我有以下代码使用标志-msse4编译GCC，但问题是pop计数只获得转换后的__m128i类型的最后4位8位。基本上我想要的是计算__m128i类型中的所有16个数字，但是我不确定在创建变量popA之后要进行的内部函数调用。不知何故，popA必须转换为包含所有128位信息的整数？我想theres _mm_cvtsi128_si64并且使用了一些shuffle几个操作，但我的操作系统是32位。是否只有shuffle方法并使用_mm_cvtsi128_si32？
编辑：如果shuffle方法是我需要帮助为我的32位操作系统实现它的唯一选项，请。
继承代码。
#include <stdio.h>
#include <smmintrin.h>
#include <emmintrin.h>
int main(void)
{
int A = 1;
__m128i popA = _mm_set_epi8( A, A, A, A, A, A, A, A, A, A, A, A, A, A, A, A);
unsigned int integer = _mm_cvtsi128_si32(popA);
//long long LONG = _mm_cvtsi128_si64(popA);//my OS is 32-bits so no luck here
printf("integer = %d\n", integer);
int pop = _mm_popcnt_u32(integer);
//int popLONG = _mm_popcnt_u64(LONG);
printf("popcount = %d\n", pop);
//printf("popcount LONG = %d\n", popLONG);
return 0;
}
编辑2 ：这个最终运行（使用GCC编译器标志-msse -msse2 -msse3 -msse4）虽然我不确定pop_count1()的输出是否正确。
输出：
pop_count1(): 1799 1799 1799 1799 1799 1799 1799 1799
pop_count2():population count for each byte: 1 1 1 1 1 1 1 1 0 1 2 3 4 5 6 7
#include <stdio.h>
#include <xmmintrin.h>
#include <emmintrin.h>
#include <mmintrin.h>
#include <stdint.h>
#include <tmmintrin.h>
void print128_num(__m128i var)
{
uint16_t *val = (uint16_t*) &var;
printf("pop_count1(): %i %i %i %i %i %i %i %i \n",
val[0], val[1], val[2], val[3], val[4], val[5],
val[6], val[7]);
}
static __m128i parallelPopcnt16bytes (__m128i xmm)//for pop_count2
{
const __m128i mask4 = _mm_set1_epi8 (0x0F);
const __m128i lookup = _mm_setr_epi8 (0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4);
__m128i low, high, count;
low = _mm_and_si128 (mask4, xmm);
high = _mm_and_si128 (mask4, _mm_srli_epi16 (xmm, 4));
count = _mm_add_epi8 (_mm_shuffle_epi8 (lookup, low), _mm_shuffle_epi8 (lookup, high));
return count;
}
void pop_count1()
{
int A = 1;
__m128i in = _mm_set_epi8( A, A, A, A, A, A, A, A, A, A, A, A, A, A, A, A);
__m128i bit0 = _mm_set1_epi8( 0x80 );
__m128i mask0 = _mm_and_si128( in, bit0 );
__m128i sum = _mm_cmpeq_epi8( mask0, _mm_setzero_si128() );
/* general pattern */
__m128i bit1 = _mm_set1_epi8( 0x40 );
__m128i mask1 = _mm_and_si128( in, bit1 );
mask1 = _mm_cmpeq_epi8( mask1, _mm_setzero_si128() );
sum = _mm_add_epi8( sum, mask1 );
/* next bit */
__m128i bit2 = _mm_set1_epi8( 0x20 );
__m128i mask2 = _mm_and_si128( in, bit2 );
mask2 = _mm_cmpeq_epi8( mask2, _mm_setzero_si128() );
sum = _mm_add_epi8( sum, mask2 );
__m128i bit3 = _mm_set1_epi8( 0x10 );
__m128i mask3 = _mm_and_si128( in, bit3 );
mask3 = _mm_cmpeq_epi8( mask3, _mm_setzero_si128() );
sum = _mm_add_epi8( sum, mask3 );
__m128i bit4 = _mm_set1_epi8( 0x08 );
__m128i mask4 = _mm_and_si128( in, bit4 );
mask4 = _mm_cmpeq_epi8( mask4, _mm_setzero_si128() );
sum = _mm_add_epi8( sum, mask4 );
__m128i bit5 = _mm_set1_epi8( 0x04 );
__m128i mask5 = _mm_and_si128( in, bit5 );
mask5 = _mm_cmpeq_epi8( mask5, _mm_setzero_si128() );
sum = _mm_add_epi8( sum, mask5 );
__m128i bit6 = _mm_set1_epi8( 0x02 );
__m128i mask6 = _mm_and_si128( in, bit6 );
mask6 = _mm_cmpeq_epi8( mask6, _mm_setzero_si128() );
sum = _mm_add_epi8( sum, mask6 );
__m128i bit7 = _mm_set1_epi8( 0x01 );
__m128i mask7 = _mm_and_si128( in, bit7 );
mask7 = _mm_cmpeq_epi8( mask7, _mm_setzero_si128() );
sum = _mm_add_epi8( sum, mask7 );
/* finish up */
sum = _mm_sub_epi8( _mm_setzero_si128(), sum );
print128_num(sum);
}
void pop_count2()
{
int index;
__m128i testVector = _mm_set_epi8 (1, 2, 4, 8, 16, 32, 64, 128, 0, 1, 3, 7, 15, 31, 63, 127);
__m128i counts = parallelPopcnt16bytes (testVector);
printf ("pop_count2():population count for each byte:");
for (index = 15; index >= 0; index--)
{
uint8_t *bytes = (void *) &counts;
printf (" %d", bytes [index]);
}
printf ("\n");
}
int main(void)
{
pop_count1();
pop_count2();
return 0;
}
答案 0 :(得分：9)
16个8位值的SSE 4 popcount可以这样并行完成：
#include <stdio.h>
#include <stdint.h>
#include <immintrin.h>
//----------------------------------------------------------------------------
//
// parallelPopcnt16bytes - find population count for 8-bit groups in xmm (16 groups)
// each byte of xmm result contains a value ranging from 0 to 8
//
static __m128i parallelPopcnt16bytes (__m128i xmm)
{
const __m128i mask4 = _mm_set1_epi8 (0x0F);
const __m128i lookup = _mm_setr_epi8 (0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4);
__m128i low, high, count;
low = _mm_and_si128 (mask4, xmm);
high = _mm_and_si128 (mask4, _mm_srli_epi16 (xmm, 4));
count = _mm_add_epi8 (_mm_shuffle_epi8 (lookup, low), _mm_shuffle_epi8 (lookup, high));
return count;
}
//----------------------------------------------------------------------------
int main (void)
{
int index;
__m128i testVector = _mm_set_epi8 (1, 2, 4, 8, 16, 32, 64, 128, 0, 1, 3, 7, 15, 31, 63, 127);
__m128i counts = parallelPopcnt16bytes (testVector);
printf ("population count for each byte:");
for (index = 15; index >= 0; index--)
{
uint8_t *bytes = (void *) &counts;
printf (" %d", bytes [index]);
}
printf ("\n");
return 0;
}
//----------------------------------------------------------------------------
答案 1 :(得分：2)
popcnt与SSE4.2 ISA扩展同时引入，但不在SSE向量寄存器上运行。每个结果都需要单独的说明。
此外，它没有为8位操作数定义。如果需要对每个字节进行计数，则需要填充到16位。
你可以在64位寄存器中一次总计8个字节，但这听起来不像你所追求的那样。
参考：The SSE4 manual。
我没有对此进行过测试，但您可以将SSE注册到0x80808080 ...以获得所有1或全0的16字节掩码。对一个字节中的所有8位重复，并对掩码求和。由于所有1都表示在二进制补码中为-1，因此取消16个字节，并且您拥有所有结果。
AND和比较操作应该能够并行运行。添加链是依赖的，但它仍然应该快速运行，并且它适合32条指令。 （只需要增加7个。）
/* init */
__m128i bit0 = _mm_set1_epi8( 0x80 );
__m128i mask0 = _mm_and_si128( in, bit0 );
__m128i sum = _mm_cmpeq_epi8( mask0, _mm_setzero_si128() );
/* general pattern */
__m128i bit1 = _mm_set1_epi8( 0x40 );
__m128i mask1 = _mm_and_si128( in, bit1 );
mask1 = _mm_cmpeq_epi8( mask1, _mm_setzero_si128() );
sum = _mm_add_epi8( sum, mask1 );
/* next bit */
__m128i bit2 = _mm_set1_epi8( 0x20 );
__m128i mask2 = _mm_and_si128( in, bit2 );
mask2 = _mm_cmpeq_epi8( mask2, _mm_setzero_si128() );
sum = _mm_add_epi8( sum, mask2 );
...
/* finish up */
sum = _mm_sub_epi8( _mm_setzero_si128(), sum );