Just some early comments. I’m sure I’ll live to regret them when CUDA eventually starts eating my 100000000 finite element grid between sips of coffee …
So, as a new arrival, I just want to see if I’m understanding installation properly, now that I have a 8600GT.
One thing to note, I have a intel quad 6600, it’s 64 bit, and I’m running a 64 bit slackware called slamd64, which is way decent.
I use the 64 bit linux NVIDIA driver despite the fact that this is an intel processor.
I run the vanilla linux CUDA .run file. I imagine it’s not 64 bit. Anyhow, I have no clue if my 6800 GT is 64 bit.
Anyhow, I go to make, and, well, I get a bunch of errors. No matter let’s try one of the projects … the simple CUBLAS. I’m OK at BLAS.
Hang on, no cublas.h … anywhere. So, … it’s not in the .run file I downloaded. SO I do a search and find it separately, in a zip.
Good post - it’s this stuff that you forget when you’ve dabbled for a while and the workarounds stick to your muscle memory. It’s also this stuff that’ll fsck you up when you deploy your CUDA application. So thanks.
I arrived at the Idea that the linux toolkit is wholly separate from the SDK. It should be obvious, 'cos one is a toolkit, and the other a dev kit. Also the toolkit is a binary while the dev kit isn’t. So, I’m coming around to this.
Is there any toolkit option for non-Fedora, non-SuSe and non-Ubuntu users? I gues not: it must be a real pain to generate binaries for so many distros. Those are the chief ones, I expect.
Probably the best option with the toolkit is just to install one of these on some free partition space. Just surmising.
in the CUBLAS zip file, there is no cublas.h, just a cublasP.h. gogling didn’t show up much. Maybe we’re meant to write it up ourselves?
You need to install the toolkit ( compiler, cublas, cufft).
The SDK is just a collection of examples.
Check with version of gcc is used in your distro and try one with similar characteristics.
It may work. If you cannot find one, the only option will be to change distro