1 // ---------------------------------------------------------------------- 2 // CycleClock 3 // A CycleClock tells you the current time in Cycles. The "time" 4 // is actually time since power-on. This is like time() but doesn't 5 // involve a system call and is much more precise. 6 // 7 // NOTE: Not all cpu/platform/kernel combinations guarantee that this 8 // clock increments at a constant rate or is synchronized across all logical 9 // cpus in a system. 10 // 11 // If you need the above guarantees, please consider using a different 12 // API. There are efforts to provide an interface which provides a millisecond 13 // granularity and implemented as a memory read. A memory read is generally 14 // cheaper than the CycleClock for many architectures. 15 // 16 // Also, in some out of order CPU implementations, the CycleClock is not 17 // serializing. So if you're trying to count at cycles granularity, your 18 // data might be inaccurate due to out of order instruction execution. 19 // ---------------------------------------------------------------------- 20 21 #ifndef BENCHMARK_CYCLECLOCK_H_ 22 #define BENCHMARK_CYCLECLOCK_H_ 23 24 #include <cstdint> 25 26 #include "benchmark/benchmark.h" 27 #include "internal_macros.h" 28 29 #if defined(BENCHMARK_OS_MACOSX) 30 #include <mach/mach_time.h> 31 #endif 32 // For MSVC, we want to use '_asm rdtsc' when possible (since it works 33 // with even ancient MSVC compilers), and when not possible the 34 // __rdtsc intrinsic, declared in <intrin.h>. Unfortunately, in some 35 // environments, <windows.h> and <intrin.h> have conflicting 36 // declarations of some other intrinsics, breaking compilation. 37 // Therefore, we simply declare __rdtsc ourselves. See also 38 // http://connect.microsoft.com/VisualStudio/feedback/details/262047 39 #if defined(COMPILER_MSVC) && !defined(_M_IX86) 40 extern "C" uint64_t __rdtsc(); 41 #pragma intrinsic(__rdtsc) 42 #endif 43 44 #ifndef BENCHMARK_OS_WINDOWS 45 #include <sys/time.h> 46 #include <time.h> 47 #endif 48 49 #ifdef BENCHMARK_OS_EMSCRIPTEN 50 #include <emscripten.h> 51 #endif 52 53 namespace benchmark { 54 // NOTE: only i386 and x86_64 have been well tested. 55 // PPC, sparc, alpha, and ia64 are based on 56 // http://peter.kuscsik.com/wordpress/?p=14 57 // with modifications by m3b. See also 58 // https://setisvn.ssl.berkeley.edu/svn/lib/fftw-3.0.1/kernel/cycle.h 59 namespace cycleclock { 60 // This should return the number of cycles since power-on. Thread-safe. 61 inline BENCHMARK_ALWAYS_INLINE int64_t Now() { 62 #if defined(BENCHMARK_OS_MACOSX) 63 // this goes at the top because we need ALL Macs, regardless of 64 // architecture, to return the number of "mach time units" that 65 // have passed since startup. See sysinfo.cc where 66 // InitializeSystemInfo() sets the supposed cpu clock frequency of 67 // macs to the number of mach time units per second, not actual 68 // CPU clock frequency (which can change in the face of CPU 69 // frequency scaling). Also note that when the Mac sleeps, this 70 // counter pauses; it does not continue counting, nor does it 71 // reset to zero. 72 return mach_absolute_time(); 73 #elif defined(BENCHMARK_OS_EMSCRIPTEN) 74 // this goes above x86-specific code because old versions of Emscripten 75 // define __x86_64__, although they have nothing to do with it. 76 return static_cast<int64_t>(emscripten_get_now() * 1e+6); 77 #elif defined(__i386__) 78 int64_t ret; 79 __asm__ volatile("rdtsc" : "=A"(ret)); 80 return ret; 81 #elif defined(__x86_64__) || defined(__amd64__) 82 uint64_t low, high; 83 __asm__ volatile("rdtsc" : "=a"(low), "=d"(high)); 84 return (high << 32) | low; 85 #elif defined(__powerpc__) || defined(__ppc__) 86 // This returns a time-base, which is not always precisely a cycle-count. 87 int64_t tbl, tbu0, tbu1; 88 asm("mftbu %0" : "=r"(tbu0)); 89 asm("mftb %0" : "=r"(tbl)); 90 asm("mftbu %0" : "=r"(tbu1)); 91 tbl &= -static_cast<int64_t>(tbu0 == tbu1); 92 // high 32 bits in tbu1; low 32 bits in tbl (tbu0 is garbage) 93 return (tbu1 << 32) | tbl; 94 #elif defined(__sparc__) 95 int64_t tick; 96 asm(".byte 0x83, 0x41, 0x00, 0x00"); 97 asm("mov %%g1, %0" : "=r"(tick)); 98 return tick; 99 #elif defined(__ia64__) 100 int64_t itc; 101 asm("mov %0 = ar.itc" : "=r"(itc)); 102 return itc; 103 #elif defined(COMPILER_MSVC) && defined(_M_IX86) 104 // Older MSVC compilers (like 7.x) don't seem to support the 105 // __rdtsc intrinsic properly, so I prefer to use _asm instead 106 // when I know it will work. Otherwise, I'll use __rdtsc and hope 107 // the code is being compiled with a non-ancient compiler. 108 _asm rdtsc 109 #elif defined(COMPILER_MSVC) 110 return __rdtsc(); 111 #elif defined(BENCHMARK_OS_NACL) 112 // Native Client validator on x86/x86-64 allows RDTSC instructions, 113 // and this case is handled above. Native Client validator on ARM 114 // rejects MRC instructions (used in the ARM-specific sequence below), 115 // so we handle it here. Portable Native Client compiles to 116 // architecture-agnostic bytecode, which doesn't provide any 117 // cycle counter access mnemonics. 118 119 // Native Client does not provide any API to access cycle counter. 120 // Use clock_gettime(CLOCK_MONOTONIC, ...) instead of gettimeofday 121 // because is provides nanosecond resolution (which is noticable at 122 // least for PNaCl modules running on x86 Mac & Linux). 123 // Initialize to always return 0 if clock_gettime fails. 124 struct timespec ts = { 0, 0 }; 125 clock_gettime(CLOCK_MONOTONIC, &ts); 126 return static_cast<int64_t>(ts.tv_sec) * 1000000000 + ts.tv_nsec; 127 #elif defined(__aarch64__) 128 // System timer of ARMv8 runs at a different frequency than the CPU's. 129 // The frequency is fixed, typically in the range 1-50MHz. It can be 130 // read at CNTFRQ special register. We assume the OS has set up 131 // the virtual timer properly. 132 int64_t virtual_timer_value; 133 asm volatile("mrs %0, cntvct_el0" : "=r"(virtual_timer_value)); 134 return virtual_timer_value; 135 #elif defined(__ARM_ARCH) 136 // V6 is the earliest arch that has a standard cyclecount 137 // Native Client validator doesn't allow MRC instructions. 138 #if (__ARM_ARCH >= 6) 139 uint32_t pmccntr; 140 uint32_t pmuseren; 141 uint32_t pmcntenset; 142 // Read the user mode perf monitor counter access permissions. 143 asm volatile("mrc p15, 0, %0, c9, c14, 0" : "=r"(pmuseren)); 144 if (pmuseren & 1) { // Allows reading perfmon counters for user mode code. 145 asm volatile("mrc p15, 0, %0, c9, c12, 1" : "=r"(pmcntenset)); 146 if (pmcntenset & 0x80000000ul) { // Is it counting? 147 asm volatile("mrc p15, 0, %0, c9, c13, 0" : "=r"(pmccntr)); 148 // The counter is set up to count every 64th cycle 149 return static_cast<int64_t>(pmccntr) * 64; // Should optimize to << 6 150 } 151 } 152 #endif 153 struct timeval tv; 154 gettimeofday(&tv, nullptr); 155 return static_cast<int64_t>(tv.tv_sec) * 1000000 + tv.tv_usec; 156 #elif defined(__mips__) 157 // mips apparently only allows rdtsc for superusers, so we fall 158 // back to gettimeofday. It's possible clock_gettime would be better. 159 struct timeval tv; 160 gettimeofday(&tv, nullptr); 161 return static_cast<int64_t>(tv.tv_sec) * 1000000 + tv.tv_usec; 162 #else 163 // The soft failover to a generic implementation is automatic only for ARM. 164 // For other platforms the developer is expected to make an attempt to create 165 // a fast implementation and use generic version if nothing better is available. 166 #error You need to define CycleTimer for your OS and CPU 167 #endif 168 } 169 } // end namespace cycleclock 170 } // end namespace benchmark 171 172 #endif // BENCHMARK_CYCLECLOCK_H_ 173
