1 files changed, 181 insertions, 34 deletions

diff --git a/linden/indra/newview/llviewerjointmesh.cpp b/linden/indra/newview/llviewerjointmesh.cpp
index c76990c..09fd015 100644
--- a/linden/indra/newview/llviewerjointmesh.cpp
+++ b/linden/indra/newview/llviewerjointmesh.cpp
@@ -31,14 +31,11 @@
 //-----------------------------------------------------------------------------
 #include "llviewerprecompiledheaders.h"
 
-#if LL_WINDOWS // For Intel vector classes
-        #include "fvec.h"
-#endif
-
 #include "imageids.h"
 #include "llfasttimer.h"
 
 #include "llagent.h"
+#include "llapr.h"
 #include "llbox.h"
 #include "lldrawable.h"
 #include "lldrawpoolavatar.h"
@@ -49,14 +46,19 @@
 #include "llglheaders.h"
 #include "lltexlayer.h"
 #include "llviewercamera.h"
+#include "llviewercontrol.h"
 #include "llviewerimagelist.h"
 #include "llviewerjointmesh.h"
 #include "llvoavatar.h"
 #include "llsky.h"
 #include "pipeline.h"
 #include "llglslshader.h"
+#include "llmath.h"
+#include "v4math.h"
+#include "m3math.h"
+#include "m4math.h"
 
-#if !LL_DARWIN && !LL_LINUX
+#if !LL_DARWIN && !LL_LINUX && !LL_SOLARIS
 extern PFNGLWEIGHTPOINTERARBPROC glWeightPointerARB;
 extern PFNGLWEIGHTFVARBPROC glWeightfvARB;
 extern PFNGLVERTEXBLENDARBPROC glVertexBlendARB;
@@ -68,6 +70,7 @@ static const U32 sRenderMask = LLVertexBuffer::MAP_VERTEX |
                                                            LLVertexBuffer::MAP_NORMAL |
                                                            LLVertexBuffer::MAP_TEXCOORD;
 
+
 //-----------------------------------------------------------------------------
 //-----------------------------------------------------------------------------
 // LLViewerJointMesh::LLSkinJoint
@@ -120,6 +123,7 @@ BOOL LLSkinJoint::setupSkinJoint( LLViewerJoint *joint)
         return TRUE;
 }
 
+
 //-----------------------------------------------------------------------------
 //-----------------------------------------------------------------------------
 // LLViewerJointMesh
@@ -414,9 +418,9 @@ const S32 NUM_AXES = 3;
 // rotation Z 0-n
 // pivot parent 0-n -- child = n+1
 
-static LLMatrix4 gJointMat[32];
-static LLMatrix3 gJointRot[32];
-static LLVector4 gJointPivot[32];
+static LLMatrix4        gJointMatUnaligned[32];
+static LLMatrix3        gJointRotUnaligned[32];
+static LLVector4        gJointPivot[32];
 
 //-----------------------------------------------------------------------------
 // uploadJointMatrices()
@@ -437,8 +441,8 @@ void LLViewerJointMesh::uploadJointMatrices()
                 {
                         joint_mat *= LLDrawPoolAvatar::getModelView();
                 }
-                gJointMat[joint_num] = joint_mat;
-                gJointRot[joint_num] = joint_mat.getMat3();
+                gJointMatUnaligned[joint_num] = joint_mat;
+                gJointRotUnaligned[joint_num] = joint_mat.getMat3();
         }
 
         BOOL last_pivot_uploaded = FALSE;
@@ -475,8 +479,8 @@ void LLViewerJointMesh::uploadJointMatrices()
         {
                 LLVector3 pivot;
                 pivot = LLVector3(gJointPivot[i]);
-                pivot = pivot * gJointRot[i];
-                gJointMat[i].translate(pivot);
+                pivot = pivot * gJointRotUnaligned[i];
+                gJointMatUnaligned[i].translate(pivot);
         }
 
         // upload matrices
@@ -487,11 +491,11 @@ void LLViewerJointMesh::uploadJointMatrices()
 
                 for (joint_num = 0; joint_num < reference_mesh->mJointRenderData.count(); joint_num++)
                 {
-                        gJointMat[joint_num].transpose();
+                        gJointMatUnaligned[joint_num].transpose();
 
                         for (S32 axis = 0; axis < NUM_AXES; axis++)
                         {
-                                F32* vector = gJointMat[joint_num].mMatrix[axis];
+                                F32* vector = gJointMatUnaligned[joint_num].mMatrix[axis];
                                 //glProgramLocalParameter4fvARB(GL_VERTEX_PROGRAM_ARB, LL_CHARACTER_MAX_JOINTS_PER_MESH * axis + joint_num+5, (GLfloat*)vector);
                                 U32 offset = LL_CHARACTER_MAX_JOINTS_PER_MESH*axis+joint_num;
                                 memcpy(mat+offset*4, vector, sizeof(GLfloat)*4);
@@ -903,21 +907,9 @@ BOOL LLViewerJointMesh::updateLOD(F32 pixel_area, BOOL activate)
         return (valid != activate);
 }
 
-
-void LLViewerJointMesh::updateGeometry()
+// static
+void LLViewerJointMesh::updateGeometryOriginal(LLFace *mFace, LLPolyMesh *mMesh)
 {
-        if (!(mValid
-                  && mMesh
-                  && mFace
-                  && mMesh->hasWeights()
-                  && mFace->mVertexBuffer.notNull()
-                  && LLShaderMgr::getVertexShaderLevel(LLShaderMgr::SHADER_AVATAR) == 0))
-        {
-                return;
-        }
-        
-        uploadJointMatrices();
-
         LLStrider<LLVector3> o_vertices;
         LLStrider<LLVector3> o_normals;
 
@@ -958,9 +950,9 @@ void LLViewerJointMesh::updateGeometry()
                 // No lerp required in this case.
                 if (w == 1.0f)
                 {
-                        gBlendMat = gJointMat[joint+1];
+                        gBlendMat = gJointMatUnaligned[joint+1];
                         o_vertices[bidx] = coords[index] * gBlendMat;
-                        gBlendRotMat = gJointRot[joint+1];
+                        gBlendRotMat = gJointRotUnaligned[joint+1];
                         o_normals[bidx] = normals[index] * gBlendRotMat;
                         continue;
                 }
@@ -968,8 +960,8 @@ void LLViewerJointMesh::updateGeometry()
                 // Try to keep all the accesses to the matrix data as close
                 // together as possible.  This function is a hot spot on the
                 // Mac. JC
-                LLMatrix4 &m0 = gJointMat[joint+1];
-                LLMatrix4 &m1 = gJointMat[joint+0];
+                LLMatrix4 &m0 = gJointMatUnaligned[joint+1];
+                LLMatrix4 &m1 = gJointMatUnaligned[joint+0];
                 
                 gBlendMat.mMatrix[VX][VX] = lerp(m1.mMatrix[VX][VX], m0.mMatrix[VX][VX], w);
                 gBlendMat.mMatrix[VX][VY] = lerp(m1.mMatrix[VX][VY], m0.mMatrix[VX][VY], w);
@@ -989,8 +981,8 @@ void LLViewerJointMesh::updateGeometry()
 
                 o_vertices[bidx] = coords[index] * gBlendMat;
                 
-                LLMatrix3 &n0 = gJointRot[joint+1];
-                LLMatrix3 &n1 = gJointRot[joint+0];
+                LLMatrix3 &n0 = gJointRotUnaligned[joint+1];
+                LLMatrix3 &n1 = gJointRotUnaligned[joint+0];
                 
                 gBlendRotMat.mMatrix[VX][VX] = lerp(n1.mMatrix[VX][VX], n0.mMatrix[VX][VX], w);
                 gBlendRotMat.mMatrix[VX][VY] = lerp(n1.mMatrix[VX][VY], n0.mMatrix[VX][VY], w);
@@ -1008,6 +1000,161 @@ void LLViewerJointMesh::updateGeometry()
         }
 }
 
+const U32 UPDATE_GEOMETRY_CALL_MASK                     = 0x1FFF; // 8K samples before overflow
+const U32 UPDATE_GEOMETRY_CALL_OVERFLOW         = ~UPDATE_GEOMETRY_CALL_MASK;
+static bool sUpdateGeometryCallPointer          = false;
+static F64 sUpdateGeometryGlobalTime            = 0.0 ;
+static F64 sUpdateGeometryElapsedTime           = 0.0 ;
+static F64 sUpdateGeometryElapsedTimeOff        = 0.0 ;
+static F64 sUpdateGeometryElapsedTimeOn         = 0.0 ;
+static F64 sUpdateGeometryRunAvgOff[10];
+static F64 sUpdateGeometryRunAvgOn[10];
+static U32 sUpdateGeometryRunCount                      = 0 ;
+static U32 sUpdateGeometryCalls                         = 0 ;
+static U32 sUpdateGeometryLastProcessor         = 0 ;
+void (*LLViewerJointMesh::sUpdateGeometryFunc)(LLFace* face, LLPolyMesh* mesh);
+
+void LLViewerJointMesh::updateGeometry()
+{
+        extern BOOL gVectorizePerfTest;
+        extern U32      gVectorizeProcessor;
+
+        if (!(mValid
+                  && mMesh
+                  && mFace
+                  && mMesh->hasWeights()
+                  && mFace->mVertexBuffer.notNull()
+                  && LLShaderMgr::getVertexShaderLevel(LLShaderMgr::SHADER_AVATAR) == 0))
+        {
+                return;
+        }
+
+        if (!gVectorizePerfTest)
+        {
+                // Once we've measured performance, just run the specified
+                // code version.
+                if(sUpdateGeometryFunc == updateGeometryOriginal)
+                        uploadJointMatrices();
+                sUpdateGeometryFunc(mFace, mMesh);
+        }
+        else
+        {
+                // At startup, measure the amount of time in skinning and choose
+                // the fastest one.
+                LLTimer ug_timer ;
+                
+                if (sUpdateGeometryCallPointer)
+                {
+                        if(sUpdateGeometryFunc == updateGeometryOriginal)
+                                uploadJointMatrices();
+                        // call accelerated version for this processor
+                        sUpdateGeometryFunc(mFace, mMesh);
+                }
+                else
+                {
+                        uploadJointMatrices();
+                        updateGeometryOriginal(mFace, mMesh);
+                }
+        
+                sUpdateGeometryElapsedTime += ug_timer.getElapsedTimeF64();
+                ++sUpdateGeometryCalls;
+                if(0 != (sUpdateGeometryCalls & UPDATE_GEOMETRY_CALL_OVERFLOW))
+                {
+                        F64 time_since_app_start = ug_timer.getElapsedSeconds();
+                        if(sUpdateGeometryGlobalTime == 0.0 
+                                || sUpdateGeometryLastProcessor != gVectorizeProcessor)
+                        {
+                                sUpdateGeometryGlobalTime               = time_since_app_start;
+                                sUpdateGeometryElapsedTime              = 0;
+                                sUpdateGeometryCalls                    = 0;
+                                sUpdateGeometryRunCount                 = 0;
+                                sUpdateGeometryLastProcessor    = gVectorizeProcessor;
+                                sUpdateGeometryCallPointer              = false;
+                                return;
+                        }
+                        F64 percent_time_in_function = 
+                                ( sUpdateGeometryElapsedTime * 100.0 ) / ( time_since_app_start - sUpdateGeometryGlobalTime ) ;
+                        sUpdateGeometryGlobalTime = time_since_app_start;
+                        if (!sUpdateGeometryCallPointer)
+                        {
+                                // First set of run data is with vectorization off.
+                                sUpdateGeometryCallPointer = true;
+                                llinfos << "profile (avg of " << sUpdateGeometryCalls << " samples) = "
+                                        << "vectorize off " << percent_time_in_function
+                                        << "% of time with "
+                                        << (sUpdateGeometryElapsedTime / (F64)sUpdateGeometryCalls)
+                                        << " seconds per call "
+                                        << llendl;
+                                sUpdateGeometryRunAvgOff[sUpdateGeometryRunCount] = percent_time_in_function;
+                                sUpdateGeometryElapsedTimeOff += sUpdateGeometryElapsedTime;
+                                sUpdateGeometryCalls = 0;
+                        }
+                        else
+                        {
+                                // Second set of run data is with vectorization on.
+                                sUpdateGeometryCallPointer = false;
+                                llinfos << "profile (avg of " << sUpdateGeometryCalls << " samples) = "
+                                        << "VEC on " << percent_time_in_function
+                                        << "% of time with "
+                                        << (sUpdateGeometryElapsedTime / (F64)sUpdateGeometryCalls)
+                                        << " seconds per call "
+                                        << llendl;
+                                sUpdateGeometryRunAvgOn[sUpdateGeometryRunCount] = percent_time_in_function ;
+                                sUpdateGeometryElapsedTimeOn += sUpdateGeometryElapsedTime;
+
+                                sUpdateGeometryCalls = 0;
+                                sUpdateGeometryRunCount++;
+                                F64 a = 0.0, b = 0.0;
+                                for(U32 i = 0; i<sUpdateGeometryRunCount; i++)
+                                {
+                                        a += sUpdateGeometryRunAvgOff[i];
+                                        b += sUpdateGeometryRunAvgOn[i];
+                                }
+                                a /= sUpdateGeometryRunCount;
+                                b /= sUpdateGeometryRunCount;
+                                F64 perf_boost = ( sUpdateGeometryElapsedTimeOff - sUpdateGeometryElapsedTimeOn ) / sUpdateGeometryElapsedTimeOn;
+                                llinfos << "run averages (" << (F64)sUpdateGeometryRunCount
+                                        << "/10) vectorize off " << a
+                                        << "% : vectorize type " << gVectorizeProcessor
+                                        << " " << b
+                                        << "% : performance boost " 
+                                        << perf_boost * 100.0
+                                        << "%"
+                                        << llendl ;
+                                if(sUpdateGeometryRunCount == 10)
+                                {
+                                        // In case user runs test again, force reset of data on
+                                        // next run.
+                                        sUpdateGeometryGlobalTime = 0.0;
+
+                                        // We have data now on which version is faster.  Switch to that
+                                        // code and save the data for next run.
+                                        gVectorizePerfTest = FALSE;
+                                        gSavedSettings.setBOOL("VectorizePerfTest", FALSE);
+
+                                        if (perf_boost > 0.0)
+                                        {
+                                                llinfos << "Vectorization improves avatar skinning performance, "
+                                                        << "keeping on for future runs."
+                                                        << llendl;
+                                                gSavedSettings.setBOOL("VectorizeSkin", TRUE);
+                                        }
+                                        else
+                                        {
+                                                // SIMD decreases performance, fall back to original code
+                                                llinfos << "Vectorization decreases avatar skinning performance, "
+                                                        << "switching back to original code."
+                                                        << llendl;
+
+                                                gSavedSettings.setBOOL("VectorizeSkin", FALSE);
+                                        }
+                                }
+                        }
+                        sUpdateGeometryElapsedTime = 0.0f;
+                }
+        }
+}
+
 void LLViewerJointMesh::dump()
 {
         if (mValid)