geqrf basic question

[jinyan]

Hi, I'm new to CULA, can someone explain to me the parameters on Sgeqrf?

m - # of rows in A
n - # of columns in A
a - pointer to matrix A
Ida - leading dimension of A, so Ida = m?
tau - scalar factors of elementary reflectors??

So if I call culaSgeqrf(...), where does the output Q and R matrices get stored?

Another question, what's the difference between culaSgeqrf and culaDeviceSgeqrf? If I use culaDeviceSgeqrf, do I have to allocate memory space/copy memory to GPU, like CUDA?

Sorry for the silly questions, the CULA programming guide wasn't very beginner-friendly.

Thanks in advance!

[kyle]

CULA's QR decomposition is implemented using Householder reflections. Check out the Wikipedia page for some decent information if you are unfamiliar with the algorithm.

m - # of rows in A
n - # of columns in A
a - pointer to matrix A
Ida - leading dimension of A, so Ida = m?
tau - scalar factors of elementary reflectors??

LDA is typically M, but doesn't have to be. TAU is used to construct (or multiply) the Q matrix after xGEQRF.

So if I call culaSgeqrf(...), where does the output Q and R matrices get stored?

After xGEQRF, R is stored in the upper triangular portion of A. Q can then be generated using xORGQR with A and TAU as the inputs. Alternatively, Q can multiplied directly with another matrix using xORMQR with A, TAU, and C as the inputs. This two step method is typical and will be seen in any package based on the LAPACK interface. When you call QR in MATLAB for example, it calls both of these functions behinds the scenes. Also, it's worth nothing that xORGQR and xORMQR are only available in CULA Premium.

Another question, what's the difference between culaSgeqrf and culaDeviceSgeqrf? If I use culaDeviceSgeqrf, do I have to allocate memory space/copy memory to GPU, like CUDA?

culaSgeqrf expects host memory and culaDeviceSgeqrf expects device memory. We always recommend the standard host interface because it's simpler and uses special allocation methods that maximizes memory throughput.

Sorry for the silly questions, the CULA programming guide wasn't very beginner-friendly.

Not a problem! The guide is more focused at developers who are familiar with LAPACK notation but new to GPU computing. Let us know if you have any other questions.

[jinyan]

Thanks for the help, Kyle!

One more question, I'm getting this error while trying to run the executable

error while loading shared libraries: libcula.so: cannot open shared object file: No such file or directory

I saw your reply on the other post "make sure to include...in LD_LIBRARY_PATH", I tried looking around on google and still couldn't figure out how to change LD_LIBRARY_PATH. Can you help? I'm fairly new to linux too

[kyle]

To add CULA to your library path, try the following command.

Code: Select all

export LD_LIBRARY_PATH=/usr/local/cula/lib64:$LD_LIBRARY_PATH


Change to 'lib' if you are on a 32-bit system.

[john]

You can add that to your .bashrc if you want to make it permanent.

[jinyan]

woohoo! okay program is working, well, only for square matrices right now. the code I wrote for CPU calculation is row-major, so the column-major implementation in CULA is causing some errors in rectangular matrices. I'll fix that later.

Another question, when I try to include cutil.h to use the cut...Timer functions, it says cutil.h cannot be found. How can i tell the program to look for the header in CUDA SDK folder?

[kyle]

Check your compiler for documentation on how to search for additional include directories.

In GCC it's:

Code: Select all

-I dir

If you are using cutil.h, remember you'll have to link against the cutil library as well.

[john]

If you would prefer to avoid cutil, you'll find some simple timing code in the CULA examples/benchmark folder.

[jinyan]

Thanks again. I ended up using clock(), it seems to be the more precise out of the two (clock() vs time()).

Is there a function in CULA that will allow me to efficiently transpose a matrix? I couldn't find one in manual.

[john]

I'm afraid there isn't. If your matrix is square or if you are willing to do an out-of-place transpose, the CUDA code is pretty simple and very fast. The Nvidia GPU SDK has an okay implementation.