/* Time routines for speed measurments. */ /* Copyright (C) 1999, 2000 Free Software Foundation, Inc. This file is part of the GNU MP Library. The GNU MP Library is free software; you can redistribute it and/or modify it under the terms of the GNU Lesser General Public License as published by the Free Software Foundation; either version 2.1 of the License, or (at your option) any later version. The GNU MP Library is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License for more details. You should have received a copy of the GNU Lesser General Public License along with the GNU MP Library; see the file COPYING.LIB. If not, write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. */ /* speed_time_init() - initialize timing things. speed_starttime() calls this if it hasn't been done yet, so you only need to call this explicitly if you want to use the global variables before the first measurement. speed_starttime() - start a time measurment. speed_endtime() - end a time measurment, return time taken, in seconds. speed_unittime - global variable with the unit of time measurement accuracy, in seconds. speed_precision - global variable which is the intended accuracy of time measurements. speed_measure() for instance runs target routines with enough repetitions so it takes at least speed_unittime*speed_precision seconds. A program can provide an option so the user can set this. speed_cycletime - the time in seconds for each CPU cycle, for example on a 100 MHz CPU this would be 1.0e-8. If the CPU frequency is unknown, speed_cycletime is 1.0. See speed_cycletime_init(). speed_time_string - a string describing the time method in use. Enhancements: Add support for accurate timing on more CPUs, machines and systems. Extend automatic CPU frequency determination to more kernels and systems. */ #include #include /* for getenv */ #if HAVE_UNISTD_H #include #endif #include #if HAVE_SYS_SYSCTL_H #include #endif #include "gmp.h" #include "gmp-impl.h" #include "longlong.h" #include "speed.h" #if HAVE_SPEED_CYCLECOUNTER #define SPEED_USE_CYCLECOUNTER 1 #else #define SPEED_USE_MICROSECOND_GETRUSAGE 0 #define SPEED_USE_MICROSECOND_GETTIMEOFDAY 1 #define SPEED_USE_TMS_UTIME 0 #endif #define TIMEVAL_SECS(tp) \ ((double) (tp)->tv_sec + (double) (tp)->tv_usec * 1.0e-6) /* Look for an environment variable for CPU clock frequency. GMP_CPU_FREQUENCY should be in Hertz, in floating point form, eg. "450e6". */ int speed_cpu_frequency_environment (void) { char *e; e = getenv ("GMP_CPU_FREQUENCY"); if (e == NULL) return 0; speed_cycletime = 1.0 / atof (e); return 1; } /* On FreeBSD 3.3 the headers have #defines like CPU_WALLCLOCK under CTL_MACHDEP but don't seem to have anything for machdep.tsc_freq or machdep.i586_freq. Using the string forms with sysctlbyname() works though, and lets libc worry about the defines and headers. FreeBSD 3.3 has tsc_freq, FreeBSD 2.2.8 has i586_freq instead. The "sysctl -a" command prints everything available. */ #if HAVE_SYSCTLBYNAME int speed_cpu_frequency_sysctlbyname (void) { unsigned val; size_t valsize; valsize = sizeof(val); if (sysctlbyname ("machdep.tsc_freq", &val, &valsize, NULL, 0) != 0 || valsize != sizeof(val)) { valsize = sizeof(val); if (sysctlbyname ("machdep.i586_freq", &val, &valsize, NULL, 0) != 0 || valsize != sizeof(val)) return 0; } speed_cycletime = 1.0 / (double) val; return 1; } #endif /* Linux doesn't seem to have any system call to get the CPU frequency, at least not in 2.0.x or 2.2.x, so it's necessary to read /proc/cpuinfo. i386 2.0.36 - "bogomips" is the CPU frequency. i386 2.2.13 - has both "cpu MHz" and "bogomips", and it's "cpu MHz" which is the frequency. alpha 2.2.5 - "cycle frequency [Hz]" seems to be right, "BogoMIPS" is very slightly different. */ int speed_cpu_frequency_proc_cpuinfo (void) { FILE *fp; char buf[128]; double val; int ret = 0; if ((fp = fopen ("/proc/cpuinfo", "r")) != NULL) { while (fgets (buf, sizeof (buf), fp) != NULL) { if (sscanf (buf, "cycle frequency [Hz] : %lf est.\n", &val) == 1) { speed_cycletime = 1.0 / val; ret = 1; break; } if (sscanf (buf, "cpu MHz : %lf\n", &val) == 1) { speed_cycletime = 1e-6 / val; ret = 1; break; } if (sscanf (buf, "bogomips : %lf\n", &val) == 1) { speed_cycletime = 1e-6 / val; ret = 1; break; } } fclose (fp); } return ret; } /* SunOS /bin/sysinfo prints a line like: cpu0 is a "75 MHz TI,TMS390Z55" CPU */ #if HAVE_POPEN int speed_cpu_frequency_sunos_sysinfo (void) { FILE *fp; char buf[128]; double val; int ret = 0; /* Error messages are sent to /dev/null in case /bin/sysinfo doesn't exist. The brackets are necessary for some shells (eg. ash). */ if ((fp = popen ("(/bin/sysinfo) 2>/dev/null", "r")) != NULL) { while (fgets (buf, sizeof (buf), fp) != NULL) { if (sscanf (buf, " cpu0 is a \"%lf MHz", &val) == 1) { speed_cycletime = 1e-6 / val; ret = 1; break; } } pclose (fp); } return ret; } #endif /* This is for Solaris. "psrinfo" is the command-line interface to processor_info(). "prtconf -vp" gives similar information. */ #if HAVE_PROCESSOR_INFO #include /* for _SC_NPROCESSORS_CONF */ #include /* for processor_info_t */ int speed_cpu_frequency_processor_info (void) { processor_info_t p; int i, n, mhz = 0; n = sysconf (_SC_NPROCESSORS_CONF); for (i = 0; i < n; i++) { if (processor_info (i, &p) != 0) continue; if (p.pi_state != P_ONLINE) continue; if (mhz != 0 && p.pi_clock != mhz) { fprintf (stderr, "speed_cpu_frequency_processor_info(): There's more than one CPU and they have different clock speeds\n"); return 0; } mhz = p.pi_clock; } speed_cycletime = 1.0e-6 / (double) mhz; return 1; } #endif /* Each function returns 1 if it succeeds in setting speed_cycletime, or 0 if not. */ static const struct { int (*fun) _PROTO ((void)); const char *description; } speed_cpu_frequency_table[] = { /* This should be first, so an environment variable can override anything the system gives. */ { speed_cpu_frequency_environment, "environment variable GMP_CPU_FREQUENCY (in Hertz)" }, #if HAVE_SYSCTLBYNAME { speed_cpu_frequency_sysctlbyname, "sysctlbyname() machdep.tsc_freq or machdep.i586_freq" }, #endif #if HAVE_PROCESSOR_INFO { speed_cpu_frequency_processor_info, "processor_info() pi_clock" }, #endif { speed_cpu_frequency_proc_cpuinfo, "linux kernel /proc/cpuinfo file, cpu MHz or bogomips" }, #if HAVE_POPEN { speed_cpu_frequency_sunos_sysinfo, "SunOS /bin/sysinfo program cpu0 output" }, #endif }; int speed_cycletime_init (void) { int i; for (i = 0; i < numberof (speed_cpu_frequency_table); i++) if ((*speed_cpu_frequency_table[i].fun)()) return 1; fprintf (stderr, "Cannot determine CPU frequency, need one of the following\n"); for (i = 0; i < numberof (speed_cpu_frequency_table); i++) fprintf (stderr, "\t- %s\n", speed_cpu_frequency_table[i].description); return 0; } /* ---------------------------------------------------------------------- */ #if SPEED_USE_CYCLECOUNTER const char *speed_time_string = "Time measurements using CPU cycle counter.\n"; /* bigish value because we have a fast timer */ int speed_precision = 10000; double speed_unittime; double speed_cycletime; static int speed_time_initialized = 0; static unsigned speed_starttime_save[2]; /* Knowing the CPU frequency is mandatory, so cycles can be converted to seconds. */ void speed_time_init (void) { if (speed_time_initialized) return; speed_time_initialized = 1; if (!speed_cycletime_init ()) exit (1); speed_unittime = speed_cycletime; } void speed_starttime (void) { if (!speed_time_initialized) speed_time_init (); speed_cyclecounter (speed_starttime_save); } #define M_2POWU ((double) (1L << (BITS_PER_INT-2)) * 4.0) #define M_2POW32 4294967296.0 double speed_endtime (void) { unsigned endtime[2], e0; double t; speed_cyclecounter (endtime); /* This still works even if speed_cyclecounter() puts a value bigger than 32-bits in the low word. The start and end values are allowed to cancel in uints in case a uint is more than the 53 bits that will normally fit in a double. */ e0 = endtime[0] - speed_starttime_save[0]; t = e0 - (e0 > endtime[0] ? M_2POWU : 0); t += (endtime[1] - speed_starttime_save[1]) * M_2POW32; return t * speed_unittime; } #endif /* ---------------------------------------------------------------------- */ #if SPEED_USE_MICROSECOND_GETRUSAGE #include #include #include const char *speed_time_string = "Time measurements using microsecond accurate getrusage.\n"; int speed_precision = 1000; double speed_unittime = 1.0e-6; double speed_cycletime = 1.0; static struct rusage speed_starttime_save; static int speed_time_initialized = 0; void speed_time_init (void) { if (speed_time_initialized) return; speed_time_initialized = 1; speed_cycletime_init (); } void speed_starttime (void) { if (!speed_time_initialized) speed_time_init (); getrusage (0, &speed_starttime_save); } double speed_endtime (void) { struct rusage r; getrusage (0, &r); return TIMEVAL_SECS (&r.ru_utime) - TIMEVAL_SECS (&speed_starttime_save.ru_utime); } #endif /* ---------------------------------------------------------------------- */ #if SPEED_USE_MICROSECOND_GETTIMEOFDAY /* This method is for systems with a microsecond accurate gettimeofday(). A dummy timezone parameter is always given to gettimeofday(), in case it doesn't allow NULL. */ #include const char *speed_time_string = "Time measurements using microsecond accurate gettimeofday.\n"; /* highish value because we have an accurate timer */ int speed_precision = 1000; double speed_unittime = 1.0e-6; double speed_cycletime = 1.0; static struct timeval speed_starttime_save; static int speed_time_initialized = 0; void speed_time_init (void) { if (speed_time_initialized) return; speed_time_initialized = 1; speed_cycletime_init (); } void speed_starttime (void) { struct timezone tz; if (!speed_time_initialized) speed_time_init (); gettimeofday (&speed_starttime_save, &tz); } double speed_endtime (void) { struct timeval t; struct timezone tz; gettimeofday (&t, &tz); return TIMEVAL_SECS (&t) - TIMEVAL_SECS (&speed_starttime_save); } #endif /* ---------------------------------------------------------------------- */ #if SPEED_USE_TMS_UTIME /* You're in trouble if you have to use this method. Speed measurments and threshold tuning are going to take a long time. */ #if STDC_HEADERS #include /* for errno */ #include /* for strerror */ #endif #if HAVE_UNISTD_H #include /* for sysconf */ #endif #include /* for times */ const char *speed_time_string = "Time measurements using tms_utime.\n"; /* lowish default value so we don't take days and days to do tuning */ int speed_precision = 200; double speed_unittime; double speed_cycletime = 1.0; static struct tms speed_starttime_save; static int speed_time_initialized = 0; void speed_time_init (void) { long clk_tck; if (speed_time_initialized) return; speed_time_initialized = 1; speed_cycletime_init (); #if HAVE_SYSCONF clk_tck = sysconf (_SC_CLK_TCK); if (clk_tck == -1L) { fprintf (stderr, "sysconf(_SC_CLK_TCK) not available: %s\n", strerror(errno)); fprintf (stderr, "\tusing CLK_TCK instead\n"); clk_tck = CLK_TCK; } #else clk_tck = CLK_TCK; #endif speed_unittime = 1.0 / (double) clk_tck; } /* Burn up CPU until a times() tms_utime tick boundary. Doing so lets you know a measurement has started on a tick boundary, effectively halving the uncertainty in the measurement. *t1 gets the start times() values the caller should use. */ void times_utime_boundary (struct tms *t1) { struct tms t2; times (&t2); do times (t1); while (t1->tms_utime == t2.tms_utime); } void speed_starttime (void) { if (!speed_time_initialized) speed_time_init (); times_utime_boundary (&speed_starttime_save); } double speed_endtime (void) { struct tms t; times (&t); return (t.tms_utime - speed_starttime_save.tms_utime) * speed_unittime; } #endif