GPU Warp-Level Parallelism in LLVM/OpenMP

[@LeDabe]

What would be the speedups when on a loop, parallel is used and simd isnt but the SPMD optimization triggered vs parallel + simd without the SPMD optimization. That is, if I have code that does not use omp simd but works fine thanks to the parallel SPMD optimization, can I expect a speedup by using simd ?

[@EricWright-qc4sp]

Getting any benefit from simd is situational. Many programming models have a comparable concept (i.e Kokkos, Alpaca, OpenACC), and for all of them warp-level parallelism is code-dependent. In general: if you have a code with an inner loop with a relatively small number of iterations, and that loop cannot be folded into an outer loop (like with the collapse directive) then you would likely see a speedup using simd, regardless if it's SPMD or not.