Vectorize LLRender UI scale and offset stack to resolve perf hotspot on arm

2 files changed, 21 insertions, 41 deletions

diff --git a/indra/llrender/llrender.cpp b/indra/llrender/llrender.cpp
index ec76aa4f36..c26e9b15b8 100644
--- a/indra/llrender/llrender.cpp
+++ b/indra/llrender/llrender.cpp
@@ -118,7 +118,7 @@ static const GLenum sGLBlendFactor[] =
 
 LLTexUnit::LLTexUnit(S32 index)
     : mCurrTexType(TT_NONE),
-    mCurrColorScale(1), mCurrAlphaScale(1), mCurrTexture(0),
+    mCurrTexture(0),
     mHasMipMaps(false),
     mIndex(index)
 {
@@ -600,26 +600,6 @@ GLint LLTexUnit::getTextureSourceType(eTextureBlendSrc src, bool isAlpha)
     }
 }
 
-void LLTexUnit::setColorScale(S32 scale)
-{
-    if (mCurrColorScale != scale || gGL.mDirty)
-    {
-        mCurrColorScale = scale;
-        gGL.flush();
-        glTexEnvi( GL_TEXTURE_ENV, GL_RGB_SCALE, scale );
-    }
-}
-
-void LLTexUnit::setAlphaScale(S32 scale)
-{
-    if (mCurrAlphaScale != scale || gGL.mDirty)
-    {
-        mCurrAlphaScale = scale;
-        gGL.flush();
-        glTexEnvi( GL_TEXTURE_ENV, GL_ALPHA_SCALE, scale );
-    }
-}
-
 // Useful for debugging that you've manually assigned a texture operation to the correct
 // texture unit based on the currently set active texture in opengl.
 void LLTexUnit::debugTextureUnit(void)
@@ -1283,9 +1263,7 @@ void LLRender::translateUI(F32 x, F32 y, F32 z)
         LL_ERRS() << "Need to push a UI translation frame before offsetting" << LL_ENDL;
     }
 
-    mUIOffset.back().mV[0] += x;
-    mUIOffset.back().mV[1] += y;
-    mUIOffset.back().mV[2] += z;
+    mUIOffset.back().add(LLVector4a(x, y, z));
 }
 
 void LLRender::scaleUI(F32 x, F32 y, F32 z)
@@ -1295,14 +1273,14 @@ void LLRender::scaleUI(F32 x, F32 y, F32 z)
         LL_ERRS() << "Need to push a UI transformation frame before scaling." << LL_ENDL;
     }
 
-    mUIScale.back().scaleVec(LLVector3(x,y,z));
+    mUIScale.back().mul(LLVector4a(x, y, z));
 }
 
 void LLRender::pushUIMatrix()
 {
     if (mUIOffset.empty())
     {
-        mUIOffset.emplace_back(0.f,0.f,0.f);
+        mUIOffset.emplace_back(0.f);
     }
     else
     {
@@ -1311,7 +1289,7 @@ void LLRender::pushUIMatrix()
 
     if (mUIScale.empty())
     {
-        mUIScale.emplace_back(1.f,1.f,1.f);
+        mUIScale.emplace_back(1.f);
     }
     else
     {
@@ -1333,18 +1311,20 @@ LLVector3 LLRender::getUITranslation()
 {
     if (mUIOffset.empty())
     {
-        return LLVector3(0,0,0);
+        return LLVector3::zero;
     }
-    return mUIOffset.back();
+
+    return LLVector3(mUIOffset.back().getF32ptr());
 }
 
 LLVector3 LLRender::getUIScale()
 {
     if (mUIScale.empty())
     {
-        return LLVector3(1,1,1);
+        return LLVector3::all_one;
     }
-    return mUIScale.back();
+
+    return LLVector3(mUIScale.back().getF32ptr());
 }
 
 
@@ -1354,8 +1334,9 @@ void LLRender::loadUIIdentity()
     {
         LL_ERRS() << "Need to push UI translation frame before clearing offset." << LL_ENDL;
     }
-    mUIOffset.back().setVec(0,0,0);
-    mUIScale.back().setVec(1,1,1);
+
+    mUIOffset.back().clear();
+    mUIScale.back().splat(1);
 }
 
 void LLRender::setColorMask(bool writeColor, bool writeAlpha)
@@ -1783,8 +1764,10 @@ void LLRender::vertex3f(const GLfloat& x, const GLfloat& y, const GLfloat& z)
     }
     else
     {
-        LLVector3 vert = (LLVector3(x,y,z)+mUIOffset.back()).scaledVec(mUIScale.back());
-        mVerticesp[mCount].set(vert.mV[VX], vert.mV[VY], vert.mV[VZ]);
+        LLVector4a vert(x, y, z);
+        vert.add(mUIOffset.back());
+        vert.mul(mUIScale.back());
+        mVerticesp[mCount] = vert;
     }
 
     mCount++;
diff --git a/indra/llrender/llrender.h b/indra/llrender/llrender.h
index 7b6bd4198b..9e68c2dcd1 100644
--- a/indra/llrender/llrender.h
+++ b/indra/llrender/llrender.h
@@ -43,6 +43,7 @@
 #include "llglheaders.h"
 #include "llmatrix4a.h"
 #include "glm/mat4x4.hpp"
+#include <boost/align/aligned_allocator.hpp>
 
 #include <array>
 #include <list>
@@ -227,13 +228,9 @@ protected:
     S32                 mIndex;
     U32                 mCurrTexture;
     eTextureType        mCurrTexType;
-    S32                 mCurrColorScale;
-    S32                 mCurrAlphaScale;
     bool                mHasMipMaps;
 
     void debugTextureUnit(void);
-    void setColorScale(S32 scale);
-    void setAlphaScale(S32 scale);
     GLint getTextureSource(eTextureBlendSrc src);
     GLint getTextureSourceType(eTextureBlendSrc src, bool isAlpha = false);
 };
@@ -526,8 +523,8 @@ private:
     eBlendFactor mCurrBlendAlphaSFactor;
     eBlendFactor mCurrBlendAlphaDFactor;
 
-    std::vector<LLVector3> mUIOffset;
-    std::vector<LLVector3> mUIScale;
+    std::vector<LLVector4a, boost::alignment::aligned_allocator<LLVector4a, 16> > mUIOffset;
+    std::vector<LLVector4a, boost::alignment::aligned_allocator<LLVector4a, 16> > mUIScale;
 };
 
 extern F32 gGLModelView[16];