Compute Accelerator Forum - Beams GPU & VecMem
Virtual
Zoom
To receive annuoncements and information about this forum please subscribe to compute-accelerator-forum-announce@cern.ch
Compute Accelerator Forum Live Notes
This is the LiveNotes for Compute Accelerator Forum meetings. Feel free to add observations on talks and to ask questions.
-
Please don’t delete things that other people add
-
Please make sure your name is next to the question, like this [Graeme]
Introduction
-
Thematic CERN School of Computing on heterogeneous programming is coming up, applications opening 9 Dec ‘21, https://indico.cern.ch/e/tCSC-2022 - deadline for applications is January 23rd, 2022 (Sebastian Lopienski)
Heterogeneous Memory Management with VecMem
-
Did you try to compare to standard managed memory (std::par parallelism)? Suggest to follow that evolution in the standards committee. (Jack Wells, )
-
No, I didn't try b/c of too much custom code.
-
-
Can you give details about “interfacing to Llama”? (Ben Morgan)
-
Try to make Llama use the pmr feature.
-
-
How does VecMem compare to Kokkos::views? (Martin Kwok)
-
Both implement fairly similar things. Reason for VecMem is on concentrating to mimic stl code on the device.
-
-
Llama tries to focus on data arrangement and tries to keep memory management out of the library (Bernhard Gruber)
Recent development and status of beam physics codes for heterogenous platform
-
Why do you do an abstraction on top of OpenCL, instead of using it as the only solution? (Attila)
-
Fear that OpenCL may disappear, at the moment seems not the case though. OpenCL does not give the control on the device that’s maybe needed in the future.
-
Cupy works nicely with large data structures. OpenCL is not comfortable for such use cases.
-
-
NVidia has a dropin for Numpy called cuNumeric ( https://developer.nvidia.com/cunumeric ) (Jack Wells)
-
Is there scope for collaboration for Beam simulation with other labs? (Ben Morgan)
-
Development started this year. Seeing interest outside the group
-