我正在尝试使用 SSE 指令通过 uint64 数组实现线性搜索。我得到了适用于 uint16 和 uint32 的东西,但是我得到了 uint64 代码的编译器错误(linux、gcc - 请参阅最后的规范)。
我正在尝试比较 2x2 64 位数字,然后以某种方式将结果转换为我的数组的索引。这适用于 uint32(学分转到 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);
}
到目前为止,这是我为 uint64 想到的。比较有效,我只是不知道如何处理结果,并且 __builtin_ia32_packssdw() 调用无法编译:
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 在顶部,在 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,而是如果有人有更好的主意来获得匹配的数组索引,也许没有整个打包的东西?
提前致谢!