Hello everyone, A little while back there was a discussion about another patch and how much coding efficiency improved speed. With a little slack time at work I decided to hot rod radeonhd's r6xx/7xx exa acceleration code and do some tests. I don't expect this patch to be accepted because its sloppy and it makes the code unmaintainable. The reason for doing it was to see what some of the potential for speed improvement actually was. Testing was done on both the unpatched driver and the patched driver at two different clock speeds. Each test was done by issuing the command "x11perf -v1.3 -all " in xorgs default window manager as root. The clock speeds for the tests were fixed at 2GHz and 500MHz and the results are listed in results_at_2GHz_clock.txt and results_at_500MHz_clock.txt respectively. There were some notable gains in the move window, put image, and circle drawing tests. There were also some slow downs for which I am at a loss to explain as no code path was slowed down by my changes. Most notable are the create unmapped window tests. As expected the gains were much more noticeable at the slower clock speed. The conclusions I have come to are the following. Gains of 5% or more are possible in some areas. Laptop users who have their clocks throttled back to save on power will benefit the most from coding efficiency improvements. There is no one magic spot to make things faster. The patch should apply cleanly to radeonhd-1.2.5 Raw data furnished upon request. -- Conn Conn O. Clark Observation: In formal computer science advances are made by standing on the shoulders of giants. Linux has proved that if there are enough of you, you can advance just as far by stepping on each others toes.
On Thu, Apr 23, 2009 at 11:24 AM, Conn Clark
Hello everyone,
<snip>
The patch should apply cleanly to radeonhd-1.2.5
Raw data furnished upon request.
-- Conn
Conn O. Clark
Observation: In formal computer science advances are made by standing on the shoulders of giants. Linux has proved that if there are enough of you, you can advance just as far by stepping on each others toes.
And of course here is the patch. <blush> -- Conn O. Clark Observation: In formal computer science advances are made by standing on the shoulders of giants. Linux has proved that if there are enough of you, you can advance just as far by stepping on each others toes. -- To unsubscribe, e-mail: radeonhd+unsubscribe@opensuse.org For additional commands, e-mail: radeonhd+help@opensuse.org
On Thu, Apr 23, 2009 at 11:33 AM, Conn Clark
On Thu, Apr 23, 2009 at 11:24 AM, Conn Clark
wrote: Hello everyone,
<snip>
The patch should apply cleanly to radeonhd-1.2.5
Raw data furnished upon request.
-- Conn
Conn O. Clark
Observation: In formal computer science advances are made by standing on the shoulders of giants. Linux has proved that if there are enough of you, you can advance just as far by stepping on each others toes.
And of course here is the patch. <blush>
--
Conn O. Clark
Observation: In formal computer science advances are made by standing on the shoulders of giants. Linux has proved that if there are enough of you, you can advance just as far by stepping on each others toes.
Gmail doesn't want to let me email the patch !?!?! here it is again under a different name -- Conn O. Clark Observation: In formal computer science advances are made by standing on the shoulders of giants. Linux has proved that if there are enough of you, you can advance just as far by stepping on each others toes. -- To unsubscribe, e-mail: radeonhd+unsubscribe@opensuse.org For additional commands, e-mail: radeonhd+help@opensuse.org
On Thu, Apr 23, 2009 at 11:36 AM, Conn Clark
On Thu, Apr 23, 2009 at 11:33 AM, Conn Clark
wrote: On Thu, Apr 23, 2009 at 11:24 AM, Conn Clark
wrote: Hello everyone,
<snip>
The patch should apply cleanly to radeonhd-1.2.5
Raw data furnished upon request.
-- Conn
Conn O. Clark
Observation: In formal computer science advances are made by standing on the shoulders of giants. Linux has proved that if there are enough of you, you can advance just as far by stepping on each others toes.
And of course here is the patch. <blush>
--
Conn O. Clark
Observation: In formal computer science advances are made by standing on the shoulders of giants. Linux has proved that if there are enough of you, you can advance just as far by stepping on each others toes.
Gmail doesn't want to let me email the patch !?!?!
here it is again under a different name
--
Conn O. Clark
Observation: In formal computer science advances are made by standing on the shoulders of giants. Linux has proved that if there are enough of you, you can advance just as far by stepping on each others toes.
permission problem :( <Blush> -- Conn O. Clark Observation: In formal computer science advances are made by standing on the shoulders of giants. Linux has proved that if there are enough of you, you can advance just as far by stepping on each others toes.
On Apr 23, 09 11:24:37 -0700, Conn Clark wrote:
A little while back there was a discussion about another patch and how much coding efficiency improved speed. With a little slack time at work I decided to hot rod radeonhd's r6xx/7xx exa acceleration code and do some tests. I don't expect this patch to be accepted because its sloppy and it makes the code unmaintainable. The reason for doing it was to see what some of the potential for speed improvement actually was.
That's a very nice experiment.
It's interesting for me to see that there is a potential of approx. 5%
speed increase. It's not exactly much, and I doubt we will go for
problematic code (as you indicated) for this speed gain - at least now.
There are much bigger potential speedup possibilities - the most
important being tiled frame buffers, then additional exa hooks. And
getting those !@#$% overlapping copy operations to actually work as
intended.
CU
Matthias
--
Matthias Hopf
On Fri, Apr 24, 2009 at 3:32 AM, Matthias Hopf
On Apr 23, 09 11:24:37 -0700, Conn Clark wrote:
A little while back there was a discussion about another patch and how much coding efficiency improved speed. With a little slack time at work I decided to hot rod radeonhd's r6xx/7xx exa acceleration code and do some tests. I don't expect this patch to be accepted because its sloppy and it makes the code unmaintainable. The reason for doing it was to see what some of the potential for speed improvement actually was.
That's a very nice experiment.
It's interesting for me to see that there is a potential of approx. 5% speed increase. It's not exactly much, and I doubt we will go for problematic code (as you indicated) for this speed gain - at least now.
There are much bigger potential speedup possibilities - the most important being tiled frame buffers, then additional exa hooks. And getting those !@#$% overlapping copy operations to actually work as intended.
CU
Matthias
-- Matthias Hopf
__ __ __ Maxfeldstr. 5 / 90409 Nuernberg (_ | | (_ |__ mat@mshopf.de Phone +49-911-74053-715 __) |_| __) |__ R & D www.mshopf.de
Thanks for the compliment. I put a lot of work into this testing. Something else occurred to me too. My benchmarks are being carried out on a radeon HD 3100 which I think runs at 300MHz. Faster GPUs should benefit more from this patch than my computer did. I added a couple of percentage points to some of the improvements last night. One quick test indicated an improvement of 16%. I have included the new patch. I should also mention that the patch should also apply cleanly to the current git as well as the 1.2.5 release. I would be interested in seeing some benchmarks before and after from people with better GPUs if anybody has some time to kill. To dangle a carrot to encourage people to try this patch, the text in gnome terminal scrolls by faster during compiles which is partially due to more cpu time being available to compile with :D -- Conn Conn O. Clark Observation: In formal computer science advances are made by standing on the shoulders of giants. Linux has proved that if there are enough of you, you can advance just as far by stepping on each others toes.
Great work, Conn! Its always good to see people working and
experimenting with performance work.
Seems to get way too little attention in the Linux development
universe, whereas MS employs thousands of profile-monkeys doing
nothing else the whole day ;)
I know this is a bit off topic, but I wonder how much speedup
compiling the driver (and maybe the whole Xorg stack) with profile
driven optimizations would yield.
I've experimented with that lately, and beside the fact that it makes
-O3 finally useable, it yields impressive results.
Sad that distributors still mostly ignore that powerful option, and
instead decide to throw i586 support away for 1% speed gain (like
fedora did for F12).
- Clemens
2009/4/24 Conn Clark
On Fri, Apr 24, 2009 at 3:32 AM, Matthias Hopf
wrote: On Apr 23, 09 11:24:37 -0700, Conn Clark wrote:
A little while back there was a discussion about another patch and how much coding efficiency improved speed. With a little slack time at work I decided to hot rod radeonhd's r6xx/7xx exa acceleration code and do some tests. I don't expect this patch to be accepted because its sloppy and it makes the code unmaintainable. The reason for doing it was to see what some of the potential for speed improvement actually was.
That's a very nice experiment.
It's interesting for me to see that there is a potential of approx. 5% speed increase. It's not exactly much, and I doubt we will go for problematic code (as you indicated) for this speed gain - at least now.
There are much bigger potential speedup possibilities - the most important being tiled frame buffers, then additional exa hooks. And getting those !@#$% overlapping copy operations to actually work as intended.
CU
Matthias
-- Matthias Hopf
__ __ __ Maxfeldstr. 5 / 90409 Nuernberg (_ | | (_ |__ mat@mshopf.de Phone +49-911-74053-715 __) |_| __) |__ R & D www.mshopf.de Thanks for the compliment. I put a lot of work into this testing.
Something else occurred to me too. My benchmarks are being carried out on a radeon HD 3100 which I think runs at 300MHz. Faster GPUs should benefit more from this patch than my computer did.
I added a couple of percentage points to some of the improvements last night. One quick test indicated an improvement of 16%. I have included the new patch. I should also mention that the patch should also apply cleanly to the current git as well as the 1.2.5 release.
I would be interested in seeing some benchmarks before and after from people with better GPUs if anybody has some time to kill.
To dangle a carrot to encourage people to try this patch, the text in gnome terminal scrolls by faster during compiles which is partially due to more cpu time being available to compile with :D
-- Conn
Conn O. Clark
Observation: In formal computer science advances are made by standing on the shoulders of giants. Linux has proved that if there are enough of you, you can advance just as far by stepping on each others toes.
-- To unsubscribe, e-mail: radeonhd+unsubscribe@opensuse.org For additional commands, e-mail: radeonhd+help@opensuse.org
participants (3)
-
Clemens Eisserer
-
Conn Clark
-
Matthias Hopf