Hi,
we have modified PoCL to enable offloading
Gromacs->SYCL->OpenCL using the PoCL cpu driver.
This required some minor changes on the Gromacs side and I would be interested in upstreaming those.
The changes are small, but I’m unsure about their potential impact on other backends.
Are the proposed changes best discussed in a pull request, or should I elaborate on them here first?
We have a developer call in 12 hours, and the next one in ~2 weeks, which is a natural place for such discussion. The agenda for this Wednesday is already finalized and is quite packed, but there’s always time for a quick chat.
Without knowing the scope of changes it’s hard to suggest the best way. But if they are “small” and, presumably, you already have the code, opening an MR at our GitLab seems fine without any further discussion.
For either MR or the live discussion, a list of hardware targeted (now and potentially) and some performance numbers would be nice to have, in addition to the patch itself. Hard to judge the code for accelerated computing not knowing how much it accelerates things.
Thanks, I will proceed with the MR then.
Currently, only the fallback-version of the kernel compiler of PoCL runs Gromacs so it’s not very efficient. We are hoping to bring the functionality to the more optimized default kernel compiler in near future.