Thomas Herault

Exploring this issue, I see the following behavior: - `__parsec_execute` selects first the RECURSIVE hook for POTRF(0) - During the preamble (generated code) of the RECURSIVE hook, we transfer the...

We looked at the code during the code review meeting, and we found a series of issues that need addressing: - the recursive device should make use of the evaluate()...

My understanding of CMP0104 is that if we want the same level of portability as the local nvcc provides, we should set `CUDA_ARCHITECTURES` to `OFF` if it is not set...

> Mhh, this happens when PaRSEC is built as part of TTG, not when I try to build it by itself. Strange. I think it's because we enable the CUDA...

I'm using the issue to keep notes... OK, after digging in the code, here is the situation: - when the task is selected for execution by a CPU scheduler (e.g....

Yes, I know. The confusion here is that every task (including tasks that will eventually run on the GPU) start to exist in the CPU scheduler queues, and are logged...

The root of the problem is that in recursive, we try to `parsec_taskpool_destroy(tp)` in the callback of the termination detector for `tp`. To allow this behvior, we should terminate (and...

I think adding noexcept is a good idea, and probably necessary to have a high performance implementation of TTG. I'm not sure about allowing tasks to raise exceptions (although, tasks...

If we agree on 1. (require process-wide unique IDs for cross-thread matching events), we should add a check in dbpreader.c that will issue a warning or a fatal error if...

It looks like we need to support both kinds of semantics then, and that only the user can decide if this type of event is cross-thread matching or per-thread matching....