Discussion Round: Desired Capabilities

[stvdwtt]

Emulated device backends to facilitate debugging: Is it feasible to have a host backend that emulates a device backend (e.g. CUDA) with a separate MemorySpace and the appropriate layout. This could make development/debugging easier for people with limited access to GPUs.

MDRange support for 1D Views: The Sundials developers tell me this is blocking them from full Kokkos support. Currently they only support 1D Views and do not want to handle the 1D and ND cases separately.

[simon-schlepphorst]

Double the number of ranks in Kokkos::Views: Probably a repeating request, once it's implemented. For high dimensional problems it would be easier for new users to start using Kokkos if Views would support more than 7 ranks. In the meantime some best practices to tackle these types of problems with the current capabilities without hurting the performance portability would be appreciated.

[dalg24]

Notes

(see MPI interop, FFT, and algorithms dedicated issues)

Misc

• Virtual device backend as a debugging facility: we have been thinking about it for a while. all attempts to provide it have failed so far.
• MDRange for Rank-1: Yes we will do it
• more than 7D view: probably possible, a good idea, but lots of work, don't expect it tomorrow
• deterministic (pseudo-)random number generator for Cuda: yes we could, getting good random numbers in this case is really hard, providing a bad one for debugging purposes is more realistic
  • https://xkcd.com/221/
• Memory pools: how to make Umpire available to kokkos applications
• Support kernel pipelining: interesting, what about more complex dataflows?
• SYCL style dependency detection: problem with that is that it kind of prevents encapsulation
• Per thread scratch memory for MDRange: could be expensive
• Access allocation label on GPU for debugging purpose: this could be extremely expensive, solution specific to debug would lead to different ABI for debug or non-debug ones
• Optimized indexing in many views with the same shape: maybe with mdspan where an external mapping could be shared