bloodline wrote:
Hell yeah... I imagine with the PPC, Copying the display updates to the Chipram would probably be 95% of the total overhead of the whole process :-D
Actually, one interesting thing I have noticed in my various bus battering experiments is that the PPC actually has considerably slower write access to Chip RAM than the 680x0 does.
For me, the ideal scenario would be:
1) Chunky RGB buffer in fastram, all rendering calls work on this. Possibly use MMU to determine the damage area. Very small changes could be "accumulated" until either a worthwile amount of required conversion is reached or a time threshold has elapsed.
2) Conversion of RGB -> HAM8 planar data in fast ram. This could be PPC accelerated where available - but only on areas that are above a certian size (as determined by the context switch time) - for small areas the context switch time could dominate, but above a certian area the PPC would rapidly overtake the 680x0 in performance terms. This should be a user preference tweak.
3) Copy converted data to CHIP ram, again if MMU was used originally, copy only the changed area. This could be done on 680x0 or PPC or even directly as part of step (2). However, it seems writing to chip ram is considerably slower on all the PPC's than for 68040/68060, so the 680x0 might make more sense.
