SYCL buffer pinning
BenchMem performance (-nb gpu -nsteps 2000 -nobackup -resethway -notunepme -ntomp 6 -ntmpi 1 -pin on), ns/day, median of three:
| before | after | |
|---|---|---|
| oneAPI/CUDA (GTX1660SUPER, IntelLLVM 0c7a1e18978) | 46 | 60 |
| hipSYCL/CUDA (GTX1660SUPER, hipSYCL 6aa58ce6 + Clang12) | 15.6 | 17.0 |
| oneAPI/OpenCL (XeMax, 2022.2) | 11.8 | 12.5 |
Supersedes !2572 (closed).
Refs #4522
Edited by Andrey Alekseenko