Can't find batched or/unmqr and or/ungqr

Open discussion for MAGMA library (Matrix Algebra on GPU and Multicore Architectures)

Can't find batched or/unmqr and or/ungqr

Postby berzjackson » Wed May 16, 2018 5:25 am

Dear MAGMA developers,

I have searched through the code of 2.3.0 and I can't seem to find batched version of or/unmqr and or/ungqr. The documentation pages are empty as well.

See:
http://icl.cs.utk.edu/projectsfiles/mag ... tched.html
http://icl.cs.utk.edu/projectsfiles/mag ... tched.html

Or maybe I'm missing something here?
berzjackson
 
Posts: 2
Joined: Wed May 16, 2018 5:05 am

Re: Can't find batched or/unmqr and or/ungqr

Postby mgates3 » Wed May 16, 2018 10:46 am

Nope, they don't exist. There is larfb batched, which is what they would be based on, so we can probably add them easily.

The documentation pages are empty placeholders. They should probably be hidden until the routines are available.

-mark
mgates3
 
Posts: 806
Joined: Fri Jan 06, 2012 2:13 pm

Re: Can't find batched or/unmqr and or/ungqr

Postby berzjackson » Tue Jun 05, 2018 6:50 am

Thank you for your reply!

I'm not familiar with larfb, but I will dig into the possibility of utilizing it to achieve what I want. Setting that possibility aside, I have some follow-up questions regarding QR factorization.

1.Are batched geqrf LAPACK compliant?

2.Are there ETA for batched or/unmqr and or/ungqr to be available in future releases?

3.Assuming they won't be ready in the near future, can I do QR factorization with batched version of geqrf and non-batched version of orgqr?
If I understand correctly, magma_dorgqr_gpu requires dT, and the user should be able to get it from an expert interface such as magma_dgeqrf_expert_batched. But looking at the docs, dT in magma_dgeqrf_expert_batched is different from dT in magma_dgeqrf_gpu and the doc for magma_dorgqr_gpu explicitly states that the dT passed in "must be the 6th argument of magma_dgeqrf_gpu". So maybe batched version of geqrf and non-batched version of orgqr can't work together due to this inconsistency?

4.If that is the case, another option is to use magma_dorgqr2 which does not need dT passed in explicitly but can be slower since T is recomputed by the CPU. Is that right?
berzjackson
 
Posts: 2
Joined: Wed May 16, 2018 5:05 am


Return to User discussion

Who is online

Users browsing this forum: No registered users and 2 guests

cron