Add block formats #8
Conversation
src/gfloat/types.py
Outdated
| @@ -201,7 +206,12 @@ def min(self) -> float: | |||
| """ | |||
| The smallest representable number, typically ``-max``. | |||
| """ | |||
| return -self.max | |||
| if self.is_signed: | |||
| return -self.max | |||
There was a problem hiding this comment.
Is this true for format_info_ocp_int8? I can't find a test for it, and it can be asymmetric (max == 1.0 + 63.0 / 64, min == -2.0).
There was a problem hiding this comment.
There is a test that code_of_min decodes to min, and code_of_min is "self-evidently" 2**k-1, so it should be fine, but... I have now added a dense test to this PR which exhaustively enumerates all values and ensures min, max, etc are consistent.
There was a problem hiding this comment.
I forgot to mention: this PR does include OCP int8, but I let it decode 0x80 as a NaN, which is allowed by [at least one reading of] the spec...
"Integer data types use a 2’s complement encoding, but the maximum negative representation
(−2) may be left unused to maintain symmetry between the maximum positive and negative
representations and avoid introducing a negative bias."
But maybe that is to be read as "0x80 should decode to -2, and downstream users can choose not to use it...."
There was a problem hiding this comment.
I'll do it that way, as the spec says "N/A" for NaN, but presumably means "not available" rather than "not applicable".
Add block formats (e.g. OCP MX)
Add OCP E8M0 type. This is an unsigned format, so we add
is_signedtoFormatInfo.Add OCP INT8 type. This represents the significand as twos-complement, so we add
is_twos_complementtoFormatInfo.Docs preview at https://gfloat--8.org.readthedocs.build/en/8/