Optimizations, SIMD and const correctness

libraries/render-utils/src/Model.cpp

@@ -1160,7 +1160,8 @@ void Model::updateClusterMatrices(glm::vec3 modelPosition, glm::quat modelOrient
     }
     _needsUpdateClusterMatrices = false;
     const FBXGeometry& geometry = getFBXGeometry();
-    glm::mat4 zeroScale(glm::vec4(0.0f, 0.0f, 0.0f, 0.0f),
+    static const glm::mat4 zeroScale(
+        glm::vec4(0.0f, 0.0f, 0.0f, 0.0f),
         glm::vec4(0.0f, 0.0f, 0.0f, 0.0f),
         glm::vec4(0.0f, 0.0f, 0.0f, 0.0f),
         glm::vec4(0.0f, 0.0f, 0.0f, 1.0f));
@@ -1170,11 +1171,17 @@ void Model::updateClusterMatrices(glm::vec3 modelPosition, glm::quat modelOrient
     for (int i = 0; i < _meshStates.size(); i++) {
         MeshState& state = _meshStates[i];
         const FBXMesh& mesh = geometry.meshes.at(i);
-
         for (int j = 0; j < mesh.clusters.size(); j++) {
             const FBXCluster& cluster = mesh.clusters.at(j);
             auto jointMatrix = _rig->getJointTransform(cluster.jointIndex);
+#if GLM_ARCH & GLM_ARCH_SSE2
+            glm::mat4 temp, out, inverseBindMatrix = cluster.inverseBindMatrix;
+            glm_mat4_mul((glm_vec4*)&modelToWorld, (glm_vec4*)&jointMatrix, (glm_vec4*)&temp);
+            glm_mat4_mul((glm_vec4*)&temp, (glm_vec4*)&inverseBindMatrix, (glm_vec4*)&out);
+            state.clusterMatrices[j] = out;
+#else 
             state.clusterMatrices[j] = modelToWorld * jointMatrix * cluster.inverseBindMatrix;
+#endif
 
             // as an optimization, don't build cautrizedClusterMatrices if the boneSet is empty.
             if (!_cauterizeBoneSet.empty()) {

libraries/render/src/render/Scene.cpp

@@ -35,7 +35,7 @@ void PendingChanges::updateItem(ItemID id, const UpdateFunctorPointer& functor)
     _updateFunctors.push_back(functor);
 }
 
-void PendingChanges::merge(PendingChanges& changes) {
+void PendingChanges::merge(const PendingChanges& changes) {
     _resetItems.insert(_resetItems.end(), changes._resetItems.begin(), changes._resetItems.end());
     _resetPayloads.insert(_resetPayloads.end(), changes._resetPayloads.begin(), changes._resetPayloads.end());
     _removedItems.insert(_removedItems.end(), changes._removedItems.begin(), changes._removedItems.end());
@@ -71,7 +71,7 @@ void Scene::enqueuePendingChanges(const PendingChanges& pendingChanges) {
 
 void consolidateChangeQueue(PendingChangesQueue& queue, PendingChanges& singleBatch) {
     while (!queue.empty()) {
-        auto pendingChanges = queue.front();
+        const auto& pendingChanges = queue.front();
         singleBatch.merge(pendingChanges);
         queue.pop();
     };

libraries/render/src/render/Scene.h

@@ -34,7 +34,7 @@ public:
     void updateItem(ItemID id, const UpdateFunctorPointer& functor);
     void updateItem(ItemID id) { updateItem(id, nullptr); }
 
-    void merge(PendingChanges& changes);
+    void merge(const PendingChanges& changes);
 
     ItemIDs _resetItems; 
     Payloads _resetPayloads;

libraries/shared/src/AABox.cpp

@@ -575,18 +575,18 @@ void AABox::transform(const Transform& transform) {
 
 // Logic based on http://clb.demon.fi/MathGeoLib/nightly/docs/AABB.cpp_code.html#471
 void AABox::transform(const glm::mat4& matrix) {
+    // FIXME use simd operations
     auto halfSize = _scale * 0.5f;
     auto center = _corner + halfSize;
     halfSize = abs(halfSize);
-    auto newCenter = transformPoint(matrix, center);
-
     auto mm = glm::transpose(glm::mat3(matrix));
     vec3 newDir = vec3(
-        glm::dot(glm::abs(vec3(mm[0])), halfSize),
-        glm::dot(glm::abs(vec3(mm[1])), halfSize),
-        glm::dot(glm::abs(vec3(mm[2])), halfSize)
+        glm::dot(glm::abs(mm[0]), halfSize),
+        glm::dot(glm::abs(mm[1]), halfSize),
+        glm::dot(glm::abs(mm[2]), halfSize)
     );
 
+    auto newCenter = transformPoint(matrix, center);
     _corner = newCenter - newDir;
     _scale = newDir * 2.0f;
 }

libraries/shared/src/GLMHelpers.cpp

@@ -502,7 +502,10 @@ glm::mat4 cancelOutRollAndPitch(const glm::mat4& m) {
 
 glm::vec3 transformPoint(const glm::mat4& m, const glm::vec3& p) {
     glm::vec4 temp = m * glm::vec4(p, 1.0f);
-    return glm::vec3(temp.x / temp.w, temp.y / temp.w, temp.z / temp.w);
+    if (temp.w != 1.0f) {
+        temp *= (1.0f / temp.w);
+    }
+    return glm::vec3(temp);
 }
 
 // does not handle non-uniform scale correctly, but it's faster then transformVectorFull

tests/shared/src/GLMHelpersTests.cpp

@@ -15,6 +15,8 @@
 #include <StreamUtils.h>
 
 #include <../QTestExtensions.h>
+#include <glm/gtc/matrix_transform.hpp>
+#include <glm/simd/matrix.h>
 
 
 QTEST_MAIN(GLMHelpersTests)
@@ -102,3 +104,39 @@ void GLMHelpersTests::testSixByteOrientationCompression() {
     testQuatCompression(-(ROT_Y_180 * ROT_Z_30 * ROT_X_90));
     testQuatCompression(-(ROT_Z_30 * ROT_X_90 * ROT_Y_180));
 }
+
+#define LOOPS 500000
+
+void GLMHelpersTests::testSimd() {
+    glm::mat4 a = glm::translate(glm::mat4(), vec3(1, 4, 9));
+    glm::mat4 b = glm::rotate(glm::mat4(), PI / 3, vec3(0, 1, 0));
+    glm::mat4 a1, b1;
+    glm::mat4 a2, b2;
+
+    a1 = a * b;
+    b1 = b * a;
+    glm_mat4_mul((glm_vec4*)&a, (glm_vec4*)&b, (glm_vec4*)&a2);
+    glm_mat4_mul((glm_vec4*)&b, (glm_vec4*)&a, (glm_vec4*)&b2);
+
+
+    {
+        QElapsedTimer timer;
+        timer.start();
+        for (size_t i = 0; i < LOOPS; ++i) {
+            a1 = a * b;
+            b1 = b * a;
+        }
+        qDebug() << "Native " << timer.elapsed();
+    }
+
+    {
+        QElapsedTimer timer;
+        timer.start();
+        for (size_t i = 0; i < LOOPS; ++i) {
+            glm_mat4_mul((glm_vec4*)&a, (glm_vec4*)&b, (glm_vec4*)&a2);
+            glm_mat4_mul((glm_vec4*)&b, (glm_vec4*)&a, (glm_vec4*)&b2);
+        }
+        qDebug() << "SIMD " << timer.elapsed();
+    }
+    qDebug() << "Done ";
+}

tests/shared/src/GLMHelpersTests.h

@@ -20,6 +20,7 @@ class GLMHelpersTests : public QObject {
 private slots:
     void testEulerDecomposition();
     void testSixByteOrientationCompression();
+    void testSimd();
 };
 
 float getErrorDifference(const float& a, const float& b);