Open ATI Specifications For R100-200

[glisse]

Not wanting to go Too offtopic, but might it not be more then justified to rewrite this application so that at the very least it coul drun in maybe dos-box under linux if anything? If there is no support or anything for this application, I hate to say this, but you'll get screwed over in the long run anyway. Costs of finding such old and exclusive hardware will keep going up and finding it will be even harder.

As for r100-r200 based hardware, I have an old IBM x226 series server that has an onboard 7000 based radeon. Having the r100 supported wouldn't be interesting in terms of 3D games or even 3D desktop, it could be interesting for media boxes I guess but maybe somewhat more for opencl. You could offload, once supported, raid calculations, ssl calculations etc to the otherwise unused GPU.

Having said that, I'm in the process of migrating the old x226 out to be replaced by an HP microserver (n36l). Not because it's faster, the dual CPU 3.0ghz xeon is quite a bit faster, but because it's much quieter and uses a LOT less electricity. The idea still stands however, it uses an M880G (radeon hd 4200) which woul dbe very interesting to use to offload certain stuff too. Once software (and driver) support is in place

r100 will never support OpenCL it just doesn't have any of the capacity. HD4200 isn't either very compute friendly. You can't ask hw which predate a specification to suddenly support it.

[oliver]

long post, all gone :S

Summarized, I always thought opencl just used the shaders to offload stuff to. Would be kinda awesome to offload raid5/6 calculations, SSL calculations to the GPU, even a little help can mean a lot on low powerd CPU's?

[DaemonFC]

r100 will never support OpenCL it just doesn't have any of the capacity. HD4200 isn't either very compute friendly. You can't ask hw which predate a specification to suddenly support it.

It seems totally unreasonable to even assume that hardware from 2000-2003 will ever support OpenCL. Most of those cards are limited to OpenGL 1.x from my understanding. USers of those cards should be glad there's open source support at all, AMD hasn't supported them in a long long time

[DeepDayze]

I always seem to be stumbling across issues with the R100 and R200 drivers where I am later told "You need to disable HyperZ" before it will work properly. Which is apparently the reason why HyperZ isn't enabled by default in the first place. However, HyperZ gives a noticeable increase in FPS for (say) glxgears, so I would really like to be able to use it properly.

HyperZ doesn't quite work well when KMS is enabled but it works alright in userland mode. Maybe with more detailed specs of these older chips then I can hang on to my old Thinkpad T42 a while longer. Also these chips are too old to be supportable by the latest fglrx, so the OSS driver is the only way to go with such chips

[BlackStar]

long post, all gone :S

Summarized, I always thought opencl just used the shaders to offload stuff to. Would be kinda awesome to offload raid5/6 calculations, SSL calculations to the GPU, even a little help can mean a lot on low powerd CPU's?

OpenCL requires R600-level shaders to work (i.e. DX10/GL3 hardware).

R100 is a DX7 GPU - it doesn't have "shaders" like those found in modern GPUs. It has "register combiners" which are programmable in a very very limited fashion (no raid or ssl calculations!)

The 4200 mentioned above is a low- (bottom-) end DX10 GPU that could, in theory, support OpenCL. Even if it does, there is a good chance it will be *slower* than the CPU.

Finally, raid and ssl calculations are not good candidates for OpenCL acceleration. The reason is that sending the calculations to the GPU and waiting for the results takes more time than running the calculations directly on the CPU - even if the GPU can perform these calculations faster. In other words, GPU communication adds latency, even if its throughput is way higher. The performance difference must be significantly higher than that latency for OpenCL to make sense.

(Note that you can always run OpenCL on the CPU).

[b15hop]

Rare? That's true, to a point...
However, there are SOME of us who are forced by circumstances beyond our control to cast our minds back to such prehistoric times (and MORE).

Here's a specific example of a (self imposed) sideline project of mine.
The systems in question are some of the earliest ever PCs that had a PCI slot (although the currently installed 'custom' cards are all ISA there's two available PCI slots!)
They are running a modified version of Concurrent DOS and the application in question is a PURE textual application (NO graphics involved at all!).
The _normal_ setup of these systems has two attached monitors - One being a hercules type MDA and the other being a VGA. (I repeat that BOTH of them are in 'text' mode).
The VGA is an IGD on the motherboard and the MDA is a separate ISA card.
The problem I have is that functional Hercules MDA monitors have become harder to find here than the Higgs Boson at LHC / CERN!!!
(And I am getting rather tired of replacing LOPTs on them!)
You can't _easily_ hook up a current LCD monitor to an MDA card (18kHz VSYNC is nowhere NEAR fast enough on most current LCDs although I guess it COULD be possible to 'butcher' an LCD TV as the sync frequencies are ALMOST in the ballpark)

After some serious 'head scratching', I remembered that VGA cards could run in 'monochrome emulation mode'.
The problem though is that co-existence of effectively TWO VGA cards, one running in MDA-text mode (BIOS mode 0x07) and the other running in VGA-text mode (BIOS mode 0x03) doesn't work right either!
The PCI setup in the system BIOS tends to only allow ONE of the two PCI cards to initialise and 'claim' the necessary video RAM (A000:0000 thru BFFF:FFFF), VGA BIOS (C000:0000 thru C000:FFFF) and the corresponding pertinent I/O space.
Interestingly enough, I can make it work perfectly using an ISA VGA card (with the BIOS ROM removed so that the system BIOS would not 'see' it was there), but I think I found the LAST ISA VGA card in existence!!!

Soooo, I am still kind of stuck for all the other machines of this group.
I then remembered the good old Radeon 7000 series. Some of them came out with TWO VGA outputs (others had a DVI + VGA + Composite).
I already know that the R100 includes TWO CRTCs along with two of most everything else...
What I am wondering is whether it will be feasible to drive BOTH monitors from a single (dual head) card.
The 'primary' head would be running in MDA emulation (BIOS mode 0x07 - Video RAM at B000:0000) and would have the control of the BIOS Data segment at 0040:0000
The 'secondary' head would be running as VGA (BIOS mode 0x03 - Video RAM at B800:0000).
The application suite performs direct writes to the VGA mode video RAM @ B800:0000 and it uses NO cursor, NO scrolling... (In the process, it violates almost every possible 'standard'... LOL)

If I had a nice clear manual on the R100 specifics (including ALL registers at the bit level), I could confirm or deny if the concept will work with a fairly good level of confidence.
Trying to 'guess' based upon the R100 register list in the sources is far less certain.
I've grabbed some new/old-stock R100 based dual head cards (just a few to start with) and I'm going to 'play' for a while.
Luckily, it's just a 'fun project' for me. I hope to learn a bit (a lot?) from it (along with refreshing my few remaining aging braincells with some OLD tech hardware).

Wish me luck!

Wow. Whatever you're working on, either must be important or you're really keen. (Even both)

I agree that OpenCL might have latancy issues on older hardware. Especially since the bus on older devices is much slower than newer architecture. The biggest issue would be that the performance gains would be negligible, even if the hardware supported some extra acceleration. I like the idea though. There is a saying though. Leave horses to courses. Or let bygones be bygones.

Once before, I tried upgrading a 1.7ghz gpu to faster and faster gpu hardware... It got to a point where the GPU was heavily limited by the cpu. 7800 GS was limited by the cpu and bus speed of AGP. Once I upgraded the machine to even just a AMD x2 3600+ (1.9ghz duel core), and 9800 GTX+ , there were huge gains. The newer 9800 wasn't much faster, only 2x possibly... The CPU / Faster pci express bus speed, along with faster sata drives etc... made the machine feel like a "real" upgrade. I've since learned that sometimes mixing new hardware that is too far apart in age with old hardware, is not always the best thing. It's still good to know it's all compatible though. You're correct though that in some rare cases, there can be benefits. It's just knowing what and where. Just make sure there is a good reason as to why you would want it in the first place. xD

[oliver]

HyperZ doesn't quite work well when KMS is enabled but it works alright in userland mode. Maybe with more detailed specs of these older chips then I can hang on to my old Thinkpad T42 a while longer. Also these chips are too old to be supportable by the latest fglrx, so the OSS driver is the only way to go with such chips

My T42 boost a Radeon 9600 mobile with 64mb ram e.g. r300 based chip

[DeepDayze]

My T42 boost a Radeon 9600 mobile with 64mb ram e.g. r300 based chip

My own T42 has the Radeon Mobility 7500 chip with just 32MB RAM :/