CPU time in Windows

[hgm]

Does anyone know how to get the CPU time used by the current process in Windows? I thought that clock() was supposed to do that, but it seems to always return a value identical to the wall-clock time from GetTickCount().

[Jim Ablett]

Does anyone know how to get the CPU time used by the current process in Windows? I thought that clock() was supposed to do that, but it seems to always return a value identical to the wall-clock time from GetTickCount().

How about getrusage >

Code: Select all

#include <sys/resource.h>
#include <errno.h>

#define WIN32_LEAN_AND_MEAN
#include <windows.h>

int __cdecl getrusage (int who, struct rusage *r_usage)
{
  FILETIME dummy;
  ULARGE_INTEGER KernelTime;
  ULARGE_INTEGER UserTime;

  if (!rusage) 
    {
      errno = EFAULT;
      return -1;
    }

  if (who != RUSAGE_SELF)
    {
      errno = EINVAL;
      return -1;
    }

  GetProcessTimes (GetCurrentProcess (), &dummy, &dummy,
                   (LPFILETIME) &KernelTime, (LPFILETIME) &UserTime);

  r_usage->ru_stime.tv_sec = (long) (0x7fffffff
				     & (unsigned long) KernelTime.QuadPart / 10000000);
  r_usage->ru_utime.tv_sec = (long) (0x7fffffff
				     & (unsigned long) UserTime.QuadPart / 10000000);
  r_usage->ru_stime.tv_usec = (long)(0x7fffffff
				     & (unsigned long) ((KernelTime.QuadPart % 10000000) / 10));
  r_usage->ru_utime.tv_usec = (long)(0x7fffffff
				     & (unsigned long) ((UserTime.QuadPart % 10000000) / 10));

  return 0;
}

Jim.

[hgm]

Thanks!

A big request:

could you make a Fruit 2.1 Windows compile for me that prints the real time and the user time? I don't manage to compile it with gcc in -mno-cygwin mode, it produces tons of error messages. (On Linux it compiled fine, and cpu_now() and real_now() also worked with good precision.) What I need is a version that prints

Code: Select all

 startTime = WallClockTime(); // time stamp (millisec)
cpuTime = CPUtime(); // Total CPU used upto now (millisec)
printf("info string times @ %u\n", startTime);

immediately after it receives a "go" command (this code is in protocol.cpp), and

Code: Select all

stopTime = WallClockTime(); // time stamp
cpuTime = CPUtime() - cpuTime; // Consumed CPU time
printf("info string times @ %u: real=%u cpu=%u\n", stopTime, stopTime - startTime, cpuTime);

just before it sends "bestmove". (Also in protocol.cpp.)

I have written a small program that extracts such messages from the debug file, and calculates from them the delay between sending a move and the opponent receiving it, and CPU time stolen by the GUI.

[Jim Ablett]

Thanks!

A big request:

could you make a Fruit 2.1 Windows compile for me that prints the real time and the user time? I don't manage to compile it with gcc in -mno-cygwin mode, it produces tons of error messages. (On Linux it compiled fine, and cpu_now() and real_now() also worked with good precision.) What I need is a version that prints

Code: Select all

 startTime = WallClockTime(); // time stamp (millisec)
cpuTime = CPUtime(); // Total CPU used upto now (millisec)
printf("info string times @ %u\n", startTime);

immediately after it receives a "go" command (this code is in protocol.cpp), and

Code: Select all

stopTime = WallClockTime(); // time stamp
cpuTime = CPUtime() - cpuTime; // Consumed CPU time
printf("info string times @ %u: real=%u cpu=%u\n", stopTime, stopTime - startTime, cpuTime);

just before it sends "bestmove". (Also in protocol.cpp.)

I have written a small program that extracts such messages from the debug file, and calculates from them the delay between sending a move and the opponent receiving it, and CPU time stolen by the GUI.

http://dl.dropbox.com/u/5047625/fruit-2 ... est%29.zip


Jim.

[Jim Ablett]

This is a better solution to Mingw32 cpu time bug >

Code: Select all

#define WIN32_LEAN_AND_MEAN
#include <windows.h>

static void
__winnt_cpu_time (long *sec, long *usec)
{
  union {
    FILETIME ft;
    unsigned long long ulltime;
  } kernel_time,  user_time;

  FILETIME unused1, unused2;
  unsigned long long total_time;

  /* No support for Win9x.  The high order bit of the DWORD
     returned by GetVersion is 0 for NT and higher. */
  if (GetVersion () >= 0x80000000)
    {
      *sec = -1;
      *usec = 0;
      return;
    }

  /* The FILETIME structs filled in by GetProcessTimes represent
     time in 100 nanosecond units. */
  GetProcessTimes (GetCurrentProcess (), &unused1, &unused2,
              	   &kernel_time.ft, &user_time.ft);
      
  total_time = (kernel_time.ulltime + user_time.ulltime)/10; 
  *sec = total_time / 1000000;
  *usec = total_time % 1000000;
}

Jim.

[Daniel Shawul]

I am surprized though why unix clock() measures the cpu time. C++ seems to define clock() as a measure of wall clock time http://www.cplusplus.com/reference/clib ... ime/clock/. It is mentioned that the reference could be different on different systems but nothing about some systems excluding I/O time.

[hgm]

Jim, could you make another Win-32 compile of Fruit for me? I have now increased the precision of the timers to 0.1 msec, for both wall-clock and CPU time.

Just use the file at http://hgm.nubati.net/protocol.cpp in stead of the normal one, that should do it.

I plan to distribute this version of Fruit 2.1 with the WinBoard binary package, and perhaps also make a separate 'timing kit' for people wanting to evaluate GUI performance, with this Fruit, a similarly adapted Fairy-Max and a tool to extract the timing lines from a debug file, and print the timing stats.

[Jim Ablett]

http://dl.dropbox.com/u/5047625/fruit-2 ... -win32.zip

Jim.

[hgm]

Wow, you are fast! Thanks!

[rbarreira]

I am surprized though why unix clock() measures the cpu time. C++ seems to define clock() as a measure of wall clock time http://www.cplusplus.com/reference/clib ... ime/clock/. It is mentioned that the reference could be different on different systems but nothing about some systems excluding I/O time.

Actually that page is even more confusing, as it talks about "processing times" as well.