/**
 * @file llmodel.cpp
 * @brief Model handling implementation
 *
 * $LicenseInfo:firstyear=2001&license=viewerlgpl$
 * Second Life Viewer Source Code
 * Copyright (C) 2010, Linden Research, Inc.
 *
 * This library is free software; you can redistribute it and/or
 * modify it under the terms of the GNU Lesser General Public
 * License as published by the Free Software Foundation;
 * version 2.1 of the License only.
 *
 * This library is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Lesser General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public
 * License along with this library; if not, write to the Free Software
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
 *
 * Linden Research, Inc., 945 Battery Street, San Francisco, CA  94111  USA
 * $/LicenseInfo$
 */

#include "linden_common.h"

#include "llmodel.h"
#include "llmemory.h"
#include "llconvexdecomposition.h"
#include "llsdserialize.h"
#include "llvector4a.h"
#include "hbxxh.h"

#ifdef LL_USESYSTEMLIBS
# include <zlib.h>
#else
# include "zlib-ng/zlib.h"
#endif

std::string model_names[] =
{
    "lowest_lod",
    "low_lod",
    "medium_lod",
    "high_lod",
    "physics_mesh"
};

const int MODEL_NAMES_LENGTH = sizeof(model_names) / sizeof(std::string);

LLModel::LLModel(const LLVolumeParams& params, F32 detail)
    : LLVolume(params, detail),
      mNormalizedScale(1,1,1),
      mNormalizedTranslation(0, 0, 0),
      mPelvisOffset( 0.0f ),
      mStatus(NO_ERRORS),
      mSubmodelID(0)
{
    mDecompID = -1;
    mLocalID = -1;
}

LLModel::~LLModel()
{
    if (mDecompID >= 0)
    {
        LLConvexDecomposition::getInstance()->deleteDecomposition(mDecompID);
    }
    mPhysics.mMesh.clear();
}

//static
std::string LLModel::getStatusString(U32 status)
{
    const static std::string status_strings[(S32)INVALID_STATUS] = {"status_no_error", "status_vertex_number_overflow","bad_element"};

    if(status < INVALID_STATUS)
    {
        if(status_strings[status] == std::string())
        {
            //LL_ERRS() << "No valid status string for this status: " << (U32)status << LL_ENDL();
        }
        return status_strings[status] ;
    }

    //LL_ERRS() << "Invalid model status: " << (U32)status << LL_ENDL();

    return std::string() ;
}


void LLModel::offsetMesh( const LLVector3& pivotPoint )
{
    LLVector4a pivot( pivotPoint[VX], pivotPoint[VY], pivotPoint[VZ] );

    for (std::vector<LLVolumeFace>::iterator faceIt = mVolumeFaces.begin(); faceIt != mVolumeFaces.end(); )
    {
        std::vector<LLVolumeFace>:: iterator currentFaceIt = faceIt++;
        LLVolumeFace& face = *currentFaceIt;
        LLVector4a *pos = (LLVector4a*) face.mPositions;

        for (S32 i=0; i<face.mNumVertices; ++i )
        {
            pos[i].add( pivot );
        }
    }
}

void LLModel::remapVolumeFaces()
{
    for (S32 i = 0; i < getNumVolumeFaces(); ++i)
    {
        mVolumeFaces[i].remap();
    }
}

void LLModel::optimizeVolumeFaces()
{
    for (S32 i = 0; i < getNumVolumeFaces(); ++i)
    {
        mVolumeFaces[i].optimize();
    }
}

struct MaterialBinding
{
    int             index;
    std::string     matName;
};

struct MaterialSort
{
    bool operator()(const MaterialBinding& lhs, const MaterialBinding& rhs)
    {
        return LLStringUtil::compareInsensitive(lhs.matName, rhs.matName) < 0;
    }
};

void LLModel::sortVolumeFacesByMaterialName()
{
    std::vector<MaterialBinding> bindings;
    bindings.resize(mVolumeFaces.size());

    for (int i = 0; i < bindings.size(); i++)
    {
        bindings[i].index = i;
        if(i < mMaterialList.size())
        {
            bindings[i].matName = mMaterialList[i];
        }
    }
    std::sort(bindings.begin(), bindings.end(), MaterialSort());
    std::vector< LLVolumeFace > new_faces;

    // remap the faces to be in the same order the mats now are...
    //
    new_faces.resize(bindings.size());
    for (int i = 0; i < bindings.size(); i++)
    {
        new_faces[i] = mVolumeFaces[bindings[i].index];
        if(i < mMaterialList.size())
        {
            mMaterialList[i] = bindings[i].matName;
        }
    }

    mVolumeFaces = new_faces;
}

void LLModel::trimVolumeFacesToSize(U32 new_count, LLVolume::face_list_t* remainder)
{
    llassert(new_count <= LL_SCULPT_MESH_MAX_FACES);

    if (new_count > 0 && ((U32)getNumVolumeFaces() > new_count))
    {
        // Copy out remaining volume faces for alternative handling, if provided
        //
        if (remainder)
        {
            (*remainder).assign(mVolumeFaces.begin() + new_count, mVolumeFaces.end());
        }

        // Trim down to the final set of volume faces (now stuffed to the gills!)
        //
        mVolumeFaces.resize(new_count);
    }
}

// Shrink the model to fit
// on a 1x1x1 cube centered at the origin.
// The positions and extents
// multiplied by  mNormalizedScale
// and offset by mNormalizedTranslation
// to be the "original" extents and position.
// Also, the positions will fit
// within the unit cube.
void LLModel::normalizeVolumeFaces()
{
    if (!mVolumeFaces.empty())
    {
        LLVector4a min, max;

        // For all of the volume faces
        // in the model, loop over
        // them and see what the extents
        // of the volume along each axis.
        min = mVolumeFaces[0].mExtents[0];
        max = mVolumeFaces[0].mExtents[1];

        for (U32 i = 1; i < mVolumeFaces.size(); ++i)
        {
            LLVolumeFace& face = mVolumeFaces[i];

            update_min_max(min, max, face.mExtents[0]);
            update_min_max(min, max, face.mExtents[1]);

            if (face.mTexCoords)
            {
                LLVector2& min_tc = face.mTexCoordExtents[0];
                LLVector2& max_tc = face.mTexCoordExtents[1];

                min_tc = face.mTexCoords[0];
                max_tc = face.mTexCoords[0];

                for (S32 j = 1; j < face.mNumVertices; ++j)
                {
                    update_min_max(min_tc, max_tc, face.mTexCoords[j]);
                }
            }
            else
            {
                face.mTexCoordExtents[0].set(0,0);
                face.mTexCoordExtents[1].set(1,1);
            }
        }

        // Now that we have the extents of the model
        // we can compute the offset needed to center
        // the model at the origin.

        // Compute center of the model
        // and make it negative to get translation
        // needed to center at origin.
        LLVector4a trans;
        trans.setAdd(min, max);
        trans.mul(-0.5f);

        // Compute the total size along all
        // axes of the model.
        LLVector4a size;
        size.setSub(max, min);

        // Prevent division by zero.
        F32 x = size[0];
        F32 y = size[1];
        F32 z = size[2];
        F32 w = size[3];
        if (fabs(x)<F_APPROXIMATELY_ZERO)
        {
            x = 1.0;
        }
        if (fabs(y)<F_APPROXIMATELY_ZERO)
        {
            y = 1.0;
        }
        if (fabs(z)<F_APPROXIMATELY_ZERO)
        {
            z = 1.0;
        }
        size.set(x,y,z,w);

        // Compute scale as reciprocal of size
        LLVector4a scale;
        scale.splat(1.f);
        scale.div(size);

        LLVector4a inv_scale(1.f);
        inv_scale.div(scale);

        for (U32 i = 0; i < mVolumeFaces.size(); ++i)
        {
            LLVolumeFace& face = mVolumeFaces[i];

            // We shrink the extents so
            // that they fall within
            // the unit cube.
            // VFExtents change
            face.mExtents[0].add(trans);
            face.mExtents[0].mul(scale);

            face.mExtents[1].add(trans);
            face.mExtents[1].mul(scale);

            // For all the positions, we scale
            // the positions to fit within the unit cube.
            LLVector4a* pos = (LLVector4a*) face.mPositions;
            LLVector4a* norm = (LLVector4a*) face.mNormals;
            LLVector4a* t = (LLVector4a*)face.mTangents;

            for (S32 j = 0; j < face.mNumVertices; ++j)
            {
                pos[j].add(trans);
                pos[j].mul(scale);
                if (norm && !norm[j].equals3(LLVector4a::getZero()))
                {
                    norm[j].mul(inv_scale);
                    norm[j].normalize3();
                }

                if (t)
                {
                    F32 w = t[j].getF32ptr()[3];
                    t[j].mul(inv_scale);
                    t[j].normalize3();
                    t[j].getF32ptr()[3] = w;
                }
            }
        }

        // mNormalizedScale is the scale at which
        // we would need to multiply the model
        // by to get the original size of the
        // model instead of the normalized size.
        LLVector4a normalized_scale;
        normalized_scale.splat(1.f);
        normalized_scale.div(scale);
        mNormalizedScale.set(normalized_scale.getF32ptr());
        mNormalizedTranslation.set(trans.getF32ptr());
        mNormalizedTranslation *= -1.f;

        // remember normalized scale so original dimensions can be recovered for mesh processing (i.e. tangent generation)
        for (auto& face : mVolumeFaces)
        {
            face.mNormalizedScale = mNormalizedScale;
        }
    }
}

void LLModel::getNormalizedScaleTranslation(LLVector3& scale_out, LLVector3& translation_out)
{
    scale_out = mNormalizedScale;
    translation_out = mNormalizedTranslation;
}

LLVector3 LLModel::getTransformedCenter(const LLMatrix4& mat)
{
    LLVector3 ret;

    if (!mVolumeFaces.empty())
    {
        LLMatrix4a m;
        m.loadu(mat);

        LLVector4a minv,maxv;

        LLVector4a t;
        m.affineTransform(mVolumeFaces[0].mPositions[0], t);
        minv = maxv = t;

        for (S32 i = 0; i < mVolumeFaces.size(); ++i)
        {
            LLVolumeFace& face = mVolumeFaces[i];

            for (S32 j = 0; j < face.mNumVertices; ++j)
            {
                m.affineTransform(face.mPositions[j],t);
                update_min_max(minv, maxv, t);
            }
        }

        minv.add(maxv);
        minv.mul(0.5f);

        ret.set(minv.getF32ptr());
    }

    return ret;
}



void LLModel::setNumVolumeFaces(S32 count)
{
    mVolumeFaces.resize(count);
}

void LLModel::setVolumeFaceData(
    S32 f,
    LLStrider<LLVector3> pos,
    LLStrider<LLVector3> norm,
    LLStrider<LLVector2> tc,
    LLStrider<U16> ind,
    U32 num_verts,
    U32 num_indices)
{
    llassert(num_indices % 3 == 0);

    LLVolumeFace& face = mVolumeFaces[f];

    face.resizeVertices(num_verts);
    face.resizeIndices(num_indices);

    LLVector4a::memcpyNonAliased16((F32*) face.mPositions, (F32*) pos.get(), num_verts*4*sizeof(F32));
    if (norm.get())
    {
        LLVector4a::memcpyNonAliased16((F32*) face.mNormals, (F32*) norm.get(), num_verts*4*sizeof(F32));
    }
    else
    {
        //ll_aligned_free_16(face.mNormals);
        face.mNormals = NULL;
    }

    if (tc.get())
    {
        U32 tex_size = (num_verts*2*sizeof(F32)+0xF)&~0xF;
        LLVector4a::memcpyNonAliased16((F32*) face.mTexCoords, (F32*) tc.get(), tex_size);
    }
    else
    {
        //ll_aligned_free_16(face.mTexCoords);
        face.mTexCoords = NULL;
    }

    U32 size = (num_indices*2+0xF)&~0xF;
    LLVector4a::memcpyNonAliased16((F32*) face.mIndices, (F32*) ind.get(), size);
}

void LLModel::addFace(const LLVolumeFace& face)
{
    if (face.mNumVertices == 0)
    {
        LL_ERRS() << "Cannot add empty face." << LL_ENDL;
    }

    mVolumeFaces.push_back(face);

    if (mVolumeFaces.size() > MAX_MODEL_FACES)
    {
        LL_ERRS() << "Model prims cannot have more than " << MAX_MODEL_FACES << " faces!" << LL_ENDL;
    }
}


void LLModel::generateNormals(F32 angle_cutoff)
{
    //generate normals for all faces by:
    // 1 - Create faceted copy of face with no texture coordinates
    // 2 - Weld vertices in faceted copy that are shared between triangles with less than "angle_cutoff" difference between normals
    // 3 - Generate smoothed set of normals based on welding results
    // 4 - Create faceted copy of face with texture coordinates
    // 5 - Copy smoothed normals to faceted copy, using closest normal to triangle normal where more than one normal exists for a given position
    // 6 - Remove redundant vertices from new faceted (now smooth) copy

    angle_cutoff = cosf(angle_cutoff);
    for (U32 j = 0; j < mVolumeFaces.size(); ++j)
    {
        LLVolumeFace& vol_face = mVolumeFaces[j];

        if (vol_face.mNumIndices > 65535)
        {
            LL_WARNS("MESHSKININFO") << "Too many vertices for normal generation to work." << LL_ENDL;
            continue;
        }

        //create faceted copy of current face with no texture coordinates (step 1)
        LLVolumeFace faceted;

        LLVector4a* src_pos = (LLVector4a*) vol_face.mPositions;
        //LLVector4a* src_norm = (LLVector4a*) vol_face.mNormals;


        faceted.resizeVertices(vol_face.mNumIndices);
        faceted.resizeIndices(vol_face.mNumIndices);
        //bake out triangles into temporary face, clearing texture coordinates
        for (S32 i = 0; i < vol_face.mNumIndices; ++i)
        {
            U32 idx = vol_face.mIndices[i];

            faceted.mPositions[i] = src_pos[idx];
            faceted.mTexCoords[i] = LLVector2(0,0);
            faceted.mIndices[i] = i;
        }

        //generate normals for temporary face
        for (S32 i = 0; i < faceted.mNumIndices; i += 3)
        { //for each triangle
            U16 i0 = faceted.mIndices[i+0];
            U16 i1 = faceted.mIndices[i+1];
            U16 i2 = faceted.mIndices[i+2];

            LLVector4a& p0 = faceted.mPositions[i0];
            LLVector4a& p1 = faceted.mPositions[i1];
            LLVector4a& p2 = faceted.mPositions[i2];

            LLVector4a& n0 = faceted.mNormals[i0];
            LLVector4a& n1 = faceted.mNormals[i1];
            LLVector4a& n2 = faceted.mNormals[i2];

            LLVector4a lhs, rhs;
            lhs.setSub(p1, p0);
            rhs.setSub(p2, p0);

            n0.setCross3(lhs, rhs);
            n0.normalize3();
            n1 = n0;
            n2 = n0;
        }

        //weld vertices in temporary face, respecting angle_cutoff (step 2)
        faceted.optimize(angle_cutoff);

        //generate normals for welded face based on new topology (step 3)

        for (S32 i = 0; i < faceted.mNumVertices; i++)
        {
            faceted.mNormals[i].clear();
        }

        for (S32 i = 0; i < faceted.mNumIndices; i += 3)
        { //for each triangle
            U16 i0 = faceted.mIndices[i+0];
            U16 i1 = faceted.mIndices[i+1];
            U16 i2 = faceted.mIndices[i+2];

            LLVector4a& p0 = faceted.mPositions[i0];
            LLVector4a& p1 = faceted.mPositions[i1];
            LLVector4a& p2 = faceted.mPositions[i2];

            LLVector4a& n0 = faceted.mNormals[i0];
            LLVector4a& n1 = faceted.mNormals[i1];
            LLVector4a& n2 = faceted.mNormals[i2];

            LLVector4a lhs, rhs;
            lhs.setSub(p1, p0);
            rhs.setSub(p2, p0);

            LLVector4a n;
            n.setCross3(lhs, rhs);

            n0.add(n);
            n1.add(n);
            n2.add(n);
        }

        //normalize normals and build point map
        LLVolumeFace::VertexMapData::PointMap point_map;

        for (S32 i = 0; i < faceted.mNumVertices; ++i)
        {
            faceted.mNormals[i].normalize3();

            LLVolumeFace::VertexMapData v;
            v.setPosition(faceted.mPositions[i]);
            v.setNormal(faceted.mNormals[i]);

            point_map[LLVector3(v.getPosition().getF32ptr())].push_back(v);
        }

        //create faceted copy of current face with texture coordinates (step 4)
        LLVolumeFace new_face;

        //bake out triangles into new face
        new_face.resizeIndices(vol_face.mNumIndices);
        new_face.resizeVertices(vol_face.mNumIndices);

        for (S32 i = 0; i < vol_face.mNumIndices; ++i)
        {
            U32 idx = vol_face.mIndices[i];
            LLVolumeFace::VertexData v;
            new_face.mPositions[i] = vol_face.mPositions[idx];
            new_face.mNormals[i].clear();
            new_face.mIndices[i] = i;
        }

        if (vol_face.mTexCoords)
        {
            for (S32 i = 0; i < vol_face.mNumIndices; i++)
            {
                U32 idx = vol_face.mIndices[i];
                new_face.mTexCoords[i] = vol_face.mTexCoords[idx];
            }
        }
        else
        {
            //ll_aligned_free_16(new_face.mTexCoords);
            new_face.mTexCoords = NULL;
        }

        //generate normals for new face
        for (S32 i = 0; i < new_face.mNumIndices; i += 3)
        { //for each triangle
            U16 i0 = new_face.mIndices[i+0];
            U16 i1 = new_face.mIndices[i+1];
            U16 i2 = new_face.mIndices[i+2];

            LLVector4a& p0 = new_face.mPositions[i0];
            LLVector4a& p1 = new_face.mPositions[i1];
            LLVector4a& p2 = new_face.mPositions[i2];

            LLVector4a& n0 = new_face.mNormals[i0];
            LLVector4a& n1 = new_face.mNormals[i1];
            LLVector4a& n2 = new_face.mNormals[i2];

            LLVector4a lhs, rhs;
            lhs.setSub(p1, p0);
            rhs.setSub(p2, p0);

            n0.setCross3(lhs, rhs);
            n0.normalize3();
            n1 = n0;
            n2 = n0;
        }

        //swap out normals in new_face with best match from point map (step 5)
        for (S32 i = 0; i < new_face.mNumVertices; ++i)
        {
            //LLVolumeFace::VertexData v = new_face.mVertices[i];

            LLVector4a ref_norm = new_face.mNormals[i];

            LLVolumeFace::VertexMapData::PointMap::iterator iter = point_map.find(LLVector3(new_face.mPositions[i].getF32ptr()));

            if (iter != point_map.end())
            {
                F32 best = -2.f;
                for (U32 k = 0; k < iter->second.size(); ++k)
                {
                    LLVector4a& n = iter->second[k].getNormal();

                    F32 cur = n.dot3(ref_norm).getF32();

                    if (cur > best)
                    {
                        best = cur;
                        new_face.mNormals[i] = n;
                    }
                }
            }
        }

        //remove redundant vertices from new face (step 6)
        new_face.optimize();

        mVolumeFaces[j] = new_face;
    }
}


std::string LLModel::getName() const
{
    return mRequestedLabel.empty() ? mLabel : mRequestedLabel;
}

//static
LLSD LLModel::writeModel(
    std::ostream& ostr,
    LLModel* physics,
    LLModel* high,
    LLModel* medium,
    LLModel* low,
    LLModel* impostor,
    const LLModel::Decomposition& decomp,
    bool upload_skin,
    bool upload_joints,
    bool lock_scale_if_joint_position,
    bool nowrite,
    bool as_slm,
    int submodel_id)
{
    LLSD mdl;

    LLModel* model[] =
    {
        impostor,
        low,
        medium,
        high,
        physics
    };

    bool skinning = upload_skin && high && !high->mSkinWeights.empty();

    if (skinning)
    { //write skinning block
        mdl["skin"] = high->mSkinInfo.asLLSD(upload_joints, lock_scale_if_joint_position);
    }

    if (!decomp.mBaseHull.empty() ||
        !decomp.mHull.empty())
    {
        mdl["physics_convex"] = decomp.asLLSD();
        if (!decomp.mHull.empty() && !as_slm)
        { //convex decomposition exists, physics mesh will not be used (unless this is an slm file)
            model[LLModel::LOD_PHYSICS] = NULL;
        }
    }
    else if (submodel_id)
    {
        const LLModel::Decomposition fake_decomp;
        mdl["secondary"] = true;
        mdl["submodel_id"] = submodel_id;
        mdl["physics_convex"] = fake_decomp.asLLSD();
        model[LLModel::LOD_PHYSICS] = NULL;
    }

    if (as_slm)
    { //save material list names
        for (U32 i = 0; i < high->mMaterialList.size(); ++i)
        {
            mdl["material_list"][i] = high->mMaterialList[i];
        }
    }

    for (U32 idx = 0; idx < MODEL_NAMES_LENGTH; ++idx)
    {
        if (model[idx] && (model[idx]->getNumVolumeFaces() > 0) && model[idx]->getVolumeFace(0).mPositions != NULL)
        {
            LLVector3 min_pos = LLVector3(model[idx]->getVolumeFace(0).mPositions[0].getF32ptr());
            LLVector3 max_pos = min_pos;

            //find position domain
            for (S32 i = 0; i < model[idx]->getNumVolumeFaces(); ++i)
            { //for each face
                const LLVolumeFace& face = model[idx]->getVolumeFace(i);
                for (S32 j = 0; j < face.mNumVertices; ++j)
                {
                    update_min_max(min_pos, max_pos, face.mPositions[j].getF32ptr());
                }
            }

            LLVector3 pos_range = max_pos - min_pos;

            for (S32 i = 0; i < model[idx]->getNumVolumeFaces(); ++i)
            { //for each face
                const LLVolumeFace& face = model[idx]->getVolumeFace(i);
                if (face.mNumVertices < 3)
                { //don't export an empty face
                    mdl[model_names[idx]][i]["NoGeometry"] = true;
                    continue;
                }
                LLSD::Binary verts(face.mNumVertices*3*2);
                LLSD::Binary tc(face.mNumVertices*2*2);
                LLSD::Binary normals(face.mNumVertices*3*2);
                LLSD::Binary tangents(face.mNumVertices * 4 * 2);
                LLSD::Binary indices(face.mNumIndices*2);

                U32 vert_idx = 0;
                U32 norm_idx = 0;
                //U32 tan_idx = 0;
                U32 tc_idx = 0;

                LLVector2* ftc = (LLVector2*) face.mTexCoords;
                LLVector2 min_tc;
                LLVector2 max_tc;

                if (ftc)
                {
                    min_tc = ftc[0];
                    max_tc = min_tc;

                    //get texture coordinate domain
                    for (S32 j = 0; j < face.mNumVertices; ++j)
                    {
                        update_min_max(min_tc, max_tc, ftc[j]);
                    }
                }

                LLVector2 tc_range = max_tc - min_tc;

                for (S32 j = 0; j < face.mNumVertices; ++j)
                { //for each vert

                    F32* pos = face.mPositions[j].getF32ptr();

                    //position
                    for (U32 k = 0; k < 3; ++k)
                    { //for each component
                        //convert to 16-bit normalized across domain
                        U16 val = (U16) (((pos[k]-min_pos.mV[k])/pos_range.mV[k])*65535);

                        U8* buff = (U8*) &val;
                        //write to binary buffer
                        verts[vert_idx++] = buff[0];
                        verts[vert_idx++] = buff[1];
                    }

                    if (face.mNormals)
                    { //normals
                        F32* norm = face.mNormals[j].getF32ptr();

                        for (U32 k = 0; k < 3; ++k)
                        { //for each component
                            //convert to 16-bit normalized
                            U16 val = (U16) ((norm[k]+1.f)*0.5f*65535);
                            U8* buff = (U8*) &val;

                            //write to binary buffer
                            normals[norm_idx++] = buff[0];
                            normals[norm_idx++] = buff[1];
                        }
                    }

#if 0 // keep this code for now in case we want to support transporting tangents with mesh assets
                    if (face.mTangents)
                    { //normals
                        F32* tangent = face.mTangents[j].getF32ptr();

                        for (U32 k = 0; k < 4; ++k)
                        { //for each component
                            //convert to 16-bit normalized
                            U16 val = (U16)((tangent[k] + 1.f) * 0.5f * 65535);
                            U8* buff = (U8*)&val;

                            //write to binary buffer
                            tangents[tan_idx++] = buff[0];
                            tangents[tan_idx++] = buff[1];
                        }
                    }
#endif

                    //texcoord
                    if (face.mTexCoords)
                    {
                        F32* src_tc = (F32*) face.mTexCoords[j].mV;

                        for (U32 k = 0; k < 2; ++k)
                        { //for each component
                            //convert to 16-bit normalized
                            U16 val = (U16) ((src_tc[k]-min_tc.mV[k])/tc_range.mV[k]*65535);

                            U8* buff = (U8*) &val;
                            //write to binary buffer
                            tc[tc_idx++] = buff[0];
                            tc[tc_idx++] = buff[1];
                        }
                    }
                }

                U32 idx_idx = 0;
                for (S32 j = 0; j < face.mNumIndices; ++j)
                {
                    U8* buff = (U8*) &(face.mIndices[j]);
                    indices[idx_idx++] = buff[0];
                    indices[idx_idx++] = buff[1];
                }

                //write out face data
                mdl[model_names[idx]][i]["PositionDomain"]["Min"] = min_pos.getValue();
                mdl[model_names[idx]][i]["PositionDomain"]["Max"] = max_pos.getValue();
                mdl[model_names[idx]][i]["NormalizedScale"] = face.mNormalizedScale.getValue();

                mdl[model_names[idx]][i]["Position"] = verts;

                if (face.mNormals)
                {
                    mdl[model_names[idx]][i]["Normal"] = normals;
                }

#if 0 // keep this code for now in case we decide to transport tangents with mesh assets
                if (face.mTangents)
                {
                    mdl[model_names[idx]][i]["Tangent"] = tangents;
                }
#endif

                if (face.mTexCoords)
                {
                    mdl[model_names[idx]][i]["TexCoord0Domain"]["Min"] = min_tc.getValue();
                    mdl[model_names[idx]][i]["TexCoord0Domain"]["Max"] = max_tc.getValue();
                    mdl[model_names[idx]][i]["TexCoord0"] = tc;
                }

                mdl[model_names[idx]][i]["TriangleList"] = indices;

                if (skinning)
                {
                    if (!model[idx]->mSkinWeights.empty())
                    {
                        //write out skin weights

                        //each influence list entry is up to 4 24-bit values
                        // first 8 bits is bone index
                        // last 16 bits is bone influence weight
                        // a bone index of 0xFF signifies no more influences for this vertex

                        std::stringstream ostr;
                        for (S32 j = 0; j < face.mNumVertices; ++j)
                        {
                            LLVector3 pos(face.mPositions[j].getF32ptr());

                            weight_list& weights = model[idx]->getJointInfluences(pos);

                            S32 count = 0;
                            for (weight_list::iterator iter = weights.begin(); iter != weights.end(); ++iter)
                            {
                                // Note joint index cannot exceed 255.
                                if (iter->mJointIdx < 255 && iter->mJointIdx >= 0)
                                {
                                    U8 idx = (U8)iter->mJointIdx;
                                    ostr.write((const char*)&idx, 1);

                                    U16 influence = (U16)(iter->mWeight * 65535);
                                    ostr.write((const char*)&influence, 2);

                                    ++count;
                                }
                            }
                            U8 end_list = 0xFF;
                            if (count < 4)
                            {
                                ostr.write((const char*)&end_list, 1);
                            }
                        }

                        //copy ostr to binary buffer
                        std::string data = ostr.str();
                        const U8* buff = (U8*)data.data();
                        U32 bytes = static_cast<U32>(data.size());

                        LLSD::Binary w(bytes);
                        for (U32 j = 0; j < bytes; ++j)
                        {
                            w[j] = buff[j];
                        }

                        mdl[model_names[idx]][i]["Weights"] = w;
                    }
                    else
                    {
                        if (idx == LLModel::LOD_PHYSICS)
                        {
                            // Ex: using "bounding box"
                            LL_DEBUGS("MESHSKININFO") << "Using physics model without skin weights" << LL_ENDL;
                        }
                        else
                        {
                            LL_WARNS("MESHSKININFO") << "Attempting to use skinning without having skin weights" << LL_ENDL;
                        }
                    }
                }
            }
        }
    }

    return writeModelToStream(ostr, mdl, nowrite, as_slm);
}

LLSD LLModel::writeModelToStream(std::ostream& ostr, LLSD& mdl, bool nowrite, bool as_slm)
{
    std::string::size_type cur_offset = 0;

    LLSD header;

    if (as_slm && mdl.has("material_list"))
    { //save material binding names to header
        header["material_list"] = mdl["material_list"];
    }

    std::string skin;

    if (mdl.has("skin"))
    { //write out skin block
        skin = zip_llsd(mdl["skin"]);

        U32 size = static_cast<U32>(skin.size());
        if (size > 0)
        {
            header["skin"]["offset"] = (LLSD::Integer) cur_offset;
            header["skin"]["size"] = (LLSD::Integer) size;
            cur_offset += size;
        }
    }

    std::string decomposition;

    if (mdl.has("physics_convex"))
    { //write out convex decomposition
        decomposition = zip_llsd(mdl["physics_convex"]);

        U32 size = static_cast<U32>(decomposition.size());
        if (size > 0)
        {
            header["physics_convex"]["offset"] = (LLSD::Integer) cur_offset;
            header["physics_convex"]["size"] = (LLSD::Integer) size;
            cur_offset += size;
        }
    }

        if (mdl.has("submodel_id"))
        { //write out submodel id
        header["submodel_id"] = (LLSD::Integer)mdl["submodel_id"];
        }

    std::string out[MODEL_NAMES_LENGTH];

    for (S32 i = 0; i < MODEL_NAMES_LENGTH; i++)
    {
        if (mdl.has(model_names[i]))
        {
            out[i] = zip_llsd(mdl[model_names[i]]);

            U32 size = static_cast<U32>(out[i].size());

            header[model_names[i]]["offset"] = (LLSD::Integer) cur_offset;
            header[model_names[i]]["size"] = (LLSD::Integer) size;
            cur_offset += size;
        }
    }

    if (!nowrite)
    {
        LLSDSerialize::toBinary(header, ostr);

        if (!skin.empty())
        { //write skin block
            ostr.write((const char*) skin.data(), header["skin"]["size"].asInteger());
        }

        if (!decomposition.empty())
        { //write decomposition block
            ostr.write((const char*) decomposition.data(), header["physics_convex"]["size"].asInteger());
        }

        for (S32 i = 0; i < MODEL_NAMES_LENGTH; i++)
        {
            if (!out[i].empty())
            {
                ostr.write((const char*) out[i].data(), header[model_names[i]]["size"].asInteger());
            }
        }
    }

    return header;
}

LLModel::weight_list& LLModel::getJointInfluences(const LLVector3& pos)
{
    //1. If a vertex has been weighted then we'll find it via pos and return its weight list
    weight_map::iterator iterPos = mSkinWeights.begin();
    weight_map::iterator iterEnd = mSkinWeights.end();

    if (mSkinWeights.empty())
    {
        // function calls iter->second on all return paths
        // everything that calls this function should precheck that there is data.
        LL_ERRS() << "called getJointInfluences with empty weights list" << LL_ENDL;
    }

    for ( ; iterPos!=iterEnd; ++iterPos )
    {
        if ( jointPositionalLookup( iterPos->first, pos ) )
        {
            return iterPos->second;
        }
    }

    //2. Otherwise we'll use the older implementation
    weight_map::iterator iter = mSkinWeights.find(pos);

    if (iter != mSkinWeights.end())
    {
        if ((iter->first - pos).magVec() > 0.1f)
        {
            LL_ERRS() << "Couldn't find weight list." << LL_ENDL;
        }

        return iter->second;
    }
    else
    {  //no exact match found, get closest point
        const F32 epsilon = 1e-5f;
        weight_map::iterator iter_up = mSkinWeights.lower_bound(pos);
        weight_map::iterator iter_down = iter_up;
        weight_map::iterator best = iter_up;
        if (iter_up != mSkinWeights.end())
        {
            iter_down = ++iter_up;
        }
        else
        {
            // Assumes that there is at least one element
            --best;
        }

        F32 min_dist = (iter->first - pos).magVec();

        bool done = false;
        while (!done)
        { //search up and down mSkinWeights from lower bound of pos until a
          //match is found within epsilon.  If no match is found within epsilon,
          //return closest match
            done = true;
            if (iter_up != mSkinWeights.end() && ++iter_up != mSkinWeights.end())
            {
                done = false;
                F32 dist = (iter_up->first - pos).magVec();

                if (dist < epsilon)
                {
                    return iter_up->second;
                }

                if (dist < min_dist)
                {
                    best = iter_up;
                    min_dist = dist;
                }
            }

            if (iter_down != mSkinWeights.begin() && --iter_down != mSkinWeights.begin())
            {
                done = false;

                F32 dist = (iter_down->first - pos).magVec();

                if (dist < epsilon)
                {
                    return iter_down->second;
                }

                if (dist < min_dist)
                {
                    best = iter_down;
                    min_dist = dist;
                }

            }
        }

        return best->second;
    }
}

void LLModel::setConvexHullDecomposition(
    const LLModel::convex_hull_decomposition& decomp)
{
    mPhysics.mHull = decomp;
    mPhysics.mMesh.clear();
    updateHullCenters();
}

void LLModel::updateHullCenters()
{
    mHullCenter.resize(mPhysics.mHull.size());
    mHullPoints = 0;
    mCenterOfHullCenters.clear();

    for (U32 i = 0; i < mPhysics.mHull.size(); ++i)
    {
        LLVector3 cur_center;

        for (U32 j = 0; j < mPhysics.mHull[i].size(); ++j)
        {
            cur_center += mPhysics.mHull[i][j];
        }
        mCenterOfHullCenters += cur_center;
        cur_center *= 1.f/mPhysics.mHull[i].size();
        mHullCenter[i] = cur_center;
        mHullPoints += static_cast<U32>(mPhysics.mHull[i].size());
    }

    if (mHullPoints > 0)
    {
        mCenterOfHullCenters *= 1.f / mHullPoints;
        llassert(mPhysics.hasHullList());
    }
}

bool LLModel::loadModel(std::istream& is)
{
    mSculptLevel = -1;  // default is an error occured

    LLSD header;
    {
        if (!LLSDSerialize::fromBinary(header, is, 1024*1024*1024))
        {
            LL_WARNS("MESHSKININFO") << "Mesh header parse error.  Not a valid mesh asset!" << LL_ENDL;
            return false;
        }
    }

    if (header.has("material_list"))
    { //load material list names
        mMaterialList.clear();
        for (U32 i = 0; i < header["material_list"].size(); ++i)
        {
            mMaterialList.push_back(header["material_list"][i].asString());
        }
    }

    mSubmodelID = header.has("submodel_id") ? header["submodel_id"].asInteger() : false;

    static const std::string lod_name[] =
    {
        "lowest_lod",
        "low_lod",
        "medium_lod",
        "high_lod",
        "physics_mesh",
    };

    const S32 MODEL_LODS = 5;

    S32 lod = llclamp((S32) mDetail, 0, MODEL_LODS);

    if (header[lod_name[lod]]["offset"].asInteger() == -1 ||
        header[lod_name[lod]]["size"].asInteger() == 0 )
    { //cannot load requested LOD
        LL_WARNS("MESHSKININFO") << "LoD data is invalid!" << LL_ENDL;
        return false;
    }

    bool has_skin = header["skin"]["offset"].asInteger() >=0 &&
                    header["skin"]["size"].asInteger() > 0;

    if ((lod == LLModel::LOD_HIGH) && !mSubmodelID)
    { //try to load skin info and decomp info
        std::ios::pos_type cur_pos = is.tellg();
        loadSkinInfo(header, is);
        is.seekg(cur_pos);
    }

    if ((lod == LLModel::LOD_HIGH || lod == LLModel::LOD_PHYSICS) && !mSubmodelID)
    {
        std::ios::pos_type cur_pos = is.tellg();
        loadDecomposition(header, is);
        is.seekg(cur_pos);
    }

    is.seekg(header[lod_name[lod]]["offset"].asInteger(), std::ios_base::cur);

    if (unpackVolumeFaces(is, header[lod_name[lod]]["size"].asInteger()))
    {
        if (has_skin)
        {
            //build out mSkinWeight from face info
            for (S32 i = 0; i < getNumVolumeFaces(); ++i)
            {
                const LLVolumeFace& face = getVolumeFace(i);

                if (face.mWeights)
                {
                    for (S32 j = 0; j < face.mNumVertices; ++j)
                    {
                        LLVector4a& w = face.mWeights[j];

                        std::vector<JointWeight> wght;

                        for (S32 k = 0; k < 4; ++k)
                        {
                            S32 idx = (S32) w[k];
                            F32 f = w[k] - idx;
                            if (f > 0.f)
                            {
                                wght.push_back(JointWeight(idx, f));
                            }
                        }

                        if (!wght.empty())
                        {
                            LLVector3 pos(face.mPositions[j].getF32ptr());
                            mSkinWeights[pos] = wght;
                        }
                    }
                }
            }
        }
        return true;
    }
    else
    {
        LL_WARNS("MESHSKININFO") << "unpackVolumeFaces failed!" << LL_ENDL;
    }

    return false;
}

bool LLModel::isMaterialListSubset( LLModel* ref )
{
    auto refCnt = ref->mMaterialList.size();
    auto modelCnt = mMaterialList.size();

    for (size_t src = 0; src < modelCnt; ++src)
    {
        bool foundRef = false;

        for (size_t dst = 0; dst < refCnt; ++dst)
        {
            //LL_INFOS()<<mMaterialList[src]<<" "<<ref->mMaterialList[dst]<<LL_ENDL;
            foundRef = mMaterialList[src] == ref->mMaterialList[dst];

            if ( foundRef )
            {
                break;
            }
        }

        if (!foundRef)
        {
            LL_INFOS("MESHSKININFO") << "Could not find material " << mMaterialList[src] << " in reference model " << ref->mLabel << LL_ENDL;
            return false;
        }
    }

    return true;
}

bool LLModel::needToAddFaces( LLModel* ref, int& refFaceCnt, int& modelFaceCnt )
{
    bool changed = false;
    if ( refFaceCnt< modelFaceCnt )
    {
        refFaceCnt += modelFaceCnt - refFaceCnt;
        changed = true;
    }
    else
    if ( modelFaceCnt < refFaceCnt )
    {
        modelFaceCnt += refFaceCnt - modelFaceCnt;
        changed = true;
    }

    return changed;
}

bool LLModel::matchMaterialOrder(LLModel* ref, int& refFaceCnt, int& modelFaceCnt )
{
    //Is this a subset?
    //LODs cannot currently add new materials, e.g.
    //1. ref = a,b,c lod1 = d,e => This is not permitted
    //2. ref = a,b,c lod1 = c => This would be permitted

    bool isASubset = isMaterialListSubset( ref );
    if ( !isASubset )
    {
        LL_INFOS("MESHSKININFO")<<"Material of model is not a subset of reference."<<LL_ENDL;
        return false;
    }

    if (mMaterialList.size() > ref->mMaterialList.size())
    {
        LL_INFOS("MESHSKININFO") << "Material of model has more materials than a reference." << LL_ENDL;
        // We passed isMaterialListSubset, so materials are a subset, but subset isn't supposed to be
        // larger than original and if we keep going, reordering will cause a crash
        return false;
    }

    std::map<std::string, U32> index_map;

    //build a map of material slot names to face indexes
    bool reorder = false;

    std::set<std::string> base_mat;
    std::set<std::string> cur_mat;

    for (U32 i = 0; i < mMaterialList.size(); i++)
    {
        index_map[ref->mMaterialList[i]] = i;
        //if any material name does not match reference, we need to reorder
        reorder |= ref->mMaterialList[i] != mMaterialList[i];
        base_mat.insert(ref->mMaterialList[i]);
        cur_mat.insert(mMaterialList[i]);
    }


    if (reorder &&  (base_mat == cur_mat)) //don't reorder if material name sets don't match
    {
        std::vector<LLVolumeFace> new_face_list;
        new_face_list.resize(mMaterialList.size());

        std::vector<std::string> new_material_list;
        new_material_list.resize(mMaterialList.size());

        //rebuild face list so materials have the same order
        //as the reference model
        for (U32 i = 0; i < mMaterialList.size(); ++i)
        {
            U32 ref_idx = index_map[mMaterialList[i]];

            if (i < mVolumeFaces.size())
            {
                new_face_list[ref_idx] = mVolumeFaces[i];
            }
            new_material_list[ref_idx] = mMaterialList[i];
        }

        llassert(new_material_list == ref->mMaterialList);

        mVolumeFaces = new_face_list;

        //override material list with reference model ordering
        mMaterialList = ref->mMaterialList;
    }

    return true;
}

bool LLModel::loadSkinInfo(LLSD& header, std::istream &is)
{
    S32 offset = header["skin"]["offset"].asInteger();
    S32 size = header["skin"]["size"].asInteger();

    if (offset >= 0 && size > 0)
    {
        is.seekg(offset, std::ios_base::cur);

        LLSD skin_data;

        if (LLUZipHelper::unzip_llsd(skin_data, is, size) == LLUZipHelper::ZR_OK)
        {
            mSkinInfo.fromLLSD(skin_data);
            return true;
        }
    }

    return false;
}

bool LLModel::loadDecomposition(LLSD& header, std::istream& is)
{
    S32 offset = header["physics_convex"]["offset"].asInteger();
    S32 size = header["physics_convex"]["size"].asInteger();

    if (offset >= 0 && size > 0 && !mSubmodelID)
    {
        is.seekg(offset, std::ios_base::cur);

        LLSD data;

        if (LLUZipHelper::unzip_llsd(data, is, size) == LLUZipHelper::ZR_OK)
        {
            mPhysics.fromLLSD(data);
            updateHullCenters();
        }
    }

    return true;
}

LLMeshSkinInfo::LLMeshSkinInfo():
    mPelvisOffset(0.0),
    mLockScaleIfJointPosition(false),
    mInvalidJointsScrubbed(false),
    mJointNumsInitialized(false)
{
}

LLMeshSkinInfo::LLMeshSkinInfo(LLSD& skin):
    mPelvisOffset(0.0),
    mLockScaleIfJointPosition(false),
    mInvalidJointsScrubbed(false),
    mJointNumsInitialized(false)
{
    fromLLSD(skin);
}

LLMeshSkinInfo::LLMeshSkinInfo(const LLUUID& mesh_id, LLSD& skin) :
    mMeshID(mesh_id),
    mPelvisOffset(0.0),
    mLockScaleIfJointPosition(false),
    mInvalidJointsScrubbed(false),
    mJointNumsInitialized(false)
{
    fromLLSD(skin);
}

void LLMeshSkinInfo::fromLLSD(LLSD& skin)
{
    if (skin.has("joint_names"))
    {
        for (U32 i = 0; i < skin["joint_names"].size(); ++i)
        {
            mJointNames.push_back(skin["joint_names"][i]);
            mJointNums.push_back(-1);
        }
    }

    if (skin.has("inverse_bind_matrix"))
    {
        for (U32 i = 0; i < skin["inverse_bind_matrix"].size(); ++i)
        {
            LLMatrix4 mat;
            for (U32 j = 0; j < 4; j++)
            {
                for (U32 k = 0; k < 4; k++)
                {
                    mat.mMatrix[j][k] = (F32)skin["inverse_bind_matrix"][i][j*4+k].asReal();
                }
            }

            mInvBindMatrix.push_back(LLMatrix4a(mat));
        }

        if (mJointNames.size() != mInvBindMatrix.size())
        {
            LL_WARNS("MESHSKININFO") << "Joints vs bind matrix count mismatch. Dropping joint bindings." << LL_ENDL;
            mJointNames.clear();
            mJointNums.clear();
            mInvBindMatrix.clear();
        }
    }

    if (skin.has("bind_shape_matrix"))
    {
        LLMatrix4 mat;
        for (U32 j = 0; j < 4; j++)
        {
            for (U32 k = 0; k < 4; k++)
            {
                mat.mMatrix[j][k] = (F32)skin["bind_shape_matrix"][j*4+k].asReal();
            }
        }
        mBindShapeMatrix.loadu(mat);
    }

    if (skin.has("alt_inverse_bind_matrix"))
    {
        for (U32 i = 0; i < skin["alt_inverse_bind_matrix"].size(); ++i)
        {
            LLMatrix4 mat;
            for (U32 j = 0; j < 4; j++)
            {
                for (U32 k = 0; k < 4; k++)
                {
                    mat.mMatrix[j][k] = (F32)skin["alt_inverse_bind_matrix"][i][j*4+k].asReal();
                }
            }

            mAlternateBindMatrix.push_back(LLMatrix4a(mat));
        }
    }

    if (skin.has("pelvis_offset"))
    {
        mPelvisOffset = (F32)skin["pelvis_offset"].asReal();
    }

    if (skin.has("lock_scale_if_joint_position"))
    {
        mLockScaleIfJointPosition = skin["lock_scale_if_joint_position"].asBoolean();
    }
    else
    {
        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();
}

LLSD LLMeshSkinInfo::asLLSD(bool include_joints, bool lock_scale_if_joint_position) const
{
    LLSD ret;

    for (U32 i = 0; i < mJointNames.size(); ++i)
    {
        ret["joint_names"][i] = mJointNames[i];

        for (U32 j = 0; j < 4; j++)
        {
            for (U32 k = 0; k < 4; k++)
            {
                ret["inverse_bind_matrix"][i][j*4+k] = mInvBindMatrix[i].mMatrix[j][k];
            }
        }
    }

    for (U32 i = 0; i < 4; i++)
    {
        for (U32 j = 0; j < 4; j++)
        {
            ret["bind_shape_matrix"][i*4+j] = mBindShapeMatrix.mMatrix[i][j];
        }
    }

    if ( include_joints && mAlternateBindMatrix.size() > 0 )
    {
        for (U32 i = 0; i < mJointNames.size(); ++i)
        {
            for (U32 j = 0; j < 4; j++)
            {
                for (U32 k = 0; k < 4; k++)
                {
                    ret["alt_inverse_bind_matrix"][i][j*4+k] = mAlternateBindMatrix[i].mMatrix[j][k];
                }
            }
        }

        if (lock_scale_if_joint_position)
        {
            ret["lock_scale_if_joint_position"] = lock_scale_if_joint_position;
        }

        ret["pelvis_offset"] = mPelvisOffset;
    }

    return ret;
}

void LLMeshSkinInfo::updateHash()
{
    //  get hash of data relevant to render batches
    HBXXH64 hash;

    //mJointNames
    for (auto& name : mJointNames)
    {
        hash.update(name);
    }

    //mJointNums
    hash.update((const void*)mJointNums.data(), sizeof(S32) * mJointNums.size());

    //mInvBindMatrix
    F32* src = mInvBindMatrix[0].getF32ptr();

    for (size_t i = 0, count = mInvBindMatrix.size() * 16; i < count; ++i)
    {
        S32 t = ll_round(src[i] * 10000.f);
        hash.update((const void*)&t, sizeof(S32));
    }
    //hash.update((const void*)mInvBindMatrix.data(), sizeof(LLMatrix4a) * mInvBindMatrix.size());

    mHash = hash.digest();
}

U32 LLMeshSkinInfo::sizeBytes() const
{
    U32 res = sizeof(LLUUID); // mMeshID

    res += sizeof(std::vector<std::string>) + sizeof(std::string) * static_cast<U32>(mJointNames.size());
    for (U32 i = 0; i < mJointNames.size(); ++i)
    {
        res += static_cast<U32>(mJointNames[i].size()); // actual size, not capacity
    }

    res += sizeof(std::vector<S32>) + sizeof(S32) * static_cast<U32>(mJointNums.size());
    res += sizeof(std::vector<LLMatrix4>) + 16 * sizeof(float) * static_cast<U32>(mInvBindMatrix.size());
    res += sizeof(std::vector<LLMatrix4>) + 16 * sizeof(float) * static_cast<U32>(mAlternateBindMatrix.size());
    res += 16 * sizeof(float); //mBindShapeMatrix
    res += sizeof(float) + 3 * sizeof(bool);

    return res;
}

LLModel::Decomposition::Decomposition(LLSD& data)
{
    fromLLSD(data);
}

void LLModel::Decomposition::fromLLSD(LLSD& decomp)
{
    if (decomp.has("HullList") && decomp.has("Positions"))
    {
        // updated for const-correctness. gcc is picky about this type of thing - Nyx
        const LLSD::Binary& hulls = decomp["HullList"].asBinary();
        const LLSD::Binary& position = decomp["Positions"].asBinary();

        U16* p = (U16*) &position[0];

        mHull.resize(hulls.size());

        LLVector3 min;
        LLVector3 max;
        LLVector3 range;

        if (decomp.has("Min"))
        {
            min.setValue(decomp["Min"]);
            max.setValue(decomp["Max"]);
        }
        else
        {
            min.set(-0.5f, -0.5f, -0.5f);
            max.set(0.5f, 0.5f, 0.5f);
        }

        range = max-min;

        for (U32 i = 0; i < hulls.size(); ++i)
        {
            U16 count = (hulls[i] == 0) ? 256 : hulls[i];

            std::set<U64> valid;

            //must have at least 4 points
            //llassert(count > 3);

            for (U32 j = 0; j < count; ++j)
            {
                U64 test = (U64) p[0] | ((U64) p[1] << 16) | ((U64) p[2] << 32);
                //point must be unique
                //llassert(valid.find(test) == valid.end());
                valid.insert(test);

                mHull[i].push_back(LLVector3(
                    (F32) p[0]/65535.f*range.mV[0]+min.mV[0],
                    (F32) p[1]/65535.f*range.mV[1]+min.mV[1],
                    (F32) p[2]/65535.f*range.mV[2]+min.mV[2]));
                p += 3;


            }

            //each hull must contain at least 4 unique points
            //llassert(valid.size() > 3);
        }
    }

    if (decomp.has("BoundingVerts"))
    {
        const LLSD::Binary& position = decomp["BoundingVerts"].asBinary();

        U16* p = (U16*) &position[0];

        LLVector3 min;
        LLVector3 max;
        LLVector3 range;

        if (decomp.has("Min"))
        {
            min.setValue(decomp["Min"]);
            max.setValue(decomp["Max"]);
        }
        else
        {
            min.set(-0.5f, -0.5f, -0.5f);
            max.set(0.5f, 0.5f, 0.5f);
        }

        range = max-min;

        U16 count = (U16)(position.size()/6);

        for (U32 j = 0; j < count; ++j)
        {
            mBaseHull.push_back(LLVector3(
                (F32) p[0]/65535.f*range.mV[0]+min.mV[0],
                (F32) p[1]/65535.f*range.mV[1]+min.mV[1],
                (F32) p[2]/65535.f*range.mV[2]+min.mV[2]));
            p += 3;
        }
    }
    else
    {
        //empty base hull mesh to indicate decomposition has been loaded
        //but contains no base hull
        mBaseHullMesh.clear();
    }
}

U32 LLModel::Decomposition::sizeBytes() const
{
    U32 res = sizeof(LLUUID); // mMeshID

    res += sizeof(LLModel::convex_hull_decomposition) + sizeof(std::vector<LLVector3>) * static_cast<U32>(mHull.size());
    for (U32 i = 0; i < mHull.size(); ++i)
    {
        res += static_cast<U32>(mHull[i].size()) * sizeof(LLVector3);
    }

    res += sizeof(LLModel::hull) + sizeof(LLVector3) * static_cast<U32>(mBaseHull.size());

    res += sizeof(std::vector<LLModel::PhysicsMesh>) + sizeof(std::vector<LLModel::PhysicsMesh>) * static_cast<U32>(mMesh.size());
    for (U32 i = 0; i < mMesh.size(); ++i)
    {
        res += mMesh[i].sizeBytes();
    }

    res += sizeof(std::vector<LLModel::PhysicsMesh>) * 2;
    res += mBaseHullMesh.sizeBytes() + mPhysicsShapeMesh.sizeBytes();

    return res;
}

bool LLModel::Decomposition::hasHullList() const
{
    return !mHull.empty() ;
}

LLSD LLModel::Decomposition::asLLSD() const
{
    LLSD ret;

    if (mBaseHull.empty() && mHull.empty())
    { //nothing to write
        return ret;
    }

    //write decomposition block
    // ["physics_convex"]["HullList"] -- list of 8 bit integers, each entry represents a hull with specified number of points
    // ["physics_convex"]["Position"] -- list of 16-bit integers to be decoded to given domain, encoded 3D points
    // ["physics_convex"]["BoundingVerts"] -- list of 16-bit integers to be decoded to given domain, encoded 3D points representing a single hull approximation of given shape

    //get minimum and maximum
    LLVector3 min;

    if (mHull.empty())
    {
        min = mBaseHull[0];
    }
    else
    {
        min = mHull[0][0];
    }

    LLVector3 max = min;

    LLSD::Binary hulls(mHull.size());

    U32 total = 0;

    for (U32 i = 0; i < mHull.size(); ++i)
    {
        U32 size = static_cast<U32>(mHull[i].size());
        total += size;
        hulls[i] = (U8) (size);

        for (U32 j = 0; j < mHull[i].size(); ++j)
        {
            update_min_max(min, max, mHull[i][j]);
        }
    }

    for (U32 i = 0; i < mBaseHull.size(); ++i)
    {
        update_min_max(min, max, mBaseHull[i]);
    }

    ret["Min"] = min.getValue();
    ret["Max"] = max.getValue();

    LLVector3 range = max-min;

    if (!hulls.empty())
    {
        ret["HullList"] = hulls;
    }

    if (total > 0)
    {
        LLSD::Binary p(total*3*2);

        U32 vert_idx = 0;

        for (U32 i = 0; i < mHull.size(); ++i)
        {
            std::set<U64> valid;

            llassert(!mHull[i].empty());

            for (U32 j = 0; j < mHull[i].size(); ++j)
            {
                U64 test = 0;
                const F32* src = mHull[i][j].mV;

                for (U32 k = 0; k < 3; k++)
                {
                    //convert to 16-bit normalized across domain
                    U16 val = (U16) (((src[k]-min.mV[k])/range.mV[k])*65535);

                    if(valid.size() < 3)
                    {
                        switch (k)
                        {
                            case 0: test = test | (U64) val; break;
                            case 1: test = test | ((U64) val << 16); break;
                            case 2: test = test | ((U64) val << 32); break;
                        };

                        valid.insert(test);
                    }

                    U8* buff = (U8*) &val;
                    //write to binary buffer
                    p[vert_idx++] = buff[0];
                    p[vert_idx++] = buff[1];

                    //makes sure we haven't run off the end of the array
                    llassert(vert_idx <= p.size());
                }
            }

            //must have at least 3 unique points
            llassert(valid.size() > 2);
        }

        ret["Positions"] = p;
    }

    //llassert(!mBaseHull.empty());

    if (!mBaseHull.empty())
    {
        LLSD::Binary p(mBaseHull.size()*3*2);

        U32 vert_idx = 0;
        for (U32 j = 0; j < mBaseHull.size(); ++j)
        {
            const F32* v = mBaseHull[j].mV;

            for (U32 k = 0; k < 3; k++)
            {
                //convert to 16-bit normalized across domain
                U16 val = (U16) (((v[k]-min.mV[k])/range.mV[k])*65535);

                U8* buff = (U8*) &val;
                //write to binary buffer
                p[vert_idx++] = buff[0];
                p[vert_idx++] = buff[1];

                if (vert_idx > p.size())
                {
                    LL_ERRS() << "Index out of bounds" << LL_ENDL;
                }
            }
        }

        ret["BoundingVerts"] = p;
    }

    return ret;
}

void LLModel::Decomposition::merge(const LLModel::Decomposition* rhs)
{
    if (!rhs)
    {
        return;
    }

    if (mMeshID != rhs->mMeshID)
    {
        LL_ERRS() << "Attempted to merge with decomposition of some other mesh." << LL_ENDL;
    }

    if (mBaseHull.empty())
    { //take base hull and decomposition from rhs
        mHull = rhs->mHull;
        mBaseHull = rhs->mBaseHull;
        mMesh = rhs->mMesh;
        mBaseHullMesh = rhs->mBaseHullMesh;
    }

    if (mPhysicsShapeMesh.empty())
    { //take physics shape mesh from rhs
        mPhysicsShapeMesh = rhs->mPhysicsShapeMesh;
    }
}

bool ll_is_degenerate(const LLVector4a& a, const LLVector4a& b, const LLVector4a& c, F32 tolerance)
{
    // small area check
    {
        LLVector4a edge1; edge1.setSub( a, b );
        LLVector4a edge2; edge2.setSub( a, c );
        //////////////////////////////////////////////////////////////////////////
        /// Linden Modified
        //////////////////////////////////////////////////////////////////////////

        // If no one edge is more than 10x longer than any other edge, we weaken
        // the tolerance by a factor of 1e-4f.

        LLVector4a edge3; edge3.setSub( c, b );
        const F32 len1sq = edge1.dot3(edge1).getF32();
        const F32 len2sq = edge2.dot3(edge2).getF32();
        const F32 len3sq = edge3.dot3(edge3).getF32();
        bool abOK = (len1sq <= 100.f * len2sq) && (len1sq <= 100.f * len3sq);
        bool acOK = (len2sq <= 100.f * len1sq) && (len1sq <= 100.f * len3sq);
        bool cbOK = (len3sq <= 100.f * len1sq) && (len1sq <= 100.f * len2sq);
        if ( abOK && acOK && cbOK )
        {
            tolerance *= 1e-4f;
        }

        //////////////////////////////////////////////////////////////////////////
        /// End Modified
        //////////////////////////////////////////////////////////////////////////

        LLVector4a cross; cross.setCross3( edge1, edge2 );

        LLVector4a edge1b; edge1b.setSub( b, a );
        LLVector4a edge2b; edge2b.setSub( b, c );
        LLVector4a crossb; crossb.setCross3( edge1b, edge2b );

        if ( ( cross.dot3(cross).getF32() < tolerance ) || ( crossb.dot3(crossb).getF32() < tolerance ))
        {
            return true;
        }
    }

    // point triangle distance check
    {
        LLVector4a Q; Q.setSub(a, b);
        LLVector4a R; R.setSub(c, b);

        const F32 QQ = dot3fpu(Q, Q);
        const F32 RR = dot3fpu(R, R);
        const F32 QR = dot3fpu(R, Q);

        volatile F32 QQRR = QQ * RR;
        volatile F32 QRQR = QR * QR;
        F32 Det = (QQRR - QRQR);

        if( Det == 0.0f )
        {
            return true;
        }
    }

    return false;
}

bool validate_face(const LLVolumeFace& face)
{
    for (S32 i = 0; i < face.mNumIndices; ++i)
    {
        if (face.mIndices[i] >= face.mNumVertices)
        {
            LL_WARNS("MESHSKININFO") << "Face has invalid index." << LL_ENDL;
            return false;
        }
    }

    if (face.mNumIndices % 3 != 0 || face.mNumIndices == 0)
    {
        LL_WARNS("MESHSKININFO") << "Face has invalid number of indices." << LL_ENDL;
        return false;
    }

    /*const LLVector4a scale(0.5f);

    for (U32 i = 0; i < face.mNumIndices; i+=3)
    {
        U16 idx1 = face.mIndices[i];
        U16 idx2 = face.mIndices[i+1];
        U16 idx3 = face.mIndices[i+2];

        LLVector4a v1; v1.setMul(face.mPositions[idx1], scale);
        LLVector4a v2; v2.setMul(face.mPositions[idx2], scale);
        LLVector4a v3; v3.setMul(face.mPositions[idx3], scale);

        if (ll_is_degenerate(v1,v2,v3))
        {
            llwarns << "Degenerate face found!" << LL_ENDL;
            return false;
        }
    }*/

    return true;
}

bool validate_model(const LLModel* mdl)
{
    if (mdl->getNumVolumeFaces() == 0)
    {
        LL_WARNS("MESHSKININFO") << "Model has no faces!" << LL_ENDL;
        return false;
    }

    for (S32 i = 0; i < mdl->getNumVolumeFaces(); ++i)
    {
        if (mdl->getVolumeFace(i).mNumVertices == 0)
        {
            LL_WARNS("MESHSKININFO") << "Face has no vertices." << LL_ENDL;
            return false;
        }

        if (mdl->getVolumeFace(i).mNumIndices == 0)
        {
            LL_WARNS("MESHSKININFO") << "Face has no indices." << LL_ENDL;
            return false;
        }

        if (!validate_face(mdl->getVolumeFace(i)))
        {
            return false;
        }
    }

    return true;
}

LLModelInstance::LLModelInstance(LLSD& data)
    : LLModelInstanceBase()
{
    mLocalMeshID = data["mesh_id"].asInteger();
    mLabel = data["label"].asString();
    mTransform.setValue(data["transform"]);

    for (U32 i = 0; i < data["material"].size(); ++i)
    {
        LLImportMaterial mat(data["material"][i]);
        mMaterial[mat.mBinding] = mat;
    }
}


LLSD LLModelInstance::asLLSD()
{
    LLSD ret;

    ret["mesh_id"] = mModel->mLocalID;
    ret["label"] = mLabel;
    ret["transform"] = mTransform.getValue();

    U32 i = 0;
    for (std::map<std::string, LLImportMaterial>::iterator iter = mMaterial.begin(); iter != mMaterial.end(); ++iter)
    {
        ret["material"][i++] = iter->second.asLLSD();
    }

    return ret;
}


LLImportMaterial::~LLImportMaterial()
{
}

LLImportMaterial::LLImportMaterial(LLSD& data)
{
    mDiffuseMapFilename = data["diffuse"]["filename"].asString();
    mDiffuseMapLabel = data["diffuse"]["label"].asString();
    mDiffuseColor.setValue(data["diffuse"]["color"]);
    mFullbright = data["fullbright"].asBoolean();
    mBinding = data["binding"].asString();
}


LLSD LLImportMaterial::asLLSD()
{
    LLSD ret;

    ret["diffuse"]["filename"] = mDiffuseMapFilename;
    ret["diffuse"]["label"] = mDiffuseMapLabel;
    ret["diffuse"]["color"] = mDiffuseColor.getValue();
    ret["fullbright"] = mFullbright;
    ret["binding"] = mBinding;

    return ret;
}

bool LLImportMaterial::operator<(const LLImportMaterial &rhs) const
{

    if (mDiffuseMapID != rhs.mDiffuseMapID)
    {
        return mDiffuseMapID < rhs.mDiffuseMapID;
    }

    if (mDiffuseMapFilename != rhs.mDiffuseMapFilename)
    {
        return mDiffuseMapFilename < rhs.mDiffuseMapFilename;
    }

    if (mDiffuseMapLabel != rhs.mDiffuseMapLabel)
    {
        return mDiffuseMapLabel < rhs.mDiffuseMapLabel;
    }

    if (mDiffuseColor != rhs.mDiffuseColor)
    {
        return mDiffuseColor < rhs.mDiffuseColor;
    }

    if (mBinding != rhs.mBinding)
    {
        return mBinding < rhs.mBinding;
    }

    return mFullbright < rhs.mFullbright;
}