template<int bpp, int incrementPtr>
STRONG_INLINE void ColorPutter<bpp, incrementPtr>::PutColorAlphaSwitch(Uint8 *&ptr, const Uint8 & R, const Uint8 & G, const Uint8 & B, const Uint8 & A)
{
    Uint *ptr2 = ptr;
    if (bpp == 4)
        ptr2++;
    switch (A)
    {
    case 255:
        ptr += bpp * incrementPtr;
        return;
    case 0:
        PutColor(ptr, R, G, B);
        return;
    case 128:  // optimized
        PutColor(ptr,   ((Uint16)R + (Uint16)ptr2[1]) >> 1,
            ((Uint16)G + (Uint16)ptr2[2]) >> 1,
            ((Uint16)B + (Uint16)ptr2[3]) >> 1);
        return;
    default:
        PutColor(ptr, R, G, B, A);
        return;
    }
}

template<int bpp, int incrementPtr>
STRONG_INLINE void ColorPutter<bpp, incrementPtr>::PutColor(Uint8 *&ptr, const Uint8 & R, const Uint8 & G, const Uint8 & B, const Uint8 & A)
{
    Uint *ptr2 = ptr;
    if (bpp == 4)
        ptr2++;
    PutColor(ptr, ((((Uint32)ptr2[1]-(Uint32)R)*(Uint32)A) >> 8 ) + (Uint32)R,
                  ((((Uint32)ptr2[2]-(Uint32)G)*(Uint32)A) >> 8 ) + (Uint32)G,
                  ((((Uint32)ptr2[3]-(Uint32)B)*(Uint32)A) >> 8 ) + (Uint32)B);
}

template<int bpp, int incrementPtr>
STRONG_INLINE void ColorPutter<bpp, incrementPtr>::PutColor(Uint8 *&ptr, const Uint8 & R, const Uint8 & G, const Uint8 & B)
{
    if(incrementPtr == 0)
    {
        Uint8 *ptr2 = ptr;
        if(bpp == 4)
            *ptr2++ = 0;

        ptr2[0] = R;
        ptr2[1] = G;
        ptr2[2] = B;
    }
    else if(incrementPtr == 1)
    {
        if(bpp == 4)
            *ptr++ = 0;

        *ptr++ = R;
        *ptr++ = G;
        *ptr++ = B;
    }
    else if(incrementPtr == -1)
    {
        *(--ptr) = B;
        *(--ptr) = G;
        *(--ptr) = R;

        if(bpp == 4)
            *(--ptr) = 0;
    }
    else
    {
        assert(0);
    }
}

template<int bpp, int incrementPtr>
STRONG_INLINE void ColorPutter<bpp, incrementPtr>::PutColorRow(Uint8 *&ptr, const SDL_Color & Color, size_t count)
{
    Uint32 pixel = ((Uint32)Color.b << 16 ) + ((Uint32)Color.g << 8) + ((Uint32)Color.r << 0);

    for (size_t i=0; i<count; i++)
    {
        memcpy(ptr, &pixel, bpp);
        if(incrementPtr == -1)
            ptr -= bpp;
        if(incrementPtr == 1)
            ptr += bpp;
    }
}