No real idea if this is an issue for LLVM or for Mesa but on my r600g GPU(6520g) LLVM crashes when compiling an opencl kernel that uses either atomic_add or atomic_or. On LLVM 3.8 and mesa 11.2
Are the atomic_add/atomic_or operations causing the crash operating against local or global memory? Last I checked, I think that atomics for r600g-based cards were only implemented in LLVM for the local memory space (while SI+ have both local/global implemented). Someone may have gotten around to implementing support for global atomics for Evergreen/NI in LLVM while I wasn't looking, but I have no knowledge of it having been done.
it's for global memory. looks like it's unimplemented according to the galliumcompute chart. Oh well.
Use of freedesktop.org services, including Bugzilla, is subject to our Code of Conduct. How we collect and use information is described in our Privacy Policy.