Question

我尝试使用SSE通过uint64数组实现线性搜索说明。我为uint16和uint32工作了，但是我得到了编译器 uint64代码的错误（linux，gcc - 最后见规范）。

我试图比较2x2 64位数字，然后以某种方式翻译结果在我的数组的索引中。这适用于uint32（积分转到 http://schani.wordpress.com/2010/04/30/linear-vs-binary-search/）：

#include <xmmintrin.h>
#include <smmintrin.h>

typedef ham_u64_t vec2uint64 __attribute__ ((vector_size (16)));
typedef ham_u32_t vec4uint32 __attribute__ ((vector_size (16)));
typedef float     vec4float  __attribute__ ((vector_size (16)));
typedef ham_u16_t vec8uint16 __attribute__ ((vector_size (16)));
typedef ham_u8_t  vec16uint8 __attribute__ ((vector_size (16)));

// ...

vec4uint32 v1 = _mm_loadu_si128((const __m128i *)&data[start + i + 0]);
vec4uint32 v2 = _mm_loadu_si128((const __m128i *)&data[start + i + 4]);
vec4uint32 v3 = _mm_loadu_si128((const __m128i *)&data[start + i + 8]);
vec4uint32 v4 = _mm_loadu_si128((const __m128i *)&data[start + i + 12]);

vec4uint32 cmp0 = _mm_cmpeq_epi32(key4, v1);
vec4uint32 cmp1 = _mm_cmpeq_epi32(key4, v2);
vec4uint32 cmp2 = _mm_cmpeq_epi32(key4, v3);
vec4uint32 cmp3 = _mm_cmpeq_epi32(key4, v4);

vec8uint16 pack01 = __builtin_ia32_packssdw128(cmp0, cmp1);
vec8uint16 pack23 = __builtin_ia32_packssdw128(cmp2, cmp3);
vec16uint8 pack0123 = __builtin_ia32_packsswb128(pack01, pack23);

int res = __builtin_ia32_pmovmskb128(pack0123);
if (res > 0) {
  int czt = __builtin_ctz(~res + 1);
  return (start + i + czt);
}

这是我到目前为止为uint64提出的建议。比较有效，我只是不知道如何处理结果，以及__builtin_ia32_packssdw（）调用不编译：

vec2uint64 v1 = _mm_loadu_si128((const __m128i *)&data[start + i + 0]);
vec2uint64 v2 = _mm_loadu_si128((const __m128i *)&data[start + i + 2]);

vec2uint64 cmp0 = _mm_cmpeq_epi64(key2, v1);
vec2uint64 cmp1 = _mm_cmpeq_epi64(key2, v2);

vec4uint32 pack01 = __builtin_ia32_packssdw(cmp0, cmp1); // error
vec4uint32 pack23 = _mm_set1_epi32(0);
vec16uint8 pack0123 = __builtin_ia32_packsswb128(pack01, pack23);

int res = __builtin_ia32_pmovmskb128(pack0123);
if (res > 0) {
  int czt = __builtin_ctz(~res + 1);
  return (start + i + czt);
}

错误说：

error: cannot convert 'vec1uint64 {aka __vector(2) long unsigned int}'
to '__vector(2) int' for argument '1' to '__vector(4) short int
__builtin_ia32_packssdw(__vector(2) int, __vector(2) int)'

（vec2uint64的typedef位于顶部，位于uint32的代码中。）

我的环境：

Linux ws4484 3.5.0-48-generic #72~precise1-Ubuntu SMP Tue Mar 11 20:09:08 UTC 2014 x86_64 x86_64 x86_64 GNU/Linux

gcc version 4.6.3 (Ubuntu/Linaro 4.6.3-1ubuntu5)

我的问题不仅仅是我如何修复编译器错误，但是如果有人有更好的想法来获得匹配的数组索引，也许没有整个包装物？

提前致谢！

Answer 1

我建议不要使用内置的内在函数和隐式向量。这只有在你不使用非GCC内在函数（例如_mm_cmpeq_epi32）并且只想坚持GCC时才有意义。你可以做你想做的事情

__m128i key2 = _mm_set1_epi64x(key);
__m128i v1 = _mm_loadu_si128((const __m128i *)&data[start + i + 0]);
__m128i v2 = _mm_loadu_si128((const __m128i *)&data[start + i + 2]);

__m128i cmp0 = _mm_cmpeq_epi64(key2, v1);
__m128i cmp1 = _mm_cmpeq_epi64(key2, v2);

__m128i low2  = _mm_shuffle_epi32(cmp0,0xD8);  
__m128i high2 = _mm_shuffle_epi32(cmp1,0xD8);      
__m128i pack = _mm_unpacklo_epi64(low2,high2);

__m128i pack01 = _mm_packs_epi32(pack, _mm_setzero_si128());
__m128i pack0123 = _mm_packs_epi16(pack01, _mm_setzero_si128());

int res =  _mm_movemask_epi8(pack0123);

您可以找到一个更有效的版本来避免打包，但是您必须使用与__builtin_ctz不同的功能。

对于32位整数，我建议

__m128i key4 = _mm_set1_epi32(key);
__m128i v1 = _mm_loadu_si128((const __m128i *)&data[start + i + 0]);
__m128i v2 = _mm_loadu_si128((const __m128i *)&data[start + i + 4]);
__m128i v3 = _mm_loadu_si128((const __m128i *)&data[start + i + 8]);
__m128i v4 = _mm_loadu_si128((const __m128i *)&data[start + i + 12]);

__m128i cmp0 = _mm_cmpeq_epi32(key4, v1);
__m128i cmp1 = _mm_cmpeq_epi32(key4, v2);
__m128i cmp2 = _mm_cmpeq_epi32(key4, v3);
__m128i cmp3 = _mm_cmpeq_epi32(key4, v4);

__m128i pack01 = _mm_packs_epi32(cmp0, cmp1);
__m128i pack23 = _mm_packs_epi32(cmp2, cmp3);
__m128i pack0123 = _mm_packs_epi16(pack01, pack23);

int res = _mm_movemask_epi8(pack0123);

Answer 2

这是一个搜索匹配64位值的完整解决方案。在这种情况下，值（namehash）是结构成员。此例程在每次迭代时比较8个64位值，并提供匹配的结构索引。

//ptr is a struct array
__m128i key2 = _mm_set1_epi64x(k); //k is the 64 bit search key
for(;;)
   {
   if(!ptr[i].namehash)return NULL;
   __m128i v1 = _mm_set_epi64x (ptr[i+1].namehash,ptr[i].namehash);
   __m128i v2 = _mm_set_epi64x (ptr[i+3].namehash,ptr[i+2].namehash);
   __m128i v3 = _mm_set_epi64x (ptr[i+5].namehash,ptr[i+4].namehash);
   __m128i v4 = _mm_set_epi64x (ptr[i+7].namehash,ptr[i+6].namehash);

   __m128i cmp0 = _mm_cmpeq_epi64(key2, v1);
   __m128i cmp1 = _mm_cmpeq_epi64(key2, v2);
   __m128i cmp2 = _mm_cmpeq_epi64(key2, v3);
   __m128i cmp3 = _mm_cmpeq_epi64(key2, v4);

   __m128i L0  = _mm_shuffle_epi32(cmp0,0xD8);
   __m128i H1 = _mm_shuffle_epi32(cmp1,0xD8);
   __m128i L2  = _mm_shuffle_epi32(cmp2,0xD8);
   __m128i H3 = _mm_shuffle_epi32(cmp3,0xD8);

   __m128i pack0 = _mm_unpacklo_epi64(L0,H1);
   __m128i pack1 = _mm_unpacklo_epi64(L2,H3);

   __m128i pack01 = _mm_packs_epi32(pack0,pack1);
   __m128i pack0123 = _mm_packs_epi16(pack01, _mm_setzero_si128());

   res =  _mm_movemask_epi8(pack0123);
   if(res > 0)break;
   i+=8;
   }

int index = i + __builtin_ctz(res);  //The struct table index to the matching struct.

重要说明：struct array length必须是8的倍数，并且至少有1个NULL尾随成员

或者，如果每次迭代只需要进行2次64位比较，则上述内容可以大大简化为：

for(;;)
   {
   if(!ptr[i].namehash)return NULL;
   __m128i v1 = _mm_set_epi64x (ptr[i+1].namehash,ptr[i].namehash);
   __m128i cmp0 = _mm_cmpeq_epi64(key2, v1);
   res =  _mm_movemask_epi8(cmp0);
   if(res > 0)break;
   i+=2;
   }
int ctz = __builtin_ctz(res);
int index = i + (ctz>>5);  //The struct table index to the matching struct.

Answer 3

我找不到任何将64位整数转换为32位整数的指令，这就是你需要使用packssdw等等。它会变得很长而且很混乱，但应该可以工作：

所以，我认为解决方案是使用位掩码（位0,1,2,3：

这些在循环之前：

vec2uint64 mask0 = { 2, 1 };
vec2uint64 mask1 = { 8, 4 };
vec2uint64 zero  = { 0, 0 };

内圈：

vec2uint64 res0 = _mm_and_si128(cmp0, mask0);
vec2uint64 res1 = _mm_and_si128(cmp1, mask1);

vec2uint64 res2 = _mm_or_si128(res0, res1);

然后我们需要将高位部分拖放到新变量的低部分：

vec2uint64 hi0 = _mm_unpackhi_epi64(res0, zero);
vec2uint64 hi1 = _mm_unpackhi_epi64(res1, zero);

vec2uint64 hi2 = _mm_or_si128(hi0, hi1);
vec2uint64 res3 = _mm_or_si128(res2, hi2);

现在，res的低64位[井，低4位，其余为零]是一个位掩码，其中一个值匹配。

int res = _mm_cvtsi128_si32(res3);

（现在你可以像之前一样计算尾随零）。

通过uint64 []与SSE进行线性搜索

3 个答案: