CULA and CUBLAS

[tang]

Dear Administrator,

I am trying to learn the new powerful numerical calculation libraries accelerated by GPU. However, there are two different libraries: CULA and CUBLAS, the former is the collection of LAPACK functions and the latter is the collection of BLAS level functions. However, it seems that I can't find some key functions used by some sparse solvers, i.e., the following functions:
DZNRM2
DNRM2
ZGEMV
ZDOTC
ZSWAP
ZTRSM
ZGEMM
ZCOPY
ZTRSV
ZGERU
ZAXPY
ZSCAL

From the CUBLAS, I can only find two functions:
DNRM2
ZGEMM

Do you have any plan to develop all BLAS level functions, include double complex data type?

Thanks,
Zhanghong Tang

[dan]

Hi Zhanghong,

The functions you're looking for are BLAS level functions. The GPU-accelerated BLAS library, CUBLAS, is an NVIDIA library and is currently actively developed. In the next version of CUDA, 3.0 (currently in beta), I believe that all of these functions will be available.

Dan

[tang]

Dear Dr. Dan,

Thank you very much for your kindly reply. Have you released the beta version of CUDA 3.0? From your website I only find the latest version 1.1. How can we get the beta version of CUDA 3.0?

Thanks,
Zhanghong Tang

[kyle]

The current beta version of NVIDIA's CUDA 3.0 is currently only available to NVIDIA Registered Developers. You can register for that program here:

http://developer.nvidia.com/page/regist ... ogram.html

[tang]

Dear Dr. Dan,

Thank you very much for your so quick reply.
However, I am still confused. So now there are 3 different libraries accelerated by GPU, i.e., CUDA, CULA and CUBLAS, could you please tell me the difference between these libraries?

Now it seems that the CUDA 3.0 is what I looked for. I have the G220 1GB Nvidia card, but from the download page
http://forums.nvidia.com/index.php?showtopic=149959
I found so many versions, which one should I download?

Thanks,
Zhanghong Tang

[john]

Hi - to summarize the differences for you:

CUDA is Nvidia's language for developing Nvidia-based GPU applications.
CUBLAS is Nvidia's CUDA-based implementation of BLAS, which are simple linear algebra routines (dot product, matrix-vector product, matrix-matrix product, etc)
CULA is our product, and is also a CUDA-based library similar to CUBLAS. CUDA accelerates high-level linear algebra concepts (system solve, QR, SVD, eigenvalues, inverse, least squares, etc.)

CUDA is where you will want to begin for developing your own GPU programs unless you need any of the concepts covered by CUBLAS/CULA (in which case I encourage you to try them). When you get CUDA you will need to select the appropriate: driver, toolkit, and SDK (SDK is optional but useful.)

[tang]

Dear Dr. John,

Thank you very much for your so detailed explain. Do you mean that you suggest me to use the CULA? But from your document I didn't find the functions I listed above. Could you please tell me which library provide these functions?

BTW: maybe latter we will use the functions provided in CULA, such as GETRF, GEEV, but from the document I only fund the functions like "culaGetrf" and "culaGeev", they still can't replace the LAPACK library directly, at least we have to write some interface code to call these functions, am I right?

Thanks,
Zhanghong Tang

[dan]

[tang]

[tang]

[john]

[tang]

[tang]

[john]

[tang]