X-Git-Url: https://git.cworth.org/git?a=blobdiff_plain;f=retrace%2Fglretrace_main.cpp;h=d215133cbff906ff3c6e912bb574ee834dfc8a7d;hb=342e9725ecc44cf882ee13f87d5099f71f2700af;hp=f9d61e05713384df80eed73cf7924b002ad2063b;hpb=d1b3611276f097f02c8379722319fcdc44ff02a4;p=apitrace

diff --git a/retrace/glretrace_main.cpp b/retrace/glretrace_main.cpp
old mode 100644
new mode 100755
index f9d61e0..d215133
--- a/retrace/glretrace_main.cpp
+++ b/retrace/glretrace_main.cpp
@@ -1,6 +1,8 @@
 /**************************************************************************
  *
  * Copyright 2011 Jose Fonseca
+ * Copyright (C) 2013 Intel Corporation. All rights reversed.
+ * Author: Shuang He <shuang.he@intel.com>
  * All Rights Reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
@@ -31,6 +33,7 @@
 #include "glstate.hpp"
 #include "glretrace.hpp"
 #include "os_time.hpp"
+#include "os_memory.hpp"
 
 /* Synchronous debug output may reduce performance however,
  * without it the callNo in the callback may be inaccurate
@@ -42,119 +45,153 @@ namespace glretrace {
 
 bool insideList = false;
 bool insideGlBeginEnd = false;
+bool supportsARBShaderObjects = false;
+
+enum {
+    GPU_START = 0,
+    GPU_DURATION,
+    OCCLUSION,
+    NUM_QUERIES,
+};
 
 struct CallQuery
 {
-    GLuint ids[3];
+    GLuint ids[NUM_QUERIES];
     unsigned call;
+    bool isDraw;
     GLuint program;
     const trace::FunctionSig *sig;
-    uint64_t start;
-    uint64_t duration;
+    int64_t cpuStart;
+    int64_t cpuEnd;
+    int64_t vsizeStart;
+    int64_t vsizeEnd;
+    int64_t rssStart;
+    int64_t rssEnd;
 };
 
 static bool supportsElapsed = true;
 static bool supportsTimestamp = true;
 static bool supportsOcclusion = true;
+static bool supportsDebugOutput = true;
 
-static bool firstFrame = true;
 static std::list<CallQuery> callQueries;
 
-static void
+static void APIENTRY
 debugOutputCallback(GLenum source, GLenum type, GLuint id, GLenum severity, GLsizei length, const GLchar* message, GLvoid* userParam);
 
 void
 checkGlError(trace::Call &call) {
     GLenum error = glGetError();
-    if (error == GL_NO_ERROR) {
-        return;
+    while (error != GL_NO_ERROR) {
+        std::ostream & os = retrace::warning(call);
+
+        os << "glGetError(";
+        os << call.name();
+        os << ") = ";
+
+        switch (error) {
+        case GL_INVALID_ENUM:
+            os << "GL_INVALID_ENUM";
+            break;
+        case GL_INVALID_VALUE:
+            os << "GL_INVALID_VALUE";
+            break;
+        case GL_INVALID_OPERATION:
+            os << "GL_INVALID_OPERATION";
+            break;
+        case GL_STACK_OVERFLOW:
+            os << "GL_STACK_OVERFLOW";
+            break;
+        case GL_STACK_UNDERFLOW:
+            os << "GL_STACK_UNDERFLOW";
+            break;
+        case GL_OUT_OF_MEMORY:
+            os << "GL_OUT_OF_MEMORY";
+            break;
+        case GL_INVALID_FRAMEBUFFER_OPERATION:
+            os << "GL_INVALID_FRAMEBUFFER_OPERATION";
+            break;
+        case GL_TABLE_TOO_LARGE:
+            os << "GL_TABLE_TOO_LARGE";
+            break;
+        default:
+            os << error;
+            break;
+        }
+        os << "\n";
+    
+        error = glGetError();
     }
+}
 
-    std::ostream & os = retrace::warning(call);
-
-    os << "glGetError(";
-    os << call.name();
-    os << ") = ";
-
-    switch (error) {
-    case GL_INVALID_ENUM:
-        os << "GL_INVALID_ENUM";
-        break;
-    case GL_INVALID_VALUE:
-        os << "GL_INVALID_VALUE";
-        break;
-    case GL_INVALID_OPERATION:
-        os << "GL_INVALID_OPERATION";
-        break;
-    case GL_STACK_OVERFLOW:
-        os << "GL_STACK_OVERFLOW";
-        break;
-    case GL_STACK_UNDERFLOW:
-        os << "GL_STACK_UNDERFLOW";
-        break;
-    case GL_OUT_OF_MEMORY:
-        os << "GL_OUT_OF_MEMORY";
-        break;
-    case GL_INVALID_FRAMEBUFFER_OPERATION:
-        os << "GL_INVALID_FRAMEBUFFER_OPERATION";
-        break;
-    case GL_TABLE_TOO_LARGE:
-        os << "GL_TABLE_TOO_LARGE";
-        break;
-    default:
-        os << error;
-        break;
+static inline int64_t
+getCurrentTime(void) {
+    if (retrace::profilingGpuTimes && supportsTimestamp) {
+        /* Get the current GL time without stalling */
+        GLint64 timestamp = 0;
+        glGetInteger64v(GL_TIMESTAMP, &timestamp);
+        return timestamp;
+    } else {
+        return os::getTime();
     }
-    os << "\n";
 }
 
-static GLuint64
-getGpuTimestamp() {
-    GLuint query = 0;
-    GLuint64 timestamp = 0;
-
+static inline int64_t
+getTimeFrequency(void) {
     if (retrace::profilingGpuTimes && supportsTimestamp) {
-        glGenQueries(1, &query);
-        glQueryCounter(query, GL_TIMESTAMP);
-        glGetQueryObjectui64vEXT(query, GL_QUERY_RESULT, &timestamp);
-        glDeleteQueries(1, &query);
+        return 1000000000;
+    } else {
+        return os::timeFrequency;
     }
+}
 
-    return timestamp;
+static inline void
+getCurrentVsize(int64_t& vsize) {
+    vsize = os::getVsize();
 }
 
-static GLuint64
-getCpuTimestamp() {
-    if (retrace::profilingCpuTimes) {
-        return os::getTime() * (1.0E9 / os::timeFrequency);
-    } else {
-        return 0;
-    }
+static inline void
+getCurrentRss(int64_t& rss) {
+    rss = os::getRss();
 }
 
 static void
 completeCallQuery(CallQuery& query) {
     /* Get call start and duration */
-    GLuint64 timestamp = 0, duration = 0, samples = 0;
+    int64_t gpuStart = 0, gpuDuration = 0, cpuDuration = 0, pixels = 0, vsizeDuration = 0, rssDuration = 0;
 
-    if (retrace::profilingGpuTimes) {
-        if (supportsTimestamp) {
-            glGetQueryObjectui64vEXT(query.ids[0], GL_QUERY_RESULT, &timestamp);
+    if (query.isDraw) {
+        if (retrace::profilingGpuTimes) {
+            if (supportsTimestamp) {
+                glGetQueryObjecti64vEXT(query.ids[GPU_START], GL_QUERY_RESULT, &gpuStart);
+            }
+
+            glGetQueryObjecti64vEXT(query.ids[GPU_DURATION], GL_QUERY_RESULT, &gpuDuration);
         }
 
-        if (supportsElapsed) {
-            glGetQueryObjectui64vEXT(query.ids[1], GL_QUERY_RESULT, &duration);
+        if (retrace::profilingPixelsDrawn) {
+            glGetQueryObjecti64vEXT(query.ids[OCCLUSION], GL_QUERY_RESULT, &pixels);
         }
+
+    } else {
+        pixels = -1;
+    }
+
+    if (retrace::profilingCpuTimes) {
+        double cpuTimeScale = 1.0E9 / getTimeFrequency();
+        cpuDuration = (query.cpuEnd - query.cpuStart) * cpuTimeScale;
+        query.cpuStart *= cpuTimeScale;
     }
 
-    if (retrace::profilingPixelsDrawn && supportsOcclusion) {
-        glGetQueryObjectui64vEXT(query.ids[2], GL_QUERY_RESULT, &samples);
+    if (retrace::profilingMemoryUsage) {
+        vsizeDuration = query.vsizeEnd - query.vsizeStart;
+        rssDuration = query.rssEnd - query.rssStart;
     }
 
-    glDeleteQueries(3, query.ids);
+    glDeleteQueries(NUM_QUERIES, query.ids);
 
     /* Add call to profile */
-    retrace::profiler.addCall(query.call, query.sig->name, query.program, samples, timestamp, duration, query.start, query.duration);
+    retrace::profiler.addCall(query.call, query.sig->name, query.program, pixels, gpuStart, gpuDuration, query.cpuStart, cpuDuration, query.vsizeStart, vsizeDuration, query.rssStart, rssDuration);
 }
 
 void
@@ -167,72 +204,95 @@ flushQueries() {
 }
 
 void
-beginProfile(trace::Call &call) {
-    if (firstFrame) {
-        frame_start();
-    }
+beginProfile(trace::Call &call, bool isDraw) {
+    glretrace::Context *currentContext = glretrace::getCurrentContext();
 
     /* Create call query */
     CallQuery query;
+    query.isDraw = isDraw;
     query.call = call.no;
     query.sig = call.sig;
-    query.program = glretrace::currentContext ? glretrace::currentContext->activeProgram : 0;
+    query.program = currentContext ? currentContext->activeProgram : 0;
 
-    glGenQueries(3, query.ids);
+    glGenQueries(NUM_QUERIES, query.ids);
 
-    if (retrace::profilingGpuTimes) {
-        if (supportsTimestamp) {
-            glQueryCounter(query.ids[0], GL_TIMESTAMP);
+    /* GPU profiling only for draw calls */
+    if (isDraw) {
+        if (retrace::profilingGpuTimes) {
+            if (supportsTimestamp) {
+                glQueryCounter(query.ids[GPU_START], GL_TIMESTAMP);
+            }
+
+            glBeginQuery(GL_TIME_ELAPSED, query.ids[GPU_DURATION]);
         }
 
-        if (supportsElapsed) {
-            glBeginQuery(GL_TIME_ELAPSED, query.ids[1]);
+        if (retrace::profilingPixelsDrawn) {
+            glBeginQuery(GL_SAMPLES_PASSED, query.ids[OCCLUSION]);
         }
     }
 
-    if (retrace::profilingPixelsDrawn && supportsOcclusion) {
-        glBeginQuery(GL_SAMPLES_PASSED, query.ids[2]);
-    }
+    callQueries.push_back(query);
 
+    /* CPU profiling for all calls */
     if (retrace::profilingCpuTimes) {
-        query.start = os::getTime();
+        CallQuery& query = callQueries.back();
+        query.cpuStart = getCurrentTime();
     }
 
-    callQueries.push_back(query);
+    if (retrace::profilingMemoryUsage) {
+        CallQuery& query = callQueries.back();
+        query.vsizeStart = os::getVsize();
+        query.rssStart = os::getRss();
+    }
 }
 
 void
-endProfile(trace::Call &call) {
+endProfile(trace::Call &call, bool isDraw) {
+
+    /* CPU profiling for all calls */
     if (retrace::profilingCpuTimes) {
         CallQuery& query = callQueries.back();
-        query.duration = (os::getTime() - query.start) * (1.0E9 / os::timeFrequency);
+        query.cpuEnd = getCurrentTime();
     }
 
-    if (retrace::profilingGpuTimes && supportsElapsed) {
-        glEndQuery(GL_TIME_ELAPSED);
+    /* GPU profiling only for draw calls */
+    if (isDraw) {
+        if (retrace::profilingGpuTimes) {
+            glEndQuery(GL_TIME_ELAPSED);
+        }
+
+        if (retrace::profilingPixelsDrawn) {
+            glEndQuery(GL_SAMPLES_PASSED);
+        }
     }
 
-    if (retrace::profilingPixelsDrawn && supportsOcclusion) {
-        glEndQuery(GL_SAMPLES_PASSED);
+    if (retrace::profilingMemoryUsage) {
+        CallQuery& query = callQueries.back();
+        query.vsizeEnd = os::getVsize();
+        query.rssEnd = os::getRss();
     }
 }
 
 void
 initContext() {
-    /* Check for extension support */
-    const char* extensions = (const char*)glGetString(GL_EXTENSIONS);
-    GLint bits;
+    glretrace::Context *currentContext = glretrace::getCurrentContext();
 
-    supportsTimestamp = glws::checkExtension("GL_ARB_timer_query", extensions);
-    supportsElapsed   = glws::checkExtension("GL_EXT_timer_query", extensions) || supportsTimestamp;
-    supportsOcclusion = glws::checkExtension("GL_ARB_occlusion_query", extensions);
+    /* Ensure we have adequate extension support */
+    assert(currentContext);
+    supportsTimestamp   = currentContext->hasExtension("GL_ARB_timer_query");
+    supportsElapsed     = currentContext->hasExtension("GL_EXT_timer_query") || supportsTimestamp;
+    supportsOcclusion   = currentContext->hasExtension("GL_ARB_occlusion_query");
+    supportsDebugOutput = currentContext->hasExtension("GL_ARB_debug_output");
+    supportsARBShaderObjects = currentContext->hasExtension("GL_ARB_shader_objects");
 
+    /* Check for timer query support */
     if (retrace::profilingGpuTimes) {
         if (!supportsTimestamp && !supportsElapsed) {
             std::cout << "Error: Cannot run profile, GL_EXT_timer_query extension is not supported." << std::endl;
             exit(-1);
         }
 
+        GLint bits = 0;
         glGetQueryiv(GL_TIME_ELAPSED, GL_QUERY_COUNTER_BITS, &bits);
 
         if (!bits) {
@@ -241,30 +301,38 @@ initContext() {
         }
     }
 
+    /* Check for occlusion query support */
     if (retrace::profilingPixelsDrawn && !supportsOcclusion) {
         std::cout << "Error: Cannot run profile, GL_ARB_occlusion_query extension is not supported." << std::endl;
         exit(-1);
     }
 
-    if (retrace::debug) {
-        bool supportsDebugOutput = glws::checkExtension("GL_ARB_debug_output", extensions);
+    /* Setup debug message call back */
+    if (retrace::debug && supportsDebugOutput) {
+        glretrace::Context *currentContext = glretrace::getCurrentContext();
+        glDebugMessageCallbackARB(&debugOutputCallback, currentContext);
 
-        if (supportsDebugOutput) {
-            glDebugMessageCallbackARB(&debugOutputCallback, currentContext);
-
-            if (DEBUG_OUTPUT_SYNCHRONOUS) {
-                glEnable(GL_DEBUG_OUTPUT_SYNCHRONOUS_ARB);
-            }
+        if (DEBUG_OUTPUT_SYNCHRONOUS) {
+            glEnable(GL_DEBUG_OUTPUT_SYNCHRONOUS_ARB);
         }
     }
-}
 
-void
-frame_start() {
-    firstFrame = false;
+    /* Sync the gpu and cpu start times */
+    if (retrace::profilingCpuTimes || retrace::profilingGpuTimes) {
+        if (!retrace::profiler.hasBaseTimes()) {
+            double cpuTimeScale = 1.0E9 / getTimeFrequency();
+            GLint64 currentTime = getCurrentTime() * cpuTimeScale;
+            retrace::profiler.setBaseCpuTime(currentTime);
+            retrace::profiler.setBaseGpuTime(currentTime);
+        }
+    }
 
-    if (retrace::profiling) {
-        retrace::profiler.addFrameStart(retrace::frameNo, getGpuTimestamp(), getCpuTimestamp());
+    if (retrace::profilingMemoryUsage) {
+        GLint64 currentVsize, currentRss;
+        getCurrentVsize(currentVsize);
+        retrace::profiler.setBaseVsizeUsage(currentVsize);
+        getCurrentRss(currentRss);
+        retrace::profiler.setBaseRssUsage(currentRss);
     }
 }
 
@@ -275,19 +343,18 @@ frame_complete(trace::Call &call) {
         flushQueries();
 
         /* Indicate end of current frame */
-        retrace::profiler.addFrameEnd(getGpuTimestamp(), getCpuTimestamp());
+        retrace::profiler.addFrameEnd();
     }
 
     retrace::frameComplete(call);
 
-    /* Indicate start of next frame */
-    frame_start();
-
-    if (!currentDrawable) {
+    glretrace::Context *currentContext = glretrace::getCurrentContext();
+    if (!currentContext) {
         return;
     }
 
-    if (retrace::debug && !currentDrawable->visible) {
+    assert(currentContext->drawable);
+    if (retrace::debug && !currentContext->drawable->visible) {
         retrace::warning(call) << "could not infer drawable size (glViewport never called)\n";
     }
 }
@@ -344,7 +411,7 @@ getDebugOutputSeverity(GLenum severity) {
     }
 }
 
-static void
+static void APIENTRY
 debugOutputCallback(GLenum source, GLenum type, GLuint id, GLenum severity, GLsizei length, const GLchar* message, GLvoid* userParam) {
     std::cerr << retrace::callNo << ": ";
     std::cerr << "glDebugOutputCallback: ";
@@ -358,9 +425,35 @@ debugOutputCallback(GLenum source, GLenum type, GLuint id, GLenum severity, GLsi
 } /* namespace glretrace */
 
 
+class GLDumper : public retrace::Dumper {
+public:
+    image::Image *
+    getSnapshot(void) {
+        if (!glretrace::getCurrentContext()) {
+            return NULL;
+        }
+        return glstate::getDrawBufferImage();
+    }
+
+    bool
+    dumpState(std::ostream &os) {
+        glretrace::Context *currentContext = glretrace::getCurrentContext();
+        if (glretrace::insideGlBeginEnd ||
+            !currentContext) {
+            return false;
+        }
+        glstate::dumpCurrentContext(os);
+        return true;
+    }
+};
+
+static GLDumper glDumper;
+
+
 void
 retrace::setUp(void) {
     glws::init();
+    dumper = &glDumper;
 }
 
 
@@ -375,39 +468,24 @@ retrace::addCallbacks(retrace::Retracer &retracer)
 }
 
 
-image::Image *
-retrace::getSnapshot(void) {
-    if (!glretrace::currentDrawable) {
-        return NULL;
-    }
-
-    return glstate::getDrawBufferImage();
-}
-
-
-bool
-retrace::dumpState(std::ostream &os)
-{
-    if (glretrace::insideGlBeginEnd ||
-        !glretrace::currentDrawable ||
-        !glretrace::currentContext) {
-        return false;
-    }
-
-    glstate::dumpCurrentContext(os);
-
-    return true;
-}
-
 void
 retrace::flushRendering(void) {
-    glretrace::flushQueries();
-    glFlush();
+    glretrace::Context *currentContext = glretrace::getCurrentContext();
+    if (currentContext) {
+        glretrace::flushQueries();
+        glFlush();
+    }
 }
 
 void
 retrace::waitForInput(void) {
+    glretrace::Context *currentContext = glretrace::getCurrentContext();
+    if (currentContext) {
+        glretrace::flushQueries();
+        glFlush();
+    }
     while (glws::processEvents()) {
+        os::sleep(100*1000);
     }
 }