CPU Profiling now includes all OpenGL calls (was only draw calls).

This only affects CPU profiling, not GPU or non-occluded pixels.

5 files changed, 94 insertions, 60 deletions

diff --git a/common/trace_profiler.cpp b/common/trace_profiler.cpp
index 2d126980..403c2609 100644
--- a/common/trace_profiler.cpp
+++ b/common/trace_profiler.cpp
@@ -55,27 +55,32 @@ void Profiler::setup(bool cpuTimes_, bool gpuTimes_, bool pixelsDrawn_)
     std::cout << "# call no gpu_start gpu_dura cpu_start cpu_dura pixels program name" << std::endl;
 }
 
-void Profiler::setBaseTimes(uint64_t gpuStart, uint64_t cpuStart)
+void Profiler::setBaseTimes(int64_t gpuStart, int64_t cpuStart)
 {
     baseCpuTime = cpuStart;
     baseGpuTime = gpuStart;
 }
 
+bool Profiler::hasBaseTimes()
+{
+    return baseCpuTime != 0 || baseGpuTime != 0;
+}
+
 void Profiler::addCall(unsigned no,
                        const char *name,
                        unsigned program,
-                       uint64_t pixels,
-                       uint64_t gpuStart, uint64_t gpuDuration,
-                       uint64_t cpuStart, uint64_t cpuDuration)
+                       int64_t pixels,
+                       int64_t gpuStart, int64_t gpuDuration,
+                       int64_t cpuStart, int64_t cpuDuration)
 {
-    if (gpuTimes) {
+    if (gpuTimes && gpuStart) {
         gpuStart -= baseGpuTime;
     } else {
         gpuStart = 0;
         gpuDuration = 0;
     }
 
-    if (cpuTimes) {
+    if (cpuTimes && cpuStart) {
         double cpuTimeScale = 1.0E9 / os::timeFrequency;
         cpuStart = (cpuStart - baseCpuTime) * cpuTimeScale;
         cpuDuration = cpuDuration * cpuTimeScale;
@@ -100,7 +105,7 @@ void Profiler::addCall(unsigned no,
               << std::endl;
 }
 
-void Profiler::addFrameStart(unsigned no, uint64_t gpuStart, uint64_t cpuStart)
+void Profiler::addFrameStart(unsigned no, int64_t gpuStart, int64_t cpuStart)
 {
     lastFrame.no = no;
     lastFrame.gpuStart = gpuStart;
@@ -126,9 +131,9 @@ void Profiler::addFrameStart(unsigned no, uint64_t gpuStart, uint64_t cpuStart)
               << std::endl;
 }
 
-void Profiler::addFrameEnd(uint64_t gpuEnd, uint64_t cpuEnd)
+void Profiler::addFrameEnd(int64_t gpuEnd, int64_t cpuEnd)
 {
-    uint64_t gpuDuration, cpuDuration;
+    int64_t gpuDuration, cpuDuration;
 
     if (gpuTimes) {
         gpuDuration = gpuEnd - lastFrame.gpuStart;
@@ -179,7 +184,9 @@ void Profiler::parseLine(const char* in, Profile* profile)
              >> call.program
              >> call.name;
 
-        profile->frames.back().calls.push_back(call);
+        if (call.pixels >= 0) {
+            profile->frames.back().calls.push_back(call);
+        }
     } else if (type.compare("frame_begin") == 0) {
         Profile::Frame frame;
         frame.gpuDuration = 0;
diff --git a/common/trace_profiler.hpp b/common/trace_profiler.hpp
index 78783f5a..dc3c4be0 100644
--- a/common/trace_profiler.hpp
+++ b/common/trace_profiler.hpp
@@ -71,23 +71,24 @@ public:
     ~Profiler();
 
     void setup(bool cpuTimes_, bool gpuTimes_, bool pixelsDrawn_);
-    void setBaseTimes(uint64_t gpuStart, uint64_t cpuStart);
+    void setBaseTimes(int64_t gpuStart, int64_t cpuStart);
+    bool hasBaseTimes();
 
-    void addFrameStart(unsigned no, uint64_t gpuStart, uint64_t cpuStart);
-    void addFrameEnd(uint64_t gpuEnd, uint64_t cpuEnd);
+    void addFrameStart(unsigned no, int64_t gpuStart, int64_t cpuStart);
+    void addFrameEnd(int64_t gpuEnd, int64_t cpuEnd);
 
     void addCall(unsigned no,
                  const char* name,
                  unsigned program,
-                 uint64_t pixels,
-                 uint64_t gpuStart, uint64_t gpuDuration,
-                 uint64_t cpuStart, uint64_t cpuDuration);
+                 int64_t pixels,
+                 int64_t gpuStart, int64_t gpuDuration,
+                 int64_t cpuStart, int64_t cpuDuration);
 
     static void parseLine(const char* line, Profile* profile);
 
 private:
-    uint64_t baseGpuTime;
-    uint64_t baseCpuTime;
+    int64_t baseGpuTime;
+    int64_t baseCpuTime;
 
     bool cpuTimes;
     bool gpuTimes;
@@ -95,8 +96,8 @@ private:
 
     struct {
         unsigned no;
-        uint64_t gpuStart;
-        uint64_t cpuStart;
+        int64_t gpuStart;
+        int64_t cpuStart;
     } lastFrame;
 };
 }
diff --git a/retrace/glretrace.hpp b/retrace/glretrace.hpp
index 27085d48..568049ec 100644
--- a/retrace/glretrace.hpp
+++ b/retrace/glretrace.hpp
@@ -90,8 +90,8 @@ void initContext();
 void updateDrawable(int width, int height);
 
 void flushQueries();
-void beginProfile(trace::Call &call);
-void endProfile(trace::Call &call);
+void beginProfile(trace::Call &call, bool isDraw);
+void endProfile(trace::Call &call, bool isDraw);
 
 } /* namespace glretrace */
 
diff --git a/retrace/glretrace.py b/retrace/glretrace.py
index 987798fe..6c73e697 100644
--- a/retrace/glretrace.py
+++ b/retrace/glretrace.py
@@ -310,9 +310,12 @@ class GlRetracer(Retracer):
         if function.name == 'glEndList':
             print r'    glretrace::insideList = false;'
 
-        if profileDraw and function.name != 'glEnd':
+        if function.name != 'glEnd':
             print r'    if (!glretrace::insideList && !glretrace::insideGlBeginEnd && retrace::profiling) {'
-            print r'        glretrace::beginProfile(call);'
+            if profileDraw:
+                print r'        glretrace::beginProfile(call, true);'
+            else:
+                print r'        glretrace::beginProfile(call, false);'
             print r'    }'
 
         if function.name == 'glCreateShaderProgramv':
@@ -349,10 +352,12 @@ class GlRetracer(Retracer):
         if function.name == "glBegin":
             print '    glretrace::insideGlBeginEnd = true;'
 
-        if profileDraw or function.name == 'glEnd':
-            print r'    if (!glretrace::insideList && !glretrace::insideGlBeginEnd && retrace::profiling) {'
-            print r'        glretrace::endProfile(call);'
-            print r'    }'
+        print r'    if (!glretrace::insideList && !glretrace::insideGlBeginEnd && retrace::profiling) {'
+        if profileDraw:
+            print r'        glretrace::endProfile(call, true);'
+        else:
+            print r'        glretrace::endProfile(call, false);'
+        print r'    }'
 
         # Error checking
         if function.name.startswith('gl'):
diff --git a/retrace/glretrace_main.cpp b/retrace/glretrace_main.cpp
index 8bd52762..c01b67d2 100755
--- a/retrace/glretrace_main.cpp
+++ b/retrace/glretrace_main.cpp
@@ -49,8 +49,8 @@ struct CallQuery
     unsigned call;
     GLuint program;
     const trace::FunctionSig *sig;
-    uint64_t cpuStart;
-    uint64_t cpuEnd;
+    int64_t cpuStart;
+    int64_t cpuEnd;
 };
 
 static bool supportsElapsed = true;
@@ -109,22 +109,22 @@ checkGlError(trace::Call &call) {
     os << "\n";
 }
 
-static inline GLuint64
+static inline GLint64
 getGpuTimestamp() {
     GLuint query = 0;
-    GLuint64 timestamp = 0;
+    GLint64 timestamp = 0;
 
     if (retrace::profilingGpuTimes && supportsTimestamp) {
         glGenQueries(1, &query);
         glQueryCounter(query, GL_TIMESTAMP);
-        glGetQueryObjectui64vEXT(query, GL_QUERY_RESULT, &timestamp);
+        glGetQueryObjecti64vEXT(query, GL_QUERY_RESULT, &timestamp);
         glDeleteQueries(1, &query);
     }
 
     return timestamp;
 }
 
-static inline GLuint64
+static inline GLint64
 getCpuTimestamp() {
     if (retrace::profilingCpuTimes) {
         return os::getTime();
@@ -136,14 +136,14 @@ getCpuTimestamp() {
 static void
 completeCallQuery(CallQuery& query) {
     /* Get call start and duration */
-    GLuint64 gpuStart = 0, gpuDuration = 0, cpuDuration = 0, samples = 0;
+    GLint64 gpuStart = 0, gpuDuration = 0, cpuDuration = 0, samples = 0;
 
     if (retrace::profilingGpuTimes) {
         if (supportsTimestamp) {
-            glGetQueryObjectui64vEXT(query.ids[0], GL_QUERY_RESULT, &gpuStart);
+            glGetQueryObjecti64vEXT(query.ids[0], GL_QUERY_RESULT, &gpuStart);
         }
 
-        glGetQueryObjectui64vEXT(query.ids[1], GL_QUERY_RESULT, &gpuDuration);
+        glGetQueryObjecti64vEXT(query.ids[1], GL_QUERY_RESULT, &gpuDuration);
     }
 
     if (retrace::profilingCpuTimes) {
@@ -151,7 +151,7 @@ completeCallQuery(CallQuery& query) {
     }
 
     if (retrace::profilingPixelsDrawn) {
-        glGetQueryObjectui64vEXT(query.ids[2], GL_QUERY_RESULT, &samples);
+        glGetQueryObjecti64vEXT(query.ids[2], GL_QUERY_RESULT, &samples);
     }
 
     glDeleteQueries(3, query.ids);
@@ -170,7 +170,7 @@ flushQueries() {
 }
 
 void
-beginProfile(trace::Call &call) {
+beginProfile(trace::Call &call, bool isDraw) {
     if (firstFrame) {
         frame_start();
     }
@@ -181,18 +181,20 @@ beginProfile(trace::Call &call) {
     query.sig = call.sig;
     query.program = glretrace::currentContext ? glretrace::currentContext->activeProgram : 0;
 
-    glGenQueries(3, query.ids);
+    if (isDraw) {
+        glGenQueries(3, query.ids);
 
-    if (retrace::profilingGpuTimes) {
-        if (supportsTimestamp) {
-            glQueryCounter(query.ids[0], GL_TIMESTAMP);
-        }
+        if (retrace::profilingGpuTimes) {
+            if (supportsTimestamp) {
+                glQueryCounter(query.ids[0], GL_TIMESTAMP);
+            }
 
-        glBeginQuery(GL_TIME_ELAPSED, query.ids[1]);
-    }
+            glBeginQuery(GL_TIME_ELAPSED, query.ids[1]);
+        }
 
-    if (retrace::profilingPixelsDrawn) {
-        glBeginQuery(GL_SAMPLES_PASSED, query.ids[2]);
+        if (retrace::profilingPixelsDrawn) {
+            glBeginQuery(GL_SAMPLES_PASSED, query.ids[2]);
+        }
     }
 
     if (retrace::profilingCpuTimes) {
@@ -203,24 +205,31 @@ beginProfile(trace::Call &call) {
 }
 
 void
-endProfile(trace::Call &call) {
+endProfile(trace::Call &call, bool isDraw) {
     if (retrace::profilingCpuTimes) {
         CallQuery& query = callQueries.back();
         query.cpuEnd = getCpuTimestamp();
-    }
 
-    if (retrace::profilingGpuTimes) {
-        glEndQuery(GL_TIME_ELAPSED);
+        if (!isDraw) {
+            retrace::profiler.addCall(query.call, query.sig->name, query.program, -1, 0, 0, query.cpuStart, query.cpuEnd - query.cpuStart);
+            callQueries.pop_back();
+        }
     }
 
-    if (retrace::profilingPixelsDrawn) {
-        glEndQuery(GL_SAMPLES_PASSED);
+    if (isDraw) {
+        if (retrace::profilingGpuTimes) {
+            glEndQuery(GL_TIME_ELAPSED);
+        }
+
+        if (retrace::profilingPixelsDrawn) {
+            glEndQuery(GL_SAMPLES_PASSED);
+        }
     }
 }
 
 void
 initContext() {
-    GLuint64 gpuTime, cpuTime;
+    GLint64 gpuTime = 0, cpuTime = 0;
     const char* extensions;
 
     extensions = (const char*)glGetString(GL_EXTENSIONS);
@@ -260,9 +269,17 @@ initContext() {
     }
 
     /* Sync the gpu and cpu start times */
-    gpuTime = getGpuTimestamp();
-    cpuTime = getCpuTimestamp();
-    retrace::profiler.setBaseTimes(gpuTime, cpuTime);
+    if (!retrace::profiler.hasBaseTimes()) {
+        if (retrace::profilingGpuTimes) {
+            gpuTime = getGpuTimestamp();
+        }
+
+        if (retrace::profilingCpuTimes) {
+            cpuTime = getCpuTimestamp();
+        }
+
+        retrace::profiler.setBaseTimes(gpuTime, cpuTime);
+    }
 }
 
 void
@@ -270,7 +287,9 @@ frame_start() {
     firstFrame = false;
 
     if (retrace::profiling) {
-        retrace::profiler.addFrameStart(retrace::frameNo, getGpuTimestamp(), getCpuTimestamp());
+        int64_t gpuStart = getGpuTimestamp();
+        int64_t cpuStart = getCpuTimestamp();
+        retrace::profiler.addFrameStart(retrace::frameNo, gpuStart, cpuStart);
     }
 }
 
@@ -281,7 +300,9 @@ frame_complete(trace::Call &call) {
         flushQueries();
 
         /* Indicate end of current frame */
-        retrace::profiler.addFrameEnd(getGpuTimestamp(), getCpuTimestamp());
+        int64_t gpuEnd = getGpuTimestamp();
+        int64_t cpuEnd = getCpuTimestamp();
+        retrace::profiler.addFrameEnd(gpuEnd, cpuEnd);
     }
 
     retrace::frameComplete(call);