#5602 deleteProbe optimization pass

glGenQueries synchronizes cpu with gpu, which is expensive

3 files changed, 44 insertions, 2 deletions

diff --git a/indra/newview/llreflectionmap.cpp b/indra/newview/llreflectionmap.cpp
index 7f5076bd56..f4f3e39f17 100644
--- a/indra/newview/llreflectionmap.cpp
+++ b/indra/newview/llreflectionmap.cpp
@@ -45,7 +45,8 @@ LLReflectionMap::~LLReflectionMap()
 {
     if (mOcclusionQuery)
     {
-        glDeleteQueries(1, &mOcclusionQuery);
+        gPipeline.mReflectionMapManager.recycleQuery(mOcclusionQuery);
+        mOcclusionQuery = 0;
     }
 }
 
@@ -341,7 +342,7 @@ void LLReflectionMap::doOcclusion(const LLVector4a& eye)
     if (mOcclusionQuery == 0)
     { // no query was previously issued, allocate one and issue
         LL_PROFILE_ZONE_NAMED_CATEGORY_PIPELINE("rmdo - glGenQueries");
-        glGenQueries(1, &mOcclusionQuery);
+        mOcclusionQuery = gPipeline.mReflectionMapManager.allocateQuery();
         do_query = true;
     }
     else
diff --git a/indra/newview/llreflectionmapmanager.cpp b/indra/newview/llreflectionmapmanager.cpp
index c6fa64753c..d2e37fb40a 100644
--- a/indra/newview/llreflectionmapmanager.cpp
+++ b/indra/newview/llreflectionmapmanager.cpp
@@ -523,6 +523,7 @@ void LLReflectionMapManager::refreshSettings()
     mRenderReflectionProbeLevel = gSavedSettings.getS32("RenderReflectionProbeLevel");
     mRenderReflectionProbeCount = gSavedSettings.getU32("RenderReflectionProbeCount");
     mRenderReflectionProbeDynamicAllocation = gSavedSettings.getS32("RenderReflectionProbeDynamicAllocation");
+    cleanupQueryPool();
 }
 
 LLReflectionMap* LLReflectionMapManager::addProbe(LLSpatialGroup* group)
@@ -570,6 +571,25 @@ U32 LLReflectionMapManager::probeMemory()
     return (mDynamicProbeCount * 6 * (mProbeResolution * mProbeResolution) * 4) / 1024 / 1024 + (mDynamicProbeCount * 6 * (LL_IRRADIANCE_MAP_RESOLUTION * LL_IRRADIANCE_MAP_RESOLUTION) * 4) / 1024 / 1024;
 }
 
+GLuint LLReflectionMapManager::allocateQuery()
+{
+    if (mQueryPool.empty())
+    {
+        GLuint query;
+        glGenQueries(1, &query);
+        return query;
+    }
+
+    GLuint query = mQueryPool.front();
+    mQueryPool.pop_front();
+    return query;
+}
+
+void LLReflectionMapManager::recycleQuery(GLuint query)
+{
+    mQueryPool.push_back(query);
+}
+
 struct CompareProbeDepth
 {
     bool operator()(const LLReflectionMap* lhs, const LLReflectionMap* rhs)
@@ -713,6 +733,7 @@ void LLReflectionMapManager::deleteProbe(U32 i)
         other->mNeighbors.erase(iter);
     }
 
+    // Probes are distance sorted, order matters.
     mProbes.erase(mProbes.begin() + i);
 }
 
@@ -1549,10 +1570,23 @@ void LLReflectionMapManager::cleanup()
     glDeleteBuffers(1, &mUBO);
     mUBO = 0;
 
+    cleanupQueryPool();
+
     // note: also called on teleport (not just shutdown), so make sure we're in a good "starting" state
     initCubeFree();
 }
 
+void LLReflectionMapManager::cleanupQueryPool()
+{
+    if (!mQueryPool.empty())
+    {
+        LL_PROFILE_ZONE_NAMED_CATEGORY_DISPLAY("cleanup query pool");
+        std::vector<GLuint> queries(mQueryPool.begin(), mQueryPool.end());
+        glDeleteQueries(static_cast<GLsizei>(queries.size()), queries.data());
+        mQueryPool.clear();
+    }
+}
+
 void LLReflectionMapManager::doOcclusion()
 {
     LLVector4a eye;
diff --git a/indra/newview/llreflectionmapmanager.h b/indra/newview/llreflectionmapmanager.h
index 5daed7d1cf..1e1f9ad1a2 100644
--- a/indra/newview/llreflectionmapmanager.h
+++ b/indra/newview/llreflectionmapmanager.h
@@ -106,6 +106,7 @@ public:
 
     // release any GL state
     void cleanup();
+    void cleanupQueryPool();
 
     // maintain reflection probes
     void update();
@@ -164,6 +165,10 @@ public:
     U32 probeCount();
     U32 probeMemory();
 
+    // glDeleteQueries is expensive, so we maintain a pool of queries
+    GLuint allocateQuery();
+    void recycleQuery(GLuint query);
+
 private:
     friend class LLPipeline;
     friend class LLHeroProbeManager;
@@ -190,6 +195,8 @@ private:
     // bind UBO used for rendering
     void setUniforms();
 
+    std::deque<GLuint>                                    mQueryPool;
+
     // render target for cube snapshots
     // used to generate mipmaps without doing a copy-to-texture
     LLRenderTarget mRenderTarget;