task2: interface

[mzuzek]

Scope

Try to give common "feeling" to both interfaces 2a and 2b

2a. Execution Space

Blas kernels to cover:

• GEMV (WIP/KK team: Gemv unified interface kokkos/kokkos-kernels#1498)
• GEMM
• SCAL
• DOT

The objective would be do things like:

auto exec = ExecSpace(); // instance

 auto A, B, C;

template<class ExecSpace, ... >
 KokkosBlas::gemm( exec, A, B, C );

2b. Parallelization level dispatch

Have parallelization level (serial, team and team-vector) as a parameter - like ArgMode in: https://github.com/kokkos/kokkos-kernels/blob/develop/src/batched/dense/KokkosBatched_Gemm_Decl.hpp#L98-L119 (inspiration):

• Serial implementation should not use any parallelism nor dispatch to TPLs, so it can be called within any parallel context;
• Team implementation should take team member argument, be callable in a functor (same where TeamThreadRange) and use TeamThreadRange+ThreadVectorRange combination;
• TeamVector: WAIT for confirmation: should use ThreadVectorRange only and be callable from inside TeamThreadRange ? Should NOT use TeamVectorRange !

Blas kernels to cover:

• GEMV (already implemented on #1433 and #1435)
• GEMM (WIP: Move functor-level GEMM routines to KokkosBlas kokkos/kokkos-kernels#1519)
• SCAL
• DOT

[mzuzek]

Parallel Contexts

• TeamThreadRange: called inside TeamPolicy functor - maps to CPU threads and GPU warps.
  • ThreadVectorRange: called inside TeamThreadRange
• TeamVectorRange: combines/flattens TeamThreadRange and ThreadVectorRange (implemented by Introduce TeamVectorRange policy kokkos/kokkos#1227 for Feature Request: TeamVectorRange kokkos/kokkos#713)
  In the functor we can call

Note: ThreadVectorRange can be also called like TeamThreadRange (not inside it) - and then works like TeamVectorRange (which is probably better choice - TODO learn what's the difference)

See also HierarchicalParallelism 8.4 in Kokkos Wiki.

Kokkos hierarchy maps to hardware threads slightly different on CPU and GPU:

Kokkos level	CPU	GPU
Team	thread group	thread group (block)
Thread	CPU thread	thread group (warp)
Vector	SIMD / intrinsics	GPU thread