OpenCL/SYCL: Use sane cluster splitting on Intel GPUs
The same OpenCL/SYCL API call returns very different things on Intel and AMD/NVIDIA, which led to too much cluster splitting on Intel GPUs.
A quick test on a single XeMax, SYCL oneAPI 2022.2.0, water box with RF electrostatics:
Num atoms | Before, ns/day | After, ns/day |
---|---|---|
1 500 | 136 | 142 |
6 000 | 125 | 119 |
48 000 | 40 | 40 |
As expected, only relevant for small systems.
Edited by Andrey Alekseenko