Performance optimization: Find low-hanging fruit

[neobrain]

All major components required for high-performance are in place (if not yet in the developer edition), but the emulator isn't quite reaching 60 FPS on anything but high-end devices yet. There must be low-hanging fruit somewhere that I've missed.

Some good candidates to look into are:

• Graphics:
  • Streaming of vertex/texture/uniform data has poor efficiency. Persistent mapping of buffers and explicit flushing should perform better.
  • framebuffer management (implement more fast paths for blits; avoid unnecessary flushing to/reloading from memory)
    • MemoryFills currently invalidate textures/render targets (causing them to be reloaded from memory on next use). Instead, they should be translated to clears of the covered resources.
• Kernel:
  • kernel scheduler (the task rotation is highly inefficient, which may add up to considerable overhead)
  • The thread timeout checks at the bottom of OS::ElapseTime is a major bottleneck that can be fixed fairly easily by storing all threads in WaitingForTimeout/Sleeping states in a list instead of traversing all threads.
    Fixed in 7d2c663.
• Audio:
  • Implement null-audio (implemented in b094815)

[neobrain]

The menu screen in Mario Kart 7 is abnormally slow, spending 30ms of host time each frame just in CPU emulation.