如何使用向量化解決方案提高IP位址解析速度？

Improving IP Address Parsing Speed

Your current code for parsing IPv4 addresses is fairly efficient, but it can be further optimized for even greater speed. One approach is to utilize vectorized solutions specifically designed for this task.

Vectorized Solution for Fast IPv4 Parsing

For x86 processors supporting SSE4.1 or SSSE3 instructions, here's a vectorized solution that significantly improves performance:

__m128i shuffleTable[65536];    //can be reduced 256x times, see @IwillnotexistIdonotexist

UINT32 MyGetIP(const char *str) {
    __m128i input = _mm_lddqu_si128((const __m128i*)str);   //"192.167.1.3"
    input = _mm_sub_epi8(input, _mm_set1_epi8('0'));        //1 9 2 254 1 6 7 254 1 254 3 208 245 0 8 40 
    __m128i cmp = input;                                    //...X...X.X.XX...  (signs)
    UINT32 mask = _mm_movemask_epi8(cmp);                   //6792 - magic index
    __m128i shuf = shuffleTable[mask];                      //10 -1 -1 -1 8 -1 -1 -1 6 5 4 -1 2 1 0 -1 
    __m128i arr = _mm_shuffle_epi8(input, shuf);            //3 0 0 0 | 1 0 0 0 | 7 6 1 0 | 2 9 1 0 
    __m128i coeffs = _mm_set_epi8(0, 100, 10, 1, 0, 100, 10, 1, 0, 100, 10, 1, 0, 100, 10, 1);
    __m128i prod = _mm_maddubs_epi16(coeffs, arr);          //3 0 | 1 0 | 67 100 | 92 100 
    prod = _mm_hadd_epi16(prod, prod);                      //3 | 1 | 167 | 192 | ? | ? | ? | ?
    __m128i imm = _mm_set_epi8(-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, 6, 4, 2, 0);
    prod = _mm_shuffle_epi8(prod, imm);                     //3 1 167 192 0 0 0 0 0 0 0 0 0 0 0 0
    return _mm_extract_epi32(prod, 0);
//  return (UINT32(_mm_extract_epi16(prod, 1)) << 16) + UINT32(_mm_extract_epi16(prod, 0)); //no SSE 4.1
}

Shuffle Table Pre-Calculation

To utilize this vectorized solution effectively, it requires a pre-calculated shuffle table, shuffleTable, that can be generated as follows:

void MyInit() {
    memset(shuffleTable, -1, sizeof(shuffleTable));
    int len[4];
    for (len[0] = 1; len[0] <= 3; len[0]++)
        for (len[1] = 1; len[1] <= 3; len[1]++)
            for (len[2] = 1; len[2] <= 3; len[2]++)
                for (len[3] = 1; len[3] <= 3; len[3]++) {
                    int slen = len[0] + len[1] + len[2] + len[3] + 4;
                    int rem = 16 - slen;
                    for (int rmask = 0; rmask < 1<<rem; rmask++) {
//                    { int rmask = (1<<rem)-1;    //note: only maximal rmask is possible if strings are zero-padded
                        int mask = 0;
                        char shuf[16] = {-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1};
                        int pos = 0;
                        for (int i = 0; i < 4; i++) {
                            for (int j = 0; j < len[i]; j++) {
                                shuf[(3-i) * 4 + (len[i]-1-j)] = pos;
                                pos++;
                            }
                            mask ^= (1<<pos);
                            pos++;
                        }
                        mask ^= (rmask<<slen);
                        _mm_store_si128(&shuffleTable[mask], _mm_loadu_si128((__m128i*)shuf));
                    }
                }
}

Performance Benchmark

On an Ivy Bridge processor, the vectorized solution demonstrates impressive performance, processing 336 million addresses per second. This is approximately 7.8 times faster than the code provided in the original question.

以上是如何使用向量化解決方案提高IP位址解析速度？的詳細內容。更多資訊請關注PHP中文網其他相關文章！