gistfile1.c

#include <xmmintrin.h>
#include <stdio.h>
#include <stdint.h>
#include <string.h>
#include <assert.h>

__m64 zero ;

inline long atol_chunk(char *str) {
  uint32_t chunk = *((uint32_t *)str);
  // 0x31323334 -> 0x01020304
  uint64_t chunk_64 = chunk & 0x0f0f0f0f;

  // Interleave the bytes from above with 0s
  // so in = 0x0001000200030004
  __m64 chunk_m = *((__m64 *)&chunk_64);
  __m64 in = _mm_unpacklo_pi8(chunk_m, zero);

  // Multiply times the bases. This adds up
  // the left half and the right half
  __m64 mult = _mm_set_pi16(1, 10, 100, 1000);
  __m64 ret_m = _mm_madd_pi16(in, mult);

  // Shift back and add together
  return ((uint64_t)ret_m >> 32) + 
          ((uint64_t)ret_m & 0xffff);
}

long my_atol(char *arg) {
  int len = strlen(arg);
  long ret = 0;

  while (len >= 4) {
    ret *= 10000;

    long chunk = atol_chunk(arg);
    arg += 4;
    len -= 4;
    ret += chunk;
  }

  switch (len) {
    case 0:
      return ret;
    case 1:
      ret *= 10;
      ret += (*arg - '0');
      return ret;
    case 2:
      ret *= 100;
      ret += (*arg++ - '0') * 10;
      ret += *arg - '0';
      return ret;
    case 3:
      ret *= 1000;
      char buf[4] = {'0','0','0','0'};
      memcpy(buf + 4 - len, arg, len);
      ret += atol_chunk(buf);
      return ret;
    default:
      assert(0 && "x");
  }
}

__inline__ uint64_t rdtsc(void) {
  uint32_t lo, hi;
  __asm__ __volatile__ (      // serialize
  "xorl %%eax,%%eax \n        cpuid"
  ::: "%rax", "%rbx", "%rcx", "%rdx");
  /* We cannot use "=A", since this would use %rax on x86_64 and return only the lower 32bits of the TSC */
  __asm__ __volatile__ ("rdtsc" : "=a" (lo), "=d" (hi));
  return (uint64_t)hi << 32 | lo;
}


int main() {
  char buf[100];
  long sum = 0;
  int i;

  zero = _mm_set_pi16(0,0,0,0);
  uint64_t st = rdtsc();
  sum = 0;
  for (i = 0; i < 10000000; i++) {
    sprintf(buf, "%d", i);
    sum += atol(buf);
  }
  uint64_t et = rdtsc();
  printf("sum : %ld took %ld ticks\n", sum, (et - st));


  st = rdtsc();
  sum = 0;
  for (i = 0; i < 10000000; i++) {
    sprintf(buf, "%d", i);
    long ret = my_atol(buf);
    sum += ret;
#ifdef CHECK
    if (ret != i) {
      printf("fail at %d != %ld\n", i, ret);
      exit(1);
    }
#endif
  }
  et = rdtsc();
  printf("sum : %ld took %ld ticks\n", sum, (et - st));
  return 0;
}