[CP2K-user] [CP2K:15498] Plans for GPU support for libint/hfx operations?

[hut... at chem.uzh.ch]

Hi

yes, that is correct, forces and response are missing. We will also
have to optimize the GPU backend of the tensor library.

regards

Juerg Hutter
--------------------------------------------------------------
Juerg Hutter                         Phone : ++41 44 635 4491
Institut für Chemie C                FAX   : ++41 44 635 6838
Universität Zürich                   E-mail: hut... at chem.uzh.ch
Winterthurerstrasse 190
CH-8057 Zürich, Switzerland
---------------------------------------------------------------

-----cp... at googlegroups.com wrote: -----
To: "cp2k" <cp... at googlegroups.com>
From: "Nicholas Winner" 
Sent by: cp... at googlegroups.com
Date: 06/02/2021 06:42PM
Subject: Re: [CP2K:15498] Plans for GPU support for libint/hfx operations?

Thanks a lot Jurg, makes sense. RI-HFX seems like a great way to go. Looking at the source, it seems like RI-HFX is mostly implemented, its just the forces that need to be implemented. Is that right?

On Wednesday, June 2, 2021 at 4:34:39 AM UTC-7 jgh wrote:
Hi 
 
there are some preliminary plans to port the HFX integral to 
Fock matrix and gradient routines to GPUs. For the integral routines 
(libint or another equivalent library) we are hoping that some other 
project will contribute. 
Another route is the RI-HFX code. It is not yet in production mode, 
but would bring GPU acceleration to HFX. It moves computation from 
basic integrals to tensor contractions. This also should reduce 
memory requirements and improve parallelization. Through the 
usage of the DBCSR tensor library, we get GPU acceleration.  
 
best regards 
 
Juerg Hutter 
-------------------------------------------------------------- 
Juerg Hutter                         Phone : ++41 44 635 4491 
Institut für Chemie C                FAX   : ++41 44 635 6838 
Universität Zürich                   E-mail: h... at chem.uzh.ch 
Winterthurerstrasse 190 
CH-8057 Zürich, Switzerland 
--------------------------------------------------------------- 
 
-----c... at googlegroups.com wrote: ----- 
To: "cp2k" <c... at googlegroups.com> 
From: "Nicholas Winner"  
Sent by: c... at googlegroups.com 
Date: 06/02/2021 12:29AM 
Subject: [CP2K:15495] Plans for GPU support for libint/hfx operations? 
 
I was wondering if the devs have plans for pushing the expensive HFX integrals/derivatives onto GPUs in the near future? There has been lots of recent discussion about GPUs that I have seen, and now DBSCR, CP2K+ELPA for diagonalization, and grid collocate/integrate are are GPU compatible, but the big one would be hfx (for myself at least). 
 
Are there plans for this, are their practical issues with implementing it beyond finding the time to write the modules? 
 
Just curious, thanks. 
 
-Nick   
  --  
 You received this message because you are subscribed to the Google Groups "cp2k" group. 
 To unsubscribe from this group and stop receiving emails from it, send an email to cp... at googlegroups.com. 
 To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/33345f64-fa80-4706-ae3b-2a3384cb9e84n%40googlegroups.com. 
  
  
  -- 
 You received this message because you are subscribed to the Google Groups "cp2k" group.
 To unsubscribe from this group and stop receiving emails from it, send an email to cp... at googlegroups.com.
 To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/53a09fa0-1e05-422b-a94a-91171b2e2c8fn%40googlegroups.com.