gistfile1.cpp

    cal3d::SSEArray<CalCoreSubmesh::Vertex> MorphSubmeshCache;

    void accumulateMorphTarget(
        const cal3d::MorphTarget* morphTarget
    ) {
        // VC++ isn't hoisting this SSE register out of the loop, so do it manually.
        CalVector4 weight(morphTarget->weight); // _mm_load1_ps

        const CalCoreMorphTarget::VertexOffsetArray& vertexOffsets = morphTarget->coreMorphTarget->vertexOffsets;
        const VertexOffset* morphVertex = cal3d::pointerFromVector(vertexOffsets);
        const VertexOffset* lastMorphVertex = morphVertex + vertexOffsets.size();
        for (; morphVertex != lastMorphVertex; ++morphVertex) {
            size_t i = morphVertex->vertexId;
            MorphSubmeshCache[i].position += weight * morphVertex->position; // mulps; addps
            MorphSubmeshCache[i].normal   += weight * morphVertex->normal;   // mulps; addps
        }
    }

    const CalCoreSubmesh::Vertex* accumulateMorphTargets(
        size_t vertexCount,
        const CalCoreSubmesh::Vertex* sourceVertices,
        const cal3d::MorphTarget* morphTarget,
        const cal3d::MorphTarget* morphTargetEnd
    ) {
        if (vertexCount > MorphSubmeshCache.size()) {
            MorphSubmeshCache.destructive_resize(vertexCount);
        }

        cal3d::verify(morphTarget->weight != 0.0f, "Don't bother accumulating morphs until you have found an active morph target");

        // In theory we could initialize the accumulation buffer with
        // "morphTarget", since we know it's active.  It would reduce
        // memory traffic but increase code complexity, and probably
        // never matters in practice.
        std::copy(sourceVertices, sourceVertices + vertexCount, MorphSubmeshCache.begin());

        // Now find active morph targets and accumulate them
        while (morphTarget != morphTargetEnd) {
            if (morphTarget->weight != 0.0f) {
                accumulateMorphTarget(morphTarget);
            }
            ++morphTarget;
        }

        return cal3d::pointerFromVector(MorphSubmeshCache);
    }