Profile guided optimizations

5 files changed, 63 insertions, 102 deletions

diff --git a/indra/newview/llface.cpp b/indra/newview/llface.cpp
index ccfef09b09..a1ec75e34b 100644
--- a/indra/newview/llface.cpp
+++ b/indra/newview/llface.cpp
@@ -1741,7 +1741,7 @@ bool LLFace::getGeometryVolume(const LLVolume& volume,
             { //bump mapped or has material, just do the whole expensive loop
                 LL_PROFILE_ZONE_NAMED_CATEGORY_FACE("getGeometryVolume - texgen default");
 
-                std::vector<LLVector2> bump_tc;
+                LLStrider<LLVector2> bump_tc;
 
                 if (mat && !mat->getNormalID().isNull())
                 { //writing out normal and specular texture coordinates, not bump offsets
@@ -1803,49 +1803,70 @@ bool LLFace::getGeometryVolume(const LLVolume& volume,
                     }
                     const bool do_xform = (xforms & xform_channel) != XFORM_NONE;
 
+                    // hold onto strider to front of TC array for use later
+                    bump_tc = dst;
 
-                    for (S32 i = 0; i < num_vertices; i++)
                     {
-                        LLVector2 tc(vf.mTexCoords[i]);
-
-                        LLVector4a& norm = vf.mNormals[i];
-
-                        LLVector4a& center = *(vf.mCenter);
-
-                        if (texgen != LLTextureEntry::TEX_GEN_DEFAULT)
+                        // NOTE: split TEX_GEN_PLANAR implementation to reduce branchiness of inner loop
+                        // These are per-vertex operations and every little bit counts
+                        if (texgen == LLTextureEntry::TEX_GEN_PLANAR)
                         {
-                            LLVector4a vec = vf.mPositions[i];
+                            LL_PROFILE_ZONE_NAMED_CATEGORY_FACE("tgd - planar");
+                            for (S32 i = 0; i < num_vertices; i++)
+                            {
+                                LLVector2 tc(vf.mTexCoords[i]);
+                                LLVector4a& norm = vf.mNormals[i];
+                                LLVector4a& center = *(vf.mCenter);
+                                LLVector4a vec = vf.mPositions[i];
 
-                            vec.mul(scalea);
+                                vec.mul(scalea);
 
-                            if (texgen == LLTextureEntry::TEX_GEN_PLANAR)
-                            {
                                 planarProjection(tc, norm, center, vec);
-                            }
-                        }
 
-                        if (tex_mode && mTextureMatrix)
-                        {
-                            LLVector3 tmp(tc.mV[0], tc.mV[1], 0.f);
-                            tmp = tmp * *mTextureMatrix;
-                            tc.mV[0] = tmp.mV[0];
-                            tc.mV[1] = tmp.mV[1];
+                                if (tex_mode && mTextureMatrix)
+                                {
+                                    LLVector3 tmp(tc.mV[0], tc.mV[1], 0.f);
+                                    tmp = tmp * *mTextureMatrix;
+                                    tc.mV[0] = tmp.mV[0];
+                                    tc.mV[1] = tmp.mV[1];
+                                }
+                                else if (do_xform)
+                                {
+                                    xform(tc, cos_ang, sin_ang, os, ot, ms, mt);
+                                }
+
+                                *dst++ = tc;
+                            }
                         }
-                        else if (do_xform)
+                        else
                         {
-                            xform(tc, cos_ang, sin_ang, os, ot, ms, mt);
-                        }
+                            LL_PROFILE_ZONE_NAMED_CATEGORY_FACE("tgd - transform");
 
-                        *dst++ = tc;
-                        if (do_bump)
-                        {
-                            bump_tc.push_back(tc);
+                            for (S32 i = 0; i < num_vertices; i++)
+                            {
+                                LLVector2 tc(vf.mTexCoords[i]);
+
+                                if (tex_mode && mTextureMatrix)
+                                {
+                                    LLVector3 tmp(tc.mV[0], tc.mV[1], 0.f);
+                                    tmp = tmp * *mTextureMatrix;
+                                    tc.mV[0] = tmp.mV[0];
+                                    tc.mV[1] = tmp.mV[1];
+                                }
+                                else if (do_xform)
+                                {
+                                    xform(tc, cos_ang, sin_ang, os, ot, ms, mt);
+                                }
+
+                                *dst++ = tc;
+                            }
                         }
                     }
                 }
 
                 if ((!mat && !gltf_mat) && do_bump)
                 {
+                    LL_PROFILE_ZONE_NAMED_CATEGORY_FACE("tgd - do bump");
                     mVertexBuffer->getTexCoord1Strider(tex_coords1, mGeomIndex, mGeomCount);
 
                     mVObjp->getVolume()->genTangents(face_index);
diff --git a/indra/newview/llskinningutil.cpp b/indra/newview/llskinningutil.cpp
index 9b4ed4c946..1c92e06700 100644
--- a/indra/newview/llskinningutil.cpp
+++ b/indra/newview/llskinningutil.cpp
@@ -315,11 +315,9 @@ void LLSkinningUtil::initJointNums(LLMeshSkinInfo* skin, LLVOAvatar *avatar)
     }
 }
 
-static LLTrace::BlockTimerStatHandle FTM_FACE_RIGGING_INFO("Face Rigging Info");
-
 void LLSkinningUtil::updateRiggingInfo(const LLMeshSkinInfo* skin, LLVOAvatar *avatar, LLVolumeFace& vol_face)
 {
-    LL_RECORD_BLOCK_TIME(FTM_FACE_RIGGING_INFO);
+    LL_PROFILE_ZONE_SCOPED_CATEGORY_AVATAR;
 
     if (vol_face.mJointRiggingInfoTab.needsUpdate())
     {
diff --git a/indra/newview/llviewertexture.cpp b/indra/newview/llviewertexture.cpp
index 681d91c945..9e1cb84bd1 100644
--- a/indra/newview/llviewertexture.cpp
+++ b/indra/newview/llviewertexture.cpp
@@ -1361,51 +1361,6 @@ void LLViewerFetchedTexture::addToCreateTexture()
     }
     else
     {
-        LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE;
-#if 1
-        //
-        //if mRequestedDiscardLevel > mDesiredDiscardLevel, we assume the required image res keep going up,
-        //so do not scale down the over qualified image.
-        //Note: scaling down image is expensensive. Do it only when very necessary.
-        //
-        if(mRequestedDiscardLevel <= mDesiredDiscardLevel && !mForceToSaveRawImage)
-        {
-            U32 w = mFullWidth >> mRawDiscardLevel;
-            U32 h = mFullHeight >> mRawDiscardLevel;
-
-            //if big image, do not load extra data
-            //scale it down to size >= LLViewerTexture::sMinLargeImageSize
-            if(w * h > LLViewerTexture::sMinLargeImageSize)
-            {
-                S32 d_level = llmin(mRequestedDiscardLevel, (S32)mDesiredDiscardLevel) - mRawDiscardLevel;
-
-                if(d_level > 0)
-                {
-                    S32 i = 0;
-                    while((d_level > 0) && ((w >> i) * (h >> i) > LLViewerTexture::sMinLargeImageSize))
-                    {
-                        i++;
-                        d_level--;
-                    }
-                    if(i > 0)
-                    {
-                        mRawDiscardLevel += i;
-                        if(mRawDiscardLevel >= getDiscardLevel() && getDiscardLevel() > 0)
-                        {
-                            mNeedsCreateTexture = false;
-                            destroyRawImage();
-                            return;
-                        }
-
-                        {
-                            //make a duplicate in case somebody else is using this raw image
-                            mRawImage = mRawImage->scaled(w >> i, h >> i);
-                        }
-                    }
-                }
-            }
-        }
-#endif
         scheduleCreateTexture();
     }
     return;
diff --git a/indra/newview/llvoavatar.cpp b/indra/newview/llvoavatar.cpp
index b04bb99b97..26cb2a573a 100644
--- a/indra/newview/llvoavatar.cpp
+++ b/indra/newview/llvoavatar.cpp
@@ -10684,14 +10684,18 @@ void LLVOAvatar::updateRiggingInfo()
 
     std::map<LLUUID, S32> curr_rigging_info_key;
 
-    // Get current rigging info key
-    for (LLVOVolume* vol : volumes)
     {
-        if (vol->isMesh() && vol->getVolume())
+        LL_PROFILE_ZONE_NAMED_CATEGORY_AVATAR("update rig info - get key")
+
+        // Get current rigging info key
+        for (LLVOVolume* vol : volumes)
         {
-            const LLUUID& mesh_id = vol->getVolume()->getParams().getSculptID();
-            S32 max_lod = llmax(vol->getLOD(), vol->mLastRiggingInfoLOD);
-            curr_rigging_info_key[mesh_id] = max_lod;
+            if (vol->isMesh() && vol->getVolume())
+            {
+                const LLUUID& mesh_id = vol->getVolume()->getParams().getSculptID();
+                S32 max_lod = llmax(vol->getLOD(), vol->mLastRiggingInfoLOD);
+                curr_rigging_info_key[mesh_id] = max_lod;
+            }
         }
     }
 
diff --git a/indra/newview/llvovolume.cpp b/indra/newview/llvovolume.cpp
index 4b63354893..4f48a070e3 100644
--- a/indra/newview/llvovolume.cpp
+++ b/indra/newview/llvovolume.cpp
@@ -6028,8 +6028,8 @@ void LLVolumeGeometryManager::rebuildMesh(LLSpatialGroup* group)
 
             group->mBuilt = 1.f;
 
-            const U32 MAX_BUFFER_COUNT = 4096;
-            LLVertexBuffer* locked_buffer[MAX_BUFFER_COUNT];
+            static std::vector<LLVertexBuffer*> locked_buffer;
+            locked_buffer.resize(0);
 
             U32 buffer_count = 0;
 
@@ -6074,8 +6074,6 @@ void LLVolumeGeometryManager::rebuildMesh(LLSpatialGroup* group)
                                     group->dirtyGeom();
                                     gPipeline.markRebuild(group);
                                 }
-
-                                buff->unmapBuffer();
                             }
                         }
                     }
@@ -6091,17 +6089,7 @@ void LLVolumeGeometryManager::rebuildMesh(LLSpatialGroup* group)
 
             {
                 LL_PROFILE_ZONE_NAMED("rebuildMesh - flush");
-                for (LLVertexBuffer** iter = locked_buffer, ** end_iter = locked_buffer+buffer_count; iter != end_iter; ++iter)
-                {
-                    (*iter)->unmapBuffer();
-                }
-
-                // don't forget alpha
-                if(group != NULL &&
-                   !group->mVertexBuffer.isNull())
-                {
-                    group->mVertexBuffer->unmapBuffer();
-                }
+                LLVertexBuffer::flushBuffers();
             }
 
             group->clearState(LLSpatialGroup::MESH_DIRTY | LLSpatialGroup::NEW_DRAWINFO);
@@ -6783,11 +6771,6 @@ U32 LLVolumeGeometryManager::genDrawInfo(LLSpatialGroup* group, U32 mask, LLFace
 
             ++face_iter;
         }
-
-        if (buffer)
-        {
-            buffer->unmapBuffer();
-        }
     }
 
     group->mBufferMap[mask].clear();