Profile guided optimization pass

indra/llmath/llrigginginfo.h

@@ -66,6 +66,10 @@ class LLJointRiggingInfoTab
     const LLJointRiggingInfo& operator[](S32 i) const { return mRigInfoPtr[i]; };
     bool needsUpdate() { return mNeedsUpdate; }
     void setNeedsUpdate(bool val) { mNeedsUpdate = val; }
+
+    LLJointRiggingInfo* begin() { return mRigInfoPtr; }
+    LLJointRiggingInfo* end() { return mRigInfoPtr + mSize; }
+
 private:
     // Not implemented
     LLJointRiggingInfoTab& operator=(const LLJointRiggingInfoTab& src);

indra/llprimitive/llmodel.cpp

@@ -1547,6 +1547,13 @@ void LLMeshSkinInfo::fromLLSD(LLSD& skin)
         mLockScaleIfJointPosition = false;
     }
 
+    // combine mBindShapeMatrix and mInvBindMatrix into mBindPoseMatrix
+    mBindPoseMatrix.resize(mInvBindMatrix.size());
+    for (U32 i = 0; i < mInvBindMatrix.size(); ++i)
+    {
+        matMul(mBindShapeMatrix, mInvBindMatrix[i], mBindPoseMatrix[i]);
+    }
+
     updateHash();
 }
 

indra/llprimitive/llmodel.h

@@ -56,12 +56,15 @@ class LLMeshSkinInfo : public LLRefCount
     LLUUID mMeshID;
     std::vector<std::string> mJointNames;
     mutable std::vector<S32> mJointNums;
-    typedef std::vector<LLMatrix4a, boost::alignment::aligned_allocator<LLMatrix4a, 16>> matrix_list_t;
+    typedef std::vector<LLMatrix4a> matrix_list_t;
     matrix_list_t mInvBindMatrix;
 
     // bones/joints position overrides
     matrix_list_t mAlternateBindMatrix;
 
+    // cached multiply of mBindShapeMatrix and mInvBindMatrix
+    matrix_list_t mBindPoseMatrix;
+
     LL_ALIGN_16(LLMatrix4a mBindShapeMatrix);
 
     float mPelvisOffset;

indra/llrender/llrendersphere.cpp

@@ -34,6 +34,8 @@
 #include "llerror.h"
 
 #include "llglheaders.h"
+#include "llvertexbuffer.h"
+#include "llglslshader.h"
 
 LLRenderSphere gSphere;
 
@@ -53,12 +55,20 @@ inline LLVector3 polar_to_cart(F32 latitude, F32 longitude)
 
 void LLRenderSphere::renderGGL()
 {
+    LL_PROFILE_ZONE_SCOPED;
     S32 const LATITUDE_SLICES = 20;
     S32 const LONGITUDE_SLICES = 30;
 
-    if (mSpherePoints.empty())
+    if (mVertexBuffer.isNull())
     {
         mSpherePoints.resize(LATITUDE_SLICES + 1);
+        mVertexBuffer = new LLVertexBuffer(LLVertexBuffer::MAP_VERTEX);
+
+        mVertexBuffer->allocateBuffer((U32)(LATITUDE_SLICES + 1) * (LONGITUDE_SLICES + 1), LATITUDE_SLICES * LONGITUDE_SLICES * 6);
+
+        LLStrider<LLVector3> v;
+        mVertexBuffer->getVertexStrider(v);
+
         for (S32 lat_i = 0; lat_i < LATITUDE_SLICES + 1; lat_i++)
         {
             mSpherePoints[lat_i].resize(LONGITUDE_SLICES + 1);
@@ -68,24 +78,52 @@ void LLRenderSphere::renderGGL()
                 F32 lon = (F32)lon_i / LONGITUDE_SLICES;
 
                 mSpherePoints[lat_i][lon_i] = polar_to_cart(lat, lon);
+                v[lat_i * (LONGITUDE_SLICES + 1) + lon_i] = mSpherePoints[lat_i][lon_i];
             }
         }
+
+        LLStrider<U16> i;
+        mVertexBuffer->getIndexStrider(i);
+
+        for (S32 lat_i = 0; lat_i < LATITUDE_SLICES; lat_i++)
+        {
+            for (S32 lon_i = 0; lon_i < LONGITUDE_SLICES; lon_i++)
+            {
+                i[(lat_i * LONGITUDE_SLICES + lon_i) * 6 + 0] = lat_i * (LONGITUDE_SLICES + 1) + lon_i;
+                i[(lat_i * LONGITUDE_SLICES + lon_i) * 6 + 1] = lat_i * (LONGITUDE_SLICES + 1) + lon_i + 1;
+                i[(lat_i * LONGITUDE_SLICES + lon_i) * 6 + 2] = (lat_i + 1) * (LONGITUDE_SLICES + 1) + lon_i;
+
+                i[(lat_i * LONGITUDE_SLICES + lon_i) * 6 + 3] = (lat_i + 1) * (LONGITUDE_SLICES + 1) + lon_i;
+                i[(lat_i * LONGITUDE_SLICES + lon_i) * 6 + 4] = lat_i * (LONGITUDE_SLICES + 1) + lon_i + 1;
+                i[(lat_i * LONGITUDE_SLICES + lon_i) * 6 + 5] = (lat_i + 1) * (LONGITUDE_SLICES + 1) + lon_i + 1;
+            }
+        }
+
+        mVertexBuffer->unmapBuffer();
     }
 
-    gGL.begin(LLRender::TRIANGLES);
 
-    for (S32 lat_i = 0; lat_i < LATITUDE_SLICES; lat_i++)
-    {
-        for (S32 lon_i = 0; lon_i < LONGITUDE_SLICES; lon_i++)
+    if (LLGLSLShader::sCurBoundShaderPtr->mAttributeMask == LLVertexBuffer::MAP_VERTEX)
+    { // shader expects only vertex positions in vertex buffer, use fast path
+        mVertexBuffer->setBuffer();
+        mVertexBuffer->drawRange(LLRender::TRIANGLES, 0, mVertexBuffer->getNumVerts(), mVertexBuffer->getNumIndices(), 0);
+    }
+    else
+    { //shader wants colors in the vertex stream, use slow path
+        gGL.begin(LLRender::TRIANGLES);
+        for (S32 lat_i = 0; lat_i < LATITUDE_SLICES; lat_i++)
         {
-            gGL.vertex3fv(mSpherePoints[lat_i][lon_i].mV);
-            gGL.vertex3fv(mSpherePoints[lat_i][lon_i+1].mV);
-            gGL.vertex3fv(mSpherePoints[lat_i+1][lon_i].mV);
+            for (S32 lon_i = 0; lon_i < LONGITUDE_SLICES; lon_i++)
+            {
+                gGL.vertex3fv(mSpherePoints[lat_i][lon_i].mV);
+                gGL.vertex3fv(mSpherePoints[lat_i][lon_i + 1].mV);
+                gGL.vertex3fv(mSpherePoints[lat_i + 1][lon_i].mV);
 
-            gGL.vertex3fv(mSpherePoints[lat_i+1][lon_i].mV);
-            gGL.vertex3fv(mSpherePoints[lat_i][lon_i+1].mV);
-            gGL.vertex3fv(mSpherePoints[lat_i+1][lon_i+1].mV);
+                gGL.vertex3fv(mSpherePoints[lat_i + 1][lon_i].mV);
+                gGL.vertex3fv(mSpherePoints[lat_i][lon_i + 1].mV);
+                gGL.vertex3fv(mSpherePoints[lat_i + 1][lon_i + 1].mV);
+            }
         }
+        gGL.end();
     }
-    gGL.end();
 }

indra/llrender/llrendersphere.h

@@ -45,6 +45,7 @@ class LLRenderSphere
 
 private:
     std::vector< std::vector<LLVector3> > mSpherePoints;
+    LLPointer<LLVertexBuffer> mVertexBuffer;
 };
 
 extern LLRenderSphere gSphere;

indra/llrender/llvertexbuffer.cpp

@@ -1341,8 +1341,6 @@ void LLVertexBuffer::unmapBuffer()
 
 void LLVertexBuffer::_mapBuffer()
 {
-    // must only be called from main thread
-    llassert(LLCoros::on_main_thread_main_coro());
     if (!mMapped)
     {
         mMapped = true;

indra/newview/llface.cpp

@@ -56,6 +56,7 @@
 #include "llvoavatar.h"
 #include "llsculptidsize.h"
 #include "llmeshrepository.h"
+#include "llskinningutil.h"
 
 #if LL_LINUX
 // Work-around spurious used before init warning on Vector4a
@@ -2184,28 +2185,88 @@ F32 LLFace::getTextureVirtualSize()
 
 bool LLFace::calcPixelArea(F32& cos_angle_to_view_dir, F32& radius)
 {
+    constexpr F32 PIXEL_AREA_UPDATE_PERIOD = 0.1f;
+    // this is an expensive operation and the result is valid (enough) for several frames
+    // don't update every frame
+    if (gFrameTimeSeconds - mLastPixelAreaUpdate < PIXEL_AREA_UPDATE_PERIOD)
+    {
+        return true;
+    }
+
     LL_PROFILE_ZONE_SCOPED_CATEGORY_FACE;
 
     //get area of circle around face
-
     LLVector4a center;
     LLVector4a size;
 
-
     if (isState(LLFace::RIGGED))
     {
-        //override with avatar bounding box
+        LL_PROFILE_ZONE_NAMED_CATEGORY_FACE("calcPixelArea - rigged");
+        //override with joint volume face joint bounding boxes
         LLVOAvatar* avatar = mVObjp->getAvatar();
+        bool hasRiggedExtents = false;
+
         if (avatar && avatar->mDrawable)
         {
-            center.load3(avatar->getPositionAgent().mV);
-            const LLVector4a* exts = avatar->mDrawable->getSpatialExtents();
-            size.setSub(exts[1], exts[0]);
+            LLVolume* volume = mVObjp->getVolume();
+            if (volume)
+            {
+                LLVolumeFace& face = volume->getVolumeFace(mTEOffset);
+
+                auto& rigInfo = face.mJointRiggingInfoTab;
+
+                if (rigInfo.needsUpdate())
+                {
+                    LLVOVolume* vo_volume = (LLVOVolume*)mVObjp.get();
+                    LLVOAvatar* avatar = mVObjp->getAvatar();
+                    const LLMeshSkinInfo* skin = vo_volume->getSkinInfo();
+                    LLSkinningUtil::updateRiggingInfo(skin, avatar, face);
+                }
+
+                // calculate the world space bounding box of the face by combining the bounding boxes of all the joints
+                LLVector4a& minp = mRiggedExtents[0];
+                LLVector4a& maxp = mRiggedExtents[1];
+                minp = LLVector4a(FLT_MAX, FLT_MAX, FLT_MAX);
+                maxp = LLVector4a(-FLT_MAX, -FLT_MAX, -FLT_MAX);
+
+                for (S32 i = 0; i < rigInfo.size(); i++)
+                {
+                    auto& jointInfo = rigInfo[i];
+                    if (jointInfo.isRiggedTo())
+                    {
+                        LLJoint* joint = avatar->getJoint(i);
+
+                        if (joint)
+                        {
+                            LLVector4a jointPos;
+
+                            LLMatrix4a worldMat;
+                            worldMat.loadu((F32*)&joint->getWorldMatrix().mMatrix[0][0]);
+
+                            LLVector4a extents[2];
+
+                            matMulBoundBox(worldMat, jointInfo.getRiggedExtents(), extents);
+
+                            minp.setMin(minp, extents[0]);
+                            maxp.setMax(maxp, extents[1]);
+                            hasRiggedExtents = true;
+                        }
+                    }
+                }
+            }
         }
-        else
+
+        if (!hasRiggedExtents)
         {
+            // no rigged extents, zero out bounding box and skip update
+            mRiggedExtents[0] = mRiggedExtents[1] = LLVector4a(0.f, 0.f, 0.f);
+
             return false;
         }
+
+        center.setAdd(mRiggedExtents[1], mRiggedExtents[0]);
+        center.mul(0.5f);
+        size.setSub(mRiggedExtents[1], mRiggedExtents[0]);
     }
     else
     {
@@ -2231,6 +2292,10 @@ bool LLFace::calcPixelArea(F32& cos_angle_to_view_dir, F32& radius)
     F32 app_angle = atanf((F32) sqrt(size_squared) / dist);
     radius = app_angle*LLDrawable::sCurPixelAngle;
     mPixelArea = radius*radius * 3.14159f;
+
+    // remember last update time, add 10% noise to avoid all faces updating at the same time
+    mLastPixelAreaUpdate = gFrameTimeSeconds + ll_frand() * PIXEL_AREA_UPDATE_PERIOD * 0.1f;
+
     LLVector4a x_axis;
     x_axis.load3(camera->getXAxis().mV);
     cos_angle_to_view_dir = lookAt.dot3(x_axis).getF32();

indra/newview/llface.h

@@ -234,8 +234,14 @@ class alignas(16) LLFace
     // return true if this face is in an alpha draw pool
     bool isInAlphaPool() const;
 public: //aligned members
+
+    // bounding box of face in drawable space
     LLVector4a      mExtents[2];
 
+    // cached bounding box of rigged face in world space
+    // calculated on-demand by LLFace::calcPixelArea and may not be up-to-date
+    LLVector4a  mRiggedExtents[2] = { LLVector4a(0,0,0), LLVector4a(0,0,0) };
+
 private:
     friend class LLViewerTextureList;
     F32         adjustPartialOverlapPixelArea(F32 cos_angle_to_view_dir, F32 radius );
@@ -301,7 +307,14 @@ class alignas(16) LLFace
     S32         mReferenceIndex;
     std::vector<S32> mRiggedIndex;
 
+    // gFrameTimeSeconds when mPixelArea was last updated
+    F32         mLastPixelAreaUpdate = 0.f;
+
+    // virtual size of face in texture area  (mPixelArea adjusted by texture repeats)
+    // used to determine desired resolution of texture
     F32         mVSize;
+
+    // pixel area face covers on screen
     F32         mPixelArea;
 
     //importance factor, in the range [0, 1.0].