Encapsulate HDF5 module #5031
Encapsulate HDF5 module #5031
Conversation
Drop support for Intel Classic. Provide a clear error message when an unsupported compiler is detected. Also update C++ code to leverage C++20 and address build issues with Clang compiler toolchains.
Rewrite h5xx and hdf5 bindings such that they no longer propagate their compiler flags to the entire codebase.
| if (!mass_unit().empty() and (m_fields & H5MD_OUT_MASS)) { | ||
| h5xx::write_attribute(datasets["particles/atoms/mass/value"], "unit", | ||
| h5xx::write_attribute(datasets.at("particles/atoms/mass/value"), "unit", |
There was a problem hiding this comment.
Is there a benefit to making it a bound-checked access?
There was a problem hiding this comment.
Muscle memory. I didn't think too much about it.
Marginal benefit: if the key doesn't exist, a default-constructed hdf5 dataset is created and returned by reference by operator[]. Your file then contains a dataset with no elements. Since this method is called only once per simulation, at file creation, overhead of exceptions isn't relevant, and is probably smaller than default-constructing an empty dataset. Plus, we get notified that the map isn't properly initialized.
Whether we need bound-checked accesses in the hot loop (File::write()) is debatable. The h5xx module will be removed from ESPResSo eventually, in favor of a yet-to-be-determined C++ wrapper library for hdf5. Whether that new library will offer bound-checked access is likely, but not guaranteed.
Description of changes:
--recursive