OSError: libcula_lapack.so, libcula_lapack_basic.so, libcula.so not found about scikit-cuda

[jantogni]

Hello again :),

I added /usr/cula/lib64 to my LD_LIBRARY_PATH and now I get a scikits.cuda.cula.culaRuntimeError: 3.

Also I tried to define my LD_LIBRARY_PATH with /usr/cula/lib and I get the first error: OSError: libcula_lapack.so, libcula_lapack_basic.so, libcula.so not found.
But there is a libcula_lapack_basic in the /usr/cula/lib.

In advance thank you,

[lebedov]

Hi, Jonathan.

What version of CULA and CUDA do you have installed and what OS platform are you using (i.e., Linux version and architecture)? If you are running 32-bit Linux, your LD_LIBRARY_PATH should contain /usr/local/cula/lib rather than /usr/local/cula/lib64.

[jantogni]

Hi,

CUDA (nvcc): Cuda compilation tools, release 5.5, V5.5.0
And CULA I have downloaded cula_dense_free_R18-linux.run from the CULA webpage.
I'm using Ubuntu 14.04 x86_64, but both directories (lib and lib64) has the libcula_lapack_basic.so

Should I install CULA from different source?

In advance thank you,

[lebedov]

Try R17 - I believe R18 requires CUDA 6.0.

[jantogni]

Hi :), I have installed R17, and I included /usr/local/cula/lib64 to my LD_LIBRARY_PATH and now is working (I tested mdot_demo.py and svd_demo.py).

Thank you very much, best regards.

[lebedov]

Great.

[ShanLuo]

Hi @jantogni , where did you download CULA? I checked the webpage of CULA but got this " As of June 14, 2013, CULA will no longer be offered as an individually licensed product. Cluster and distribution licenses will be available. Please contact us for additional details." Is there any other way to obtain CULA?

@lebedov Any advice?

Thanks in advance!

[lebedov]

@ShanLuo, I'm not aware of any legal way to download CULA at the present time. Some of the functionality of CULA is present in NVIDIA's CUSOLVER library, while other comparable functions are in MAGMA. What function(s) in CULA do you need?

[ShanLuo]

@lebedov. Thanks for your reply. For now, I want to use CULA to do the SVD computation, as shown in one of the demo examples of scikit-cuda. Is there any way to do it using CUSOLVER or MAGMA instead? Thanks.

[lebedov]

@ShanLuo: yes, skcuda.linalg.svd() in the latest revision of scikit-cuda on GitHub supports the use of CUSOLVER instead of CULA - just set lib='cusolver' when you call the function.

[ShanLuo]

@lebedov Great! Thanks for your advice! But I don't how to I set lib='cusolver'.. Can you give me more details? Thanks.

[ShanLuo]

@lebedov I think I find the solution: skcuda.linalg.svd(a_gpu, jobu=’A’, jobvt=’A’, lib=’cusolver’). It works, thanks!