1 /*
2 * Copyright (c) 2003, 2007 Matteo Frigo
3 * Copyright (c) 2003, 2007 Massachusetts Institute of Technology
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining
6 * a copy of this software and associated documentation files (the
7 * "Software"), to deal in the Software without restriction, including
8 * without limitation the rights to use, copy, modify, merge, publish,
9 * distribute, sublicense, and/or sell copies of the Software, and to
10 * permit persons to whom the Software is furnished to do so, subject to
11 * the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be
14 * included in all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
17 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
18 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
19 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE
20 * LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
21 * OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
22 * WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
23 *
24 */
25
26
27 /* machine-dependent cycle counters code. Needs to be inlined. */
28
29 /***************************************************************************/
30 /* To use the cycle counters in your code, simply #include "cycle.h" (this
31 file), and then use the functions/macros:
32
33 ticks getticks(void);
34
35 ticks is an opaque typedef defined below, representing the current time.
36 You extract the elapsed time between two calls to gettick() via:
37
38 double elapsed(ticks t1, ticks t0);
39
40 which returns a double-precision variable in arbitrary units. You
41 are not expected to convert this into human units like seconds; it
42 is intended only for *comparisons* of time intervals.
43
44 (In order to use some of the OS-dependent timer routines like
45 Solaris' gethrtime, you need to paste the autoconf snippet below
46 into your configure.ac file and #include "config.h" before cycle.h,
47 or define the relevant macros manually if you are not using autoconf.)
48 */
49
50 /***************************************************************************/
51 /* This file uses macros like HAVE_GETHRTIME that are assumed to be
52 defined according to whether the corresponding function/type/header
53 is available on your system. The necessary macros are most
54 conveniently defined if you are using GNU autoconf, via the tests:
55
56 dnl ---------------------------------------------------------------------
57
58 AC_C_INLINE
59 AC_HEADER_TIME
60 AC_CHECK_HEADERS([sys/time.h c_asm.h intrinsics.h mach/mach_time.h])
61
62 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
63 #include <sys/time.h>
64 #endif])
65
66 AC_CHECK_FUNCS([gethrtime read_real_time time_base_to_time clock_gettime mach_absolute_time])
67
68 dnl Cray UNICOS _rtc() (real-time clock) intrinsic
69 AC_MSG_CHECKING([for _rtc intrinsic])
70 rtc_ok=yes
71 AC_TRY_LINK([#ifdef HAVE_INTRINSICS_H
72 #include <intrinsics.h>
73 #endif], [_rtc()], [AC_DEFINE(HAVE__RTC,1,[Define if you have the UNICOS _rtc() intrinsic.])], [rtc_ok=no])
74 AC_MSG_RESULT($rtc_ok)
75
76 dnl ---------------------------------------------------------------------
77 */
78
79 /***************************************************************************/
80
81 #if TIME_WITH_SYS_TIME
82 # include <sys/time.h>
83 # include <time.h>
84 #else
85 # if HAVE_SYS_TIME_H
86 # include <sys/time.h>
87 # else
88 # include <time.h>
89 # endif
90 #endif
91
92 #define INLINE_ELAPSED(INL) static INL double elapsed(ticks t1, ticks t0) \
93 { \
94 return (double)t1 - (double)t0; \
95 }
96
97 /*----------------------------------------------------------------*/
98 /* Solaris */
99 #if defined(HAVE_GETHRTIME) && defined(HAVE_HRTIME_T) && !defined(HAVE_TICK_COUNTER)
100 typedef hrtime_t ticks;
101
102 #define getticks gethrtime
103
104 INLINE_ELAPSED(inline)
105
106 #define HAVE_TICK_COUNTER
107 #endif
108
109 /*----------------------------------------------------------------*/
110 /* AIX v. 4+ routines to read the real-time clock or time-base register */
111 #if defined(HAVE_READ_REAL_TIME) && defined(HAVE_TIME_BASE_TO_TIME) && !defined(HAVE_TICK_COUNTER)
112 typedef timebasestruct_t ticks;
113
getticks(void)114 static __inline ticks getticks(void)
115 {
116 ticks t;
117 read_real_time(&t, TIMEBASE_SZ);
118 return t;
119 }
120
elapsed(ticks t1,ticks t0)121 static __inline double elapsed(ticks t1, ticks t0) /* time in nanoseconds */
122 {
123 time_base_to_time(&t1, TIMEBASE_SZ);
124 time_base_to_time(&t0, TIMEBASE_SZ);
125 return (((double)t1.tb_high - (double)t0.tb_high) * 1.0e9 +
126 ((double)t1.tb_low - (double)t0.tb_low));
127 }
128
129 #define HAVE_TICK_COUNTER
130 #endif
131
132 /*----------------------------------------------------------------*/
133 /*
134 * PowerPC ``cycle'' counter using the time base register.
135 */
136 #if ((((defined(__GNUC__) && (defined(__powerpc__) || defined(__ppc__))) || (defined(__MWERKS__) && defined(macintosh)))) || (defined(__IBM_GCC_ASM) && (defined(__powerpc__) || defined(__ppc__)))) && !defined(HAVE_TICK_COUNTER)
137 typedef unsigned long long ticks;
138
getticks(void)139 static __inline__ ticks getticks(void)
140 {
141 unsigned int tbl, tbu0, tbu1;
142
143 do {
144 __asm__ __volatile__ ("mftbu %0" : "=r"(tbu0));
145 __asm__ __volatile__ ("mftb %0" : "=r"(tbl));
146 __asm__ __volatile__ ("mftbu %0" : "=r"(tbu1));
147 } while (tbu0 != tbu1);
148
149 return (((unsigned long long)tbu0) << 32) | tbl;
150 }
151
152 INLINE_ELAPSED(__inline__)
153
154 #define HAVE_TICK_COUNTER
155 #endif
156
157 /* MacOS/Mach (Darwin) time-base register interface (unlike UpTime,
158 from Carbon, requires no additional libraries to be linked). */
159 #if defined(HAVE_MACH_ABSOLUTE_TIME) && defined(HAVE_MACH_MACH_TIME_H) && !defined(HAVE_TICK_COUNTER)
160 #include <mach/mach_time.h>
161 typedef uint64_t ticks;
162 #define getticks mach_absolute_time
163 INLINE_ELAPSED(__inline__)
164 #define HAVE_TICK_COUNTER
165 #endif
166
167 /*----------------------------------------------------------------*/
168 /*
169 * Pentium cycle counter
170 */
171 #if (defined(__GNUC__) || defined(__ICC)) && defined(__i386__) && !defined(HAVE_TICK_COUNTER)
172 typedef unsigned long long ticks;
173
getticks(void)174 static __inline__ ticks getticks(void)
175 {
176 ticks ret;
177
178 __asm__ __volatile__("rdtsc": "=A" (ret));
179 /* no input, nothing else clobbered */
180 return ret;
181 }
182
183 INLINE_ELAPSED(__inline__)
184
185 #define HAVE_TICK_COUNTER
186 #define TIME_MIN 5000.0 /* unreliable pentium IV cycle counter */
187 #endif
188
189 /* Visual C++ -- thanks to Morten Nissov for his help with this */
190 #if _MSC_VER >= 1200 && _M_IX86 >= 500 && !defined(HAVE_TICK_COUNTER)
191 #include <windows.h>
192 typedef LARGE_INTEGER ticks;
193 #define RDTSC __asm __emit 0fh __asm __emit 031h /* hack for VC++ 5.0 */
194
getticks(void)195 static __inline ticks getticks(void)
196 {
197 ticks retval;
198
199 __asm {
200 RDTSC
201 mov retval.HighPart, edx
202 mov retval.LowPart, eax
203 }
204 return retval;
205 }
206
elapsed(ticks t1,ticks t0)207 static __inline double elapsed(ticks t1, ticks t0)
208 {
209 return (double)t1.QuadPart - (double)t0.QuadPart;
210 }
211
212 #define HAVE_TICK_COUNTER
213 #define TIME_MIN 5000.0 /* unreliable pentium IV cycle counter */
214 #endif
215
216 /*----------------------------------------------------------------*/
217 /*
218 * X86-64 cycle counter
219 */
220 #if (defined(__GNUC__) || defined(__ICC) || defined(__SUNPRO_C)) && defined(__x86_64__) && !defined(HAVE_TICK_COUNTER)
221 typedef unsigned long long ticks;
222
getticks(void)223 static __inline__ ticks getticks(void)
224 {
225 unsigned a, d;
226 asm volatile("rdtsc" : "=a" (a), "=d" (d));
227 return ((ticks)a) | (((ticks)d) << 32);
228 }
229
230 INLINE_ELAPSED(__inline__)
231
232 #define HAVE_TICK_COUNTER
233 #endif
234
235 /* PGI compiler, courtesy Cristiano Calonaci, Andrea Tarsi, & Roberto Gori.
236 NOTE: this code will fail to link unless you use the -Masmkeyword compiler
237 option (grrr). */
238 #if defined(__PGI) && defined(__x86_64__) && !defined(HAVE_TICK_COUNTER)
239 typedef unsigned long long ticks;
getticks(void)240 static ticks getticks(void)
241 {
242 asm(" rdtsc; shl $0x20,%rdx; mov %eax,%eax; or %rdx,%rax; ");
243 }
244 INLINE_ELAPSED(__inline__)
245 #define HAVE_TICK_COUNTER
246 #endif
247
248 /* Visual C++, courtesy of Dirk Michaelis */
249 #if _MSC_VER >= 1400 && (defined(_M_AMD64) || defined(_M_X64)) && !defined(HAVE_TICK_COUNTER)
250
251 #include <intrin.h>
252 #pragma intrinsic(__rdtsc)
253 typedef unsigned __int64 ticks;
254 #define getticks __rdtsc
255 INLINE_ELAPSED(__inline)
256
257 #define HAVE_TICK_COUNTER
258 #endif
259
260 /*----------------------------------------------------------------*/
261 /*
262 * IA64 cycle counter
263 */
264
265 /* intel's icc/ecc compiler */
266 #if (defined(__EDG_VERSION) || defined(__ECC)) && defined(__ia64__) && !defined(HAVE_TICK_COUNTER)
267 typedef unsigned long ticks;
268 #include <ia64intrin.h>
269
getticks(void)270 static __inline__ ticks getticks(void)
271 {
272 return __getReg(_IA64_REG_AR_ITC);
273 }
274
275 INLINE_ELAPSED(__inline__)
276
277 #define HAVE_TICK_COUNTER
278 #endif
279
280 /* gcc */
281 #if defined(__GNUC__) && defined(__ia64__) && !defined(HAVE_TICK_COUNTER)
282 typedef unsigned long ticks;
283
getticks(void)284 static __inline__ ticks getticks(void)
285 {
286 ticks ret;
287
288 __asm__ __volatile__ ("mov %0=ar.itc" : "=r"(ret));
289 return ret;
290 }
291
292 INLINE_ELAPSED(__inline__)
293
294 #define HAVE_TICK_COUNTER
295 #endif
296
297 /* HP/UX IA64 compiler, courtesy Teresa L. Johnson: */
298 #if defined(__hpux) && defined(__ia64) && !defined(HAVE_TICK_COUNTER)
299 #include <machine/sys/inline.h>
300 typedef unsigned long ticks;
301
getticks(void)302 static inline ticks getticks(void)
303 {
304 ticks ret;
305
306 ret = _Asm_mov_from_ar (_AREG_ITC);
307 return ret;
308 }
309
310 INLINE_ELAPSED(inline)
311
312 #define HAVE_TICK_COUNTER
313 #endif
314
315 /* Microsoft Visual C++ */
316 #if defined(_MSC_VER) && defined(_M_IA64) && !defined(HAVE_TICK_COUNTER)
317 typedef unsigned __int64 ticks;
318
319 # ifdef __cplusplus
320 extern "C"
321 # endif
322 ticks __getReg(int whichReg);
323 #pragma intrinsic(__getReg)
324
getticks(void)325 static __inline ticks getticks(void)
326 {
327 volatile ticks temp;
328 temp = __getReg(3116);
329 return temp;
330 }
331
332 INLINE_ELAPSED(inline)
333
334 #define HAVE_TICK_COUNTER
335 #endif
336
337 /*----------------------------------------------------------------*/
338 /*
339 * PA-RISC cycle counter
340 */
341 #if defined(__hppa__) || defined(__hppa) && !defined(HAVE_TICK_COUNTER)
342 typedef unsigned long ticks;
343
344 # ifdef __GNUC__
getticks(void)345 static __inline__ ticks getticks(void)
346 {
347 ticks ret;
348
349 __asm__ __volatile__("mfctl 16, %0": "=r" (ret));
350 /* no input, nothing else clobbered */
351 return ret;
352 }
353 # else
354 # include <machine/inline.h>
getticks(void)355 static inline unsigned long getticks(void)
356 {
357 register ticks ret;
358 _MFCTL(16, ret);
359 return ret;
360 }
361 # endif
362
363 INLINE_ELAPSED(inline)
364
365 #define HAVE_TICK_COUNTER
366 #endif
367
368 /*----------------------------------------------------------------*/
369 /* S390, courtesy of James Treacy */
370 #if defined(__GNUC__) && defined(__s390__) && !defined(HAVE_TICK_COUNTER)
371 typedef unsigned long long ticks;
372
getticks(void)373 static __inline__ ticks getticks(void)
374 {
375 ticks cycles;
376 __asm__("stck 0(%0)" : : "a" (&(cycles)) : "memory", "cc");
377 return cycles;
378 }
379
380 INLINE_ELAPSED(__inline__)
381
382 #define HAVE_TICK_COUNTER
383 #endif
384 /*----------------------------------------------------------------*/
385 #if defined(__GNUC__) && defined(__alpha__) && !defined(HAVE_TICK_COUNTER)
386 /*
387 * The 32-bit cycle counter on alpha overflows pretty quickly,
388 * unfortunately. A 1GHz machine overflows in 4 seconds.
389 */
390 typedef unsigned int ticks;
391
getticks(void)392 static __inline__ ticks getticks(void)
393 {
394 unsigned long cc;
395 __asm__ __volatile__ ("rpcc %0" : "=r"(cc));
396 return (cc & 0xFFFFFFFF);
397 }
398
399 INLINE_ELAPSED(__inline__)
400
401 #define HAVE_TICK_COUNTER
402 #endif
403
404 /*----------------------------------------------------------------*/
405 #if defined(__GNUC__) && defined(__sparc_v9__) && !defined(HAVE_TICK_COUNTER)
406 typedef unsigned long ticks;
407
getticks(void)408 static __inline__ ticks getticks(void)
409 {
410 ticks ret;
411 __asm__ __volatile__("rd %%tick, %0" : "=r" (ret));
412 return ret;
413 }
414
415 INLINE_ELAPSED(__inline__)
416
417 #define HAVE_TICK_COUNTER
418 #endif
419
420 /*----------------------------------------------------------------*/
421 #if (defined(__DECC) || defined(__DECCXX)) && defined(__alpha) && defined(HAVE_C_ASM_H) && !defined(HAVE_TICK_COUNTER)
422 # include <c_asm.h>
423 typedef unsigned int ticks;
424
getticks(void)425 static __inline ticks getticks(void)
426 {
427 unsigned long cc;
428 cc = asm("rpcc %v0");
429 return (cc & 0xFFFFFFFF);
430 }
431
432 INLINE_ELAPSED(__inline)
433
434 #define HAVE_TICK_COUNTER
435 #endif
436 /*----------------------------------------------------------------*/
437 /* SGI/Irix */
438 #if defined(HAVE_CLOCK_GETTIME) && defined(CLOCK_SGI_CYCLE) && !defined(HAVE_TICK_COUNTER)
439 typedef struct timespec ticks;
440
getticks(void)441 static inline ticks getticks(void)
442 {
443 struct timespec t;
444 clock_gettime(CLOCK_SGI_CYCLE, &t);
445 return t;
446 }
447
elapsed(ticks t1,ticks t0)448 static inline double elapsed(ticks t1, ticks t0)
449 {
450 return ((double)t1.tv_sec - (double)t0.tv_sec) * 1.0E9 +
451 ((double)t1.tv_nsec - (double)t0.tv_nsec);
452 }
453 #define HAVE_TICK_COUNTER
454 #endif
455
456 /*----------------------------------------------------------------*/
457 /* Cray UNICOS _rtc() intrinsic function */
458 #if defined(HAVE__RTC) && !defined(HAVE_TICK_COUNTER)
459 #ifdef HAVE_INTRINSICS_H
460 # include <intrinsics.h>
461 #endif
462
463 typedef long long ticks;
464
465 #define getticks _rtc
466
467 INLINE_ELAPSED(inline)
468
469 #define HAVE_TICK_COUNTER
470 #endif
471
472 /*----------------------------------------------------------------*/
473 /* MIPS ZBus */
474 #if HAVE_MIPS_ZBUS_TIMER
475 #if defined(__mips__) && !defined(HAVE_TICK_COUNTER)
476 #include <sys/mman.h>
477 #include <unistd.h>
478 #include <fcntl.h>
479
480 typedef uint64_t ticks;
481
getticks(void)482 static inline ticks getticks(void)
483 {
484 static uint64_t* addr = 0;
485
486 if (addr == 0)
487 {
488 uint32_t rq_addr = 0x10030000;
489 int fd;
490 int pgsize;
491
492 pgsize = getpagesize();
493 fd = open ("/dev/mem", O_RDONLY | O_SYNC, 0);
494 if (fd < 0) {
495 perror("open");
496 return NULL;
497 }
498 addr = mmap(0, pgsize, PROT_READ, MAP_SHARED, fd, rq_addr);
499 close(fd);
500 if (addr == (uint64_t *)-1) {
501 perror("mmap");
502 return NULL;
503 }
504 }
505
506 return *addr;
507 }
508
509 INLINE_ELAPSED(inline)
510
511 #define HAVE_TICK_COUNTER
512 #endif
513 #endif /* HAVE_MIPS_ZBUS_TIMER */
514
515