将 __m256i 的前 N 位或后 N 位设置为 1,其余设置为 0 的有效方法
Efficient way to set first N or last N bits of __m256i to 1, the rest to 0
如何使用 AVX2
有效地设置为 1
- 前
N
位
- 最后
N
位
of __m256i
,将其余设置为 0
?
当范围可能在 __m256i
值的中间开始和结束时,这是针对位范围的尾部和头部的 2 个单独操作。范围中占据完整 __m256i
值的部分使用所有-0
或所有-1
掩码进行处理。
AVX2 移位指令 vpsllvd
和 vpsrlvd
具有很好的 属性 移位计数
大于或等于 32 导致 ymm 寄存器中的整数为零。
换句话说:相比之下,班次计数没有被掩盖
到 x86 标量移位指令的移位计数。
因此代码相当简单:
/*
gcc -O3 -m64 -Wall -mavx2 -march=broadwell avx2_bit_mask.c
*/
#include <immintrin.h>
#include <stdio.h>
__m256i bit_mask_avx2_msb(unsigned int n)
{
__m256i ones = _mm256_set1_epi32(-1);
__m256i cnst32_256 = _mm256_set_epi32(32,64,96,128, 160,192,224,256);
__m256i shift = _mm256_set1_epi32(n);
shift = _mm256_subs_epu16(cnst32_256,shift);
return _mm256_sllv_epi32(ones,shift);
}
__m256i bit_mask_avx2_lsb(unsigned int n)
{
__m256i ones = _mm256_set1_epi32(-1);
__m256i cnst32_256 = _mm256_set_epi32(256,224,192,160, 128,96,64,32);
__m256i shift = _mm256_set1_epi32(n);
shift = _mm256_subs_epu16(cnst32_256,shift);
return _mm256_srlv_epi32(ones,shift);
}
int print_avx2_hex(__m256i ymm)
{
long unsigned int x[4];
_mm256_storeu_si256((__m256i*)x,ymm);
printf("%016lX %016lX %016lX %016lX\n", x[3],x[2],x[1],x[0]);
return 0;
}
int main()
{
unsigned int i;
for (i=0;i<259;i++){
printf("bit_mask_avx2_lsb(%3d) ",i);
print_avx2_hex(bit_mask_avx2_lsb(i));
}
printf("\n");
for (i=0;i<259;i++){
printf("bit_mask_avx2_msb(%3d) ",i);
print_avx2_hex(bit_mask_avx2_msb(i));
}
printf("\n");
return 0;
}
结果是:
$ ./a.out
bit_mask_avx2_lsb( 0) 0000000000000000 0000000000000000 0000000000000000 0000000000000000
bit_mask_avx2_lsb( 1) 0000000000000000 0000000000000000 0000000000000000 0000000000000001
bit_mask_avx2_lsb( 2) 0000000000000000 0000000000000000 0000000000000000 0000000000000003
bit_mask_avx2_lsb( 3) 0000000000000000 0000000000000000 0000000000000000 0000000000000007
bit_mask_avx2_lsb( 4) 0000000000000000 0000000000000000 0000000000000000 000000000000000F
bit_mask_avx2_lsb( 5) 0000000000000000 0000000000000000 0000000000000000 000000000000001F
bit_mask_avx2_lsb( 6) 0000000000000000 0000000000000000 0000000000000000 000000000000003F
bit_mask_avx2_lsb( 7) 0000000000000000 0000000000000000 0000000000000000 000000000000007F
bit_mask_avx2_lsb( 8) 0000000000000000 0000000000000000 0000000000000000 00000000000000FF
bit_mask_avx2_lsb( 9) 0000000000000000 0000000000000000 0000000000000000 00000000000001FF
bit_mask_avx2_lsb( 10) 0000000000000000 0000000000000000 0000000000000000 00000000000003FF
bit_mask_avx2_lsb( 11) 0000000000000000 0000000000000000 0000000000000000 00000000000007FF
...
bit_mask_avx2_lsb(124) 0000000000000000 0000000000000000 0FFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(125) 0000000000000000 0000000000000000 1FFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(126) 0000000000000000 0000000000000000 3FFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(127) 0000000000000000 0000000000000000 7FFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(128) 0000000000000000 0000000000000000 FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(129) 0000000000000000 0000000000000001 FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(130) 0000000000000000 0000000000000003 FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(131) 0000000000000000 0000000000000007 FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(132) 0000000000000000 000000000000000F FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
...
bit_mask_avx2_lsb(248) 00FFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(249) 01FFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(250) 03FFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(251) 07FFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(252) 0FFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(253) 1FFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(254) 3FFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(255) 7FFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(256) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(257) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(258) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_msb( 0) 0000000000000000 0000000000000000 0000000000000000 0000000000000000
bit_mask_avx2_msb( 1) 8000000000000000 0000000000000000 0000000000000000 0000000000000000
bit_mask_avx2_msb( 2) C000000000000000 0000000000000000 0000000000000000 0000000000000000
bit_mask_avx2_msb( 3) E000000000000000 0000000000000000 0000000000000000 0000000000000000
bit_mask_avx2_msb( 4) F000000000000000 0000000000000000 0000000000000000 0000000000000000
bit_mask_avx2_msb( 5) F800000000000000 0000000000000000 0000000000000000 0000000000000000
bit_mask_avx2_msb( 6) FC00000000000000 0000000000000000 0000000000000000 0000000000000000
bit_mask_avx2_msb( 7) FE00000000000000 0000000000000000 0000000000000000 0000000000000000
bit_mask_avx2_msb( 8) FF00000000000000 0000000000000000 0000000000000000 0000000000000000
bit_mask_avx2_msb( 9) FF80000000000000 0000000000000000 0000000000000000 0000000000000000
bit_mask_avx2_msb( 10) FFC0000000000000 0000000000000000 0000000000000000 0000000000000000
bit_mask_avx2_msb( 11) FFE0000000000000 0000000000000000 0000000000000000 0000000000000000
...
bit_mask_avx2_msb(124) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFF0 0000000000000000 0000000000000000
bit_mask_avx2_msb(125) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFF8 0000000000000000 0000000000000000
bit_mask_avx2_msb(126) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFC 0000000000000000 0000000000000000
bit_mask_avx2_msb(127) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFE 0000000000000000 0000000000000000
bit_mask_avx2_msb(128) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF 0000000000000000 0000000000000000
bit_mask_avx2_msb(129) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF 8000000000000000 0000000000000000
bit_mask_avx2_msb(130) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF C000000000000000 0000000000000000
bit_mask_avx2_msb(131) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF E000000000000000 0000000000000000
bit_mask_avx2_msb(132) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF F000000000000000 0000000000000000
...
bit_mask_avx2_msb(248) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFF00
bit_mask_avx2_msb(249) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFF80
bit_mask_avx2_msb(250) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFC0
bit_mask_avx2_msb(251) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFE0
bit_mask_avx2_msb(252) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFF0
bit_mask_avx2_msb(253) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFF8
bit_mask_avx2_msb(254) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFC
bit_mask_avx2_msb(255) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFE
bit_mask_avx2_msb(256) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_msb(257) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_msb(258) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
对于值 n
,256<=n
<=65535,所有位都设置为 1,正如人们所期望的那样。
65535的上限是由于_mm256_subs_epu16()
的16位饱和运算。
n
=65536 位掩码(输出值)为零。可以修改代码,使所有位都设置为 1
对于 256<=n
<=INT_MAX
的范围。
这可以通过替换来实现
shift = _mm256_subs_epu16(cnst32_256,shift);
和
__m256i mask = _mm256_cmpgt_epi32(cnst32_256,shift);
shift = _mm256_sub_epi32(cnst32_256,shift);
shift = _mm256_and_si256(shift,mask);
这三个内在函数或多或少地模拟了 _mm256_subs_epu32(cnst32_256,shift)
,后者并不存在。
如何使用 AVX2
有效地设置为1
- 前
N
位 - 最后
N
位
of __m256i
,将其余设置为 0
?
当范围可能在 __m256i
值的中间开始和结束时,这是针对位范围的尾部和头部的 2 个单独操作。范围中占据完整 __m256i
值的部分使用所有-0
或所有-1
掩码进行处理。
AVX2 移位指令 vpsllvd
和 vpsrlvd
具有很好的 属性 移位计数
大于或等于 32 导致 ymm 寄存器中的整数为零。
换句话说:相比之下,班次计数没有被掩盖
到 x86 标量移位指令的移位计数。
因此代码相当简单:
/*
gcc -O3 -m64 -Wall -mavx2 -march=broadwell avx2_bit_mask.c
*/
#include <immintrin.h>
#include <stdio.h>
__m256i bit_mask_avx2_msb(unsigned int n)
{
__m256i ones = _mm256_set1_epi32(-1);
__m256i cnst32_256 = _mm256_set_epi32(32,64,96,128, 160,192,224,256);
__m256i shift = _mm256_set1_epi32(n);
shift = _mm256_subs_epu16(cnst32_256,shift);
return _mm256_sllv_epi32(ones,shift);
}
__m256i bit_mask_avx2_lsb(unsigned int n)
{
__m256i ones = _mm256_set1_epi32(-1);
__m256i cnst32_256 = _mm256_set_epi32(256,224,192,160, 128,96,64,32);
__m256i shift = _mm256_set1_epi32(n);
shift = _mm256_subs_epu16(cnst32_256,shift);
return _mm256_srlv_epi32(ones,shift);
}
int print_avx2_hex(__m256i ymm)
{
long unsigned int x[4];
_mm256_storeu_si256((__m256i*)x,ymm);
printf("%016lX %016lX %016lX %016lX\n", x[3],x[2],x[1],x[0]);
return 0;
}
int main()
{
unsigned int i;
for (i=0;i<259;i++){
printf("bit_mask_avx2_lsb(%3d) ",i);
print_avx2_hex(bit_mask_avx2_lsb(i));
}
printf("\n");
for (i=0;i<259;i++){
printf("bit_mask_avx2_msb(%3d) ",i);
print_avx2_hex(bit_mask_avx2_msb(i));
}
printf("\n");
return 0;
}
结果是:
$ ./a.out
bit_mask_avx2_lsb( 0) 0000000000000000 0000000000000000 0000000000000000 0000000000000000
bit_mask_avx2_lsb( 1) 0000000000000000 0000000000000000 0000000000000000 0000000000000001
bit_mask_avx2_lsb( 2) 0000000000000000 0000000000000000 0000000000000000 0000000000000003
bit_mask_avx2_lsb( 3) 0000000000000000 0000000000000000 0000000000000000 0000000000000007
bit_mask_avx2_lsb( 4) 0000000000000000 0000000000000000 0000000000000000 000000000000000F
bit_mask_avx2_lsb( 5) 0000000000000000 0000000000000000 0000000000000000 000000000000001F
bit_mask_avx2_lsb( 6) 0000000000000000 0000000000000000 0000000000000000 000000000000003F
bit_mask_avx2_lsb( 7) 0000000000000000 0000000000000000 0000000000000000 000000000000007F
bit_mask_avx2_lsb( 8) 0000000000000000 0000000000000000 0000000000000000 00000000000000FF
bit_mask_avx2_lsb( 9) 0000000000000000 0000000000000000 0000000000000000 00000000000001FF
bit_mask_avx2_lsb( 10) 0000000000000000 0000000000000000 0000000000000000 00000000000003FF
bit_mask_avx2_lsb( 11) 0000000000000000 0000000000000000 0000000000000000 00000000000007FF
...
bit_mask_avx2_lsb(124) 0000000000000000 0000000000000000 0FFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(125) 0000000000000000 0000000000000000 1FFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(126) 0000000000000000 0000000000000000 3FFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(127) 0000000000000000 0000000000000000 7FFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(128) 0000000000000000 0000000000000000 FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(129) 0000000000000000 0000000000000001 FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(130) 0000000000000000 0000000000000003 FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(131) 0000000000000000 0000000000000007 FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(132) 0000000000000000 000000000000000F FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
...
bit_mask_avx2_lsb(248) 00FFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(249) 01FFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(250) 03FFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(251) 07FFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(252) 0FFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(253) 1FFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(254) 3FFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(255) 7FFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(256) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(257) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_lsb(258) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_msb( 0) 0000000000000000 0000000000000000 0000000000000000 0000000000000000
bit_mask_avx2_msb( 1) 8000000000000000 0000000000000000 0000000000000000 0000000000000000
bit_mask_avx2_msb( 2) C000000000000000 0000000000000000 0000000000000000 0000000000000000
bit_mask_avx2_msb( 3) E000000000000000 0000000000000000 0000000000000000 0000000000000000
bit_mask_avx2_msb( 4) F000000000000000 0000000000000000 0000000000000000 0000000000000000
bit_mask_avx2_msb( 5) F800000000000000 0000000000000000 0000000000000000 0000000000000000
bit_mask_avx2_msb( 6) FC00000000000000 0000000000000000 0000000000000000 0000000000000000
bit_mask_avx2_msb( 7) FE00000000000000 0000000000000000 0000000000000000 0000000000000000
bit_mask_avx2_msb( 8) FF00000000000000 0000000000000000 0000000000000000 0000000000000000
bit_mask_avx2_msb( 9) FF80000000000000 0000000000000000 0000000000000000 0000000000000000
bit_mask_avx2_msb( 10) FFC0000000000000 0000000000000000 0000000000000000 0000000000000000
bit_mask_avx2_msb( 11) FFE0000000000000 0000000000000000 0000000000000000 0000000000000000
...
bit_mask_avx2_msb(124) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFF0 0000000000000000 0000000000000000
bit_mask_avx2_msb(125) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFF8 0000000000000000 0000000000000000
bit_mask_avx2_msb(126) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFC 0000000000000000 0000000000000000
bit_mask_avx2_msb(127) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFE 0000000000000000 0000000000000000
bit_mask_avx2_msb(128) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF 0000000000000000 0000000000000000
bit_mask_avx2_msb(129) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF 8000000000000000 0000000000000000
bit_mask_avx2_msb(130) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF C000000000000000 0000000000000000
bit_mask_avx2_msb(131) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF E000000000000000 0000000000000000
bit_mask_avx2_msb(132) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF F000000000000000 0000000000000000
...
bit_mask_avx2_msb(248) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFF00
bit_mask_avx2_msb(249) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFF80
bit_mask_avx2_msb(250) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFC0
bit_mask_avx2_msb(251) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFE0
bit_mask_avx2_msb(252) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFF0
bit_mask_avx2_msb(253) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFF8
bit_mask_avx2_msb(254) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFC
bit_mask_avx2_msb(255) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFE
bit_mask_avx2_msb(256) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_msb(257) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
bit_mask_avx2_msb(258) FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF FFFFFFFFFFFFFFFF
对于值 n
,256<=n
<=65535,所有位都设置为 1,正如人们所期望的那样。
65535的上限是由于_mm256_subs_epu16()
的16位饱和运算。
n
=65536 位掩码(输出值)为零。可以修改代码,使所有位都设置为 1
对于 256<=n
<=INT_MAX
的范围。
这可以通过替换来实现
shift = _mm256_subs_epu16(cnst32_256,shift);
和
__m256i mask = _mm256_cmpgt_epi32(cnst32_256,shift);
shift = _mm256_sub_epi32(cnst32_256,shift);
shift = _mm256_and_si256(shift,mask);
这三个内在函数或多或少地模拟了 _mm256_subs_epu32(cnst32_256,shift)
,后者并不存在。