Slow 2D performance with Fx6600GT AGP Returns + culprit found

[_john_i_]

The slow 2D performance with fx6600gt agp has reared it's ugly head again, and I have found the culprit!

This was previously discussed in this thread:
http://www.nvnews.net/vbulletin/showthread.php?t=44317

I had found the upgrading to xorg-6.9.0-rc2 fixed the problem. Well, the final xorg-6.9.0 was just released and I was shocked to find that the problem was back!

So to find the offending code, I generated a patch set between the 6.9.0-rc2 source and the current 6.9.0 source. I then looked for anything I thought might be the cause.

What I found is a patch submitted for bug #5258 was the culprit. The bug info can be found here:
https://bugs.freedesktop.org/show_bug.cgi?id=5258

Reversing this patch and recompiling xorg fixes the problem.

What is even more interesting is that this patch was submitted by Nvidia! The purpose was restore binary compatibility with 6.8.2.

Furthermore I don't have to fully reverse the patch to fix it. I can simply insert a spare bit after the first in the bitfield. Simply shifting over the remaining bits gives me a speedup. I think one of the bits in this field must be manipulated frequently, and shifting it to an even position makes a profound difference in performance.

The attached patch will add the extra unused bit, speeding things back up.


Lonnie: You worked with me on this problem when I entered a bug report. Should I open another bug and attach this info? I can also submit this to x.org, but it will I think undo what nvidia was trying to accomplish with it's patch. The offending patch was submitted by Aaron Plattner at nvidia. in the xorg bug #5258 referenced above.


John

[PrakashP]

Well, if you use subpixel hinting, try setting to grey hinting. This helps for me regarding speed, but fonts don't look that nice on LCDs anymore...

[_john_i_]

It's fast again with my patch, so I'll leave my hinting like it is

I suppose if this isn't addressed in xorg or the Nvidia drivers, I'll just keep patching it. I compile most of my own stuff anyway.

[PrakashP]

Well, reverting the patch is *wrong*, as it leaves some things in undefined state which could lead to all kinds of problems. Furthermore your subpixel hinting will be broken anyway using that patch, if you look closely. So this would be the wrong way to go...

[_john_i_]

No, my subpixel render *IS* working. Just to prove this to myself I did a screen shot, and magnified several fonts. It is clearly doing subpixel rendering with various color pixels, not just grey.

If you read the bug report from which the original patch came from, some of the xorg developers were questioning the need for the patch in the first place.

[PrakashP]

Well, interesting, because for me subpixel hinting wasn't working without above patch in kde apps. Firefox etc seems mostly fine, but konqueror for example not.

[_john_i_]

ahhhhhhhhhhh, you are correct. I magnified a konqueror window, and all of the hinting is grey. When I reported it as working was looking at a firefox window, where the hinting properly has all of the colors.

I wonder why non-kde apps are not affected. This is really odd.

The speedup makes konqueror and other KDE apps run fast again, so I can live with the grey rendering for now in kde apps.

For anyone who wants to see what we are seeing, I'm attaching a zoomed screenshot. On the left is firefox with proper subpixel rendering. On the right is the same page in konqueror, where the subpixel rendering is all grey.