SIMD and Avx2

比眉伴天荒 2022-10-05 00:49 237阅读 0赞

SIMD 一条指令可以执行多个数据group的计算和输出。对于SIMD相对应的SISD.
intel SSE2 , AVX2, AVX-512

假设有一个任务是统计字符串中每一个字符出现的次数，我们可以用128bit 的SISD指令进行统计。每8个bit代表一个字符，所以只需要两个SIMD指令（move mask、pop count）。
在这里插入图片描述

详细测试:

  1. #include <stdio.h>
  2. #include <thread>
  3. #define INC_TO 1000000 // one million...
  4. #include <mutex>
  5. #include <functional>
  6. #include <atomic>
  7. #include <vector>
  8. #include <sstream>
  9. #include <iostream>
  10. #include <emmintrin.h>
  11. #include <immintrin.h>
  12. #include <assert.h> 
  13. #include <x86intrin.h>
  15. struct StringView {
  16.   char* buffer;
  17.   size_t len;
  18. };
  20. void RandomGeneratorFile(const char* filename) {
  21.   FILE* fp =  fopen(filename, "w");
  22.   const size_t numbers = 16*8*1000;
  23.   size_t count =0;
  24.   do {
  25.     for (char i = 'A'; i < 'Z'; i++) {
  26.       fputc(i, fp);
  27.     }
  28.     count++;
  29.   } while (count < numbers);
  30.   fclose(fp);
  31. }
  33. StringView* GetFileContent(const char* filename) {
  34.   FILE* fp = fopen(filename, "r");
  35.   /*Move file point at the end of file.*/
  36.   fseek(fp,0,SEEK_END);
  37.   /*Get the current position of the file pointer.*/
  38.   size_t size=ftell(fp);
  39.   printf("file size:%d\n", size);
  40.   char * buffer = new char[size];
  41.   fseek(fp, 0, SEEK_SET);
  42.   fread(buffer, size, size, fp);
  43.   //  printf("content of buffer:%s\n", buffer);
  44.   fclose(fp);
  45.   StringView* str = new StringView();
  46.   str->buffer = buffer;
  47.   str->len = size;
  48.   return str;
  49. }
  51. // 正常统计字符串
  52. size_t count_chars_8(const char* data, size_t size, const char ch)
  53. {
  54.   size_t total = 0;
  55.   while (size) {
  56.     if (*data == ch)
  57.       total += 1;
  58.     data += 1;
  59.     size -= 1;
  60.   }
  61.   return total;
  62. }
  64. // SIMD
  65. size_t count_chars_128(const char* data, size_t size, const char ch)
  66. {
  67.   size_t total = 0;
  68.   assert(size % 16 == 0);
  69.   // 将ch广播16次
  70.   __m128i tocmp = _mm_set1_epi8(ch);
  71.   while (size) {
  72.     int mask = 0;
  73.     // 从memory 取出128bit数据
  74.     __m128i chunk = _mm_load_si128 ((__m128i const*)data);
  75.     // 对128bit数据进行比较， 返回16bit
  76.     __m128i results = _mm_cmpeq_epi8(chunk, tocmp);
  77.     // 
  78.     mask = _mm_movemask_epi8(results);
  79.     //统计int32的bit位是1的值
  80.     // _mm_ prefix, because it does not operate on 128-bit registers, it just operates on standard 64-bit registers.
  81.     total += _popcnt32(mask);
  82.     data += 16;
  83.     size -= 16;
  84.   }
  85.   return total;
  86. }
  88. // AVX
  89. size_t count_chars_avx(const char* data, size_t size, const char ch)
  90. {
  91.   size_t total = 0;
  92.   assert(size % 16 == 0);
  93.   __m256i tocmp = _mm256_set1_epi8(ch);
  94.   while(size) {
  95.     __m256i chunk = _mm256_loadu_si256((__m256i*)data);
  96.     __m256i results = _mm256_cmpeq_epi8(tocmp, chunk);
  97.     unsigned mask = _mm256_movemask_epi8(results);
  98.     total += __builtin_popcount(mask);
  99.     data += 32;
  100.     size -= 32;
  101.   }
  102.   // printf("count:%d\n", total);
  103.   return total;
  104. }
  106. static void print_time_us(const char* name, 
  107.       size_t(*fn)(const char*, size_t, const char), const char* a,
  108.       size_t size, const char ch) {
  109.     struct timespec start, end;
  110.     clock_gettime(CLOCK_MONOTONIC_RAW, &start);
  111.     fn(a, size, ch);
  112.     clock_gettime(CLOCK_MONOTONIC_RAW, &end);
  113.     uint64_t delta_us = (end.tv_sec - start.tv_sec) * 1000000 + (end.tv_nsec - start.tv_nsec) / 1000;
  114.     printf("Running: '%s' took %llu u/s\n", name, delta_us);
  115. }
  117. int main()
  118. {
  119.   RandomGeneratorFile("test.file");
  120.   StringView* str = GetFileContent("test.file");
  121.   print_time_us("NORMAL", count_chars_8, str->buffer, str->len, 'A');
  122.   print_time_us("SIMD", count_chars_128, str->buffer, str->len, 'A');
  123.   print_time_us("AVX", count_chars_avx, str->buffer, str->len, 'A');
  124.   return 0;
  125. }

编译命令：g++ -std=c++14 main.cc -o main -mavx -mavx2 -O2

输出：

  1. file size:3200000
  2. Running: 'NORMAL' took 2505 u/s
  3. Running: 'SIMD' took 192 u/s
  4. Running: 'AVX' took 97 u/s

发表评论

表情：
评论列表 （有 0 条评论，237人围观）

还没有评论，来说两句吧...

相关阅读