Tobias Ribizel

Some first results here from our Titan X vs. cuBLAS Tuning parameter is the oversubscription, i.e. number of launched warps / max number of active warps  For small inputs,...

I guess that is the crux of 2 vs. 3: Does solver->apply(...) have a read/write dependency on itself? With 3. we could avoid that, with 2 we need to consider...

I prefer 1., since it produces the clearest interface without any indirection, since `double` and `const LinOp*` are not really alternative types I would expect inside a Sum Type. Since...

Yes, it should be a replacement (as a feature-extended specialization), otherwise we get diamond inheritance.

Yes, anything interface-breaking would only be in the scope of 2.0, and I think we can do this without obvious downsides without the need to change EnableLinOp directly. Some types...

@MarcelKoch I like the idea! I'm just not sure whether it is better to be explicit with a different type, or do this implicitly based on value_type, which might be...

It seems we need to do some more work, we are not getting any speedup so far.

Just for completeness, can you tell us which CMake version you are using, and attach your CMakeCache.txt? Is the environment entirely Easybuild based?

I'm having a hard time reproducing your issue in a similar environment. What happens if you run `make ginkgo_device`? Does the same issue occur if you replace `make` by `ninja`?...

Thanks for the details, this is most likely an issue with our RPATH settings: https://stackoverflow.com/questions/69881222/why-rpath-check-in-cmake-deletes-executable