So I wondered about time related bogons and I found very odd looking code
in i810_accel.c - do a grep for GetTimeInMillis

Now consider what happens close to wrap time

Firstly:
On a reload now could reload with 0. Tripping the start = 0 case next loop.
We then fall through into the last_head = ring->head load. The same occurs 
during a wrap event where start=0x7FFFFFFFF now=0x80000000 (minus loads).
We mistakenly see it as progress. Fortunately it appears to fail safe as we
merely reset the timer.

Secondly GetTimeInMillis appears to be unsigned, its stuffed into
variables. That doesn't itself seem fatal as the (now-start) maths seems to
wrap safely - but wants checking

Thirdly there is a small delay loop. Or rather in egcs-1.1.2 there is. gcc
2.96 will optimise that to "i=2000" - i must be volatile if that loop is
needed.

Can you confirm what that loop is for and if its required ?

Alan