bosilca

We never build without MPI, so the code is really not handling it very well. The rest of the warnings have simple fixes or can be ignores. I need to...

Right, a whole lot of code move for a small benefit. I find it strange that we now query a device capabilities in `parsec_cuda_all_devices_attached` instead of the place where the...

We tried hard to avoid any context level structures for managing tasks in order to avoid contentions between all the threads. This PR bypasses the scheduler and creates exactly what...

Then you dont need an additional mechanism, simply push the tasks back into the parsec scheduler. As long as the tasks are marked in the completion state, that should be...

Set the distance to 1 and the tasks shall not be pushed into the local queue. If that's not the case, the code needs to be fixed.

With distance 1 they should never be pushed into the GPU manager local queue, but hierarchically from there. But you are right they will eventually end in the global queue...

Unfortunately there is no fix for this issue, because the problem is not the `const` but the fact that for the outside world all components are seen as the base...

To add one point to this discussion, let's look at the potential benefits this huge effort could potentially bring. There is a reason we are creating these structures, to standardize...

Let me be extremely clear here, the bug is in the LTO implementation, it tries to outsmart smart humans. The `static-component.h` file is autogenerated and it is included, on purpose,...

We LOVE optimizations, we are writing code to optimize for micro-seconds, sometimes in assembly with [godbolt](https://godbolt.org/) by our side. So, you're preaching to the right crowd, but for the wrong...