testing/fast_apply_bench.c

   1 #include <mach/mach.h>
   2 #include <mach/mach_time.h>
   3 #include <stdio.h>
   4 #include <stdlib.h>
   5 #include <assert.h>
   6 #include <errno.h>
   7 #include <string.h>
   8
   9 static inline uint64_t
  10 rdtsc(void)
  11 {
  12         uint32_t lo, hi;
  13
  14         asm volatile("rdtsc" : "=a" (lo), "=d" (hi));
  15
  16         return (uint64_t)hi << 32 | lo;
  17 }
  18
  19 __attribute__((noinline)) void
  20 apply_p(void (^b)(size_t), size_t offset, size_t count)
  21 {
  22         /* This would feed through to the existing dispatch_apply() */
  23         abort();
  24 }
  25
  26 /* a dynamically variable to eventually be added to the kernel/user 'commpage' */
  27 size_t total_active_cpus = 8;
  28
  29 __attribute__((noinline)) void
  30 apply(void (^b)(size_t), size_t offset, size_t count)
  31 {
  32         const size_t too_long = 100000; /* 100 us */
  33         const size_t laps = 16;
  34         uint64_t delta, tmp, now;
  35         size_t i;
  36
  37         if (total_active_cpus == 1) {
  38                 for (i = 0; i < count; i++) {
  39                         b(offset + i);
  40                 }
  41                 return;
  42         }
  43
  44         now = mach_absolute_time();
  45
  46         for (i = 0; i < count; i++) {
  47                 b(offset + i);
  48
  49                 if (i % laps) {
  50                         continue;
  51                 }
  52
  53                 tmp = mach_absolute_time();
  54                 delta = tmp - now;
  55                 now = tmp;
  56
  57                 if (delta > (too_long * laps) || (i == 0 && delta > too_long)) {
  58                         apply_p(b, offset + i + 1, count - (i + 1));
  59                         return;
  60                 }
  61         }
  62 }
  63
  64 int
  65 main(void)
  66 {
  67         void (^b)(size_t) = ^(size_t index) {
  68                 asm volatile(""); /* defeat compiler optimizations */
  69         };
  70         const size_t laps = 10000000;
  71         mach_timebase_info_data_t tbi;
  72         kern_return_t kr;
  73         long double dd;
  74         uint64_t s, e;
  75         size_t i;
  76
  77         kr = mach_timebase_info(&tbi);
  78         assert(kr == 0);
  79         assert(tbi.numer == tbi.denom); /* This will fail on PowerPC. */
  80
  81         s = mach_absolute_time();
  82         for (i = 0; i < laps; i++) {
  83                 b(i);
  84         }
  85         e = mach_absolute_time();
  86         dd = e - s;
  87         dd /= laps;
  88         printf("direct:\t%Lf ns\n", dd);
  89
  90         s = mach_absolute_time();
  91         apply(b, 0, laps);
  92         e = mach_absolute_time();
  93         dd = e - s;
  94         dd /= laps;
  95         printf("apply:\t%Lf ns\n", dd);
  96
  97         return 0;
  98 }