Use a std::array instead of a C-style array for tensor elements. #16474
Conversation
| ar &values; | ||
| if constexpr (rank_ > 1) | ||
| ar &values; | ||
| else | ||
| ar &boost::serialization::make_array(&values[0], dim); |
There was a problem hiding this comment.
Let me add here that we could have used the first case also for the std::array in the base case (rank==1) but that I expose the array explicitly because I would like to replace the std::array in the base case by a VectorizedArray and I don't want to change the serialization output again at that point.
(VectorizedArray does not have the infrastructure to serialize itself, but in any case, I can only use it when available and need to fall back to std::array otherwise. This would lead to platform-dependent serialization output, which I can avoid by just describing the elements here as an explicit array.)
There was a problem hiding this comment.
I am looking forward to that - albeit I would note that we need to look out for some places, e.g. for what happens with Tensor<1, dim, VectorizedArray<double>>. (It should just work fine, because VectorizedArray<VectorizedArray<double>> is valid; nonetheless, we might need to have a closer look there.)
There was a problem hiding this comment.
Right. std::conditional is what is necessary, and I know how to do it :-)
|
I checked the library size (admittedly, I have a few PRs from this afternoon in my difference, but nothing is really substantial ); the debug library went from 641.6 MB to 642.9 MB, the release library even decreased from 303.95 MB to 303.93 MB. So nothing worth worrying, 👍 for the improvement. |
This gets rid of the awkwardness of the C-style array which we need to make sure has size greater than zero. This changes all of the serialization output.
See #16465.