istream support for cfloat

[davidmallasen]

Hello,

When reading numbers from a file to the fp16 format using std::ifstream, I'm getting an error:

undefined reference to `std::istream& sw::universal::operator>><16u, 5u, unsigned short, true, false, false>(std::istream&, sw::universal::cfloat<16u, 5u, unsigned short, true, false, false>&)'

I think this could be related to

universal/include/universal/number/cfloat/cfloat_impl.hpp

Line 3151 in 4c57ec1

// istream input: TBD

Any insights on this? If this would be hard to implement, how could I hack the reading fp16 numbers from a file?

Thanks!

[ghost]

What is the format of the FP16 in the file?

WIthin Universal we have been serializing using a hex format that carries the configuration of the type as prefix. Problem with that is that it doesn't work efficiently for matrix dumps as the format prefix should be set at the matrix level. That hurdle we have yet to cross.

Any preference?

[davidmallasen]

Text. For now I'm reading into a double and then doing the casting down to a fp16 manually, so it's not a blocking issue, but it would be a nice addition.

[ghost]

can you give me an example of the text you are getting in? For an FP16, we can just leverage the default double input and then use the assignment operator to cast to the underlying cfloat<>

[ghost]

@davidmallasen I added the istream operator>> for cfloat in this commit 487c6ad

it takes the short cut of using the native operator>>(istream&, double) and then doing the conversion. The prototype was also incorrect, so fixed that too. It is in a new branch v3.70. If this works for you I can quickly merge it into main if needed.

[davidmallasen]

I tested it and it works for me. Thanks @theo-lemurian !

[Ravenwater]

TBD: we need to expand this to the big cfloats through a native implementation.