In less than 2 weeks we’ll meet for the LLVM-Dev Meeting, and on the workshop day, Oct 10, you can join us to discuss GPUs & Offloading!
We have some “speakers” already, but are looking mostly for open and constructive discussions.
If you want to present your project or interest in a short <= 5min talk, please reach out (or respond here). If you have topics for discussion, w/ or w/o presentation, please also reach out.
Scheduling details:
October 10th, 8:30 AM - 12:30 PM [GPU Offloading with LLVM]
(Session Details: 2023 LLVM Developers' Meeting)
(Rough and Preliminary) Agenda
- Welcome (@arsenm @jdoerfert)
- On the fly agenda additions.
- 4 x 5min lighting talks:
- 40 min open discussion
- 4 x 5min lighting talks:
- 40 min open discussion
- break
- 4 x 5min lighting talks:
- open discussion
Open discussion, potential topics:
- Bringing libcxx to the device, who’s up for it?
- Everyone hates language X, should we provide a generic “LLVM offloading” API/runtime?
- Unifying drivers: is the “new driver” good enough for everyone?
- Unifying compilation: Patching in AS late for NVIDIA, and other spurious divergences we should eliminate (see [RFC] Cleaning up the NVIDIA (and potentially AMD) GPU backend).
- Buildbots, and other ways of testing.
- Inclusion of “libdevice” in LLVM.
- Can we have a “target independent” GPU IR?
- Linking PTX with lld, anyone?
I’ll update this post as things evolve. Keep returning 