Description
On nvidia, a pointer being member of a class is triggering a page fault. This does not occur if a local pointer is set on this member pointer right before the parallel region, i.e. the member pointer is more or less aliased locally.
- see discussion in forum https://forums.developer.nvidia.com/t/nvc-25-1-stdpar-htod-for-already-fetched-member-variable/324618
- is this also happening on shared memory architecture of AMD ?
AMD (e.g. Hunter)
- visualize with old version : https://ui.perfetto.dev/v46.0-35b3d9845/#!
NVIDIA (AIA / JUWELS Booster)
- visualize on aia
fe2
with/opt/nvidia/nsight-compute/2025.1.0/ncu-ui post/aia_report.nsys-rep