Difference between revisions of "Teleconference 2015-04-08"
(Created page with "Attendees: Ken Moreland (SNL), Rob Maynard (Kitware), Hank Childs (OU), James Kress (OU), Chris Sewell (LANL), Dave Pugmire (ORNL) Teleconference 2015-03-25|Last teleconfer...") |
|||
| Line 2: | Line 2: | ||
[[Teleconference 2015-03-25|Last teleconference]] there was some discussion about whether the unstructured grid data model should group cells by type or leave them ungrouped. There are advantages and disadvantages to each approach. It was suggested to use the ungrouped version first. This is probably easier; it has fewer gotchas and is a more general solution. We can use it to find real performance issues (instead of speculating) and specialize as necessary. We should get more progress on this after SC submissions have passed. | [[Teleconference 2015-03-25|Last teleconference]] there was some discussion about whether the unstructured grid data model should group cells by type or leave them ungrouped. There are advantages and disadvantages to each approach. It was suggested to use the ungrouped version first. This is probably easier; it has fewer gotchas and is a more general solution. We can use it to find real performance issues (instead of speculating) and specialize as necessary. We should get more progress on this after SC submissions have passed. | ||
| + | |||
| + | The presentations as GTC went well. VTK-m was discussed in 2 sessions plus in a hands on lab. There appeared to be a lot of interest. The presentations are available on the GTC website. The hands-on session material can be made available on request. | ||
| + | |||
| + | Rob has put in a request to Google Summer of Code for a student to work on VTK-m. We won't know if that happens until Google announces the results, but if that happens Rob will mentor the student with Ken as a backup. | ||
| + | |||
| + | There is some talk about moving the VTK-m repo to Kitware's gitlab (much like VTK and other Kitware repositories). Generally this sounds like a good idea, but not particularly important right now. The question is when is a good time to make the transition so as to no interfere with development. Probably not during the summer when several students might be working on the project. | ||
| + | |||
| + | Rob is working on a native CUDA scheduler (as opposed to using thrust's parallel_for). This allows us to perform 3D ijk scheduling, which can save time used to convert flat indices to 3D indices. | ||
Latest revision as of 14:28, 8 April 2015
Attendees: Ken Moreland (SNL), Rob Maynard (Kitware), Hank Childs (OU), James Kress (OU), Chris Sewell (LANL), Dave Pugmire (ORNL)
Last teleconference there was some discussion about whether the unstructured grid data model should group cells by type or leave them ungrouped. There are advantages and disadvantages to each approach. It was suggested to use the ungrouped version first. This is probably easier; it has fewer gotchas and is a more general solution. We can use it to find real performance issues (instead of speculating) and specialize as necessary. We should get more progress on this after SC submissions have passed.
The presentations as GTC went well. VTK-m was discussed in 2 sessions plus in a hands on lab. There appeared to be a lot of interest. The presentations are available on the GTC website. The hands-on session material can be made available on request.
Rob has put in a request to Google Summer of Code for a student to work on VTK-m. We won't know if that happens until Google announces the results, but if that happens Rob will mentor the student with Ken as a backup.
There is some talk about moving the VTK-m repo to Kitware's gitlab (much like VTK and other Kitware repositories). Generally this sounds like a good idea, but not particularly important right now. The question is when is a good time to make the transition so as to no interfere with development. Probably not during the summer when several students might be working on the project.
Rob is working on a native CUDA scheduler (as opposed to using thrust's parallel_for). This allows us to perform 3D ijk scheduling, which can save time used to convert flat indices to 3D indices.
