Direct link to the kernel in the git tree:
https://github.com/ckolivas/cgminer/blob/master/poclbm120214.cl

It's using a worksize of 256, 1 vector (i.e. no vectors) and intensity 11.

Yes I remember this old graph. It would appear to me that cgminer's defaults of v2 w128 are the most robust for all tunings unless the user goes out of his way to tweak things manually, in which case he'll be choosing his own settings, thanks.

Hmm what do the 58xx report their name as? Cypress? It sounds like cgminer should default them to worksize 256 instead of 128. Would that be a fair assessment?

With Diapolo's help we finally got the diakgcn kernel working on cgminer. I've just committed code to the git tree which makes it work. Alas at the same engine and clock speeds on the 7970, diakgcn gives me 699 MHash while my customised kernel gives me 717 MHash. But now that it's working, he may be able to tweak it further...

cgminer sets clocks all back to default on exit... if it exits cleanly, and of course on windows it's a miracle when it does.

Did you check what it says when you run "cgminer -n" or start it with debugging enabled as it says in the readme? --verbose -D -T as well as your regular commands. Might give a clue. Also perhaps the hardware stuff got lost for 5xxx in later ATI drivers and worked on the ones somewhere in between. Knowing AMD that's entirely possible  Like it might work in 11.6 and then got broken in 11.11 or something.

Never blame the new version of cgminer. ALWAYS blame something else first. 

Update on 7970 performance.

I did a little calculation to see how my 7970 is performing with the latest cgminer.

Clock per clock it is precisely 1/3 faster than the 6970 running at the equivalent clock speed.
What's hilarious about this is that the number of stream processors is precisely 1/3 more than the 6970 (2048 vs 1536). This is despite the 6970 running on phatk on sdk 2.5 and 7970 running on my customised poclbm on sdk 2.6. As I said before, at least on the 6970 which I don't grossly underclock the memory (only -125) I'm getting slightly better performance on poclbm with 2.6 now than phatk with 2.5.

So in terms of raw performance, I think the current cgminer release and kernel is doing pretty good.  

Alas none of the magic GCN changes that aren't just more processors appear to have been exploited, and every thread I've read on suggestions to exploit them on these forums has only led to worse performance, often disastrously worse.

Right, that's why cgminer reports back the *actual* values because the driver can happily accept some changes and the card can happily ignore them.

A shortcut won't work unless you're "in" the directory where the kernel files are kept. If you're writing a .bat file make sure to cd into the cgminer directory before starting and make SURE all the files it came with are there.  Also make sure to use the latest version (currently 2.2.6).

export DISPLAY=:0

Also you will get more overclocking ability with 11.6 instead of 11.4.

The PCIE spec dictates up to 1.5cm freeplay or something like that. Use spacers:

http://ck.kolivas.org/pictures/Mining/IMG_1240.JPG

cgminer 2.2.6 out. 2+ more mhash on 7970. It's getting harder and harder to extract much more

1200/1050+5% clocks, intensity 11 - 717 Mhash.

11.6 needs dummy plugs on windows. You need a later driver but beware of SDK fuckage.

New release: Version 2.2.6 - February 16, 2012

Links in top post as always.

Trying hard to make the most current release as stable as possible at the moment.

Updated customised poclbm kernel for 7970 worth 2 more Mhash, and same on any other GPUs when run on SDK2.6.
BIG FAT WARNINGS if SDK 2.6 is detected.
Ability to tune dynamic mode with a new --gpu-dyninterval parameter. This tells cgminer how long maximum it can keep the GPU busy for. Default was 7ms but can now be set to any value. Setting it higher will increase your hashrate and make your desktop less responsive. This is in response to requests to improve hashrate back to what it was when 2 threads were in use.
The display should keep all logs now from startup when the curses formatted display begins.
The bug where dynamic mode increases stale rate by submitting duplicate work should be fixed.

Full changelog:

 Provide warning on each startup about sdk 2.6
- Fix unused warnings on win32.
- bitforce: Simplify BFopen WIN32 ifdef/else
- Fix initialization warning with jansson 1.3
- bitforce: Cleanup extraneous TODO that isn't needed
- Move tcsetattr (and new tcflush) into *nix BFopen to simplify things a bit
- Add message explaining 2nd thread disabling for dynamic mode and how to tune
it.
- Move logwindow down once number of devices is known.
- Automatically choose phatk kernel for bitalign non-gcn ATI cards, and then
only select poclbm if SDK2.6 is detected.
- Allow the refresh interval to be adjusted in dynamic intensity with a
--gpu-dyninterval parameter.
- Make curses display visible right from the beginning and fix the window sizes
so the initial messages don't get lost once the status window is drawn.
- The amount of work scanned can fluctuate when intensity changes and since we
do this one cycle behind, we increment the work more than enough to prevent
repeati
- bitforce: Set a 30 second timeout for serial port on Windows, since the
default is undefined
- Use PreVal4addT1 instead of PreVal4 in poclbm kernel.
- Import PreVal4 and PreVal0 into poclbm kernel.
- Import more prepared constants into poclbm kernel.
- Keep variables in one array but use Vals[] name for consistency with other
kernel designs.
- Replace constants that are mandatorily added in poclbm kernel with one value.
- Remove addition of final constant before testing for result in poclbm kernel.
- Hand optimise variable addition order.
- Hand optimise first variable declaration order in poclbm kernel.
- Radical reordering machine based first pass to change variables as late as
possible, bringing their usage close together.
- fix strcpy NULL pointer if env HOME unset.
- bitforce: Disable automatic scanning when at least one device is specified
manually
- Unroll all poclbm additions to enable further optimisations.

This is the first time you've reported it?

Suggestions:
Perhaps your pool is actually asking you to submit the old shares even if they're stale. Newer cgminer supports the "submitold" feature now.

I get .3-.4% rejects with static intensities so I can't reproduce your problem here.

p2pool also has a significantly higher reject rate if that's what you're on/have moved to.

Thanks. In the meantime, I've audited the code and can't find where this might happen, so a fix for this is not coming in the next version I'm afraid. I'll keep looking.

Hmm if I didn't know any better, it looks like the auto-fan is not being enabled on the other devices.
Can you click on GPU setting, Change setting and select each device in order and see if they report:
Thanks

While it's running, can you go into settings and enable debug mode and see what it's saying about temperatures and fan speeds? It will spew all sorts of info so you'll have to watch for a while to find it.

You have hardware errors with cgminer which means something is horribly wrong with your install and this has nothing to do with p2pool. Start with a fresh install of cgminer (2.2.5 is latest).