16-bit floating numbers for media and image processing
targeting embedded architecture and Vision-SoC
Origin
Recently, 16-bit floating point format has been introduced in the OpenEXP format and in the Cg language defined by NVIDIA.
This format, called 'half' is used by ILM for color coding, and by nVidia for 3D computations.
Half format has a smaller accuracy thant 32-bit FP number, but a smaller size too...
F16 at University Paris Sud
F16 format is now used to optimize some algoritms on RISC but also, and especially to make possible FP implementation possible on FPGA.
This research is done with two other labs:
Daniel Etiemble from LRI: for the original FPGA port and algorithms implementations and optimisations
Michel Kieffer from LSS: for Interval Computing and embedded localisation