Skip to contentSkip to conversation

memcpy to videoram eats too much CPU on ATI cards (cache trashing?)

DISCUSSION Peter Surda 4 participants 7 messages Aug 27, 2001 – Sep 6, 2001

lore
Share & tools
lore
#0 lore raw

Dear kernel gurus,

First of all I want to apologise for writing here, but I think this is the place with the greatest chance of getting some help. I have done extensive research on the problem and talked to many people including driver developers and was unable to find any solutions yet.

So, fist a little intro: when watching videos with Xv under XFree86, a certain function is called to transfer the data from system RAM to video RAM. This function is driver specific, but for all the drivers I checked (mga, tdfx, mach64, r128) it has the same contents, looks like pasted. It basically does a for (h--) memcpy (blah, blah, blah).

The point is that with mga, tdfx and what I heard nvidia too, this doesn't cause any CPU load (or more precisely, non-measurable load). However, with mach64 and r128, it DOES. I did some more research.

memcpy-ing 380kB at 25fps takes about 5ms per frame and causes X to eat 1% cpu time (time measurements were done by tsc)
memcpy-ing 760kB at 25fps takes about 11ms per frame, but instead of eating 2% CPU time, it eats 35% (yes, that's 35 times more)

The speed isn't the real problem (when you multiply it you get about 70MB/s and that's definitely enough). The problem is that this eats CPU time, and that shouldn't (or at least not so much).

This happens on both of my systems, one with PIIMobile/366 and mach64, and one with Duron 650 with r128. I had a voodoo before for tests, and CPU load wasn't measurable, from what I heard mga and nvidia as well, so it is something ATI-specific. Some other people having ATI cards have the same problem (from what I read on gatos-devel list), but I have never heard someone explicitely say "the problem doesn't exist on my ATI".

MTRR is enabled correctly, disabling it only worsens the problem.

I have been in close contact with the driver developer and XFree86 maintainer, but none of them seem to know exactly how to solve it. Current theory is that this is caused by some cache being trashed, but I have no idea how to fix it.

Oh yes, I already tried using a memcpy written in assembly utilizing MMX, but it didn't show any change.

I would be very grateful for ideas.

Please CC me, I'm not on the list.

Bye,

Peter Surda (Shurdeek) <[email protected]>, ICQ 10236103, +436505122023

--
Failure is not an option. It comes bundled with your Microsoft product.

#1 lore raw
Quoted 2 lines

cause any CPU load (or more precisely, non-measurable load). However, with mach64 and r128, it DOES. I did some more research.

Makes sense

Quoted 4 lines

memcpy-ing 380kB at 25fps takes about 5ms per frame and causes X to eat 1% cpu time (time measurements were done by tsc)
memcpy-ing 760kB at 25fps takes about 11ms per frame, but instead of eating 2% CPU time, it eats 35% (yes, that's 35 times more)

So presumably at this bandwidth you are beginning collide with the graphics controllers and it will stall your cycles, at which point its going to stall the CPU so effectively eat CPU time

Sounds like a reasonable explanation to me

#2 lore raw
Quoted 7 lines · Alan Cox
Quoted 2 lines

cause any CPU load (or more precisely, non-measurable load). However, with mach64 and r128, it DOES. I did some more research.

Makes sense

Quoted 2 lines

memcpy-ing 380kB at 25fps takes about 5ms per frame and causes X to eat 1% cpu time (time measurements were done by tsc)

1%? at 5ms/f * 25 f/s = 125 ms/ second. = 12% of your CPU.

However, as the X server manages to finish doing what it has to do before the next timer tick, it will almost never get a timer tick accounted to it.

Quoted 2 lines
Quoted 2 lines

memcpy-ing 760kB at 25fps takes about 11ms per frame, but instead of eating 2% CPU time, it eats 35% (yes, that's 35 times more)

So at 2.2 times as much CPU time, I'd expect around 27% real usage. But instead of managing to miss the timer tick almost every time, it is now managing to hit slightly more than average on the timer ticks. Or it's using a little bit more than you measured.

By doing stuff that takes on the order of a timer tick, and to trigger them off a timer event, means that CPU time measurements can become highly inaccurate.

If you buy CPUtime on a large unix computer, and have a tight mainloop, you can measure the number of iterations you can do inside your mainloop in 9ms. Then schedule an "usleep (1)" every time you hit that many iterations. Cheap computing... ;-)

Roger.

Signature · 4 lines
** [email protected] ** http://www.BitWizard.nl/ ** +31-15-2137555 **
*-- BitWizard writes Linux device drivers for any device you may have! --*
* There are old pilots, and there are bold pilots. 
* There are also old, bald pilots. 
#3 lore raw
In reply to Alan Cox
Quoted 2 lines · Rogier Wolff

However, as the X server manages to finish doing what it has to do before the next timer tick, it will almost never get a timer tick accounted to it.

Yes, I also realised that, and other people also seem to think it is this way.

So the conclusion is basically that the card can't chew data that fast and I should use busmastering instead of memcpy (and other drivers should do that too because "hidden load" occurs anyway). I'm working on it.

Thanks to all who replied, I am as always pleased by the cooperation in open-source world and wish everyone good luck. Yeah, and linux rocks of course :-)

Quoted 1 line

Roger.

Bye,

Peter Surda (Shurdeek) <[email protected]>, ICQ 10236103, +436505122023

--
I believe the technical term is "Oops!"

#4 lore raw

(solved) memcpy to videoram eats too much CPU on ATI cards

In reply to Alan Cox
Quoted 3 lines · Peter Surda

So the conclusion is basically that the card can't chew data that fast and I should use busmastering instead of memcpy (and other drivers should do that too because "hidden load" occurs anyway). I'm working on it.

Just to end this thread in a victorous manner ;-), thanks to Michel Dänzer <[email protected]> and me, there is now a working implementation of busmastered video transfers for the r128 driver, and it has been submitted to all relevant lists and maintainers. It indeed solved the problem, CPU time eaten by video transfers is negligible and DVD and "DivX ;-)" playback was never so smooth. With software-only DVD decoder, watching fullscreen DVD leaves 50-60% CPU time idle on a Duron 650, even on action-packed scenes. If I catch someone claiming again that Linux isn't suitable for multimedia, I can just laugh now :-).

Bye,

Peter Surda (Shurdeek) <[email protected]>, ICQ 10236103, +436505122023

--
Dudes! May the Open Source be with you.

#5 lore raw

Re: (solved) memcpy to videoram eats too much CPU on ATI cards

Am 06 Sep 2001 08:18:43 +0200 schrieb Peter Surda:

Quoted 4 lines

Just to end this thread in a victorous manner ;-), thanks to Michel Dänzer <[email protected]> and me, there is now a working implementation of busmastered video transfers for the r128 driver, and it has been submitted to. all relevant lists and maintainers.

I kept on checking the relevant mailinglist and project CVSes but failed to find the described implementation, would you please provide additional hints where to get it?

Servus,
Daniel

#6 lore raw

Re: (solved) memcpy to videoram eats too much CPU on ATI cards

In reply to Alan Cox
Quoted 7 lines · Daniel Egger
Quoted 4 lines

Just to end this thread in a victorous manner ;-), thanks to Michel Dänzer <[email protected]> and me, there is now a working implementation of busmastered video transfers for the r128 driver, and it has been submitted to. all relevant lists and maintainers.

I kept on checking the relevant mailinglist and project CVSes but failed to find the described implementation, would you please provide additional hints where to get it?

dri-devel should have it, as well as livid-gatos (which unfortunately seems to have full disks and is offline at the moment).

For very lazy ones I found a link to the annoucement in dri-devel:
http://www.geocrawler.com/lists/3/SourceForge/680/0/6536416/

Quoted 2 lines

Servus,
Daniel

Mit freundlichen Grüßen

Peter Surda (Shurdeek) <[email protected]>, ICQ 10236103, +436505122023

--
"One world, one web, one program" -- Microsoft promotional ad
"Ein Volk, ein Reich, ein Fuehrer" -- Adolf Hitler