RFC 100 text: Add RFC for float16 support

doc/source/development/rfc/index.rst

@@ -106,3 +106,4 @@ RFC list
     rfc97_feature_and_fielddefn_sealing
     rfc98_build_requirements_gdal_3_9
     rfc99_geometry_coordinate_precision
+    rfc100_float16_support

doc/source/development/rfc/rfc100_float16_support.rst

@@ -0,0 +1,141 @@
+.. _rfc-100:
+
+=============================
+RFC 100: Support float16 type
+=============================
+
+============== =============================================
+Author:        Erik Schnetter
+Contact:       eschnetter @ perimeterinstitute.ca
+Started:       2024-Jun-05
+Status:        Proposed
+Target:        GDAL 3.11
+============== =============================================
+
+Summary
+-------
+
+This RFC adds support for the IEEE 16-bit floating point data type
+(aka ``half``, ``float16``). It adds new pixel data types
+``GDT_Float16`` and ``GDT_CFloat16``, backed by a new ``GFloat16`` C++
+type. This type will be emulated in software if not available
+natively.
+
+Some drivers, in particular Zarr, will be extended to support this
+type. Drivers that do not support this type will be checked to ensure
+they report an appropriate errors when necessary.
+
+Motivation
+----------
+
+- The float16 type is increasingly common. It is supported by GPUs, by
+  modern CPUs, and by modern compilers (e.g. GCC 12).
+
+- Using ``uint16`` values to shuffle ``float16`` values around is
+  inconvenient and awkward. There is no mechanism in GDAL to mark
+  attributes or datasets which ``uint16`` values should be interpreted
+  as ``float16``.
+
+- Some drivers (at least the HDF5, GTiff, and Zarr) already handle
+  float16 by exposing it as float32, using software conversion
+  routines. This type should be supported without converting to
+  float32.
+
+- C++23 will introduce native support for ``std::float16_t``. However,
+  it will likely be several years until C++23 will be a requirement
+  for GDAL. A shorter-term solution is needed.
+
+- Many other projects and storage libraries support float16. GDAL will
+  fall behind if it doesn't.
+
+Details
+-------
+
+A new type ``GFloat16`` will be defined in C++. This type will be a
+thin wrapper around a native float16 type if one is available, such as
+e.g. ``_Float16`` in modern versions of GCC. Otherwise, a this type
+will be emulated (transparently to the user), and operations will be
+performed as ``float``.
+
+The following pixel data types are added:
+- ``GDT_Float16``  --> ``_Float16``
+- ``GDT_CFloat16`` --> ``std::complex<_Float16``
+
+Some drivers (at least the HDF5, GTiff, and Zarr) already handle
+float16 by exposing it as float32, using software conversion routines.
+float16 is now supported directly, i.e., without converting to
+float32. In other drivers, the current behaviour is retained, which
+automatically converts float16 to float32.
+
+For simplicity there will be no new functions handling attributes for
+multidimensional datasets. Attributes of type float16 can still be
+read/written as raw values.
+
+Impacts in the code base
+------------------------
+
+It is likely that a large fraction of the code base will be impacted.
+
+SWIG bindings
+-------------
+
+The SWIG bindings are extended where possible. Unfortunately, SWIG
+does not support the native float16 Python type, but it does support
+the float16 numpy type. This means that not all SWIG Python wrappers
+can support float16.
+
+Backward compatibility
+----------------------
+
+C and C++ API and ABI are impacted. This design is backward-compatible
+manner, i.e. it does not break the ABI, and can thus be implemented in
+GDAL 3.x.
+
+Main impacts are:
+
+* Two new enum values for ``GDALDataType``
+
+* Some drivers already support float16 by converting from/to float32.
+  This capability remains in a backward-compatible manner. In
+  addition, these drivers might (if implemented) now also support
+  reading/writing float16 values from/to float16 arrays in memory.
+
+* C Multidimension API impacts (and equivalent C++ methods):
+
+  - :cpp:func:`GDALGroupCreateAttribute`
+  - :cpp:func:`GDALMDArrayRead`
+  - :cpp:func:`GDALMDArrayWrite`
+
+Risks
+-----
+
+The changes of this RFC add new features without removing or disabling
+any current features. The risk should be low.
+
+Documentation
+-------------
+
+Documentation will be added.
+
+Testing
+-------
+
+The C++ test suite will be updated. Tests will be implemented in C++
+and Python.
+
+Related issues and PRs
+----------------------
+
+- https://github.com/OSGeo/gdal/issues/10144: Supporting float16
+
+For comparison:
+
+- https://github.com/OSGeo/gdal/pull/5257: [FEATURE] Add (initial)
+  support Int64 and UInt64 raster data types
+
+No candidate implementation exists yet.
+
+Voting history
+--------------
+
+TBD