Using FP16 data in MATLAB

I have an embedded application that generates half precision (fp16) output that I would like to import and use in MATLAB. Would someone please advise me the easiest way to do that?

[Tony]

I should clarify, I'm storing the out put in binary file format, and reading them into MATLAB using fread (which doesn't support "half" as a precision. I can't seem to cast the binary data into a floating point number.

[James Tursa]

What version of MATLAB are you using?

Answer 1

[James Tursa]

If you have R2018b or later, you can fread as uint16 and then typecast into half type. E.g.,

% Generate some sample data
>> d = [-inf -pi 0 pi inf nan]
d =
      -Inf   -3.1416         0    3.1416       Inf       NaN
>> h = half(d)
h = 
  1×6 half row vector
      -Inf   -3.1406         0    3.1406       Inf       NaN
>> u = storedInteger(h)
u =
  1×6 uint16 row vector
   64512   49736       0   16968   31744   65024
% Convert the uint16 values to half precision
>> H = half.typecast(u)
H = 
  1×6 half row vector
      -Inf   -3.1406         0    3.1406       Inf       NaN
      

If you don't have R2018b or later, then the half type is not availalbe and you will be stuck with converting the values to single or double precision if you want to work with them in MATLAB. In that case, use fread as uint16 and then use the following FEX submission to turn the values into single or double:

https://www.mathworks.com/matlabcentral/fileexchange/23173-ieee-754r-half-precision-floating-point-converter

E.g.,

>> S = halfprecision(u,'single')
S =
  1×6 single row vector
      -Inf   -3.1406         0    3.1406       Inf       NaN

[Tony]

That works well, thank you!