2014-09-25 108 views
1

我有我正在分配的两项SSE整数如下:洗牌注册

__m128i m1 = _mm_set_epi32(4,3,2,1); 
    __m128i m2 = _mm_set_epi32(40,30,20,10); 

现在,我要做的这两个寄存器之间的一些洗牌,并把结果保存在另外两个寄存器,从而输出如下:

 m3 = (30,3,10,1); 
    m4 = (40,4,20,2); 

有什么办法可以实现这个!

感谢

+0

但我想用SSE洗牌 – 2014-09-25 20:38:27

+0

@MarcoA做。这是用来在单个寄存器中洗牌......我想在两个寄存器之间洗牌 – 2014-09-25 20:46:05

+0

对,对不起。我应该仔细阅读 – 2014-09-25 20:46:58

回答

2

可以是这样做的:

t1 = _mm_shuffle_epi32(m1, 0xd8); 
t2 = _mm_shuffle_epi32(m2, 0xd8);  
m4 = _mm_unpackhi_epi32(t1,t2); 
m3 = _mm_unpacklo_epi32(t1,t2); 

下面是一个完整的例子

#include <x86intrin.h> 
#include <stdio.h> 

int main() { 
     __m128i m1 = _mm_set_epi32(4,3,2,1); 
     __m128i m2 = _mm_set_epi32(40,30,20,10); 

     __m128i m3, m4, t1, t2; 
     t1 = _mm_shuffle_epi32(m1, 0xd8); 
     t2 = _mm_shuffle_epi32(m2, 0xd8);  
     m4 = _mm_unpackhi_epi32(t1,t2); 
     m3 = _mm_unpacklo_epi32(t1,t2); 

     int out3[4], out4[4]; 
     _mm_store_si128((__m128i*)out3, m3); 
     _mm_store_si128((__m128i*)out4, m4); 
     printf("%d %d %d %d\n", out3[3], out3[2], out3[1], out3[0]); 
     printf("%d %d %d %d\n", out4[3], out4[2], out4[1], out4[0]); 
} 

输出

30 3 10 1 
40 4 20 2 
+1

你让我的一天...非常感谢你 – 2014-09-25 23:55:52

+0

BTW 0xd8 = 3120在base-4(和二进制11011000)。这告诉你我正在交换的数字。我知道base-4不是典型的思维方式,但在这种情况下很方便。 – 2014-09-26 07:52:54

1

我想出了一个稍微不同的Z Boson解决方案:

#include <emmintrin.h> 
#include <iostream> 
#include <cstring> 

void print_data(const char *name, __m128i v) 
{ 
    struct { int a, b, c, d; } unpacked; 

    std::memcpy((void *)&unpacked, (void *)&v, sizeof(v)); 
    std::cout << name << ":"; 
    std::cout << unpacked.d << " " << unpacked.c << " " 
      << unpacked.b << " " << unpacked.a << std::endl; 
} 

int main() 
{ 

    __m128i m1 = _mm_set_epi32(4,3,2,1); 
    __m128i m2 = _mm_set_epi32(40,30,20,10); 
    __m128i mask = _mm_set_epi32(-1,0,-1,0); 

    /* 
    m3 = (30,3,10,1); 
    m4 = (40,4,20,2); 
    */ 

    __m128i tmp1 = _mm_shuffle_epi32(m2, 0x80); 
    __m128i tmp2 = _mm_shuffle_epi32(m1, 0x31); 
    __m128i m3 = _mm_or_si128(_mm_and_si128(mask, tmp1), _mm_andnot_si128(mask, m1)); 
    __m128i m4 = _mm_or_si128(_mm_and_si128(mask, m2), _mm_andnot_si128(mask, tmp2)); 

    print_data("m3", m3); 
    print_data("m4", m4); 
} 

我敢肯定,解压缩变形稍微好一点...