git.cworth.org Git - apitrace/blob - retrace/glretrace_main.cpp

   1 /**************************************************************************
   2  *
   3  * Copyright 2011 Jose Fonseca
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a copy
   7  * of this software and associated documentation files (the "Software"), to deal
   8  * in the Software without restriction, including without limitation the rights
   9  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
  10  * copies of the Software, and to permit persons to whom the Software is
  11  * furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice shall be included in
  14  * all copies or substantial portions of the Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  22  * THE SOFTWARE.
  23  *
  24  **************************************************************************/
  25
  26
  27 #include <string.h>
  28
  29 #include "retrace.hpp"
  30 #include "glproc.hpp"
  31 #include "glstate.hpp"
  32 #include "glretrace.hpp"
  33 #include "os_time.hpp"
  34
  35 /* Synchronous debug output may reduce performance however,
  36  * without it the callNo in the callback may be inaccurate
  37  * as the callback may be called at any time.
  38  */
  39 #define DEBUG_OUTPUT_SYNCHRONOUS 0
  40
  41 namespace glretrace {
  42
  43 bool insideList = false;
  44 bool insideGlBeginEnd = false;
  45
  46 enum {
  47     GPU_START = 0,
  48     GPU_DURATION,
  49     OCCLUSION,
  50     NUM_QUERIES,
  51 };
  52
  53 struct CallQuery
  54 {
  55     GLuint ids[NUM_QUERIES];
  56     unsigned call;
  57     bool isDraw;
  58     GLuint program;
  59     const trace::FunctionSig *sig;
  60     int64_t cpuStart;
  61     int64_t cpuEnd;
  62 };
  63
  64 static bool supportsElapsed = true;
  65 static bool supportsTimestamp = true;
  66 static bool supportsOcclusion = true;
  67 static bool supportsDebugOutput = true;
  68
  69 static std::list<CallQuery> callQueries;
  70
  71 static void APIENTRY
  72 debugOutputCallback(GLenum source, GLenum type, GLuint id, GLenum severity, GLsizei length, const GLchar* message, GLvoid* userParam);
  73
  74 void
  75 checkGlError(trace::Call &call) {
  76     GLenum error = glGetError();
  77     while (error != GL_NO_ERROR) {
  78         std::ostream & os = retrace::warning(call);
  79
  80         os << "glGetError(";
  81         os << call.name();
  82         os << ") = ";
  83
  84         switch (error) {
  85         case GL_INVALID_ENUM:
  86             os << "GL_INVALID_ENUM";
  87             break;
  88         case GL_INVALID_VALUE:
  89             os << "GL_INVALID_VALUE";
  90             break;
  91         case GL_INVALID_OPERATION:
  92             os << "GL_INVALID_OPERATION";
  93             break;
  94         case GL_STACK_OVERFLOW:
  95             os << "GL_STACK_OVERFLOW";
  96             break;
  97         case GL_STACK_UNDERFLOW:
  98             os << "GL_STACK_UNDERFLOW";
  99             break;
 100         case GL_OUT_OF_MEMORY:
 101             os << "GL_OUT_OF_MEMORY";
 102             break;
 103         case GL_INVALID_FRAMEBUFFER_OPERATION:
 104             os << "GL_INVALID_FRAMEBUFFER_OPERATION";
 105             break;
 106         case GL_TABLE_TOO_LARGE:
 107             os << "GL_TABLE_TOO_LARGE";
 108             break;
 109         default:
 110             os << error;
 111             break;
 112         }
 113         os << "\n";
 114
 115         error = glGetError();
 116     }
 117 }
 118
 119 static inline int64_t
 120 getCurrentTime(void) {
 121     if (retrace::profilingGpuTimes && supportsTimestamp) {
 122         /* Get the current GL time without stalling */
 123         GLint64 timestamp = 0;
 124         glGetInteger64v(GL_TIMESTAMP, &timestamp);
 125         return timestamp;
 126     } else {
 127         return os::getTime();
 128     }
 129 }
 130
 131 static inline int64_t
 132 getTimeFrequency(void) {
 133     if (retrace::profilingGpuTimes && supportsTimestamp) {
 134         return 1000000000;
 135     } else {
 136         return os::timeFrequency;
 137     }
 138 }
 139
 140 static void
 141 getCurrentTimes(int64_t& cpuTime, int64_t& gpuTime) {
 142     GLuint query = 0;
 143
 144     if (retrace::profilingGpuTimes && supportsTimestamp) {
 145         glGenQueries(1, &query);
 146         glQueryCounter(query, GL_TIMESTAMP);
 147         GLint64 timestamp = 0;
 148         glGetQueryObjecti64vEXT(query, GL_QUERY_RESULT, &timestamp);
 149         gpuTime = timestamp;
 150     } else {
 151         gpuTime = 0;
 152     }
 153
 154     if (retrace::profilingCpuTimes) {
 155         cpuTime = getCurrentTime();
 156     } else {
 157         cpuTime = 0;
 158     }
 159
 160     if (retrace::profilingGpuTimes && supportsTimestamp) {
 161         glDeleteQueries(1, &query);
 162     }
 163 }
 164
 165 static void
 166 completeCallQuery(CallQuery& query) {
 167     /* Get call start and duration */
 168     int64_t gpuStart = 0, gpuDuration = 0, cpuDuration = 0, pixels = 0;
 169
 170     if (query.isDraw) {
 171         if (retrace::profilingGpuTimes) {
 172             if (supportsTimestamp) {
 173                 glGetQueryObjecti64vEXT(query.ids[GPU_START], GL_QUERY_RESULT, &gpuStart);
 174             }
 175
 176             glGetQueryObjecti64vEXT(query.ids[GPU_DURATION], GL_QUERY_RESULT, &gpuDuration);
 177         }
 178
 179         if (retrace::profilingPixelsDrawn) {
 180             glGetQueryObjecti64vEXT(query.ids[OCCLUSION], GL_QUERY_RESULT, &pixels);
 181         }
 182
 183     } else {
 184         pixels = -1;
 185     }
 186
 187     if (retrace::profilingCpuTimes) {
 188         cpuDuration = query.cpuEnd - query.cpuStart;
 189     }
 190
 191     glDeleteQueries(NUM_QUERIES, query.ids);
 192
 193     /* Add call to profile */
 194     retrace::profiler.addCall(query.call, query.sig->name, query.program, pixels, gpuStart, gpuDuration, query.cpuStart, cpuDuration);
 195 }
 196
 197 void
 198 flushQueries() {
 199     for (std::list<CallQuery>::iterator itr = callQueries.begin(); itr != callQueries.end(); ++itr) {
 200         completeCallQuery(*itr);
 201     }
 202
 203     callQueries.clear();
 204 }
 205
 206 void
 207 beginProfile(trace::Call &call, bool isDraw) {
 208     glretrace::Context *currentContext = glretrace::getCurrentContext();
 209
 210     /* Create call query */
 211     CallQuery query;
 212     query.isDraw = isDraw;
 213     query.call = call.no;
 214     query.sig = call.sig;
 215     query.program = currentContext ? currentContext->activeProgram : 0;
 216
 217     glGenQueries(NUM_QUERIES, query.ids);
 218
 219     /* GPU profiling only for draw calls */
 220     if (isDraw) {
 221         if (retrace::profilingGpuTimes) {
 222             if (supportsTimestamp) {
 223                 glQueryCounter(query.ids[GPU_START], GL_TIMESTAMP);
 224             }
 225
 226             glBeginQuery(GL_TIME_ELAPSED, query.ids[GPU_DURATION]);
 227         }
 228
 229         if (retrace::profilingPixelsDrawn) {
 230             glBeginQuery(GL_SAMPLES_PASSED, query.ids[OCCLUSION]);
 231         }
 232     }
 233
 234     callQueries.push_back(query);
 235
 236     /* CPU profiling for all calls */
 237     if (retrace::profilingCpuTimes) {
 238         CallQuery& query = callQueries.back();
 239         query.cpuStart = getCurrentTime();
 240     }
 241 }
 242
 243 void
 244 endProfile(trace::Call &call, bool isDraw) {
 245
 246     /* CPU profiling for all calls */
 247     if (retrace::profilingCpuTimes) {
 248         CallQuery& query = callQueries.back();
 249         query.cpuEnd = getCurrentTime();
 250     }
 251
 252     /* GPU profiling only for draw calls */
 253     if (isDraw) {
 254         if (retrace::profilingGpuTimes) {
 255             glEndQuery(GL_TIME_ELAPSED);
 256         }
 257
 258         if (retrace::profilingPixelsDrawn) {
 259             glEndQuery(GL_SAMPLES_PASSED);
 260         }
 261     }
 262 }
 263
 264 void
 265 initContext() {
 266     glretrace::Context *currentContext = glretrace::getCurrentContext();
 267
 268     /* Ensure we have adequate extension support */
 269     assert(currentContext);
 270     supportsTimestamp   = currentContext->hasExtension("GL_ARB_timer_query");
 271     supportsElapsed     = currentContext->hasExtension("GL_EXT_timer_query") || supportsTimestamp;
 272     supportsOcclusion   = currentContext->hasExtension("GL_ARB_occlusion_query");
 273     supportsDebugOutput = currentContext->hasExtension("GL_ARB_debug_output");
 274
 275     /* Check for timer query support */
 276     if (retrace::profilingGpuTimes) {
 277         if (!supportsTimestamp && !supportsElapsed) {
 278             std::cout << "Error: Cannot run profile, GL_EXT_timer_query extension is not supported." << std::endl;
 279             exit(-1);
 280         }
 281
 282         GLint bits = 0;
 283         glGetQueryiv(GL_TIME_ELAPSED, GL_QUERY_COUNTER_BITS, &bits);
 284
 285         if (!bits) {
 286             std::cout << "Error: Cannot run profile, GL_QUERY_COUNTER_BITS == 0." << std::endl;
 287             exit(-1);
 288         }
 289     }
 290
 291     /* Check for occlusion query support */
 292     if (retrace::profilingPixelsDrawn && !supportsOcclusion) {
 293         std::cout << "Error: Cannot run profile, GL_ARB_occlusion_query extension is not supported." << std::endl;
 294         exit(-1);
 295     }
 296
 297     /* Setup debug message call back */
 298     if (retrace::debug && supportsDebugOutput) {
 299         glretrace::Context *currentContext = glretrace::getCurrentContext();
 300         glDebugMessageCallbackARB(&debugOutputCallback, currentContext);
 301
 302         if (DEBUG_OUTPUT_SYNCHRONOUS) {
 303             glEnable(GL_DEBUG_OUTPUT_SYNCHRONOUS_ARB);
 304         }
 305     }
 306
 307     /* Sync the gpu and cpu start times */
 308     if (retrace::profilingCpuTimes || retrace::profilingGpuTimes) {
 309         if (!retrace::profiler.hasBaseTimes()) {
 310             GLint64 gpuTime, cpuTime;
 311
 312             getCurrentTimes(cpuTime, gpuTime);
 313             retrace::profiler.setBaseCpuTime(cpuTime);
 314             retrace::profiler.setBaseGpuTime(gpuTime);
 315         }
 316     }
 317 }
 318
 319 void
 320 frame_complete(trace::Call &call) {
 321     if (retrace::profiling) {
 322         /* Complete any remaining queries */
 323         flushQueries();
 324
 325         /* GPU time drifts due to being relative times, not absolute and can be
 326          * affected by the gpu switch between processes.
 327          *
 328          * To attempt to compensate we resynchronise on frame end however there is
 329          * still noticeable drift within a single frame which we do not account for.
 330          */
 331         if (retrace::profilingCpuTimes || retrace::profilingGpuTimes) {
 332             int64_t cpuTime, gpuTime, error;
 333
 334             getCurrentTimes(cpuTime, gpuTime);
 335             cpuTime = cpuTime - retrace::profiler.getBaseCpuTime();
 336             gpuTime = gpuTime - retrace::profiler.getBaseGpuTime();
 337             error   = gpuTime - cpuTime * (1.0E9 / getTimeFrequency());
 338             std::cerr << "error = " << error << "\n";
 339
 340             if (0) {
 341                 retrace::profiler.setBaseGpuTime(retrace::profiler.getBaseGpuTime() + error);
 342             }
 343         }
 344
 345         /* Indicate end of current frame */
 346         retrace::profiler.addFrameEnd();
 347     }
 348
 349     retrace::frameComplete(call);
 350
 351     glretrace::Context *currentContext = glretrace::getCurrentContext();
 352     if (!currentContext) {
 353         return;
 354     }
 355
 356     assert(currentContext->drawable);
 357     if (retrace::debug && !currentContext->drawable->visible) {
 358         retrace::warning(call) << "could not infer drawable size (glViewport never called)\n";
 359     }
 360 }
 361
 362 static const char*
 363 getDebugOutputSource(GLenum source) {
 364     switch(source) {
 365     case GL_DEBUG_SOURCE_API_ARB:
 366         return "API";
 367     case GL_DEBUG_SOURCE_WINDOW_SYSTEM_ARB:
 368         return "Window System";
 369     case GL_DEBUG_SOURCE_SHADER_COMPILER_ARB:
 370         return "Shader Compiler";
 371     case GL_DEBUG_SOURCE_THIRD_PARTY_ARB:
 372         return "Third Party";
 373     case GL_DEBUG_SOURCE_APPLICATION_ARB:
 374         return "Application";
 375     case GL_DEBUG_SOURCE_OTHER_ARB:
 376     default:
 377         return "";
 378     }
 379 }
 380
 381 static const char*
 382 getDebugOutputType(GLenum type) {
 383     switch(type) {
 384     case GL_DEBUG_TYPE_ERROR_ARB:
 385         return "error";
 386     case GL_DEBUG_TYPE_DEPRECATED_BEHAVIOR_ARB:
 387         return "deprecated behaviour";
 388     case GL_DEBUG_TYPE_UNDEFINED_BEHAVIOR_ARB:
 389         return "undefined behaviour";
 390     case GL_DEBUG_TYPE_PORTABILITY_ARB:
 391         return "portability issue";
 392     case GL_DEBUG_TYPE_PERFORMANCE_ARB:
 393         return "performance issue";
 394     case GL_DEBUG_TYPE_OTHER_ARB:
 395     default:
 396         return "unknown issue";
 397     }
 398 }
 399
 400 static const char*
 401 getDebugOutputSeverity(GLenum severity) {
 402     switch(severity) {
 403     case GL_DEBUG_SEVERITY_HIGH_ARB:
 404         return "High";
 405     case GL_DEBUG_SEVERITY_MEDIUM_ARB:
 406         return "Medium";
 407     case GL_DEBUG_SEVERITY_LOW_ARB:
 408         return "Low";
 409     default:
 410         return "usnknown";
 411     }
 412 }
 413
 414 static void APIENTRY
 415 debugOutputCallback(GLenum source, GLenum type, GLuint id, GLenum severity, GLsizei length, const GLchar* message, GLvoid* userParam) {
 416     std::cerr << retrace::callNo << ": ";
 417     std::cerr << "glDebugOutputCallback: ";
 418     std::cerr << getDebugOutputSeverity(severity) << " severity ";
 419     std::cerr << getDebugOutputSource(source) << " " << getDebugOutputType(type);
 420     std::cerr << " " << id;
 421     std::cerr << ", " << message;
 422     std::cerr << std::endl;
 423 }
 424
 425 } /* namespace glretrace */
 426
 427
 428 class GLDumper : public retrace::Dumper {
 429 public:
 430     image::Image *
 431     getSnapshot(void) {
 432         if (!glretrace::getCurrentContext()) {
 433             return NULL;
 434         }
 435         return glstate::getDrawBufferImage();
 436     }
 437
 438     bool
 439     dumpState(std::ostream &os) {
 440         glretrace::Context *currentContext = glretrace::getCurrentContext();
 441         if (glretrace::insideGlBeginEnd ||
 442             !currentContext) {
 443             return false;
 444         }
 445         glstate::dumpCurrentContext(os);
 446         return true;
 447     }
 448 };
 449
 450 static GLDumper glDumper;
 451
 452
 453 void
 454 retrace::setUp(void) {
 455     glws::init();
 456     dumper = &glDumper;
 457 }
 458
 459
 460 void
 461 retrace::addCallbacks(retrace::Retracer &retracer)
 462 {
 463     retracer.addCallbacks(glretrace::gl_callbacks);
 464     retracer.addCallbacks(glretrace::glx_callbacks);
 465     retracer.addCallbacks(glretrace::wgl_callbacks);
 466     retracer.addCallbacks(glretrace::cgl_callbacks);
 467     retracer.addCallbacks(glretrace::egl_callbacks);
 468 }
 469
 470
 471 void
 472 retrace::flushRendering(void) {
 473     glretrace::Context *currentContext = glretrace::getCurrentContext();
 474     if (currentContext) {
 475         glretrace::flushQueries();
 476         glFlush();
 477     }
 478 }
 479
 480 void
 481 retrace::waitForInput(void) {
 482     while (glws::processEvents()) {
 483         os::sleep(100*1000);
 484     }
 485 }
 486
 487 void
 488 retrace::cleanUp(void) {
 489     glws::cleanup();
 490 }