55 lines
2.0 KiB
C
55 lines
2.0 KiB
C
#include "StgSharpC.h"
|
|
#include "ssc_intrinsic.h"
|
|
|
|
int index_pair_sse(short const *str, uint32_t target, int length)
|
|
{
|
|
int index = 0, result = 0, size = length, end = size - 8;
|
|
__m128i s_even = { .m128i_u32 = {
|
|
target,
|
|
target,
|
|
target,
|
|
target,
|
|
} };
|
|
__m128i vec, com;
|
|
|
|
while (index < end) {
|
|
vec = _mm_loadu_epi32(str + index);
|
|
com = _mm_cmpeq_epi16(vec, s_even);
|
|
result = _mm_movemask_epi8(com);
|
|
if (result) {
|
|
for (size_t i = 0; i < 12; i++) {
|
|
result >>= 2;
|
|
if (((result & 0xF) == 0xF)) {
|
|
return index + i;
|
|
}
|
|
if (result == 0) {
|
|
break;
|
|
}
|
|
}
|
|
}
|
|
index += 1;
|
|
vec = _mm_loadu_epi32(str + index);
|
|
com = _mm_cmpeq_epi16(vec, s_even);
|
|
result = _mm_movemask_epi8(com);
|
|
if (result) {
|
|
for (size_t i = 0; i < 12; i++) {
|
|
result >>= 2;
|
|
if (((result & 0xF) == 0xF)) {
|
|
return index + i;
|
|
}
|
|
if (result == 0) {
|
|
break;
|
|
}
|
|
}
|
|
}
|
|
index += 7;
|
|
}
|
|
for (; (index < size - 1) && (!result); index += 2) {
|
|
result = *(int *)(str + index) == target;
|
|
}
|
|
if (result) {
|
|
return index;
|
|
}
|
|
return -1;
|
|
}
|