home *** CD-ROM | disk | FTP | other *** search
-
- > Too much and too spread out. There isn't much repeated accessing in time for
- > the cache to still have the data around.
- > Almost everything but the texure data is kept in registers all the time.
-
- On this subject, I've been experimenting with the 68040 PMMU and it's cache
- mode flags (for individual pages).
-
- Since 16-bit writes tend to invalidate 'good' cache lines and therefore reduce
- cacheing effectiveness, I decided to mark the screen as non-cacheable. The
- result was a significant improvement in drawing times. Using this on the desktop
- in bitplane modes (which should actually BENEFIT from destination-cacheing) GemBench
- indicated a rise from 900% to 950% in mono, and similar improvements in all the
- other colour modes. Truecolour modes produced the most significant increases.
-
- I haven't looked at the 68030 PMMU flags for a while, but it may also have this
- paged cache mode capability.
-
- It is a bit low-level for something like Bad Mood, but an auto-folder based PMMU
- driver (a bit like my 68040 driver) could possibly speed up standard Falcons under
- various conditions of use. Particularly on the Desktop, and especially for Bad Mood.
- This does of course depend on the capabilities of the 68030 PMMU and the relatively
- small 68030 data cache. Reducing write-invalidations might not produce a notable
- change with such an easily trashed cache...
-
- Doug.
-
-