The driver is very slow with Kicad

[Alex Deucher]

On Mon, Nov 28, 2011 at 4:23 PM, Element Green
<element at elementsofsound.org> wrote:
> On Mon, Nov 28, 2011 at 12:53 PM, Alex Deucher <alexdeucher at gmail.com>
> wrote:
>>
>> On Mon, Nov 28, 2011 at 3:33 PM, Element Green
>> <element at elementsofsound.org> wrote:
>> > Hello,
>> > I happened upon this thread when searching for information on slow
>> > performance with Kicad and the open source radeon driver.
>> > http://lists.x.org/archives/xorg-driver-ati/2011-August/021077.html
>> > I have a ATI/AMD HD 5800 series video card in my laptop.  I've also been
>> > rather impressed with the open source driver, even compared to Catalyst
>> > (which seems to suffer from many problems, especially with gnome 3).  So
>> > I'm
>> > very keen on getting Kicad to be usable with it.
>> > After doing a lot of profiling with sysprof, oprofile and gprof, I
>> > discovered that the open source radeon driver works acceptably when the
>> > radeon Kernel module is not loaded.  I discovered this when I was trying
>> > to
>> > switch back to the open source driver from Catalyst and there was an
>> > issue
>> > with the Kernel module loading.
>> > From what I can tell from /var/log/Xorg.0.log there is some rather
>> > significant initialization differences when the radeon Kernel module has
>> > not
>> > been loaded.  I tried disabling dri and dri2 to see if they were the
>> > culprit, but that makes Xorg non functional (weird square block for a
>> > mouse
>> > cursor over a black or corrupted display).  I also tried disabling many
>> > other extensions and modules (Render, glx, Composite), but none of them
>> > seemed to be causing the issue.
>> > Does this lend any information in regards to what could be causing the
>> > poor
>> > performance with Kicad?  Any ideas what different code paths could be
>>
>> When you don't load KMS, you disable acceleration so you are getting
>> software rendering.  Adding:
>> Option "NoAccel" "True"
>> to the device section of your xorg.conf should produce the same
>> effect.  IIRC, Kicad uses certain operations that are not accelerated
>> which causes a fallback to software rendering.  This ends up being
>> slower than pure sw rendering due to ping-ponging between GPU and CPU
>> memory and waiting for the gfx pipe to drain.
>>
>> Alex
>>
>
> Hello Alex,
> I just gave that a try just before getting your email (a diff between the 2
> Xorg logs made it obvious) and it did indeed speed things up.  Previously
> using xtrace I got a list of X protocol drawing primitives being used
> (listed below, with the number of calls in my test run).  What extension or
> module would acceleration of these operations fall under?  EXA, Render, or

EXA is not an extension itself.  It's a framework drivers can use to
accelerate core X rendering and the render protocol.

> ??  I'm a pretty competent programmer, but have never worked with Xorg and
> find the sheer number of different extensions, deprecated extensions, etc to
> be rather confusing.  I'd like to have a crack at adding acceleration for
> the missing operations.  Any idea why they are missing for this card?  Has
> AMD released enough info for this type of effort?

Yes.

> Also if you have any tips
> on how I can tell if a given operation is accelerated or not, that would be
> great.

Profiling the app and finding out where most of the time is being
spent should help you narrow down where the problem is.  I could be a
special case that the driver does not support, or more likely, a case
that EXA itself does not support.

Alex

> Thanks for the helpful info!
> Element Green
>
> Some of the protocol operations gleaned using xtrace for Kicad test run:
>    1035 CopyArea
>    1072 op=Over(0x03)
>    1168 GetInputFocus
>    1193 FreePicture
>    1231 CreatePicture
>    1439 FillRectangles
>    1616 FreePixmap
>    1655 CreatePixmap
>    2064 coordinate-mode=Origin(0x00)
>    3235 PolyFillArc
>    4467 SetClipRectangles
>    8443 FillPoly
>    9526 PolyRectangle
>   10374 PolyFillRectangle
>   13094 PolyArc
>   31302 PolySegment
>   39269 PolyPoint
>   54587 ChangeGC
>   59700 PolyLine
>