Mihai Preda

When closing this issue (when fixed), it would be nice to leave a link here to the LLVM commit with the fix, thanks!

I didn't see this bug with recent ROCm, including ROCm 6.0.0. I think it's been fixed for a while now. Closing.

I hit the same issue with most recent ROCm

It looks like there are too many unrelated changes in this PR; I'm not inclined to merge it as-is. Maybe some small individual fixes can be extracted as separate PR.

Propose a cute logo, will be considered.

For PRPLL, the logo theme is a purple cat :)

Closing this as I'm not investigating OpenCL 1.2 support ATM.

The above happens when using barrier(CLK_LOCAL_MEM_FENCE); If instead barrier(0) is used, the generated ISA is as expected: ``` s_waitcnt lgkmcnt(0) s_barrier ``` So, IMO, something is wrong with barrier(CLK_LOCAL_MEM_FENCE) as...

@jamesxu2 I'm unable to dump OpenCL ISA the way I used to due to #2940 (as soon as -save-temps works again I'll confirm the behavior of the most recent ROCm...

@jamesxu2 you mentioned > The first s_waitcnt combined with the barrier would seem to implicitly guarantee the consistency of local memory (but not global memory) But s_waitcnt lgkmcnt(0) is concerned...