[1963] | 1 | |
---|
| 2 | /*************************************************************************************************** |
---|
| 3 | ** |
---|
| 4 | ** Real-Time Hierarchical Profiling for Game Programming Gems 3 |
---|
| 5 | ** |
---|
| 6 | ** by Greg Hjelstrom & Byon Garrabrant |
---|
| 7 | ** |
---|
| 8 | ***************************************************************************************************/ |
---|
| 9 | |
---|
| 10 | // Credits: The Clock class was inspired by the Timer classes in |
---|
| 11 | // Ogre (www.ogre3d.org). |
---|
| 12 | |
---|
[2430] | 13 | |
---|
| 14 | |
---|
[1963] | 15 | #ifndef QUICK_PROF_H |
---|
| 16 | #define QUICK_PROF_H |
---|
| 17 | |
---|
[2430] | 18 | //To disable built-in profiling, please comment out next line |
---|
| 19 | //#define BT_NO_PROFILE 1 |
---|
| 20 | #ifndef BT_NO_PROFILE |
---|
| 21 | |
---|
[1963] | 22 | #include "btScalar.h" |
---|
| 23 | #include "LinearMath/btAlignedAllocator.h" |
---|
| 24 | #include <new> |
---|
| 25 | |
---|
| 26 | |
---|
[2430] | 27 | |
---|
| 28 | |
---|
[1963] | 29 | //if you don't need btClock, you can comment next line |
---|
| 30 | #define USE_BT_CLOCK 1 |
---|
| 31 | |
---|
| 32 | #ifdef USE_BT_CLOCK |
---|
| 33 | #ifdef __CELLOS_LV2__ |
---|
| 34 | #include <sys/sys_time.h> |
---|
| 35 | #include <sys/time_util.h> |
---|
| 36 | #include <stdio.h> |
---|
| 37 | #endif |
---|
| 38 | |
---|
| 39 | #if defined (SUNOS) || defined (__SUNOS__) |
---|
| 40 | #include <stdio.h> |
---|
| 41 | #endif |
---|
| 42 | |
---|
| 43 | #if defined(WIN32) || defined(_WIN32) |
---|
| 44 | |
---|
| 45 | #define USE_WINDOWS_TIMERS |
---|
| 46 | #define WIN32_LEAN_AND_MEAN |
---|
| 47 | #define NOWINRES |
---|
| 48 | #define NOMCX |
---|
| 49 | #define NOIME |
---|
| 50 | #ifdef _XBOX |
---|
| 51 | #include <Xtl.h> |
---|
| 52 | #else |
---|
| 53 | #include <windows.h> |
---|
| 54 | #endif |
---|
| 55 | #include <time.h> |
---|
| 56 | |
---|
| 57 | #else |
---|
| 58 | #include <sys/time.h> |
---|
| 59 | #endif |
---|
| 60 | |
---|
| 61 | #define mymin(a,b) (a > b ? a : b) |
---|
| 62 | |
---|
| 63 | ///The btClock is a portable basic clock that measures accurate time in seconds, use for profiling. |
---|
| 64 | class btClock |
---|
| 65 | { |
---|
| 66 | public: |
---|
| 67 | btClock() |
---|
| 68 | { |
---|
| 69 | #ifdef USE_WINDOWS_TIMERS |
---|
| 70 | QueryPerformanceFrequency(&mClockFrequency); |
---|
| 71 | #endif |
---|
| 72 | reset(); |
---|
| 73 | } |
---|
| 74 | |
---|
| 75 | ~btClock() |
---|
| 76 | { |
---|
| 77 | } |
---|
| 78 | |
---|
| 79 | /// Resets the initial reference time. |
---|
| 80 | void reset() |
---|
| 81 | { |
---|
| 82 | #ifdef USE_WINDOWS_TIMERS |
---|
| 83 | QueryPerformanceCounter(&mStartTime); |
---|
| 84 | mStartTick = GetTickCount(); |
---|
| 85 | mPrevElapsedTime = 0; |
---|
| 86 | #else |
---|
| 87 | #ifdef __CELLOS_LV2__ |
---|
| 88 | |
---|
| 89 | typedef uint64_t ClockSize; |
---|
| 90 | ClockSize newTime; |
---|
| 91 | //__asm __volatile__( "mftb %0" : "=r" (newTime) : : "memory"); |
---|
| 92 | SYS_TIMEBASE_GET( newTime ); |
---|
| 93 | mStartTime = newTime; |
---|
| 94 | #else |
---|
| 95 | gettimeofday(&mStartTime, 0); |
---|
| 96 | #endif |
---|
| 97 | |
---|
| 98 | #endif |
---|
| 99 | } |
---|
| 100 | |
---|
| 101 | /// Returns the time in ms since the last call to reset or since |
---|
| 102 | /// the btClock was created. |
---|
| 103 | unsigned long int getTimeMilliseconds() |
---|
| 104 | { |
---|
| 105 | #ifdef USE_WINDOWS_TIMERS |
---|
| 106 | LARGE_INTEGER currentTime; |
---|
| 107 | QueryPerformanceCounter(¤tTime); |
---|
| 108 | LONGLONG elapsedTime = currentTime.QuadPart - |
---|
| 109 | mStartTime.QuadPart; |
---|
| 110 | |
---|
| 111 | // Compute the number of millisecond ticks elapsed. |
---|
| 112 | unsigned long msecTicks = (unsigned long)(1000 * elapsedTime / |
---|
| 113 | mClockFrequency.QuadPart); |
---|
| 114 | |
---|
| 115 | // Check for unexpected leaps in the Win32 performance counter. |
---|
| 116 | // (This is caused by unexpected data across the PCI to ISA |
---|
| 117 | // bridge, aka south bridge. See Microsoft KB274323.) |
---|
| 118 | unsigned long elapsedTicks = GetTickCount() - mStartTick; |
---|
| 119 | signed long msecOff = (signed long)(msecTicks - elapsedTicks); |
---|
| 120 | if (msecOff < -100 || msecOff > 100) |
---|
| 121 | { |
---|
| 122 | // Adjust the starting time forwards. |
---|
| 123 | LONGLONG msecAdjustment = mymin(msecOff * |
---|
| 124 | mClockFrequency.QuadPart / 1000, elapsedTime - |
---|
| 125 | mPrevElapsedTime); |
---|
| 126 | mStartTime.QuadPart += msecAdjustment; |
---|
| 127 | elapsedTime -= msecAdjustment; |
---|
| 128 | |
---|
| 129 | // Recompute the number of millisecond ticks elapsed. |
---|
| 130 | msecTicks = (unsigned long)(1000 * elapsedTime / |
---|
| 131 | mClockFrequency.QuadPart); |
---|
| 132 | } |
---|
| 133 | |
---|
| 134 | // Store the current elapsed time for adjustments next time. |
---|
| 135 | mPrevElapsedTime = elapsedTime; |
---|
| 136 | |
---|
| 137 | return msecTicks; |
---|
| 138 | #else |
---|
| 139 | |
---|
| 140 | #ifdef __CELLOS_LV2__ |
---|
| 141 | uint64_t freq=sys_time_get_timebase_frequency(); |
---|
| 142 | double dFreq=((double) freq) / 1000.0; |
---|
| 143 | typedef uint64_t ClockSize; |
---|
| 144 | ClockSize newTime; |
---|
| 145 | SYS_TIMEBASE_GET( newTime ); |
---|
| 146 | //__asm __volatile__( "mftb %0" : "=r" (newTime) : : "memory"); |
---|
| 147 | |
---|
| 148 | return (unsigned long int)((double(newTime-mStartTime)) / dFreq); |
---|
| 149 | #else |
---|
| 150 | |
---|
| 151 | struct timeval currentTime; |
---|
| 152 | gettimeofday(¤tTime, 0); |
---|
| 153 | return (currentTime.tv_sec - mStartTime.tv_sec) * 1000 + |
---|
| 154 | (currentTime.tv_usec - mStartTime.tv_usec) / 1000; |
---|
| 155 | #endif //__CELLOS_LV2__ |
---|
| 156 | #endif |
---|
| 157 | } |
---|
| 158 | |
---|
| 159 | /// Returns the time in us since the last call to reset or since |
---|
| 160 | /// the Clock was created. |
---|
| 161 | unsigned long int getTimeMicroseconds() |
---|
| 162 | { |
---|
| 163 | #ifdef USE_WINDOWS_TIMERS |
---|
| 164 | LARGE_INTEGER currentTime; |
---|
| 165 | QueryPerformanceCounter(¤tTime); |
---|
| 166 | LONGLONG elapsedTime = currentTime.QuadPart - |
---|
| 167 | mStartTime.QuadPart; |
---|
| 168 | |
---|
| 169 | // Compute the number of millisecond ticks elapsed. |
---|
| 170 | unsigned long msecTicks = (unsigned long)(1000 * elapsedTime / |
---|
| 171 | mClockFrequency.QuadPart); |
---|
| 172 | |
---|
| 173 | // Check for unexpected leaps in the Win32 performance counter. |
---|
| 174 | // (This is caused by unexpected data across the PCI to ISA |
---|
| 175 | // bridge, aka south bridge. See Microsoft KB274323.) |
---|
| 176 | unsigned long elapsedTicks = GetTickCount() - mStartTick; |
---|
| 177 | signed long msecOff = (signed long)(msecTicks - elapsedTicks); |
---|
| 178 | if (msecOff < -100 || msecOff > 100) |
---|
| 179 | { |
---|
| 180 | // Adjust the starting time forwards. |
---|
| 181 | LONGLONG msecAdjustment = mymin(msecOff * |
---|
| 182 | mClockFrequency.QuadPart / 1000, elapsedTime - |
---|
| 183 | mPrevElapsedTime); |
---|
| 184 | mStartTime.QuadPart += msecAdjustment; |
---|
| 185 | elapsedTime -= msecAdjustment; |
---|
| 186 | } |
---|
| 187 | |
---|
| 188 | // Store the current elapsed time for adjustments next time. |
---|
| 189 | mPrevElapsedTime = elapsedTime; |
---|
| 190 | |
---|
| 191 | // Convert to microseconds. |
---|
| 192 | unsigned long usecTicks = (unsigned long)(1000000 * elapsedTime / |
---|
| 193 | mClockFrequency.QuadPart); |
---|
| 194 | |
---|
| 195 | return usecTicks; |
---|
| 196 | #else |
---|
| 197 | |
---|
| 198 | #ifdef __CELLOS_LV2__ |
---|
| 199 | uint64_t freq=sys_time_get_timebase_frequency(); |
---|
| 200 | double dFreq=((double) freq)/ 1000000.0; |
---|
| 201 | typedef uint64_t ClockSize; |
---|
| 202 | ClockSize newTime; |
---|
| 203 | //__asm __volatile__( "mftb %0" : "=r" (newTime) : : "memory"); |
---|
| 204 | SYS_TIMEBASE_GET( newTime ); |
---|
| 205 | |
---|
| 206 | return (unsigned long int)((double(newTime-mStartTime)) / dFreq); |
---|
| 207 | #else |
---|
| 208 | |
---|
| 209 | struct timeval currentTime; |
---|
| 210 | gettimeofday(¤tTime, 0); |
---|
| 211 | return (currentTime.tv_sec - mStartTime.tv_sec) * 1000000 + |
---|
| 212 | (currentTime.tv_usec - mStartTime.tv_usec); |
---|
| 213 | #endif//__CELLOS_LV2__ |
---|
| 214 | #endif |
---|
| 215 | } |
---|
| 216 | |
---|
| 217 | private: |
---|
| 218 | #ifdef USE_WINDOWS_TIMERS |
---|
| 219 | LARGE_INTEGER mClockFrequency; |
---|
| 220 | DWORD mStartTick; |
---|
| 221 | LONGLONG mPrevElapsedTime; |
---|
| 222 | LARGE_INTEGER mStartTime; |
---|
| 223 | #else |
---|
| 224 | #ifdef __CELLOS_LV2__ |
---|
| 225 | uint64_t mStartTime; |
---|
| 226 | #else |
---|
| 227 | struct timeval mStartTime; |
---|
| 228 | #endif |
---|
| 229 | #endif //__CELLOS_LV2__ |
---|
| 230 | |
---|
| 231 | }; |
---|
| 232 | |
---|
| 233 | #endif //USE_BT_CLOCK |
---|
| 234 | |
---|
| 235 | |
---|
| 236 | |
---|
| 237 | |
---|
| 238 | ///A node in the Profile Hierarchy Tree |
---|
| 239 | class CProfileNode { |
---|
| 240 | |
---|
| 241 | public: |
---|
| 242 | CProfileNode( const char * name, CProfileNode * parent ); |
---|
| 243 | ~CProfileNode( void ); |
---|
| 244 | |
---|
| 245 | CProfileNode * Get_Sub_Node( const char * name ); |
---|
| 246 | |
---|
| 247 | CProfileNode * Get_Parent( void ) { return Parent; } |
---|
| 248 | CProfileNode * Get_Sibling( void ) { return Sibling; } |
---|
| 249 | CProfileNode * Get_Child( void ) { return Child; } |
---|
| 250 | |
---|
| 251 | void CleanupMemory(); |
---|
| 252 | void Reset( void ); |
---|
| 253 | void Call( void ); |
---|
| 254 | bool Return( void ); |
---|
| 255 | |
---|
| 256 | const char * Get_Name( void ) { return Name; } |
---|
| 257 | int Get_Total_Calls( void ) { return TotalCalls; } |
---|
| 258 | float Get_Total_Time( void ) { return TotalTime; } |
---|
| 259 | |
---|
| 260 | protected: |
---|
| 261 | |
---|
| 262 | const char * Name; |
---|
| 263 | int TotalCalls; |
---|
| 264 | float TotalTime; |
---|
| 265 | unsigned long int StartTime; |
---|
| 266 | int RecursionCounter; |
---|
| 267 | |
---|
| 268 | CProfileNode * Parent; |
---|
| 269 | CProfileNode * Child; |
---|
| 270 | CProfileNode * Sibling; |
---|
| 271 | }; |
---|
| 272 | |
---|
| 273 | ///An iterator to navigate through the tree |
---|
| 274 | class CProfileIterator |
---|
| 275 | { |
---|
| 276 | public: |
---|
| 277 | // Access all the children of the current parent |
---|
| 278 | void First(void); |
---|
| 279 | void Next(void); |
---|
| 280 | bool Is_Done(void); |
---|
| 281 | bool Is_Root(void) { return (CurrentParent->Get_Parent() == 0); } |
---|
| 282 | |
---|
| 283 | void Enter_Child( int index ); // Make the given child the new parent |
---|
| 284 | void Enter_Largest_Child( void ); // Make the largest child the new parent |
---|
| 285 | void Enter_Parent( void ); // Make the current parent's parent the new parent |
---|
| 286 | |
---|
| 287 | // Access the current child |
---|
| 288 | const char * Get_Current_Name( void ) { return CurrentChild->Get_Name(); } |
---|
| 289 | int Get_Current_Total_Calls( void ) { return CurrentChild->Get_Total_Calls(); } |
---|
| 290 | float Get_Current_Total_Time( void ) { return CurrentChild->Get_Total_Time(); } |
---|
| 291 | |
---|
| 292 | // Access the current parent |
---|
| 293 | const char * Get_Current_Parent_Name( void ) { return CurrentParent->Get_Name(); } |
---|
| 294 | int Get_Current_Parent_Total_Calls( void ) { return CurrentParent->Get_Total_Calls(); } |
---|
| 295 | float Get_Current_Parent_Total_Time( void ) { return CurrentParent->Get_Total_Time(); } |
---|
| 296 | |
---|
| 297 | protected: |
---|
| 298 | |
---|
| 299 | CProfileNode * CurrentParent; |
---|
| 300 | CProfileNode * CurrentChild; |
---|
| 301 | |
---|
| 302 | CProfileIterator( CProfileNode * start ); |
---|
| 303 | friend class CProfileManager; |
---|
| 304 | }; |
---|
| 305 | |
---|
| 306 | |
---|
| 307 | ///The Manager for the Profile system |
---|
| 308 | class CProfileManager { |
---|
| 309 | public: |
---|
| 310 | static void Start_Profile( const char * name ); |
---|
| 311 | static void Stop_Profile( void ); |
---|
| 312 | |
---|
| 313 | static void CleanupMemory(void) |
---|
| 314 | { |
---|
| 315 | Root.CleanupMemory(); |
---|
| 316 | } |
---|
| 317 | |
---|
| 318 | static void Reset( void ); |
---|
| 319 | static void Increment_Frame_Counter( void ); |
---|
| 320 | static int Get_Frame_Count_Since_Reset( void ) { return FrameCounter; } |
---|
| 321 | static float Get_Time_Since_Reset( void ); |
---|
| 322 | |
---|
| 323 | static CProfileIterator * Get_Iterator( void ) |
---|
| 324 | { |
---|
| 325 | |
---|
| 326 | return new CProfileIterator( &Root ); |
---|
| 327 | } |
---|
| 328 | static void Release_Iterator( CProfileIterator * iterator ) { delete ( iterator); } |
---|
| 329 | |
---|
[2430] | 330 | static void dumpRecursive(CProfileIterator* profileIterator, int spacing); |
---|
| 331 | |
---|
| 332 | static void dumpAll(); |
---|
| 333 | |
---|
[1963] | 334 | private: |
---|
| 335 | static CProfileNode Root; |
---|
| 336 | static CProfileNode * CurrentNode; |
---|
| 337 | static int FrameCounter; |
---|
| 338 | static unsigned long int ResetTime; |
---|
| 339 | }; |
---|
| 340 | |
---|
| 341 | |
---|
| 342 | ///ProfileSampleClass is a simple way to profile a function's scope |
---|
| 343 | ///Use the BT_PROFILE macro at the start of scope to time |
---|
| 344 | class CProfileSample { |
---|
| 345 | public: |
---|
| 346 | CProfileSample( const char * name ) |
---|
| 347 | { |
---|
| 348 | CProfileManager::Start_Profile( name ); |
---|
| 349 | } |
---|
| 350 | |
---|
| 351 | ~CProfileSample( void ) |
---|
| 352 | { |
---|
| 353 | CProfileManager::Stop_Profile(); |
---|
| 354 | } |
---|
| 355 | }; |
---|
| 356 | |
---|
[2430] | 357 | |
---|
[1963] | 358 | #define BT_PROFILE( name ) CProfileSample __profile( name ) |
---|
[2430] | 359 | |
---|
[1963] | 360 | #else |
---|
[2430] | 361 | |
---|
[1963] | 362 | #define BT_PROFILE( name ) |
---|
| 363 | |
---|
[2430] | 364 | #endif //#ifndef BT_NO_PROFILE |
---|
[1963] | 365 | |
---|
| 366 | |
---|
| 367 | |
---|
| 368 | #endif //QUICK_PROF_H |
---|
| 369 | |
---|
| 370 | |
---|