perf/perf.h

   1 /*
   2  * Written in 2013 by Gregor Pintar <grpintar@gmail.com>
   3  *
   4  * To the extent possible under law, the author(s) have dedicated
   5  * all copyright and related and neighboring rights to this software
   6  * to the public domain worldwide.
   7  *
   8  * This software is distributed without any warranty.
   9  *
  10  * You should have received a copy of the CC0 Public Domain Dedication.
  11  * If not, see <http://creativecommons.org/publicdomain/zero/1.0/>.
  12  */
  13
  14 #include <stdio.h>
  15
  16 #if defined(PERF_UNIX)
  17
  18 #include <unistd.h>
  19 #include <sched.h>
  20 #include <time.h>
  21
  22 #elif defined(PERF_WINDOWS)
  23
  24 #include <windows.h>
  25 #include <PowrProf.h>
  26 #include <ntstatus.h>
  27
  28 #endif
  29
  30 #if (defined(__GNUC__) || defined(__clang__)) && (defined(__i386__) || defined(__x86_64__))
  31 #define PERF_RDTSC
  32 #endif
  33
  34 #if defined(PERF_WINDOWS) || defined(PERF_RDTSC)
  35
  36 #include <stdint.h>
  37
  38 typedef uint64_t perf_int;
  39 #define PERF_INT_MAX UINT64_MAX
  40
  41 #else
  42
  43 #include <time.h>
  44
  45 typedef clock_t perf_int;
  46 #define PERF_INT_MAX (clock_t)UINT64_MAX
  47
  48 #endif
  49
  50 inline perf_int perf_clock(void)
  51 {
  52         #if defined(PERF_WINDOWS) && defined(PERF_QPC)
  53
  54         LARGE_INTEGER x;
  55
  56         QueryPerformanceCounter(&x);
  57
  58         return x.QuadPart;
  59
  60         #elif defined(PERF_RDTSC)
  61
  62         uint32_t hi;
  63         uint32_t lo;
  64
  65         __asm__ __volatile__
  66         (
  67                 "cpuid\n"
  68                 "rdtsc"
  69                 : "=a"(lo), "=d"(hi)
  70                 : "a"(0)
  71                 : "%ebx", "%ecx"
  72         );
  73
  74         return (((uint64_t)hi) << 32) | lo;
  75
  76         #elif defined(_POSIX_CPUTIME)
  77
  78         struct timespec x;
  79
  80         if(clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &x))
  81                 perror("clock_gettime()");
  82
  83         return (((uint64_t)x.tv_sec) << 32) | x.tv_nsec;
  84
  85         #else
  86
  87         return clock();
  88
  89         #endif
  90 }
  91
  92 #if defined(PERF_AVG)
  93
  94 #define PERF_START                                              \
  95 {                                                                               \
  96         unsigned int i;                                         \
  97         cycles = perf_clock();                          \
  98         for(i = 0; i < 1000000; i++)            \
  99         {
 100
 101 #define PERF_STOP                                                               \
 102         }                                                                                       \
 103         cycles = (perf_clock() - cycles) / 1000000;     \
 104 }
 105
 106 #else
 107
 108 #define PERF_START                                      \
 109 {                                                                       \
 110         perf_int t0;                                    \
 111         perf_int t1;                                    \
 112         unsigned int i;                                 \
 113         cycles = PERF_INT_MAX;                  \
 114         for(i = 0; i < 1000000; i++)    \
 115         {                                                               \
 116                 t0 = perf_clock();
 117
 118 #define PERF_STOP                                                       \
 119                 t1 = perf_clock() - t0 - perf_c;        \
 120                 if(cycles > t1) cycles = t1;            \
 121         }                                                                               \
 122 }
 123
 124 #endif
 125
 126 perf_int perf_c;
 127
 128 void perf_init(void)
 129 {
 130         perf_int cycles;
 131
 132         #if defined(PERF_UNIX)
 133
 134         struct sched_param p;
 135
 136         #ifdef _GNU_SOURCE
 137         cpu_set_t set;
 138
 139         /* lock process to one core */
 140         CPU_ZERO(&set);
 141         CPU_SET(0, &set);
 142         if(sched_setaffinity(0, CPU_ALLOC_SIZE(1), &set))
 143                 perror("sched_setaffinity()");
 144         #endif
 145
 146         puts("For better results disable dynamic CPU frequency scaling!");
 147
 148         /* set highest possible priority */
 149         for(p.sched_priority = sched_get_priority_max(SCHED_FIFO);
 150                 p.sched_priority; p.sched_priority--)
 151         {
 152                 if(!sched_setscheduler(0, SCHED_FIFO, &p)) break;
 153         }
 154         printf("SCHED_FIFO with priority: %u\n", p.sched_priority);
 155
 156         #elif defined(PERF_WINDOWS)
 157
 158         SYSTEM_POWER_CAPABILITIES s;
 159
 160         /* disable dynamic CPU frequency scaling */
 161         if(CallNtPowerInformation(SystemPowerCapabilities, 0, 0, &s, sizeof(s)) == STATUS_SUCCESS)
 162         {
 163                 if(s.ProcessorMinThrottle != 100 || s.ProcessorMaxThrottle != 100)
 164                 {
 165                         s.ProcessorMinThrottle = 100;
 166                         s.ProcessorMaxThrottle = 100;
 167                         if(CallNtPowerInformation(SystemPowerCapabilities, &s, sizeof(s), 0, 0) != STATUS_SUCCESS)
 168                                 puts("Can't disable dynamic CPU frequency scaling!");
 169                 }
 170         }
 171         else puts("Can't disable dynamic CPU frequency scaling!");
 172
 173         /* set priority */
 174         if(!SetPriorityClass(GetCurrentProcess(), REALTIME_PRIORITY_CLASS))
 175                 puts("SetPriorityClass() failed!");
 176
 177         /* lock process to one core */
 178         if(!SetProcessAffinityMask(GetCurrentProcess(), 1))
 179                 puts("SetProcessAffinityMask() failed!");
 180
 181         #endif
 182
 183         /* calibrate */
 184         perf_c = 0;
 185         PERF_START
 186         PERF_STOP
 187         perf_c = cycles;
 188 }
 189
 190 void perf_rest(void)
 191 {
 192         #if defined(PERF_UNIX)
 193         (void)sleep(1);
 194         #elif defined(PERF_WINDOWS)
 195         (void)Sleep(1);
 196         #endif
 197 }