Mikael Simberg

Add thread sanitizer CI pipeline

24

Nothing to see here.

Workarounds for broken complex type operator overloads in HIP 5.7 onwards

5

Nothing to see here yet. This is an ugly, hacky attempt at not using the broken upstream operators, but I'm unsure if 1. it works correctly, 2. it generates the...

Add test stage for ROCm configuration

7

This won't work yet... - [ ] Add testing also for stdexec configuration

See https://github.com/ROCm/HIP/issues/3377#issuecomment-1853040754 for context. This would allow compiling with HIP 5.7.

It's currently hard-coded to only be enablable with CUDA, but it should in theory be possible to enable it with HIP as well. First this has to be done: https://github.com/eth-cscs/DLA-Future/issues/734.

Attempt to further split explicit template instantiations into separate files

1

e.g. by element type. One benefit would be to expose more parallelism in the build. A potential downside is even higher memory consumption during builds (which might anyway limit the...

Avoid pika `detail` functionality for GPU assertions

1

`assertExtendInfo` currently relies on implementation details of pika: https://github.com/eth-cscs/DLA-Future/blob/4754f309956d722128dd4962132b9ddf1d623bf3/include/dlaf/lapack/tile.h#L385-L393. Ideally this would be refactored to not use `detail` at all, but this might be difficult, at least if the same...

While investigating the slowdowns from newer HIP versions I found that the eigensolver miniapp fails immediately on the first iteration (consistently) with: ``` terminate called after throwing an instance of...