Teleconference 2017-02-08
Attendees: Ken Moreland (SNL), Tom Fogal (NVIDIA), Tom Otahal (SNL), Ollie Lo (LANL), Brett Lesley (UO), Hank Childs (UO), Dave Pugmire (ORNL), Mark Kim (ORNL), David Rogers (LANL), Rob Maynard (Kitware), David Lonie (Kitware), Alberto Villarreal (Intel)
Tom O working on the vector magnitude filter. Cleaned up warning messages. Probably ready for merge.
Ollie picked up Chris' work on unify memory. Adding an allocator to use cuda managed allocate. Created a thrust device vector that allocates all memory as managed. Use the managed memory instead of GPU memory. Most of the tests work. Performance for unified memory ranges from on par to 4x slower. Managed memory is about 2x as the streaming example. But the streaming code only support simple examples. Rob mentioned that recent versions of the device adapter have gotten rid of using thrust vectors. This should make Ollie's life easier. It might also speed things up since thrust vectors like to zero out arrays.
UO folks are gearing up to implement streamlines under ECP. Will probably get going early March when funding is in. Will be work in collaboration with ORNL.
Brett has been looking at graph algorithms like maximal clique within VTK-m. Brett needs an ability to have an array of variable length Vecs. Ken suggested the ArrayHandleGroupVecVariable, which allows you to specify such a think at runtime.
Dave P looking at the uniform grid streamline worklet as a starting point for their streamline work.
Dave L added a special array handle that is the equivalent to /dev/null. It discards anything written to it in case you don't care about an output. Also been reorganizing the exception classes.
Gitlab has been updated to support LFS.
Alberto working with VTK-m code on Xeon Phi. Coordinating with Tom Otahal.
We had a brief conversation about having another VTK-m code sprint. It looks like the beginning of April works well for most people. Hank will look into hosting at U of Oregon @ Eugene. We will look into space the week of April 4 or perhaps push that back a week if space is more available.
