[C++] Cast to/from halffloat not implemented

[asfimport]

library(reticulate)
pa <- import("pyarrow")
pa$array(c(1,2,3))$cast(pa$float16())
#> Error in py_call_impl(callable, dots$args, dots$keywords): ArrowNotImplementedError: No cast implemented from double to halffloat
#> 
#> Detailed traceback: 
#> File "pyarrow/array.pxi", line 277, in pyarrow.lib.Array.cast (/Users/travis/build/BryanCutler/arrow-dist/arrow/python/build/temp.macosx-10.6-intel-3.6/lib.cxx:30459)
#> File "pyarrow/error.pxi", line 85, in pyarrow.lib.check_status (/Users/travis/build/BryanCutler/arrow-dist/arrow/python/build/temp.macosx-10.6-intel-3.6/lib.cxx:8570)

Reporter: Romain Francois / @romainfrancois
Watchers: Rok Mihevc / @rok

Related issues:

• [R] Implement halffloat -> R conversion (blocks)
• [C++] C++ array kernels framework and execution buildout (umbrella issue) (is a child of)
• [Python] Unable to cast HalffloatArray to Float or Double (supercedes)
• [C++] vendor a half precision floating point library (depends upon)

_{Note: This issue was originally created as ARROW-3802. Please see the migration documentation for further details.}

[asfimport]

Antoine Pitrou / @pitrou:
We lack an internal facility to interpret half-floats. Currently we merely see them as opaque 16-bit ints.

[asfimport]

Wes McKinney / @wesm:
TensorFlow uses the Eigen::half type to deal with float16. I presume that NumPy has its own implementations of things, since float16 is not implemented in hardware?

[asfimport]

Antoine Pitrou / @pitrou:
Yes, Numpy has dedicated routines.

[pitrou]

We should revive this now that we do have a half-float library available.