diff --git a/components/files/constrainedfilestream.cpp b/components/files/constrainedfilestream.cpp
index 4f76139d9..b239ec6a1 100644
--- a/components/files/constrainedfilestream.cpp
+++ b/components/files/constrainedfilestream.cpp
@@ -8,7 +8,7 @@
 namespace
 {
 // somewhat arbitrary though 64KB buffers didn't seem to improve performance any
-const size_t sBufferSize = 4096;
+const size_t sBufferSize = 8192;
 }
 
 namespace Files
diff --git a/components/nif/nifstream.cpp b/components/nif/nifstream.cpp
index d0fc9bab0..08a7466fa 100644
--- a/components/nif/nifstream.cpp
+++ b/components/nif/nifstream.cpp
@@ -6,138 +6,8 @@ namespace Nif
 {
 
 //Private functions
-uint8_t NIFStream::read_byte()
-{
-    uint8_t byte;
-    inp->read((char*)&byte, 1);
-    return byte;
-}
-uint16_t NIFStream::read_le16()
-{
-    uint8_t buffer[2];
-    inp->read((char*)buffer, 2);
-    return buffer[0] | (buffer[1]<<8);
-}
-uint32_t NIFStream::read_le32()
-{
-    uint8_t buffer[4];
-    inp->read((char*)buffer, 4);
-    return buffer[0] | (buffer[1]<<8) | (buffer[2]<<16) | (buffer[3]<<24);
-}
-float NIFStream::read_le32f()
-{
-    union {
-        uint32_t i;
-        float f;
-    } u = { read_le32() };
-    return u.f;
-}
 
-//Public functions
-osg::Vec2f NIFStream::getVector2()
-{
-    osg::Vec2f vec;
-    for(size_t i = 0;i < 2;i++)
-        vec._v[i] = getFloat();
-    return vec;
-}
-osg::Vec3f NIFStream::getVector3()
-{
-    osg::Vec3f vec;
-    for(size_t i = 0;i < 3;i++)
-        vec._v[i] = getFloat();
-    return vec;
-}
-osg::Vec4f NIFStream::getVector4()
-{
-    osg::Vec4f vec;
-    for(size_t i = 0;i < 4;i++)
-        vec._v[i] = getFloat();
-    return vec;
-}
-Matrix3 NIFStream::getMatrix3()
-{
-    Matrix3 mat;
-    for(size_t i = 0;i < 3;i++)
-    {
-        for(size_t j = 0;j < 3;j++)
-            mat.mValues[i][j] = getFloat();
-    }
-    return mat;
-}
-osg::Quat NIFStream::getQuaternion()
-{
-    osg::Quat quat;
-    quat.w() = getFloat();
-    quat.x() = getFloat();
-    quat.y() = getFloat();
-    quat.z() = getFloat();
-    return quat;
-}
-Transformation NIFStream::getTrafo()
-{
-    Transformation t;
-    t.pos = getVector3();
-    t.rotation = getMatrix3();
-    t.scale = getFloat();
-    return t;
-}
 
-std::string NIFStream::getString(size_t length)
-{
-    std::vector<char> str (length+1, 0);
-
-    inp->read(&str[0], length);
-
-    return &str[0];
-}
-std::string NIFStream::getString()
-{
-    size_t size = read_le32();
-    return getString(size);
-}
-std::string NIFStream::getVersionString()
-{
-    std::string result;
-    std::getline(*inp, result);
-    return result;
-}
-
-void NIFStream::getUShorts(std::vector<unsigned short> &vec, size_t size)
-{
-    vec.resize(size);
-    for(size_t i = 0;i < vec.size();i++)
-        vec[i] = getUShort();
-}
-void NIFStream::getFloats(std::vector<float> &vec, size_t size)
-{
-    vec.resize(size);
-    for(size_t i = 0;i < vec.size();i++)
-        vec[i] = getFloat();
-}
-void NIFStream::getVector2s(std::vector<osg::Vec2f> &vec, size_t size)
-{
-    vec.resize(size);
-    for(size_t i = 0;i < vec.size();i++)
-        vec[i] = getVector2();
-}
-void NIFStream::getVector3s(std::vector<osg::Vec3f> &vec, size_t size)
-{
-    vec.resize(size);
-    for(size_t i = 0;i < vec.size();i++)
-        vec[i] = getVector3();
-}
-void NIFStream::getVector4s(std::vector<osg::Vec4f> &vec, size_t size)
-{
-    vec.resize(size);
-    for(size_t i = 0;i < vec.size();i++)
-        vec[i] = getVector4();
-}
-void NIFStream::getQuaternions(std::vector<osg::Quat> &quat, size_t size)
-{
-    quat.resize(size);
-    for(size_t i = 0;i < quat.size();i++)
-        quat[i] = getQuaternion();
-}
+//Public functions
 
 }
diff --git a/components/nif/nifstream.hpp b/components/nif/nifstream.hpp
index 860c62e64..d00069be7 100644
--- a/components/nif/nifstream.hpp
+++ b/components/nif/nifstream.hpp
@@ -21,16 +21,70 @@ namespace Nif
 
 class NIFFile;
 
+/* 
+    readLittleEndianBufferOfType: This template should only be used with non POD data types
+*/
+template <uint32_t numInstances, typename T, typename IntegerT> inline void readLittleEndianBufferOfType(Files::IStreamPtr &pIStream, T* dest)
+{
+#if defined(__x86_64__) || defined(_M_X64) || defined(__i386) || defined(_M_IX86)
+    pIStream->read((char*)dest, numInstances * sizeof(T));
+#else
+    uint8_t* destByteBuffer = (uint8_t*)dest;
+    pIStream->read((char*)dest, numInstances * sizeof(T));
+    /*
+        Due to the loop iterations being known at compile time,
+        this nested loop will most likely be unrolled
+        For example, for 2 instances of a 4 byte data type, you should get the below result
+    */
+    union {
+        IntegerT i;
+        T t;
+    } u;
+    for (uint32_t i = 0; i < numInstances; i++)
+    {
+        u = { 0 };
+        for (uint32_t byte = 0; byte < sizeof(T); byte++)
+            u.i |= (((IntegerT)destByteBuffer[i * sizeof(T) + byte]) << (byte * 8));
+        dest[i] = u.t;
+    }
+#endif
+}
+
+/*
+    readLittleEndianDynamicBufferOfType: This template should only be used with non POD data types
+*/
+template <typename T, typename IntegerT> inline void readLittleEndianDynamicBufferOfType(Files::IStreamPtr &pIStream, T* dest, uint32_t numInstances)
+{
+#if defined(__x86_64__) || defined(_M_X64) || defined(__i386) || defined(_M_IX86)
+    pIStream->read((char*)dest, numInstances * sizeof(T));
+#else
+    uint8_t* destByteBuffer = (uint8_t*)dest;
+    pIStream->read((char*)dest, numInstances * sizeof(T));
+    union {
+        IntegerT i;
+        T t;
+    } u;
+    for (uint32_t i = 0; i < numInstances; i++)
+    {
+        u.i = 0;
+        for (uint32_t byte = 0; byte < sizeof(T); byte++)
+            u.i |= ((IntegerT)destByteBuffer[i * sizeof(T) + byte]) << (byte * 8);
+        dest[i] = u.t;
+    }
+#endif
+}
+template<typename type, typename IntegerT> type inline readLittleEndianType(Files::IStreamPtr &pIStream)
+{
+    type val;
+    readLittleEndianBufferOfType<1,type,IntegerT>(pIStream, (type*)&val);
+    return val;
+}
+
 class NIFStream {
 
     /// Input stream
     Files::IStreamPtr inp;
 
-    uint8_t read_byte();
-    uint16_t read_le16();
-    uint32_t read_le32();
-    float read_le32f();
-
 public:
 
     NIFFile * const file;
@@ -39,33 +93,117 @@ public:
 
     void skip(size_t size) { inp->ignore(size); }
 
-    char getChar() { return read_byte(); }
-    short getShort() { return read_le16(); }
-    unsigned short getUShort() { return read_le16(); }
-    int getInt() { return read_le32(); }
-    unsigned int getUInt() { return read_le32(); }
-    float getFloat() { return read_le32f(); }
-
-    osg::Vec2f getVector2();
-    osg::Vec3f getVector3();
-    osg::Vec4f getVector4();
-    Matrix3 getMatrix3();
-    osg::Quat getQuaternion();
-    Transformation getTrafo();
+    char getChar() 
+    {
+        return readLittleEndianType<char,char>(inp); 
+    }
+    short getShort() 
+    { 
+        return readLittleEndianType<short,short>(inp);
+    }
+    unsigned short getUShort() 
+    { 
+        return readLittleEndianType<unsigned short,unsigned short>(inp);
+    }
+    int getInt() 
+    {
+        return readLittleEndianType<int,int>(inp);
+    }
+    unsigned int getUInt() 
+    { 
+        return readLittleEndianType<unsigned int,unsigned int>(inp);
+    }
+    float getFloat() 
+    { 
+        return readLittleEndianType<float,uint32_t>(inp);
+    }
+
+    osg::Vec2f getVector2() {
+        osg::Vec2f vec;
+        readLittleEndianBufferOfType<2,float,uint32_t>(inp, (float*)&vec._v[0]);
+        return vec;
+    }
+    osg::Vec3f getVector3() {
+        osg::Vec3f vec;
+        readLittleEndianBufferOfType<3, float,uint32_t>(inp, (float*)&vec._v[0]);
+        return vec;
+    }
+    osg::Vec4f getVector4() {
+        osg::Vec4f vec;
+        readLittleEndianBufferOfType<4, float,uint32_t>(inp, (float*)&vec._v[0]);
+        return vec;
+    }
+    Matrix3 getMatrix3() {
+        Matrix3 mat;
+        readLittleEndianBufferOfType<9, float,uint32_t>(inp, (float*)&mat.mValues);
+        return mat;
+    }
+    osg::Quat getQuaternion() {
+        float f[4];
+        readLittleEndianBufferOfType<4, float,uint32_t>(inp, (float*)&f);
+        osg::Quat quat;
+        quat.w() = f[0];
+        quat.x() = f[1];
+        quat.y() = f[2];
+        quat.z() = f[3];
+        return quat;
+    }
+    Transformation getTrafo() {
+        Transformation t;
+        t.pos = getVector3();
+        t.rotation = getMatrix3();
+        t.scale = getFloat();
+        return t;
+    }
 
     ///Read in a string of the given length
-    std::string getString(size_t length);
+    std::string getString(size_t length) {
+        std::vector<char> str(length + 1, 0);
+
+        inp->read(&str[0], length);
+
+        return &str[0];
+    }
     ///Read in a string of the length specified in the file
-    std::string getString();
+    std::string getString() {
+        size_t size = readLittleEndianType<uint32_t,uint32_t>(inp);
+        return getString(size);
+    }
     ///This is special since the version string doesn't start with a number, and ends with "\n"
-    std::string getVersionString();
-
-    void getUShorts(std::vector<unsigned short> &vec, size_t size);
-    void getFloats(std::vector<float> &vec, size_t size);
-    void getVector2s(std::vector<osg::Vec2f> &vec, size_t size);
-    void getVector3s(std::vector<osg::Vec3f> &vec, size_t size);
-    void getVector4s(std::vector<osg::Vec4f> &vec, size_t size);
-    void getQuaternions(std::vector<osg::Quat> &quat, size_t size);
+    std::string getVersionString() {
+        std::string result;
+        std::getline(*inp, result);
+        return result;
+    }
+
+    void getUShorts(std::vector<unsigned short> &vec, size_t size) {
+        vec.resize(size);
+        readLittleEndianDynamicBufferOfType<unsigned short,unsigned short>(inp, &vec.front(), size);
+    }
+    void getFloats(std::vector<float> &vec, size_t size) {
+        vec.resize(size);
+        readLittleEndianDynamicBufferOfType<float,uint32_t>(inp, &vec.front(), size);
+    }
+    void getVector2s(std::vector<osg::Vec2f> &vec, size_t size) {
+        vec.resize(size);
+        /* The packed storage of each Vec2f is 2 floats exactly */
+        readLittleEndianDynamicBufferOfType<float,uint32_t>(inp,(float*) &vec.front(), size*2);
+    }
+    void getVector3s(std::vector<osg::Vec3f> &vec, size_t size) {
+        vec.resize(size);
+        /* The packed storage of each Vec3f is 3 floats exactly */
+        readLittleEndianDynamicBufferOfType<float,uint32_t>(inp, (float*) &vec.front(), size*3);
+    }
+    void getVector4s(std::vector<osg::Vec4f> &vec, size_t size) {
+        vec.resize(size);
+        /* The packed storage of each Vec4f is 4 floats exactly */
+        readLittleEndianDynamicBufferOfType<float,uint32_t>(inp, (float*) &vec.front(), size*4);
+    }
+    void getQuaternions(std::vector<osg::Quat> &quat, size_t size) {
+        quat.resize(size);
+        for (size_t i = 0;i < quat.size();i++)
+            quat[i] = getQuaternion();
+    }
 };
 
 }