c++ SIMD AVX2比较 例子

大家好，又见面了，我是你们的朋友全栈君。

示例代码含义：记目标字符串中有多少个目标字符。
linux代码(例子)如下：

#include <iostream>
#include <x86intrin.h>
#include <fstream>
#include <chrono>
using namespace std;
struct StringView { 

const char* p;
const size_t len;
};
StringView FileSize(const char* fileName) { 

ifstream ifstr(fileName);
const auto b = ifstr.tellg();
ifstr.seekg(0, ios::end);
const auto e = ifstr.tellg();
const size_t fileSize = e - b;
ifstr.seekg(0, ios::beg);
char *p = new char[fileSize];
ifstr.read(p, fileSize);
return { 
p, fileSize};
}
// Normal function
size_t count_c_normal(const StringView& str, const uint8_t c) { 

uint32_t num = 0;
for (uint32_t i = 0; i < str.len; ++i) { 

if (c == *(str.p + i)) { 

++num;
}
}
return num;
}
// SIMD function
size_t count_c_simd(const StringView& str, const uint8_t c) { 

__m128i ch = _mm_set1_epi8(c); // char ch[16] = { c, c, ..., c }
size_t cnt = 0;
uint32_t i = 0;
for (; i < str.len; i+=16) { 

// char t[16] = { (str+i)[0], (str+i)[1], ... }
__m128i t = _mm_loadu_si128((__m128i *)(str.p + i));
__m128i res = _mm_cmpeq_epi8(t, ch);
// res[16] = { 0xFF, 0x00, 0xFF ... }
unsigned mask = _mm_movemask_epi8(res);
// bits[16] = 0...1101
cnt += __builtin_popcount(mask);
}
// free cnt .
for (; i < str.len; ++i) { 

if (c == *(str.p + i))
{ 

++cnt;
}
}
return cnt;
}
// AVX function
size_t count_c_avx256(const StringView& str, const uint8_t c) { 

__m256i ch = _mm256_set1_epi8(c); // char ch[16] = { c, c, ..., c }
size_t cnt = 0;
uint32_t i = 0;
for (; i < str.len; i+=32) { 

// char t[16] = { (str+i)[0], (str+i)[1], ... }
__m256i t = _mm256_loadu_si256((__m256i *)(str.p + i));
__m256i res = _mm256_cmpeq_epi8(t, ch);
// res[16] = { 0xFF, 0x00, 0xFF ... }
unsigned mask = _mm256_movemask_epi8(res);
// bits[16] = 0...1101
cnt += __builtin_popcount(mask);
}
// free cnt .
for (; i < str.len; ++i) { 

if (c == *(str.p + i))
{ 

++cnt;
}
}
return cnt;
}
int main() { 

const auto ret = FileSize("./test_file");
size_t cnt1 = 0, cnt2 = 0, cnt3 = 0;
const auto t1 = std::chrono::steady_clock::now();
cnt1 = count_c_normal(ret, uint8_t('1'));
const auto t2 = std::chrono::steady_clock::now();
cnt2 = count_c_simd(ret, uint8_t('1'));
const auto t3 = std::chrono::steady_clock::now();
cnt3 = count_c_avx256(ret, uint8_t('1'));
const auto t4 = std::chrono::steady_clock::now();
std::cout << "cnt1:" << cnt1 << ",cnt2:" << cnt2 << ",cnt3:" << cnt3 << std::endl;
const auto d1 = std::chrono::duration_cast<std::chrono::milliseconds>(t2-t1).count();
const auto d2 = std::chrono::duration_cast<std::chrono::milliseconds>(t3-t2).count();
const auto d3 = std::chrono::duration_cast<std::chrono::milliseconds>(t4-t3).count();
std::cout << "time1:" << d1 << ",time2:" << d2 << ",time3:" << d3 << std::endl;
return 0;
}

生成随机文件代码详见：https://blog.csdn.net/weixin_41644391/article/details/113526563

编译命令：g++ -std=c++11 main.cc -o main -mavx -mavx2 -O2
性能：

普通O2:1890ms,simd:509ms,avx2:253ms

因为编译命令中加了avx2，怀疑simd的代码也被avx2优化了。纯simd结果可见：https://blog.csdn.net/weixin_41644391/article/details/113526563

其他：基于avx512的测试因为不支持gcc4.8.5，所以需要等一段时间才能出来。