dbl_to_lofi.RdPack a double into the lower bits of an int32 with the given bits for sign, exponent and mantissa.
dbl_to_lofi(dbl, float_name = "bfloat16", float_bits = NULL) lofi_to_dbl(lofi, float_name = "bfloat16", float_bits = NULL)
| dbl | 64 bit R double |
|---|---|
| float_name | 'single', 'half', 'bfloat16'. Default: 'bfloat16' |
| float_bits | length (in number of bits) of sign, exponent and mantissa.
Default: NULL. If this value is not null, then it will override
anything the user may have specified for |
| lofi | low-bit representation |
32 bit integer with lower bits set to represent the quantized floating point value
By packing into a low fidelity bit representation you will definitely lose precision i.e. converting back into full 64 bit precision will not give you back the number you started with.
Packing a double into low-fidelity format has no explicit support for special values
such as NaN, NA or Inf. These values may get converted
to other numeric values or other special values. The result is undefined.
Operate on special values at your own risk.