Standards like OpenMP, OpenACC will stay and compile to different architectures.

https://www.openacc.org/get-started

Also Parallel STL for C++17, C++20

# Alpaka

Alpaka is still documented as unstable.

# Kokkos

Kokkos has the positive of an OpenMP backend. It might be enough to survive NVidia's monopolistic practices and seems stable.

- Last commit 6 months ago? Bad...

- Performance of big libraries like autodiff, BLAS or Tensorflow?

# SYCL

SYCL is a Khronos standard and it has the backing of Intel. It should generate good code (see TBB). And Intel is working to incorporate it Clang.

Slight negative: depends on OpenCL, not a priority inside Nvidia for a long time. That may have changed. also Khronos seems over bureaucratic.

# Others (that CMS will never consider)

- Radical option: Julia https://juliaacademy.com/p/parallel-computing

Parallelism defined in the language from 0.2. Tested in Tensorflow, SIMD, GPU...

- Swift: see swift tensorflow.

- Rust

. check Rust SIMD and

. and Weld (link to paper from Wld (NumPy, Rust and LLVM)

C++/Root never made an impact outside CERN, which helps to grow and gain traction. I would choose a pair language/library which shows promise outside CMS/Atlas.

raul

On 24/09/2019 14:18, David Colling wrote: