Chris@10: /*
Chris@10:  * Copyright (c) 2003, 2007-11 Matteo Frigo
Chris@10:  * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
Chris@10:  *
Chris@10:  * Permission is hereby granted, free of charge, to any person obtaining
Chris@10:  * a copy of this software and associated documentation files (the
Chris@10:  * "Software"), to deal in the Software without restriction, including
Chris@10:  * without limitation the rights to use, copy, modify, merge, publish,
Chris@10:  * distribute, sublicense, and/or sell copies of the Software, and to
Chris@10:  * permit persons to whom the Software is furnished to do so, subject to
Chris@10:  * the following conditions:
Chris@10:  *
Chris@10:  * The above copyright notice and this permission notice shall be
Chris@10:  * included in all copies or substantial portions of the Software.
Chris@10:  *
Chris@10:  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
Chris@10:  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
Chris@10:  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
Chris@10:  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE
Chris@10:  * LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
Chris@10:  * OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
Chris@10:  * WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
Chris@10:  *
Chris@10:  */
Chris@10: 
Chris@10: 
Chris@10: /* machine-dependent cycle counters code. Needs to be inlined. */
Chris@10: 
Chris@10: /***************************************************************************/
Chris@10: /* To use the cycle counters in your code, simply #include "cycle.h" (this
Chris@10:    file), and then use the functions/macros:
Chris@10: 
Chris@10:                  ticks getticks(void);
Chris@10: 
Chris@10:    ticks is an opaque typedef defined below, representing the current time.
Chris@10:    You extract the elapsed time between two calls to gettick() via:
Chris@10: 
Chris@10:                  double elapsed(ticks t1, ticks t0);
Chris@10: 
Chris@10:    which returns a double-precision variable in arbitrary units.  You
Chris@10:    are not expected to convert this into human units like seconds; it
Chris@10:    is intended only for *comparisons* of time intervals.
Chris@10: 
Chris@10:    (In order to use some of the OS-dependent timer routines like
Chris@10:    Solaris' gethrtime, you need to paste the autoconf snippet below
Chris@10:    into your configure.ac file and #include "config.h" before cycle.h,
Chris@10:    or define the relevant macros manually if you are not using autoconf.)
Chris@10: */
Chris@10: 
Chris@10: /***************************************************************************/
Chris@10: /* This file uses macros like HAVE_GETHRTIME that are assumed to be
Chris@10:    defined according to whether the corresponding function/type/header
Chris@10:    is available on your system.  The necessary macros are most
Chris@10:    conveniently defined if you are using GNU autoconf, via the tests:
Chris@10:    
Chris@10:    dnl ---------------------------------------------------------------------
Chris@10: 
Chris@10:    AC_C_INLINE
Chris@10:    AC_HEADER_TIME
Chris@10:    AC_CHECK_HEADERS([sys/time.h c_asm.h intrinsics.h mach/mach_time.h])
Chris@10: 
Chris@10:    AC_CHECK_TYPE([hrtime_t],[AC_DEFINE(HAVE_HRTIME_T, 1, [Define to 1 if hrtime_t is defined in <sys/time.h>])],,[#if HAVE_SYS_TIME_H
Chris@10: #include <sys/time.h>
Chris@10: #endif])
Chris@10: 
Chris@10:    AC_CHECK_FUNCS([gethrtime read_real_time time_base_to_time clock_gettime mach_absolute_time])
Chris@10: 
Chris@10:    dnl Cray UNICOS _rtc() (real-time clock) intrinsic
Chris@10:    AC_MSG_CHECKING([for _rtc intrinsic])
Chris@10:    rtc_ok=yes
Chris@10:    AC_TRY_LINK([#ifdef HAVE_INTRINSICS_H
Chris@10: #include <intrinsics.h>
Chris@10: #endif], [_rtc()], [AC_DEFINE(HAVE__RTC,1,[Define if you have the UNICOS _rtc() intrinsic.])], [rtc_ok=no])
Chris@10:    AC_MSG_RESULT($rtc_ok)
Chris@10: 
Chris@10:    dnl ---------------------------------------------------------------------
Chris@10: */
Chris@10: 
Chris@10: /***************************************************************************/
Chris@10: 
Chris@10: #if TIME_WITH_SYS_TIME
Chris@10: # include <sys/time.h>
Chris@10: # include <time.h>
Chris@10: #else
Chris@10: # if HAVE_SYS_TIME_H
Chris@10: #  include <sys/time.h>
Chris@10: # else
Chris@10: #  include <time.h>
Chris@10: # endif
Chris@10: #endif
Chris@10: 
Chris@10: #define INLINE_ELAPSED(INL) static INL double elapsed(ticks t1, ticks t0) \
Chris@10: {									  \
Chris@10:      return (double)t1 - (double)t0;					  \
Chris@10: }
Chris@10: 
Chris@10: /*----------------------------------------------------------------*/
Chris@10: /* Solaris */
Chris@10: #if defined(HAVE_GETHRTIME) && defined(HAVE_HRTIME_T) && !defined(HAVE_TICK_COUNTER)
Chris@10: typedef hrtime_t ticks;
Chris@10: 
Chris@10: #define getticks gethrtime
Chris@10: 
Chris@10: INLINE_ELAPSED(inline)
Chris@10: 
Chris@10: #define HAVE_TICK_COUNTER
Chris@10: #endif
Chris@10: 
Chris@10: /*----------------------------------------------------------------*/
Chris@10: /* AIX v. 4+ routines to read the real-time clock or time-base register */
Chris@10: #if defined(HAVE_READ_REAL_TIME) && defined(HAVE_TIME_BASE_TO_TIME) && !defined(HAVE_TICK_COUNTER)
Chris@10: typedef timebasestruct_t ticks;
Chris@10: 
Chris@10: static __inline ticks getticks(void)
Chris@10: {
Chris@10:      ticks t;
Chris@10:      read_real_time(&t, TIMEBASE_SZ);
Chris@10:      return t;
Chris@10: }
Chris@10: 
Chris@10: static __inline double elapsed(ticks t1, ticks t0) /* time in nanoseconds */
Chris@10: {
Chris@10:      time_base_to_time(&t1, TIMEBASE_SZ);
Chris@10:      time_base_to_time(&t0, TIMEBASE_SZ);
Chris@10:      return (((double)t1.tb_high - (double)t0.tb_high) * 1.0e9 + 
Chris@10: 	     ((double)t1.tb_low - (double)t0.tb_low));
Chris@10: }
Chris@10: 
Chris@10: #define HAVE_TICK_COUNTER
Chris@10: #endif
Chris@10: 
Chris@10: /*----------------------------------------------------------------*/
Chris@10: /*
Chris@10:  * PowerPC ``cycle'' counter using the time base register.
Chris@10:  */
Chris@10: #if ((((defined(__GNUC__) && (defined(__powerpc__) || defined(__ppc__))) || (defined(__MWERKS__) && defined(macintosh)))) || (defined(__IBM_GCC_ASM) && (defined(__powerpc__) || defined(__ppc__))))  && !defined(HAVE_TICK_COUNTER)
Chris@10: typedef unsigned long long ticks;
Chris@10: 
Chris@10: static __inline__ ticks getticks(void)
Chris@10: {
Chris@10:      unsigned int tbl, tbu0, tbu1;
Chris@10: 
Chris@10:      do {
Chris@10: 	  __asm__ __volatile__ ("mftbu %0" : "=r"(tbu0));
Chris@10: 	  __asm__ __volatile__ ("mftb %0" : "=r"(tbl));
Chris@10: 	  __asm__ __volatile__ ("mftbu %0" : "=r"(tbu1));
Chris@10:      } while (tbu0 != tbu1);
Chris@10: 
Chris@10:      return (((unsigned long long)tbu0) << 32) | tbl;
Chris@10: }
Chris@10: 
Chris@10: INLINE_ELAPSED(__inline__)
Chris@10: 
Chris@10: #define HAVE_TICK_COUNTER
Chris@10: #endif
Chris@10: 
Chris@10: /* MacOS/Mach (Darwin) time-base register interface (unlike UpTime,
Chris@10:    from Carbon, requires no additional libraries to be linked). */
Chris@10: #if defined(HAVE_MACH_ABSOLUTE_TIME) && defined(HAVE_MACH_MACH_TIME_H) && !defined(HAVE_TICK_COUNTER)
Chris@10: #include <mach/mach_time.h>
Chris@10: typedef uint64_t ticks;
Chris@10: #define getticks mach_absolute_time
Chris@10: INLINE_ELAPSED(__inline__)
Chris@10: #define HAVE_TICK_COUNTER
Chris@10: #endif
Chris@10: 
Chris@10: /*----------------------------------------------------------------*/
Chris@10: /*
Chris@10:  * Pentium cycle counter 
Chris@10:  */
Chris@10: #if (defined(__GNUC__) || defined(__ICC)) && defined(__i386__)  && !defined(HAVE_TICK_COUNTER)
Chris@10: typedef unsigned long long ticks;
Chris@10: 
Chris@10: static __inline__ ticks getticks(void)
Chris@10: {
Chris@10:      ticks ret;
Chris@10: 
Chris@10:      __asm__ __volatile__("rdtsc": "=A" (ret));
Chris@10:      /* no input, nothing else clobbered */
Chris@10:      return ret;
Chris@10: }
Chris@10: 
Chris@10: INLINE_ELAPSED(__inline__)
Chris@10: 
Chris@10: #define HAVE_TICK_COUNTER
Chris@10: #define TIME_MIN 5000.0   /* unreliable pentium IV cycle counter */
Chris@10: #endif
Chris@10: 
Chris@10: /* Visual C++ -- thanks to Morten Nissov for his help with this */
Chris@10: #if _MSC_VER >= 1200 && _M_IX86 >= 500 && !defined(HAVE_TICK_COUNTER)
Chris@10: #include <windows.h>
Chris@10: typedef LARGE_INTEGER ticks;
Chris@10: #define RDTSC __asm __emit 0fh __asm __emit 031h /* hack for VC++ 5.0 */
Chris@10: 
Chris@10: static __inline ticks getticks(void)
Chris@10: {
Chris@10:      ticks retval;
Chris@10: 
Chris@10:      __asm {
Chris@10: 	  RDTSC
Chris@10: 	  mov retval.HighPart, edx
Chris@10: 	  mov retval.LowPart, eax
Chris@10:      }
Chris@10:      return retval;
Chris@10: }
Chris@10: 
Chris@10: static __inline double elapsed(ticks t1, ticks t0)
Chris@10: {  
Chris@10:      return (double)t1.QuadPart - (double)t0.QuadPart;
Chris@10: }  
Chris@10: 
Chris@10: #define HAVE_TICK_COUNTER
Chris@10: #define TIME_MIN 5000.0   /* unreliable pentium IV cycle counter */
Chris@10: #endif
Chris@10: 
Chris@10: /*----------------------------------------------------------------*/
Chris@10: /*
Chris@10:  * X86-64 cycle counter
Chris@10:  */
Chris@10: #if (defined(__GNUC__) || defined(__ICC) || defined(__SUNPRO_C)) && defined(__x86_64__)  && !defined(HAVE_TICK_COUNTER)
Chris@10: typedef unsigned long long ticks;
Chris@10: 
Chris@10: static __inline__ ticks getticks(void)
Chris@10: {
Chris@10:      unsigned a, d; 
Chris@10:      asm volatile("rdtsc" : "=a" (a), "=d" (d)); 
Chris@10:      return ((ticks)a) | (((ticks)d) << 32); 
Chris@10: }
Chris@10: 
Chris@10: INLINE_ELAPSED(__inline__)
Chris@10: 
Chris@10: #define HAVE_TICK_COUNTER
Chris@10: #endif
Chris@10: 
Chris@10: /* PGI compiler, courtesy Cristiano Calonaci, Andrea Tarsi, & Roberto Gori.
Chris@10:    NOTE: this code will fail to link unless you use the -Masmkeyword compiler
Chris@10:    option (grrr). */
Chris@10: #if defined(__PGI) && defined(__x86_64__) && !defined(HAVE_TICK_COUNTER) 
Chris@10: typedef unsigned long long ticks;
Chris@10: static ticks getticks(void)
Chris@10: {
Chris@10:     asm(" rdtsc; shl    $0x20,%rdx; mov    %eax,%eax; or     %rdx,%rax;    ");
Chris@10: }
Chris@10: INLINE_ELAPSED(__inline__)
Chris@10: #define HAVE_TICK_COUNTER
Chris@10: #endif
Chris@10: 
Chris@10: /* Visual C++, courtesy of Dirk Michaelis */
Chris@10: #if _MSC_VER >= 1400 && (defined(_M_AMD64) || defined(_M_X64)) && !defined(HAVE_TICK_COUNTER)
Chris@10: 
Chris@10: #include <intrin.h>
Chris@10: #pragma intrinsic(__rdtsc)
Chris@10: typedef unsigned __int64 ticks;
Chris@10: #define getticks __rdtsc
Chris@10: INLINE_ELAPSED(__inline)
Chris@10: 
Chris@10: #define HAVE_TICK_COUNTER
Chris@10: #endif
Chris@10: 
Chris@10: /*----------------------------------------------------------------*/
Chris@10: /*
Chris@10:  * IA64 cycle counter
Chris@10:  */
Chris@10: 
Chris@10: /* intel's icc/ecc compiler */
Chris@10: #if (defined(__EDG_VERSION) || defined(__ECC)) && defined(__ia64__) && !defined(HAVE_TICK_COUNTER)
Chris@10: typedef unsigned long ticks;
Chris@10: #include <ia64intrin.h>
Chris@10: 
Chris@10: static __inline__ ticks getticks(void)
Chris@10: {
Chris@10:      return __getReg(_IA64_REG_AR_ITC);
Chris@10: }
Chris@10:  
Chris@10: INLINE_ELAPSED(__inline__)
Chris@10:  
Chris@10: #define HAVE_TICK_COUNTER
Chris@10: #endif
Chris@10: 
Chris@10: /* gcc */
Chris@10: #if defined(__GNUC__) && defined(__ia64__) && !defined(HAVE_TICK_COUNTER)
Chris@10: typedef unsigned long ticks;
Chris@10: 
Chris@10: static __inline__ ticks getticks(void)
Chris@10: {
Chris@10:      ticks ret;
Chris@10: 
Chris@10:      __asm__ __volatile__ ("mov %0=ar.itc" : "=r"(ret));
Chris@10:      return ret;
Chris@10: }
Chris@10: 
Chris@10: INLINE_ELAPSED(__inline__)
Chris@10: 
Chris@10: #define HAVE_TICK_COUNTER
Chris@10: #endif
Chris@10: 
Chris@10: /* HP/UX IA64 compiler, courtesy Teresa L. Johnson: */
Chris@10: #if defined(__hpux) && defined(__ia64) && !defined(HAVE_TICK_COUNTER)
Chris@10: #include <machine/sys/inline.h>
Chris@10: typedef unsigned long ticks;
Chris@10: 
Chris@10: static inline ticks getticks(void)
Chris@10: {
Chris@10:      ticks ret;
Chris@10: 
Chris@10:      ret = _Asm_mov_from_ar (_AREG_ITC);
Chris@10:      return ret;
Chris@10: }
Chris@10: 
Chris@10: INLINE_ELAPSED(inline)
Chris@10: 
Chris@10: #define HAVE_TICK_COUNTER
Chris@10: #endif
Chris@10: 
Chris@10: /* Microsoft Visual C++ */
Chris@10: #if defined(_MSC_VER) && defined(_M_IA64) && !defined(HAVE_TICK_COUNTER)
Chris@10: typedef unsigned __int64 ticks;
Chris@10: 
Chris@10: #  ifdef __cplusplus
Chris@10: extern "C"
Chris@10: #  endif
Chris@10: ticks __getReg(int whichReg);
Chris@10: #pragma intrinsic(__getReg)
Chris@10: 
Chris@10: static __inline ticks getticks(void)
Chris@10: {
Chris@10:      volatile ticks temp;
Chris@10:      temp = __getReg(3116);
Chris@10:      return temp;
Chris@10: }
Chris@10: 
Chris@10: INLINE_ELAPSED(inline)
Chris@10: 
Chris@10: #define HAVE_TICK_COUNTER
Chris@10: #endif
Chris@10: 
Chris@10: /*----------------------------------------------------------------*/
Chris@10: /*
Chris@10:  * PA-RISC cycle counter 
Chris@10:  */
Chris@10: #if defined(__hppa__) || defined(__hppa) && !defined(HAVE_TICK_COUNTER)
Chris@10: typedef unsigned long ticks;
Chris@10: 
Chris@10: #  ifdef __GNUC__
Chris@10: static __inline__ ticks getticks(void)
Chris@10: {
Chris@10:      ticks ret;
Chris@10: 
Chris@10:      __asm__ __volatile__("mfctl 16, %0": "=r" (ret));
Chris@10:      /* no input, nothing else clobbered */
Chris@10:      return ret;
Chris@10: }
Chris@10: #  else
Chris@10: #  include <machine/inline.h>
Chris@10: static inline unsigned long getticks(void)
Chris@10: {
Chris@10:      register ticks ret;
Chris@10:      _MFCTL(16, ret);
Chris@10:      return ret;
Chris@10: }
Chris@10: #  endif
Chris@10: 
Chris@10: INLINE_ELAPSED(inline)
Chris@10: 
Chris@10: #define HAVE_TICK_COUNTER
Chris@10: #endif
Chris@10: 
Chris@10: /*----------------------------------------------------------------*/
Chris@10: /* S390, courtesy of James Treacy */
Chris@10: #if defined(__GNUC__) && defined(__s390__) && !defined(HAVE_TICK_COUNTER)
Chris@10: typedef unsigned long long ticks;
Chris@10: 
Chris@10: static __inline__ ticks getticks(void)
Chris@10: {
Chris@10:      ticks cycles;
Chris@10:      __asm__("stck 0(%0)" : : "a" (&(cycles)) : "memory", "cc");
Chris@10:      return cycles;
Chris@10: }
Chris@10: 
Chris@10: INLINE_ELAPSED(__inline__)
Chris@10: 
Chris@10: #define HAVE_TICK_COUNTER
Chris@10: #endif
Chris@10: /*----------------------------------------------------------------*/
Chris@10: #if defined(__GNUC__) && defined(__alpha__) && !defined(HAVE_TICK_COUNTER)
Chris@10: /*
Chris@10:  * The 32-bit cycle counter on alpha overflows pretty quickly, 
Chris@10:  * unfortunately.  A 1GHz machine overflows in 4 seconds.
Chris@10:  */
Chris@10: typedef unsigned int ticks;
Chris@10: 
Chris@10: static __inline__ ticks getticks(void)
Chris@10: {
Chris@10:      unsigned long cc;
Chris@10:      __asm__ __volatile__ ("rpcc %0" : "=r"(cc));
Chris@10:      return (cc & 0xFFFFFFFF);
Chris@10: }
Chris@10: 
Chris@10: INLINE_ELAPSED(__inline__)
Chris@10: 
Chris@10: #define HAVE_TICK_COUNTER
Chris@10: #endif
Chris@10: 
Chris@10: /*----------------------------------------------------------------*/
Chris@10: #if defined(__GNUC__) && defined(__sparc_v9__) && !defined(HAVE_TICK_COUNTER)
Chris@10: typedef unsigned long ticks;
Chris@10: 
Chris@10: static __inline__ ticks getticks(void)
Chris@10: {
Chris@10:      ticks ret;
Chris@10:      __asm__ __volatile__("rd %%tick, %0" : "=r" (ret));
Chris@10:      return ret;
Chris@10: }
Chris@10: 
Chris@10: INLINE_ELAPSED(__inline__)
Chris@10: 
Chris@10: #define HAVE_TICK_COUNTER
Chris@10: #endif
Chris@10: 
Chris@10: /*----------------------------------------------------------------*/
Chris@10: #if (defined(__DECC) || defined(__DECCXX)) && defined(__alpha) && defined(HAVE_C_ASM_H) && !defined(HAVE_TICK_COUNTER)
Chris@10: #  include <c_asm.h>
Chris@10: typedef unsigned int ticks;
Chris@10: 
Chris@10: static __inline ticks getticks(void)
Chris@10: {
Chris@10:      unsigned long cc;
Chris@10:      cc = asm("rpcc %v0");
Chris@10:      return (cc & 0xFFFFFFFF);
Chris@10: }
Chris@10: 
Chris@10: INLINE_ELAPSED(__inline)
Chris@10: 
Chris@10: #define HAVE_TICK_COUNTER
Chris@10: #endif
Chris@10: /*----------------------------------------------------------------*/
Chris@10: /* SGI/Irix */
Chris@10: #if defined(HAVE_CLOCK_GETTIME) && defined(CLOCK_SGI_CYCLE) && !defined(HAVE_TICK_COUNTER)
Chris@10: typedef struct timespec ticks;
Chris@10: 
Chris@10: static inline ticks getticks(void)
Chris@10: {
Chris@10:      struct timespec t;
Chris@10:      clock_gettime(CLOCK_SGI_CYCLE, &t);
Chris@10:      return t;
Chris@10: }
Chris@10: 
Chris@10: static inline double elapsed(ticks t1, ticks t0)
Chris@10: {
Chris@10:      return ((double)t1.tv_sec - (double)t0.tv_sec) * 1.0E9 +
Chris@10: 	  ((double)t1.tv_nsec - (double)t0.tv_nsec);
Chris@10: }
Chris@10: #define HAVE_TICK_COUNTER
Chris@10: #endif
Chris@10: 
Chris@10: /*----------------------------------------------------------------*/
Chris@10: /* Cray UNICOS _rtc() intrinsic function */
Chris@10: #if defined(HAVE__RTC) && !defined(HAVE_TICK_COUNTER)
Chris@10: #ifdef HAVE_INTRINSICS_H
Chris@10: #  include <intrinsics.h>
Chris@10: #endif
Chris@10: 
Chris@10: typedef long long ticks;
Chris@10: 
Chris@10: #define getticks _rtc
Chris@10: 
Chris@10: INLINE_ELAPSED(inline)
Chris@10: 
Chris@10: #define HAVE_TICK_COUNTER
Chris@10: #endif
Chris@10: 
Chris@10: /*----------------------------------------------------------------*/
Chris@10: /* MIPS ZBus */
Chris@10: #if HAVE_MIPS_ZBUS_TIMER
Chris@10: #if defined(__mips__) && !defined(HAVE_TICK_COUNTER)
Chris@10: #include <sys/mman.h>
Chris@10: #include <unistd.h>
Chris@10: #include <fcntl.h>
Chris@10: 
Chris@10: typedef uint64_t ticks;
Chris@10: 
Chris@10: static inline ticks getticks(void)
Chris@10: {
Chris@10:   static uint64_t* addr = 0;
Chris@10: 
Chris@10:   if (addr == 0)
Chris@10:   {
Chris@10:     uint32_t rq_addr = 0x10030000;
Chris@10:     int fd;
Chris@10:     int pgsize;
Chris@10: 
Chris@10:     pgsize = getpagesize();
Chris@10:     fd = open ("/dev/mem", O_RDONLY | O_SYNC, 0);
Chris@10:     if (fd < 0) {
Chris@10:       perror("open");
Chris@10:       return NULL;
Chris@10:     }
Chris@10:     addr = mmap(0, pgsize, PROT_READ, MAP_SHARED, fd, rq_addr);
Chris@10:     close(fd);
Chris@10:     if (addr == (uint64_t *)-1) {
Chris@10:       perror("mmap");
Chris@10:       return NULL;
Chris@10:     }
Chris@10:   }
Chris@10: 
Chris@10:   return *addr;
Chris@10: }
Chris@10: 
Chris@10: INLINE_ELAPSED(inline)
Chris@10: 
Chris@10: #define HAVE_TICK_COUNTER
Chris@10: #endif
Chris@10: #endif /* HAVE_MIPS_ZBUS_TIMER */
Chris@10: