h>
#include<immintrin.h>
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;
x_min.i = 0x0010000000000000ull;
x_max.i = 0x7FD0000000000000ull;
mask_i = 0x7FF0000000000000ull;
x.d = t;
x.i = x.i & mask_i;
x.d = (x.d < x_min.d) ? x_min.d : x.d;
x.d = (x.d > x_max.d) ? x_max.d : x.d;
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);
}
double inv_of_scale(double t){
union dbl_int64 x;
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;
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;
}
输出结果看起来很好:
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
向量化
如果编译器支持自动向量化,则函数get_scale
应该进行向量化。以下代码片段在clang下向量化效果很好(无需编写SSE/AVX指令代码)。
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]);
}
}
很不幸,gcc找不到vmaxpd
和vminpd
指令。
_mm512_getexp_pd
(提取指数作为double
)和_mm512_scalef_pd
(https://software.intel.com/sites/landingpage/IntrinsicsGuide/#expand=2403,6062,4147,4841,4841&techs=SSE2,SSE4_2,AVX,AVX2,AVX_512,Other&text=_mm512_scalef_pd),它执行dst [63:0]:= tmp_src1 [63:0] * POW(2,FLOOR(tmp_src2 [63:0]))
(即将双精度浮点数的整数部分加到另一个浮点数的指数域中)? - Peter Cordesminsd
、minpd
、maxsd
、maxpd
,一些方面上clang比gcc表现更好。Godbolt链接 - wim