You need to sign in or sign up before continuing.
Decide on explicit vs. implicit synchronization in SYCL
In SYCL, it is possible to avoid explicitly synchronizing computations and data copying by relying on the data-dependency graph and hoping the runtime is smart enough.
It does work ok for NBNXM, but there are the following concerns, raised by @pszilard:
- Divergence between CUDA/OpenCL/HIP and SYCL in code organization and task scheduling, potentially causing hard-to-catch bugs.
- Efficiency of the scheduling by the SYCL runtime.
Benchmarking and thorough analysis in light of the proposed GPU code merger is necessary.
Related: #3847, #3311 (closed) (#3020, #3323 (closed))