Casting between widths

Widening is exact

Every sf8 value is also an sf16 value (the code is shifted left by 8). Widening therefore never loses anything and never raises. It happens automatically when you mix widths:

import sf

sf.sf8(0.5) + sf.sf16(0.25)        # sf16(0.75)
sf.cast(sf.sf8(0.1), sf.sf16)      # sf16(0.1015625): the sf8 rounding is kept
sf.sf16(sf.sf8(0.5))               # sf16(0.5)

Note the second line: sf.sf8(0.1) is 13/128 = 0.1015625, and widening preserves that, not the original 0.1.

Narrowing must be explicit

Going to fewer bits re-quantizes: the value is rounded to the nearest code of the narrower format (ties to even), and if that rounding overflows or underflows, the event goes through the policy.

sf.cast(sf.sf16(0.1), sf.sf8)      # sf8(0.1015625)
sf.cast(sf.sf16(0.999), sf.sf8)
# SFOverflowError: sf16(0.998992919921875) is not representable in sf8 (max 0.9921875)
sf.cast(sf.sf16(0.001), sf.sf8)    # sf8(0.0)   SFUnderflowWarning

Narrowing never happens implicitly. Passing a wider value where a narrower one is required raises SFTypeError:

sf.sf8(sf.sf16(0.5))
# SFTypeError: cannot store sf16 into sf8 implicitly; use sf.cast(x, sf.sf8)

This is deliberate: a silent narrowing could drop precision or saturate somewhere you did not expect.

Arrays

sf.cast(arr, dtype) and arr.astype(dtype) do the same element by element, reporting events with counts. Arrays: creating →

Which widths exist

All of sf4 to sf16 on the host. The device accepts sf16 only; placing a narrower array on a device raises SFTypeError.

Get started

Concepts

Scalars

Arrays

Errors

Testing

Widening is exact

Narrowing must be explicit

Arrays

Which widths exist

​Widening is exact

​Narrowing must be explicit

​Arrays

​Which widths exist

Widening is exact

Narrowing must be explicit

Arrays

Which widths exist