Rice - actually, with the Xenon's XGPU, you can't directly access the framebuffer. You can get a window into the XGPU memory space, but you can only use it for reading / writing texture data and other data that doesn't directly touch any of the four main FBs or final FB. You could have the XGPU render the FB contents into a texture within an area of memory that you've marked as write-combined (both the XGPU and CPU can't blindly access the full memory space, you have to mark shared memory as such), or you can use the CPU to write a texture to the aforementioned area of memory, and then render that texture in screenspace to the FB, but you cannot explicitly read bytes out of the FB. So in any case, there's no free lunch, and you wind up with about the same performance hit as you would with DMA setup time on the PS3. Furthermore, you can command the RSX on the PS3 to pull texture data from main RAM, but you have a slight performance hit when doing so.
What I'd be more curious about is, as Doom / squall mentioned, would be emulating the RDP completely on a consumer video card's GPU. You'd need a bleeding-edge card, but it might be possible, and in fact it might be marginally more realistic in terms of the way the RDP works. You pass the RDP packets themselves to the GPU, which in turn processes the packet and renders the triangle (or performs the necessary function). There would be an additional command set implemented to write or read back only the emulated FB space (rather than the full framebuffer), and you could simply read back the FB at the start of the frame, run the emulation until the end of the frame (with any FB reads and writes happening with the cached FB contents), then at the end of the frame you write the emulated FB back to your program running on the GPU. Slow, but not quite as slow.
wierd.
