testing/yet-another-apply-test.c

   1 #include <dispatch/dispatch.h>
   2 #include <sys/stat.h>
   3 #include <sys/mman.h>
   4 #include <mach/mach.h>
   5 #include <mach/mach_time.h>
   6 #include <stdlib.h>
   7 #include <stdio.h>
   8 #include <fcntl.h>
   9 #include <assert.h>
  10
  11 static uint64_t total;
  12
  13 #define LAPS (256 * 1024 * 1024)
  14 #define SIZE (LAPS * sizeof(int))
  15
  16 int
  17 main(int argc, char *argv[])
  18 {
  19         dispatch_queue_t cq = dispatch_get_concurrent_queue(0);
  20         struct stat sb;
  21         long double cycles;
  22         uint64_t s, e, d;
  23         uint64_t tmp_total;
  24         int r, fd;
  25         const int *nums;
  26         size_t i, stride;
  27
  28         if (argc != 2) {
  29                 fprintf(stderr, "usage: %s <file>\n", argv[0]);
  30                 exit(EXIT_FAILURE);
  31         }
  32
  33         // make sure to have 2GB + of RAM installed and run this before hand:
  34         // dd if=/dev/random bs=1024k count=1024 of=/tmp/testfile
  35         fd = open(argv[1], O_RDONLY);
  36         assert(fd != -1);
  37
  38         r = fstat(fd, &sb);
  39         assert(r != -1);
  40         assert(sb.st_size >= (off_t)SIZE);
  41
  42         nums = mmap(NULL, SIZE, PROT_READ, MAP_FILE|MAP_PRIVATE, fd, 0);
  43         assert(nums != MAP_FAILED);
  44
  45         // force pages to be faulted in
  46         for (i = 0; i < LAPS; i++) {
  47                 total += nums[i];
  48         }
  49
  50         for (stride = 1; stride < (LAPS + 1); stride <<= 1) {
  51                 total = 0;
  52                 s = mach_absolute_time();
  53                 dispatch_apply(LAPS / stride, cq, ^(size_t idx) {
  54                         const int *nums2 = nums + (idx * stride);
  55                         uint64_t ptotal = 0;
  56                         size_t idx2 = 0;
  57
  58                         // assert(stride > 0);
  59                         do {
  60                                 ptotal += nums2[idx2++];
  61                         } while (idx2 < stride);
  62
  63                         __sync_fetch_and_add(&total, ptotal);
  64                 });
  65                 e = mach_absolute_time();
  66                 d = e - s;
  67                 cycles = d;
  68                 cycles /= LAPS;
  69                 printf("da%lu:\t%Lf ns\n", stride, cycles);
  70         }
  71
  72         tmp_total = 0;
  73         total = 0;
  74         s = mach_absolute_time();
  75         for (i = 0; i < LAPS; i++) {
  76                 tmp_total += nums[i];
  77         }
  78         total = tmp_total;
  79         e = mach_absolute_time();
  80         d = e - s;
  81         cycles = d;
  82         cycles /= LAPS;
  83         printf("naïve:\t%Lf ns\n", cycles);
  84
  85
  86         tmp_total = 0;
  87         total = 0;
  88         s = mach_absolute_time();
  89 #pragma omp parallel for reduction(+:tmp_total)
  90         for (i = 0; i < LAPS; i++) {
  91                 tmp_total += nums[i];
  92         }
  93         total = tmp_total;
  94         e = mach_absolute_time();
  95         d = e - s;
  96         cycles = d;
  97         cycles /= LAPS;
  98         printf("OpenMP:\t%Lf ns\n", cycles);
  99
 100         exit(EXIT_SUCCESS);
 101 }