Re: [hatari-devel] DSP performance

[ Thread Index | Date Index | More Archives ]


Is NOP=4 cycles measured 100% accurately (using external test hardware) or using some loop where it is too easy to get bad results, for example if there is multiple NOPs back to back and only every other NOP needs to do memory fetch.

When I gather timing results (in software) on a Falcon, I use several measurements.

- cached ops
- uncached ops
- cached ups where misses are deliberately incurred
- cached ops where the cache state is inhbited

For a nop (with the cache on), the timing behaviour must be as you suggest - occasional spikes to fetch a longword. In this case only an average cycle count is determined and it isn't a tidy integral.

For a nop with the cache off, the measurement is exactly 4 cycles (not 6 or 8). The 030 is fetching words from the bus in this mode, not longwords. If a word-fetch is really 4 cycles then it must be overlapping with the execute of 2 cycles.


Mail converted by MHonArc 2.6.19+