Let's call it 2.3 then.

EDIT: No one has worked on cgminer's kernel in a while. For some reason everyone wants to work on something else. Ironically I'm not a cl kernel expert so have made no progress in the kernel area since the last time someone did something for cgminer.

You need sdk 2.4+ for full hardware monitoring support. It's called phatk for simplicity but really it's phatk2.1.

In the git tree, I added it to verbose mode, and also to --ndevs which will allow us to ask people "what does --devs show?"

Something like this:
[2012-01-29 00:37:48] CL Platform vendor: Advanced Micro Devices, Inc.
[2012-01-29 00:37:48] CL Platform name: ATI Stream
[2012-01-29 00:37:48] CL Platform version: OpenCL 1.0 ATI-Stream-v2.1 (145)
[2012-01-29 00:37:48] 6 GPU devices detected

On the other hand, I may have to eat my words...

A little debugging found this:
[2012-01-28 22:43:25] PLATFORM OpenCL 1.0 ATI-Stream-v2.1 (145)                   

I wish it were only that bad 

AMD:

I'll check but not as far as I'm aware it doesn't.

There is a reason cgminer reports back the CURRENT settings after you change it: It's because the driver will often accept a value you give it, but the hardware will NOT. So you may set it to 300 and it will merrily ignore you and set it back to the default. Try setting it to only 125 lower than the gpu engine speed and if it suddenly starts running cooler than when you "set" it to 300, that is what's happened.

Some people say the 5xxx cards were faster with SDK 2.1, however this changed with the phatk2+ kernel in cgminer, where 2.1, 2.4 and 2.5 are pretty much equal. 2.2 and 2.3 have various problems with various kernels and various hardware, the details of which I don't know. 2.6 is basically crap for everything but 7970, where it's mandatory.

New release: Version 2.2.0

NOTE: GPU Device order may change with this release with ATI GPUs as cgminer
now can enumerate them according to their Bus ID which means the values should
now correlate with their physical position on the motherboard.


Links in top post as always.

Given the number of changes in this version, I have a human readable summary of user-visible changes first. CPU mining binaries will no longer be distributed by myself:

New features:
Bitforce FPGA support
Net Delay option to not overload slow routers
Socks proxy support.
More RPC API interfaces
Preliminary GCN support
Linking of fan speed and control on dual GPU cards.

Changes:
Reordering of devices where possible to match PCI order.
Intensity support up to 14
Better damping of fanspeed with auto-fan
GPU temperature of 2nd GPU in dual GPU  cards affects fanspeed
CPU mining being deprecated and off by default
Parsing of configuration files copes much better with problems
Less false positives for network delays.

Bugfixes:
Much less likely to fail to build binaries on first startup. (completely fixed?)
Fixed potential for false positive of a dead pool.

New bugs:
Unknown  

Full changelog:
- Default to poclbm kernel on Tahiti (7970) since phatk does not work, even
though performance is sub-standard so that at least it will mine successfully by
defau
- Retain cl program after every possible place we might build the program.
- Update ADL SDK URL.
- Fix potential overflow.
- Map GPU devices to virtual devices in their true physical order based on
BusNumber.
- Change the warning that comes with failure to init cl on a device to be more
generic and accurate.
- Advertise longpoll support in X-Mining-Extensions
- Detect dual GPU cards by iterating through all GPUs, finding ones without
fanspeed and matching twins with fanspeed one bus ID apart.
- Do not attempt to build the program that becomes the kernel twice. This could
have been leading to failures on initialising cl.
- Some opencl compilers have issues with no spaces after -D in the compiler
options.
- Allow intensity up to 14.
- Use calloced stack memory for CompilerOptions to ensure sprintf writes to the
beginning of the char.
- Whitelist 79x0 cards to prefer no vectors as they perform better without.
- Adjust fan speed gently while in the optimal range when temperature is
drifting to minimise overshoot in either direction.
- Detect dual GPU cards via the indirect information of - 1st card has a fan
controller. 2nd card does not have a fan controller, cards share the same device
name
- Instead of using the BFI_INT patching hack on any device reporting
cl_amd_media_ops, create a whitelist of devices that need it. This should enable
GCN architec
- Fixed API compiling issue on OS X
- Add more explanation of JSON format and the 'save' command
- Return an error if using ADL API commands when it's not available
- Read off lpThermalControllerInfo from each ADL device.
- Add ADL_Overdrive5_ThermalDevices_Enum interface.
- Add API commands: config, switchpool, gpu settings, save
- Implement socks4 proxy support.
- Fix send() for JSON strings
- Introduce a --net-delay option which guarantees at least 250ms between any
networking requests to not overload slow routers.
- Generalise locking init code.
- Allow invalid values to be in the configuration file, just skipping over them
provided the rest of the file is valid JSON. This will allow older configurat
- Allow CPU mining explicitly enable only if other mining support is built in.
- BitForce FPGA support
- Configure out building and support of all CPU mining code unless
--enable-cpumining is enabled.
- Allow parsed values to be zero which will allow 0 values in the config file to
work.
- Advertise that we can make our own midstate, so the pool can skip generating
it for us
- Refactor the CPU scanhash_* functions to use a common API. Fixes bugs.
- Don't consider a pool lagging if a request has only just been filed. This
should decrease the false positives for "pool not providing work fast enough".
- Invalidating work after longpoll made hash_pop return no work giving a false
positive for dead pool. Rework hash_pop to retry while finds no staged work u
- Remove TCP_NODELAY from curl options as many small packets may be contributing
to network overload, when --net-delay is enabled.
- Refactor miner_thread to be common code for any kind of device
- Simplify submit_nonce loop and avoid potentially missing FOUND - 1 entry.
Reported by Luke-Jr.
- Micro-optimisation in sha256_sse2 code courtesy of Guido Ascioti
- Refactor to abstract device-specific code

Anyone who's ever coded crossplatform will have to laugh at this, so I just had to share it.

Apologies, this order change did not benefit linux which usually showed the correct order, it only swapped the order around, so I'll be disabling it from linux and only using it on windows.

I assume you mean 2.6 is garbage. 2.5 worked identically to 2.4 in my experience.

The reorder branch has now been merged into the master git branch. Putting the finishing touches on the code, readme etc. to release the new version today.

WARNING
This means that the ordering of your cards will almost certainly change from here on with cgminer into the "correct" order based on physical bus ID location so if you have different settings for different GPUs, you will need to update all your configuration files accordingly.

Hmm thanks for the link, but that doesn't even appear in my code anywhere. Presumably that's ufa they're talking about. Meanwhile cgminer mines fine on my nvidia laptop so 

Thanks very much. Yes the order will often be reversed due to mobo+os+drivr discrepancies.. The order should now most accurately represent what is happening physically with the cards.

Okay, here is the first windows executable with PCI bus based reordering of GPU devices. Put it into an existing 2.1.2 installation.

Please test it out to see if it works and what it does to your advertised card order!

http://ck.kolivas.org/apps/cgminer/temp/cgminer.exe

Linux build works from "reorder" branch on git.

On further inspection this code won't work...

Back to the drawing board. 

Ok I've added a git branch called "reorder". If you git pull and then "git checkout reorder" you can try it.

Theoretically it should even fix card ordering of any sort, not just dual GPUs that are split up:


Here it is completely untested... so please test it  
http://ck.kolivas.org/apps/cgminer/temp/cgminer.exe

EDIT: tested, doesn't work. Hold on...

Just use as per normal.

Yes I understood that. I just didn't expect it was possible, but this extra debugging has shown me it is. I just need to actually code it up now.