include/gmx_cyclecounter.h

   1 /* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
   2  *
   3  *
   4  * This file is part of Gromacs        Copyright (c) 1991-2006
   5  * David van der Spoel, Erik Lindahl, Berk Hess, University of Groningen.
   6  *
   7  * This program is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU General Public License
   9  * as published by the Free Software Foundation; either version 2
  10  * of the License, or (at your option) any later version.
  11  *
  12  * To help us fund GROMACS development, we humbly ask that you cite
  13  * the research papers on the package. Check out http://www.gromacs.org
  14  *
  15  * And Hey:
  16  * Gnomes, ROck Monsters And Chili Sauce
  17  */
  18
  19 #ifndef _GMX_CYCLECOUNTER_H_
  20 #define _GMX_CYCLECOUNTER_H_
  21
  22 /*
  23  * define HAVE_RDTSCP to use the serializing rdtscp instruction instead of rdtsc.
  24  * This is only supported on newer Intel/AMD hardware, but provides better accuracy.
  25  */
  26
  27
  28 /** @file gmx_cyclecounter.h
  29  *
  30  *  @brief High-resolution timestamp or CPU clock cycle counters.
  31  *
  32  *  After reading the current value with gmx_cycles_read() you can add or
  33  *  subtract these numbers as normal integers of type gmx_cycles_t.
  34  */
  35
  36 #ifdef _MSC_VER
  37 #include <intrin.h>
  38 #endif
  39
  40 #ifdef __cplusplus
  41 extern "C"
  42 {
  43 #endif
  44 #if 0
  45 } /* fixes auto-indentation problems */
  46 #endif
  47
  48
  49
  50 /* Minor implementation note:
  51  *
  52  * I like to use these counters in other programs too, so to avoid making
  53  * it dependent on other Gromacs definitions I use the #ifdef's to set
  54  * architecture-specific inline macros instead of using gmx_inline from
  55  * gmx_types.h /Erik 2005-12-10
  56  */
  57
  58 #if ((defined(__GNUC__) || defined(__INTEL_COMPILER) || defined(__PATHSCALE__) || defined(__PGIC__)) && \
  59      (defined(__i386__) || defined(__x86_64__)))
  60 /* x86 or x86-64 with GCC inline assembly */
  61 typedef unsigned long long
  62 gmx_cycles_t;
  63
  64 #elif defined(_MSC_VER)
  65 #include <windows.h>
  66 typedef __int64
  67 gmx_cycles_t;
  68
  69 #elif (defined(__hpux) || defined(__HP_cc)) && defined(__ia64)
  70 /* HP compiler on ia64 */
  71 #include <machine/sys/inline.h>
  72 typedef unsigned long
  73 gmx_cycles_t;
  74
  75 #elif (defined(__INTEL_COMPILER) || defined(__ECC)) && defined(__ia64__)
  76 /* Intel compiler on ia64 */
  77 #include <ia64intrin.h>
  78 typedef unsigned long
  79 gmx_cycles_t;
  80
  81 #elif defined(__GNUC__) && defined(__ia64__)
  82 /* ia64 with GCC inline assembly */
  83 typedef unsigned long
  84 gmx_cycles_t;
  85
  86 #elif ((defined(__hppa__) || defined(__hppa)) && defined (__GNUC__))
  87 /* HP PA-RISC, inline asm with gcc */
  88 typedef unsigned long
  89 gmx_cycles_t;
  90
  91 #elif ((defined(__hppa__) || defined(__hppa)) && defined (__hpux))
  92 /* HP PA-RISC, instruction when using HP compiler */
  93 #include <machine/inline.h>
  94 typedef unsigned long
  95 gmx_cycles_t;
  96
  97 #elif defined(__GNUC__) && defined(__s390__)
  98 /* S390, taken from FFTW who got it from James Treacy */
  99 typedef unsigned long long
 100 gmx_cycles_t;
 101
 102 #elif defined(__GNUC__) && defined(__alpha__)
 103 /* gcc inline assembly on alpha CPUs */
 104 typedef unsigned long
 105 gmx_cycles_t;
 106
 107 #elif defined(__GNUC__) && defined(__sparc_v9__)
 108 /* gcc inline assembly on sparc v9 */
 109 typedef unsigned long
 110 gmx_cycles_t;
 111
 112 #elif defined(__DECC) && defined(__alpha)
 113 /* Digital GEM C compiler on alpha */
 114 #include <c_asm.h>
 115 typedef unsigned long
 116 gmx_cycles_t;
 117
 118 #elif (defined(__sgi) && defined(CLOCK_SGI_CYCLE))
 119 /* Irix compilers on SGI hardware. Get nanoseconds from struct timespec */
 120 typedef unsigned long long
 121 gmx_cycles_t;
 122
 123 #elif (defined(__SVR4) && defined (__SUNPRO_CC))
 124 /* Solaris high-resolution timers */
 125 typedef hrtime_t
 126 gmx_cycles_t;
 127
 128 #elif defined(__xlC__) && defined (_AIX)
 129 /* AIX compilers */
 130 #include <sys/time.h>
 131 #include <sys/systemcfg.h>
 132 typedef unsigned long long
 133 gmx_cycles_t;
 134
 135 #elif ( ( defined(__GNUC__) || defined(__IBM_GCC_ASM) || defined(__IBM_STDCPP_ASM) ) && \
 136         ( defined(__powerpc__) || defined(__ppc__) ) )
 137 /* PowerPC using gcc inline assembly (also works on xlc>=7.0 with -qasm=gcc) */
 138 typedef unsigned long long
 139 gmx_cycles_t;
 140
 141 #elif (defined(__MWERKS__) && (defined(MAC) || defined(macintosh)))
 142 /* Metrowerks on macintosh */
 143 typedef unsigned long long
 144 gmx_cycles_t;
 145
 146 #elif defined(__sun) && defined(__sparcv9)
 147
 148 typedef unsigned long
 149 gmx_cycles_t;
 150
 151 #else
 152 /*! \brief Integer-like datatype for cycle counter values
 153  *
 154  *  Depending on your system this will usually be something like long long,
 155  *  or a special cycle datatype from the system header files. It is NOT
 156  *  necessarily real processor cycles - many systems count in nanoseconds
 157  *  or a special external time register at fixed frequency (not the CPU freq.)
 158  *
 159  *  You can subtract or add gmx_cycle_t types just as normal integers, and if
 160  *  you run the calibration routine you can also multiply it with a factor to
 161  *  translate the cycle data to seconds.
 162  */
 163 typedef long
 164 gmx_cycles_t;
 165
 166 #endif
 167
 168
 169
 170 /*! \brief Check if high-resolution cycle counters are available
 171  *
 172  *  Not all architectures provide any way to read timestep counters
 173  *  in the CPU, and on some it is broken. Although we refer to it
 174  *  as cycle counters, it is not necessarily given in units of
 175  *  cycles.
 176  *
 177  *  If you notice that system is missing, implement support for it,
 178  *  find out how to detect the system during preprocessing, and send us a
 179  *  patch.
 180  *
 181  *  \return 1 if cycle counters are available, 0 if not.
 182  *
 183  * \note This functions not need to be in the header for performance
 184  *       reasons, but it is very important that we get exactly the
 185  *       same detection as for gmx_cycles_read() routines. If you
 186  *       compile the library with one compiler, and then use a different
 187  *       one when later linking to the library it might happen that the
 188  *       library supports cyclecounters but not the headers, or vice versa.
 189  */
 190 #if ((defined(__GNUC__) || defined(__INTEL_COMPILER) || defined(__PATHSCALE__) || defined(__PGIC__)) && \
 191      (defined(__i386__) || defined(__x86_64__)))
 192 static __inline__ int gmx_cycles_have_counter(void)
 193 {
 194         /* x86 or x86-64 with GCC inline assembly - pentium TSC register */
 195         return 1;
 196 }
 197 #elif (defined(_MSC_VER))
 198 static __inline int gmx_cycles_have_counter(void)
 199 {
 200         return 1;
 201 }
 202 #elif (defined(__hpux) || defined(__HP_cc)) && defined(__ia64)
 203 static inline int gmx_cycles_have_counter(void)
 204 {
 205         /* HP compiler on ia64, use special instruction to read ITC */
 206         return 1;
 207 }
 208 #elif (defined(__INTEL_COMPILER) || defined(__ECC)) && defined(__ia64__)
 209 static __inline__ int gmx_cycles_have_counter(void)
 210 {
 211         /* Intel compiler on ia64, use special instruction to read ITC */
 212         return 1;
 213 }
 214 #elif defined(__GNUC__) && defined(__ia64__)
 215 static __inline__ int gmx_cycles_have_counter(void)
 216 {
 217         /* AMD64 with GCC inline assembly - TSC register */
 218         return 1;
 219 }
 220 #elif ((defined(__hppa__) || defined(__hppa)) && defined (__GNUC__))
 221 static __inline__ int gmx_cycles_have_counter(void)
 222 {
 223         /* HP PA-RISC, inline asm with gcc */
 224         return 1;
 225 }
 226 #elif ((defined(__hppa__) || defined(__hppa)) && defined (__hpux))
 227 static inline int gmx_cycles_have_counter(void)
 228 {
 229         /* HP PA-RISC, instruction when using HP compiler */
 230         return 1;
 231 }
 232 #elif defined(__GNUC__) && defined(__s390__)
 233 static __inline__ int gmx_cycles_have_counter(void)
 234 {
 235     /* S390, taken from FFTW who got it from James Treacy */
 236         return 1;
 237 }
 238 #elif defined(__GNUC__) && defined(__alpha__)
 239 static __inline__ int gmx_cycles_have_counter(void)
 240 {
 241     /* gcc inline assembly on alpha CPUs */
 242     return 1;
 243 }
 244 #elif defined(__GNUC__) && defined(__sparc_v9__)
 245 static __inline__ int gmx_cycles_have_counter(void)
 246 {
 247     /* gcc inline assembly on sparc v9 */
 248     return 1;
 249 }
 250 #elif defined(__DECC) && defined(__alpha)
 251 static __inline int gmx_cycles_have_counter(void)
 252 {
 253     /* Digital GEM C compiler on alpha */
 254     return 1;
 255 }
 256 #elif (defined(__sgi) && defined(CLOCK_SGI_CYCLE))
 257 static __inline int gmx_cycles_have_counter(void)
 258 {
 259     /* Irix compilers on SGI hardware */
 260     return 1;
 261 }
 262 #elif (defined(__SVR4) && defined (__SUNPRO_CC))
 263 static inline int gmx_cycles_have_counter(void)
 264 {
 265     /* Solaris high-resolution timers */
 266     return 1;
 267 }
 268 #elif defined(__xlC__) && defined (_AIX)
 269 static inline int gmx_cycles_have_counter(void)
 270 {
 271     /* AIX compilers */
 272     return 1;
 273 }
 274 #elif ( ( defined(__GNUC__) || defined(__IBM_GCC_ASM) || defined(__IBM_STDCPP_ASM) ) && \
 275         ( defined(__powerpc__) || defined(__ppc__) ) )
 276 static __inline__ int gmx_cycles_have_counter(void)
 277 {
 278     /* PowerPC using gcc inline assembly (and xlc>=7.0 with -qasm=gcc) */
 279     return 1;
 280 }
 281 #elif (defined(__MWERKS__) && (defined(MAC) || defined(macintosh)))
 282 static __inline__ int gmx_cycles_have_counter(void)
 283 {
 284     /* Metrowerks on macintosh */
 285     return 1;
 286 }
 287 #elif defined(__sun) && defined(__sparcv9)
 288
 289 static __inline__ int gmx_cycles_have_counter(void)
 290 {
 291     /* Solaris on SPARC*/
 292     return 1;
 293 }
 294 #else
 295 static int gmx_cycles_have_counter(void)
 296 {
 297     /* No cycle counter that we know of on this system */
 298     return 0;
 299 }
 300 #endif
 301
 302
 303
 304
 305
 306
 307 /*! \brief Read CPU cycle counter
 308 *
 309 *  This routine returns an abstract datatype containing a
 310 *  cycle counter timestamp.
 311 *
 312 *  \return Opaque data corresponding to a cycle reading.
 313 *
 314 *  Please note that on most systems it takes several cycles
 315 *  to read and return the cycle counters. If you are measuring
 316 *  small intervals, you can compensate for this time by calling
 317 *  the routine twice and calculating what the difference is.
 318 *  Subtract this from your other measurements to get an accurate result.
 319 *
 320 *  Use gmx_cycles_difference() to get a real number corresponding to
 321 *  the difference between two gmx_cycles_t values returned from this
 322 *  routine.
 323 */
 324 #if ((defined(__GNUC__) || defined(__INTEL_COMPILER) || defined(__PATHSCALE__) || defined(__PGIC__)) && \
 325      (defined(__i386__) || defined(__x86_64__)))
 326 static __inline__ gmx_cycles_t gmx_cycles_read(void)
 327 {
 328     /* x86 with GCC inline assembly - pentium TSC register */
 329     gmx_cycles_t   cycle;
 330     unsigned       low,high;
 331
 332 #ifdef HAVE_RDTSCP
 333     __asm__ __volatile__("rdtscp" : "=a" (low), "=d" (high) :: "ecx" );
 334 #else
 335     __asm__ __volatile__("rdtsc" : "=a" (low), "=d" (high));
 336 #endif
 337
 338     cycle = ((unsigned long long)low) | (((unsigned long long)high)<<32);
 339
 340     return cycle;
 341 }
 342 #elif defined(_MSC_VER)
 343 static __inline gmx_cycles_t gmx_cycles_read(void)
 344 {
 345 #ifdef HAVE_RDTSCP
 346     unsigned int ui;
 347     return __rdtscp(&ui);
 348 #else
 349         return __rdtsc();
 350 #endif
 351 }
 352 #elif (defined(__hpux) || defined(__HP_cc)) && defined(__ia64)
 353 static inline gmx_cycles_t gmx_cycles_read(void)
 354 {
 355     /* HP compiler on ia64 */
 356     gmx_cycles_t ret;
 357     ret = _Asm_mov_from_ar (_AREG_ITC);
 358     return ret;
 359 }
 360 #elif (defined(__INTEL_COMPILER) && defined(__ia64__))
 361 static __inline__ gmx_cycles_t gmx_cycles_read(void)
 362 {
 363     /* Intel compiler on ia64 */
 364     return __getReg(_IA64_REG_AR_ITC);
 365 }
 366 #elif defined(__GNUC__) && defined(__ia64__)
 367 static __inline__ gmx_cycles_t gmx_cycles_read(void)
 368 {
 369     /* ia64 with GCC inline assembly */
 370     gmx_cycles_t ret;
 371     __asm__ __volatile__ ("mov %0=ar.itc" : "=r"(ret));
 372     return ret;
 373 }
 374 #elif ((defined(__hppa__) || defined(__hppa)) && defined (__GNUC__))
 375 static __inline__ gmx_cycles_t gmx_cycles_read(void)
 376 {
 377     /* HP PA-RISC, inline asm with gcc */
 378     gmx_cycles_t ret;
 379     __asm__ __volatile__("mfctl 16, %0": "=r" (ret));
 380     /* no input, nothing else clobbered */
 381     return ret;
 382 }
 383 #elif ((defined(__hppa__) || defined(__hppa)) && defined (__hpux))
 384 static inline gmx_cycles_t gmx_cycles_read(void)
 385 {
 386     /* HP PA-RISC, instruction when using HP compiler */
 387     gmx_cycles_t ret;
 388     _MFCTL(16, ret);
 389     return ret;
 390 }
 391 #elif defined(__GNUC__) && defined(__s390__)
 392 static __inline__ gmx_cycles_t gmx_cycles_read(void)
 393 {
 394     /* S390, taken from FFTW who got it from James Treacy */
 395     gmx_cycles_t cycle;
 396     __asm__("stck 0(%0)" : : "a" (&(cycle)) : "memory", "cc");
 397     return cycle;
 398 }
 399 #elif defined(__GNUC__) && defined(__alpha__)
 400 static __inline__ gmx_cycles_t gmx_cycles_read(void)
 401 {
 402     /* gcc inline assembly on alpha CPUs */
 403     unsigned long cycle;
 404     __asm__ __volatile__ ("rpcc %0" : "=r"(cycle));
 405     return (cycle & 0xFFFFFFFF);
 406 }
 407 #elif defined(__GNUC__) && defined(__sparc_v9__)
 408 static __inline__ gmx_cycles_t gmx_cycles_read(void)
 409 {
 410     /* gcc inline assembly on sparc v9 */
 411     unsigned long ret;
 412     __asm__("rd %%tick, %0" : "=r" (ret));
 413     return ret;
 414 }
 415 #elif defined(__DECC) && defined(__alpha)
 416 static __inline gmx_cycles_t gmx_cycles_read(void)
 417 {
 418     /* Digital GEM C compiler on alpha */
 419     unsigned long cycle;
 420     cycle = asm("rpcc %v0");
 421     return (cycle & 0xFFFFFFFF);
 422 }
 423 #elif (defined(__sgi) && defined(CLOCK_SGI_CYCLE))
 424 static __inline gmx_cycles_t gmx_cycles_read(void)
 425 {
 426     /* Irix compilers on SGI hardware */
 427     struct timespec t;
 428     clock_gettime(CLOCK_SGI_CYCLE, &t);
 429     /* Return the number of nanoseconds, so we can subtract/add */
 430     return ((unsigned long long)t.tv_sec)*1000000000+
 431         (unsigned long long)t.tv_nsec;
 432 }
 433 #elif (defined(__SVR4) && defined (__SUNPRO_CC))
 434 static inline gmx_cycles_t gmx_cycles_read(void)
 435 {
 436     /* Solaris high-resolution timers */
 437     return gethrtime();
 438 }
 439 #elif defined(__xlC__) && defined (_AIX)
 440 static inline gmx_cycles_t gmx_cycles_read(void)
 441 {
 442     /* AIX compilers. Inline the calculation instead of using library functions */
 443     timebasestruct_t t1;
 444     read_real_time(&t1, TIMEBASE_SZ);
 445     /* POWER returns real time (seconds + nanoseconds),
 446      * POWER_PC returns high/low 32 bits of a counter.
 447      */
 448     if(t1.flag==RTC_POWER_PC)
 449     {
 450         return ((gmx_cycles_t)t1.tb_high)<<32 | (gmx_cycles_t)t1.tb_low;
 451     }
 452     else
 453     {
 454         return ((gmx_cycles_t)t1.tb_high)*1000000000+(gmx_cycles_t)t1.tb_low;
 455     }
 456 }
 457 #elif ( ( defined(__GNUC__) || defined(__IBM_GCC_ASM) || defined(__IBM_STDCPP_ASM) ) && \
 458         ( defined(__powerpc__) || defined(__ppc__) ) )
 459 static __inline__ gmx_cycles_t gmx_cycles_read(void)
 460 {
 461     /* PowerPC using gcc inline assembly (and xlC>=7.0 with -qasm=gcc) */
 462     unsigned long low, high1, high2;
 463     do
 464     {
 465         __asm__ __volatile__ ("mftbu %0" : "=r" (high1) : );
 466         __asm__ __volatile__ ("mftb %0" : "=r" (low) : );
 467         __asm__ __volatile__ ("mftbu %0" : "=r" (high2) : );
 468     }
 469     while (high1 != high2);
 470
 471     return (((gmx_cycles_t)high2) << 32) | (gmx_cycles_t)low;
 472 }
 473 #elif (defined(__MWERKS__) && (defined(MAC) || defined(macintosh)))
 474 static __inline__ gmx_cycles_t gmx_cycles_read(void)
 475 {
 476     /* Metrowerks on macintosh */
 477     unsigned int long low, high1, high2;
 478     do
 479     {
 480         __asm__ __volatile__ ("mftbu %0" : "=r" (high1) : );
 481         __asm__ __volatile__ ("mftb %0" : "=r" (low) : );
 482         __asm__ __volatile__ ("mftbu %0" : "=r" (high2) : );
 483     }
 484     while (high1 != high2);
 485
 486     return (((gmx_cycles_t)high2) << 32) | (gmx_cycles_t)low;
 487 }
 488 #elif defined(__sun) && defined(__sparcv9)
 489
 490 static __inline__ gmx_cycles_t gmx_cycles_read(void)
 491 {
 492      gmx_cycles_t ret;
 493      __asm__ __volatile__("rd %%tick, %0" : "=r" (ret));
 494      return ret;
 495 }
 496
 497 #else
 498 static gmx_cycles_t gmx_cycles_read(void)
 499 {
 500     return 0;
 501 }
 502 #endif
 503
 504
 505
 506
 507
 508
 509
 510
 511 /*! \brief Calculate number of seconds per cycle tick on host
 512 *
 513 *  This routine runs a timer loop to calibrate the number of
 514 *  seconds per the units returned from gmx_cycles_difference()
 515 *
 516 *  To calculate the time used, call gmx_cycles_read() twice,
 517 *  and then use this routine to calculate the difference as a double
 518 *  precision floating-point number.
 519 *
 520 *  \param  sampletime Minimum number of seconds to sample.
 521 *          One second should give you a reasonably accurate calibration.
 522 *  \return Number of seconds per cycle unit. If it is not possible to
 523 *          calculate on this system (for whatever reason) the return value
 524 *          will be -1, so check that it is positive before using it.
 525 */
 526 double
 527 gmx_cycles_calibrate(double sampletime);
 528
 529
 530 #ifdef __cplusplus
 531 }
 532 #endif
 533
 534
 535
 536 #endif /* _GMX_CYCLECOUNTER_H_ */
 537