The old texture code was being bottle necked by the texture reads. We mirrored the real3d texture memory directly, including the mipmaps in a single large texture. I *think* most h/w has some sort of texture cache for a 2x2 or 4x4 block of pixels for a texture. What we were doing was reading the base texture, then reading the mipmap data from a totally separate part of the same texture which I can only assume flushed this cache. What I did was to create mipmap chains for the texture sheet, then copy the mipmap data there. Doing this basically doubles performance.
Some games update the tilegen after the ping_ping bit has flipped at 66% of the frame, so we need to split the tilegen drawing up into two stages to get some effects to work. So having the tilegen draw independantly of the 3d chip can make this happen.
