Archives Discussions

[Raistmer]

SDK installation requirement unappropriate for app distribution!

When could we expect OpenCL support to be included into driver itself and the need to install SDK will be removed?
My application requires unattended installation by another process (it's BOINC-related application) so any additional requirements that need user's intervention (and BOINC can't afford to download and install SDK by itself) are absolutely unacceptable.

[kbrafford]

The only options I've seen for that are either to move to Windows 7 and use DirectCompute (since DirectCompute is nicely built into DirectX 11), or move to Mac OS X Snow Leopard (since Snow Leopard comes with OpenCL functionality built-in).

[Raistmer]

Well, both not look like acceptable options for already written application that needs to be distributed under windows [all windows (and Linux too) versions, not just Win7. Perhaps you don't know what is BOINC ]
If NV can embed OpenCL support into driver I see no reasons why ATI can't. Surely they can, main question - when ?

[omkaranathan]

If NV can embed OpenCL support into driver I see no reasons why ATI can't.

Nvidia only have GPU implementation. 

Surely they can, main question - when ?

We do have plans to include OpenCL runtime into driver, but I'm afraid I don't have an ETA for you now. 

[nou]

on linux you can try my package. it should work even for Debian.

[Raistmer]

Originally posted by: nou

on linux you can try my package. it should work even for Debian.

Thanks, when Linux-related bug will be fixed...

[Lev]

I am sorry, is using CAL option for your? Looks like CAL run time is included in the driver.

[Raistmer]

No, it should work on too many hardware platforms to be bound to proprietary API. CAL is not an option. Also, do you know working FFT implementation for ATI written on CAL ? hardwired 1024 FFT sample ATI putted in last SDKs no an option too.

[Lev]

Why do not use cuda for nvidia and CAL for AMD, anyway algortihm is the same, only interface is different.

Sorry, I do not know any FFT. AFAIK there is no such for opencl still, but porting from cuda should not be a problem.

[Raistmer]

1) what do you mean "CAL for AMD" then? CUDA is pretty high-level C-like language. What you propose to use for kernels on AMD, Brook? I tried - very limited abilities. Or IL ? Completely different language level, more like assembler - I have no resources to learn such language, especially for doing unpaid job! It's not practical.
2) There is OpenCL FFT sample from Apple I adopted it to my app. And it works. Cuda port - what CUDA-based sources to port? Volkov's samples? They too limited in offered FFT size, sometimes I need 32k FFT sometimes even more. And CUFFT supplied by NV w/o sources AFAIK.

[Lev]

Sorry, I thought CAL was more high level, a bit more like CUDA driver API.

[MicahVillmow]

Lev,
The CAL API is similiar to the CUDA driver API, but the IL language itself is closer to PTX and not CUDA.

[Raistmer]

Yes, that's why "Use CAL for AMD" sounds ambiguous. Yes, I could manage buffer allocations and kernel calls in proprietary API like CAL, but what to do with kernels themselves in this case...
So, hope ATI could manage OpenCL requirements better and includes runtime support into driver in next release