Optionally set data type

[aidanheerdegen]

When reading in namelists I would like to be able to preserve the precision of the values being read in using the python decimal library. If numerical values are converted from string to a numerical type makes it more difficult to do this.

Options to support this might be to specify only strings to be returned when read in, or specify decimal.Decimal as one of the types to cast to.

[marshallward]

Reads ought to be in double precision, is that not happening?

[aidanheerdegen]

Yeah, but you can't represent some numbers perfectly with floating point, so you get weird results in some cases. See the linked payu issues.

[marshallward]

Oh, I see, you do not want to use base2 floating point...
That is a tall order, we might be better off doing some in-person discussion.

[aidanheerdegen]

Not possible to make adding to/modifying this array user settable through API?

https://github.com/marshallward/f90nml/blob/master/f90nml/parser.py#L725

[marshallward]

It may be possible, but a bit against the spirit of the module. For one thing, there's no way to indicate that the number is base10 in the namelist.

(Theoretically one could assign a base10 "kind"; however, kind cannot be assigned in a namelist, nor have I ever seen a compiler actually do this, so let's ignore that possibility).

So we would have to introduce new infrastructure which lets us pre-assign a type to this data, presumably with some sort of metadata during initialization of the Parser. Even then, we'd be storing this number in some format which Fortran does not understand, and going back to base2 would probably just create another problem.

Before considering this, would something like this work for you?

• read in the value at full precision: e.g. 0.10000000000000001

• Write it to a string in some truncated precision which represents your decimal accuracy (say 10 digits)

• pass this string to decimal.Decimal, do your processing, and generate whatever integer you are trying to get out of this thing.

Would something like that work? If not, then let's keep talking.

[marshallward]

BTW this has come up before, and I resisted it back then, but perhaps I should be more open to allowing users to add new types to that list, and providing some infrastructure to access them.

In my opinion, such an entity would no longer be a namelist, but I can see how it could be useful.

[aidanheerdegen]

Clearly you've thought this through before, and I can see what you're saying, this is input into a program which is using base2 floating point.

I'd say leave it and just point me back to this issue when I forget I've raised this before and whinge again!

[marshallward]

BTW it was not you who had suggested it before!

I am coming around to the idea of letting users add their own types, I will keep that one in mind.