Bug#648222: Significant 2D performance regression with ColorTiling

[Alex Deucher]

On Sun, Nov 13, 2011 at 1:34 PM, Iustin Pop <iustin at debian.org> wrote:
> On Thu, Nov 10, 2011 at 06:15:43PM +0100, Michel Dänzer wrote:
>> On Don, 2011-11-10 at 18:01 +0100, Michal Suchanek wrote:
>> > On 10 November 2011 17:46, Alex Deucher <alexdeucher at gmail.com> wrote:
>> > > On Wed, Nov 9, 2011 at 12:36 PM, Iustin Pop <iustin at debian.org> wrote:
>> > >>
>> > >> The recent upgrade of xserver-xorg-video-radeon from 1:6.14.2-2 to
>> > >> 1:6.14.3-1 enabled ColorTiling for my card, which in turn caused a
>> > >> significant performance degradation in 2D (yes, I understand it should
>> > >> make 3D faster, but I didn't know it should slow down 2D applications).
>> > >>
>> > >> I'm using plain 2D environment (openbox, no compositing, anything) and
>> > >> plain xterm (bitmap fonts, no AA, etc.). The speed of display text has
>> > >> changed significantly enough that I can "see" my mutt refreshing the
>> > >> inbox and drawing the lines.
>> > >
>> > > Tiling will speed up all rendering (2D and 3D).  However, it sounds
>> > > like you are using an environment that is mostly software rendering.
>> > > As such in order for the CPU to access tiled buffers, the GPU has to
>> > > copy them to a linear buffer before CPU can access it properly.
>> >
>> > FWIW I have color tiling enabled and have no speed issues in urxvt -
>> > TrueType fonts, AA enabled, etc.
>> >
>> > Unlike xterm urxvt (rxvt-unicode) uses some special font-rendering
>> > libraries, however.
>> >
>> > If I understand it correctly xterm would use the in-server bitmap font
>> > rendering which the X server can accelerate as much as it wants.
>>
>> Core bitmap fonts are completely unaccelerated so far with EXA. xterm
>> can also use Xft for font rendering via the -fa option though, which is
>> well accelerated.
>
> Hi all, thanks for the replies.
>
> I've tested further and this seems to be a problem specific to xterm
> (and possibly other software? not sure how to test e.g. firefox's UI
> speed):
>
> Test file: ~20K lines. I've chosen the font sizes so as to have the same
> number of lines in full-screen. The timing is simply "time cat file".
>
> ColorTiling disabled:
>
> xterm -fn fixed:      ~3.3s
> xterm -fa Mono -fs 8: ~8.6s
> rxvt-unicode -fn fixed: ~0.07s
> rxvt-unicode -fn xft:Mono:pixelsize=11: ~0.05s
>
> ColorTiling enabled:
>
> xterm -fn fixed:      ~21s
> xterm -fa Mono -fs 8: ~7.8s
> rxvt-unicode -fn fixed: ~0.6s usually, sometimes ~0.1s??
> rxvt-unicode -fn xft:Mono:pixelsize=11: ~0.05s
>
> So it seems to me that:
>
> a) whatever xterm does, it is very sub-optimal
> b) while xterm's -fa mode is indeed sped up by ColorTiling, it's still
>   many times slower than with ColorTiling disabled and using core
>   fonts
> c) also rxvt-unicode with core fonts is slowed down by color tiling,
>   sometimes very much so (10x)
>
> Based on this I think this is mostly an xterm bug (-fa should/could be
> much faster), but I wonder if it also applies to other purely software
> 2D rendering.

It likely does.  We optimize the accelerated paths rather than the slow paths.

Alex