unable to execute cubin file

[Jianbin Fang]

Hi,

I am trying to use asfermi now. However, I encountered a problem
listed below.

I download the test example in svn/test/ and then:

(1) assemble the 1 (or 2, 3, 4) to cubin file, namely, 1.k (or 2.k,
3.k, 4.k) using asfermi;
(2) compiler the file main.cu using nvcc (v4.1, and my local os is
ubuntu 10.04 (64bit) and GPUs is Quadro5000);
(3) run the app: ./main 1.k k

Then it reports:
----------------------------------------
Success at 100
Success at 95
./main: symbol lookup error: ./main: undefined symbol: cuCtxCreate_v2
----------------------------------------

When I run the other cubin files (2.k, 3.k, and 4.k), the results
reported are the same.

Could anyone tell me the reason(s)? Is it because of my systems or
configurations?

Regards,

Jianbin

[Hou Yunqing]

Can you successfully execute ordinary cuda executables? Maybe you need to set LD_LIBRARY_PATH properly (usually to /use/local/cuda/lib or something similar)

Sent from my iPod

[Dmitry N. Mikushin]

Hi,

Could you be using mismatching versions of CUDA driver and Toolkit? In cuda/include/cuda.h you can see how _v2-s are selected.

- D.

2012/4/26 Jianbin Fang <j.fa...@gmail.com>

[Jianbin Fang]

Yes, it is because of the mismatch (versions of Driver was lower than
that of SDK, and it should be the other way around). Thanks Yunqing
and Dmitry for the help. Actually, it didn't work for the normal CUDA
programs (Usually, I use OpenCL support, and it is working all the
time. It is still weird, right? That means OCL performs no
restrictions on the versions matching?).

Anyway, problem solved, and thanks very much.

Jianbin

On Apr 27, 8:46 am, "Dmitry N. Mikushin" <maemar...@gmail.com> wrote:
> Hi,
>
> Could you be using mismatching versions of CUDA driver and Toolkit? In
> cuda/include/cuda.h you can see how _v2-s are selected.
>
> - D.
>

> 2012/4/26 Jianbin Fang <j.fang...@gmail.com>