DRTVWR-559 Optimization pass, make it so profileAvatar can read back GPU timer without a frame stall.

2 files changed, 64 insertions, 24 deletions

diff --git a/indra/llrender/llglslshader.cpp b/indra/llrender/llglslshader.cpp
index 04ac2476a7..b7f08aa9af 100644
--- a/indra/llrender/llglslshader.cpp
+++ b/indra/llrender/llglslshader.cpp
@@ -243,9 +243,9 @@ void LLGLSLShader::stopProfile()
     }
 }
 
-void LLGLSLShader::placeProfileQuery()
+void LLGLSLShader::placeProfileQuery(bool for_runtime)
 {
-    if (sProfileEnabled)
+    if (sProfileEnabled || for_runtime)
     {
         if (mTimerQuery == 0)
         {
@@ -254,42 +254,70 @@ void LLGLSLShader::placeProfileQuery()
             glGenQueries(1, &mPrimitivesQuery);
         }
 
-        glBeginQuery(GL_SAMPLES_PASSED, mSamplesQuery);
         glBeginQuery(GL_TIME_ELAPSED, mTimerQuery);
-        glBeginQuery(GL_PRIMITIVES_GENERATED, mPrimitivesQuery);
+
+        if (!for_runtime)
+        {
+            glBeginQuery(GL_SAMPLES_PASSED, mSamplesQuery);
+            glBeginQuery(GL_PRIMITIVES_GENERATED, mPrimitivesQuery);
+        }
     }
 }
 
-void LLGLSLShader::readProfileQuery()
+bool LLGLSLShader::readProfileQuery(bool for_runtime, bool force_read)
 {
-    if (sProfileEnabled)
+    if (sProfileEnabled || for_runtime)
     {
-        glEndQuery(GL_TIME_ELAPSED);
-        glEndQuery(GL_SAMPLES_PASSED);
-        glEndQuery(GL_PRIMITIVES_GENERATED);
+        if (!mProfilePending)
+        {
+            glEndQuery(GL_TIME_ELAPSED);
+            if (!for_runtime)
+            {
+                glEndQuery(GL_SAMPLES_PASSED);
+                glEndQuery(GL_PRIMITIVES_GENERATED);
+            }
+            mProfilePending = for_runtime;
+        }
+
+        if (mProfilePending && for_runtime && !force_read)
+        {
+            GLuint64 result = 0;
+            glGetQueryObjectui64v(mTimerQuery, GL_QUERY_RESULT_AVAILABLE, &result);
+
+            if (result != GL_TRUE)
+            {
+                return false;
+            }
+        }
 
         GLuint64 time_elapsed = 0;
         glGetQueryObjectui64v(mTimerQuery, GL_QUERY_RESULT, &time_elapsed);
+        mTimeElapsed += time_elapsed;
+        mProfilePending = false;
 
-        GLuint64 samples_passed = 0;
-        glGetQueryObjectui64v(mSamplesQuery, GL_QUERY_RESULT, &samples_passed);
+        if (!for_runtime)
+        {
+            GLuint64 samples_passed = 0;
+            glGetQueryObjectui64v(mSamplesQuery, GL_QUERY_RESULT, &samples_passed);
 
-        U64 primitives_generated = 0;
-        glGetQueryObjectui64v(mPrimitivesQuery, GL_QUERY_RESULT, &primitives_generated);
-        sTotalTimeElapsed += time_elapsed;
-        mTimeElapsed += time_elapsed;
+            U64 primitives_generated = 0;
+            glGetQueryObjectui64v(mPrimitivesQuery, GL_QUERY_RESULT, &primitives_generated);
+            sTotalTimeElapsed += time_elapsed;
 
-        sTotalSamplesDrawn += samples_passed;
-        mSamplesDrawn += samples_passed;
+            sTotalSamplesDrawn += samples_passed;
+            mSamplesDrawn += samples_passed;
 
-        U32 tri_count = (U32)primitives_generated / 3;
+            U32 tri_count = (U32)primitives_generated / 3;
 
-        mTrianglesDrawn += tri_count;
-        sTotalTrianglesDrawn += tri_count;
+            mTrianglesDrawn += tri_count;
+            sTotalTrianglesDrawn += tri_count;
 
-        sTotalBinds++;
-        mBinds++;
+            sTotalBinds++;
+            mBinds++;
+        }
     }
+
+    return true;
 }
 
 
diff --git a/indra/llrender/llglslshader.h b/indra/llrender/llglslshader.h
index 9d187c972c..3e7dae6669 100644
--- a/indra/llrender/llglslshader.h
+++ b/indra/llrender/llglslshader.h
@@ -168,8 +168,16 @@ public:
     void unload();
     void clearStats();
     void dumpStats();
-    void placeProfileQuery();
-    void readProfileQuery();
+
+    // place query objects for profiling if profiling is enabled
+    // if for_runtime is true, will place timer query only whether or not profiling is enabled
+    void placeProfileQuery(bool for_runtime = false);
+
+    // Readback query objects if profiling is enabled
+    // If for_runtime is true, will readback timer query iff query is available
+    // Will return false if a query is pending (try again later)
+    // If force_read is true, will force an immediate readback (severe performance penalty)
+    bool readProfileQuery(bool for_runtime = false, bool force_read = false);
 
     BOOL createShader(std::vector<LLStaticHashedString>* attributes,
         std::vector<LLStaticHashedString>* uniforms,
@@ -292,6 +300,7 @@ public:
     defines_map_t mDefines;
 
     //statistics for profiling shader performance
+    bool mProfilePending = false;
     U32 mTimerQuery;
     U32 mSamplesQuery;
     U32 mPrimitivesQuery;
@@ -308,6 +317,9 @@ public:
     // this pointer should be set to whichever shader represents this shader's rigged variant
     LLGLSLShader* mRiggedVariant = nullptr;
 
+    // hacky flag used for optimization in LLDrawPoolAlpha
+    bool mCanBindFast = false;
+
 #ifdef LL_PROFILER_ENABLE_RENDER_DOC
     void setLabel(const char* label);
 #endif