this creates an uninitialized "compatible" evaluation that is
compatible with its argument. For primitive floating point types and
statically-sized Evaluations, this is identical to calling the default
constructor, for dynamically sized ones, it creates an uninitialized
Evaluation object of identical size as the argument.
thanks to [at]GitPaen for the heads up.
e.g., looping over the wrong range or an infinite loop. also, the
dense-AD unit test is shortend to test one specialization and the
unspecialized class.
this allows some flexibility of how an Evaluation object is
represented internally. the main motivation for this is to allow using
SIMD instruction which expect special data types as their operants.
`getValue(x)` should be better called `value(x)` but this leads to
really cryptic compiler errors which seem to be related to the fact
that the Evaluation class has a `value` attribute (this is for GCC
5.2.1).
with this, evaluation objects can be decayed via
`Opm::decay<Scalar>(x)`. since function templates do not require a
'template' qualifier, it is a bigger advantage for this function that
for the mathematical ones. (the conventional way to do this is way
more verbose:
```c++
Opm::MathToolbox<Evaluation>::template decay<Scalar>(x)
```
)
this PR was inspired by one of @atgeirr's recent comments. it allows
to get rid if the `MathToolbox<Eval>` detour for the dense AD code in
most cases: e.g. instead of writing
```c++
template <class Eval>
Eval f(const Eval& val)
{ return Opm::MathToolbox<Eval>::sqrt(val); }
```
one can simply write
```c++
template <class Eval>
Eval f(const Eval& val)
{ return Opm::sqrt(val); }
```
and it will work transparently with DenseAD `Evaluation` objects and
primitive floating point values.
one complication of this is that the type of the `Evaluation` object
does not need to be explicitly defined. for functions which take more
than one argument (like e.g. `pow()`, `min()` and `max()`), it is thus
assumed that the type of the result is the same as the type of the
first argument.
another drawback is that when both, the contents of the `Opm::` and
the `std::` namespaces are implicitly available, it is not clear to me
what's used for primitive floating point values. (it seems to compile
but IMO it could lead to surprising behaviour. thus, please only merge
if you consider the benefits of these convenience functions to be
greater than their drawbacks.)
Since "dense automatic differentiation" describes what this code is
all about much better in my opinion. ("Local AD" is just a possible
use case in the context of PDE discretization.)
