diff options
Diffstat (limited to 'indra')
| -rwxr-xr-x | indra/llcommon/llfasttimer.h | 151 | ||||
| -rwxr-xr-x[-rw-r--r--] | indra/llcommon/llfasttimer_class.cpp | 164 |
2 files changed, 167 insertions, 148 deletions
diff --git a/indra/llcommon/llfasttimer.h b/indra/llcommon/llfasttimer.h index 5c2df877b0..c177027f4e 100755 --- a/indra/llcommon/llfasttimer.h +++ b/indra/llcommon/llfasttimer.h @@ -27,155 +27,10 @@ #ifndef LL_FASTTIMER_H #define LL_FASTTIMER_H +// Temporarily(?) de-inlined these functions to simplify diagnosis of problems. +// Implementation of getCPUClockCount32() and getCPUClockCount64 are now in llfastertimer_class.cpp. + // pull in the actual class definition #include "llfasttimer_class.h" -// -// Important note: These implementations must be FAST! -// - -#if LL_WINDOWS -// -// Windows implementation of CPU clock -// - -// -// NOTE: put back in when we aren't using platform sdk anymore -// -// because MS has different signatures for these functions in winnt.h -// need to rename them to avoid conflicts -//#define _interlockedbittestandset _renamed_interlockedbittestandset -//#define _interlockedbittestandreset _renamed_interlockedbittestandreset -//#include <intrin.h> -//#undef _interlockedbittestandset -//#undef _interlockedbittestandreset - -//inline U32 LLFastTimer::getCPUClockCount32() -//{ -// U64 time_stamp = __rdtsc(); -// return (U32)(time_stamp >> 8); -//} -// -//// return full timer value, *not* shifted by 8 bits -//inline U64 LLFastTimer::getCPUClockCount64() -//{ -// return __rdtsc(); -//} - -// shift off lower 8 bits for lower resolution but longer term timing -// on 1Ghz machine, a 32-bit word will hold ~1000 seconds of timing -#ifdef USE_RDTSC -inline U32 LLFastTimer::getCPUClockCount32() -{ - U32 ret_val; - __asm - { - _emit 0x0f - _emit 0x31 - shr eax,8 - shl edx,24 - or eax, edx - mov dword ptr [ret_val], eax - } - return ret_val; -} - -// return full timer value, *not* shifted by 8 bits -inline U64 LLFastTimer::getCPUClockCount64() -{ - U64 ret_val; - __asm - { - _emit 0x0f - _emit 0x31 - mov eax,eax - mov edx,edx - mov dword ptr [ret_val+4], edx - mov dword ptr [ret_val], eax - } - return ret_val; -} -#else -LL_COMMON_API U64 get_clock_count(); // in lltimer.cpp -// These use QueryPerformanceCounter, which is arguably fine and also works on amd architectures. -inline U32 LLFastTimer::getCPUClockCount32() -{ - return (U32)(get_clock_count()>>8); -} - -inline U64 LLFastTimer::getCPUClockCount64() -{ - return get_clock_count(); -} -#endif - -#endif - - -#if (LL_LINUX || LL_SOLARIS) && !(defined(__i386__) || defined(__amd64__)) -// -// Linux and Solaris implementation of CPU clock - non-x86. -// This is accurate but SLOW! Only use out of desperation. -// -// Try to use the MONOTONIC clock if available, this is a constant time counter -// with nanosecond resolution (but not necessarily accuracy) and attempts are -// made to synchronize this value between cores at kernel start. It should not -// be affected by CPU frequency. If not available use the REALTIME clock, but -// this may be affected by NTP adjustments or other user activity affecting -// the system time. -inline U64 LLFastTimer::getCPUClockCount64() -{ - struct timespec tp; - -#ifdef CLOCK_MONOTONIC // MONOTONIC supported at build-time? - if (-1 == clock_gettime(CLOCK_MONOTONIC,&tp)) // if MONOTONIC isn't supported at runtime then ouch, try REALTIME -#endif - clock_gettime(CLOCK_REALTIME,&tp); - - return (tp.tv_sec*LLFastTimer::sClockResolution)+tp.tv_nsec; -} - -inline U32 LLFastTimer::getCPUClockCount32() -{ - return (U32)(LLFastTimer::getCPUClockCount64() >> 8); -} -#endif // (LL_LINUX || LL_SOLARIS) && !(defined(__i386__) || defined(__amd64__)) - - -#if (LL_LINUX || LL_SOLARIS || LL_DARWIN) && (defined(__i386__) || defined(__amd64__)) -// -// Mac+Linux+Solaris FAST x86 implementation of CPU clock -inline U32 LLFastTimer::getCPUClockCount32() -{ - U64 x; - __asm__ volatile (".byte 0x0f, 0x31": "=A"(x)); - return (U32)(x >> 8); -} - -inline U64 LLFastTimer::getCPUClockCount64() -{ - U64 x; - __asm__ volatile (".byte 0x0f, 0x31": "=A"(x)); - return x; -} -#endif - - -#if ( LL_DARWIN && !(defined(__i386__) || defined(__amd64__))) -// -// Mac PPC (deprecated) implementation of CPU clock -// -// Just use gettimeofday implementation for now - -inline U32 LLFastTimer::getCPUClockCount32() -{ - return (U32)(get_clock_count()>>8); -} - -inline U64 LLFastTimer::getCPUClockCount64() -{ - return get_clock_count(); -} -#endif - #endif // LL_LLFASTTIMER_H diff --git a/indra/llcommon/llfasttimer_class.cpp b/indra/llcommon/llfasttimer_class.cpp index c45921cdec..a3e006d70b 100644..100755 --- a/indra/llcommon/llfasttimer_class.cpp +++ b/indra/llcommon/llfasttimer_class.cpp @@ -35,7 +35,9 @@ #include <boost/bind.hpp> + #if LL_WINDOWS +#include "lltimer.h" #elif LL_LINUX || LL_SOLARIS #include <sys/time.h> #include <sched.h> @@ -481,6 +483,19 @@ void LLFastTimer::NamedTimer::resetFrame() { if (sLog) { //output current frame counts to performance log + + static S32 call_count = 0; + if (call_count % 100 == 0) + { + llinfos << "countsPerSecond (32 bit): " << countsPerSecond() << llendl; + llinfos << "get_clock_count (64 bit): " << get_clock_count() << llendl; + llinfos << "LLProcessorInfo().getCPUFrequency() " << LLProcessorInfo().getCPUFrequency() << llendl; + llinfos << "getCPUClockCount32() " << getCPUClockCount32() << llendl; + llinfos << "getCPUClockCount64() " << getCPUClockCount64() << llendl; + llinfos << "elapsed sec " << ((F64)getCPUClockCount64())/((F64)LLProcessorInfo().getCPUFrequency()*1000000.0) << llendl; + } + call_count++; + F64 iclock_freq = 1000.0 / countsPerSecond(); // good place to calculate clock frequency F64 total_time = 0; @@ -762,3 +777,152 @@ LLFastTimer::LLFastTimer(LLFastTimer::FrameState* state) ////////////////////////////////////////////////////////////////////////////// +// +// Important note: These implementations must be FAST! +// + + +#if LL_WINDOWS +// +// Windows implementation of CPU clock +// + +// +// NOTE: put back in when we aren't using platform sdk anymore +// +// because MS has different signatures for these functions in winnt.h +// need to rename them to avoid conflicts +//#define _interlockedbittestandset _renamed_interlockedbittestandset +//#define _interlockedbittestandreset _renamed_interlockedbittestandreset +//#include <intrin.h> +//#undef _interlockedbittestandset +//#undef _interlockedbittestandreset + +//inline U32 LLFastTimer::getCPUClockCount32() +//{ +// U64 time_stamp = __rdtsc(); +// return (U32)(time_stamp >> 8); +//} +// +//// return full timer value, *not* shifted by 8 bits +//inline U64 LLFastTimer::getCPUClockCount64() +//{ +// return __rdtsc(); +//} + +// shift off lower 8 bits for lower resolution but longer term timing +// on 1Ghz machine, a 32-bit word will hold ~1000 seconds of timing +#ifdef USE_RDTSC +inline U32 LLFastTimer::getCPUClockCount32() +{ + U32 ret_val; + __asm + { + _emit 0x0f + _emit 0x31 + shr eax,8 + shl edx,24 + or eax, edx + mov dword ptr [ret_val], eax + } + return ret_val; +} + +// return full timer value, *not* shifted by 8 bits +inline U64 LLFastTimer::getCPUClockCount64() +{ + U64 ret_val; + __asm + { + _emit 0x0f + _emit 0x31 + mov eax,eax + mov edx,edx + mov dword ptr [ret_val+4], edx + mov dword ptr [ret_val], eax + } + return ret_val; +} +#else +//LL_COMMON_API U64 get_clock_count(); // in lltimer.cpp +// These use QueryPerformanceCounter, which is arguably fine and also works on amd architectures. +inline U32 LLFastTimer::getCPUClockCount32() +{ + return (U32)(get_clock_count()>>8); +} + +inline U64 LLFastTimer::getCPUClockCount64() +{ + return get_clock_count(); +} +#endif + +#endif + + +#if (LL_LINUX || LL_SOLARIS) && !(defined(__i386__) || defined(__amd64__)) +// +// Linux and Solaris implementation of CPU clock - non-x86. +// This is accurate but SLOW! Only use out of desperation. +// +// Try to use the MONOTONIC clock if available, this is a constant time counter +// with nanosecond resolution (but not necessarily accuracy) and attempts are +// made to synchronize this value between cores at kernel start. It should not +// be affected by CPU frequency. If not available use the REALTIME clock, but +// this may be affected by NTP adjustments or other user activity affecting +// the system time. +inline U64 LLFastTimer::getCPUClockCount64() +{ + struct timespec tp; + +#ifdef CLOCK_MONOTONIC // MONOTONIC supported at build-time? + if (-1 == clock_gettime(CLOCK_MONOTONIC,&tp)) // if MONOTONIC isn't supported at runtime then ouch, try REALTIME +#endif + clock_gettime(CLOCK_REALTIME,&tp); + + return (tp.tv_sec*LLFastTimer::sClockResolution)+tp.tv_nsec; +} + +inline U32 LLFastTimer::getCPUClockCount32() +{ + return (U32)(LLFastTimer::getCPUClockCount64() >> 8); +} +#endif // (LL_LINUX || LL_SOLARIS) && !(defined(__i386__) || defined(__amd64__)) + + +#if (LL_LINUX || LL_SOLARIS || LL_DARWIN) && (defined(__i386__) || defined(__amd64__)) +// +// Mac+Linux+Solaris FAST x86 implementation of CPU clock +inline U32 LLFastTimer::getCPUClockCount32() +{ + U64 x; + __asm__ volatile (".byte 0x0f, 0x31": "=A"(x)); + return (U32)(x >> 8); +} + +inline U64 LLFastTimer::getCPUClockCount64() +{ + U64 x; + __asm__ volatile (".byte 0x0f, 0x31": "=A"(x)); + return x; +} +#endif + + +#if ( LL_DARWIN && !(defined(__i386__) || defined(__amd64__))) +// +// Mac PPC (deprecated) implementation of CPU clock +// +// Just use gettimeofday implementation for now + +inline U32 LLFastTimer::getCPUClockCount32() +{ + return (U32)(get_clock_count()>>8); +} + +inline U64 LLFastTimer::getCPUClockCount64() +{ + return get_clock_count(); +} +#endif + |