A rewritten Summary.cpp with some minor header modifications. Synposis
of the new implementation:
* Uses unordered_map< string, std::function > for dispatch, instead of
multiple functions and a switch
* Some poor man's function composition support has been added
(privately) to avoid a lot of reptition in the post processing.
* Functions assume they work over lists of wells instead of single wells
being special cased - this means groups of well etc. can share
implementation with single wells and field keywords.
* Unsupported keywords are not written in the Summary file.
Furthermore, some comments on special cases and overall approach and
a generally more declarative implementation. This change is invisible to
downstream developers. Users will obviously see no more garbage
keywords.
- Removed class Init completely - writeint an init file is handled by
the method EclipseWriter::writeINITFile( )
- Seperated writing of INIT file and EGRID file better.
Simulators might modify the grid post EclipseState creation, so the Grid
fetched from there is unreliable. Copy the Deck-provided grid and apply
the manipulations at EclipseWriter construction time to ensure it uses
the same dimensions and has the same properties as the simulator.
Ert is a hard dependency that provides much of the core functionality of
opm-output, whereas boost::filesystem only provided create-directory
features. Since Ert conveniently provides the same features with a
near-identical interface and is used for file-writing anyway, the
responsibility for interacting with the file system has been put on ert.
Boost is now only used for testing in opm-output.
Since production is seen as negative injection, and production rates
often are zero, negative zeros would be returned. Likewise, occasionally
numerical noise gives slightly-below-zero values which are now rounded
to be plain zeros.
It's perfectly legitimate for clients to "default" a well result, e.g.
when it hasn't started up yet, in which case most output will default to
0. Avoid crashing when this happens, and instead calculate based on a
dummy well.
Basing the dest durations on days over seconds make them less noisy,
more readable and closer to actual idiomatic use. Removes some subtle
complexities from the tests.
The to_si/from_si functions were moved into UnitSystem which now manages
table lookup for clients, rather than having to store the raw
conversion tables.
Reading the start time from EclipseState means it's no longer required
as a parameter to every method, and that the current time can be
calculated as start time + time elapsed.
This is-a relationship is never used, and the vtable is never leveraged.
The augmented output writers already use uniquely named output writers
anyway, so there is no reason to go via the hurdle of inheritance.
This gives a lot of flexibility when it comes to interface, since
Eclipse output can make Eclipse-specific assumptions.
The boost object was only used to calculate elapsed posix time anyway,
which is required (and already managed) by the callers. This makes the
interface slightly less complicated and removes an otherwise pointless
dependency.
