Open syamajala opened 2 weeks ago
@elliottslaughter i put a reproducer for this bug on frontier here: /lustre/orion/cmb103/world-shared/seshuy/prefix
llvm 18 and rocm 6 do not fix this error. Think its a bug in Regent.
I have reproduced this on Frontier.
Because the CUDA version passes, this is either a bug in HIP-specific code generation or something of a mismatch when applying generic algorithms to HIP.
We're seeing S3D diverge when moving a task to gpu. The task uses
__parallel_prefix
and it looks the output changes. I will try to find a minimal reproducer.