Subject: Re: More on rcons performance woes: pmax profiling
To: None <firstname.lastname@example.org>
From: Jonathan Stone <jonathan@DSG.Stanford.EDU>
Date: 03/05/1999 17:53:19
I reran the `head -1000 /usr/share/misc/termacp > /dev/rcons' tests,
with the stupid, stupid bcopy() speedups I posted earlier-- this time
using the tuned mips kernel bcopy.
Elapsed time, on a non-profiled kernel, went from 6:21 to 5:04, or
roughly a 20% speedup. This, again, is doing single-char-line
jumpscrolling, with no atempt to do the Sun-console-style, multi-line
Subjectively, scrolling is now fast enough that the time difference in
painting `short' lines versus `long' lines is, visibly, very
noticeable. That strongly suggests using similar tricks to speed up
painting individual char-cells would be a win.
I think the issue of how best to handle framebuffers with no direct
text mode -- both accelerated and non-accelerated -- is still open.
that said -- Andy Doran has changes to the rcons code, to add in hooks
to use an outboard blit engine on a DEC accelerated fb. Is anyone
willing to review them?